fix: queen lifecycle

Added methods to control tabs via the Chrome extension:
- create_tab(groupId, url) - create and navigate tabs in user's Chrome
- close_tab(tabId) - close tabs
- list_tabs(groupId?) - list tabs
- cdp_attach(tabId) - attach CDP for automation
- cdp_send(tabId, method, params) - send CDP commands

These enable browser automation through the extension when Playwright
can't connect directly to the user's Chrome.

Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>

.claude/settings.local.json

@@ -1,44 +0,0 @@
-{
-  "permissions": {
-    "allow": [
-      "Bash(npm install:*)",
-      "Bash(npm test:*)",
-      "Skill(building-agents-construction)",
-      "Skill(building-agents-construction:*)",
-      "Bash(PYTHONPATH=core:exports pytest:*)",
-      "mcp__agent-builder__create_session",
-      "mcp__agent-builder__get_session_status",
-      "mcp__agent-builder__set_goal",
-      "mcp__agent-builder__list_mcp_servers",
-      "mcp__agent-builder__test_node",
-      "mcp__agent-builder__add_node",
-      "mcp__agent-builder__add_edge",
-      "mcp__agent-builder__validate_graph",
-      "Bash(ruff check:*)",
-      "Bash(PYTHONPATH=core:exports python:*)",
-      "mcp__agent-builder__list_tests",
-      "mcp__agent-builder__generate_constraint_tests",
-      "Bash(python -m agent:*)",
-      "Bash(python agent.py:*)",
-      "Bash(python -c:*)",
-      "Bash(done)",
-      "Bash(xargs cat:*)",
-      "mcp__agent-builder__list_mcp_tools",
-      "mcp__agent-builder__add_mcp_server",
-      "mcp__agent-builder__check_missing_credentials",
-      "mcp__agent-builder__store_credential",
-      "mcp__agent-builder__list_stored_credentials",
-      "mcp__agent-builder__delete_stored_credential",
-      "mcp__agent-builder__verify_credentials",
-      "Bash(PYTHONPATH=/home/timothy/oss/hive/core:/home/timothy/oss/hive/exports python:*)",
-      "Bash(PYTHONPATH=core:exports:tools/src python -m hubspot_input:*)",
-      "mcp__agent-builder__export_graph",
-      "Bash(python3:*)"
-    ]
-  },
-  "enableAllProjectMcpServers": true,
-  "enabledMcpjsonServers": [
-    "agent-builder",
-    "tools"
-  ]
-}

.claude/settings.local.json.example

@@ -0,0 +1,16 @@
+{
+  "permissions": {
+    "allow": [
+      "Bash(git status:*)",
+      "Bash(gh run view:*)",
+      "Bash(uv run:*)",
+      "Bash(env:*)",
+      "Bash(python -m py_compile:*)",
+      "Bash(python -m pytest:*)",
+      "Bash(source:*)",
+      "Bash(find:*)",
+      "Bash(PYTHONPATH=core:exports:tools/src uv run pytest:*)"
+    ]
+  },
+  "enabledMcpjsonServers": ["tools"]
+}

.claude/skills/agent-workflow/SKILL.md

@@ -1,463 +0,0 @@
----
-name: agent-workflow
-description: Complete workflow for building, implementing, and testing goal-driven agents. Orchestrates building-agents-* and testing-agent skills. Use when starting a new agent project, unsure which skill to use, or need end-to-end guidance.
-license: Apache-2.0
-metadata:
-  author: hive
-  version: "2.0"
-  type: workflow-orchestrator
-  orchestrates:
-    - building-agents-core
-    - building-agents-construction
-    - building-agents-patterns
-    - testing-agent
-    - setup-credentials
----
-
-# Agent Development Workflow
-
-Complete Standard Operating Procedure (SOP) for building production-ready goal-driven agents.
-
-## Overview
-
-This workflow orchestrates specialized skills to take you from initial concept to production-ready agent:
-
-1. **Understand Concepts** → `/building-agents-core` (optional)
-2. **Build Structure** → `/building-agents-construction`
-3. **Optimize Design** → `/building-agents-patterns` (optional)
-4. **Setup Credentials** → `/setup-credentials` (if agent uses tools requiring API keys)
-5. **Test & Validate** → `/testing-agent`
-
-## When to Use This Workflow
-
-Use this meta-skill when:
-- Starting a new agent from scratch
-- Unclear which skill to use first
-- Need end-to-end guidance for agent development
-- Want consistent, repeatable agent builds
-
-**Skip this workflow** if:
-- You only need to test an existing agent → use `/testing-agent` directly
-- You know exactly which phase you're in → use specific skill directly
-
-## Quick Decision Tree
-
-```
-"Need to understand agent concepts" → building-agents-core
-"Build a new agent" → building-agents-construction
-"Optimize my agent design" → building-agents-patterns
-"Set up API keys for my agent" → setup-credentials
-"Test my agent" → testing-agent
-"Not sure what I need" → Read phases below, then decide
-"Agent has structure but needs implementation" → See agent directory STATUS.md
-```
-
-## Phase 0: Understand Concepts (Optional)
-
-**Duration**: 5-10 minutes
-**Skill**: `/building-agents-core`
-**Input**: Questions about agent architecture
-
-### When to Use
-
-- First time building an agent
-- Need to understand node types, edges, goals
-- Want to validate tool availability
-- Learning about pause/resume architecture
-
-### What This Phase Provides
-
-- Architecture overview (Python packages, not JSON)
-- Core concepts (Goal, Node, Edge, Pause/Resume)
-- Tool discovery and validation procedures
-- Workflow overview
-
-**Skip this phase** if you already understand agent fundamentals.
-
-## Phase 1: Build Agent Structure
-
-**Duration**: 15-30 minutes
-**Skill**: `/building-agents-construction`
-**Input**: User requirements ("Build an agent that...")
-
-### What This Phase Does
-
-Creates the complete agent architecture:
-- Package structure (`exports/agent_name/`)
-- Goal with success criteria and constraints
-- Workflow graph (nodes and edges)
-- Node specifications
-- CLI interface
-- Documentation
-
-### Process
-
-1. **Create package** - Directory structure with skeleton files
-2. **Define goal** - Success criteria and constraints written to agent.py
-3. **Design nodes** - Each node approved and written incrementally
-4. **Connect edges** - Workflow graph with conditional routing
-5. **Finalize** - Agent class, exports, and documentation
-
-### Outputs
-
-- ✅ `exports/agent_name/` package created
-- ✅ Goal defined in agent.py
-- ✅ 3-5 success criteria defined
-- ✅ 1-5 constraints defined
-- ✅ 5-10 nodes specified in nodes/__init__.py
-- ✅ 8-15 edges connecting workflow
-- ✅ Validated structure (passes `python -m agent_name validate`)
-- ✅ README.md with usage instructions
-- ✅ CLI commands (info, validate, run, shell)
-
-### Success Criteria
-
-You're ready for Phase 2 when:
-- Agent structure validates without errors
-- All nodes and edges are defined
-- CLI commands work (info, validate)
-- You see: "Agent complete: exports/agent_name/"
-
-### Common Outputs
-
-The building-agents-construction skill produces:
-```
-exports/agent_name/
-├── __init__.py          (package exports)
-├── __main__.py          (CLI interface)
-├── agent.py             (goal, graph, agent class)
-├── nodes/__init__.py    (node specifications)
-├── config.py            (configuration)
-├── implementations.py   (may be created for Python functions)
-└── README.md            (documentation)
-```
-
-### Next Steps
-
-**If structure complete and validated:**
-→ Check `exports/agent_name/STATUS.md` or `IMPLEMENTATION_GUIDE.md`
-→ These files explain implementation options
-→ You may need to add Python functions or MCP tools (not covered by current skills)
-
-**If want to optimize design:**
-→ Proceed to Phase 1.5 (building-agents-patterns)
-
-**If ready to test:**
-→ Proceed to Phase 2
-
-## Phase 1.5: Optimize Design (Optional)
-
-**Duration**: 10-15 minutes
-**Skill**: `/building-agents-patterns`
-**Input**: Completed agent structure
-
-### When to Use
-
-- Want to add pause/resume functionality
-- Need error handling patterns
-- Want to optimize performance
-- Need examples of complex routing
-- Want best practices guidance
-
-### What This Phase Provides
-
-- Practical examples and patterns
-- Pause/resume architecture
-- Error handling strategies
-- Anti-patterns to avoid
-- Performance optimization techniques
-
-**Skip this phase** if your agent design is straightforward.
-
-## Phase 2: Test & Validate
-
-**Duration**: 20-40 minutes
-**Skill**: `/testing-agent`
-**Input**: Working agent from Phase 1
-
-### What This Phase Does
-
-Creates comprehensive test suite:
-- Constraint tests (verify hard requirements)
-- Success criteria tests (measure goal achievement)
-- Edge case tests (handle failures gracefully)
-- Integration tests (end-to-end workflows)
-
-### Process
-
-1. **Analyze agent** - Read goal, constraints, success criteria
-2. **Generate tests** - Create pytest files in `exports/agent_name/tests/`
-3. **User approval** - Review and approve each test
-4. **Run evaluation** - Execute tests and collect results
-5. **Debug failures** - Identify and fix issues
-6. **Iterate** - Repeat until all tests pass
-
-### Outputs
-
-- ✅ Test files in `exports/agent_name/tests/`
-- ✅ Test report with pass/fail metrics
-- ✅ Coverage of all success criteria
-- ✅ Coverage of all constraints
-- ✅ Edge case handling verified
-
-### Success Criteria
-
-You're done when:
-- All tests pass
-- All success criteria validated
-- All constraints verified
-- Agent handles edge cases
-- Test coverage is comprehensive
-
-### Next Steps
-
-**Agent ready for:**
-- Production deployment
-- Integration into larger systems
-- Documentation and handoff
-- Continuous monitoring
-
-## Phase Transitions
-
-### From Phase 1 to Phase 2
-
-**Trigger signals:**
-- "Agent complete: exports/..."
-- Structure validation passes
-- README indicates implementation complete
-
-**Before proceeding:**
-- Verify agent can be imported: `from exports.agent_name import default_agent`
-- Check if implementation is needed (see STATUS.md or IMPLEMENTATION_GUIDE.md)
-- Confirm agent executes without import errors
-
-### Skipping Phases
-
-**When to skip Phase 1:**
-- Agent structure already exists
-- Only need to add tests
-- Modifying existing agent
-
-**When to skip Phase 2:**
-- Prototyping or exploring
-- Agent not production-bound
-- Manual testing sufficient
-
-## Common Patterns
-
-### Pattern 1: Complete New Build (Simple)
-
-```
-User: "Build an agent that monitors files"
-→ Use /building-agents-construction
-→ Agent structure created
-→ Use /testing-agent
-→ Tests created and passing
-→ Done: Production-ready agent
-```
-
-### Pattern 1b: Complete New Build (With Learning)
-
-```
-User: "Build an agent (first time)"
-→ Use /building-agents-core (understand concepts)
-→ Use /building-agents-construction (build structure)
-→ Use /building-agents-patterns (optimize design)
-→ Use /testing-agent (validate)
-→ Done: Production-ready agent
-```
-
-### Pattern 2: Test Existing Agent
-
-```
-User: "Test my agent at exports/my_agent"
-→ Skip Phase 1
-→ Use /testing-agent directly
-→ Tests created
-→ Done: Validated agent
-```
-
-### Pattern 3: Iterative Development
-
-```
-User: "Build an agent"
-→ Use /building-agents-construction (Phase 1)
-→ Implementation needed (see STATUS.md)
-→ [User implements functions]
-→ Use /testing-agent (Phase 2)
-→ Tests reveal bugs
-→ [Fix bugs manually]
-→ Re-run tests
-→ Done: Working agent
-```
-
-### Pattern 4: Complex Agent with Patterns
-
-```
-User: "Build an agent with multi-turn conversations"
-→ Use /building-agents-core (learn pause/resume)
-→ Use /building-agents-construction (build structure)
-→ Use /building-agents-patterns (implement pause/resume pattern)
-→ Use /testing-agent (validate conversation flows)
-→ Done: Complex conversational agent
-```
-
-## Skill Dependencies
-
-```
-agent-workflow (meta-skill)
-    │
-    ├── building-agents-core (foundational)
-    │   ├── Architecture concepts
-    │   ├── Node/Edge/Goal definitions
-    │   ├── Tool discovery procedures
-    │   └── Workflow overview
-    │
-    ├── building-agents-construction (procedural)
-    │   ├── Creates package structure
-    │   ├── Defines goal
-    │   ├── Adds nodes incrementally
-    │   ├── Connects edges
-    │   ├── Finalizes agent class
-    │   └── Requires: building-agents-core
-    │
-    ├── building-agents-patterns (reference)
-    │   ├── Best practices
-    │   ├── Pause/resume patterns
-    │   ├── Error handling
-    │   ├── Anti-patterns
-    │   └── Performance optimization
-    │
-    └── testing-agent
-        ├── Reads agent goal
-        ├── Generates tests
-        ├── Runs evaluation
-        └── Reports results
-```
-
-## Troubleshooting
-
-### "Agent structure won't validate"
-
-- Check node IDs match between nodes/__init__.py and agent.py
-- Verify all edges reference valid node IDs
-- Ensure entry_node exists in nodes list
-- Run: `PYTHONPATH=core:exports python -m agent_name validate`
-
-### "Agent has structure but won't run"
-
-- Check for STATUS.md or IMPLEMENTATION_GUIDE.md in agent directory
-- Implementation may be needed (Python functions or MCP tools)
-- This is expected - building-agents-construction creates structure, not implementation
-- See implementation guide for completion options
-
-### "Tests are failing"
-
-- Review test output for specific failures
-- Check agent goal and success criteria
-- Verify constraints are met
-- Use `/testing-agent` to debug and iterate
-- Fix agent code and re-run tests
-
-### "Not sure which phase I'm in"
-
-Run these checks:
-
-```bash
-# Check if agent structure exists
-ls exports/my_agent/agent.py
-
-# Check if it validates
-PYTHONPATH=core:exports python -m my_agent validate
-
-# Check if tests exist
-ls exports/my_agent/tests/
-
-# If structure exists and validates → Phase 2 (testing)
-# If structure doesn't exist → Phase 1 (building)
-# If tests exist but failing → Debug phase
-```
-
-## Best Practices
-
-### For Phase 1 (Building)
-
-1. **Start with clear requirements** - Know what the agent should do
-2. **Define success criteria early** - Measurable goals drive design
-3. **Keep nodes focused** - One responsibility per node
-4. **Use descriptive names** - Node IDs should explain purpose
-5. **Validate incrementally** - Check structure after each major addition
-
-### For Phase 2 (Testing)
-
-1. **Test constraints first** - Hard requirements must pass
-2. **Mock external dependencies** - Use mock mode for LLMs/APIs
-3. **Cover edge cases** - Test failures, not just success paths
-4. **Iterate quickly** - Fix one test at a time
-5. **Document test patterns** - Future tests follow same structure
-
-### General Workflow
-
-1. **Use version control** - Git commit after each phase
-2. **Document decisions** - Update README with changes
-3. **Keep iterations small** - Build → Test → Fix → Repeat
-4. **Preserve working states** - Tag successful iterations
-5. **Learn from failures** - Failed tests reveal design issues
-
-## Exit Criteria
-
-You're done with the workflow when:
-
-✅ Agent structure validates
-✅ All tests pass
-✅ Success criteria met
-✅ Constraints verified
-✅ Documentation complete
-✅ Agent ready for deployment
-
-## Additional Resources
-
-- **building-agents-core**: See `.claude/skills/building-agents-core/SKILL.md`
-- **building-agents-construction**: See `.claude/skills/building-agents-construction/SKILL.md`
-- **building-agents-patterns**: See `.claude/skills/building-agents-patterns/SKILL.md`
-- **testing-agent**: See `.claude/skills/testing-agent/SKILL.md`
-- **Agent framework docs**: See `core/README.md`
-- **Example agents**: See `exports/` directory
-
-## Summary
-
-This workflow provides a proven path from concept to production-ready agent:
-
-1. **Learn** with `/building-agents-core` → Understand fundamentals (optional)
-2. **Build** with `/building-agents-construction` → Get validated structure
-3. **Optimize** with `/building-agents-patterns` → Apply best practices (optional)
-4. **Test** with `/testing-agent` → Get verified functionality
-
-The workflow is **flexible** - skip phases as needed, iterate freely, and adapt to your specific requirements. The goal is **production-ready agents** built with **consistent, repeatable processes**.
-
-## Skill Selection Guide
-
-**Choose building-agents-core when:**
-- First time building agents
-- Need to understand architecture
-- Validating tool availability
-- Learning about node types and edges
-
-**Choose building-agents-construction when:**
-- Actually building an agent
-- Have clear requirements
-- Ready to write code
-- Want step-by-step guidance
-
-**Choose building-agents-patterns when:**
-- Agent structure complete
-- Need advanced patterns
-- Implementing pause/resume
-- Optimizing performance
-- Want best practices
-
-**Choose testing-agent when:**
-- Agent structure complete
-- Ready to validate functionality
-- Need comprehensive test coverage
-- Debugging agent behavior

.claude/skills/agent-workflow/examples/file-monitor-example.md

@@ -1,199 +0,0 @@
-# Example: File Monitor Agent
-
-This example shows the complete agent-workflow in action for building a file monitoring agent.
-
-## Initial Request
-
-```
-User: "Build an agent that monitors ~/Downloads and copies new files to ~/Documents"
-```
-
-## Phase 1: Building (20 minutes)
-
-### Step 1: Create Structure
-
-Agent invokes `/building-agents` skill and:
-
-1. Creates `exports/file_monitor_agent/` package
-2. Writes skeleton files (__init__.py, __main__.py, agent.py, etc.)
-
-**Output**: Package structure visible immediately
-
-### Step 2: Define Goal
-
-```python
-goal = Goal(
-    id="file-monitor-copy",
-    name="Automated File Monitor & Copy",
-    success_criteria=[
-        # 100% detection rate
-        # 100% copy success
-        # 100% conflict resolution
-        # >99% uptime
-    ],
-    constraints=[
-        # Preserve originals
-        # Handle errors gracefully
-        # Track state
-        # Respect permissions
-    ]
-)
-```
-
-**Output**: Goal written to agent.py
-
-### Step 3: Design Nodes
-
-7 nodes approved and written incrementally:
-
-1. `initialize-state` - Set up tracking
-2. `list-downloads` - Scan directory
-3. `identify-new-files` - Find new files
-4. `check-for-new-files` - Router
-5. `copy-files` - Copy with conflict resolution
-6. `update-state` - Mark as processed
-7. `wait-interval` - Sleep between cycles
-
-**Output**: All nodes in nodes/__init__.py
-
-### Step 4: Connect Edges
-
-8 edges connecting the workflow loop:
-
-```
-initialize → list → identify → check
-                                ↓  ↓
-                              copy  wait
-                                ↓    ↑
-                              update ↓
-                                ↓    ↓
-                              wait → list (loop)
-```
-
-**Output**: Edges written to agent.py
-
-### Step 5: Finalize
-
-```bash
-$ PYTHONPATH=core:exports python -m file_monitor_agent validate
-✓ Agent is valid
-
-$ PYTHONPATH=core:exports python -m file_monitor_agent info
-Agent: File Monitor & Copy Agent
-Nodes: 7
-Edges: 8
-```
-
-**Phase 1 Complete**: Structure validated ✅
-
-### Status After Phase 1
-
-```
-exports/file_monitor_agent/
-├── __init__.py          ✅ (exports)
-├── __main__.py          ✅ (CLI)
-├── agent.py             ✅ (goal, graph, agent class)
-├── nodes/__init__.py    ✅ (7 nodes)
-├── config.py            ✅ (configuration)
-├── implementations.py   ✅ (Python functions)
-├── README.md            ✅ (documentation)
-├── IMPLEMENTATION_GUIDE.md ✅ (next steps)
-└── STATUS.md            ✅ (current state)
-```
-
-**Note**: Implementation gap exists - data flow needs connection (covered in STATUS.md)
-
-## Phase 2: Testing (25 minutes)
-
-### Step 1: Analyze Agent
-
-Agent invokes `/testing-agent` skill and:
-
-1. Reads goal from `exports/file_monitor_agent/agent.py`
-2. Identifies 4 success criteria to test
-3. Identifies 4 constraints to verify
-4. Plans test coverage
-
-### Step 2: Generate Tests
-
-Creates test files:
-
-```
-exports/file_monitor_agent/tests/
-├── conftest.py              (fixtures)
-├── test_constraints.py      (4 constraint tests)
-├── test_success_criteria.py (4 success tests)
-└── test_edge_cases.py       (error handling)
-```
-
-Tests approved incrementally by user.
-
-### Step 3: Run Tests
-
-```bash
-$ PYTHONPATH=core:exports pytest exports/file_monitor_agent/tests/
-
-test_constraints.py::test_preserves_originals     PASSED
-test_constraints.py::test_handles_errors          PASSED
-test_constraints.py::test_tracks_state            PASSED
-test_constraints.py::test_respects_permissions    PASSED
-
-test_success_criteria.py::test_detects_all_files  PASSED
-test_success_criteria.py::test_copies_all_files   PASSED
-test_success_criteria.py::test_resolves_conflicts PASSED
-test_success_criteria.py::test_continuous_run     PASSED
-
-test_edge_cases.py::test_empty_directory          PASSED
-test_edge_cases.py::test_permission_denied        PASSED
-test_edge_cases.py::test_disk_full                PASSED
-test_edge_cases.py::test_large_files              PASSED
-
-========================== 12 passed in 3.42s ==========================
-```
-
-**Phase 2 Complete**: All tests pass ✅
-
-## Final Output
-
-**Production-Ready Agent:**
-
-```bash
-# Run the agent
-./RUN_AGENT.sh
-
-# Or manually
-PYTHONPATH=core:exports:tools/src python -m file_monitor_agent run
-```
-
-**Capabilities:**
-- Monitors ~/Downloads continuously
-- Copies new files to ~/Documents
-- Resolves conflicts with timestamps
-- Handles errors gracefully
-- Tracks processed files
-- Runs as background service
-
-**Total Time**: ~45 minutes from concept to production
-
-## Key Learnings
-
-1. **Incremental building** - Files written immediately, visible throughout
-2. **Validation early** - Structure validated before moving to implementation
-3. **Test-driven** - Tests reveal real behavior
-4. **Documentation included** - README, STATUS, and guides auto-generated
-5. **Repeatable process** - Same workflow for any agent type
-
-## Variations
-
-**For simpler agents:**
-- Fewer nodes (3-5 instead of 7)
-- Simpler workflow (linear instead of looping)
-- Faster build time (10-15 minutes)
-
-**For complex agents:**
-- More nodes (10-15+)
-- Multiple subgraphs
-- Pause/resume points for human-in-the-loop
-- Longer build time (45-60 minutes)
-
-The workflow scales to your needs!

.claude/skills/browser-edge-cases/SKILL.md

@@ -0,0 +1,241 @@
+---
+name: browser-edge-cases
+description: SOP for debugging browser automation failures on complex websites. Use when browser tools fail on specific sites like LinkedIn, Twitter/X, SPAs, or sites with Shadow DOM.
+license: MIT
+---
+
+# Browser Tool Edge Cases
+
+Standard Operating Procedure for debugging and fixing browser automation failures on complex websites.
+
+## When to Use This Skill
+
+- `browser_scroll` succeeds but page doesn't move
+- `browser_click` succeeds but no action triggered
+- `browser_type` text disappears or doesn't work
+- `browser_snapshot` hangs or returns stale content
+- `browser_navigate` loads wrong content
+
+## SOP: Debugging Browser Tool Failures
+
+### Phase 1: Reproduce & Isolate
+
+```
+1. Create minimal test case demonstrating failure
+2. Test against simple site (example.com) to verify tool works
+3. Test against problematic site to confirm issue
+```
+
+**Quick isolation test:**
+```python
+# Test 1: Does the tool work at all?
+await browser_navigate(tab_id, "https://example.com")
+result = await browser_scroll(tab_id, "down", 100)
+# Should work on simple sites
+
+# Test 2: Does it fail on the problematic site?
+await browser_navigate(tab_id, "https://linkedin.com/feed")
+result = await browser_scroll(tab_id, "down", 100)
+# If this fails but example.com works → site-specific edge case
+```
+
+### Phase 2: Analyze Root Cause
+
+**Step 2a: Check console for errors**
+```python
+console = await browser_console(tab_id)
+# Look for: CSP violations, React errors, JavaScript exceptions
+```
+
+**Step 2b: Inspect DOM structure**
+```python
+html = await browser_html(tab_id)
+snapshot = await browser_snapshot(tab_id)
+# Look for:
+# - Nested scrollable divs (overflow: scroll/auto)
+# - Shadow DOM roots
+# - iframes
+# - Custom widgets
+```
+
+**Step 2c: Identify the pattern**
+
+| Symptom | Likely Cause | Check |
+|---------|--------------|-------|
+| Scroll doesn't move | Nested scroll container | Look for `overflow: scroll` divs |
+| Click no effect | Element covered | Check `getBoundingClientRect` vs viewport |
+| Type clears | Autocomplete/React | Check for event listeners on input |
+| Snapshot hangs | Huge DOM | Check node count in snapshot |
+| Snapshot stale | SPA hydration | Wait after navigation |
+
+### Phase 3: Implement Multi-Layer Fix
+
+**Pattern: Always have fallbacks**
+
+```python
+async def robust_operation(tab_id):
+    # Method 1: Primary approach
+    try:
+        result = await primary_method(tab_id)
+        if verify_success(result):
+            return result
+    except Exception:
+        pass
+
+    # Method 2: CDP fallback
+    try:
+        result = await cdp_fallback(tab_id)
+        if verify_success(result):
+            return result
+    except Exception:
+        pass
+
+    # Method 3: JavaScript fallback
+    return await javascript_fallback(tab_id)
+```
+
+**Pattern: Always add timeouts**
+
+```python
+# Bad - can hang forever
+result = await browser_snapshot(tab_id)
+
+# Good - fails fast with useful error
+try:
+    result = await browser_snapshot(tab_id, timeout_s=10.0)
+except asyncio.TimeoutError:
+    # Handle timeout gracefully
+    result = await fallback_snapshot(tab_id)
+```
+
+### Phase 4: Verify Fix
+
+```
+1. Run against problematic site → should work
+2. Run against simple site → should still work (regression check)
+3. Document in registry.md
+```
+
+## Pattern Library
+
+### P1: Nested Scrollable Containers
+
+**Sites:** LinkedIn, Twitter/X, any SPA with scrollable feeds
+
+**Detection:**
+```javascript
+// Find largest scrollable container
+const candidates = [];
+document.querySelectorAll('*').forEach(el => {
+    const style = getComputedStyle(el);
+    if (style.overflow.includes('scroll') || style.overflow.includes('auto')) {
+        const rect = el.getBoundingClientRect();
+        if (rect.width > 100 && rect.height > 100) {
+            candidates.push({el, area: rect.width * rect.height});
+        }
+    }
+});
+candidates.sort((a, b) => b.area - a.area);
+return candidates[0]?.el;
+```
+
+**Fix:** Dispatch scroll events at container's center, not viewport center.
+
+### P2: Element Covered by Overlay
+
+**Sites:** Modals, tooltips, SPAs with loading overlays
+
+**Detection:**
+```javascript
+const rect = element.getBoundingClientRect();
+const centerX = rect.left + rect.width / 2;
+const centerY = rect.top + rect.height / 2;
+const topElement = document.elementFromPoint(centerX, centerY);
+return topElement === element || element.contains(topElement);
+```
+
+**Fix:** Wait for overlay to disappear, or use JavaScript click.
+
+### P3: React Synthetic Events
+
+**Sites:** React SPAs, modern web apps
+
+**Detection:** If CDP click doesn't trigger handler but manual click works.
+
+**Fix:** Use JavaScript click as primary:
+```javascript
+element.click();
+```
+
+### P4: Huge DOM / Accessibility Tree
+
+**Sites:** LinkedIn, Facebook, Twitter (feeds with 1000s of nodes)
+
+**Detection:**
+```javascript
+document.querySelectorAll('*').length > 5000
+```
+
+**Fix:**
+1. Add timeout to snapshot operation
+2. Truncate tree at 2000 nodes
+3. Fall back to DOM-based snapshot if accessibility tree too large
+
+### P5: SPA Hydration Delay
+
+**Sites:** React, Vue, Angular SPAs after navigation
+
+**Detection:**
+```javascript
+// Check if React app has hydrated
+document.querySelector('[data-reactroot]') ||
+document.querySelector('[data-reactid]')
+```
+
+**Fix:** Wait for specific selector after navigation:
+```python
+await browser_navigate(tab_id, url, wait_until="load")
+await browser_wait(tab_id, selector='[data-testid="content"]', timeout_ms=5000)
+```
+
+### P6: Shadow DOM
+
+**Sites:** Components using Shadow DOM, Lit elements
+
+**Detection:**
+```javascript
+document.querySelectorAll('*').some(el => el.shadowRoot)
+```
+
+**Fix:** Pierce shadow root:
+```javascript
+function queryShadow(selector) {
+    const parts = selector.split('>>>');
+    let node = document;
+    for (const part of parts) {
+        if (node.shadowRoot) {
+            node = node.shadowRoot.querySelector(part.trim());
+        } else {
+            node = node.querySelector(part.trim());
+        }
+    }
+    return node;
+}
+```
+
+## Quick Reference
+
+| Issue | Primary Fix | Fallback |
+|-------|-------------|----------|
+| Scroll not working | Find scrollable container | Mouse wheel at container center |
+| Click no effect | JavaScript click() | CDP mouse events |
+| Type clears | Add delay_ms | Use execCommand |
+| Snapshot hangs | Add timeout_s | DOM snapshot fallback |
+| Stale content | Wait for selector | Increase wait_until timeout |
+| Shadow DOM | Pierce selector | JavaScript traversal |
+
+## References
+
+- [registry.md](registry.md) - Full list of known edge cases
+- [scripts/test_case.py](scripts/test_case.py) - Template for testing new cases
+- [BROWSER_USE_PATTERNS.md](../../tools/BROWSER_USE_PATTERNS.md) - Implementation patterns from browser-use

.claude/skills/browser-edge-cases/registry.md

@@ -0,0 +1,261 @@
+# Browser Edge Case Registry
+
+Curated list of known browser automation edge cases with symptoms, causes, and fixes.
+
+---
+
+## Scroll Issues
+
+### #1: LinkedIn Nested Scroll Container
+
+| Attribute | Value |
+|-----------|-------|
+| **Site** | LinkedIn (linkedin.com/feed) |
+| **Symptom** | `browser_scroll()` returns `{ok: true}` but page doesn't move |
+| **Root Cause** | Content is in a nested scrollable div (`overflow: scroll`), not the main window |
+| **Detection** | `document.querySelectorAll('*')` with `overflow: scroll/auto` has large candidates |
+| **Fix** | JavaScript finds largest scrollable container, uses `container.scrollBy()` |
+| **Code** | `bridge.py:808-891` - smart scroll with container detection |
+| **Verified** | 2026-04-03 ✓ |
+
+### #2: Twitter/X Lazy Loading
+
+| Attribute | Value |
+|-----------|-------|
+| **Site** | Twitter/X (x.com) |
+| **Symptom** | Infinite scroll doesn't load new content |
+| **Root Cause** | Lazy loading requires content to be visible before loading more |
+| **Detection** | Scroll position at bottom but no new `[data-testid="tweet"]` elements |
+| **Fix** | Add `wait_for_selector` between scroll calls with 1s delay |
+| **Code** | Test file: `tests/test_x_page_load_repro.py` |
+| **Verified** | - |
+
+### #3: Modal/Dialog Scroll Container
+
+| Attribute | Value |
+|-----------|-------|
+| **Site** | Any site with modal dialogs |
+| **Symptom** | Scroll scrolls background page, not modal content |
+| **Root Cause** | Modal has its own scroll container with `overflow: scroll` |
+| **Detection** | Visible element with `position: fixed` and scrollable content |
+| **Fix** | Find visible modal container (highest z-index scrollable), scroll that |
+| **Code** | - |
+| **Verified** | - |
+
+---
+
+## Click Issues
+
+### #4: Element Covered by Overlay
+
+| Attribute | Value |
+|-----------|-------|
+| **Site** | SPAs, sites with loading overlays |
+| **Symptom** | Click succeeds but no action triggered |
+| **Root Cause** | Element is covered by transparent overlay, tooltip, or iframe |
+| **Detection** | `document.elementFromPoint(x, y) !== target` |
+| **Fix** | Wait for overlay to disappear, or use JavaScript `element.click()` |
+| **Code** | `bridge.py:394-591` - JavaScript click as primary |
+| **Verified** | - |
+
+### #5: React Synthetic Events
+
+| Attribute | Value |
+|-----------|-------|
+| **Site** | React applications |
+| **Symptom** | CDP click doesn't trigger React handler |
+| **Root Cause** | React uses synthetic events that don't respond to CDP events |
+| **Detection** | Site uses React (check for `__reactFiber$` or `data-reactroot`) |
+| **Fix** | Use JavaScript `element.click()` as primary method |
+| **Code** | `bridge.py:394-591` - JavaScript-first click |
+| **Verified** | - |
+
+### #6: Shadow DOM Elements
+
+| Attribute | Value |
+|-----------|-------|
+| **Site** | Components using Shadow DOM, Lit elements |
+| **Symptom** | `querySelector` can't find element |
+| **Root Cause** | Element is inside a shadow root, not main DOM tree |
+| **Detection** | `element.shadowRoot !== null` on parent elements |
+| **Fix** | Use piercing selector (`host >>> target`) or traverse shadow roots |
+| **Code** | See SKILL.md P6 pattern |
+| **Verified** | 2026-04-03 ✓ |
+
+---
+
+## Input Issues
+
+### #7: ContentEditable / Rich Text Editors
+
+| Attribute | Value |
+|-----------|-------|
+| **Site** | Rich text editors (Notion, Slack web, etc.) |
+| **Symptom** | `browser_type()` doesn't insert text |
+| **Root Cause** | Element is `contenteditable`, not an `<input>` or `<textarea>` |
+| **Detection** | `element.contentEditable === 'true'` |
+| **Fix** | Focus via JavaScript, use `execCommand('insertText')` or `Input.dispatchKeyEvent` |
+| **Code** | `bridge.py:616-694` - contentEditable handling |
+| **Verified** | 2026-04-03 ✓ |
+
+### #8: Autocomplete Field Clearing
+
+| Attribute | Value |
+|-----------|-------|
+| **Site** | Search fields with autocomplete, address forms |
+| **Symptom** | Typed text gets cleared immediately |
+| **Root Cause** | Field expects realistic keystroke timing for autocomplete |
+| **Detection** | Field has autocomplete listeners or dropdown appears |
+| **Fix** | Add `delay_ms=50` between keystrokes |
+| **Code** | `bridge.py:type()` - delay_ms parameter |
+| **Verified** | 2026-04-03 ✓ |
+
+### #9: Custom Date Pickers
+
+| Attribute | Value |
+|-----------|-------|
+| **Site** | Forms with custom date widgets |
+| **Symptom** | Can't type date into date field |
+| **Root Cause** | Custom widget intercepts and blocks keyboard input |
+| **Detection** | Typing doesn't change field value |
+| **Fix** | Click calendar widget icon, select date from dropdown |
+| **Code** | - |
+| **Verified** | - |
+
+---
+
+## Snapshot Issues
+
+### #10: LinkedIn Huge DOM Tree
+
+| Attribute | Value |
+|-----------|-------|
+| **Site** | LinkedIn, Facebook, Twitter feeds |
+| **Symptom** | `browser_snapshot()` hangs forever |
+| **Root Cause** | 10k+ DOM nodes, accessibility tree has 50k+ nodes |
+| **Detection** | `document.querySelectorAll('*').length > 5000` |
+| **Fix** | Add `timeout_s` param with `asyncio.timeout()`, proper error handling |
+| **Code** | `bridge.py:1041-1028` - snapshot with timeout protection |
+| **Verified** | 2026-04-03 ✓ (0.08s on LinkedIn) |
+
+### #11: SPA Hydration Delay
+
+| Attribute | Value |
+|-----------|-------|
+| **Site** | React/Vue/Angular SPAs |
+| **Symptom** | Snapshot shows old content after navigation |
+| **Root Cause** | Client-side hydration hasn't completed when snapshot runs |
+| **Detection** | `document.readyState === 'complete'` but content missing |
+| **Fix** | Wait for specific selector after navigation |
+| **Code** | Test file: `tests/test_x_page_load_repro.py` |
+| **Verified** | - |
+
+### #12: iframe Content Missing
+
+| Attribute | Value |
+|-----------|-------|
+| **Site** | Sites with embedded content |
+| **Symptom** | Snapshot missing iframe content |
+| **Root Cause** | Accessibility tree doesn't include iframe content |
+| **Detection** | `document.querySelectorAll('iframe')` has results |
+| **Fix** | Use `DOM.getFrameOwner` + separate snapshot for each iframe |
+| **Code** | - |
+| **Verified** | - |
+
+---
+
+## Navigation Issues
+
+### #13: SPA Navigation Events
+
+| Attribute | Value |
+|-----------|-------|
+| **Site** | React Router, Vue Router SPAs |
+| **Symptom** | `wait_until="load"` fires before content ready |
+| **Root Cause** | SPA uses client-side routing, no full page load |
+| **Detection** | URL changes but `load` event already fired |
+| **Fix** | Use `wait_until="networkidle"` or `wait_for_selector` |
+| **Code** | `bridge.py:navigate()` - wait_until options |
+| **Verified** | - |
+
+### #14: Cross-Origin Redirects
+
+| Attribute | Value |
+|-----------|-------|
+| **Site** | OAuth flows, SSO logins |
+| **Symptom** | Navigation fails during redirect |
+| **Root Cause** | Cross-origin security prevents CDP tracking |
+| **Detection** | URL changes to different domain |
+| **Fix** | Use `wait_for_url` with pattern matching instead of exact URL |
+| **Code** | - |
+| **Verified** | - |
+
+---
+
+## Screenshot Issues
+
+### #15: Selector Screenshot Not Implemented
+
+| Attribute | Value |
+|-----------|-------|
+| **Site** | Any site |
+| **Symptom** | `browser_screenshot(selector="h1")` takes full viewport instead of element |
+| **Root Cause** | `selector` param existed in signature but was silently ignored in both `bridge.py` and `inspection.py` |
+| **Detection** | Screenshot with selector same byte size as screenshot without selector |
+| **Fix** | Use CDP `Runtime.evaluate` to call `getBoundingClientRect()` on the element, pass result as `clip` to `Page.captureScreenshot` |
+| **Code** | `bridge.py:1315-1344` - selector clip logic; `inspection.py:94-96` - pass selector to bridge |
+| **Verified** | 2026-04-03 ✓ (JS rect query returns correct viewport coords; requires server restart) |
+
+### #16: Stale Browser Context (Group ID Mismatch)
+
+| Attribute | Value |
+|-----------|-------|
+| **Site** | Any |
+| **Symptom** | `browser_open()` returns `"No group with id: XXXXXXX"` even though `browser_status` shows `running: true` |
+| **Root Cause** | In-memory `_contexts` dict has a stale `groupId` from a Chrome tab group that was closed outside the tool (e.g. user closed the tab group) |
+| **Detection** | `browser_status` returns `running: true` but `browser_open` fails with "No group with id" |
+| **Fix** | Call `browser_stop()` to clear stale context from `_contexts`, then `browser_start()` again |
+| **Code** | `tools/lifecycle.py:144-160` - `already_running` check uses cached dict without validating against Chrome |
+| **Verified** | 2026-04-03 ✓ |
+
+---
+
+## How to Add New Edge Cases
+
+1. **Reproduce** the issue with minimal test case
+2. **Document** using the template below
+3. **Implement** fix with multi-layer fallback
+4. **Verify** against both problematic and simple sites
+5. **Submit** by appending to this file
+
+### Template
+
+```markdown
+### #N: [Short Title]
+
+| Attribute | Value |
+|-----------|-------|
+| **Site** | [URL or site type] |
+| **Symptom** | [What the user observes] |
+| **Root Cause** | [Technical explanation] |
+| **Detection** | [JavaScript to detect this case] |
+| **Fix** | [Solution approach] |
+| **Code** | [File:line reference if implemented] |
+| **Verified** | [Date or "pending"] |
+```
+
+---
+
+## Statistics
+
+| Category | Count |
+|----------|-------|
+| Scroll Issues | 3 |
+| Click Issues | 3 |
+| Input Issues | 3 |
+| Snapshot Issues | 3 |
+| Navigation Issues | 2 |
+| Screenshot Issues | 2 |
+| **Total** | **16** |
+
+Last updated: 2026-04-03

.claude/skills/browser-edge-cases/scripts/test_02_twitter_scroll.py

@@ -0,0 +1,111 @@
+#!/usr/bin/env python
+"""
+Test #2: Twitter/X Lazy Loading Scroll
+
+Symptom: Infinite scroll doesn't load new content
+Root Cause: Lazy loading requires content to be visible before loading more
+Fix: Add wait_for_selector between scroll calls
+"""
+
+import asyncio
+import sys
+import time
+from pathlib import Path
+
+sys.path.insert(0, str(Path(__file__).parent.parent.parent.parent / "tools" / "src"))
+
+from gcu.browser.bridge import BeelineBridge
+
+BRIDGE_PORT = 9229
+CONTEXT_NAME = "twitter-scroll-test"
+
+
+async def test_twitter_lazy_scroll():
+    """Test that repeated scrolls with waits load new content."""
+    print("=" * 70)
+    print("TEST #2: Twitter/X Lazy Loading Scroll")
+    print("=" * 70)
+
+    bridge = BeelineBridge()
+
+    try:
+        await bridge.start()
+
+        for i in range(10):
+            await asyncio.sleep(1)
+            if bridge.is_connected:
+                print("✓ Extension connected!")
+                break
+            print(f"Waiting for extension... ({i+1}/10)")
+        else:
+            print("✗ Extension not connected")
+            return
+
+        context = await bridge.create_context(CONTEXT_NAME)
+        tab_id = context.get("tabId")
+        group_id = context.get("groupId")
+        print(f"✓ Created tab: {tab_id}")
+
+        # Navigate to Twitter/X
+        print("\n--- Navigating to X.com ---")
+        await bridge.navigate(tab_id, "https://x.com", wait_until="networkidle", timeout_ms=30000)
+        print("✓ Page loaded")
+
+        # Wait for tweets to appear
+        print("\n--- Waiting for tweets ---")
+        await bridge.wait_for_selector(tab_id, '[data-testid="tweet"]', timeout_ms=10000)
+
+        # Count initial tweets
+        initial_count = await bridge.evaluate(
+            tab_id,
+            '(function() { return document.querySelectorAll(\'[data-testid="tweet"]\').length; })()'
+        )
+        print(f"Initial tweet count: {initial_count.get('result', 0)}")
+
+        # Take screenshot of initial state
+        screenshot = await bridge.screenshot(tab_id)
+        print(f"Screenshot: {len(screenshot.get('data', ''))} bytes")
+
+        # Scroll multiple times with waits
+        print("\n--- Scrolling with waits ---")
+        for i in range(3):
+            result = await bridge.scroll(tab_id, "down", 500)
+            print(f"  Scroll {i+1}: {result.get('method', 'unknown')} method")
+
+            # Wait for new content to load
+            await asyncio.sleep(2)
+
+            # Count tweets after scroll
+            count_result = await bridge.evaluate(
+                tab_id,
+                '(function() { return document.querySelectorAll(\'[data-testid="tweet"]\').length; })()'
+            )
+            count = count_result.get('result', 0)
+            print(f"  Tweet count after scroll: {count}")
+
+        # Final count
+        final_count = await bridge.evaluate(
+            tab_id,
+            '(function() { return document.querySelectorAll(\'[data-testid="tweet"]\').length; })()'
+        )
+        final = final_count.get('result', 0)
+        initial = initial_count.get('result', 0)
+
+        print(f"\n--- Results ---")
+        print(f"Initial tweets: {initial}")
+        print(f"Final tweets: {final}")
+
+        if final > initial:
+            print(f"✓ PASS: Loaded {final - initial} new tweets")
+        else:
+            print("✗ FAIL: No new tweets loaded (may need login)")
+
+        await bridge.destroy_context(group_id)
+        print("\n✓ Context destroyed")
+
+    finally:
+        await bridge.stop()
+
+
+if __name__ == "__main__":
+    asyncio.run(test_twitter_lazy_scroll())

.claude/skills/browser-edge-cases/scripts/test_03_modal_scroll.py

@@ -0,0 +1,97 @@
+#!/usr/bin/env python
+"""
+Test #3: Modal/Dialog Scroll Container
+
+Symptom: Scroll scrolls background page, not modal content
+Root Cause: Modal has its own scroll container with overflow: scroll
+Fix: Find visible modal container (highest z-index scrollable), scroll that
+"""
+
+import asyncio
+import sys
+import time
+from pathlib import Path
+
+sys.path.insert(0, str(Path(__file__).parent.parent.parent.parent / "tools" / "src"))
+
+from gcu.browser.bridge import BeelineBridge
+
+BRIDGE_PORT = 9229
+CONTEXT_NAME = "modal-scroll-test"
+
+# Test site with modal - using a demo site
+MODAL_DEMO_URL = "https://www.w3schools.com/howto/howto_css_modals.asp"
+
+
+async def test_modal_scroll():
+    """Test that scroll targets modal content, not background."""
+    print("=" * 70)
+    print("TEST #3: Modal/Dialog Scroll Container")
+    print("=" * 70)
+
+    bridge = BeelineBridge()
+
+    try:
+        await bridge.start()
+
+        for i in range(10):
+            await asyncio.sleep(1)
+            if bridge.is_connected:
+                print("✓ Extension connected!")
+                break
+        else:
+            print("✗ Extension not connected")
+            return
+
+        context = await bridge.create_context(CONTEXT_NAME)
+        tab_id = context.get("tabId")
+        group_id = context.get("groupId")
+        print(f"✓ Created tab: {tab_id}")
+
+        # Navigate to modal demo
+        print("\n--- Navigating to modal demo ---")
+        await bridge.navigate(tab_id, MODAL_DEMO_URL, wait_until="load")
+        print("✓ Page loaded")
+
+        # Take screenshot before
+        screenshot_before = await bridge.screenshot(tab_id)
+        print(f"Screenshot before: {len(screenshot_before.get('data', ''))} bytes")
+
+        # Click button to open modal
+        print("\n--- Opening modal ---")
+        # Find and click the "Open Modal" button
+        result = await bridge.click(tab_id, '.ws-btn', timeout_ms=5000)
+        print(f"Click result: {result}")
+
+        await asyncio.sleep(1)
+
+        # Take screenshot with modal open
+        screenshot_modal = await bridge.screenshot(tab_id)
+        print(f"Screenshot modal open: {len(screenshot_modal.get('data', ''))} bytes")
+
+        # Try to scroll within modal
+        print("\n--- Scrolling modal content ---")
+        result = await bridge.scroll(tab_id, "down", 100)
+        print(f"Scroll result: {result}")
+
+        await asyncio.sleep(0.5)
+
+        # Take screenshot after scroll
+        screenshot_after = await bridge.screenshot(tab_id)
+        print(f"Screenshot after scroll: {len(screenshot_after.get('data', ''))} bytes")
+
+        # Check if modal content scrolled (not background)
+        # This is a visual check - we can verify by comparing screenshots
+        print("\n--- Results ---")
+        print(f"Modal scroll test completed. Method used: {result.get('method', 'unknown')}")
+        print("Visual verification needed: Check if modal content scrolled vs background")
+
+        await bridge.destroy_context(group_id)
+        print("\n✓ Context destroyed")
+
+    finally:
+        await bridge.stop()
+
+
+if __name__ == "__main__":
+    asyncio.run(test_modal_scroll())

.claude/skills/browser-edge-cases/scripts/test_04_element_covered.py

@@ -0,0 +1,123 @@
+#!/usr/bin/env python
+"""
+Test #4: Element Covered by Overlay
+
+Symptom: Click succeeds but no action triggered
+Root Cause: Element is covered by transparent overlay, tooltip, or iframe
+Detection: document.elementFromPoint(x, y) !== target
+Fix: Wait for overlay to disappear, or use JavaScript element.click()
+"""
+
+import asyncio
+import sys
+from pathlib import Path
+
+sys.path.insert(0, str(Path(__file__).parent.parent.parent.parent / "tools" / "src"))
+
+from gcu.browser.bridge import BeelineBridge
+
+CONTEXT_NAME = "overlay-click-test"
+
+
+async def test_overlay_click():
+    """Test clicking elements that are covered by overlays."""
+    print("=" * 70)
+    print("TEST #4: Element Covered by Overlay")
+    print("=" * 70)
+
+    bridge = BeelineBridge()
+
+    try:
+        await bridge.start()
+
+        for i in range(10):
+            await asyncio.sleep(1)
+            if bridge.is_connected:
+                print("✓ Extension connected!")
+                break
+        else:
+            print("✗ Extension not connected")
+            return
+
+        context = await bridge.create_context(CONTEXT_NAME)
+        tab_id = context.get("tabId")
+        group_id = context.get("groupId")
+        print(f"✓ Created tab: {tab_id}")
+
+        # Create a test page with overlay
+        print("\n--- Creating test page with overlay ---")
+        test_html = """
+        <!DOCTYPE html>
+        <html>
+        <head><title>Overlay Test</title></head>
+        <body>
+            <button id="target-btn" onclick="alert('Clicked!')">Click Me</button>
+            <div id="overlay" style="position:fixed;top:0;left:0;width:100%;height:100%;background:rgba(0,0,0,0.3);z-index:1000;"></div>
+            <script>
+                window.clickCount = 0;
+                document.getElementById('target-btn').addEventListener('click', () => {
+                    window.clickCount++;
+                });
+            </script>
+        </body>
+        </html>
+        """
+
+        # Navigate to data URL
+        import base64
+        data_url = f"data:text/html;base64,{base64.b64encode(test_html.encode()).decode()}"
+        await bridge.navigate(tab_id, data_url, wait_until="load")
+
+        # Screenshot before
+        screenshot = await bridge.screenshot(tab_id)
+        print(f"Screenshot: {len(screenshot.get('data', ''))} bytes")
+
+        # Try to click the covered button
+        print("\n--- Attempting to click covered button ---")
+
+        # First, check if element is covered
+        coverage_check = await bridge.evaluate(
+            tab_id,
+            """
+            (function() {
+                const btn = document.getElementById('target-btn');
+                const rect = btn.getBoundingClientRect();
+                const centerX = rect.left + rect.width / 2;
+                const centerY = rect.top + rect.height / 2;
+                const topElement = document.elementFromPoint(centerX, centerY);
+                return {
+                    isCovered: topElement !== btn && !btn.contains(topElement),
+                    topElement: topElement?.tagName,
+                    targetElement: btn.tagName
+                };
+            })();
+        """
+        )
+        print(f"Coverage check: {coverage_check.get('result', {})}")
+
+        # Try CDP click (may fail due to overlay)
+        click_result = await bridge.click(tab_id, "#target-btn", timeout_ms=5000)
+        print(f"Click result: {click_result}")
+
+        # Check if click registered
+        count_result = await bridge.evaluate(
+            tab_id,
+            "(function() { return window.clickCount; })()"
+        )
+        count = count_result.get("result", 0)
+        print(f"Click count after CDP click: {count}")
+
+        if count > 0:
+            print("✓ PASS: JavaScript click penetrated overlay")
+        else:
+            print("✗ FAIL: Click did not reach button (overlay blocked it)")
+
+        await bridge.destroy_context(group_id)
+        print("\n✓ Context destroyed")
+
+    finally:
+        await bridge.stop()
+
+
+if __name__ == "__main__":
+    asyncio.run(test_overlay_click())

.claude/skills/browser-edge-cases/scripts/test_06_shadow_dom.py

@@ -0,0 +1,154 @@
+#!/usr/bin/env python
+"""
+Test #6: Shadow DOM Elements
+
+Symptom: querySelector can't find element
+Root Cause: Element is inside a shadow root, not main DOM tree
+Detection: element.shadowRoot !== null on parent elements
+Fix: Use piercing selector (host >>> target) or traverse shadow roots
+"""
+
+import asyncio
+import sys
+import base64
+from pathlib import Path
+
+sys.path.insert(0, str(Path(__file__).parent.parent.parent.parent / "tools" / "src"))
+
+from gcu.browser.bridge import BeelineBridge
+
+CONTEXT_NAME = "shadow-dom-test"
+
+
+async def test_shadow_dom():
+    """Test clicking elements inside Shadow DOM."""
+    print("=" * 70)
+    print("TEST #6: Shadow DOM Elements")
+    print("=" * 70)
+
+    bridge = BeelineBridge()
+
+    try:
+        await bridge.start()
+
+        for i in range(10):
+            await asyncio.sleep(1)
+            if bridge.is_connected:
+                print("✓ Extension connected!")
+                break
+        else:
+            print("✗ Extension not connected")
+            return
+
+        context = await bridge.create_context(CONTEXT_NAME)
+        tab_id = context.get("tabId")
+        group_id = context.get("groupId")
+        print(f"✓ Created tab: {tab_id}")
+
+        # Create test page with Shadow DOM
+        print("\n--- Creating test page with Shadow DOM ---")
+        test_html = """
+        <!DOCTYPE html>
+        <html>
+        <head><title>Shadow DOM Test</title></head>
+        <body>
+            <div id="shadow-host"></div>
+            <script>
+                const host = document.getElementById('shadow-host');
+                const shadow = host.attachShadow({ mode: 'open' });
+                shadow.innerHTML = `
+                    <style>
+                        button { padding: 10px 20px; font-size: 16px; }
+                    </style>
+                    <button id="shadow-btn">Shadow Button</button>
+                `;
+                shadow.getElementById('shadow-btn').addEventListener('click', () => {
+                    window.shadowClickCount = (window.shadowClickCount || 0) + 1;
+                    console.log('Shadow button clicked:', window.shadowClickCount);
+                });
+            </script>
+        </body>
+        </html>
+        """
+
+        # Write to file and use file:// URL (data: URLs don't work well with extension)
+        test_file = Path("/tmp/shadow_dom_test.html")
+        test_file.write_text(test_html.strip())
+        file_url = f"file://{test_file}"
+        await bridge.navigate(tab_id, file_url, wait_until="load")
+        print("✓ Page loaded")
+
+        # Screenshot
+        screenshot = await bridge.screenshot(tab_id)
+        print(f"Screenshot: {len(screenshot.get('data', ''))} bytes")
+
+        # Detect Shadow DOM
+        print("\n--- Detecting Shadow DOM ---")
+        detection = await bridge.evaluate(
+            tab_id,
+            """
+            (function() {
+                const hosts = [];
+                document.querySelectorAll('*').forEach(el => {
+                    if (el.shadowRoot) {
+                        hosts.push({
+                            tag: el.tagName,
+                            id: el.id,
+                            hasButton: el.shadowRoot.querySelector('button') !== null
+                        });
+                    }
+                });
+                return { count: hosts.length, hosts };
+            })();
+        """
+        )
+        print(f"Shadow DOM detection: {detection.get('result', {})}")
+
+        # Try to click shadow button using regular selector (should fail)
+        print("\n--- Attempting click with regular selector ---")
+        try:
+            result = await bridge.click(tab_id, "#shadow-btn", timeout_ms=3000)
+            print(f"Result: {result}")
+        except Exception as e:
+            print(f"Expected failure: {e}")
+
+        # Try to click using JavaScript that pierces shadow DOM
+        print("\n--- Clicking via JavaScript shadow piercing ---")
+        click_result = await bridge.evaluate(
+            tab_id,
+            """
+            (function() {
+                const host = document.getElementById('shadow-host');
+                const btn = host.shadowRoot.getElementById('shadow-btn');
+                if (btn) {
+                    btn.click();
+                    return { success: true, clicked: 'shadow-btn' };
+                }
+                return { success: false, error: 'Button not found' };
+            })();
+        """
+        )
+        print(f"JS click result: {click_result.get('result', {})}")
+
+        # Verify click was registered
+        count_result = await bridge.evaluate(
+            tab_id,
+            "(function() { return window.shadowClickCount || 0; })()"
+        )
+        count = count_result.get("result") or 0
+        print(f"Shadow click count: {count}")
+
+        if count and count > 0:
+            print("✓ PASS: Shadow DOM element clicked successfully")
+        else:
+            print("✗ FAIL: Could not click Shadow DOM element")
+
+        await bridge.destroy_context(group_id)
+        print("\n✓ Context destroyed")
+
+    finally:
+        await bridge.stop()
+
+
+if __name__ == "__main__":
+    asyncio.run(test_shadow_dom())

.claude/skills/browser-edge-cases/scripts/test_07_contenteditable.py

@@ -0,0 +1,178 @@
+#!/usr/bin/env python
+"""
+Test #7: ContentEditable / Rich Text Editors
+
+Symptom: browser_type() doesn't insert text
+Root Cause: Element is contenteditable, not an <input> or <textarea>
+Detection: element.contentEditable === 'true'
+Fix: Focus via JavaScript, use execCommand('insertText') or Input.dispatchKeyEvent
+"""
+
+import asyncio
+import sys
+import base64
+from pathlib import Path
+
+sys.path.insert(0, str(Path(__file__).parent.parent.parent.parent / "tools" / "src"))
+
+from gcu.browser.bridge import BeelineBridge
+
+CONTEXT_NAME = "contenteditable-test"
+
+
+async def test_contenteditable():
+    """Test typing into contenteditable elements."""
+    print("=" * 70)
+    print("TEST #7: ContentEditable / Rich Text Editors")
+    print("=" * 70)
+
+    bridge = BeelineBridge()
+
+    try:
+        await bridge.start()
+
+        for i in range(10):
+            await asyncio.sleep(1)
+            if bridge.is_connected:
+                print("✓ Extension connected!")
+                break
+        else:
+            print("✗ Extension not connected")
+            return
+
+        context = await bridge.create_context(CONTEXT_NAME)
+        tab_id = context.get("tabId")
+        group_id = context.get("groupId")
+        print(f"✓ Created tab: {tab_id}")
+
+        # Create test page with contenteditable
+        test_html = """
+        <!DOCTYPE html>
+        <html>
+        <head><title>ContentEditable Test</title></head>
+        <body>
+            <h2>ContentEditable Test</h2>
+
+            <h3>1. Simple contenteditable div</h3>
+            <div id="editor1" contenteditable="true" style="border:1px solid #ccc;padding:10px;min-height:50px;">Start text</div>
+
+            <h3>2. Rich text editor (like Notion)</h3>
+            <div id="editor2" contenteditable="true" style="border:1px solid #ccc;padding:10px;min-height:50px;">
+                <p>Type here...</p>
+            </div>
+
+            <h3>3. Regular input (for comparison)</h3>
+            <input id="input1" type="text" placeholder="Regular input" />
+
+            <script>
+                // Track content changes
+                window.editor1Content = '';
+                window.editor2Content = '';
+
+                document.getElementById('editor1').addEventListener('input', (e) => {
+                    window.editor1Content = e.target.innerText;
+                });
+                document.getElementById('editor2').addEventListener('input', (e) => {
+                    window.editor2Content = e.target.innerText;
+                });
+            </script>
+        </body>
+        </html>
+        """
+
+        # Write to file and use file:// URL (data: URLs don't work well with extension)
+        test_file = Path("/tmp/contenteditable_test.html")
+        test_file.write_text(test_html.strip())
+        file_url = f"file://{test_file}"
+        await bridge.navigate(tab_id, file_url, wait_until="load")
+        print("✓ Page loaded")
+
+        # Screenshot with timeout protection
+        try:
+            screenshot = await asyncio.wait_for(bridge.screenshot(tab_id), timeout=10.0)
+            print(f"Screenshot: {len(screenshot.get('data', ''))} bytes")
+        except asyncio.TimeoutError:
+            print("Screenshot timed out (skipping)")
+
+        # Detect contenteditable
+        print("\n--- Detecting contenteditable elements ---")
+        detection = await bridge.evaluate(
+            tab_id,
+            """
+            (function() {
+                const editables = document.querySelectorAll('[contenteditable="true"]');
+                return {
+                    count: editables.length,
+                    ids: Array.from(editables).map(el => el.id)
+                };
+            })();
+        """
+        )
+        print(f"Contenteditable detection: {detection.get('result', {})}")
+
+        # Test 1: Type into regular input (baseline)
+        print("\n--- Test 1: Regular input ---")
+        await bridge.click(tab_id, "#input1")
+        await bridge.type_text(tab_id, "#input1", "Hello input")
+        input_result = await bridge.evaluate(
+            tab_id,
+            "(function() { return document.getElementById('input1').value; })()"
+        )
+        print(f"Input value: {input_result.get('result', '')}")
+
+        # Test 2: Type into contenteditable div
+        print("\n--- Test 2: Contenteditable div ---")
+        await bridge.click(tab_id, "#editor1")
+        await bridge.type_text(tab_id, "#editor1", "Hello contenteditable", clear_first=True)
+        editor_result = await bridge.evaluate(
+            tab_id,
+            "(function() { return document.getElementById('editor1').innerText; })()"
+        )
+        print(f"Editor1 innerText: {editor_result.get('result', '')}")
+
+        # Test 3: Use JavaScript insertText for rich editor
+        print("\n--- Test 3: JavaScript insertText for rich editor ---")
+        insert_result = await bridge.evaluate(
+            tab_id,
+            """
+            (function() {
+                const editor = document.getElementById('editor2');
+                editor.focus();
+                document.execCommand('selectAll', false, null);
+                document.execCommand('insertText', false, 'Hello from execCommand');
+                return editor.innerText;
+            })();
+        """
+        )
+        print(f"Editor2 after execCommand: {insert_result.get('result', '')}")
+
+        # Screenshot after with timeout protection
+        try:
+            screenshot_after = await asyncio.wait_for(bridge.screenshot(tab_id), timeout=10.0)
+            print(f"Screenshot after: {len(screenshot_after.get('data', ''))} bytes")
+        except asyncio.TimeoutError:
+            print("Screenshot after timed out (skipping)")
+
+        # Results
+        print("\n--- Results ---")
+        input_val = input_result.get("result", "")
+        editor1_val = editor_result.get("result", "")
+        editor2_val = insert_result.get("result", "")
+
+        input_pass = "Hello input" in input_val
+        editor1_pass = "Hello contenteditable" in editor1_val
+        editor2_pass = "execCommand" in editor2_val
+
+        print(f"Input: {'✓ PASS' if input_pass else '✗ FAIL'} - {input_val}")
+        print(f"Editor1: {'✓ PASS' if editor1_pass else '✗ FAIL'} - {editor1_val}")
+        print(f"Editor2: {'✓ PASS' if editor2_pass else '✗ FAIL'} - {editor2_val}")
+
+        await bridge.destroy_context(group_id)
+        print("\n✓ Context destroyed")
+
+    finally:
+        await bridge.stop()
+
+
+if __name__ == "__main__":
+    asyncio.run(test_contenteditable())

.claude/skills/browser-edge-cases/scripts/test_08_autocomplete.py

@@ -0,0 +1,233 @@
+#!/usr/bin/env python
+"""
+Test #8: Autocomplete Field Clearing
+
+Symptom: Typed text gets cleared immediately
+Root Cause: Field expects realistic keystroke timing for autocomplete
+Detection: Field has autocomplete listeners or dropdown appears
+Fix: Add delay_ms between keystrokes
+"""
+
+import asyncio
+import sys
+import base64
+from pathlib import Path
+
+sys.path.insert(0, str(Path(__file__).parent.parent.parent.parent / "tools" / "src"))
+
+from gcu.browser.bridge import BeelineBridge
+
+CONTEXT_NAME = "autocomplete-test"
+
+
+async def test_autocomplete():
+    """Test typing into fields with autocomplete behavior."""
+    print("=" * 70)
+    print("TEST #8: Autocomplete Field Clearing")
+    print("=" * 70)
+
+    bridge = BeelineBridge()
+
+    try:
+        await bridge.start()
+
+        for i in range(10):
+            await asyncio.sleep(1)
+            if bridge.is_connected:
+                print("✓ Extension connected!")
+                break
+        else:
+            print("✗ Extension not connected")
+            return
+
+        context = await bridge.create_context(CONTEXT_NAME)
+        tab_id = context.get("tabId")
+        group_id = context.get("groupId")
+        print(f"✓ Created tab: {tab_id}")
+
+        # Create test page with autocomplete behavior
+        test_html = """
+        <!DOCTYPE html>
+        <html>
+        <head><title>Autocomplete Test</title>
+        <style>
+            .autocomplete-items {
+                position: absolute;
+                border: 1px solid #d4d4d4;
+                border-top: none;
+                z-index: 99;
+                top: 100%;
+                left: 0;
+                right: 0;
+                max-height: 200px;
+                overflow-y: auto;
+                background: white;
+            }
+            .autocomplete-items div {
+                padding: 10px;
+                cursor: pointer;
+            }
+            .autocomplete-items div:hover {
+                background-color: #e9e9e9;
+            }
+            .autocomplete-active {
+                background-color: DodgerBlue !important;
+                color: white;
+            }
+            .autocomplete { position: relative; display: inline-block; }
+            input { width: 300px; padding: 10px; font-size: 16px; }
+        </style></head>
+        <body>
+            <h2>Autocomplete Test</h2>
+
+            <div class="autocomplete">
+                <input id="search" type="text" placeholder="Search countries..." autocomplete="off">
+            </div>
+
+            <div id="log" style="margin-top:20px;font-family:monospace;"></div>
+
+            <script>
+                const countries = ["Afghanistan","Albania","Algeria","Andorra","Angola","Argentina","Armenia","Australia","Austria","Azerbaijan","Bahamas","Bahrain","Bangladesh","Belarus","Belgium","Belize","Benin","Bhutan","Bolivia","Brazil","Canada","China","Colombia","Denmark","Egypt","France","Germany","India","Indonesia","Italy","Japan","Mexico","Netherlands","Nigeria","Norway","Pakistan","Peru","Philippines","Poland","Portugal","Russia","Spain","Sweden","Switzerland","Thailand","Turkey","Ukraine","United Kingdom","United States","Vietnam"];
+
+                const input = document.getElementById('search');
+                const log = document.getElementById('log');
+                let currentFocus = -1;
+                let typingTimeout = null;
+
+                // Track events for testing
+                window.inputEvents = [];
+                window.inputValue = '';
+
+                function logEvent(type, value) {
+                    window.inputEvents.push({ type, value, time: Date.now() });
+                    const entry = document.createElement('div');
+                    entry.textContent = type + ': ' + value;
+                    log.insertBefore(entry, log.firstChild);
+                }
+
+                // Simulate autocomplete that clears fast typing
+                input.addEventListener('input', function(e) {
+                    const val = this.value;
+
+                    // Clear previous dropdown
+                    closeAllLists();
+
+                    if (!val) return;
+
+                    // If typing too fast (autocomplete-style), clear and restart
+                    clearTimeout(typingTimeout);
+                    typingTimeout = setTimeout(() => {
+                        logEvent('input', val);
+                        window.inputValue = val;
+
+                        // Create dropdown
+                        const div = document.createElement('div');
+                        div.setAttribute('id', this.id + 'autocomplete-list');
+                        div.setAttribute('class', 'autocomplete-items');
+                        this.parentNode.appendChild(div);
+
+                        countries.filter(c => c.substr(0, val.length).toUpperCase() === val.toUpperCase())
+                            .slice(0, 5)
+                            .forEach(country => {
+                                const item = document.createElement('div');
+                                item.innerHTML = '<strong>' + country.substr(0, val.length) + '</strong>' + country.substr(val.length);
+                                item.addEventListener('click', function() {
+                                    input.value = country;
+                                    closeAllLists();
+                                    logEvent('select', country);
+                                    window.inputValue = country;
+                                });
+                                div.appendChild(item);
+                            });
+                    }, 100); // 100ms debounce
+                });
+
+                function closeAllLists() {
+                    document.querySelectorAll('.autocomplete-items').forEach(el => el.remove());
+                }
+
+                document.addEventListener('click', function() {
+                    closeAllLists();
+                });
+            </script>
+        </body>
+        </html>
+        """
+
+        # Write to file and use file:// URL (data: URLs don't work well with extension)
+        test_file = Path("/tmp/autocomplete_test.html")
+        test_file.write_text(test_html.strip())
+        file_url = f"file://{test_file}"
+        await bridge.navigate(tab_id, file_url, wait_until="load")
+        print("✓ Page loaded")
+
+        # Screenshot
+        screenshot = await bridge.screenshot(tab_id)
+        print(f"Screenshot: {len(screenshot.get('data', ''))} bytes")
+
+        # Test 1: Fast typing (no delay) - may fail
+        print("\n--- Test 1: Fast typing (delay_ms=0) ---")
+        await bridge.click(tab_id, "#search")
+        await bridge.type_text(tab_id, "#search", "Ger", clear_first=True, delay_ms=0)
+        await asyncio.sleep(0.5)
+
+        fast_result = await bridge.evaluate(
+            tab_id,
+            "(function() { return document.getElementById('search').value; })()"
+        )
+        fast_value = fast_result.get("result", "")
+        print(f"Value after fast typing: '{fast_value}'")
+
+        # Check events
+        events_result = await bridge.evaluate(
+            tab_id,
+            "(function() { return window.inputEvents; })()"
+        )
+        print(f"Events logged: {events_result.get('result', [])}")
+
+        # Test 2: Slow typing (with delay) - should work
+        print("\n--- Test 2: Slow typing (delay_ms=100) ---")
+        await bridge.click(tab_id, "#search")
+        await bridge.type_text(tab_id, "#search", "United", clear_first=True, delay_ms=100)
+        await asyncio.sleep(0.5)
+
+        slow_result = await bridge.evaluate(
+            tab_id,
+            "(function() { return document.getElementById('search').value; })()"
+        )
+        slow_value = slow_result.get("result", "")
+        print(f"Value after slow typing: '{slow_value}'")
+
+        # Check if dropdown appeared
+        dropdown_result = await bridge.evaluate(
+            tab_id,
+            "(function() { return document.querySelectorAll('.autocomplete-items div').length; })()"
+        )
+        dropdown_count = dropdown_result.get("result", 0)
+        print(f"Dropdown items: {dropdown_count}")
+
+        # Screenshot with dropdown
+        screenshot_dropdown = await bridge.screenshot(tab_id)
+        print(f"Screenshot with dropdown: {len(screenshot_dropdown.get('data', ''))} bytes")
+
+        # Results
+        print("\n--- Results ---")
+        if "United" in slow_value:
+            print("✓ PASS: Slow typing with delay_ms worked")
+        else:
+            print("✗ FAIL: Slow typing still didn't work")
+
+        if dropdown_count > 0:
+            print("✓ PASS: Autocomplete dropdown appeared")
+        else:
+            print("⚠ WARNING: No autocomplete dropdown")
+
+        await bridge.destroy_context(group_id)
+        print("\n✓ Context destroyed")
+
+    finally:
+        await bridge.stop()
+
+
+if __name__ == "__main__":
+    asyncio.run(test_autocomplete())

.claude/skills/browser-edge-cases/scripts/test_10_huge_dom.py

@@ -0,0 +1,162 @@
+#!/usr/bin/env python
+"""
+Test #10: LinkedIn Huge DOM Tree
+
+Symptom: browser_snapshot() hangs forever
+Root Cause: 10k+ DOM nodes, accessibility tree has 50k+ nodes
+Detection: document.querySelectorAll('*').length > 5000
+Fix: Add timeout (10s default), truncate tree at 2000 nodes
+"""
+
+import asyncio
+import sys
+import time
+import base64
+from pathlib import Path
+
+sys.path.insert(0, str(Path(__file__).parent.parent.parent.parent / "tools" / "src"))
+
+from gcu.browser.bridge import BeelineBridge
+
+CONTEXT_NAME = "huge-dom-test"
+
+
+async def test_huge_dom():
+    """Test snapshot performance on huge DOM trees."""
+    print("=" * 70)
+    print("TEST #10: Huge DOM Tree (LinkedIn-style)")
+    print("=" * 70)
+
+    bridge = BeelineBridge()
+
+    try:
+        await bridge.start()
+
+        for i in range(10):
+            await asyncio.sleep(1)
+            if bridge.is_connected:
+                print("✓ Extension connected!")
+                break
+        else:
+            print("✗ Extension not connected")
+            return
+
+        context = await bridge.create_context(CONTEXT_NAME)
+        tab_id = context.get("tabId")
+        group_id = context.get("groupId")
+        print(f"✓ Created tab: {tab_id}")
+
+        # Test 1: Small DOM (baseline)
+        print("\n--- Test 1: Small DOM (baseline) ---")
+        small_html = """
+        <!DOCTYPE html>
+        <html><body>
+            <h1>Small Page</h1>
+            <p>A few elements</p>
+            <button>Click me</button>
+        </body></html>
+        """
+        data_url = f"data:text/html;base64,{base64.b64encode(small_html.encode()).decode()}"
+        await bridge.navigate(tab_id, data_url, wait_until="load")
+
+        start = time.perf_counter()
+        snapshot = await bridge.snapshot(tab_id, timeout_s=5.0)
+        elapsed = time.perf_counter() - start
+        tree_len = len(snapshot.get("tree", ""))
+        print(f"Small DOM snapshot: {elapsed:.3f}s, {tree_len} chars")
+
+        # Test 2: Generate huge DOM
+        print("\n--- Test 2: Huge DOM (5000+ elements) ---")
+        huge_html = """
+        <!DOCTYPE html>
+        <html><body>
+        <h1>Huge DOM Test</h1>
+        <div id="container"></div>
+        <script>
+            const container = document.getElementById('container');
+            for (let i = 0; i < 5000; i++) {
+                const div = document.createElement('div');
+                div.className = 'item-' + i;
+                div.innerHTML = '<span>Item ' + i + '</span><button>Action</button>';
+                container.appendChild(div);
+            }
+        </script>
+        </body></html>
+        """
+        data_url = f"data:text/html;base64,{base64.b64encode(huge_html.encode()).decode()}"
+        await bridge.navigate(tab_id, data_url, wait_until="load")
+
+        # Count elements
+        count_result = await bridge.evaluate(
+            tab_id,
+            "(function() { return document.querySelectorAll('*').length; })()"
+        )
+        elem_count = count_result.get("result", 0)
+        print(f"DOM elements: {elem_count}")
+
+        # Skip screenshot on huge DOM - it can timeout
+        # Instead verify page loaded by checking DOM
+        print("✓ Page verified (skipping screenshot on huge DOM)")
+
+        # Test snapshot with timeout
+        print("\n--- Testing snapshot with 10s timeout ---")
+        start = time.perf_counter()
+        try:
+            snapshot = await bridge.snapshot(tab_id, timeout_s=10.0)
+            elapsed = time.perf_counter() - start
+            tree_len = len(snapshot.get("tree", ""))
+            truncated = "(truncated)" in snapshot.get("tree", "")
+            print(f"✓ Huge DOM snapshot: {elapsed:.3f}s, {tree_len} chars, truncated={truncated}")
+
+            if elapsed < 5.0:
+                print("✓ PASS: Snapshot completed quickly")
+            else:
+                print(f"⚠ WARNING: Snapshot took {elapsed:.1f}s")
+
+            if truncated:
+                print("✓ PASS: Tree was truncated to prevent hang")
+            else:
+                print("⚠ WARNING: Tree not truncated (may need adjustment)")
+
+        except asyncio.TimeoutError:
+            print("✗ FAIL: Snapshot timed out (this shouldn't happen)")
+
+        # Test 3: Real LinkedIn
+        print("\n--- Test 3: Real LinkedIn Feed ---")
+        await bridge.navigate(tab_id, "https://www.linkedin.com/feed", wait_until="load", timeout_ms=30000)
+        await asyncio.sleep(2)
+
+        count_result = await bridge.evaluate(
+            tab_id,
+            "(function() { return document.querySelectorAll('*').length; })()"
+        )
+        elem_count = count_result.get("result", 0)
+        print(f"LinkedIn DOM elements: {elem_count}")
+
+        start = time.perf_counter()
+        try:
+            snapshot = await bridge.snapshot(tab_id, timeout_s=15.0)
+            elapsed = time.perf_counter() - start
+            tree_len = len(snapshot.get("tree", ""))
+            truncated = "(truncated)" in snapshot.get("tree", "")
+            print(f"LinkedIn snapshot: {elapsed:.3f}s, {tree_len} chars, truncated={truncated}")
+
+            if elapsed < 5.0:
+                print("✓ PASS: LinkedIn snapshot fast enough")
+            elif elapsed < 15.0:
+                print("⚠ WARNING: LinkedIn snapshot slow but within timeout")
+            else:
+                print("✗ FAIL: LinkedIn snapshot too slow")
+
+        except asyncio.TimeoutError:
+            print("✗ FAIL: LinkedIn snapshot timed out")
+
+        await bridge.destroy_context(group_id)
+        print("\n✓ Context destroyed")
+
+    finally:
+        await bridge.stop()
+
+
+if __name__ == "__main__":
+    asyncio.run(test_huge_dom())

.claude/skills/browser-edge-cases/scripts/test_13_spa_navigation.py

@@ -0,0 +1,187 @@
+#!/usr/bin/env python
+"""
+Test #13: SPA Navigation Events
+
+Symptom: wait_until="load" fires before content ready
+Root Cause: SPA uses client-side routing, no full page load
+Detection: URL changes but load event already fired
+Fix: Use wait_until="networkidle" or wait_for_selector
+"""
+
+import asyncio
+import sys
+import time
+import base64
+from pathlib import Path
+
+sys.path.insert(0, str(Path(__file__).parent.parent.parent.parent / "tools" / "src"))
+
+from gcu.browser.bridge import BeelineBridge
+
+CONTEXT_NAME = "spa-nav-test"
+
+
+async def test_spa_navigation():
+    """Test navigation timing on SPA pages."""
+    print("=" * 70)
+    print("TEST #13: SPA Navigation Events")
+    print("=" * 70)
+
+    bridge = BeelineBridge()
+
+    try:
+        await bridge.start()
+
+        for i in range(10):
+            await asyncio.sleep(1)
+            if bridge.is_connected:
+                print("✓ Extension connected!")
+                break
+        else:
+            print("✗ Extension not connected")
+            return
+
+        context = await bridge.create_context(CONTEXT_NAME)
+        tab_id = context.get("tabId")
+        group_id = context.get("groupId")
+        print(f"✓ Created tab: {tab_id}")
+
+        # Create a test SPA
+        spa_html = """
+        <!DOCTYPE html>
+        <html>
+        <head>
+            <title>SPA Test</title>
+            <style>
+                nav a { margin-right: 10px; }
+                .page { padding: 20px; border: 1px solid #ccc; margin-top: 10px; }
+            </style>
+        </head>
+        <body>
+            <nav>
+                <a href="#home" onclick="navigate('home')">Home</a>
+                <a href="#about" onclick="navigate('about')">About</a>
+                <a href="#contact" onclick="navigate('contact')">Contact</a>
+            </nav>
+            <div id="app" class="page">
+                <h1>Loading...</h1>
+            </div>
+            <script>
+                // Simulate SPA routing
+                let currentPage = '';
+
+                async function navigate(page) {
+                    event.preventDefault();
+                    currentPage = page;
+
+                    // Show loading state
+                    document.getElementById('app').innerHTML = '<h1>Loading...</h1>';
+
+                    // Simulate async content loading (like real SPAs)
+                    await new Promise(r => setTimeout(r, 500));
+
+                    // Render content
+                    const content = {
+                        home: '<h1>Home Page</h1><p>Welcome to the SPA!</p><button id="home-btn">Home Action</button>',
+                        about: '<h1>About Page</h1><p>This is a simulated SPA.</p><button id="about-btn">About Action</button>',
+                        contact: '<h1>Contact Page</h1><p>Contact us at test@example.com</p><button id="contact-btn">Contact Action</button>'
+                    };
+
+                    document.getElementById('app').innerHTML = content[page] || '<h1>404</h1>';
+                    window.location.hash = page;
+                }
+
+                // Initial load with delay (simulates SPA hydration)
+                setTimeout(() => {
+                    navigate('home');
+                }, 1000);
+
+                // Track for testing
+                window.pageLoads = [];
+                window.addEventListener('hashchange', () => {
+                    window.pageLoads.push(window.location.hash);
+                });
+            </script>
+        </body>
+        </html>
+        """
+
+        # Write to file and use file:// URL (data: URLs don't work well with extension)
+        test_file = Path("/tmp/spa_test.html")
+        test_file.write_text(spa_html.strip())
+        file_url = f"file://{test_file}"
+
+        # Test 1: wait_until="load" - may fire before content ready
+        print("\n--- Test 1: wait_until='load' ---")
+        start = time.perf_counter()
+        await bridge.navigate(tab_id, file_url, wait_until="load")
+        elapsed = time.perf_counter() - start
+        print(f"Navigation completed in {elapsed:.3f}s")
+
+        # Check content immediately
+        content = await bridge.evaluate(
+            tab_id,
+            "(function() { return document.getElementById('app').innerText; })()"
+        )
+        print(f"Content immediately after load: '{content.get('result', '')}'")
+
+        # Screenshot
+        screenshot = await bridge.screenshot(tab_id)
+        print(f"Screenshot: {len(screenshot.get('data', ''))} bytes")
+
+        # Wait for content
+        print("\n--- Waiting for content to hydrate ---")
+        await bridge.wait_for_selector(tab_id, "#home-btn", timeout_ms=5000)
+        print("✓ Content loaded")
+
+        # Check content after wait
+        content_after = await bridge.evaluate(
+            tab_id,
+            "(function() { return document.getElementById('app').innerText; })()"
+        )
+        print(f"Content after wait: '{content_after.get('result', '')}'")
+
+        # Test 2: SPA navigation (no full page load)
+        print("\n--- Test 2: SPA client-side navigation ---")
+
+        # Click "About" link
+        await bridge.click(tab_id, 'a[href="#about"]')
+        await asyncio.sleep(1)
+
+        # Check if content changed
+        about_content = await bridge.evaluate(
+            tab_id,
+            "(function() { return document.getElementById('app').innerText; })()"
+        )
+        print(f"Content after SPA nav: '{about_content.get('result', '')}'")
+
+        if "About Page" in about_content.get("result", ""):
+            print("✓ PASS: SPA navigation worked")
+        else:
+            print("✗ FAIL: SPA navigation didn't update content")
+
+        # Test 3: wait_until="networkidle"
+        print("\n--- Test 3: wait_until='networkidle' ---")
+        await bridge.navigate(tab_id, file_url, wait_until="networkidle", timeout_ms=10000)
+
+        # Check content immediately
+        content_networkidle = await bridge.evaluate(
+            tab_id,
+            "(function() { return document.getElementById('app').innerText; })()"
+        )
+        print(f"Content after networkidle: '{content_networkidle.get('result', '')}'")
+
+        if "Home Page" in content_networkidle.get("result", ""):
+            print("✓ PASS: networkidle waited for content")
+        else:
+            print("⚠ WARNING: networkidle didn't wait long enough")
+
+        await bridge.destroy_context(group_id)
+        print("\n✓ Context destroyed")
+
+    finally:
+        await bridge.stop()
+
+
+if __name__ == "__main__":
+    asyncio.run(test_spa_navigation())

.claude/skills/browser-edge-cases/scripts/test_15_screenshot.py

@@ -0,0 +1,262 @@
+#!/usr/bin/env python
+"""
+Test #15: Screenshot Functionality
+
+Tests browser_screenshot across multiple scenarios:
+- Basic viewport screenshot
+- Full-page screenshot
+- Selector-based screenshot
+- Screenshot on complex DOM
+- Timeout handling
+
+Category: screenshot
+"""
+
+import asyncio
+import base64
+import sys
+import time
+from pathlib import Path
+
+sys.path.insert(0, str(Path(__file__).parent.parent.parent.parent / "tools" / "src"))
+
+from gcu.browser.bridge import BeelineBridge
+
+CONTEXT_NAME = "screenshot-test"
+
+SIMPLE_HTML = """<!DOCTYPE html>
+<html>
+<head><style>
+  body { margin: 0; background: #fff; font-family: sans-serif; }
+  h1 { color: #333; padding: 20px; }
+  .box { width: 200px; height: 100px; background: #4a90e2; margin: 20px; }
+  .long-content { height: 2000px; background: linear-gradient(blue, red); }
+</style></head>
+<body>
+  <h1 id="title">Screenshot Test Page</h1>
+  <div class="box" id="target-box">Target Box</div>
+  <div class="long-content"></div>
+</body>
+</html>"""
+
+
+def check_png(data: str) -> bool:
+    """Verify that base64 data decodes to a valid PNG."""
+    try:
+        raw = base64.b64decode(data)
+        return raw[:8] == b'\x89PNG\r\n\x1a\n'
+    except Exception:
+        return False
+
+
+async def test_basic_screenshot(bridge: BeelineBridge, tab_id: int, data_url: str):
+    print("\n--- Test 1: Basic Viewport Screenshot ---")
+    await bridge.navigate(tab_id, data_url, wait_until="load")
+    await asyncio.sleep(0.5)
+
+    start = time.perf_counter()
+    result = await bridge.screenshot(tab_id)
+    elapsed = time.perf_counter() - start
+
+    ok = result.get("ok")
+    data = result.get("data", "")
+    mime = result.get("mimeType", "")
+
+    print(f"  ok={ok}, mimeType={mime}, elapsed={elapsed:.3f}s")
+    print(f"  data length: {len(data)} chars")
+
+    if ok and data:
+        valid_png = check_png(data)
+        print(f"  valid PNG: {valid_png}")
+        if valid_png:
+            raw = base64.b64decode(data)
+            print(f"  PNG size: {len(raw)} bytes")
+            print("  ✓ PASS: Basic screenshot works")
+            return True
+        else:
+            print("  ✗ FAIL: Data is not a valid PNG")
+    else:
+        print(f"  ✗ FAIL: {result.get('error', 'no data')}")
+    return False
+
+
+async def test_full_page_screenshot(bridge: BeelineBridge, tab_id: int, data_url: str):
+    print("\n--- Test 2: Full Page Screenshot ---")
+    await bridge.navigate(tab_id, data_url, wait_until="load")
+    await asyncio.sleep(0.5)
+
+    viewport_result = await bridge.screenshot(tab_id, full_page=False)
+    full_result = await bridge.screenshot(tab_id, full_page=True)
+
+    v_data = viewport_result.get("data", "")
+    f_data = full_result.get("data", "")
+
+    if not v_data or not f_data:
+        print(f"  ✗ FAIL: viewport ok={viewport_result.get('ok')}, full ok={full_result.get('ok')}")
+        return False
+
+    v_size = len(base64.b64decode(v_data))
+    f_size = len(base64.b64decode(f_data))
+    print(f"  Viewport PNG: {v_size} bytes")
+    print(f"  Full page PNG: {f_size} bytes")
+
+    if f_size > v_size:
+        print("  ✓ PASS: Full page larger than viewport")
+        return True
+    else:
+        print("  ✗ FAIL: Full page not larger than viewport (may not capture long pages)")
+        return False
+
+
+async def test_selector_screenshot(bridge: BeelineBridge, tab_id: int, data_url: str):
+    print("\n--- Test 3: Selector Screenshot ---")
+    await bridge.navigate(tab_id, data_url, wait_until="load")
+    await asyncio.sleep(0.5)
+
+    # selector param exists in signature but may not be implemented
+    result = await bridge.screenshot(tab_id, selector="#target-box")
+
+    ok = result.get("ok")
+    data = result.get("data", "")
+
+    if ok and data:
+        # If implemented, the box screenshot should be smaller than a full viewport screenshot
+        full_result = await bridge.screenshot(tab_id)
+        full_data = full_result.get("data", "")
+
+        if full_data:
+            sel_size = len(base64.b64decode(data))
+            full_size = len(base64.b64decode(full_data))
+            print(f"  Selector PNG: {sel_size} bytes")
+            print(f"  Full page PNG: {full_size} bytes")
+            if sel_size < full_size:
+                print("  ✓ PASS: Selector screenshot smaller than full page")
+                return True
+            else:
+                print("  ⚠ WARNING: Selector screenshot not smaller (may be full page)")
+                return False
+    else:
+        print(f"  ⚠ NOT IMPLEMENTED: selector param ignored (returns full page) - error={result.get('error')}")
+        print("  NOTE: selector parameter exists in signature but is not used in implementation")
+        return False
+
+
+async def test_screenshot_url_metadata(bridge: BeelineBridge, tab_id: int):
+    print("\n--- Test 4: Screenshot URL Metadata ---")
+    await bridge.navigate(tab_id, "https://example.com", wait_until="load")
+    await asyncio.sleep(1)
+
+    result = await bridge.screenshot(tab_id)
+    url = result.get("url", "")
+    tab = result.get("tabId")
+
+    print(f"  url={url!r}, tabId={tab}")
+
+    if "example.com" in url:
+        print("  ✓ PASS: URL metadata captured correctly")
+        return True
+    else:
+        print(f"  ✗ FAIL: Expected example.com in URL, got {url!r}")
+        return False
+
+
+async def test_screenshot_timeout(bridge: BeelineBridge, tab_id: int, data_url: str):
+    print("\n--- Test 5: Timeout Handling ---")
+    await bridge.navigate(tab_id, data_url, wait_until="load")
+
+    # Very short timeout - likely still completes since simple page
+    start = time.perf_counter()
+    result = await bridge.screenshot(tab_id, timeout_s=0.001)
+    elapsed = time.perf_counter() - start
+
+    if not result.get("ok"):
+        err = result.get("error", "")
+        if "timed out" in err or "cancelled" in err:
+            print(f"  ✓ PASS: Timeout handled gracefully: {err!r}")
+            return True
+        else:
+            print(f"  ⚠ Fast enough to beat timeout: {err!r} in {elapsed:.3f}s")
+            return True  # Not a failure, just fast
+    else:
+        print(f"  ⚠ Screenshot completed before timeout ({elapsed:.3f}s) - too fast to test timeout")
+        return True  # Still ok, just very fast
+
+
+async def test_screenshot_complex_site(bridge: BeelineBridge, tab_id: int):
+    print("\n--- Test 6: Complex Site (example.com) ---")
+    await bridge.navigate(tab_id, "https://example.com", wait_until="load")
+    await asyncio.sleep(1)
+
+    start = time.perf_counter()
+    result = await bridge.screenshot(tab_id)
+    elapsed = time.perf_counter() - start
+
+    ok = result.get("ok")
+    data = result.get("data", "")
+
+    print(f"  ok={ok}, elapsed={elapsed:.3f}s, data_len={len(data)}")
+    if ok and check_png(data):
+        print("  ✓ PASS: Screenshot on real site works")
+        return True
+    else:
+        print(f"  ✗ FAIL: {result.get('error', 'bad data')}")
+        return False
+
+
+async def main():
+    print("=" * 70)
+    print("TEST #15: Screenshot Functionality")
+    print("=" * 70)
+
+    bridge = BeelineBridge()
+
+    try:
+        await bridge.start()
+
+        for i in range(10):
+            await asyncio.sleep(1)
+            if bridge.is_connected:
+                print("✓ Extension connected!")
+                break
+            print(f"Waiting for extension... ({i+1}/10)")
+        else:
+            print("✗ Extension not connected. Ensure Chrome with Beeline extension is running.")
+            return
+
+        context = await bridge.create_context(CONTEXT_NAME)
+        tab_id = context.get("tabId")
+        group_id = context.get("groupId")
+        print(f"✓ Created tab: {tab_id}")
+
+        data_url = f"data:text/html;base64,{base64.b64encode(SIMPLE_HTML.encode()).decode()}"
+
+        results = {
+            "basic": await test_basic_screenshot(bridge, tab_id, data_url),
+            "full_page": await test_full_page_screenshot(bridge, tab_id, data_url),
+            "selector": await test_selector_screenshot(bridge, tab_id, data_url),
+            "metadata": await test_screenshot_url_metadata(bridge, tab_id),
+            "timeout": await test_screenshot_timeout(bridge, tab_id, data_url),
+            "complex_site": await test_screenshot_complex_site(bridge, tab_id),
+        }
+
+        print("\n" + "=" * 70)
+        print("SUMMARY")
+        print("=" * 70)
+        for name, passed in results.items():
+            status = "✓ PASS" if passed else "✗ FAIL"
+            print(f"  {status}: {name}")
+
+        passed_count = sum(1 for v in results.values() if v)
+        total = len(results)
+        print(f"\n  {passed_count}/{total} tests passed")
+
+        await bridge.destroy_context(group_id)
+        print("\n✓ Context destroyed")
+
+    finally:
+        await bridge.stop()
+        print("✓ Bridge stopped")
+
+
+if __name__ == "__main__":
+    asyncio.run(main())

.claude/skills/browser-edge-cases/scripts/test_case.py

@@ -0,0 +1,327 @@
+#!/usr/bin/env python
+"""
+Browser Edge Case Test Template
+
+This script provides a template for testing and debugging browser tool failures
+on specific websites. Use this to reproduce, isolate, and verify fixes.
+
+Usage:
+    1. Copy this file: cp test_case.py test_#[number]_[site].py
+    2. Fill in the CONFIG section with your test details
+    3. Run: uv run python test_#[number]_[site].py
+
+Example:
+    uv run python test_01_linkedin_scroll.py
+"""
+
+import asyncio
+import sys
+import time
+from pathlib import Path
+
+# Add tools to path
+sys.path.insert(0, str(Path(__file__).parent.parent.parent.parent / "tools" / "src"))
+
+from gcu.browser.bridge import BeelineBridge
+
+# ═══════════════════════════════════════════════════════════════════════════════
+# CONFIG: Fill in these values for your test case
+# ═══════════════════════════════════════════════════════════════════════════════
+
+TEST_CASE = {
+    "number": 1,
+    "name": "LinkedIn Nested Scroll Container",
+    "site": "https://www.linkedin.com/feed",
+    "simple_site": "https://example.com",
+    "category": "scroll",  # scroll, click, input, snapshot, navigation
+    "symptom": "scroll() returns success but page doesn't move",
+}
+
+BRIDGE_PORT = 9229
+CONTEXT_NAME = "edge-case-test"
+
+
+# ═══════════════════════════════════════════════════════════════════════════════
+# TEST FUNCTIONS
+# ═══════════════════════════════════════════════════════════════════════════════
+
+
+async def test_simple_site(bridge: BeelineBridge, tab_id: int) -> dict:
+    """Test that the tool works on a simple site (baseline)."""
+    print("\n--- Baseline Test (Simple Site) ---")
+
+    await bridge.navigate(tab_id, TEST_CASE["simple_site"], wait_until="load")
+    await asyncio.sleep(1)
+
+    # Adjust this based on category
+    if TEST_CASE["category"] == "scroll":
+        result = await bridge.scroll(tab_id, "down", 100)
+        print(f"  Scroll result: {result}")
+        return result
+    elif TEST_CASE["category"] == "click":
+        # Add click test
+        pass
+    elif TEST_CASE["category"] == "snapshot":
+        result = await bridge.snapshot(tab_id, timeout_s=5.0)
+        print(f"  Snapshot length: {len(result.get('tree', ''))}")
+        return result
+
+    return {"ok": True}
+
+
+async def test_problematic_site(bridge: BeelineBridge, tab_id: int) -> dict:
+    """Test the tool on the problematic site."""
+    print("\n--- Problem Site Test ---")
+
+    await bridge.navigate(tab_id, TEST_CASE["site"], wait_until="load", timeout_ms=30000)
+    await asyncio.sleep(2)
+
+    # Adjust this based on category
+    if TEST_CASE["category"] == "scroll":
+        # Get scroll positions before
+        before = await bridge.evaluate(
+            tab_id,
+            """
+            (function() {
+                const results = { window: { y: window.scrollY } };
+                document.querySelectorAll('*').forEach((el, i) => {
+                    const style = getComputedStyle(el);
+                    if ((style.overflowY === 'scroll' || style.overflowY === 'auto') &&
+                        el.scrollHeight > el.clientHeight) {
+                        results['el_' + i] = {
+                            tag: el.tagName,
+                            scrollTop: el.scrollTop,
+                            class: el.className.substring(0, 30)
+                        };
+                    }
+                });
+                return results;
+            })();
+        """
+        )
+        print(f"  Before scroll: {before.get('result', {})}")
+
+        # Try to scroll
+        result = await bridge.scroll(tab_id, "down", 500)
+        print(f"  Scroll result: {result}")
+
+        await asyncio.sleep(1)
+
+        # Get scroll positions after
+        after = await bridge.evaluate(
+            tab_id,
+            """
+            (function() {
+                const results = { window: { y: window.scrollY } };
+                document.querySelectorAll('*').forEach((el, i) => {
+                    const style = getComputedStyle(el);
+                    if ((style.overflowY === 'scroll' || style.overflowY === 'auto') &&
+                        el.scrollHeight > el.clientHeight) {
+                        results['el_' + i] = {
+                            tag: el.tagName,
+                            scrollTop: el.scrollTop,
+                            class: el.className.substring(0, 30)
+                        };
+                    }
+                });
+                return results;
+            })();
+        """
+        )
+        print(f"  After scroll: {after.get('result', {})}")
+
+        # Check if anything changed
+        before_data = before.get("result", {}) or {}
+        after_data = after.get("result", {}) or {}
+
+        changed = False
+        for key in after_data:
+            if key in before_data:
+                b_val = before_data[key].get("scrollTop", 0) if isinstance(before_data[key], dict) else 0
+                a_val = after_data[key].get("scrollTop", 0) if isinstance(after_data[key], dict) else 0
+                if a_val != b_val:
+                    print(f"  ✓ CHANGE DETECTED: {key} scrolled from {b_val} to {a_val}")
+                    changed = True
+
+        if not changed:
+            print("  ✗ NO CHANGE: Scroll did not affect any container")
+
+        return {"ok": changed, "scroll_result": result}
+
+    elif TEST_CASE["category"] == "snapshot":
+        start = time.perf_counter()
+        try:
+            result = await bridge.snapshot(tab_id, timeout_s=15.0)
+            elapsed = time.perf_counter() - start
+            tree_len = len(result.get("tree", ""))
+            print(f"  Snapshot completed in {elapsed:.2f}s, {tree_len} chars")
+            return {"ok": True, "elapsed": elapsed, "tree_length": tree_len}
+        except asyncio.TimeoutError:
+            print("  ✗ SNAPSHOT TIMED OUT")
+            return {"ok": False, "error": "timeout"}
+
+    return {"ok": True}
+
+
+async def detect_root_cause(bridge: BeelineBridge, tab_id: int) -> dict:
+    """Run detection scripts to identify the root cause."""
+    print("\n--- Root Cause Detection ---")
+
+    detections = {}
+
+    # Detection 1: Nested scrollable containers
+    scroll_check = await bridge.evaluate(
+        tab_id,
+        """
+        (function() {
+            const candidates = [];
+            document.querySelectorAll('*').forEach(el => {
+                const style = getComputedStyle(el);
+                if (style.overflow.includes('scroll') || style.overflow.includes('auto')) {
+                    const rect = el.getBoundingClientRect();
+                    if (rect.width > 100 && rect.height > 100) {
+                        candidates.push({
+                            tag: el.tagName,
+                            area: rect.width * rect.height,
+                            class: el.className.substring(0, 30)
+                        });
+                    }
+                }
+            });
+            candidates.sort((a, b) => b.area - a.area);
+            return {
+                count: candidates.length,
+                largest: candidates[0]
+            };
+        })();
+    """
+    )
+    detections["nested_scroll"] = scroll_check.get("result", {})
+    print(f"  Nested scroll containers: {detections['nested_scroll']}")
+
+    # Detection 2: Shadow DOM
+    shadow_check = await bridge.evaluate(
+        tab_id,
+        """
+        (function() {
+            const withShadow = [];
+            document.querySelectorAll('*').forEach(el => {
+                if (el.shadowRoot) {
+                    withShadow.push(el.tagName);
+                }
+            });
+            return { count: withShadow.length, elements: withShadow.slice(0, 5) };
+        })();
+    """
+    )
+    detections["shadow_dom"] = shadow_check.get("result", {})
+    print(f"  Shadow DOM: {detections['shadow_dom']}")
+
+    # Detection 3: iframes
+    iframe_check = await bridge.evaluate(
+        tab_id,
+        """
+        (function() {
+            const iframes = document.querySelectorAll('iframe');
+            return { count: iframes.length };
+        })();
+    """
+    )
+    detections["iframes"] = iframe_check.get("result", {})
+    print(f"  iframes: {detections['iframes']}")
+
+    # Detection 4: DOM size
+    dom_check = await bridge.evaluate(
+        tab_id,
+        """
+        (function() {
+            return {
+                elements: document.querySelectorAll('*').length,
+                body_children: document.body.children.length
+            };
+        })();
+    """
+    )
+    detections["dom_size"] = dom_check.get("result", {})
+    print(f"  DOM size: {detections['dom_size']}")
+
+    # Detection 5: Framework detection
+    framework_check = await bridge.evaluate(
+        tab_id,
+        """
+        (function() {
+            return {
+                react: !!document.querySelector('[data-reactroot], [data-reactid]'),
+                vue: !!document.querySelector('[data-v-]'),
+                angular: !!document.querySelector('[ng-app], [ng-version]')
+            };
+        })();
+    """
+    )
+    detections["frameworks"] = framework_check.get("result", {})
+    print(f"  Frameworks: {detections['frameworks']}")
+
+    return detections
+
+
+# ═══════════════════════════════════════════════════════════════════════════════
+# MAIN
+# ═══════════════════════════════════════════════════════════════════════════════
+
+
+async def main():
+    print("=" * 70)
+    print(f"EDGE CASE TEST #{TEST_CASE['number']}: {TEST_CASE['name']}")
+    print("=" * 70)
+    print(f"Site: {TEST_CASE['site']}")
+    print(f"Category: {TEST_CASE['category']}")
+    print(f"Symptom: {TEST_CASE['symptom']}")
+
+    bridge = BeelineBridge()
+
+    try:
+        print("\n--- Starting Bridge ---")
+        await bridge.start()
+
+        # Wait for extension connection
+        for i in range(10):
+            await asyncio.sleep(1)
+            if bridge.is_connected:
+                print("✓ Extension connected!")
+                break
+            print(f"Waiting for extension... ({i+1}/10)")
+        else:
+            print("✗ Extension not connected. Ensure Chrome with Beeline extension is running.")
+            return
+
+        # Create browser context
+        context = await bridge.create_context(CONTEXT_NAME)
+        tab_id = context.get("tabId")
+        group_id = context.get("groupId")
+        print(f"✓ Created tab: {tab_id}")
+
+        # Run tests
+        baseline_result = await test_simple_site(bridge, tab_id)
+        problem_result = await test_problematic_site(bridge, tab_id)
+        detections = await detect_root_cause(bridge, tab_id)
+
+        # Summary
+        print("\n" + "=" * 70)
+        print("SUMMARY")
+        print("=" * 70)
+        print(f"Baseline test: {'✓ PASS' if baseline_result.get('ok') else '✗ FAIL'}")
+        print(f"Problem test: {'✓ PASS' if problem_result.get('ok') else '✗ FAIL'}")
+        print(f"Root cause indicators: {list(k for k, v in detections.items() if v)}")
+
+        # Cleanup
+        print("\n--- Cleanup ---")
+        await bridge.destroy_context(group_id)
+        print("✓ Context destroyed")
+
+    finally:
+        await bridge.stop()
+        print("✓ Bridge stopped")
+
+
+if __name__ == "__main__":
+    asyncio.run(main())

.claude/skills/building-agents-construction/SKILL.md

@@ -1,361 +0,0 @@
----
-name: building-agents-construction
-description: Step-by-step guide for building goal-driven agents. Creates package structure, defines goals, adds nodes, connects edges, and finalizes agent class. Use when actively building an agent.
-license: Apache-2.0
-metadata:
-  author: hive
-  version: "2.0"
-  type: procedural
-  part_of: building-agents
-  requires: building-agents-core
----
-
-# Agent Construction - EXECUTE THESE STEPS
-
-**THIS IS AN EXECUTABLE WORKFLOW. DO NOT DISPLAY THIS FILE. EXECUTE THE STEPS BELOW.**
-
-When this skill is loaded, IMMEDIATELY begin executing Step 1. Do not explain what you will do - just do it.
-
----
-
-## STEP 1: Initialize Build Environment
-
-**EXECUTE THESE TOOL CALLS NOW:**
-
-1. Register the hive-tools MCP server:
-
-```
-mcp__agent-builder__add_mcp_server(
-    name="hive-tools",
-    transport="stdio",
-    command="python",
-    args='["mcp_server.py", "--stdio"]',
-    cwd="tools",
-    description="Hive tools MCP server"
-)
-```
-
-2. Create a build session (replace AGENT_NAME with the user's requested agent name in snake_case):
-
-```
-mcp__agent-builder__create_session(name="AGENT_NAME")
-```
-
-3. Discover available tools:
-
-```
-mcp__agent-builder__list_mcp_tools()
-```
-
-4. Create the package directory:
-
-```
-mkdir -p exports/AGENT_NAME/nodes
-```
-
-**AFTER completing these calls**, tell the user:
-
-> ✅ Build environment initialized
->
-> - Session created
-> - Available tools: [list the tools from step 3]
->
-> Proceeding to define the agent goal...
-
-**THEN immediately proceed to STEP 2.**
-
----
-
-## STEP 2: Define and Approve Goal
-
-**PROPOSE a goal to the user.** Based on what they asked for, propose:
-
-- Goal ID (kebab-case)
-- Goal name
-- Goal description
-- 3-5 success criteria (each with: id, description, metric, target, weight)
-- 2-4 constraints (each with: id, description, constraint_type, category)
-
-**FORMAT your proposal as a clear summary, then ask for approval:**
-
-> **Proposed Goal: [Name]**
->
-> [Description]
->
-> **Success Criteria:**
->
-> 1. [criterion 1]
-> 2. [criterion 2]
->    ...
->
-> **Constraints:**
->
-> 1. [constraint 1]
-> 2. [constraint 2]
->    ...
-
-**THEN call AskUserQuestion:**
-
-```
-AskUserQuestion(questions=[{
-    "question": "Do you approve this goal definition?",
-    "header": "Goal",
-    "options": [
-        {"label": "Approve", "description": "Goal looks good, proceed"},
-        {"label": "Modify", "description": "I want to change something"}
-    ],
-    "multiSelect": false
-}])
-```
-
-**WAIT for user response.**
-
-- If **Approve**: Call `mcp__agent-builder__set_goal(...)` with the goal details, then proceed to STEP 3
-- If **Modify**: Ask what they want to change, update proposal, ask again
-
----
-
-## STEP 3: Design Node Workflow
-
-**BEFORE designing nodes**, review the available tools from Step 1. Nodes can ONLY use tools that exist.
-
-**DESIGN the workflow** as a series of nodes. For each node, determine:
-
-- node_id (kebab-case)
-- name
-- description
-- node_type: `"llm_generate"` (no tools) or `"llm_tool_use"` (uses tools)
-- input_keys (what data this node receives)
-- output_keys (what data this node produces)
-- tools (ONLY tools that exist - empty list for llm_generate)
-- system_prompt
-
-**PRESENT the workflow to the user:**
-
-> **Proposed Workflow: [N] nodes**
->
-> 1. **[node-id]** - [description]
->
->    - Type: [llm_generate/llm_tool_use]
->    - Input: [keys]
->    - Output: [keys]
->    - Tools: [tools or "none"]
->
-> 2. **[node-id]** - [description]
->    ...
->
-> **Flow:** node1 → node2 → node3 → ...
-
-**THEN call AskUserQuestion:**
-
-```
-AskUserQuestion(questions=[{
-    "question": "Do you approve this workflow design?",
-    "header": "Workflow",
-    "options": [
-        {"label": "Approve", "description": "Workflow looks good, proceed to build nodes"},
-        {"label": "Modify", "description": "I want to change the workflow"}
-    ],
-    "multiSelect": false
-}])
-```
-
-**WAIT for user response.**
-
-- If **Approve**: Proceed to STEP 4
-- If **Modify**: Ask what they want to change, update design, ask again
-
----
-
-## STEP 4: Build Nodes One by One
-
-**FOR EACH node in the approved workflow:**
-
-1. **Call** `mcp__agent-builder__add_node(...)` with the node details
-
-   - input_keys and output_keys must be JSON strings: `'["key1", "key2"]'`
-   - tools must be a JSON string: `'["tool1"]'` or `'[]'`
-
-2. **Call** `mcp__agent-builder__test_node(...)` to validate:
-
-```
-mcp__agent-builder__test_node(
-    node_id="the-node-id",
-    test_input='{"key": "test value"}',
-    mock_llm_response='{"output_key": "test output"}'
-)
-```
-
-3. **Check result:**
-
-   - If valid: Tell user "✅ Node [id] validated" and continue to next node
-   - If invalid: Show errors, fix the node, re-validate
-
-4. **Show progress** after each node:
-
-```
-mcp__agent-builder__get_session_status()
-```
-
-> ✅ Node [X] of [Y] complete: [node-id]
-
-**AFTER all nodes are added and validated**, proceed to STEP 5.
-
----
-
-## STEP 5: Connect Edges
-
-**DETERMINE the edges** based on the workflow flow. For each connection:
-
-- edge_id (kebab-case)
-- source (node that outputs)
-- target (node that receives)
-- condition: `"on_success"`, `"always"`, `"on_failure"`, or `"conditional"`
-- condition_expr (Python expression, only if conditional)
-- priority (integer, lower = higher priority)
-
-**FOR EACH edge, call:**
-
-```
-mcp__agent-builder__add_edge(
-    edge_id="source-to-target",
-    source="source-node-id",
-    target="target-node-id",
-    condition="on_success",
-    condition_expr="",
-    priority=1
-)
-```
-
-**AFTER all edges are added, validate the graph:**
-
-```
-mcp__agent-builder__validate_graph()
-```
-
-- If valid: Tell user "✅ Graph structure validated" and proceed to STEP 6
-- If invalid: Show errors, fix edges, re-validate
-
----
-
-## STEP 6: Generate Agent Package
-
-**EXPORT the graph data:**
-
-```
-mcp__agent-builder__export_graph()
-```
-
-This returns JSON with all the goal, nodes, edges, and MCP server configurations.
-
-**THEN write the Python package files** using the exported data. Create these files in `exports/AGENT_NAME/`:
-
-1. `config.py` - Runtime configuration with model settings
-2. `nodes/__init__.py` - All NodeSpec definitions
-3. `agent.py` - Goal, edges, graph config, and agent class
-4. `__init__.py` - Package exports
-5. `__main__.py` - CLI interface
-6. `mcp_servers.json` - MCP server configurations
-7. `README.md` - Usage documentation
-
-**IMPORTANT entry_points format:**
-
-- MUST be: `{"start": "first-node-id"}`
-- NOT: `{"first-node-id": ["input_keys"]}` (WRONG)
-- NOT: `{"first-node-id"}` (WRONG - this is a set)
-
-**Use the example agent** at `.claude/skills/building-agents-construction/examples/online_research_agent/` as a template for file structure and patterns.
-
-**AFTER writing all files, tell the user:**
-
-> ✅ Agent package created: `exports/AGENT_NAME/`
->
-> **Files generated:**
->
-> - `__init__.py` - Package exports
-> - `agent.py` - Goal, nodes, edges, agent class
-> - `config.py` - Runtime configuration
-> - `__main__.py` - CLI interface
-> - `nodes/__init__.py` - Node definitions
-> - `mcp_servers.json` - MCP server config
-> - `README.md` - Usage documentation
->
-> **Test your agent:**
->
-> ```bash
-> cd /home/timothy/oss/hive
-> PYTHONPATH=core:exports python -m AGENT_NAME validate
-> PYTHONPATH=core:exports python -m AGENT_NAME info
-> ```
-
----
-
-## STEP 7: Verify and Test
-
-**RUN validation:**
-
-```bash
-cd /home/timothy/oss/hive && PYTHONPATH=core:exports python -m AGENT_NAME validate
-```
-
-- If valid: Agent is complete!
-- If errors: Fix the issues and re-run
-
-**SHOW final session summary:**
-
-```
-mcp__agent-builder__get_session_status()
-```
-
-**TELL the user the agent is ready** and suggest next steps:
-
-- Run with mock mode to test without API calls
-- Use `/testing-agent` skill for comprehensive testing
-- Use `/setup-credentials` if the agent needs API keys
-
----
-
-## REFERENCE: Node Types
-
-| Type           | tools param            | Use when                                       |
-| -------------- | ---------------------- | ---------------------------------------------- |
-| `llm_generate` | `'[]'`                 | Pure reasoning, JSON output, no external calls |
-| `llm_tool_use` | `'["tool1", "tool2"]'` | Needs to call MCP tools                        |
-
----
-
-## REFERENCE: Edge Conditions
-
-| Condition     | When edge is followed                 |
-| ------------- | ------------------------------------- |
-| `on_success`  | Source node completed successfully    |
-| `on_failure`  | Source node failed                    |
-| `always`      | Always, regardless of success/failure |
-| `conditional` | When condition_expr evaluates to True |
-
----
-
-## REFERENCE: System Prompt Best Practice
-
-For nodes with JSON output, include this in the system_prompt:
-
-```
-CRITICAL: Return ONLY raw JSON. NO markdown, NO code blocks.
-Just the JSON object starting with { and ending with }.
-
-Return this exact structure:
-{
-  "key1": "...",
-  "key2": "..."
-}
-```
-
----
-
-## COMMON MISTAKES TO AVOID
-
-1. **Using tools that don't exist** - Always check `mcp__agent-builder__list_mcp_tools()` first
-2. **Wrong entry_points format** - Must be `{"start": "node-id"}`, NOT a set or list
-3. **Skipping validation** - Always validate nodes and graph before proceeding
-4. **Not waiting for approval** - Always ask user before major steps
-5. **Displaying this file** - Execute the steps, don't show documentation

.claude/skills/building-agents-construction/examples/online_research_agent/README.md

@@ -1,80 +0,0 @@
-# Online Research Agent
-
-Deep-dive research agent that searches 10+ sources and produces comprehensive narrative reports with citations.
-
-## Features
-
-- Generates multiple search queries from a topic
-- Searches and fetches 15+ web sources
-- Evaluates and ranks sources by relevance
-- Synthesizes findings into themes
-- Writes narrative report with numbered citations
-- Quality checks for uncited claims
-- Saves report to local markdown file
-
-## Usage
-
-### CLI
-
-```bash
-# Show agent info
-python -m online_research_agent info
-
-# Validate structure
-python -m online_research_agent validate
-
-# Run research on a topic
-python -m online_research_agent run --topic "impact of AI on healthcare"
-
-# Interactive shell
-python -m online_research_agent shell
-```
-
-### Python API
-
-```python
-from online_research_agent import default_agent
-
-# Simple usage
-result = await default_agent.run({"topic": "climate change solutions"})
-
-# Check output
-if result.success:
-    print(f"Report saved to: {result.output['file_path']}")
-    print(result.output['final_report'])
-```
-
-## Workflow
-
-```
-parse-query → search-sources → fetch-content → evaluate-sources
-                                                      ↓
-                                write-report ← synthesize-findings
-                                      ↓
-                               quality-check → save-report
-```
-
-## Output
-
-Reports are saved to `./research_reports/` as markdown files with:
-
-1. Executive Summary
-2. Introduction
-3. Key Findings (by theme)
-4. Analysis
-5. Conclusion
-6. References
-
-## Requirements
-
-- Python 3.11+
-- LLM provider API key (Groq, Cerebras, etc.)
-- Internet access for web search/fetch
-
-## Configuration
-
-Edit `config.py` to change:
-
-- `model`: LLM model (default: groq/moonshotai/kimi-k2-instruct-0905)
-- `temperature`: Generation temperature (default: 0.7)
-- `max_tokens`: Max tokens per response (default: 16384)

.claude/skills/building-agents-construction/examples/online_research_agent/__init__.py

@@ -1,23 +0,0 @@
-"""
-Online Research Agent - Deep-dive research with narrative reports.
-
-Research any topic by searching multiple sources, synthesizing information,
-and producing a well-structured narrative report with citations.
-"""
-
-from .agent import OnlineResearchAgent, default_agent, goal, nodes, edges
-from .config import RuntimeConfig, AgentMetadata, default_config, metadata
-
-__version__ = "1.0.0"
-
-__all__ = [
-    "OnlineResearchAgent",
-    "default_agent",
-    "goal",
-    "nodes",
-    "edges",
-    "RuntimeConfig",
-    "AgentMetadata",
-    "default_config",
-    "metadata",
-]

.claude/skills/building-agents-construction/examples/online_research_agent/agent.py

@@ -1,429 +0,0 @@
-"""Agent graph construction for Online Research Agent."""
-
-from framework.graph import EdgeSpec, EdgeCondition, Goal, SuccessCriterion, Constraint
-from framework.graph.edge import GraphSpec
-from framework.graph.executor import ExecutionResult
-from framework.runtime.agent_runtime import AgentRuntime, create_agent_runtime
-from framework.runtime.execution_stream import EntryPointSpec
-from framework.llm import LiteLLMProvider
-from framework.runner.tool_registry import ToolRegistry
-
-from .config import default_config, metadata
-from .nodes import (
-    parse_query_node,
-    search_sources_node,
-    fetch_content_node,
-    evaluate_sources_node,
-    synthesize_findings_node,
-    write_report_node,
-    quality_check_node,
-    save_report_node,
-)
-
-# Goal definition
-goal = Goal(
-    id="comprehensive-online-research",
-    name="Comprehensive Online Research",
-    description="Research any topic by searching multiple sources, synthesizing information, and producing a well-structured narrative report with citations.",
-    success_criteria=[
-        SuccessCriterion(
-            id="source-coverage",
-            description="Query 10+ diverse sources",
-            metric="source_count",
-            target=">=10",
-            weight=0.20,
-        ),
-        SuccessCriterion(
-            id="relevance",
-            description="All sources directly address the query",
-            metric="relevance_score",
-            target="90%",
-            weight=0.25,
-        ),
-        SuccessCriterion(
-            id="synthesis",
-            description="Synthesize findings into coherent narrative",
-            metric="coherence_score",
-            target="85%",
-            weight=0.25,
-        ),
-        SuccessCriterion(
-            id="citations",
-            description="Include citations for all claims",
-            metric="citation_coverage",
-            target="100%",
-            weight=0.15,
-        ),
-        SuccessCriterion(
-            id="actionable",
-            description="Report answers the user's question",
-            metric="answer_completeness",
-            target="90%",
-            weight=0.15,
-        ),
-    ],
-    constraints=[
-        Constraint(
-            id="no-hallucination",
-            description="Only include information found in sources",
-            constraint_type="quality",
-            category="accuracy",
-        ),
-        Constraint(
-            id="source-attribution",
-            description="Every factual claim must cite its source",
-            constraint_type="quality",
-            category="accuracy",
-        ),
-        Constraint(
-            id="recency-preference",
-            description="Prefer recent sources when relevant",
-            constraint_type="quality",
-            category="relevance",
-        ),
-        Constraint(
-            id="no-paywalled",
-            description="Avoid sources that require payment to access",
-            constraint_type="functional",
-            category="accessibility",
-        ),
-    ],
-)
-
-# Node list
-nodes = [
-    parse_query_node,
-    search_sources_node,
-    fetch_content_node,
-    evaluate_sources_node,
-    synthesize_findings_node,
-    write_report_node,
-    quality_check_node,
-    save_report_node,
-]
-
-# Edge definitions
-edges = [
-    EdgeSpec(
-        id="parse-to-search",
-        source="parse-query",
-        target="search-sources",
-        condition=EdgeCondition.ON_SUCCESS,
-        priority=1,
-    ),
-    EdgeSpec(
-        id="search-to-fetch",
-        source="search-sources",
-        target="fetch-content",
-        condition=EdgeCondition.ON_SUCCESS,
-        priority=1,
-    ),
-    EdgeSpec(
-        id="fetch-to-evaluate",
-        source="fetch-content",
-        target="evaluate-sources",
-        condition=EdgeCondition.ON_SUCCESS,
-        priority=1,
-    ),
-    EdgeSpec(
-        id="evaluate-to-synthesize",
-        source="evaluate-sources",
-        target="synthesize-findings",
-        condition=EdgeCondition.ON_SUCCESS,
-        priority=1,
-    ),
-    EdgeSpec(
-        id="synthesize-to-write",
-        source="synthesize-findings",
-        target="write-report",
-        condition=EdgeCondition.ON_SUCCESS,
-        priority=1,
-    ),
-    EdgeSpec(
-        id="write-to-quality",
-        source="write-report",
-        target="quality-check",
-        condition=EdgeCondition.ON_SUCCESS,
-        priority=1,
-    ),
-    EdgeSpec(
-        id="quality-to-save",
-        source="quality-check",
-        target="save-report",
-        condition=EdgeCondition.ON_SUCCESS,
-        priority=1,
-    ),
-]
-
-# Graph configuration
-entry_node = "parse-query"
-entry_points = {"start": "parse-query"}
-pause_nodes = []
-terminal_nodes = ["save-report"]
-
-
-class OnlineResearchAgent:
-    """
-    Online Research Agent - Deep-dive research with narrative reports.
-
-    Uses AgentRuntime for multi-entrypoint support with HITL pause/resume.
-    """
-
-    def __init__(self, config=None):
-        self.config = config or default_config
-        self.goal = goal
-        self.nodes = nodes
-        self.edges = edges
-        self.entry_node = entry_node
-        self.entry_points = entry_points
-        self.pause_nodes = pause_nodes
-        self.terminal_nodes = terminal_nodes
-        self._runtime: AgentRuntime | None = None
-        self._graph: GraphSpec | None = None
-
-    def _build_entry_point_specs(self) -> list[EntryPointSpec]:
-        """Convert entry_points dict to EntryPointSpec list."""
-        specs = []
-        for ep_id, node_id in self.entry_points.items():
-            if ep_id == "start":
-                trigger_type = "manual"
-                name = "Start"
-            elif "_resume" in ep_id:
-                trigger_type = "resume"
-                name = f"Resume from {ep_id.replace('_resume', '')}"
-            else:
-                trigger_type = "manual"
-                name = ep_id.replace("-", " ").title()
-
-            specs.append(
-                EntryPointSpec(
-                    id=ep_id,
-                    name=name,
-                    entry_node=node_id,
-                    trigger_type=trigger_type,
-                    isolation_level="shared",
-                )
-            )
-        return specs
-
-    def _create_runtime(self, mock_mode=False) -> AgentRuntime:
-        """Create AgentRuntime instance."""
-        import json
-        from pathlib import Path
-
-        # Persistent storage in ~/.hive for telemetry and run history
-        storage_path = Path.home() / ".hive" / "online_research_agent"
-        storage_path.mkdir(parents=True, exist_ok=True)
-
-        tool_registry = ToolRegistry()
-
-        # Load MCP servers (always load, needed for tool validation)
-        agent_dir = Path(__file__).parent
-        mcp_config_path = agent_dir / "mcp_servers.json"
-
-        if mcp_config_path.exists():
-            with open(mcp_config_path) as f:
-                mcp_servers = json.load(f)
-
-            for server_config in mcp_servers.get("servers", []):
-                # Resolve relative cwd paths
-                cwd = server_config.get("cwd")
-                if cwd and not Path(cwd).is_absolute():
-                    server_config["cwd"] = str(agent_dir / cwd)
-                tool_registry.register_mcp_server(server_config)
-
-        llm = None
-        if not mock_mode:
-            # LiteLLMProvider uses environment variables for API keys
-            llm = LiteLLMProvider(
-                model=self.config.model,
-                api_key=self.config.api_key,
-                api_base=self.config.api_base,
-            )
-
-        self._graph = GraphSpec(
-            id="online-research-agent-graph",
-            goal_id=self.goal.id,
-            version="1.0.0",
-            entry_node=self.entry_node,
-            entry_points=self.entry_points,
-            terminal_nodes=self.terminal_nodes,
-            pause_nodes=self.pause_nodes,
-            nodes=self.nodes,
-            edges=self.edges,
-            default_model=self.config.model,
-            max_tokens=self.config.max_tokens,
-        )
-
-        # Create AgentRuntime with all entry points
-        self._runtime = create_agent_runtime(
-            graph=self._graph,
-            goal=self.goal,
-            storage_path=storage_path,
-            entry_points=self._build_entry_point_specs(),
-            llm=llm,
-            tools=list(tool_registry.get_tools().values()),
-            tool_executor=tool_registry.get_executor(),
-        )
-
-        return self._runtime
-
-    async def start(self, mock_mode=False) -> None:
-        """Start the agent runtime."""
-        if self._runtime is None:
-            self._create_runtime(mock_mode=mock_mode)
-        await self._runtime.start()
-
-    async def stop(self) -> None:
-        """Stop the agent runtime."""
-        if self._runtime is not None:
-            await self._runtime.stop()
-
-    async def trigger(
-        self,
-        entry_point: str,
-        input_data: dict,
-        correlation_id: str | None = None,
-        session_state: dict | None = None,
-    ) -> str:
-        """
-        Trigger execution at a specific entry point (non-blocking).
-
-        Args:
-            entry_point: Entry point ID (e.g., "start", "pause-node_resume")
-            input_data: Input data for the execution
-            correlation_id: Optional ID to correlate related executions
-            session_state: Optional session state to resume from (with paused_at, memory)
-
-        Returns:
-            Execution ID for tracking
-        """
-        if self._runtime is None or not self._runtime.is_running:
-            raise RuntimeError("Agent runtime not started. Call start() first.")
-        return await self._runtime.trigger(
-            entry_point, input_data, correlation_id, session_state=session_state
-        )
-
-    async def trigger_and_wait(
-        self,
-        entry_point: str,
-        input_data: dict,
-        timeout: float | None = None,
-        session_state: dict | None = None,
-    ) -> ExecutionResult | None:
-        """
-        Trigger execution and wait for completion.
-
-        Args:
-            entry_point: Entry point ID
-            input_data: Input data for the execution
-            timeout: Maximum time to wait (seconds)
-            session_state: Optional session state to resume from (with paused_at, memory)
-
-        Returns:
-            ExecutionResult or None if timeout
-        """
-        if self._runtime is None or not self._runtime.is_running:
-            raise RuntimeError("Agent runtime not started. Call start() first.")
-        return await self._runtime.trigger_and_wait(
-            entry_point, input_data, timeout, session_state=session_state
-        )
-
-    async def run(
-        self, context: dict, mock_mode=False, session_state=None
-    ) -> ExecutionResult:
-        """
-        Run the agent (convenience method for simple single execution).
-
-        For more control, use start() + trigger_and_wait() + stop().
-        """
-        await self.start(mock_mode=mock_mode)
-        try:
-            # Determine entry point based on session_state
-            if session_state and "paused_at" in session_state:
-                paused_node = session_state["paused_at"]
-                resume_key = f"{paused_node}_resume"
-                if resume_key in self.entry_points:
-                    entry_point = resume_key
-                else:
-                    entry_point = "start"
-            else:
-                entry_point = "start"
-
-            result = await self.trigger_and_wait(
-                entry_point, context, session_state=session_state
-            )
-            return result or ExecutionResult(success=False, error="Execution timeout")
-        finally:
-            await self.stop()
-
-    async def get_goal_progress(self) -> dict:
-        """Get goal progress across all executions."""
-        if self._runtime is None:
-            raise RuntimeError("Agent runtime not started")
-        return await self._runtime.get_goal_progress()
-
-    def get_stats(self) -> dict:
-        """Get runtime statistics."""
-        if self._runtime is None:
-            return {"running": False}
-        return self._runtime.get_stats()
-
-    def info(self):
-        """Get agent information."""
-        return {
-            "name": metadata.name,
-            "version": metadata.version,
-            "description": metadata.description,
-            "goal": {
-                "name": self.goal.name,
-                "description": self.goal.description,
-            },
-            "nodes": [n.id for n in self.nodes],
-            "edges": [e.id for e in self.edges],
-            "entry_node": self.entry_node,
-            "entry_points": self.entry_points,
-            "pause_nodes": self.pause_nodes,
-            "terminal_nodes": self.terminal_nodes,
-            "multi_entrypoint": True,
-        }
-
-    def validate(self):
-        """Validate agent structure."""
-        errors = []
-        warnings = []
-
-        node_ids = {node.id for node in self.nodes}
-        for edge in self.edges:
-            if edge.source not in node_ids:
-                errors.append(f"Edge {edge.id}: source '{edge.source}' not found")
-            if edge.target not in node_ids:
-                errors.append(f"Edge {edge.id}: target '{edge.target}' not found")
-
-        if self.entry_node not in node_ids:
-            errors.append(f"Entry node '{self.entry_node}' not found")
-
-        for terminal in self.terminal_nodes:
-            if terminal not in node_ids:
-                errors.append(f"Terminal node '{terminal}' not found")
-
-        for pause in self.pause_nodes:
-            if pause not in node_ids:
-                errors.append(f"Pause node '{pause}' not found")
-
-        # Validate entry points
-        for ep_id, node_id in self.entry_points.items():
-            if node_id not in node_ids:
-                errors.append(
-                    f"Entry point '{ep_id}' references unknown node '{node_id}'"
-                )
-
-        return {
-            "valid": len(errors) == 0,
-            "errors": errors,
-            "warnings": warnings,
-        }
-
-
-# Create default instance
-default_agent = OnlineResearchAgent()

.claude/skills/building-agents-construction/examples/online_research_agent/nodes/__init__.py

@@ -1,396 +0,0 @@
-"""Node definitions for Online Research Agent."""
-
-from framework.graph import NodeSpec
-
-# Node 1: Parse Query
-parse_query_node = NodeSpec(
-    id="parse-query",
-    name="Parse Query",
-    description="Analyze the research topic and generate 3-5 diverse search queries to cover different aspects",
-    node_type="llm_generate",
-    input_keys=["topic"],
-    output_keys=["search_queries", "research_focus", "key_aspects"],
-    output_schema={
-        "research_focus": {
-            "type": "string",
-            "required": True,
-            "description": "Brief statement of what we're researching",
-        },
-        "key_aspects": {
-            "type": "array",
-            "required": True,
-            "description": "List of 3-5 key aspects to investigate",
-        },
-        "search_queries": {
-            "type": "array",
-            "required": True,
-            "description": "List of 3-5 search queries",
-        },
-    },
-    system_prompt="""\
-You are a research query strategist. Given a research topic, analyze it and generate search queries.
-
-Your task:
-1. Understand the core research question
-2. Identify 3-5 key aspects to investigate
-3. Generate 3-5 diverse search queries that will find comprehensive information
-
-CRITICAL: Return ONLY raw JSON. NO markdown, NO code blocks.
-
-Return this JSON structure:
-{
-  "research_focus": "Brief statement of what we're researching",
-  "key_aspects": ["aspect1", "aspect2", "aspect3"],
-  "search_queries": [
-    "query 1 - broad overview",
-    "query 2 - specific angle",
-    "query 3 - recent developments",
-    "query 4 - expert opinions",
-    "query 5 - data/statistics"
-  ]
-}
-""",
-    tools=[],
-    max_retries=3,
-)
-
-# Node 2: Search Sources
-search_sources_node = NodeSpec(
-    id="search-sources",
-    name="Search Sources",
-    description="Execute web searches using the generated queries to find 15+ source URLs",
-    node_type="llm_tool_use",
-    input_keys=["search_queries", "research_focus"],
-    output_keys=["source_urls", "search_results_summary"],
-    output_schema={
-        "source_urls": {
-            "type": "array",
-            "required": True,
-            "description": "List of source URLs found",
-        },
-        "search_results_summary": {
-            "type": "string",
-            "required": True,
-            "description": "Brief summary of what was found",
-        },
-    },
-    system_prompt="""\
-You are a research assistant executing web searches. Use the web_search tool to find sources.
-
-Your task:
-1. Execute each search query using web_search tool
-2. Collect URLs from search results
-3. Aim for 15+ diverse sources
-
-After searching, return JSON with found sources:
-{
-  "source_urls": ["url1", "url2", ...],
-  "search_results_summary": "Brief summary of what was found"
-}
-""",
-    tools=["web_search"],
-    max_retries=3,
-)
-
-# Node 3: Fetch Content
-fetch_content_node = NodeSpec(
-    id="fetch-content",
-    name="Fetch Content",
-    description="Fetch and extract content from the discovered source URLs",
-    node_type="llm_tool_use",
-    input_keys=["source_urls", "research_focus"],
-    output_keys=["fetched_sources", "fetch_errors"],
-    output_schema={
-        "fetched_sources": {
-            "type": "array",
-            "required": True,
-            "description": "List of fetched source objects with url, title, content",
-        },
-        "fetch_errors": {
-            "type": "array",
-            "required": True,
-            "description": "List of URLs that failed to fetch",
-        },
-    },
-    system_prompt="""\
-You are a content fetcher. Use web_scrape tool to retrieve content from URLs.
-
-Your task:
-1. Fetch content from each source URL using web_scrape tool
-2. Extract the main content relevant to the research focus
-3. Track any URLs that failed to fetch
-
-After fetching, return JSON:
-{
-  "fetched_sources": [
-    {"url": "...", "title": "...", "content": "extracted text..."},
-    ...
-  ],
-  "fetch_errors": ["url that failed", ...]
-}
-""",
-    tools=["web_scrape"],
-    max_retries=3,
-)
-
-# Node 4: Evaluate Sources
-evaluate_sources_node = NodeSpec(
-    id="evaluate-sources",
-    name="Evaluate Sources",
-    description="Score sources for relevance and quality, filter to top 10",
-    node_type="llm_generate",
-    input_keys=["fetched_sources", "research_focus", "key_aspects"],
-    output_keys=["ranked_sources", "source_analysis"],
-    output_schema={
-        "ranked_sources": {
-            "type": "array",
-            "required": True,
-            "description": "List of ranked sources with scores",
-        },
-        "source_analysis": {
-            "type": "string",
-            "required": True,
-            "description": "Overview of source quality and coverage",
-        },
-    },
-    system_prompt="""\
-You are a source evaluator. Assess each source for quality and relevance.
-
-Scoring criteria:
-- Relevance to research focus (1-10)
-- Source credibility (1-10)
-- Information depth (1-10)
-- Recency if relevant (1-10)
-
-Your task:
-1. Score each source
-2. Rank by combined score
-3. Select top 10 sources
-4. Note what each source uniquely contributes
-
-Return JSON:
-{
-  "ranked_sources": [
-    {"url": "...", "title": "...", "content": "...", "score": 8.5, "unique_value": "..."},
-    ...
-  ],
-  "source_analysis": "Overview of source quality and coverage"
-}
-""",
-    tools=[],
-    max_retries=3,
-)
-
-# Node 5: Synthesize Findings
-synthesize_findings_node = NodeSpec(
-    id="synthesize-findings",
-    name="Synthesize Findings",
-    description="Extract key facts from sources and identify common themes",
-    node_type="llm_generate",
-    input_keys=["ranked_sources", "research_focus", "key_aspects"],
-    output_keys=["key_findings", "themes", "source_citations"],
-    output_schema={
-        "key_findings": {
-            "type": "array",
-            "required": True,
-            "description": "List of key findings with sources and confidence",
-        },
-        "themes": {
-            "type": "array",
-            "required": True,
-            "description": "List of themes with descriptions and supporting sources",
-        },
-        "source_citations": {
-            "type": "object",
-            "required": True,
-            "description": "Map of facts to supporting URLs",
-        },
-    },
-    system_prompt="""\
-You are a research synthesizer. Analyze multiple sources to extract insights.
-
-Your task:
-1. Identify key facts from each source
-2. Find common themes across sources
-3. Note contradictions or debates
-4. Build a citation map (fact -> source URL)
-
-Return JSON:
-{
-  "key_findings": [
-    {"finding": "...", "sources": ["url1", "url2"], "confidence": "high/medium/low"},
-    ...
-  ],
-  "themes": [
-    {"theme": "...", "description": "...", "supporting_sources": ["url1", ...]},
-    ...
-  ],
-  "source_citations": {
-    "fact or claim": ["supporting url1", "url2"],
-    ...
-  }
-}
-""",
-    tools=[],
-    max_retries=3,
-)
-
-# Node 6: Write Report
-write_report_node = NodeSpec(
-    id="write-report",
-    name="Write Report",
-    description="Generate a narrative report with proper citations",
-    node_type="llm_generate",
-    input_keys=[
-        "key_findings",
-        "themes",
-        "source_citations",
-        "research_focus",
-        "ranked_sources",
-    ],
-    output_keys=["report_content", "references"],
-    output_schema={
-        "report_content": {
-            "type": "string",
-            "required": True,
-            "description": "Full markdown report text with citations",
-        },
-        "references": {
-            "type": "array",
-            "required": True,
-            "description": "List of reference objects with number, url, title",
-        },
-    },
-    system_prompt="""\
-You are a research report writer. Create a well-structured narrative report.
-
-Report structure:
-1. Executive Summary (2-3 paragraphs)
-2. Introduction (context and scope)
-3. Key Findings (organized by theme)
-4. Analysis (synthesis and implications)
-5. Conclusion
-6. References (numbered list of all sources)
-
-Citation format: Use numbered citations like [1], [2] that correspond to the References section.
-
-IMPORTANT:
-- Every factual claim MUST have a citation
-- Write in clear, professional prose
-- Be objective and balanced
-- Highlight areas of consensus and debate
-
-Return JSON:
-{
-  "report_content": "Full markdown report text with citations...",
-  "references": [
-    {"number": 1, "url": "...", "title": "..."},
-    ...
-  ]
-}
-""",
-    tools=[],
-    max_retries=3,
-)
-
-# Node 7: Quality Check
-quality_check_node = NodeSpec(
-    id="quality-check",
-    name="Quality Check",
-    description="Verify all claims have citations and report is coherent",
-    node_type="llm_generate",
-    input_keys=["report_content", "references", "source_citations"],
-    output_keys=["quality_score", "issues", "final_report"],
-    output_schema={
-        "quality_score": {
-            "type": "number",
-            "required": True,
-            "description": "Quality score 0-1",
-        },
-        "issues": {
-            "type": "array",
-            "required": True,
-            "description": "List of issues found and fixed",
-        },
-        "final_report": {
-            "type": "string",
-            "required": True,
-            "description": "Corrected full report",
-        },
-    },
-    system_prompt="""\
-You are a quality assurance reviewer. Check the research report for issues.
-
-Check for:
-1. Uncited claims (factual statements without [n] citation)
-2. Broken citations (references to non-existent numbers)
-3. Coherence (logical flow between sections)
-4. Completeness (all key aspects covered)
-5. Accuracy (claims match source content)
-
-If issues found, fix them in the final report.
-
-Return JSON:
-{
-  "quality_score": 0.95,
-  "issues": [
-    {"type": "uncited_claim", "location": "paragraph 3", "fixed": true},
-    ...
-  ],
-  "final_report": "Corrected full report with all issues fixed..."
-}
-""",
-    tools=[],
-    max_retries=3,
-)
-
-# Node 8: Save Report
-save_report_node = NodeSpec(
-    id="save-report",
-    name="Save Report",
-    description="Write the final report to a local markdown file",
-    node_type="llm_tool_use",
-    input_keys=["final_report", "references", "research_focus"],
-    output_keys=["file_path", "save_status"],
-    output_schema={
-        "file_path": {
-            "type": "string",
-            "required": True,
-            "description": "Path where report was saved",
-        },
-        "save_status": {
-            "type": "string",
-            "required": True,
-            "description": "Status of save operation",
-        },
-    },
-    system_prompt="""\
-You are a file manager. Save the research report to disk.
-
-Your task:
-1. Generate a filename from the research focus (slugified, with date)
-2. Use the write_to_file tool to save the report as markdown
-3. Save to the ./research_reports/ directory
-
-Filename format: research_YYYY-MM-DD_topic-slug.md
-
-Return JSON:
-{
-  "file_path": "research_reports/research_2026-01-23_topic-name.md",
-  "save_status": "success"
-}
-""",
-    tools=["write_to_file"],
-    max_retries=3,
-)
-
-__all__ = [
-    "parse_query_node",
-    "search_sources_node",
-    "fetch_content_node",
-    "evaluate_sources_node",
-    "synthesize_findings_node",
-    "write_report_node",
-    "quality_check_node",
-    "save_report_node",
-]

.claude/skills/building-agents-core/SKILL.md

@@ -1,303 +0,0 @@
----
-name: building-agents-core
-description: Core concepts for goal-driven agents - architecture, node types, tool discovery, and workflow overview. Use when starting agent development or need to understand agent fundamentals.
-license: Apache-2.0
-metadata:
-  author: hive
-  version: "1.0"
-  type: foundational
-  part_of: building-agents
----
-
-# Building Agents - Core Concepts
-
-Foundational knowledge for building goal-driven agents as Python packages.
-
-## Architecture: Python Services (Not JSON Configs)
-
-Agents are built as Python packages:
-
-```
-exports/my_agent/
-├── __init__.py          # Package exports
-├── __main__.py          # CLI (run, info, validate, shell)
-├── agent.py             # Graph construction (goal, edges, agent class)
-├── nodes/__init__.py    # Node definitions (NodeSpec)
-├── config.py            # Runtime config
-└── README.md            # Documentation
-```
-
-**Key Principle: Agent is visible and editable during build**
-
-- ✅ Files created immediately as components are approved
-- ✅ User can watch files grow in their editor
-- ✅ No session state - just direct file writes
-- ✅ No "export" step - agent is ready when build completes
-
-## Core Concepts
-
-### Goal
-
-Success criteria and constraints (written to agent.py)
-
-```python
-goal = Goal(
-    id="research-goal",
-    name="Technical Research Agent",
-    description="Research technical topics thoroughly",
-    success_criteria=[
-        SuccessCriterion(
-            id="completeness",
-            description="Cover all aspects of topic",
-            metric="coverage_score",
-            target=">=0.9",
-            weight=0.4,
-        ),
-        # 3-5 success criteria total
-    ],
-    constraints=[
-        Constraint(
-            id="accuracy",
-            description="All information must be verified",
-            constraint_type="hard",
-            category="quality",
-        ),
-        # 1-5 constraints total
-    ],
-)
-```
-
-### Node
-
-Unit of work (written to nodes/__init__.py)
-
-**Node Types:**
-
-- `llm_generate` - Text generation, parsing
-- `llm_tool_use` - Actions requiring tools
-- `router` - Conditional branching
-- `function` - Deterministic operations
-
-```python
-search_node = NodeSpec(
-    id="search-web",
-    name="Search Web",
-    description="Search for information online",
-    node_type="llm_tool_use",
-    input_keys=["query"],
-    output_keys=["search_results"],
-    system_prompt="Search the web for: {query}",
-    tools=["web_search"],
-    max_retries=3,
-)
-```
-
-### Edge
-
-Connection between nodes (written to agent.py)
-
-**Edge Conditions:**
-
-- `on_success` - Proceed if node succeeds
-- `on_failure` - Handle errors
-- `always` - Always proceed
-- `conditional` - Based on expression
-
-```python
-EdgeSpec(
-    id="search-to-analyze",
-    source="search-web",
-    target="analyze-results",
-    condition=EdgeCondition.ON_SUCCESS,
-    priority=1,
-)
-```
-
-### Pause/Resume
-
-Multi-turn conversations
-
-- **Pause nodes** - Stop execution, wait for user input
-- **Resume entry points** - Continue from pause with user's response
-
-```python
-# Example pause/resume configuration
-pause_nodes = ["request-clarification"]
-entry_points = {
-    "start": "analyze-request",
-    "request-clarification_resume": "process-clarification"
-}
-```
-
-## Tool Discovery & Validation
-
-**CRITICAL:** Before adding a node with tools, you MUST verify the tools exist.
-
-Tools are provided by MCP servers. Never assume a tool exists - always discover dynamically.
-
-### Step 1: Register MCP Server (if not already done)
-
-```python
-mcp__agent-builder__add_mcp_server(
-    name="tools",
-    transport="stdio",
-    command="python",
-    args='["mcp_server.py", "--stdio"]',
-    cwd="../tools"
-)
-```
-
-### Step 2: Discover Available Tools
-
-```python
-# List all tools from all registered servers
-mcp__agent-builder__list_mcp_tools()
-
-# Or list tools from a specific server
-mcp__agent-builder__list_mcp_tools(server_name="tools")
-```
-
-This returns available tools with their descriptions and parameters:
-
-```json
-{
-  "success": true,
-  "tools_by_server": {
-    "tools": [
-      {
-        "name": "web_search",
-        "description": "Search the web...",
-        "parameters": ["query"]
-      },
-      {
-        "name": "web_scrape",
-        "description": "Scrape a URL...",
-        "parameters": ["url"]
-      }
-    ]
-  },
-  "total_tools": 14
-}
-```
-
-### Step 3: Validate Before Adding Nodes
-
-Before writing a node with `tools=[...]`:
-
-1. Call `list_mcp_tools()` to get available tools
-2. Check each tool in your node exists in the response
-3. If a tool doesn't exist:
-   - **DO NOT proceed** with the node
-   - Inform the user: "The tool 'X' is not available. Available tools are: ..."
-   - Ask if they want to use an alternative or proceed without the tool
-
-### Tool Validation Anti-Patterns
-
-❌ **Never assume a tool exists** - always call `list_mcp_tools()` first
-❌ **Never write a node with unverified tools** - validate before writing
-❌ **Never silently drop tools** - if a tool doesn't exist, inform the user
-❌ **Never guess tool names** - use exact names from discovery response
-
-### Example Validation Flow
-
-```python
-# 1. User requests: "Add a node that searches the web"
-# 2. Discover available tools
-tools_response = mcp__agent-builder__list_mcp_tools()
-
-# 3. Check if web_search exists
-available = [t["name"] for tools in tools_response["tools_by_server"].values() for t in tools]
-if "web_search" not in available:
-    # Inform user and ask how to proceed
-    print("❌ 'web_search' not available. Available tools:", available)
-else:
-    # Proceed with node creation
-    # ...
-```
-
-## Workflow Overview: Incremental File Construction
-
-```
-1. CREATE PACKAGE → mkdir + write skeletons
-2. DEFINE GOAL → Write to agent.py + config.py
-3. FOR EACH NODE:
-   - Propose design
-   - User approves
-   - Write to nodes/__init__.py IMMEDIATELY ← FILE WRITTEN
-   - (Optional) Validate with test_node ← MCP VALIDATION
-   - User can open file and see it
-4. CONNECT EDGES → Update agent.py ← FILE WRITTEN
-   - (Optional) Validate with validate_graph ← MCP VALIDATION
-5. FINALIZE → Write agent class to agent.py ← FILE WRITTEN
-6. DONE - Agent ready at exports/my_agent/
-```
-
-**Files written immediately. MCP tools optional for validation/testing bookkeeping.**
-
-### The Key Difference
-
-**OLD (Bad):**
-
-```
-MCP add_node → Session State → MCP add_node → Session State → ...
-                                                                ↓
-                                                     MCP export_graph
-                                                                ↓
-                                                       Files appear
-```
-
-**NEW (Good):**
-
-```
-Write node to file → (Optional: MCP test_node) → Write node to file → ...
-       ↓                                               ↓
-  File visible                                    File visible
-  immediately                                     immediately
-```
-
-**Bottom line:** Use Write/Edit for construction, MCP for validation if needed.
-
-## When to Use This Skill
-
-Use building-agents-core when:
-- Starting a new agent project and need to understand fundamentals
-- Need to understand agent architecture before building
-- Want to validate tool availability before proceeding
-- Learning about node types, edges, and graph execution
-
-**Next Steps:**
-- Ready to build? → Use `building-agents-construction` skill
-- Need patterns and examples? → Use `building-agents-patterns` skill
-
-## MCP Tools for Validation
-
-After writing files, optionally use MCP tools for validation:
-
-**test_node** - Validate node configuration with mock inputs
-```python
-mcp__agent-builder__test_node(
-    node_id="search-web",
-    test_input='{"query": "test query"}',
-    mock_llm_response='{"results": "mock output"}'
-)
-```
-
-**validate_graph** - Check graph structure
-```python
-mcp__agent-builder__validate_graph()
-# Returns: unreachable nodes, missing connections, etc.
-```
-
-**create_session** - Track session state for bookkeeping
-```python
-mcp__agent-builder__create_session(session_name="my-build")
-```
-
-**Key Point:** Files are written FIRST. MCP tools are for validation only.
-
-## Related Skills
-
-- **building-agents-construction** - Step-by-step building process
-- **building-agents-patterns** - Best practices and examples
-- **agent-workflow** - Complete workflow orchestrator
-- **testing-agent** - Test and validate completed agents

.claude/skills/building-agents-patterns/SKILL.md

@@ -1,497 +0,0 @@
----
-name: building-agents-patterns
-description: Best practices, patterns, and examples for building goal-driven agents. Includes pause/resume architecture, hybrid workflows, anti-patterns, and handoff to testing. Use when optimizing agent design.
-license: Apache-2.0
-metadata:
-  author: hive
-  version: "1.0"
-  type: reference
-  part_of: building-agents
----
-
-# Building Agents - Patterns & Best Practices
-
-Design patterns, examples, and best practices for building robust goal-driven agents.
-
-**Prerequisites:** Complete agent structure using `building-agents-construction`.
-
-## Practical Example: Hybrid Workflow
-
-How to build a node using both direct file writes and optional MCP validation:
-
-```python
-# 1. WRITE TO FILE FIRST (Primary - makes it visible)
-node_code = '''
-search_node = NodeSpec(
-    id="search-web",
-    node_type="llm_tool_use",
-    input_keys=["query"],
-    output_keys=["search_results"],
-    system_prompt="Search the web for: {query}",
-    tools=["web_search"],
-)
-'''
-
-Edit(
-    file_path="exports/research_agent/nodes/__init__.py",
-    old_string="# Nodes will be added here",
-    new_string=node_code
-)
-
-print("✅ Added search_node to nodes/__init__.py")
-print("📁 Open exports/research_agent/nodes/__init__.py to see it!")
-
-# 2. OPTIONALLY VALIDATE WITH MCP (Secondary - bookkeeping)
-validation = mcp__agent-builder__test_node(
-    node_id="search-web",
-    test_input='{"query": "python tutorials"}',
-    mock_llm_response='{"search_results": [...mock results...]}'
-)
-
-print(f"✓ Validation: {validation['success']}")
-```
-
-**User experience:**
-
-- Immediately sees node in their editor (from step 1)
-- Gets validation feedback (from step 2)
-- Can edit the file directly if needed
-
-This combines visibility (files) with validation (MCP tools).
-
-## Pause/Resume Architecture
-
-For agents needing multi-turn conversations with user interaction:
-
-### Basic Pause/Resume Flow
-
-```python
-# Define pause nodes - execution stops at these nodes
-pause_nodes = ["request-clarification", "await-approval"]
-
-# Define entry points - where to resume from each pause
-entry_points = {
-    "start": "analyze-request",  # Initial entry
-    "request-clarification_resume": "process-clarification",  # Resume from clarification
-    "await-approval_resume": "execute-action",  # Resume from approval
-}
-```
-
-### Example: Multi-Turn Research Agent
-
-```python
-# Nodes
-nodes = [
-    NodeSpec(id="analyze-request", ...),
-    NodeSpec(id="request-clarification", ...),  # PAUSE NODE
-    NodeSpec(id="process-clarification", ...),
-    NodeSpec(id="generate-results", ...),
-    NodeSpec(id="await-approval", ...),  # PAUSE NODE
-    NodeSpec(id="execute-action", ...),
-]
-
-# Edges with resume flows
-edges = [
-    EdgeSpec(
-        id="analyze-to-clarify",
-        source="analyze-request",
-        target="request-clarification",
-        condition=EdgeCondition.CONDITIONAL,
-        condition_expr="needs_clarification == true",
-    ),
-    # When resumed, goes to process-clarification
-    EdgeSpec(
-        id="clarify-to-process",
-        source="request-clarification",
-        target="process-clarification",
-        condition=EdgeCondition.ALWAYS,
-    ),
-    EdgeSpec(
-        id="results-to-approval",
-        source="generate-results",
-        target="await-approval",
-        condition=EdgeCondition.ALWAYS,
-    ),
-    # When resumed, goes to execute-action
-    EdgeSpec(
-        id="approval-to-execute",
-        source="await-approval",
-        target="execute-action",
-        condition=EdgeCondition.ALWAYS,
-    ),
-]
-
-# Configuration
-pause_nodes = ["request-clarification", "await-approval"]
-entry_points = {
-    "start": "analyze-request",
-    "request-clarification_resume": "process-clarification",
-    "await-approval_resume": "execute-action",
-}
-```
-
-### Running Pause/Resume Agents
-
-```python
-# Initial run - will pause at first pause node
-result1 = await agent.run(
-    context={"query": "research topic"},
-    session_state=None
-)
-
-# Check if paused
-if result1.paused_at:
-    print(f"Paused at: {result1.paused_at}")
-
-    # Resume with user input
-    result2 = await agent.run(
-        context={"user_response": "clarification details"},
-        session_state=result1.session_state  # Pass previous state
-    )
-```
-
-## Anti-Patterns
-
-### What NOT to Do
-
-❌ **Don't rely on `export_graph`** - Write files immediately, not at end
-```python
-# BAD: Building in session state, exporting at end
-mcp__agent-builder__add_node(...)
-mcp__agent-builder__add_node(...)
-mcp__agent-builder__export_graph()  # Files appear only now
-
-# GOOD: Writing files immediately
-Write(file_path="...", content=node_code)  # File visible now
-Write(file_path="...", content=node_code)  # File visible now
-```
-
-❌ **Don't hide code in session** - Write to files as components approved
-```python
-# BAD: Accumulating changes invisibly
-session.add_component(component1)
-session.add_component(component2)
-# User can't see anything yet
-
-# GOOD: Incremental visibility
-Edit(file_path="...", ...)  # User sees change 1
-Edit(file_path="...", ...)  # User sees change 2
-```
-
-❌ **Don't wait to write files** - Agent visible from first step
-```python
-# BAD: Building everything before writing
-design_all_nodes()
-design_all_edges()
-write_everything_at_once()
-
-# GOOD: Write as you go
-write_package_structure()  # Visible
-write_goal()  # Visible
-write_node_1()  # Visible
-write_node_2()  # Visible
-```
-
-❌ **Don't batch everything** - Write incrementally
-```python
-# BAD: Batching all nodes
-nodes = [design_node_1(), design_node_2(), ...]
-write_all_nodes(nodes)
-
-# GOOD: One at a time with user feedback
-write_node_1()  # User approves
-write_node_2()  # User approves
-write_node_3()  # User approves
-```
-
-### MCP Tools - Correct Usage
-
-**MCP tools OK for:**
-✅ `test_node` - Validate node configuration with mock inputs
-✅ `validate_graph` - Check graph structure
-✅ `create_session` - Track session state for bookkeeping
-✅ Other validation tools
-
-**Just don't:** Use MCP as the primary construction method or rely on export_graph
-
-## Best Practices
-
-### 1. Show Progress After Each Write
-
-```python
-# After writing a node
-print("✅ Added analyze_request_node to nodes/__init__.py")
-print("📊 Progress: 1/6 nodes added")
-print("📁 Open exports/my_agent/nodes/__init__.py to see it!")
-```
-
-### 2. Let User Open Files During Build
-
-```python
-# Encourage file inspection
-print("✅ Goal written to agent.py")
-print("")
-print("💡 Tip: Open exports/my_agent/agent.py in your editor to see the goal!")
-```
-
-### 3. Write Incrementally - One Component at a Time
-
-```python
-# Good flow
-write_package_structure()
-show_user("Package created")
-
-write_goal()
-show_user("Goal written")
-
-for node in nodes:
-    get_approval(node)
-    write_node(node)
-    show_user(f"Node {node.id} written")
-```
-
-### 4. Test As You Build
-
-```python
-# After adding several nodes
-print("💡 You can test current state with:")
-print("  PYTHONPATH=core:exports python -m my_agent validate")
-print("  PYTHONPATH=core:exports python -m my_agent info")
-```
-
-### 5. Keep User Informed
-
-```python
-# Clear status updates
-print("🔨 Creating package structure...")
-print("✅ Package created: exports/my_agent/")
-print("")
-print("📝 Next: Define agent goal")
-```
-
-## Continuous Monitoring Agents
-
-For agents that run continuously without terminal nodes:
-
-```python
-# No terminal nodes - loops forever
-terminal_nodes = []
-
-# Workflow loops back to start
-edges = [
-    EdgeSpec(id="monitor-to-check", source="monitor", target="check-condition"),
-    EdgeSpec(id="check-to-wait", source="check-condition", target="wait"),
-    EdgeSpec(id="wait-to-monitor", source="wait", target="monitor"),  # Loop
-]
-
-# Entry node only
-entry_node = "monitor"
-entry_points = {"start": "monitor"}
-pause_nodes = []
-```
-
-**Example: File Monitor**
-
-```python
-nodes = [
-    NodeSpec(id="list-files", ...),
-    NodeSpec(id="check-new-files", node_type="router", ...),
-    NodeSpec(id="process-files", ...),
-    NodeSpec(id="wait-interval", node_type="function", ...),
-]
-
-edges = [
-    EdgeSpec(id="list-to-check", source="list-files", target="check-new-files"),
-    EdgeSpec(
-        id="check-to-process",
-        source="check-new-files",
-        target="process-files",
-        condition=EdgeCondition.CONDITIONAL,
-        condition_expr="new_files_count > 0",
-    ),
-    EdgeSpec(
-        id="check-to-wait",
-        source="check-new-files",
-        target="wait-interval",
-        condition=EdgeCondition.CONDITIONAL,
-        condition_expr="new_files_count == 0",
-    ),
-    EdgeSpec(id="process-to-wait", source="process-files", target="wait-interval"),
-    EdgeSpec(id="wait-to-list", source="wait-interval", target="list-files"),  # Loop back
-]
-
-terminal_nodes = []  # No terminal - runs forever
-```
-
-## Complex Routing Patterns
-
-### Multi-Condition Router
-
-```python
-router_node = NodeSpec(
-    id="decision-router",
-    node_type="router",
-    input_keys=["analysis_result"],
-    output_keys=["decision"],
-    system_prompt="""
-    Based on the analysis result, decide the next action:
-    - If confidence > 0.9: route to "execute"
-    - If 0.5 <= confidence <= 0.9: route to "review"
-    - If confidence < 0.5: route to "clarify"
-
-    Return: {"decision": "execute|review|clarify"}
-    """,
-)
-
-# Edges for each route
-edges = [
-    EdgeSpec(
-        id="router-to-execute",
-        source="decision-router",
-        target="execute-action",
-        condition=EdgeCondition.CONDITIONAL,
-        condition_expr="decision == 'execute'",
-        priority=1,
-    ),
-    EdgeSpec(
-        id="router-to-review",
-        source="decision-router",
-        target="human-review",
-        condition=EdgeCondition.CONDITIONAL,
-        condition_expr="decision == 'review'",
-        priority=2,
-    ),
-    EdgeSpec(
-        id="router-to-clarify",
-        source="decision-router",
-        target="request-clarification",
-        condition=EdgeCondition.CONDITIONAL,
-        condition_expr="decision == 'clarify'",
-        priority=3,
-    ),
-]
-```
-
-## Error Handling Patterns
-
-### Graceful Failure with Fallback
-
-```python
-# Primary node with error handling
-nodes = [
-    NodeSpec(id="api-call", max_retries=3, ...),
-    NodeSpec(id="fallback-cache", ...),
-    NodeSpec(id="report-error", ...),
-]
-
-edges = [
-    # Success path
-    EdgeSpec(
-        id="api-success",
-        source="api-call",
-        target="process-results",
-        condition=EdgeCondition.ON_SUCCESS,
-    ),
-    # Fallback on failure
-    EdgeSpec(
-        id="api-to-fallback",
-        source="api-call",
-        target="fallback-cache",
-        condition=EdgeCondition.ON_FAILURE,
-        priority=1,
-    ),
-    # Report if fallback also fails
-    EdgeSpec(
-        id="fallback-to-error",
-        source="fallback-cache",
-        target="report-error",
-        condition=EdgeCondition.ON_FAILURE,
-        priority=1,
-    ),
-]
-```
-
-## Performance Optimization
-
-### Parallel Node Execution
-
-```python
-# Use multiple edges from same source for parallel execution
-edges = [
-    EdgeSpec(
-        id="start-to-search1",
-        source="start",
-        target="search-source-1",
-        condition=EdgeCondition.ALWAYS,
-    ),
-    EdgeSpec(
-        id="start-to-search2",
-        source="start",
-        target="search-source-2",
-        condition=EdgeCondition.ALWAYS,
-    ),
-    EdgeSpec(
-        id="start-to-search3",
-        source="start",
-        target="search-source-3",
-        condition=EdgeCondition.ALWAYS,
-    ),
-    # Converge results
-    EdgeSpec(
-        id="search1-to-merge",
-        source="search-source-1",
-        target="merge-results",
-    ),
-    EdgeSpec(
-        id="search2-to-merge",
-        source="search-source-2",
-        target="merge-results",
-    ),
-    EdgeSpec(
-        id="search3-to-merge",
-        source="search-source-3",
-        target="merge-results",
-    ),
-]
-```
-
-## Handoff to Testing
-
-When agent is complete, transition to testing phase:
-
-```python
-print("""
-✅ Agent complete: exports/my_agent/
-
-Next steps:
-1. Switch to testing-agent skill
-2. Generate and approve tests
-3. Run evaluation
-4. Debug any failures
-
-Command: "Test the agent at exports/my_agent/"
-""")
-```
-
-### Pre-Testing Checklist
-
-Before handing off to testing-agent:
-
-- [ ] Agent structure validates: `python -m agent_name validate`
-- [ ] All nodes defined in nodes/__init__.py
-- [ ] All edges connect valid nodes
-- [ ] Entry node specified
-- [ ] Agent can be imported: `from exports.agent_name import default_agent`
-- [ ] README.md with usage instructions
-- [ ] CLI commands work (info, validate)
-
-## Related Skills
-
-- **building-agents-core** - Fundamental concepts
-- **building-agents-construction** - Step-by-step building
-- **testing-agent** - Test and validate agents
-- **agent-workflow** - Complete workflow orchestrator
-
----
-
-**Remember: Agent is actively constructed, visible the whole time. No hidden state. No surprise exports. Just transparent, incremental file building.**

.claude/skills/setup-credentials/SKILL.md

@@ -1,572 +0,0 @@
----
-name: setup-credentials
-description: Set up and install credentials for an agent. Detects missing credentials from agent config, collects them from the user, and stores them securely in the encrypted credential store at ~/.hive/credentials.
-license: Apache-2.0
-metadata:
-  author: hive
-  version: "2.1"
-  type: utility
----
-
-# Setup Credentials
-
-Interactive credential setup for agents with multiple authentication options. Detects what's missing, offers auth method choices, validates with health checks, and stores credentials securely.
-
-## When to Use
-
-- Before running or testing an agent for the first time
-- When `AgentRunner.run()` fails with "missing required credentials"
-- When a user asks to configure credentials for an agent
-- After building a new agent that uses tools requiring API keys
-
-## Workflow
-
-### Step 1: Identify the Agent
-
-Determine which agent needs credentials. The user will either:
-
-- Name the agent directly (e.g., "set up credentials for hubspot-agent")
-- Have an agent directory open (check `exports/` for agent dirs)
-- Be working on an agent in the current session
-
-Locate the agent's directory under `exports/{agent_name}/`.
-
-### Step 2: Detect Required Credentials
-
-Read the agent's configuration to determine which tools and node types it uses:
-
-```python
-from core.framework.runner import AgentRunner
-
-runner = AgentRunner.load("exports/{agent_name}")
-validation = runner.validate()
-
-# validation.missing_credentials contains env var names
-# validation.warnings contains detailed messages with help URLs
-```
-
-Alternatively, check the credential store directly:
-
-```python
-from core.framework.credentials import CredentialStore
-
-# Use encrypted storage (default: ~/.hive/credentials)
-store = CredentialStore.with_encrypted_storage()
-
-# Check what's available
-available = store.list_credentials()
-print(f"Available credentials: {available}")
-
-# Check if specific credential exists
-if store.is_available("hubspot"):
-    print("HubSpot credential found")
-else:
-    print("HubSpot credential missing")
-```
-
-To see all known credential specs (for help URLs and setup instructions):
-
-```python
-from aden_tools.credentials import CREDENTIAL_SPECS
-
-for name, spec in CREDENTIAL_SPECS.items():
-    print(f"{name}: env_var={spec.env_var}, aden={spec.aden_supported}")
-```
-
-### Step 3: Present Auth Options for Each Missing Credential
-
-For each missing credential, check what authentication methods are available:
-
-```python
-from aden_tools.credentials import CREDENTIAL_SPECS
-
-spec = CREDENTIAL_SPECS.get("hubspot")
-if spec:
-    # Determine available auth options
-    auth_options = []
-    if spec.aden_supported:
-        auth_options.append("aden")
-    if spec.direct_api_key_supported:
-        auth_options.append("direct")
-    auth_options.append("custom")  # Always available
-
-    # Get setup info
-    setup_info = {
-        "env_var": spec.env_var,
-        "description": spec.description,
-        "help_url": spec.help_url,
-        "api_key_instructions": spec.api_key_instructions,
-    }
-```
-
-Present the available options using AskUserQuestion:
-
-```
-Choose how to configure HUBSPOT_ACCESS_TOKEN:
-
-  1) Aden Authorization Server (Recommended)
-     Secure OAuth2 flow via integration.adenhq.com
-     - Quick setup with automatic token refresh
-     - No need to manage API keys manually
-
-  2) Direct API Key
-     Enter your own API key manually
-     - Requires creating a HubSpot Private App
-     - Full control over scopes and permissions
-
-  3) Custom Credential Store (Advanced)
-     Programmatic configuration for CI/CD
-     - For automated deployments
-     - Requires manual API calls
-```
-
-### Step 4: Execute Auth Flow Based on User Choice
-
-#### Option 1: Aden Authorization Server
-
-This is the recommended flow for supported integrations (HubSpot, etc.).
-
-**How Aden OAuth Works:**
-
-The ADEN_API_KEY represents a user who has already completed OAuth authorization on Aden's platform. When users sign up and connect integrations on Aden, those OAuth tokens are stored server-side. Having an ADEN_API_KEY means:
-
-1. User has an Aden account
-2. User has already authorized integrations (HubSpot, etc.) via OAuth on Aden
-3. We just need to sync those credentials down to the local credential store
-
-**4.1a. Check for ADEN_API_KEY**
-
-```python
-import os
-aden_key = os.environ.get("ADEN_API_KEY")
-```
-
-If not set, guide user to get one from Aden (this is where they do OAuth):
-
-```python
-from aden_tools.credentials import open_browser, get_aden_setup_url
-
-# Open browser to Aden - user will sign up and connect integrations there
-url = get_aden_setup_url()  # https://integration.adenhq.com/setup
-success, msg = open_browser(url)
-
-print("Please sign in to Aden and connect your integrations (HubSpot, etc.).")
-print("Once done, copy your API key and return here.")
-```
-
-Ask user to provide the ADEN_API_KEY they received.
-
-**4.1b. Save ADEN_API_KEY to Shell Config**
-
-With user approval, persist ADEN_API_KEY to their shell config:
-
-```python
-from aden_tools.credentials import (
-    detect_shell,
-    add_env_var_to_shell_config,
-    get_shell_source_command,
-)
-
-shell_type = detect_shell()  # 'bash', 'zsh', or 'unknown'
-
-# Ask user for approval before modifying shell config
-# If approved:
-success, config_path = add_env_var_to_shell_config(
-    "ADEN_API_KEY",
-    user_provided_key,
-    comment="Aden authorization server API key"
-)
-
-if success:
-    source_cmd = get_shell_source_command()
-    print(f"Saved to {config_path}")
-    print(f"Run: {source_cmd}")
-```
-
-Also save to `~/.hive/configuration.json` for the framework:
-
-```python
-import json
-from pathlib import Path
-
-config_path = Path.home() / ".hive" / "configuration.json"
-config = json.loads(config_path.read_text()) if config_path.exists() else {}
-
-config["aden"] = {
-    "api_key_configured": True,
-    "api_url": "https://api.adenhq.com"
-}
-
-config_path.parent.mkdir(parents=True, exist_ok=True)
-config_path.write_text(json.dumps(config, indent=2))
-```
-
-**4.1c. Sync Credentials from Aden Server**
-
-Since the user has already authorized integrations on Aden, use the one-liner factory method:
-
-```python
-from core.framework.credentials import CredentialStore
-
-# This single call handles everything:
-# - Creates encrypted local storage at ~/.hive/credentials
-# - Configures Aden client from ADEN_API_KEY env var
-# - Syncs all credentials from Aden server automatically
-store = CredentialStore.with_aden_sync(
-    base_url="https://api.adenhq.com",
-    auto_sync=True,  # Syncs on creation
-)
-
-# Check what was synced
-synced = store.list_credentials()
-print(f"Synced credentials: {synced}")
-
-# If the required credential wasn't synced, the user hasn't authorized it on Aden yet
-if "hubspot" not in synced:
-    print("HubSpot not found in your Aden account.")
-    print("Please visit https://integration.adenhq.com to connect HubSpot, then try again.")
-```
-
-For more control over the sync process:
-
-```python
-from core.framework.credentials import CredentialStore
-from core.framework.credentials.aden import (
-    AdenCredentialClient,
-    AdenClientConfig,
-    AdenSyncProvider,
-)
-
-# Create client (API key loaded from ADEN_API_KEY env var)
-client = AdenCredentialClient(AdenClientConfig(
-    base_url="https://api.adenhq.com",
-))
-
-# Create provider and store
-provider = AdenSyncProvider(client=client)
-store = CredentialStore.with_encrypted_storage()
-
-# Manual sync
-synced_count = provider.sync_all(store)
-print(f"Synced {synced_count} credentials from Aden")
-```
-
-**4.1d. Run Health Check**
-
-```python
-from aden_tools.credentials import check_credential_health
-
-# Get the token from the store
-cred = store.get_credential("hubspot")
-token = cred.keys["access_token"].value.get_secret_value()
-
-result = check_credential_health("hubspot", token)
-if result.valid:
-    print("HubSpot credentials validated successfully!")
-else:
-    print(f"Validation failed: {result.message}")
-    # Offer to retry the OAuth flow
-```
-
-#### Option 2: Direct API Key
-
-For users who prefer manual API key management.
-
-**4.2a. Show Setup Instructions**
-
-```python
-from aden_tools.credentials import CREDENTIAL_SPECS
-
-spec = CREDENTIAL_SPECS.get("hubspot")
-if spec and spec.api_key_instructions:
-    print(spec.api_key_instructions)
-# Output:
-# To get a HubSpot Private App token:
-# 1. Go to HubSpot Settings > Integrations > Private Apps
-# 2. Click "Create a private app"
-# 3. Name your app (e.g., "Hive Agent")
-# ...
-
-if spec and spec.help_url:
-    print(f"More info: {spec.help_url}")
-```
-
-**4.2b. Collect API Key from User**
-
-Use AskUserQuestion to securely collect the API key:
-
-```
-Please provide your HubSpot access token:
-(This will be stored securely in ~/.hive/credentials)
-```
-
-**4.2c. Run Health Check Before Storing**
-
-```python
-from aden_tools.credentials import check_credential_health
-
-result = check_credential_health("hubspot", user_provided_token)
-if not result.valid:
-    print(f"Warning: {result.message}")
-    # Ask user if they want to:
-    # 1. Try a different token
-    # 2. Continue anyway (not recommended)
-```
-
-**4.2d. Store in Encrypted Credential Store**
-
-```python
-from core.framework.credentials import CredentialStore, CredentialObject, CredentialKey
-from pydantic import SecretStr
-
-store = CredentialStore.with_encrypted_storage()
-
-cred = CredentialObject(
-    id="hubspot",
-    name="HubSpot Access Token",
-    keys={
-        "access_token": CredentialKey(
-            name="access_token",
-            value=SecretStr(user_provided_token),
-        )
-    },
-)
-store.save_credential(cred)
-```
-
-**4.2e. Export to Current Session**
-
-```bash
-export HUBSPOT_ACCESS_TOKEN="the-value"
-```
-
-#### Option 3: Custom Credential Store (Advanced)
-
-For programmatic/CI/CD setups.
-
-**4.3a. Show Documentation**
-
-```
-For advanced credential management, you can use the CredentialStore API directly:
-
-  from core.framework.credentials import CredentialStore, CredentialObject, CredentialKey
-  from pydantic import SecretStr
-
-  store = CredentialStore.with_encrypted_storage()
-
-  cred = CredentialObject(
-      id="hubspot",
-      name="HubSpot Access Token",
-      keys={"access_token": CredentialKey(name="access_token", value=SecretStr("..."))}
-  )
-  store.save_credential(cred)
-
-For CI/CD environments:
-  - Set HIVE_CREDENTIAL_KEY for encryption
-  - Pre-populate ~/.hive/credentials programmatically
-  - Or use environment variables directly (HUBSPOT_ACCESS_TOKEN)
-
-Documentation: See core/framework/credentials/README.md
-```
-
-### Step 5: Record Configuration Method
-
-Track which auth method was used for each credential in `~/.hive/configuration.json`:
-
-```python
-import json
-from pathlib import Path
-from datetime import datetime
-
-config_path = Path.home() / ".hive" / "configuration.json"
-config = json.loads(config_path.read_text()) if config_path.exists() else {}
-
-if "credential_methods" not in config:
-    config["credential_methods"] = {}
-
-config["credential_methods"]["hubspot"] = {
-    "method": "aden",  # or "direct" or "custom"
-    "configured_at": datetime.now().isoformat(),
-}
-
-config_path.write_text(json.dumps(config, indent=2))
-```
-
-### Step 6: Verify All Credentials
-
-Run validation again to confirm everything is set:
-
-```python
-runner = AgentRunner.load("exports/{agent_name}")
-validation = runner.validate()
-assert not validation.missing_credentials, "Still missing credentials!"
-```
-
-Report the result to the user.
-
-## Health Check Reference
-
-Health checks validate credentials by making lightweight API calls:
-
-| Credential     | Endpoint                                | What It Checks                    |
-| -------------- | --------------------------------------- | --------------------------------- |
-| `hubspot`      | `GET /crm/v3/objects/contacts?limit=1`  | Bearer token validity, CRM scopes |
-| `brave_search` | `GET /res/v1/web/search?q=test&count=1` | API key validity                  |
-
-```python
-from aden_tools.credentials import check_credential_health, HealthCheckResult
-
-result: HealthCheckResult = check_credential_health("hubspot", token_value)
-# result.valid: bool
-# result.message: str
-# result.details: dict (status_code, rate_limited, etc.)
-```
-
-## Encryption Key (HIVE_CREDENTIAL_KEY)
-
-The encrypted credential store requires `HIVE_CREDENTIAL_KEY` to encrypt/decrypt credentials.
-
-- If the user doesn't have one, `EncryptedFileStorage` will auto-generate one and log it
-- The user MUST persist this key (e.g., in `~/.bashrc` or a secrets manager)
-- Without this key, stored credentials cannot be decrypted
-- This is the ONLY secret that should live in `~/.bashrc` or environment config
-
-If `HIVE_CREDENTIAL_KEY` is not set:
-
-1. Let the store generate one
-2. Tell the user to save it: `export HIVE_CREDENTIAL_KEY="{generated_key}"`
-3. Recommend adding it to `~/.bashrc` or their shell profile
-
-## Security Rules
-
-- **NEVER** log, print, or echo credential values in tool output
-- **NEVER** store credentials in plaintext files, git-tracked files, or agent configs
-- **NEVER** hardcode credentials in source code
-- **ALWAYS** use `SecretStr` from Pydantic when handling credential values in Python
-- **ALWAYS** use the encrypted credential store (`~/.hive/credentials`) for persistence
-- **ALWAYS** run health checks before storing credentials (when possible)
-- **ALWAYS** verify credentials were stored by re-running validation, not by reading them back
-- When modifying `~/.bashrc` or `~/.zshrc`, confirm with the user first
-
-## Credential Sources Reference
-
-All credential specs are defined in `tools/src/aden_tools/credentials/`:
-
-| File              | Category      | Credentials                                   | Aden Supported |
-| ----------------- | ------------- | --------------------------------------------- | -------------- |
-| `llm.py`          | LLM Providers | `anthropic`                                   | No             |
-| `search.py`       | Search Tools  | `brave_search`, `google_search`, `google_cse` | No             |
-| `integrations.py` | Integrations  | `hubspot`                                     | Yes            |
-
-**Note:** Additional LLM providers (Cerebras, Groq, OpenAI) are handled by LiteLLM via environment
-variables (`CEREBRAS_API_KEY`, `GROQ_API_KEY`, `OPENAI_API_KEY`) but are not yet in CREDENTIAL_SPECS.
-Add them to `llm.py` as needed.
-
-To check what's registered:
-
-```python
-from aden_tools.credentials import CREDENTIAL_SPECS
-for name, spec in CREDENTIAL_SPECS.items():
-    print(f"{name}: aden={spec.aden_supported}, direct={spec.direct_api_key_supported}")
-```
-
-## Migration: CredentialManager → CredentialStore
-
-**CredentialManager is deprecated.** Use CredentialStore instead.
-
-| Old (Deprecated)                          | New (Recommended)                                                    |
-| ----------------------------------------- | -------------------------------------------------------------------- |
-| `CredentialManager()`                     | `CredentialStore.with_encrypted_storage()`                           |
-| `creds.get("hubspot")`                    | `store.get("hubspot")` or `store.get_key("hubspot", "access_token")` |
-| `creds.validate_for_tools(tools)`         | Use `store.is_available(cred_id)` per credential                     |
-| `creds.get_auth_options("hubspot")`       | Check `CREDENTIAL_SPECS["hubspot"].aden_supported`                   |
-| `creds.get_setup_instructions("hubspot")` | Access `CREDENTIAL_SPECS["hubspot"]` directly                        |
-
-**Why migrate?**
-
-- **CredentialStore** supports encrypted storage, multi-key credentials, template resolution, and automatic token refresh
-- **CredentialManager** only reads from environment variables and .env files (no encryption, no refresh)
-- **CredentialStoreAdapter** exists for backward compatibility during migration
-
-```python
-# Old way (deprecated)
-from aden_tools.credentials import CredentialManager
-creds = CredentialManager()
-token = creds.get("hubspot")
-
-# New way (recommended)
-from core.framework.credentials import CredentialStore
-store = CredentialStore.with_encrypted_storage()
-token = store.get("hubspot")
-
-# With Aden sync (recommended for OAuth integrations)
-store = CredentialStore.with_aden_sync()
-token = store.get_key("hubspot", "access_token")
-```
-
-## Example Session
-
-```
-User: /setup-credentials for my hubspot-agent
-
-Agent: Let me check what credentials your hubspot-agent needs.
-
-[Runs validation, finds ANTHROPIC_API_KEY and HUBSPOT_ACCESS_TOKEN missing]
-
-Agent: Your hubspot-agent requires 2 credentials:
-
-  1. ANTHROPIC_API_KEY - Anthropic API key for LLM calls
-  2. HUBSPOT_ACCESS_TOKEN - HubSpot access token
-
-Let's set these up one at a time.
-
---- Setting up ANTHROPIC_API_KEY ---
-
-This credential only supports direct API key entry.
-
-[Shows instructions for getting Anthropic API key]
-[AskUserQuestion: "Please provide your Anthropic API key:"]
-[User provides key]
-[Stores in credential store, exports to env]
-
---- Setting up HUBSPOT_ACCESS_TOKEN ---
-
-[AskUserQuestion with 3 options]
-
-Choose how to configure HUBSPOT_ACCESS_TOKEN:
-  1) Aden Authorization Server (Recommended)
-  2) Direct API Key
-  3) Custom Credential Store (Advanced)
-
-User: 1 (Aden)
-
-Agent: Great! Let me check if you have an Aden API key configured...
-
-[Checks for ADEN_API_KEY, not found]
-[Opens browser to integration.adenhq.com/setup]
-
-Agent: I've opened Aden's setup page. Please:
-  1. Sign in or create an account
-  2. Connect your HubSpot account (OAuth happens on Aden's site)
-  3. Copy the API key shown after setup
-
-[AskUserQuestion: "Please provide your Aden API key:"]
-[User provides key]
-
-Agent: Would you like me to save this to your shell config (~/.zshrc) for future sessions?
-
-[User approves]
-[Saves to ~/.zshrc and ~/.hive/configuration.json]
-
-Agent: Let me sync your credentials from Aden...
-
-[Syncs credentials from Aden server - OAuth already done on Aden's side]
-[Runs health check]
-
-Agent: HubSpot credentials validated successfully!
-
-All credentials are now configured:
-  - ANTHROPIC_API_KEY: Stored in encrypted credential store
-  - HUBSPOT_ACCESS_TOKEN: Synced from Aden (OAuth completed on Aden)
-  - Validation passed - your agent is ready to run!
-```

.claude/skills/test-reporting/SKILL.md

@@ -0,0 +1,225 @@
+# Integration Test Reporting Skill
+
+Run the Level 2 dummy agent integration test suite and produce a detailed HTML report with per-test input → outcome analysis.
+
+## Trigger
+
+User wants to run integration tests and see results:
+- `/test-reporting`
+- `/test-reporting test_component_queen_live.py`
+- `/test-reporting --all`
+
+## SOP: Running Tests
+
+### Step 1: Select Scope
+
+If the user provides a specific test file or pattern, use it. Otherwise run the full suite.
+
+```bash
+# Full suite
+cd core && echo "1" | uv run python tests/dummy_agents/run_all.py --interactive 2>&1
+
+# Specific file (requires manual provider setup)
+cd core && uv run python -c "
+import sys
+sys.path.insert(0, '.')
+from tests.dummy_agents.run_all import detect_available
+from tests.dummy_agents.conftest import set_llm_selection
+
+avail = detect_available()
+claude = [p for p in avail if 'Claude Code' in p['name']]
+if not claude:
+    avail_names = [p['name'] for p in avail]
+    raise RuntimeError(f'No Claude Code subscription. Available: {avail_names}')
+provider = claude[0]
+set_llm_selection(
+    model=provider['model'],
+    api_key=provider['api_key'],
+    extra_headers=provider.get('extra_headers'),
+    api_base=provider.get('api_base'),
+)
+
+import pytest
+sys.exit(pytest.main([
+    'tests/dummy_agents/TEST_FILE_HERE',
+    '-v', '--override-ini=asyncio_mode=auto', '--no-header', '--tb=long',
+    '--log-cli-level=WARNING', '--junitxml=/tmp/hive_test_results.xml',
+]))
+"
+```
+
+### Step 2: Collect Results
+
+After the test run completes, collect:
+1. **JUnit XML** from `--junitxml` output (if available)
+2. **stdout/stderr** from the run
+3. **Summary table** from `run_all.py` output (the Unicode table)
+
+### Step 3: Generate HTML Report
+
+Write the report to `/tmp/hive_integration_test_report.html`.
+
+The report MUST include these sections:
+
+#### Header
+- Run timestamp (ISO 8601)
+- Provider used (model name, source)
+- Total tests / passed / failed / skipped
+- Total wall-clock time
+- Overall verdict: PASS (all green) or FAIL (with count)
+
+#### Per-Test Table
+
+For EVERY test (not just failures), include a row with:
+
+| Column | Description |
+|--------|-------------|
+| Component | Test file grouping (e.g., `component_queen_live`) |
+| Test Name | Function name (e.g., `test_queen_starts_in_planning_without_worker`) |
+| Status | PASS / FAIL / SKIP / ERROR with color badge |
+| Duration | Wall-clock seconds |
+| What | One-line description of what the test verifies |
+| How | How it works (setup → action → assertion) |
+| Why | Why this test matters (what bug/behavior it catches) |
+| Input | The input data or configuration (graph spec, initial prompt, phase, etc.) |
+| Expected Outcome | What the test asserts |
+| Actual Outcome | What actually happened (PASS: matches expected / FAIL: actual vs expected) |
+| Failure Detail | For failures only: full traceback + diagnosis |
+
+#### What / How / Why Descriptions
+
+These MUST be derived from the test function's docstring and code. Read each test file to extract:
+- **What**: From the docstring first line
+- **How**: From the test body (what fixtures, what graph, what assertions)
+- **Why**: From the docstring body or "Why this matters" section in the test module
+
+Use these mappings for the component test files:
+
+```
+test_component_llm.py          → "LLM Provider" — streaming, tool calling, tokens
+test_component_tools.py        → "Tool Registry + MCP" — connection, execution
+test_component_event_loop.py   → "EventLoopNode" — iteration, output, stall
+test_component_edges.py        → "Edge Evaluation" — conditional, priority
+test_component_conversation.py → "Conversation Persistence" — storage, cursor
+test_component_escalation.py   → "Escalation Flow" — worker→queen signaling
+test_component_continuous.py   → "Continuous Mode" — conversation threading
+test_component_queen.py        → "Queen Phase (Unit)" — phase state, tools, events
+test_component_queen_live.py   → "Queen Phase (Live)" — real queen, real LLM
+test_component_queen_state_machine.py → "Queen State Machine" — edge cases, races
+test_component_worker_comms.py → "Worker Communication" — events, data flow
+test_component_strict_outcomes.py → "Strict Outcomes" — exact path, output, quality
+```
+
+#### HTML Template
+
+Use this structure:
+
+```html
+<!DOCTYPE html>
+<html lang="en">
+<head>
+<meta charset="utf-8">
+<title>Hive Integration Test Report — {timestamp}</title>
+<style>
+  :root { --pass: #22c55e; --fail: #ef4444; --skip: #f59e0b; --bg: #0f172a; --surface: #1e293b; --text: #e2e8f0; --muted: #94a3b8; --border: #334155; }
+  * { box-sizing: border-box; margin: 0; padding: 0; }
+  body { font-family: 'SF Mono', 'Fira Code', monospace; background: var(--bg); color: var(--text); padding: 2rem; line-height: 1.6; }
+  h1, h2, h3 { font-weight: 600; }
+  h1 { font-size: 1.5rem; margin-bottom: 1rem; }
+  h2 { font-size: 1.2rem; margin: 2rem 0 1rem; border-bottom: 1px solid var(--border); padding-bottom: 0.5rem; }
+  .summary { display: grid; grid-template-columns: repeat(auto-fit, minmax(150px, 1fr)); gap: 1rem; margin-bottom: 2rem; }
+  .card { background: var(--surface); padding: 1rem; border-radius: 8px; border: 1px solid var(--border); }
+  .card .label { color: var(--muted); font-size: 0.75rem; text-transform: uppercase; }
+  .card .value { font-size: 1.5rem; font-weight: 700; margin-top: 0.25rem; }
+  .card .value.pass { color: var(--pass); }
+  .card .value.fail { color: var(--fail); }
+  table { width: 100%; border-collapse: collapse; font-size: 0.8rem; }
+  th { background: var(--surface); position: sticky; top: 0; text-align: left; padding: 0.5rem; border-bottom: 2px solid var(--border); color: var(--muted); text-transform: uppercase; font-size: 0.7rem; }
+  td { padding: 0.5rem; border-bottom: 1px solid var(--border); vertical-align: top; }
+  tr:hover { background: rgba(255,255,255,0.03); }
+  .badge { display: inline-block; padding: 2px 8px; border-radius: 4px; font-size: 0.7rem; font-weight: 700; }
+  .badge.pass { background: rgba(34,197,94,0.2); color: var(--pass); }
+  .badge.fail { background: rgba(239,68,68,0.2); color: var(--fail); }
+  .badge.skip { background: rgba(245,158,11,0.2); color: var(--skip); }
+  .detail { background: #1a1a2e; padding: 0.75rem; border-radius: 4px; margin-top: 0.5rem; font-size: 0.75rem; white-space: pre-wrap; overflow-x: auto; max-height: 200px; overflow-y: auto; }
+  .component-header { background: var(--surface); padding: 0.75rem 0.5rem; font-weight: 600; font-size: 0.85rem; }
+  .meta { color: var(--muted); font-size: 0.75rem; }
+</style>
+</head>
+<body>
+<h1>Hive Integration Test Report</h1>
+<p class="meta">Generated: {timestamp} | Provider: {provider} | Duration: {duration}s</p>
+
+<div class="summary">
+  <div class="card"><div class="label">Total</div><div class="value">{total}</div></div>
+  <div class="card"><div class="label">Passed</div><div class="value pass">{passed}</div></div>
+  <div class="card"><div class="label">Failed</div><div class="value fail">{failed}</div></div>
+  <div class="card"><div class="label">Verdict</div><div class="value {verdict_class}">{verdict}</div></div>
+</div>
+
+<h2>Test Results</h2>
+<table>
+<thead>
+<tr>
+  <th>Component</th>
+  <th>Test</th>
+  <th>Status</th>
+  <th>Time</th>
+  <th>What</th>
+  <th>Input → Expected → Actual</th>
+</tr>
+</thead>
+<tbody>
+<!-- For each test: -->
+<tr>
+  <td>{component}</td>
+  <td>{test_name}</td>
+  <td><span class="badge {status_class}">{status}</span></td>
+  <td>{duration}s</td>
+  <td>{what_description}</td>
+  <td>
+    <strong>Input:</strong> {input_description}<br>
+    <strong>Expected:</strong> {expected_outcome}<br>
+    <strong>Actual:</strong> {actual_outcome}
+    <!-- If failed: -->
+    <div class="detail">{failure_traceback}</div>
+  </td>
+</tr>
+</tbody>
+</table>
+
+<h2>Failure Analysis</h2>
+<!-- Only if there are failures -->
+<p>For each failure, provide:</p>
+<ul>
+  <li><strong>Root cause:</strong> Why it failed</li>
+  <li><strong>Impact:</strong> What this means for the system</li>
+  <li><strong>Suggested fix:</strong> How to address it</li>
+</ul>
+
+</body>
+</html>
+```
+
+### Step 4: Output
+
+1. Write the HTML file to `/tmp/hive_integration_test_report.html`
+2. Print the file path so the user can open it
+3. Print a concise summary to the terminal:
+   ```
+   Test Report: /tmp/hive_integration_test_report.html
+   Result: 74/76 PASSED (2 failures)
+   Failures:
+     - parallel_merge::test_parallel_disjoint_output_keys
+     - worker::test_worker_timestamped_note_artifact
+   ```
+
+## Key Rules
+
+1. ALWAYS use `--junitxml` when running pytest to get structured results
+2. ALWAYS read the test source files to populate What/How/Why columns — do not guess
+3. For Input/Expected/Actual, extract from the test's graph spec, assertions, and result
+4. Color-code everything: green for pass, red for fail, amber for skip
+5. Include the full traceback for failures in a scrollable `<div class="detail">`
+6. Group tests by component (file name) with a visual separator
+7. The report must be self-contained HTML (no external CSS/JS dependencies)

.claude/skills/testing-agent/examples/testing-youtube-agent.md

@@ -1,351 +0,0 @@
-# Example: Testing a YouTube Research Agent
-
-This example walks through testing a YouTube research agent that finds relevant videos based on a topic.
-
-## Prerequisites
-
-- Agent built with building-agents skill at `exports/youtube-research/`
-- Goal defined with success criteria and constraints
-
-## Step 1: Load the Goal
-
-First, load the goal that was defined during the Goal stage:
-
-```json
-{
-    "id": "youtube-research",
-    "name": "YouTube Research Agent",
-    "description": "Find relevant YouTube videos on a given topic",
-    "success_criteria": [
-        {
-            "id": "find_videos",
-            "description": "Find 3-5 relevant videos",
-            "metric": "video_count",
-            "target": "3-5",
-            "weight": 1.0
-        },
-        {
-            "id": "relevance",
-            "description": "Videos must be relevant to the topic",
-            "metric": "relevance_score",
-            "target": ">0.8",
-            "weight": 0.8
-        }
-    ],
-    "constraints": [
-        {
-            "id": "api_limits",
-            "description": "Must not exceed YouTube API rate limits",
-            "constraint_type": "hard",
-            "category": "technical"
-        },
-        {
-            "id": "content_safety",
-            "description": "Must filter out inappropriate content",
-            "constraint_type": "hard",
-            "category": "safety"
-        }
-    ]
-}
-```
-
-## Step 2: Get Constraint Test Guidelines
-
-During the Goal stage (or early Eval), get test guidelines for constraints:
-
-```python
-result = generate_constraint_tests(
-    goal_id="youtube-research",
-    goal_json='<goal JSON above>',
-    agent_path="exports/youtube-research"
-)
-```
-
-**The result contains guidelines (not generated tests):**
-- `output_file`: Where to write tests
-- `file_header`: Imports and fixtures to use
-- `test_template`: Format for test functions
-- `constraints_formatted`: The constraints to test
-- `test_guidelines`: Rules for writing tests
-
-## Step 3: Write Constraint Tests
-
-Using the guidelines, write tests directly with the Write tool:
-
-```python
-# Write constraint tests using the provided file_header and guidelines
-Write(
-    file_path="exports/youtube-research/tests/test_constraints.py",
-    content='''
-"""Constraint tests for youtube-research agent."""
-
-import os
-import pytest
-from exports.youtube_research import default_agent
-
-
-pytestmark = pytest.mark.skipif(
-    not os.environ.get("ANTHROPIC_API_KEY") and not os.environ.get("MOCK_MODE"),
-    reason="API key required for real testing."
-)
-
-
-@pytest.mark.asyncio
-async def test_constraint_api_limits_respected():
-    """Verify API rate limits are not exceeded."""
-    import time
-    mock_mode = bool(os.environ.get("MOCK_MODE"))
-
-    for i in range(10):
-        result = await default_agent.run({"topic": f"test_{i}"}, mock_mode=mock_mode)
-        time.sleep(0.1)
-
-    # Should complete without rate limit errors
-    assert "rate limit" not in str(result).lower()
-
-
-@pytest.mark.asyncio
-async def test_constraint_content_safety_filter():
-    """Verify inappropriate content is filtered."""
-    mock_mode = bool(os.environ.get("MOCK_MODE"))
-    result = await default_agent.run({"topic": "general topic"}, mock_mode=mock_mode)
-
-    for video in result.videos:
-        assert video.safe_for_work is True
-        assert video.age_restricted is False
-'''
-)
-```
-
-## Step 4: Get Success Criteria Test Guidelines
-
-After the agent is built, get success criteria test guidelines:
-
-```python
-result = generate_success_tests(
-    goal_id="youtube-research",
-    goal_json='<goal JSON>',
-    node_names="search_node,filter_node,rank_node,format_node",
-    tool_names="youtube_search,video_details,channel_info",
-    agent_path="exports/youtube-research"
-)
-```
-
-## Step 5: Write Success Criteria Tests
-
-Using the guidelines, write success criteria tests:
-
-```python
-Write(
-    file_path="exports/youtube-research/tests/test_success_criteria.py",
-    content='''
-"""Success criteria tests for youtube-research agent."""
-
-import os
-import pytest
-from exports.youtube_research import default_agent
-
-
-pytestmark = pytest.mark.skipif(
-    not os.environ.get("ANTHROPIC_API_KEY") and not os.environ.get("MOCK_MODE"),
-    reason="API key required for real testing."
-)
-
-
-@pytest.mark.asyncio
-async def test_find_videos_happy_path():
-    """Test finding videos for a common topic."""
-    mock_mode = bool(os.environ.get("MOCK_MODE"))
-    result = await default_agent.run({"topic": "machine learning"}, mock_mode=mock_mode)
-
-    assert result.success
-    assert 3 <= len(result.videos) <= 5
-    assert all(v.title for v in result.videos)
-    assert all(v.video_id for v in result.videos)
-
-
-@pytest.mark.asyncio
-async def test_find_videos_minimum_boundary():
-    """Test at minimum threshold (3 videos)."""
-    mock_mode = bool(os.environ.get("MOCK_MODE"))
-    result = await default_agent.run({"topic": "niche topic xyz"}, mock_mode=mock_mode)
-
-    assert len(result.videos) >= 3
-
-
-@pytest.mark.asyncio
-async def test_relevance_score_threshold():
-    """Test relevance scoring meets threshold."""
-    mock_mode = bool(os.environ.get("MOCK_MODE"))
-    result = await default_agent.run({"topic": "python programming"}, mock_mode=mock_mode)
-
-    for video in result.videos:
-        assert video.relevance_score > 0.8
-
-
-@pytest.mark.asyncio
-async def test_find_videos_no_results_graceful():
-    """Test graceful handling of no results."""
-    mock_mode = bool(os.environ.get("MOCK_MODE"))
-    result = await default_agent.run({"topic": "xyznonexistent123"}, mock_mode=mock_mode)
-
-    # Should not crash, return empty or message
-    assert result.videos == [] or result.message
-'''
-)
-```
-
-## Step 6: Run All Tests
-
-Execute all tests:
-
-```python
-result = run_tests(
-    goal_id="youtube-research",
-    agent_path="exports/youtube-research",
-    test_types='["all"]',
-    parallel=4
-)
-```
-
-**Results:**
-
-```json
-{
-    "goal_id": "youtube-research",
-    "overall_passed": false,
-    "summary": {
-        "total": 6,
-        "passed": 5,
-        "failed": 1,
-        "pass_rate": "83.3%"
-    },
-    "duration_ms": 4521,
-    "results": [
-        {"test_id": "test_constraint_api_001", "passed": true, "duration_ms": 1234},
-        {"test_id": "test_constraint_content_001", "passed": true, "duration_ms": 456},
-        {"test_id": "test_success_001", "passed": true, "duration_ms": 789},
-        {"test_id": "test_success_002", "passed": true, "duration_ms": 654},
-        {"test_id": "test_success_003", "passed": true, "duration_ms": 543},
-        {"test_id": "test_success_004", "passed": false, "duration_ms": 845,
-         "error_category": "IMPLEMENTATION_ERROR",
-         "error_message": "TypeError: 'NoneType' object has no attribute 'videos'"}
-    ]
-}
-```
-
-## Step 7: Debug the Failed Test
-
-```python
-result = debug_test(
-    goal_id="youtube-research",
-    test_name="test_find_videos_no_results_graceful",
-    agent_path="exports/youtube-research"
-)
-```
-
-**Debug Output:**
-
-```json
-{
-    "test_id": "test_success_004",
-    "test_name": "test_find_videos_no_results_graceful",
-    "input": {"topic": "xyznonexistent123"},
-    "expected": "Empty list or message",
-    "actual": {"error": "TypeError: 'NoneType' object has no attribute 'videos'"},
-    "passed": false,
-    "error_message": "TypeError: 'NoneType' object has no attribute 'videos'",
-    "error_category": "IMPLEMENTATION_ERROR",
-    "stack_trace": "Traceback (most recent call last):\n  File \"filter_node.py\", line 42\n    for video in result.videos:\nTypeError: 'NoneType' object has no attribute 'videos'",
-    "logs": [
-        {"timestamp": "2026-01-20T10:00:01", "node": "search_node", "level": "INFO", "msg": "Searching for: xyznonexistent123"},
-        {"timestamp": "2026-01-20T10:00:02", "node": "search_node", "level": "WARNING", "msg": "No results found"},
-        {"timestamp": "2026-01-20T10:00:02", "node": "filter_node", "level": "ERROR", "msg": "NoneType error"}
-    ],
-    "runtime_data": {
-        "execution_path": ["start", "search_node", "filter_node"],
-        "node_outputs": {
-            "search_node": null
-        }
-    },
-    "suggested_fix": "Add null check in filter_node before accessing .videos attribute",
-    "iteration_guidance": {
-        "stage": "Agent",
-        "action": "Fix the code in nodes/edges",
-        "restart_required": false,
-        "description": "The goal is correct, but filter_node doesn't handle null results from search_node."
-    }
-}
-```
-
-## Step 8: Iterate Based on Category
-
-Since this is an **IMPLEMENTATION_ERROR**, we:
-
-1. **Don't restart** the Goal → Agent → Eval flow
-2. **Fix the agent** using building-agents skill:
-   - Modify `filter_node` to handle null results
-3. **Re-run Eval** (tests only)
-
-### Fix in building-agents:
-
-```python
-# Update the filter_node to handle null
-add_node(
-    node_id="filter_node",
-    name="Filter Node",
-    description="Filter and rank videos",
-    node_type="function",
-    input_keys=["search_results"],
-    output_keys=["filtered_videos"],
-    system_prompt="""
-    Filter videos by relevance.
-    IMPORTANT: Handle case where search_results is None or empty.
-    Return empty list if no results.
-    """
-)
-```
-
-### Re-export and re-test:
-
-```python
-# Re-export the fixed agent
-export_graph(path="exports/youtube-research")
-
-# Re-run tests
-result = run_tests(
-    goal_id="youtube-research",
-    agent_path="exports/youtube-research",
-    test_types='["all"]'
-)
-```
-
-**Updated Results:**
-
-```json
-{
-    "goal_id": "youtube-research",
-    "overall_passed": true,
-    "summary": {
-        "total": 6,
-        "passed": 6,
-        "failed": 0,
-        "pass_rate": "100.0%"
-    }
-}
-```
-
-## Summary
-
-1. **Got guidelines** for constraint tests during Goal stage
-2. **Wrote** constraint tests using Write tool
-3. **Got guidelines** for success criteria tests during Eval stage
-4. **Wrote** success criteria tests using Write tool
-5. **Ran** tests in parallel
-6. **Debugged** the one failure
-7. **Categorized** as IMPLEMENTATION_ERROR
-8. **Fixed** the agent (not the goal)
-9. **Re-ran** Eval only (didn't restart full flow)
-10. **Passed** all tests
-
-The agent is now validated and ready for production use.

.cursor/mcp.json

@@ -1,20 +0,0 @@
-{
-  "mcpServers": {
-    "agent-builder": {
-      "command": "python",
-      "args": ["-m", "framework.mcp.agent_builder_server"],
-      "cwd": "core",
-      "env": {
-        "PYTHONPATH": "../tools/src"
-      }
-    },
-    "tools": {
-      "command": "python",
-      "args": ["mcp_server.py", "--stdio"],
-      "cwd": "tools",
-      "env": {
-        "PYTHONPATH": "src"
-      }
-    }
-  }
-}

.cursor/skills/agent-workflow

@@ -1 +0,0 @@
-../../.claude/skills/agent-workflow

.cursor/skills/building-agents-construction

@@ -1 +0,0 @@
-../../.claude/skills/building-agents-construction

.cursor/skills/building-agents-core

@@ -1 +0,0 @@
-../../.claude/skills/building-agents-core

.cursor/skills/building-agents-patterns

@@ -1 +0,0 @@
-../../.claude/skills/building-agents-patterns

.cursor/skills/testing-agent

@@ -1 +0,0 @@
-../../.claude/skills/testing-agent

.github/ISSUE_TEMPLATE/bug_report.md

@@ -1,9 +1,10 @@
 ---
 name: Bug Report
 about: Report a bug to help us improve
-title: '[Bug]: '
-labels: bug
+title: "[Bug]: "
+labels: bug, enhancement
 assignees: ''
+
 ---
 
 ## Describe the Bug

.github/ISSUE_TEMPLATE/feature_request.md

@@ -1,9 +1,10 @@
 ---
 name: Feature Request
 about: Suggest a new feature or enhancement
-title: '[Feature]: '
+title: "[Feature]: "
 labels: enhancement
 assignees: ''
+
 ---
 
 ## Problem Statement

.github/ISSUE_TEMPLATE/integration-bounty.yml

@@ -0,0 +1,89 @@
+name: Integration Bounty
+description: A bounty task for the integration contribution program
+title: "[Bounty]: "
+labels: []
+body:
+  - type: markdown
+    attributes:
+      value: |
+        ## Integration Bounty
+
+        This issue is part of the [Integration Bounty Program](../../docs/bounty-program/README.md).
+        **Claim this bounty** by commenting below — a maintainer will assign you within 24 hours.
+
+  - type: dropdown
+    id: bounty-type
+    attributes:
+      label: Bounty Type
+      options:
+        - "Test a Tool (20 pts)"
+        - "Write Docs (20 pts)"
+        - "Code Contribution (30 pts)"
+        - "New Integration (75 pts)"
+    validations:
+      required: true
+
+  - type: dropdown
+    id: difficulty
+    attributes:
+      label: Difficulty
+      options:
+        - Easy
+        - Medium
+        - Hard
+    validations:
+      required: true
+
+  - type: input
+    id: tool-name
+    attributes:
+      label: Tool Name
+      description: The integration this bounty targets (e.g., `airtable`, `salesforce`)
+      placeholder: e.g., airtable
+    validations:
+      required: true
+
+  - type: textarea
+    id: description
+    attributes:
+      label: Description
+      description: What needs to be done to complete this bounty.
+      placeholder: |
+        Describe the specific task, including:
+        - What the contributor needs to do
+        - Links to relevant files in the repo
+        - Any setup requirements (API keys, accounts, etc.)
+    validations:
+      required: true
+
+  - type: textarea
+    id: acceptance-criteria
+    attributes:
+      label: Acceptance Criteria
+      description: What "done" looks like. The PR or report must meet all criteria.
+      placeholder: |
+        - [ ] Criterion 1
+        - [ ] Criterion 2
+        - [ ] CI passes
+    validations:
+      required: true
+
+  - type: textarea
+    id: relevant-files
+    attributes:
+      label: Relevant Files
+      description: Links to tool directory, credential spec, health check file, etc.
+      placeholder: |
+        - Tool: `tools/src/aden_tools/tools/{tool_name}/`
+        - Credential spec: `tools/src/aden_tools/credentials/{category}.py`
+        - Health checks: `tools/src/aden_tools/credentials/health_check.py`
+
+  - type: textarea
+    id: resources
+    attributes:
+      label: Resources
+      description: Links to API docs, examples, or guides that will help the contributor.
+      placeholder: |
+        - [Building Tools Guide](../../tools/BUILDING_TOOLS.md)
+        - [Tool README Template](../../docs/bounty-program/templates/tool-readme-template.md)
+        - API docs: https://...

.github/ISSUE_TEMPLATE/integration-request.md

@@ -0,0 +1,71 @@
+---
+name: Integration Request
+about: Suggest a new integration
+title: "[Integration]:"
+labels: ''
+assignees: ''
+
+---
+
+## Service                                                                                      
+                                                                                                 
+ Name and brief description of the service and what it enables agents to do.                     
+                                                                                                 
+ **Description:** [e.g., "API key for Slack Bot" — short one-liner for the credential spec]      
+                                                                                                 
+ ## Credential Identity                                                                          
+                                                                                                 
+ - **credential_id:** [e.g., `slack`]                                                            
+ - **env_var:** [e.g., `SLACK_BOT_TOKEN`]                                                        
+ - **credential_key:** [e.g., `access_token`, `api_key`, `bot_token`]                            
+                                                                                                 
+ ## Tools                                                                                        
+                                                                                                 
+ Tool function names that require this credential:                                               
+                                                                                                 
+ - [e.g., `slack_send_message`]                                                                  
+ - [e.g., `slack_list_channels`]                                                                 
+                                                                                                 
+ ## Auth Methods                                                                                 
+                                                                                                 
+ - **Direct API key supported:** Yes / No                                                        
+ - **Aden OAuth supported:** Yes / No                                                            
+                                                                                                 
+ If Aden OAuth is supported, describe the OAuth scopes/permissions required.                     
+                                                                                                 
+ ## How to Get the Credential                                                                    
+                                                                                                 
+ Link where users obtain the key/token:                                                          
+                                                                                                 
+ [e.g., https://api.slack.com/apps]                                                              
+                                                                                                 
+ Step-by-step instructions:                                                                      
+                                                                                                 
+ 1. Go to ...                                                                                    
+ 2. Create a ...                                                                                 
+ 3. Select scopes/permissions: ...                                                               
+ 4. Copy the key/token                                                                           
+                                                                                                 
+ ## Health Check                                                                                 
+                                                                                                 
+ A lightweight API call to validate the credential (no writes, no charges).                      
+                                                                                                 
+ - **Endpoint:** [e.g., `https://slack.com/api/auth.test`]                                       
+ - **Method:** [e.g., `GET` or `POST`]                                                           
+ - **Auth header:** [e.g., `Authorization: Bearer {token}` or `X-Api-Key: {key}`]                
+ - **Parameters (if any):** [e.g., `?limit=1`]                                                   
+ - **200 means:** [e.g., key is valid]                                                           
+ - **401 means:** [e.g., invalid or expired]                                                     
+ - **429 means:** [e.g., rate limited but key is valid]                                          
+                                                                                                 
+ ## Credential Group                                                                             
+                                                                                                 
+ Does this require multiple credentials configured together? (e.g., Google Custom Search needs   
+ both an API key and a CSE ID)                                                                   
+                                                                                                 
+ - [ ] No, single credential                                                                     
+ - [ ] Yes — list the other credential IDs in the group:                                         
+                                                                                                 
+ ## Additional Context                                                                           
+                                                                                                 
+ Links to API docs, rate limits, free tier availability, or anything else relevant.

.github/ISSUE_TEMPLATE/standard-bounty.yml

@@ -0,0 +1,78 @@
+name: Standard Bounty
+description: A bounty task for general framework contributions (not integration-specific)
+title: "[Bounty]: "
+labels: []
+body:
+  - type: markdown
+    attributes:
+      value: |
+        ## Standard Bounty
+
+        This issue is part of the [Bounty Program](../../docs/bounty-program/README.md).
+        **Claim this bounty** by commenting below — a maintainer will assign you within 24 hours.
+
+  - type: dropdown
+    id: bounty-size
+    attributes:
+      label: Bounty Size
+      options:
+        - "Small (10 pts)"
+        - "Medium (30 pts)"
+        - "Large (75 pts)"
+        - "Extreme (150 pts)"
+    validations:
+      required: true
+
+  - type: dropdown
+    id: difficulty
+    attributes:
+      label: Difficulty
+      options:
+        - Easy
+        - Medium
+        - Hard
+    validations:
+      required: true
+
+  - type: textarea
+    id: description
+    attributes:
+      label: Description
+      description: What needs to be done to complete this bounty.
+      placeholder: |
+        Describe the specific task, including:
+        - What the contributor needs to do
+        - Links to relevant files in the repo
+        - Any context or motivation for the change
+    validations:
+      required: true
+
+  - type: textarea
+    id: acceptance-criteria
+    attributes:
+      label: Acceptance Criteria
+      description: What "done" looks like. The PR must meet all criteria.
+      placeholder: |
+        - [ ] Criterion 1
+        - [ ] Criterion 2
+        - [ ] CI passes
+    validations:
+      required: true
+
+  - type: textarea
+    id: relevant-files
+    attributes:
+      label: Relevant Files
+      description: Links to files or directories related to this bounty.
+      placeholder: |
+        - `path/to/file.py`
+        - `path/to/directory/`
+
+  - type: textarea
+    id: resources
+    attributes:
+      label: Resources
+      description: Links to docs, issues, or external references that will help.
+      placeholder: |
+        - Related issue: #XXXX
+        - Docs: https://...

.github/workflows/bounty-completed.yml

@@ -0,0 +1,47 @@
+name: Bounty completed
+description: Awards points and notifies Discord when a bounty PR is merged
+
+on:
+  pull_request_target:
+    types: [closed]
+
+  workflow_dispatch:
+    inputs:
+      pr_number:
+        description: "PR number to process (for missed bounties)"
+        required: true
+        type: number
+
+jobs:
+  bounty-notify:
+    if: >
+      github.event_name == 'workflow_dispatch' ||
+      (github.event.pull_request.merged == true &&
+       contains(join(github.event.pull_request.labels.*.name, ','), 'bounty:'))
+    runs-on: ubuntu-latest
+    timeout-minutes: 5
+    permissions:
+      contents: read
+      pull-requests: read
+
+    steps:
+      - name: Checkout repository
+        uses: actions/checkout@v4
+
+      - name: Setup Bun
+        uses: oven-sh/setup-bun@v2
+        with:
+          bun-version: latest
+
+      - name: Award XP and notify Discord
+        run: bun run scripts/bounty-tracker.ts notify
+        env:
+          GITHUB_TOKEN: ${{ secrets.GITHUB_TOKEN }}
+          GITHUB_REPOSITORY_OWNER: ${{ github.repository_owner }}
+          GITHUB_REPOSITORY_NAME: ${{ github.event.repository.name }}
+          DISCORD_WEBHOOK_URL: ${{ secrets.DISCORD_BOUNTY_WEBHOOK_URL }}
+          BOT_API_URL: ${{ secrets.BOT_API_URL }}
+          BOT_API_KEY: ${{ secrets.BOT_API_KEY }}
+          LURKR_API_KEY: ${{ secrets.LURKR_API_KEY }}
+          LURKR_GUILD_ID: ${{ secrets.LURKR_GUILD_ID }}
+          PR_NUMBER: ${{ inputs.pr_number || github.event.pull_request.number }}

.github/workflows/ci.yml

@@ -5,7 +5,7 @@ on:
     branches: [main]
   pull_request:
     branches: [main]
-
+    
 concurrency:
   group: ${{ github.workflow }}-${{ github.ref }}
   cancel-in-progress: true
@@ -21,23 +21,24 @@ jobs:
         uses: actions/setup-python@v5
         with:
           python-version: '3.11'
-          cache: 'pip'
+
+      - name: Install uv
+        uses: astral-sh/setup-uv@v4
+        with:
+          enable-cache: true
 
       - name: Install dependencies
-        run: |
-          cd core
-          pip install -e .
-          pip install -r requirements-dev.txt
+        run: uv sync --project core --group dev
 
       - name: Ruff lint
         run: |
-          ruff check core/
-          ruff check tools/
+          uv run --project core ruff check core/
+          uv run --project core ruff check tools/
 
       - name: Ruff format
         run: |
-          ruff format --check core/
-          ruff format --check tools/
+          uv run --project core ruff format --check core/
+          uv run --project core ruff format --check tools/
 
   test:
     name: Test Python Framework
@@ -52,23 +53,24 @@ jobs:
         uses: actions/setup-python@v5
         with:
           python-version: '3.11'
-          cache: 'pip'
 
-      - name: Install dependencies
+      - name: Install uv
+        uses: astral-sh/setup-uv@v4
+        with:
+          enable-cache: true
+
+      - name: Install dependencies and run tests
+        working-directory: core
         run: |
-          cd core
-          pip install -e .
-          pip install -r requirements-dev.txt
+          uv sync
+          uv run pytest tests/ -v
 
-      - name: Run tests
-        run: |
-          cd core
-          pytest tests/ -v
-
-  validate:
-    name: Validate Agent Exports
-    runs-on: ubuntu-latest
-    needs: [lint, test]
+  test-tools:
+    name: Test Tools (${{ matrix.os }})
+    runs-on: ${{ matrix.os }}
+    strategy:
+      matrix:
+        os: [ubuntu-latest, windows-latest]
     steps:
       - uses: actions/checkout@v4
 
@@ -76,13 +78,39 @@ jobs:
         uses: actions/setup-python@v5
         with:
           python-version: '3.11'
-          cache: 'pip'
 
-      - name: Install dependencies
+      - name: Install uv
+        uses: astral-sh/setup-uv@v4
+        with:
+          enable-cache: true
+
+      - name: Install dependencies and run tests
+        working-directory: tools
         run: |
-          cd core
-          pip install -e .
-          pip install -r requirements-dev.txt
+          uv sync --extra dev
+          uv run pytest tests/ -v
+
+  validate:
+    name: Validate Agent Exports
+    runs-on: ubuntu-latest
+    needs: [lint, test, test-tools]
+    steps:
+      - uses: actions/checkout@v4
+
+      - name: Setup Python
+        uses: actions/setup-python@v5
+        with:
+          python-version: '3.11'
+
+      - name: Install uv
+        uses: astral-sh/setup-uv@v4
+        with:
+          enable-cache: true
+            
+      - name: Install dependencies
+        working-directory: core
+        run: |
+          uv sync
 
       - name: Validate exported agents
         run: |
@@ -105,7 +133,7 @@ jobs:
           for agent_dir in "${agent_dirs[@]}"; do
             if [ -f "$agent_dir/agent.json" ]; then
               echo "Validating $agent_dir"
-              python -c "import json; json.load(open('$agent_dir/agent.json'))"
+              uv run python -c "import json; json.load(open('$agent_dir/agent.json'))"
               validated=$((validated + 1))
             fi
           done

.github/workflows/claude-issue-triage.yml

@@ -80,7 +80,13 @@ jobs:
             - help wanted: Extra attention is needed (if issue needs community input)
             - backlog: Tracked for the future, but not currently planned or prioritized
 
-            You may apply multiple labels if appropriate (e.g., "bug" and "help wanted").
+            ### 6. Estimate size (if NOT a duplicate, spam, or invalid)
+            Apply exactly ONE size label to help contributors match their capacity to the task:
+            - "size: small": Docs, typos, single-file fixes, config changes
+            - "size: medium": Bug fixes with tests, adding a single tool, changes within one package
+            - "size: large": Cross-package changes (core + tools), new modules, complex logic, architectural refactors
+
+            You may apply multiple labels if appropriate (e.g., "bug", "size: small", and "good first issue").
 
             ## Tools Available:
             - mcp__github__get_issue: Get issue details

.github/workflows/pr-requirements-enforce.yml

@@ -0,0 +1,54 @@
+# Closes PRs that still have the `pr-requirements-warning` label
+# after contributors were warned in pr-requirements.yml.
+name: PR Requirements Enforcement
+on:
+  schedule:
+    - cron: "0 0 * * *"   # runs every day once at midnight 
+jobs:
+  enforce:
+    name: Close PRs still failing contribution requirements
+    runs-on: ubuntu-latest
+    permissions:
+      pull-requests: write
+      issues: write
+    steps:
+      - name: Close PRs still failing requirements
+        uses: actions/github-script@v7
+        with:
+          script: |
+            const { owner, repo } = context.repo;
+            const prs = await github.paginate(github.rest.pulls.list, {
+              owner,
+              repo,
+              state: "open",
+              per_page: 100
+            });
+            for (const pr of prs) {
+              // Skip draft PRs — author may still be actively working toward compliance
+              if (pr.draft) continue;
+              const labels = pr.labels.map(l => l.name);
+              if (!labels.includes("pr-requirements-warning")) continue;
+              const gracePeriod = 24 * 60 * 60 * 1000;
+              const lastUpdated = new Date(pr.created_at);
+              const now = new Date();
+              if (now - lastUpdated < gracePeriod) {
+                console.log(`Skipping PR #${pr.number} — still within grace period`);
+                continue;
+              }
+              const prNumber = pr.number;
+              const prAuthor = pr.user.login;
+              await github.rest.issues.createComment({
+                owner,
+                repo,
+                issue_number: prNumber,
+                body: `Closing PR because the contribution requirements were not resolved within the 24-hour grace period.
+                If this was closed in error, feel free to reopen the PR after fixing the requirements.`
+              });
+              await github.rest.pulls.update({
+                owner,
+                repo,
+                pull_number: prNumber,
+                state: "closed"
+              });
+              console.log(`Closed PR #${prNumber} by ${prAuthor} (PR requirements were not met)`);
+            }

.github/workflows/pr-requirements.yml

@@ -43,9 +43,10 @@ jobs:
             console.log(`  Found issue references: ${issueNumbers.length > 0 ? issueNumbers.join(', ') : 'none'}`);
 
             if (issueNumbers.length === 0) {
-              const message = `## PR Closed - Requirements Not Met
+              const message = `## PR Requirements Warning
 
-            This PR has been automatically closed because it doesn't meet the requirements.
+            This PR does not meet the contribution requirements.
+            If the issue is not fixed within ~24 hours, it may be automatically closed.
 
             **Missing:** No linked issue found.
 
@@ -67,14 +68,15 @@ jobs:
 
             **Why is this required?** See #472 for details.`;
 
-              const comments = await github.rest.issues.listComments({
+              const comments = await github.paginate(github.rest.issues.listComments, {
                 owner: context.repo.owner,
                 repo: context.repo.repo,
                 issue_number: prNumber,
+                per_page: 100,
               });
 
-              const botComment = comments.data.find(
-                (c) => c.user.type === 'Bot' && c.body.includes('PR Closed - Requirements Not Met')
+              const botComment = comments.find(
+                (c) => c.user.type === 'Bot' && c.body.includes('PR Requirements Warning')
               );
 
               if (!botComment) {
@@ -86,11 +88,11 @@ jobs:
                 });
               }
 
-              await github.rest.pulls.update({
+              await github.rest.issues.addLabels({
                 owner: context.repo.owner,
                 repo: context.repo.repo,
-                pull_number: prNumber,
-                state: 'closed',
+                issue_number: prNumber,
+                labels: ['pr-requirements-warning'],
               });
 
               core.setFailed('PR must reference an issue');
@@ -132,9 +134,10 @@ jobs:
                 `#${i.number} (assignees: ${i.assignees.length > 0 ? i.assignees.join(', ') : 'none'})`
               ).join(', ');
 
-              const message = `## PR Closed - Requirements Not Met
+              const message = `## PR Requirements Warning
 
-            This PR has been automatically closed because it doesn't meet the requirements.
+            This PR does not meet the contribution requirements.
+            If the issue is not fixed within ~24 hours, it may be automatically closed.
 
             **PR Author:** @${prAuthor}
             **Found issues:** ${issueList}
@@ -157,14 +160,15 @@ jobs:
 
             **Why is this required?** See #472 for details.`;
 
-              const comments = await github.rest.issues.listComments({
+              const comments = await github.paginate(github.rest.issues.listComments, {
                 owner: context.repo.owner,
                 repo: context.repo.repo,
                 issue_number: prNumber,
+                per_page: 100,
               });
 
-              const botComment = comments.data.find(
-                (c) => c.user.type === 'Bot' && c.body.includes('PR Closed - Requirements Not Met')
+              const botComment = comments.find(
+                (c) => c.user.type === 'Bot' && c.body.includes('PR Requirements Warning')
               );
 
               if (!botComment) {
@@ -176,14 +180,24 @@ jobs:
                 });
               }
 
-              await github.rest.pulls.update({
+              await github.rest.issues.addLabels({
                 owner: context.repo.owner,
                 repo: context.repo.repo,
-                pull_number: prNumber,
-                state: 'closed',
+                issue_number: prNumber,
+                labels: ['pr-requirements-warning'],
               });
 
               core.setFailed('PR author must be assigned to the linked issue');
             } else {
               console.log(`PR requirements met! Issue #${issueWithAuthorAssigned} has ${prAuthor} as assignee.`);
-            }
+              try {
+                await github.rest.issues.removeLabel({
+                  owner: context.repo.owner,
+                  repo: context.repo.repo,
+                  issue_number: prNumber,
+                  name: "pr-requirements-warning"
+                });
+              }catch (error){
+                //ignore if label doesn't exist
+              }
+            }

.github/workflows/release.yml

@@ -21,18 +21,19 @@ jobs:
         uses: actions/setup-python@v5
         with:
           python-version: '3.11'
-          cache: 'pip'
+
+      - name: Install uv
+        uses: astral-sh/setup-uv@v4
 
       - name: Install dependencies
         run: |
           cd core
-          pip install -e .
-          pip install -r requirements-dev.txt
+          uv sync
 
       - name: Run tests
         run: |
           cd core
-          pytest tests/ -v
+          uv run pytest tests/ -v
 
       - name: Generate changelog
         id: changelog

.github/workflows/weekly-leaderboard.yml

@@ -0,0 +1,42 @@
+name: Weekly bounty leaderboard
+description: Posts the integration bounty leaderboard to Discord every Monday
+
+on:
+  schedule:
+    # Every Monday at 9:00 UTC
+    - cron: "0 9 * * 1"
+  workflow_dispatch:
+    inputs:
+      since_date:
+        description: "Only count PRs merged after this date (YYYY-MM-DD). Leave empty for all-time."
+        required: false
+
+jobs:
+  leaderboard:
+    runs-on: ubuntu-latest
+    timeout-minutes: 5
+    permissions:
+      contents: read
+      pull-requests: read
+
+    steps:
+      - name: Checkout repository
+        uses: actions/checkout@v4
+
+      - name: Setup Bun
+        uses: oven-sh/setup-bun@v2
+        with:
+          bun-version: latest
+
+      - name: Post leaderboard to Discord
+        run: bun run scripts/bounty-tracker.ts leaderboard
+        env:
+          GITHUB_TOKEN: ${{ secrets.GITHUB_TOKEN }}
+          GITHUB_REPOSITORY_OWNER: ${{ github.repository_owner }}
+          GITHUB_REPOSITORY_NAME: ${{ github.event.repository.name }}
+          DISCORD_WEBHOOK_URL: ${{ secrets.DISCORD_BOUNTY_WEBHOOK_URL }}
+          BOT_API_URL: ${{ secrets.BOT_API_URL }}
+          BOT_API_KEY: ${{ secrets.BOT_API_KEY }}
+          LURKR_API_KEY: ${{ secrets.LURKR_API_KEY }}
+          LURKR_GUILD_ID: ${{ secrets.LURKR_GUILD_ID }}
+          SINCE_DATE: ${{ github.event.inputs.since_date || '' }}

.gitignore

@@ -13,6 +13,10 @@ out/
 .env
 .env.local
 .env.*.local
+.venv
+/venv
+tools/src/uv.lock
+
 
 # User configuration (copied from .example)
 config.yaml
@@ -46,6 +50,7 @@ coverage/
 
 # TypeScript
 *.tsbuildinfo
+vite.config.d.ts
 
 # Python
 __pycache__/
@@ -54,7 +59,6 @@ __pycache__/
 *.egg-info/
 .eggs/
 *.egg
-uv.lock
 
 # Generated runtime data
 core/data/
@@ -67,9 +71,11 @@ temp/
 
 exports/*
 
-.agent-builder-sessions/*
-
-.venv
+.claude/settings.local.json
 
 docs/github-issues/*
 core/tests/*dumps/*
+
+screenshots/*
+
+.gemini/*

.mcp.json

@@ -1,20 +1,3 @@
 {
-  "mcpServers": {
-    "agent-builder": {
-      "command": ".venv/bin/python",
-      "args": ["-m", "framework.mcp.agent_builder_server"],
-      "cwd": "core",
-      "env": {
-        "PYTHONPATH": "../tools/src"
-      }
-    },
-    "tools": {
-      "command": ".venv/bin/python",
-      "args": ["mcp_server.py", "--stdio"],
-      "cwd": "tools",
-      "env": {
-        "PYTHONPATH": "src:../core"
-      }
-    }
-  }
+  "mcpServers": {}
 }

.pre-commit-config.yaml

@@ -1,6 +1,6 @@
 repos:
   - repo: https://github.com/astral-sh/ruff-pre-commit
-    rev: v0.8.6
+    rev: v0.15.0
     hooks:
       - id: ruff
         name: ruff lint (core)

.vscode/extensions.json

@@ -1,7 +0,0 @@
-{
-  "recommendations": [
-    "charliermarsh.ruff",
-    "editorconfig.editorconfig",
-    "ms-python.python"
-  ]
-}

AGENTS.md

@@ -0,0 +1,30 @@
+# Repository Guidelines
+
+Shared agent instructions for this workspace.
+
+## Coding Agent Notes
+
+- 
+- When working on a GitHub Issue or PR, print the full URL at the end of the task.
+- When answering questions, respond with high-confidence answers only: verify in code; do not guess.
+- Do not update dependencies casually. Version bumps, patched dependencies, overrides, or vendored dependency changes require explicit approval.
+- Add brief comments for tricky logic. Keep files reasonably small when practical; split or refactor large files instead of growing them indefinitely.
+- If shared guardrails are available locally, review them; otherwise follow this repo's guidance.
+- Use `uv` for Python execution and package management. Do not use `python` or `python3` directly unless the user explicitly asks for it.
+- Prefer `uv run` for scripts and tests, and `uv pip` for package operations.
+
+
+## Multi-Agent Safety
+
+- Do not create, apply, or drop `git stash` entries unless explicitly requested.
+- Do not create, remove, or modify `git worktree` checkouts unless explicitly requested.
+- Do not switch branches or check out a different branch unless explicitly requested.
+- When the user says `push`, you may `git pull --rebase` to integrate latest changes, but never discard other in-progress work.
+- When the user says `commit`, commit only your changes. When the user says `commit all`, commit everything in grouped chunks.
+- When you see unrecognized files or unrelated changes, keep going and focus on your scoped changes.
+
+## Change Hygiene
+
+- If staged and unstaged diffs are formatting-only, resolve them without asking.
+- If a commit or push was already requested, include formatting-only follow-up changes in that same commit when practical.
+- Only stop to ask for confirmation when changes are semantic and may alter behavior.

CHANGELOG.md

@@ -1,41 +1,330 @@
-# Changelog
+# Release Notes
 
-All notable changes to this project will be documented in this file.
+## v0.7.1
 
-The format is based on [Keep a Changelog](https://keepachangelog.com/en/1.1.0/),
-and this project adheres to [Semantic Versioning](https://semver.org/spec/v2.0.0.html).
+**Release Date:** March 13, 2026
+**Tag:** v0.7.1
 
-## [Unreleased]
+### Chrome-Native Browser Control
 
-### Added
-- Initial project structure
-- React frontend (honeycomb) with Vite and TypeScript
-- Node.js backend (hive) with Express and TypeScript
-- Docker Compose configuration for local development
-- Configuration system via `config.yaml`
-- GitHub Actions CI/CD workflows
-- Comprehensive documentation
+v0.7.1 replaces Playwright with direct Chrome DevTools Protocol (CDP) integration. The GCU now launches the user's system Chrome via `open -n` on macOS, connects over CDP, and manages browser lifecycle end-to-end -- no extra browser binary required.
 
-### Changed
-- N/A
+---
 
-### Deprecated
-- N/A
+### Highlights
 
-### Removed
-- N/A
+#### System Chrome via CDP
 
+The entire GCU browser stack has been rewritten:
 
-### Fixed
-- tools: Fixed web_scrape tool attempting to parse non-HTML content (PDF, JSON) as HTML (#487)
+- **Chrome finder & launcher** -- New `chrome_finder.py` discovers installed Chrome and `chrome_launcher.py` manages process lifecycle with `--remote-debugging-port`
+- **Coexist with user's browser** -- `open -n` on macOS launches a separate Chrome instance so the user's tabs stay untouched
+- **Dynamic viewport sizing** -- Viewport auto-sizes to the available display area, suppressing Chrome warning bars
+- **Orphan cleanup** -- Chrome processes are killed on GCU server shutdown to prevent leaks
+- **`--no-startup-window`** -- Chrome launches headlessly by default until a page is needed
 
-### Security
-- N/A
+#### Per-Subagent Browser Isolation
 
-## [0.1.0] - 2025-01-13
+Each GCU subagent gets its own Chrome user-data directory, preventing cookie/session cross-contamination:
 
-### Added
-- Initial release
+- Unique browser profiles injected per subagent
+- Profiles cleaned up after top-level GCU node execution
+- Tab origin and age metadata tracked per subagent
 
-[Unreleased]: https://github.com/adenhq/hive/compare/v0.1.0...HEAD
-[0.1.0]: https://github.com/adenhq/hive/releases/tag/v0.1.0
+#### Dummy Agent Testing Framework
+
+A comprehensive test suite for validating agent graph patterns without LLM calls:
+
+- 8 test modules covering echo, pipeline, branch, parallel merge, retry, feedback loop, worker, and GCU subagent patterns
+- Shared fixtures and a `run_all.py` runner for CI integration
+- Subagent lifecycle tests
+
+---
+
+### What's New
+
+#### GCU Browser
+
+- **Switch from Playwright to system Chrome via CDP** -- Direct CDP connection replaces Playwright dependency. (@bryanadenhq)
+- **Chrome finder and launcher modules** -- `chrome_finder.py` and `chrome_launcher.py` for cross-platform Chrome discovery and process management. (@bryanadenhq)
+- **Dynamic viewport sizing** -- Auto-size viewport and suppress Chrome warning bar. (@bryanadenhq)
+- **Per-subagent browser profile isolation** -- Unique user-data directories per subagent with cleanup. (@bryanadenhq)
+- **Tab origin/age metadata** -- Track which subagent opened each tab and when. (@bryanadenhq)
+- **`browser_close_all` tool** -- Bulk tab cleanup for agents managing many pages. (@bryanadenhq)
+- **Auto-track popup pages** -- Popups are automatically captured and tracked. (@bryanadenhq)
+- **Auto-snapshot from browser interactions** -- Browser interaction tools return screenshots automatically. (@bryanadenhq)
+- **Kill orphaned Chrome processes** -- GCU server shutdown cleans up lingering Chrome instances. (@bryanadenhq)
+- **`--no-startup-window` Chrome flag** -- Prevent empty window on launch. (@bryanadenhq)
+- **Launch Chrome via `open -n` on macOS** -- Coexist with the user's running browser. (@bryanadenhq)
+
+#### Framework & Runtime
+
+- **Session resume fix for new agents** -- Correctly resume sessions when a new agent is loaded. (@bryanadenhq)
+- **Queen upsert fix** -- Prevent duplicate queen entries on session restore. (@bryanadenhq)
+- **Anchor worker monitoring to queen's session ID on cold-restore** -- Worker monitors reconnect to the correct queen after restart. (@bryanadenhq)
+- **Update meta.json when loading workers** -- Worker metadata stays in sync with runtime state. (@RichardTang-Aden)
+- **Generate worker MCP file correctly** -- Fix MCP config generation for spawned workers. (@RichardTang-Aden)
+- **Share event bus so tool events are visible to parent** -- Tool execution events propagate up to parent graphs. (@bryanadenhq)
+- **Subagent activity tracking in queen status** -- Queen instructions include live subagent status. (@bryanadenhq)
+- **GCU system prompt updates** -- Auto-snapshots, batching, popup tracking, and close_all guidance. (@bryanadenhq)
+
+#### Frontend
+
+- **Loading spinner in draft panel** -- Shows spinner during planning phase instead of blank panel. (@bryanadenhq)
+- **Fix credential modal errors** -- Modal no longer eats errors; banner stays visible. (@bryanadenhq)
+- **Fix credentials_required loop** -- Stop clearing the flag on modal close to prevent infinite re-prompting. (@bryanadenhq)
+- **Fix "Add tab" dropdown overflow** -- Dropdown no longer hidden when many agents are open. (@prasoonmhwr)
+
+#### Testing
+
+- **Dummy agent test framework** -- 8 test modules (echo, pipeline, branch, parallel merge, retry, feedback loop, worker, GCU subagent) with shared fixtures and CI runner. (@bryanadenhq)
+- **Subagent lifecycle tests** -- Validate subagent spawn and completion flows. (@bryanadenhq)
+
+#### Documentation & Infrastructure
+
+- **MCP integration PRD** -- Product requirements for MCP server registry. (@TimothyZhang7)
+- **Skills registry PRD** -- Product requirements for skill registry system. (@bryanadenhq)
+- **Bounty program updates** -- Standard bounty issue template and updated contributor guide. (@bryanadenhq)
+- **Windows quickstart** -- Add default context limit for PowerShell setup. (@bryanadenhq)
+- **Remove deprecated files** -- Clean up `setup_mcp.py`, `verify_mcp.py`, `antigravity-setup.md`, and `setup-antigravity-mcp.sh`. (@bryanadenhq)
+
+---
+
+### Bug Fixes
+
+- Fix credential modal eating errors and banner staying open
+- Stop clearing `credentials_required` on modal close to prevent infinite loop
+- Share event bus so tool events are visible to parent graph
+- Use lazy %-formatting in subagent completion log to avoid f-string in logger
+- Anchor worker monitoring to queen's session ID on cold-restore
+- Update meta.json when loading workers
+- Generate worker MCP file correctly
+- Fix "Add tab" dropdown partially hidden when creating multiple agents
+
+---
+
+### Community Contributors
+
+- **Prasoon Mahawar** (@prasoonmhwr) -- Fix UI overflow on agent tab dropdown
+- **Richard Tang** (@RichardTang-Aden) -- Worker MCP generation and meta.json fixes
+
+---
+
+### Upgrading
+
+```bash
+git pull origin main
+uv sync
+```
+
+The Playwright dependency is no longer required for GCU browser operations. Chrome must be installed on the host system.
+
+---
+
+## v0.7.0
+
+**Release Date:** March 5, 2026
+**Tag:** v0.7.0
+
+Session management refactor release.
+
+---
+
+## v0.5.1
+
+**Release Date:** February 18, 2026
+**Tag:** v0.5.1
+
+### The Hive Gets a Brain
+
+v0.5.1 is our most ambitious release yet. Hive agents can now **build other agents** -- the new Hive Coder meta-agent writes, tests, and fixes agent packages from natural language. The runtime grows multi-graph support so one session can orchestrate multiple agents simultaneously. The TUI gets a complete overhaul with an in-app agent picker, live streaming, and seamless escalation to the Coder. And we're now provider-agnostic: Claude Code subscriptions, OpenAI-compatible endpoints, and any LiteLLM-supported model work out of the box.
+
+---
+
+### Highlights
+
+#### Hive Coder -- The Agent That Builds Agents
+
+A native meta-agent that lives inside the framework at `core/framework/agents/hive_coder/`. Give it a natural-language specification and it produces a complete agent package -- goal definition, node prompts, edge routing, MCP tool wiring, tests, and all boilerplate files.
+
+```bash
+# Launch the Coder directly
+hive code
+
+# Or escalate from any running agent (TUI)
+Ctrl+E  # or /coder in chat
+```
+
+The Coder ships with:
+
+- **Reference documentation** -- anti-patterns, construction guide, and design patterns baked into its system prompt
+- **Guardian watchdog** -- an event-driven monitor that catches agent failures and triggers automatic remediation
+- **Coder Tools MCP server** -- file I/O, fuzzy-match editing, git snapshots, and sandboxed shell execution (`tools/coder_tools_server.py`)
+- **Test generation** -- structural tests for forever-alive agents that don't hang on `runner.run()`
+
+#### Multi-Graph Agent Runtime
+
+`AgentRuntime` now supports loading, managing, and switching between multiple agent graphs within a single session. Six new lifecycle tools give agents (and the TUI) full control:
+
+```python
+# Load a second agent into the runtime
+await runtime.add_graph("exports/deep_research_agent")
+
+# Tools available to agents:
+# load_agent, unload_agent, start_agent, restart_agent, list_agents, get_user_presence
+```
+
+The Hive Coder uses multi-graph internally -- when you escalate from a worker agent, the Coder loads as a separate graph while the worker stays alive in the background.
+
+#### TUI Revamp
+
+The Terminal UI gets a ground-up rebuild with five major additions:
+
+- **Agent Picker** (Ctrl+A) -- tabbed modal screen for browsing Your Agents, Framework agents, and Examples with metadata badges (node count, tool count, session count, tags)
+- **Runtime-optional startup** -- TUI launches without a pre-loaded agent, showing the picker on first open
+- **Live streaming pane** -- dedicated RichLog widget shows LLM tokens as they arrive, replacing the old one-token-per-line display
+- **PDF attachments** -- `/attach` and `/detach` commands with native OS file dialog (macOS, Linux, Windows)
+- **Multi-graph commands** -- `/graphs`, `/graph <id>`, `/load <path>`, `/unload <id>` for managing agent graphs in-session
+
+#### Provider-Agnostic LLM Support
+
+Hive is no longer Anthropic-only. v0.5.1 adds first-class support for:
+
+- **Claude Code subscriptions** -- `use_claude_code_subscription: true` in `~/.hive/configuration.json` reads OAuth tokens from `~/.claude/.credentials.json` with automatic refresh
+- **OpenAI-compatible endpoints** -- `api_base` config routes traffic through any compatible API (Azure OpenAI, vLLM, Ollama, etc.)
+- **Any LiteLLM model** -- `RuntimeConfig` now passes `api_key`, `api_base`, and `extra_kwargs` through to LiteLLM
+
+The quickstart script auto-detects Claude Code subscriptions and ZAI Code installations.
+
+---
+
+### What's New
+
+#### Architecture & Runtime
+
+- **Hive Coder meta-agent** -- Natural-language agent builder with reference docs, guardian watchdog, and `hive code` CLI command. (@TimothyZhang7)
+- **Multi-graph agent sessions** -- `add_graph`/`remove_graph` on AgentRuntime with 6 lifecycle tools (`load_agent`, `unload_agent`, `start_agent`, `restart_agent`, `list_agents`, `get_user_presence`). (@TimothyZhang7)
+- **Claude Code subscription support** -- OAuth token refresh via `use_claude_code_subscription` config, auto-detection in quickstart, LiteLLM header patching. (@TimothyZhang7)
+- **OpenAI-compatible endpoint support** -- `api_base` and `extra_kwargs` in `RuntimeConfig` for any OpenAI-compatible API. (@TimothyZhang7)
+- **Remove deprecated node types** -- Delete `FlexibleGraphExecutor`, `WorkerNode`, `HybridJudge`, `CodeSandbox`, `Plan`, `FunctionNode`, `LLMNode`, `RouterNode`. Deprecated types (`llm_tool_use`, `llm_generate`, `function`, `router`, `human_input`) now raise `RuntimeError` with migration guidance. (@TimothyZhang7)
+- **Interactive credential setup** -- Guided `CredentialSetupSession` with health checks and encrypted storage, accessible via `hive setup-credentials` or automatic prompting on credential errors. (@RichardTang-Aden)
+- **Pre-start confirmation prompt** -- Interactive prompt before agent execution allowing credential updates or abort. (@RichardTang-Aden)
+- **Event bus multi-graph support** -- `graph_id` on events, `filter_graph` on subscriptions, `ESCALATION_REQUESTED` event type, `exclude_own_graph` filter. (@TimothyZhang7)
+
+#### TUI Improvements
+
+- **In-app agent picker** (Ctrl+A) -- Tabbed modal for browsing agents with metadata badges (nodes, tools, sessions, tags). (@TimothyZhang7)
+- **Runtime-optional TUI startup** -- Launches without a pre-loaded agent, shows agent picker on startup. (@TimothyZhang7)
+- **Hive Coder escalation** (Ctrl+E) -- Escalate to Hive Coder and return; also available via `/coder` and `/back` chat commands. (@TimothyZhang7)
+- **PDF attachment support** -- `/attach` and `/detach` commands with native OS file dialog. (@TimothyZhang7)
+- **Streaming output pane** -- Dedicated RichLog widget for live LLM token streaming. (@TimothyZhang7)
+- **Multi-graph TUI commands** -- `/graphs`, `/graph <id>`, `/load <path>`, `/unload <id>`. (@TimothyZhang7)
+- **Agent Guardian watchdog** -- Event-driven monitor that catches secondary agent failures and triggers automatic remediation, with `--no-guardian` CLI flag. (@TimothyZhang7)
+
+#### New Tool Integrations
+
+| Tool                   | Description                                                                                                                                                            | Contributor        |
+| ---------------------- | ---------------------------------------------------------------------------------------------------------------------------------------------------------------------- | ------------------ |
+| **Discord**            | 4 MCP tools (`discord_list_guilds`, `discord_list_channels`, `discord_send_message`, `discord_get_messages`) with rate-limit retry and channel filtering               | @mishrapravin114   |
+| **Exa Search API**     | 4 AI-powered search tools (`exa_search`, `exa_find_similar`, `exa_get_contents`, `exa_answer`) with neural/keyword search, domain filters, and citation-backed answers | @JeetKaria06       |
+| **Razorpay**           | 6 payment processing tools for payments, invoices, payment links, and refunds with HTTP Basic Auth                                                                     | @shivamshahi07     |
+| **Google Docs**        | Document creation, reading, and editing with OAuth credential support                                                                                                  | @haliaeetusvocifer |
+| **Gmail enhancements** | Expanded mail operations for inbox management                                                                                                                          | @bryanadenhq       |
+
+#### Infrastructure
+
+- **Default node type → `event_loop`** -- `NodeSpec.node_type` defaults to `"event_loop"` instead of `"llm_tool_use"`. (@TimothyZhang7)
+- **Default `max_node_visits` → 0 (unlimited)** -- Nodes default to unlimited visits, reducing friction for feedback loops and forever-alive agents. (@TimothyZhang7)
+- **Remove `function` field from NodeSpec** -- Follows deprecation of `FunctionNode`. (@TimothyZhang7)
+- **LiteLLM OAuth patch** -- Correct header construction for OAuth tokens (remove `x-api-key` when Bearer token is present). (@TimothyZhang7)
+- **Orchestrator config centralization** -- Reads `api_key`, `api_base`, `extra_kwargs` from centralized `~/.hive/configuration.json`. (@TimothyZhang7)
+- **System prompt datetime injection** -- All system prompts now include current date/time for time-aware agent behavior. (@TimothyZhang7)
+- **Utils module exports** -- Proper `__init__.py` exports for the utils module. (@Siddharth2624)
+- **Increased default max_tokens** -- Opus 4.6 defaults to 32768, Sonnet 4.5 to 16384 (up from 8192). (@TimothyZhang7)
+
+---
+
+### Bug Fixes
+
+- Flush WIP accumulator outputs on cancel/failure so edge conditions see correct values on resume
+- Stall detection state preserved across resume (no more resets on checkpoint restore)
+- Skip client-facing blocking for event-triggered executions (timer/webhook)
+- Executor retry override scoped to actual EventLoopNode instances only
+- Add `_awaiting_input` flag to EventLoopNode to prevent input injection race conditions
+- Fix TUI streaming display (tokens no longer appear one-per-line)
+- Fix `_return_from_escalation` crash when ChatRepl widgets not yet mounted
+- Fix tools registration problems for Google Docs credentials (@RichardTang-Aden)
+- Fix email agent version conflicts (@RichardTang-Aden)
+- Fix coder tool timeouts (120s for tests, 300s cap for commands)
+
+### Documentation
+
+- Clarify installation and prevent root pip install misuse (@paarths-collab)
+
+---
+
+### Agent Updates
+
+- **Email Inbox Management** -- Consolidate `gmail_inbox_guardian` and `inbox_management` into a single unified agent with updated prompts and config. (@RichardTang-Aden, @bryanadenhq)
+- **Job Hunter** -- Updated node prompts, config, and agent metadata; added PDF resume selection. (@bryanadenhq)
+- **Deep Research Agent** -- Revised node implementations with updated prompts and output handling.
+- **Tech News Reporter** -- Revised node prompts for improved output quality.
+- **Vulnerability Assessment** -- Expanded prompts with more detailed assessment instructions. (@bryanadenhq)
+
+---
+
+### Breaking Changes
+
+- **Deprecated node types raise `RuntimeError`** -- `llm_tool_use`, `llm_generate`, `function`, `router`, `human_input` now fail instead of warning. Migrate to `event_loop`.
+- **`NodeSpec.node_type` defaults to `"event_loop"`** (was `"llm_tool_use"`)
+- **`NodeSpec.max_node_visits` defaults to `0` / unlimited** (was `1`)
+- **`NodeSpec.function` field removed** -- `FunctionNode` is deleted; use event_loop nodes with tools instead.
+
+---
+
+### Community Contributors
+
+A huge thank you to everyone who contributed to this release:
+
+- **Richard Tang** (@RichardTang-Aden) -- Interactive credential setup, pre-start confirmation, email agent consolidation, tool registration fixes, lint and formatting
+- **Pravin Mishra** (@mishrapravin114) -- Discord integration with 4 MCP tools
+- **Jeet Karia** (@JeetKaria06) -- Exa Search API integration with 4 AI-powered search tools
+- **Shivam Shahi** (@shivamshahi07) -- Razorpay payment processing integration
+- **Siddharth Varshney** (@Siddharth2624) -- Utils module exports
+- **@haliaeetusvocifer** -- Google Docs integration with OAuth support
+- **Bryan** (@bryanadenhq) -- PDF selection, inbox agent fixes, Job Hunter and Vulnerability Assessment updates
+- **@paarths-collab** -- Documentation improvements
+
+---
+
+### Upgrading
+
+```bash
+git pull origin main
+uv sync
+```
+
+#### Migration Guide
+
+If your agents use deprecated node types, update them:
+
+```python
+# Before (v0.5.0) -- these now raise RuntimeError
+NodeSpec(node_type="llm_tool_use", ...)
+NodeSpec(node_type="function", function=my_func, ...)
+
+# After (v0.5.1) -- use event_loop for everything
+NodeSpec(node_type="event_loop", ...)  # or just omit node_type (it's the default now)
+```
+
+If your agents set `max_node_visits=1` explicitly, they'll still work. The only change is the _default_ -- new agents without an explicit value now get unlimited visits.
+
+To try the new Hive Coder:
+
+```bash
+# Launch Coder directly
+hive code
+
+# Or from TUI -- press Ctrl+E to escalate
+hive tui
+```

CLAUDE.md

@@ -0,0 +1 @@
+AGENTS.md

ENVIRONMENT_SETUP.md

@@ -1,495 +0,0 @@
-# Agent Development Environment Setup
-
-Complete setup guide for building and running goal-driven agents with the Aden Agent Framework.
-
-## Quick Setup
-
-```bash
-# Run the automated setup script
-./quickstart.sh
-```
-
-> **Note for Windows Users:**  
-> Running the setup script on native Windows shells (PowerShell / Git Bash) may sometimes fail due to Python App Execution Aliases.  
-> It is **strongly recommended to use WSL (Windows Subsystem for Linux)** for a smoother setup experience.
-
-This will:
-
-- Check Python version (requires 3.11+)
-- Install the core framework package (`framework`)
-- Install the tools package (`aden_tools`)
-- Fix package compatibility issues (openai + litellm)
-- Verify all installations
-
-## Alpine Linux Setup
-
-If you are using Alpine Linux (e.g., inside a Docker container), you must install system dependencies and use a virtual environment before running the setup script:
-
-1. Install System Dependencies:
-```bash
-apk update
-apk add bash git python3 py3-pip nodejs npm curl build-base python3-dev linux-headers libffi-dev
-```
-2. Set up Virtual Environment (Required for Python 3.12+):
-```
-python3 -m venv venv
-source venv/bin/activate
-pip install --upgrade pip setuptools wheel
-```
-3. Run the Quickstart Script:
-```
-./quickstart.sh
-```
-
-## Manual Setup (Alternative)
-
-If you prefer to set up manually or the script fails:
-
-### 1. Install Core Framework
-
-```bash
-cd core
-pip install -e .
-```
-
-### 2. Install Tools Package
-
-```bash
-cd tools
-pip install -e .
-```
-
-### 3. Upgrade OpenAI Package
-
-```bash
-# litellm requires openai >= 1.0.0
-pip install --upgrade "openai>=1.0.0"
-```
-
-### 4. Verify Installation
-
-```bash
-python -c "import framework; print('✓ framework OK')"
-python -c "import aden_tools; print('✓ aden_tools OK')"
-python -c "import litellm; print('✓ litellm OK')"
-```
-
-> **Windows Tip:**  
-> On Windows, if the verification commands fail, ensure you are running them in **WSL** or after **disabling Python App Execution Aliases** in Windows Settings → Apps → App Execution Aliases.
-
-## Requirements
-
-### Python Version
-
-- **Minimum:** Python 3.11
-- **Recommended:** Python 3.11 or 3.12
-- **Tested on:** Python 3.11, 3.12, 3.13
-
-### System Requirements
-
-- pip (latest version)
-- 2GB+ RAM
-- Internet connection (for LLM API calls)
-- For Windows users: WSL 2 is recommended for full compatibility.
-
-### API Keys (Optional)
-
-For running agents with real LLMs:
-
-```bash
-export ANTHROPIC_API_KEY="your-key-here"
-```
-
-## Running Agents
-
-All agent commands must be run from the project root with `PYTHONPATH` set:
-
-```bash
-# From /hive/ directory
-PYTHONPATH=core:exports python -m agent_name COMMAND
-```
-
-### Example Commands
-
-After building an agent via `/building-agents-construction`, use these commands:
-
-```bash
-# Validate agent structure
-PYTHONPATH=core:exports python -m your_agent_name validate
-
-# Show agent information
-PYTHONPATH=core:exports python -m your_agent_name info
-
-# Run agent with input
-PYTHONPATH=core:exports python -m your_agent_name run --input '{
-  "task": "Your input here"
-}'
-
-# Run in mock mode (no LLM calls)
-PYTHONPATH=core:exports python -m your_agent_name run --mock --input '{...}'
-```
-
-## Building New Agents and Run Flow
-
-Build and run an agent using Claude Code CLI with the agent building skills:
-
-### 1. Install Claude Skills (One-time)
-
-```bash
-./quickstart.sh
-```
-
-This verifies agent-related Claude Code skills are available:
-
-- `/building-agents-construction` - Step-by-step build guide
-- `/building-agents-core` - Fundamental concepts
-- `/building-agents-patterns` - Best practices
-- `/testing-agent` - Test and validate agents
-- `/agent-workflow` - Complete workflow
-
-### 2. Build an Agent
-
-```
-claude> /building-agents-construction
-```
-
-Follow the prompts to:
-
-1. Define your agent's goal
-2. Design the workflow nodes
-3. Connect nodes with edges
-4. Generate the agent package under `exports/`
-
-This step creates the initial agent structure required for further development.
-
-### 3. Define Agent Logic
-
-```
-claude> /building-agents-core
-```
-
-Follow the prompts to:
-
-1. Understand the agent architecture and file structure
-2. Define the agent's goal, success criteria, and constraints
-3. Learn node types (LLM, tool-use, router, function)
-4. Discover and validate available tools before use
-
-This step establishes the core concepts and rules needed before building an agent.
-
-### 4. Apply Agent Patterns
-
-```
-claude> /building-agents-patterns
-```
-
-Follow the prompts to:
-
-1. Apply best-practice agent design patterns
-2. Add pause/resume flows for multi-turn interactions
-3. Improve robustness with routing, fallbacks, and retries
-4. Avoid common anti-patterns during agent construction
-
-This step helps optimize agent design before final testing.
-
-### 5. Test Your Agent
-
-```
-claude> /testing-agent
-```
-Follow the prompts to:
-
-1. Generate test guidelines for constraints and success criteria
-2. Write agent tests directly under `exports/{agent}/tests/`
-3. Run goal-based evaluation tests
-4. Debug failing tests and iterate on agent improvements
-
-This step verifies that the agent meets its goals before production use.
-
-### 6. Agent Development Workflow (End-to-End)
-
-```
-claude> /agent-workflow
-```
-
-Follow the guided flow to:
-
-1. Understand core agent concepts (optional)
-2. Build the agent structure step by step
-3. Apply best-practice design patterns (optional)
-4. Test and validate the agent against its goals
-
-This workflow orchestrates all agent-building skills to take you from idea → production-ready agent.
-
-## Troubleshooting
-
-### "externally-managed-environment" error (PEP 668)
-
-**Cause:** Python 3.12+ on macOS/Homebrew, WSL, or some Linux distros prevents system-wide pip installs.
-
-**Solution:** Create and use a virtual environment:
-
-```bash
-# Create virtual environment
-python3 -m venv .venv
-
-# Activate it
-source .venv/bin/activate  # macOS/Linux
-# .venv\Scripts\activate   # Windows
-
-# Then run setup
-./quickstart.sh
-```
-
-Always activate the venv before running agents:
-
-```bash
-source .venv/bin/activate
-PYTHONPATH=core:exports python -m your_agent_name demo
-```
-
-### "ModuleNotFoundError: No module named 'framework'"
-
-**Solution:** Install the core package:
-
-```bash
-cd core && pip install -e .
-```
-
-### "ModuleNotFoundError: No module named 'aden_tools'"
-
-**Solution:** Install the tools package:
-
-```bash
-cd tools && pip install -e .
-```
-
-Or run the setup script:
-
-```bash
-./quickstart.sh
-```
-
-### "ModuleNotFoundError: No module named 'openai.\_models'"
-
-**Cause:** Outdated `openai` package (0.27.x) incompatible with `litellm`
-
-**Solution:** Upgrade openai:
-
-```bash
-pip install --upgrade "openai>=1.0.0"
-```
-
-### "No module named 'your_agent_name'"
-
-**Cause:** Not running from project root, missing PYTHONPATH, or agent not yet created
-
-**Solution:** Ensure you're in the project root directory, have built an agent, and use:
-
-```bash
-PYTHONPATH=core:exports python -m your_agent_name validate
-```
-
-### Agent imports fail with "broken installation"
-
-**Symptom:** `pip list` shows packages pointing to non-existent directories
-
-**Solution:** Reinstall packages properly:
-
-```bash
-# Remove broken installations
-pip uninstall -y framework tools
-
-# Reinstall correctly
-./quickstart.sh
-```
-
-## Package Structure
-
-The Hive framework consists of three Python packages:
-
-```
-hive/
-├── core/                    # Core framework (runtime, graph executor, LLM providers)
-│   ├── framework/
-│   ├── .venv/              # Created by quickstart.sh
-│   └── pyproject.toml
-│
-├── tools/                   # Tools and MCP servers
-│   ├── src/
-│   │   └── aden_tools/     # Actual package location
-│   ├── .venv/              # Created by quickstart.sh
-│   └── pyproject.toml
-│
-└── exports/                 # Agent packages (user-created, gitignored)
-    └── your_agent_name/     # Created via /building-agents-construction
-```
-
-## Separate Virtual Environments
-
-The project uses **separate virtual environments** for `core` and `tools` packages to:
-
-- Isolate dependencies and avoid conflicts
-- Allow independent development and testing of each package
-- Enable MCP servers to run with their specific dependencies
-
-### How It Works
-
-When you run `./quickstart.sh` or `uv sync` in each directory:
-
-1. **core/.venv/** - Contains the `framework` package and its dependencies (anthropic, litellm, mcp, etc.)
-2. **tools/.venv/** - Contains the `aden_tools` package and its dependencies (beautifulsoup4, pandas, etc.)
-
-### Cross-Package Imports
-
-The `core` and `tools` packages are **intentionally independent**:
-
-- **No cross-imports**: `framework` does not import `aden_tools` directly, and vice versa
-- **Communication via MCP**: Tools are exposed to agents through MCP servers, not direct Python imports
-- **Runtime integration**: The agent runner loads tools via the MCP protocol at runtime
-
-If you need to use both packages in a single script (e.g., for testing), you have two options:
-
-```bash
-# Option 1: Install both in a shared environment
-python -m venv .venv
-source .venv/bin/activate
-pip install -e core/ -e tools/
-
-# Option 2: Use PYTHONPATH (for quick testing)
-PYTHONPATH=core:tools/src python your_script.py
-```
-
-### MCP Server Configuration
-
-The `.mcp.json` at project root configures MCP servers to use their respective virtual environments:
-
-```json
-{
-  "mcpServers": {
-    "agent-builder": {
-      "command": "core/.venv/bin/python",
-      "args": ["-m", "framework.mcp.agent_builder_server"]
-    },
-    "tools": {
-      "command": "tools/.venv/bin/python",
-      "args": ["-m", "aden_tools.mcp_server", "--stdio"]
-    }
-  }
-}
-```
-
-This ensures each MCP server runs with its correct dependencies.
-
-### Why PYTHONPATH is Required
-
-The packages are installed in **editable mode** (`pip install -e`), which means:
-
-- `framework` and `aden_tools` are globally importable (no PYTHONPATH needed)
-- `exports` is NOT installed as a package (PYTHONPATH required)
-
-This design allows agents in `exports/` to be:
-
-- Developed independently
-- Version controlled separately
-- Deployed as standalone packages
-
-## Development Workflow
-
-### 1. Setup (Once)
-
-```bash
-./quickstart.sh
-```
-
-### 2. Build Agent (Claude Code)
-
-```
-claude> /building-agents-construction
-Enter goal: "Build an agent that processes customer support tickets"
-```
-
-### 3. Validate Agent
-
-```bash
-PYTHONPATH=core:exports python -m your_agent_name validate
-```
-
-### 4. Test Agent
-
-```
-claude> /testing-agent
-```
-
-### 5. Run Agent
-
-```bash
-PYTHONPATH=core:exports python -m your_agent_name run --input '{...}'
-```
-
-## IDE Setup
-
-### VSCode
-
-Add to `.vscode/settings.json`:
-
-```json
-{
-  "python.analysis.extraPaths": [
-    "${workspaceFolder}/core",
-    "${workspaceFolder}/exports"
-  ],
-  "python.autoComplete.extraPaths": [
-    "${workspaceFolder}/core",
-    "${workspaceFolder}/exports"
-  ]
-}
-```
-
-### PyCharm
-
-1. Open Project Settings → Project Structure
-2. Mark `core` as Sources Root
-3. Mark `exports` as Sources Root
-
-## Environment Variables
-
-### Required for LLM Operations
-
-```bash
-export ANTHROPIC_API_KEY="sk-ant-..."
-```
-
-### Optional Configuration
-
-```bash
-# Credentials storage location (default: ~/.aden/credentials)
-export ADEN_CREDENTIALS_PATH="/custom/path"
-
-# Agent storage location (default: /tmp)
-export AGENT_STORAGE_PATH="/custom/storage"
-```
-
-## Additional Resources
-
-- **Framework Documentation:** [core/README.md](core/README.md)
-- **Tools Documentation:** [tools/README.md](tools/README.md)
-- **Example Agents:** [exports/](exports/)
-- **Agent Building Guide:** [.claude/skills/building-agents-construction/SKILL.md](.claude/skills/building-agents-construction/SKILL.md)
-- **Testing Guide:** [.claude/skills/testing-agent/SKILL.md](.claude/skills/testing-agent/SKILL.md)
-
-## Contributing
-
-When contributing agent packages:
-
-1. Place agents in `exports/agent_name/`
-2. Follow the standard agent structure (see existing agents)
-3. Include README.md with usage instructions
-4. Add tests if using `/testing-agent`
-5. Document required environment variables
-
-## Support
-
-- **Issues:** https://github.com/adenhq/hive/issues
-- **Discord:** https://discord.com/invite/MXE49hrKDk
-- **Documentation:** https://docs.adenhq.com/

Makefile

@@ -1,26 +1,56 @@
-.PHONY: lint format check test install-hooks help
+.PHONY: lint format check test test-tools test-live test-all install-hooks help frontend-install frontend-dev frontend-build
+
+# ── Ensure uv is findable in Git Bash on Windows ──────────────────────────────
+# uv installs to ~/.local/bin on Windows/Linux/macOS. Git Bash may not include
+# this in PATH by default, so we prepend it here.
+export PATH := $(HOME)/.local/bin:$(PATH)
+
+# ── Targets ───────────────────────────────────────────────────────────────────
 
 help: ## Show this help
 	@grep -E '^[a-zA-Z_-]+:.*?## .*$$' $(MAKEFILE_LIST) | \
 		awk 'BEGIN {FS = ":.*?## "}; {printf "  \033[36m%-15s\033[0m %s\n", $$1, $$2}'
 
-lint: ## Run ruff linter (with auto-fix)
-	cd core && ruff check --fix .
-	cd tools && ruff check --fix .
+lint: ## Run ruff linter and formatter (with auto-fix)
+	cd core && uv run ruff check --fix .
+	cd tools && uv run ruff check --fix .
+	cd core && uv run ruff format .
+	cd tools && uv run ruff format .
 
 format: ## Run ruff formatter
-	cd core && ruff format .
-	cd tools && ruff format .
+	cd core && uv run ruff format .
+	cd tools && uv run ruff format .
 
 check: ## Run all checks without modifying files (CI-safe)
-	cd core && ruff check .
-	cd tools && ruff check .
-	cd core && ruff format --check .
-	cd tools && ruff format --check .
+	cd core && uv run ruff check .
+	cd tools && uv run ruff check .
+	cd core && uv run ruff format --check .
+	cd tools && uv run ruff format --check .
 
-test: ## Run all tests
-	cd core && python -m pytest tests/ -v
+test: ## Run all tests (core + tools, excludes live)
+	cd core && uv run python -m pytest tests/ -v
+	cd tools && uv run python -m pytest -v
+
+test-tools: ## Run tool tests only (mocked, no credentials needed)
+	cd tools && uv run python -m pytest -v
+
+test-live: ## Run live integration tests (requires real API credentials)
+	cd tools && uv run python -m pytest -m live -s -o "addopts=" --log-cli-level=INFO
+
+test-all: ## Run everything including live tests
+	cd core && uv run python -m pytest tests/ -v
+	cd tools && uv run python -m pytest -v
+	cd tools && uv run python -m pytest -m live -s -o "addopts=" --log-cli-level=INFO
 
 install-hooks: ## Install pre-commit hooks
-	pip install pre-commit
+	uv pip install pre-commit
 	pre-commit install
+
+frontend-install: ## Install frontend npm packages
+	cd core/frontend && npm install
+
+frontend-dev: ## Start frontend dev server
+	cd core/frontend && npm run dev
+
+frontend-build: ## Build frontend for production
+	cd core/frontend && npm run build

PR.md

@@ -1,51 +0,0 @@
-## Summary
-- **Added HubSpot integration** — new HubSpot MCP tool with search, get, create, and update operations for contacts, companies, and deals. Includes OAuth2 provider for HubSpot credentials and credential store adapter for the tools layer.
-- **Replaced web_scrape tool with Playwright + stealth** — swapped httpx/BeautifulSoup for a headless Chromium browser using `playwright` (async API) and `playwright-stealth`, enabling JS-rendered page scraping and bot detection evasion
-- **Added empty response retry logic** — LLM provider now detects empty responses (e.g. Gemini returning 200 with no content on rate limit) and retries with exponential backoff, preventing hallucinated output from the cleanup LLM
-- **Added context-aware input compaction** — LLM nodes now estimate input token count before calling the model and progressively truncate the largest values if they exceed the context window budget
-- **Increased rate limit retries to 10** with verbose `[retry]` and `[compaction]` logging that includes model name, finish reason, and attempt count
-- **Updated setup scripts** — `scripts/setup-python.sh` now installs Playwright Chromium browser automatically for web scraping support
-- **Interactive quickstart onboarding** — `quickstart.sh` rewritten as bee-themed interactive wizard that detects existing API keys (including Claude Code subscription), lets user pick ONE default LLM provider, and saves configuration to `~/.hive/configuration.json`
-- **Fixed lint errors** across `hubspot_tool.py` (line length) and `agent_builder_server.py` (unused variable)
-
-## Changed files
-
-### HubSpot Integration
-- `tools/src/aden_tools/tools/hubspot_tool/` — New MCP tool: contacts, companies, and deals CRUD
-- `tools/src/aden_tools/tools/__init__.py` — Registered HubSpot tools
-- `tools/src/aden_tools/credentials/integrations.py` — HubSpot credential integration
-- `tools/src/aden_tools/credentials/__init__.py` — Updated credential exports
-- `core/framework/credentials/oauth2/hubspot_provider.py` — HubSpot OAuth2 provider
-- `core/framework/credentials/oauth2/__init__.py` — Registered HubSpot OAuth2 provider
-- `core/framework/runner/runner.py` — Updated runner for credential support
-
-### Web Scrape Rewrite
-- `tools/src/aden_tools/tools/web_scrape_tool/web_scrape_tool.py` — Playwright async rewrite
-- `tools/src/aden_tools/tools/web_scrape_tool/README.md` — Updated docs
-- `tools/pyproject.toml` — Added `playwright`, `playwright-stealth` deps
-- `tools/Dockerfile` — Added `playwright install chromium --with-deps`
-- `scripts/setup-python.sh` — Added Playwright Chromium browser install step
-
-### LLM Reliability
-- `core/framework/llm/litellm.py` — Empty response retry + max retries 10 + verbose logging
-- `core/framework/graph/node.py` — Input compaction via `_compact_inputs()`, `_estimate_tokens()`, `_get_context_limit()`
-
-### Quickstart & Setup
-- `quickstart.sh` — Interactive bee-themed onboarding wizard with single provider selection
-- `~/.hive/configuration.json` — New user config file for default LLM provider/model
-
-### Fixes
-- `core/framework/mcp/agent_builder_server.py` — Removed unused variable
-- `tools/src/aden_tools/tools/hubspot_tool/hubspot_tool.py` — Fixed E501 line length violations
-
-## Test plan
-- [ ] Run `make lint` — passes clean
-- [ ] Run `./quickstart.sh` and verify interactive flow works, config saved to `~/.hive/configuration.json`
-- [ ] Run `./scripts/setup-python.sh` and verify Playwright Chromium installs
-- [ ] Run `pytest tests/tools/test_web_scrape_tool.py -v`
-- [ ] Run agent against a JS-heavy site and verify `web_scrape` returns rendered content
-- [ ] Set `HUBSPOT_ACCESS_TOKEN` and verify HubSpot tool CRUD operations work
-- [ ] Trigger rate limit and verify `[retry]` logs appear with correct attempt counts
-- [ ] Run agent with large inputs and verify `[compaction]` logs show truncation
-
-🤖 Generated with [Claude Code](https://claude.com/claude-code)

README.md

@@ -1,5 +1,5 @@
 <p align="center">
-  <img width="100%" alt="Hive Banner" src="https://storage.googleapis.com/aden-prod-assets/website/aden-title-card.png" />
+  <img width="100%" alt="Hive Banner" src="https://github.com/user-attachments/assets/a027429b-5d3c-4d34-88e4-0feaeaabbab3" />
 </p>
 
 <p align="center">
@@ -13,116 +13,154 @@
   <a href="docs/i18n/ko.md">한국어</a>
 </p>
 
-[![Apache 2.0 License](https://img.shields.io/badge/License-Apache%202.0-blue.svg)](https://github.com/adenhq/hive/blob/main/LICENSE)
-[![Y Combinator](https://img.shields.io/badge/Y%20Combinator-Aden-orange)](https://www.ycombinator.com/companies/aden)
-[![Docker Pulls](https://img.shields.io/docker/pulls/adenhq/hive?logo=Docker&labelColor=%23528bff)](https://hub.docker.com/u/adenhq)
-[![Discord](https://img.shields.io/discord/1172610340073242735?logo=discord&labelColor=%235462eb&logoColor=%23f5f5f5&color=%235462eb)](https://discord.com/invite/MXE49hrKDk)
-[![Twitter Follow](https://img.shields.io/twitter/follow/teamaden?logo=X&color=%23f5f5f5)](https://x.com/aden_hq)
-[![LinkedIn](https://custom-icon-badges.demolab.com/badge/LinkedIn-0A66C2?logo=linkedin-white&logoColor=fff)](https://www.linkedin.com/company/teamaden/)
+<p align="center">
+  <a href="https://github.com/aden-hive/hive/blob/main/LICENSE"><img src="https://img.shields.io/badge/License-Apache%202.0-blue.svg" alt="Apache 2.0 License" /></a>
+  <a href="https://www.ycombinator.com/companies/aden"><img src="https://img.shields.io/badge/Y%20Combinator-Aden-orange" alt="Y Combinator" /></a>
+  <a href="https://discord.com/invite/MXE49hrKDk"><img src="https://img.shields.io/discord/1172610340073242735?logo=discord&labelColor=%235462eb&logoColor=%23f5f5f5&color=%235462eb" alt="Discord" /></a>
+  <a href="https://x.com/aden_hq"><img src="https://img.shields.io/twitter/follow/teamaden?logo=X&color=%23f5f5f5" alt="Twitter Follow" /></a>
+  <a href="https://www.linkedin.com/company/teamaden/"><img src="https://custom-icon-badges.demolab.com/badge/LinkedIn-0A66C2?logo=linkedin-white&logoColor=fff" alt="LinkedIn" /></a>
+  <img src="https://img.shields.io/badge/MCP-102_Tools-00ADD8?style=flat-square" alt="MCP" />
+</p>
 
 <p align="center">
+  <img src="https://img.shields.io/badge/Agent_Harness-Runtime_Layer-ff6600?style=flat-square" alt="Agent Harness" />
   <img src="https://img.shields.io/badge/AI_Agents-Self--Improving-brightgreen?style=flat-square" alt="AI Agents" />
   <img src="https://img.shields.io/badge/Multi--Agent-Systems-blue?style=flat-square" alt="Multi-Agent" />
-  <img src="https://img.shields.io/badge/Goal--Driven-Development-purple?style=flat-square" alt="Goal-Driven" />
+  <img src="https://img.shields.io/badge/Headless-Development-purple?style=flat-square" alt="Headless" />
   <img src="https://img.shields.io/badge/Human--in--the--Loop-orange?style=flat-square" alt="HITL" />
-  <img src="https://img.shields.io/badge/Production--Ready-red?style=flat-square" alt="Production" />
+  <img src="https://img.shields.io/badge/Browser-Use-red?style=flat-square" alt="Browser Use" />
 </p>
 <p align="center">
   <img src="https://img.shields.io/badge/OpenAI-supported-412991?style=flat-square&logo=openai" alt="OpenAI" />
   <img src="https://img.shields.io/badge/Anthropic-supported-d4a574?style=flat-square" alt="Anthropic" />
   <img src="https://img.shields.io/badge/Google_Gemini-supported-4285F4?style=flat-square&logo=google" alt="Gemini" />
-  <img src="https://img.shields.io/badge/MCP-19_Tools-00ADD8?style=flat-square" alt="MCP" />
 </p>
 
+<p align="center"><em>The agent harness for production workloads — state management, failure recovery, observability, and human oversight so your agents actually run.</em></p>
+
 ## Overview
 
-Build reliable, self-improving AI agents without hardcoding workflows. Define your goal through conversation with a coding agent, and the framework generates a node graph with dynamically created connection code. When things break, the framework captures failure data, evolves the agent through the coding agent, and redeploys. Built-in human-in-the-loop nodes, credential management, and real-time monitoring give you control without sacrificing adaptability.
+Hive is a runtime harness for AI agents in production. You describe your goal in natural language; a coding agent (the queen) generates the agent graph and connection code to achieve it. During execution, the harness manages state isolation, checkpoint-based crash recovery, cost enforcement, and real-time observability. When agents fail, the framework captures failure data, evolves the graph through the coding agent, and redeploys automatically. Built-in human-in-the-loop nodes, browser control, credential management, and parallel execution give you production reliability without sacrificing adaptability.
 
 Visit [adenhq.com](https://adenhq.com) for complete documentation, examples, and guides.
 
-## What is Aden
+Visit [HoneyComb](http://honeycomb.open-hive.com/) to see what jobs are being automated by AI. It’s a stock market for jobs, driven by our community’s AI agent progress. You can long and short jobs (with no real money but compute token)based on how much you think a job is going to be replaced by AI.
 
-<p align="center">
-  <img width="100%" alt="Aden Architecture" src="docs/assets/aden-architecture-diagram.jpg" />
-</p>
+https://github.com/user-attachments/assets/bf10edc3-06ba-48b6-98ba-d069b15fb69d
 
-Aden is a platform for building, deploying, operating, and adapting AI agents:
 
-- **Build** - A Coding Agent generates specialized Worker Agents (Sales, Marketing, Ops) from natural language goals
-- **Deploy** - Headless deployment with CI/CD integration and full API lifecycle management
-- **Operate** - Real-time monitoring, observability, and runtime guardrails keep agents reliable
-- **Adapt** - Continuous evaluation, supervision, and adaptation ensure agents improve over time
-- **Infra** - Shared memory, LLM integrations, tools, and skills power every agent
+## Who Is Hive For?
+
+Hive is the harness layer for teams moving AI agents from prototype to production. Models are getting better on their own — the bottleneck is the infrastructure around them: state management, failure recovery, cost control, and observability.
+
+Hive is a good fit if you:
+
+- Want AI agents that **execute real business processes**, not demos
+- Need a **runtime that handles state, recovery, and parallel execution** at scale
+- Need **self-healing and adaptive agents** that improve over time
+- Require **human-in-the-loop control**, observability, and cost limits
+- Plan to run agents in **production** where uptime, cost, and auditability matter
+
+Hive may not be the best fit if you’re only experimenting with simple agent chains or one-off scripts.
+
+## When Should You Use Hive?
+
+Use Hive when the bottleneck is no longer the model but the harness around it:
+
+- Long-running agents that need **state persistence and crash recovery**
+- Production workloads requiring **cost enforcement, observability, and audit trails**
+- Agents that **self-heal** through failure capture and graph evolution
+- Multi-agent coordination with **session isolation and shared buffers**
+- A framework that **scales with model improvements** rather than fighting them
 
 ## Quick Links
 
 - **[Documentation](https://docs.adenhq.com/)** - Complete guides and API reference
 - **[Self-Hosting Guide](https://docs.adenhq.com/getting-started/quickstart)** - Deploy Hive on your infrastructure
-- **[Changelog](https://github.com/adenhq/hive/releases)** - Latest updates and releases
-<!-- - **[Roadmap](https://adenhq.com/roadmap)** - Upcoming features and plans -->
-- **[Report Issues](https://github.com/adenhq/hive/issues)** - Bug reports and feature requests
+- **[Changelog](https://github.com/aden-hive/hive/releases)** - Latest updates and releases
+- **[Roadmap](docs/roadmap.md)** - Upcoming features and plans
+- **[Report Issues](https://github.com/aden-hive/hive/issues)** - Bug reports and feature requests
+- **[Contributing](CONTRIBUTING.md)** - How to contribute and submit PRs
 
 ## Quick Start
 
 ### Prerequisites
 
-- [Python 3.11+](https://www.python.org/downloads/) for agent development
-- Claude Code or Cursor for utilizing agent skills
+- Python 3.11+ for agent development
+- An LLM provider that powers the agents
+- **ripgrep (optional, recommended on Windows):** The `search_files` tool uses ripgrep for faster file search. If not installed, a Python fallback is used. On Windows: `winget install BurntSushi.ripgrep` or `scoop install ripgrep`
+
+> **Windows Users:** Native Windows is supported via `quickstart.ps1` and `hive.ps1`. Run these in PowerShell 5.1+. WSL is also an option but not required.
 
 ### Installation
 
+> **Note**
+> Hive uses a `uv` workspace layout and is not installed with `pip install`.
+> Running `pip install -e .` from the repository root will create a placeholder package and Hive will not function correctly.
+> Please use the quickstart script below to set up the environment.
+
 ```bash
 # Clone the repository
-git clone https://github.com/adenhq/hive.git
+git clone https://github.com/aden-hive/hive.git
 cd hive
 
-# Run quickstart setup
+# Run quickstart setup (macOS/Linux)
 ./quickstart.sh
+
+# Windows (PowerShell)
+.\quickstart.ps1
 ```
 
 This sets up:
+
 - **framework** - Core agent runtime and graph executor (in `core/.venv`)
 - **aden_tools** - MCP tools for agent capabilities (in `tools/.venv`)
-- All required Python dependencies
+- **credential store** - Encrypted API key storage (`~/.hive/credentials`)
+- **LLM provider** - Interactive default model configuration, including Hive LLM and OpenRouter
+- All required Python dependencies with `uv`
+
+- Finally, it will open the Hive interface in your browser
+
+> **Tip:** To reopen the dashboard later, run `hive open` from the project directory.
 
 ### Build Your First Agent
 
-```bash
-# Build an agent using Claude Code
-claude> /building-agents-construction
+Type the agent you want to build in the home input box. The queen is going to ask you questions and work out a solution with you.
 
-# Test your agent
-claude> /testing-agent
+<img width="2500" height="1214" alt="Image" src="https://github.com/user-attachments/assets/1ce19141-a78b-46f5-8d64-dbf987e048f4" />
 
-# Run your agent
-PYTHONPATH=core:exports python -m your_agent_name run --input '{...}'
-```
+### Use Template Agents
 
-**[📖 Complete Setup Guide](ENVIRONMENT_SETUP.md)** - Detailed instructions for agent development
+Click "Try a sample agent" and check the templates. You can run a template directly or choose to build your version on top of the existing template.
 
-### Cursor IDE Support
+### Run Agents
 
-Skills are also available in Cursor. To enable:
+Now you can run an agent by selecting the agent (either an existing agent or example agent). You can click the Run button on the top left, or talk to the queen agent and it can run the agent for you.
 
-1. Open Command Palette (`Cmd+Shift+P` / `Ctrl+Shift+P`)
-2. Run `MCP: Enable` to enable MCP servers
-3. Restart Cursor to load the MCP servers from `.cursor/mcp.json`
-4. Type `/` in Agent chat and search for skills (e.g., `/building-agents-construction`)
+<img width="2549" height="1174" alt="Screenshot 2026-03-12 at 9 27 36 PM" src="https://github.com/user-attachments/assets/7c7d30fa-9ceb-4c23-95af-b1caa405547d" />
 
 ## Features
 
-- **Goal-Driven Development** - Define objectives in natural language; the coding agent generates the agent graph and connection code to achieve them
-- **Adaptiveness** - Framework captures failures, calibrates according to the objectives, and evolves the agent graph
-- **Dynamic Node Connections** - No predefined edges; connection code is generated by any capable LLM based on your goals
-- **SDK-Wrapped Nodes** - Every node gets shared memory, local RLM memory, monitoring, tools, and LLM access out of the box
-- **Human-in-the-Loop** - Intervention nodes that pause execution for human input with configurable timeouts and escalation
+- **Browser-Use** - Control the browser on your computer to achieve hard tasks
+- **Parallel Execution** - Execute the generated graph in parallel. This way you can have multiple agents completing the jobs for you
+- **[Goal-Driven Generation](docs/key_concepts/goals_outcome.md)** - Define objectives in natural language; the coding agent generates the agent graph and connection code to achieve them
+- **[Adaptiveness](docs/key_concepts/evolution.md)** - Framework captures failures, calibrates according to the objectives, and evolves the agent graph
+- **[Dynamic Node Connections](docs/key_concepts/graph.md)** - No predefined edges; connection code is generated by any capable LLM based on your goals
+- **SDK-Wrapped Nodes** - Every node gets a shared data buffer, local RLM memory, monitoring, tools, and LLM access out of the box
+- **[Human-in-the-Loop](docs/key_concepts/graph.md#human-in-the-loop)** - Intervention nodes that pause execution for human input with configurable timeouts and escalation
 - **Real-time Observability** - WebSocket streaming for live monitoring of agent execution, decisions, and node-to-node communication
-- **Cost & Budget Control** - Set spending limits, throttles, and automatic model degradation policies
-- **Production-Ready** - Self-hostable, built for scale and reliability
 
-## Why Aden
+## Integration
 
-Hive focuses on generating agents that run real business processes rather than generic agents. Instead of requiring you to manually design workflows, define agent interactions, and handle failures reactively, Hive flips the paradigm: **you describe outcomes, and the system builds itself**—delivering an outcome-driven, adaptive experience with an easy-to-use set of tools and integrations.
+<a href="https://github.com/aden-hive/hive/tree/main/tools/src/aden_tools/tools"><img width="100%" alt="Integration" src="https://github.com/user-attachments/assets/a1573f93-cf02-4bb8-b3d5-b305b05b1e51" /></a>
+Hive is built to be model-agnostic and system-agnostic.
+
+- **LLM flexibility** - Hive Framework supports Anthropic, OpenAI, OpenRouter, Hive LLM, and other hosted or local models through LiteLLM-compatible providers.
+- **Business system connectivity** - Hive Framework is designed to connect to all kinds of business systems as tools, such as CRM, support, messaging, data, file, and internal APIs via MCP.
+
+## Why Hive
+
+As models improve, the upper bound of what agents can do rises — but their reliability and production value are determined by the harness. Hive focuses on generating agents that run real business processes rather than generic agents. Instead of requiring you to manually design workflows, define agent interactions, and handle failures reactively, Hive flips the paradigm: **you describe outcomes, and the system builds itself**—delivering an outcome-driven, adaptive experience with an easy-to-use set of tools and integrations.
 
 ```mermaid
 flowchart LR
@@ -156,161 +194,162 @@ flowchart LR
     style V6 fill:#fff,stroke:#ed8c00,stroke-width:1px,color:#cc5d00
 ```
 
-### The Aden Advantage
+### The Hive Advantage
 
-| Traditional Frameworks     | Aden                                   |
+| Typical Agent Frameworks   | Hive                                   |
 | -------------------------- | -------------------------------------- |
+| Focus on model orchestration | **Production harness**: state, recovery, observability |
 | Hardcode agent workflows   | Describe goals in natural language     |
 | Manual graph definition    | Auto-generated agent graphs            |
-| Reactive error handling    | Outcome-evaluation and adaptiveness               |
+| Reactive error handling    | Outcome-evaluation and adaptiveness    |
 | Static tool configurations | Dynamic SDK-wrapped nodes              |
 | Separate monitoring setup  | Built-in real-time observability       |
 | DIY budget management      | Integrated cost controls & degradation |
 
 ### How It Works
 
-1. **Define Your Goal** → Describe what you want to achieve in plain English
-2. **Coding Agent Generates** → Creates the agent graph, connection code, and test cases
-3. **Workers Execute** → SDK-wrapped nodes run with full observability and tool access
+1. **[Define Your Goal](docs/key_concepts/goals_outcome.md)** → Describe what you want to achieve in plain English
+2. **Coding Agent Generates** → Creates the [agent graph](docs/key_concepts/graph.md), connection code, and test cases
+3. **[Workers Execute](docs/key_concepts/worker_agent.md)** → SDK-wrapped nodes run with full observability and tool access
 4. **Control Plane Monitors** → Real-time metrics, budget enforcement, policy management
-5. **Adaptiveness** → On failure, the system evolves the graph and redeploys automatically
-
-## Run pre-built Agents (Coming Soon)
-
-### Run a sample agent
-Aden Hive provides a list of featured agents that you can use and build on top of.
-
-### Run an agent shared by others
-Put the agent in `exports/` and run `PYTHONPATH=core:exports python -m your_agent_name run --input '{...}'`
-
-
-For building and running goal-driven agents with the framework:
-
-```bash
-# One-time setup
-./quickstart.sh
-
-# This sets up:
-# - framework package (core runtime)
-# - aden_tools package (MCP tools)
-# - All Python dependencies
-
-# Build new agents using Claude Code skills
-claude> /building-agents-construction
-
-# Test agents
-claude> /testing-agent
-
-# Run agents
-PYTHONPATH=core:exports python -m agent_name run --input '{...}'
-```
-
-See [ENVIRONMENT_SETUP.md](ENVIRONMENT_SETUP.md) for complete setup instructions.
+5. **[Adaptiveness](docs/key_concepts/evolution.md)** → On failure, the system evolves the graph and redeploys automatically
 
 ## Documentation
 
-- **[Developer Guide](DEVELOPER.md)** - Comprehensive guide for developers
+- **[Developer Guide](docs/developer-guide.md)** - Comprehensive guide for developers
 - [Getting Started](docs/getting-started.md) - Quick setup instructions
 - [Configuration Guide](docs/configuration.md) - All configuration options
 - [Architecture Overview](docs/architecture/README.md) - System design and structure
 
 ## Roadmap
 
-Aden Hive Agent Framework aims to help developers build outcome-oriented, self-adaptive agents. See [ROADMAP.md](ROADMAP.md) for details.
+Aden Hive Agent Framework aims to help developers build outcome-oriented, self-adaptive agents. See [roadmap.md](docs/roadmap.md) for details.
 
 ```mermaid
-flowchart TD
-subgraph Foundation
-    direction LR
-    subgraph arch["Architecture"]
-        a1["Node-Based Architecture"]:::done
-        a2["Python SDK"]:::done
-        a3["LLM Integration"]:::done
-        a4["Communication Protocol"]:::done
-    end
-    subgraph ca["Coding Agent"]
-        b1["Goal Creation Session"]:::done
-        b2["Worker Agent Creation"]
-        b3["MCP Tools"]:::done
-    end
-    subgraph wa["Worker Agent"]
-        c1["Human-in-the-Loop"]:::done
-        c2["Callback Handlers"]:::done
-        c3["Intervention Points"]:::done
-        c4["Streaming Interface"]
-    end
-    subgraph cred["Credentials"]
-        d1["Setup Process"]:::done
-        d2["Pluggable Sources"]:::done
-        d3["Enterprise Secrets"]
-        d4["Integration Tools"]:::done
-    end
-    subgraph tools["Tools"]
-        e1["File Use"]:::done
-        e2["Memory STM/LTM"]:::done
-        e3["Web Search/Scraper"]:::done
-        e4["CSV/PDF"]:::done
-        e5["Excel/Email"]
-    end
-    subgraph core["Core"]
-        f1["Eval System"]
-        f2["Pydantic Validation"]:::done
-        f3["Documentation"]:::done
-        f4["Adaptiveness"]
-        f5["Sample Agents"]
-    end
-end
+flowchart TB
+    %% Main Entity
+    User([User])
 
-subgraph Expansion
-    direction LR
-    subgraph intel["Intelligence"]
-        g1["Guardrails"]
-        g2["Streaming Mode"]
-        g3["Image Generation"]
-        g4["Semantic Search"]
+    %% =========================================
+    %% EXTERNAL EVENT SOURCES
+    %% =========================================
+    subgraph ExtEventSource [External Event Source]
+        E_Sch["Schedulers"]
+        E_WH["Webhook"]
+        E_SSE["SSE"]
     end
-    subgraph mem["Memory Iteration"]
-        h1["Message Model & Sessions"]
-        h2["Storage Migration"]
-        h3["Context Building"]
-        h4["Proactive Compaction"]
-        h5["Token Tracking"]
-    end
-    subgraph evt["Event System"]
-        i1["Event Bus for Nodes"]
-    end
-    subgraph cas["Coding Agent Support"]
-        j1["Claude Code"]
-        j2["Cursor"]
-        j3["Opencode"]
-        j4["Antigravity"]
-    end
-    subgraph plat["Platform"]
-        k1["JavaScript/TypeScript SDK"]
-        k2["Custom Tool Integrator"]
-        k3["Windows Support"]
-    end
-    subgraph dep["Deployment"]
-        l1["Self-Hosted"]
-        l2["Cloud Services"]
-        l3["CI/CD Pipeline"]
-    end
-    subgraph tmpl["Templates"]
-        m1["Sales Agent"]
-        m2["Marketing Agent"]
-        m3["Analytics Agent"]
-        m4["Training Agent"]
-        m5["Smart Form Agent"]
-    end
-end
 
-classDef done fill:#9e9e9e,color:#fff,stroke:#757575
+    %% =========================================
+    %% SYSTEM NODES
+    %% =========================================
+    subgraph WorkerBees [Worker Bees]
+        WB_C["Conversation"]
+        WB_SP["System prompt"]
+
+        subgraph Graph [Graph]
+            direction TB
+            N1["Node"] --> N2["Node"] --> N3["Node"]
+            N1 -.-> AN["Active Node"]
+            N2 -.-> AN
+            N3 -.-> AN
+
+            %% Nested Event Loop Node
+            subgraph EventLoopNode [Event Loop Node]
+                ELN_L["listener"]
+                ELN_SP["System Prompt<br/>(Task)"]
+                ELN_EL["Event loop"]
+                ELN_C["Conversation"]
+            end
+        end
+    end
+
+    subgraph JudgeNode [Judge]
+        J_C["Criteria"]
+        J_P["Principles"]
+        J_EL["Event loop"] <--> J_S["Scheduler"]
+    end
+
+    subgraph QueenBee [Queen Bee]
+        QB_SP["System prompt"]
+        QB_EL["Event loop"]
+        QB_C["Conversation"]
+    end
+
+    subgraph Infra [Infra]
+        SA["Sub Agent"]
+        TR["Tool Registry"]
+        WTM["Write through Conversation Memory<br/>(Logs/RAM/Harddrive)"]
+        SM["Shared Memory<br/>(State/Harddrive)"]
+        EB["Event Bus<br/>(RAM)"]
+        CS["Credential Store<br/>(Harddrive/Cloud)"]
+    end
+
+    subgraph PC [PC]
+        B["Browser"]
+        CB["Codebase<br/>v 0.0.x ... v n.n.n"]
+    end
+
+    %% =========================================
+    %% CONNECTIONS & DATA FLOW
+    %% =========================================
+
+    %% External Event Routing
+    E_Sch --> ELN_L
+    E_WH --> ELN_L
+    E_SSE --> ELN_L
+    ELN_L -->|"triggers"| ELN_EL
+
+    %% User Interactions
+    User -->|"Talk"| WB_C
+    User -->|"Talk"| QB_C
+    User -->|"Read/Write Access"| CS
+
+    %% Inter-System Logic
+    ELN_C <-->|"Mirror"| WB_C
+    WB_C -->|"Focus"| AN
+
+    WorkerBees -->|"Inquire"| JudgeNode
+    JudgeNode -->|"Approve"| WorkerBees
+
+    %% Judge Alignments
+    J_C <-.->|"aligns"| WB_SP
+    J_P <-.->|"aligns"| QB_SP
+
+    %% Escalate path
+    J_EL -->|"Report (Escalate)"| QB_EL
+
+    %% Pub/Sub Logic
+    AN -->|"publish"| EB
+    EB -->|"subscribe"| QB_C
+
+    %% Infra and Process Spawning
+    ELN_EL -->|"Spawn"| SA
+    SA -->|"Inform"| ELN_EL
+    SA -->|"Starts"| B
+    B -->|"Report"| ELN_EL
+    TR -->|"Assigned"| ELN_EL
+    CB -->|"Modify Worker Bee"| WB_C
+
+    %% =========================================
+    %% SHARED MEMORY & LOGS ACCESS
+    %% =========================================
+
+    %% Worker Bees Access (link to node inside Graph subgraph)
+    AN <-->|"Read/Write"| WTM
+    AN <-->|"Read/Write"| SM
+
+    %% Queen Bee Access
+    QB_C <-->|"Read/Write"| WTM
+    QB_EL <-->|"Read/Write"| SM
+
+    %% Credentials Access
+    CS -->|"Read Access"| QB_C
 ```
+
 ## Contributing
+We welcome contributions from the community! We’re especially looking for help building tools, integrations, and example agents for the framework ([check #2805](https://github.com/aden-hive/hive/issues/2805)). If you’re interested in extending its functionality, this is the perfect place to start. Please see [CONTRIBUTING.md](CONTRIBUTING.md) for guidelines.
 
-We welcome contributions from the community! We’re especially looking for help building tools, integrations, and example agents for the framework ([check #2805](https://github.com/adenhq/hive/issues/2805)). If you’re interested in extending its functionality, this is the perfect place to start. Please see [CONTRIBUTING.md](CONTRIBUTING.md) for guidelines.
-
-**Important:** Please get assigned to an issue before submitting a PR. Comment on an issue to claim it, and a maintainer will assign you. Issues with reproducible steps and proposals are prioritized. This helps prevent duplicate work. 
+**Important:** Please get assigned to an issue before submitting a PR. Comment on an issue to claim it, and a maintainer will assign you. Issues with reproducible steps and proposals are prioritized. This helps prevent duplicate work.
 
 1. Find or create an issue and get assigned
 2. Fork the repository
@@ -343,13 +382,9 @@ This project is licensed under the Apache License 2.0 - see the [LICENSE](LICENS
 
 ## Frequently Asked Questions (FAQ)
 
-**Q: Does Hive depend on LangChain or other agent frameworks?**
-
-No. Hive is built from the ground up with no dependencies on LangChain, CrewAI, or other agent frameworks. The framework is designed to be lean and flexible, generating agent graphs dynamically rather than relying on predefined components.
-
 **Q: What LLM providers does Hive support?**
 
-Hive supports 100+ LLM providers through LiteLLM integration, including OpenAI (GPT-4, GPT-4o), Anthropic (Claude models), Google Gemini, DeepSeek, Mistral, Groq, and many more. Simply set the appropriate API key environment variable and specify the model name.
+Hive supports 100+ LLM providers through LiteLLM integration, including OpenAI (GPT-4, GPT-4o), Anthropic (Claude models), Google Gemini, DeepSeek, Mistral, Groq, OpenRouter, and Hive LLM. Simply set the appropriate API key environment variable and specify the model name. See [docs/configuration.md](docs/configuration.md) for provider-specific configuration examples.
 
 **Q: Can I use Hive with local AI models like Ollama?**
 
@@ -357,37 +392,21 @@ Yes! Hive supports local models through LiteLLM. Simply use the model name forma
 
 **Q: What makes Hive different from other agent frameworks?**
 
-Hive generates your entire agent system from natural language goals using a coding agent—you don't hardcode workflows or manually define graphs. When agents fail, the framework automatically captures failure data, evolves the agent graph, and redeploys. This self-improving loop is unique to Aden.
+Hive is an agent harness, not just an orchestration framework. It provides the production runtime layer — session isolation, checkpoint-based crash recovery, cost enforcement, real-time observability, and human-in-the-loop controls — that makes agents reliable enough to run real workloads. On top of that, Hive generates your entire agent system from natural language goals and automatically [evolves the graph](docs/key_concepts/evolution.md) when agents fail. The combination of a robust harness with self-improving generation is what sets Hive apart.
 
 **Q: Is Hive open-source?**
 
 Yes, Hive is fully open-source under the Apache License 2.0. We actively encourage community contributions and collaboration.
 
-**Q: Does Hive collect data from users?**
-
-Hive collects telemetry data for monitoring and observability purposes, including token usage, latency metrics, and cost tracking. Content capture (prompts and responses) is configurable and stored with team-scoped data isolation. All data stays within your infrastructure when self-hosted.
-
-**Q: What deployment options does Hive support?**
-
-Hive supports self-hosted deployments via Python packages. See the [Environment Setup Guide](ENVIRONMENT_SETUP.md) for installation instructions. Cloud deployment options and Kubernetes-ready configurations are on the roadmap.
-
-**Q: Can Hive handle complex, production-scale use cases?**
-
-Yes. Hive is explicitly designed for production environments with features like automatic failure recovery, real-time observability, cost controls, and horizontal scaling support. The framework handles both simple automations and complex multi-agent workflows.
-
 **Q: Does Hive support human-in-the-loop workflows?**
 
-Yes, Hive fully supports human-in-the-loop workflows through intervention nodes that pause execution for human input. These include configurable timeouts and escalation policies, allowing seamless collaboration between human experts and AI agents.
-
-**Q: What monitoring and debugging tools does Hive provide?**
-
-Hive includes comprehensive observability features: real-time WebSocket streaming for live agent execution monitoring, TimescaleDB-powered analytics for cost and performance metrics, health check endpoints for Kubernetes integration, and MCP tools for agent execution, including file operations, web search, data processing, and more.
+Yes, Hive fully supports [human-in-the-loop](docs/key_concepts/graph.md#human-in-the-loop) workflows through intervention nodes that pause execution for human input. These include configurable timeouts and escalation policies, allowing seamless collaboration between human experts and AI agents.
 
 **Q: What programming languages does Hive support?**
 
 The Hive framework is built in Python. A JavaScript/TypeScript SDK is on the roadmap.
 
-**Q: Can Aden agents interact with external tools and APIs?**
+**Q: Can Hive agents interact with external tools and APIs?**
 
 Yes. Aden's SDK-wrapped nodes provide built-in tool access, and the framework supports flexible tool ecosystems. Agents can integrate with external APIs, databases, and services through the node architecture.
 
@@ -397,23 +416,21 @@ Hive provides granular budget controls including spending limits, throttles, and
 
 **Q: Where can I find examples and documentation?**
 
-Visit [docs.adenhq.com](https://docs.adenhq.com/) for complete guides, API reference, and getting started tutorials. The repository also includes documentation in the `docs/` folder and a comprehensive [DEVELOPER.md](DEVELOPER.md) guide.
+Visit [docs.adenhq.com](https://docs.adenhq.com/) for complete guides, API reference, and getting started tutorials. The repository also includes documentation in the `docs/` folder and a comprehensive [developer guide](docs/developer-guide.md).
 
 **Q: How can I contribute to Aden?**
 
 Contributions are welcome! Fork the repository, create your feature branch, implement your changes, and submit a pull request. See [CONTRIBUTING.md](CONTRIBUTING.md) for detailed guidelines.
 
-**Q: When will my team start seeing results from Aden's adaptive agents?**
+## Star History
 
-Aden's adaptation loop begins working from the first execution. When an agent fails, the framework captures the failure data, helping developers evolve the agent graph through the coding agent. How quickly this translates to measurable results depends on the complexity of your use case, the quality of your goal definitions, and the volume of executions generating feedback.
-
-**Q: How does Hive compare to other agent frameworks?**
-
-Hive focuses on generating agents that run real business processes, rather than generic agents. This vision emphasizes outcome-driven design, adaptability, and an easy-to-use set of tools and integrations.
-
-**Q: Does Aden offer enterprise support?**
-
-For enterprise inquiries, contact the Aden team through [adenhq.com](https://adenhq.com) or join our [Discord community](https://discord.com/invite/MXE49hrKDk) for support and discussions.
+<a href="https://star-history.com/#aden-hive/hive&Date">
+ <picture>
+   <source media="(prefers-color-scheme: dark)" srcset="https://api.star-history.com/svg?repos=aden-hive/hive&type=Date&theme=dark" />
+   <source media="(prefers-color-scheme: light)" srcset="https://api.star-history.com/svg?repos=aden-hive/hive&type=Date" />
+   <img alt="Star History Chart" src="https://api.star-history.com/svg?repos=aden-hive/hive&type=Date" />
+ </picture>
+</a>
 
 ---
 

ROADMAP.md

@@ -1,299 +0,0 @@
-# Product Roadmap
-
-Aden Agent Framework aims to help developers build outcome oriented, self-adaptive agents. Please find our roadmap here
-
-```mermaid
-flowchart TD
-subgraph Foundation
-    direction LR
-    subgraph arch["Architecture"]
-        a1["Node-Based Architecture"]:::done
-        a2["Python SDK"]:::done
-        a3["LLM Integration"]:::done
-        a4["Communication Protocol"]:::done
-    end
-    subgraph ca["Coding Agent"]
-        b1["Goal Creation Session"]:::done
-        b2["Worker Agent Creation"]
-        b3["MCP Tools"]:::done
-    end
-    subgraph wa["Worker Agent"]
-        c1["Human-in-the-Loop"]:::done
-        c2["Callback Handlers"]:::done
-        c3["Intervention Points"]:::done
-        c4["Streaming Interface"]
-    end
-    subgraph cred["Credentials"]
-        d1["Setup Process"]:::done
-        d2["Pluggable Sources"]:::done
-        d3["Enterprise Secrets"]
-        d4["Integration Tools"]:::done
-    end
-    subgraph tools["Tools"]
-        e1["File Use"]:::done
-        e2["Memory STM/LTM"]:::done
-        e3["Web Search/Scraper"]:::done
-        e4["CSV/PDF"]:::done
-        e5["Excel/Email"]
-    end
-    subgraph core["Core"]
-        f1["Eval System"]
-        f2["Pydantic Validation"]:::done
-        f3["Documentation"]:::done
-        f4["Adaptiveness"]
-        f5["Sample Agents"]
-    end
-end
-
-subgraph Expansion
-    direction LR
-    subgraph intel["Intelligence"]
-        g1["Guardrails"]
-        g2["Streaming Mode"]
-        g3["Image Generation"]
-        g4["Semantic Search"]
-    end
-    subgraph mem["Memory Iteration"]
-        h1["Message Model & Sessions"]
-        h2["Storage Migration"]
-        h3["Context Building"]
-        h4["Proactive Compaction"]
-        h5["Token Tracking"]
-    end
-    subgraph evt["Event System"]
-        i1["Event Bus for Nodes"]
-    end
-    subgraph cas["Coding Agent Support"]
-        j1["Claude Code"]
-        j2["Cursor"]
-        j3["Opencode"]
-        j4["Antigravity"]
-    end
-    subgraph plat["Platform"]
-        k1["JavaScript/TypeScript SDK"]
-        k2["Custom Tool Integrator"]
-        k3["Windows Support"]
-    end
-    subgraph dep["Deployment"]
-        l1["Self-Hosted"]
-        l2["Cloud Services"]
-        l3["CI/CD Pipeline"]
-    end
-    subgraph tmpl["Templates"]
-        m1["Sales Agent"]
-        m2["Marketing Agent"]
-        m3["Analytics Agent"]
-        m4["Training Agent"]
-        m5["Smart Form Agent"]
-    end
-end
-
-classDef done fill:#9e9e9e,color:#fff,stroke:#757575
-```
-
----
-
-## Phase 1: Foundation
-
-### Backbone Architecture
-- [ ] **Node-Based Architecture (Agent as a node)**
-    - [x] Object schema definition
-    - [x] Node wrapper SDK
-    - [x] Shared memory access
-    - [ ] Default monitoring hooks
-    - [x] Tool access layer
-    - [x] LLM integration layer (Natively supports all mainstream LLMs through LiteLLM)
-        - [x] Anthropic
-        - [x] OpenAI
-        - [x] Google
-- [x] **Communication protocol between nodes**
-- [x] **[Coding Agent] Goal Creation Session** (separate from coding session)
-    - [x] Instruction back and forth
-    - [x] Goal Object schema definition
-    - [x] Being able to generate the test cases
-    - [x] Test case validation for worker agent (Outcome driven)
-- [ ] **[Coding Agent] Worker Agent Creation**
-    - [x] Coding Agent tools
-    - [ ] Use Template Agent as a start
-    - [x] Use our MCP tools
-- [ ] **[Worker Agent] Human-in-the-Loop**
-    - [x] Worker Agents request with questions and options
-    - [x] Callback Handler System to receive events throughout execution
-    - [x] Tool-Based Intervention Points (tool to pause execution and request human input)
-    - [x] Multiple entrypoint for different event source (e.g. Human input, webhook)
-    - [ ] Streaming Interface for Real-time Monitoring
-    - [x] Request State Management
-
-### Credential Management
-- [x] **Credentials Setup Process**
-    - [x] Install Credential MCP
-- [x] **Pluggable Credential Sources**
-    - [x] **Abstraction & Local Sources**
-        - [x] Introduce `CredentialSource` base class
-        - [x] Refactor existing logic into `EnvVarSource`
-        - [x] Implementation of Source Priority Chain mechanism
-        - [ ] Foundation unit tests
-    - [ ] **Enterprise Secret Managers**
-        - [x] `VaultSource` (HashiCorp Vault)
-        - [ ] `AWSSecretsSource` (AWS Secrets Manager)
-        - [ ] `AzureKeyVaultSource` (Azure Key Vault)
-        - [ ] Management of optional provider dependencies
-    - [ ] **Advanced Features**
-        - [x] Credential expiration and auto-refresh
-        - [ ] Audit logging for compliance/tracking
-        - [ ] Per-environment configuration support
-    - [ ] **Documentation & DX**
-        - [ ] Comprehensive source documentation
-        - [ ] Example configurations for all providers
-    - [x] **Integration as tools coverage**
-        - [x] Gsuite Tools
-        - [x] Social Media
-            - [ ] Twitter(X)
-            - [x] Github
-            - [ ] Instagram
-        - [ ] SAAS
-            - [ ] Hubspot
-            - [ ] Slack
-            - [ ] Teams
-            - [ ] Zoom
-            - [ ] Stripe
-            - [ ] Salesforce
-
-> [!IMPORTANT]
-> **Community Contribution Wanted**: We appreciate help from the community to expand the "Integration as tools" capability. Leave an issue of the integration you want to support via Hive!
-
-### Essential Tools
-- [x] **File Use Tool Kit**
-- [X] **Memory Tools**
-    - [x] STM Layer Tool (state-based short-term memory)
-    - [x] LTM Layer Tool (RLM - long-term memory)
-- [ ] **Infrastructure Tools**
-    - [x] Runtime Log Tool (logs for coding agent)
-    - [x] Web Search
-    - [x] Web Scraper
-    - [x] CSV tools
-    - [x] PDF tools
-    - [ ] Excel tools
-    - [ ] Email Tools
-    - [ ] Recipe for "Add your own tools"
-
-### Memory & File System
-- [x] DB for long-term persistent memory (Filesystem as durable scratchpad pattern)
-- [x] Session Local memory isolation
-
-### Eval System (Basic)
-- [x] Test Driven - Run test case for all agent iteration
-- [ ] Failure recording mechanism
-- [ ] SDK for defining failure conditions
-- [ ] Basic observability hooks
-- [ ] User-driven log analysis (OSS approach)
-
-### Data Validation
-- [x] Natively Support data validation of LLMs output with Pydantic
-
-### Developer Experience
-- [ ] **MVP Features**
-    - [ ] Debugging mode
-    - [ ] CLI tools for memory management
-    - [ ] CLI tools for credential management
-- [ ] **MVP Resources & Documentation**
-    - [x] Quick start guide
-    - [x] Goal creation guide
-    - [x] Agent creation guide
-    - [x] GitHub Page setup
-    - [x] README with examples
-    - [x] Contributing guidelines
-    - [ ] Introduction Video
-
-### Adaptiveness
-- [ ] Runtime data feedback loop
-- [ ] Instant Developer Feedback for improvement
-
-### Sample Agents
-- [ ] Knowledge Agent
-- [ ] Blog Writer Agent
-- [ ] SDR Agent
-
----
-
-## Phase 2: Expansion
-
-### Basic Guardrails
-- [ ] Support Basic Monitoring from Agent node SDK
-- [ ] SDK guardrail implementation (in node)
-- [ ] Guardrail type support (Determined Condition as Guardrails)
-
-### Agent Capability
-- [ ] Streaming mode support
-- [ ] Image Generation support
-- [ ] Take end user input Image and flatfile understand capability
-
-### Event-loop For Nodes (Opencode-style)
-- [ ] **Event bus**
-
-### Memory System Iteration
-- [ ] **Message Model & Session Management**
-    - [ ] Introduce `Message` class with structured content types
-    - [ ] Implement `Session` classes for conversation state
-- [ ] **Storage Migration**
-    - [ ] Implement granular per-message file persistence (`/message/[agentID]/...`)
-    - [ ] Migrate from monolithic run storage
-- [ ] **Context Building & Conversation Loop**
-    - [ ] Implement `Message.stream(sessionID)`
-    - [ ] Update `LLMNode.execute()` for full context building
-    - [ ] Implement `Message.toModelMessages()` conversion
-- [ ] **Proactive Compaction**
-    - [ ] Implement proactive overflow detection
-    - [ ] Develop backward-scanning pruning strategy (e.g., clearing old tool outputs)
-- [ ] **Enhanced Token Tracking**
-    - [ ] Extend `LLMResponse` to track reasoning and cache tokens
-    - [ ] Integrate granular token metrics into compaction logic
-
-### Coding Agent Support
-- [ ] Claude Code
-- [ ] Cursor
-- [ ] Opencode
-- [ ] Antigravity
-
-### File System Enhancement
-- [ ] Semantic Search integration
-- [ ] Interactive File System in product (frontend integration)
-
-### More Worker Tools
-- [ ] Custom Tool Integrator
-- [ ] Integration as a tool (Credential Store & Support)
-- [ ] **Core Agent Tools**
-    - [ ] Node Discovery Tool (find other agents in the graph)
-    - [ ] HITL Tool (pause execution for human approval)
-    - [ ] Wake-up Tool (resume agent tasks)
-
-### Deployment (Self-Hosted)
-- [ ] Docker container standardization
-- [ ] Headless backend execution
-- [ ] Exposed API for frontend attachment
-- [ ] Local monitoring & observability
-- [ ] Basic lifecycle APIs (Start, Stop, Pause, Resume)
-
-### Deployment (Cloud)
-- [ ] Cloud Service Options
-- [ ] Support deployment to 3rd-party platforms
-- [ ] Self-deploy + orchestrator connection
-- [ ] **CI/CD Pipeline**
-    - [ ] Automated test execution
-    - [ ] Agent version control
-    - [ ] All tests must pass for deployment
-
-### Developer Experience Enhancement
-- [ ] Tool usage documentation
-- [ ] Discord Support Channel
-
-### More Agent Templates
-- [ ] GTM Sales Agent (workflow)
-- [ ] GTM Marketing Agent (workflow)
-- [ ] Analytics Agent
-- [ ] Training Agent
-- [ ] Smart Entry / Form Agent (self-evolution emphasis)
-
-### Cross-Platform
-- [ ] JavaScript / TypeScript Version SDK
-- [ ] Better windows support

SECURITY.md

@@ -39,8 +39,8 @@ We consider security research conducted in accordance with this policy to be:
 ## Security Best Practices for Users
 
 1. **Keep Updated**: Always run the latest version
-2. **Secure Configuration**: Review `config.yaml` settings, especially in production
-3. **Environment Variables**: Never commit `.env` files or `config.yaml` with secrets
+2. **Secure Configuration**: Review your `~/.hive/configuration.json`, `.mcp.json`, and environment variable settings, especially in production
+3. **Environment Variables**: Never commit `.env` files or any configuration files that contain secrets
 4. **Network Security**: Use HTTPS in production, configure firewalls appropriately
 5. **Database Security**: Use strong passwords, limit network access
 

core/.mcp.json

@@ -1,10 +1,5 @@
 {
   "mcpServers": {
-    "agent-builder": {
-      "command": "python",
-      "args": ["-m", "framework.mcp.agent_builder_server"],
-      "cwd": "core"
-    },
     "tools": {
       "command": "python",
       "args": ["-m", "aden_tools.mcp_server", "--stdio"],

core/MCP_BUILDER_TOOLS_GUIDE.md

@@ -82,7 +82,7 @@ Register an MCP server as a tool source for your agent.
     "example_tool"
   ],
   "total_mcp_servers": 1,
-  "note": "MCP server 'tools' registered with 6 tools. These tools can now be used in llm_tool_use nodes."
+  "note": "MCP server 'tools' registered with 6 tools. These tools can now be used in event_loop nodes."
 }
 ```
 
@@ -149,7 +149,7 @@ List tools available from registered MCP servers.
     ]
   },
   "total_tools": 6,
-  "note": "Use these tool names in the 'tools' parameter when adding llm_tool_use nodes"
+  "note": "Use these tool names in the 'tools' parameter when adding event_loop nodes"
 }
 ```
 
@@ -246,7 +246,7 @@ Here's a complete workflow for building an agent with MCP tools:
     "node_id": "web-searcher",
     "name": "Web Search",
     "description": "Search the web for information",
-    "node_type": "llm_tool_use",
+    "node_type": "event_loop",
     "input_keys": "[\"query\"]",
     "output_keys": "[\"search_results\"]",
     "system_prompt": "Search for {query} using the web_search tool",

core/MCP_INTEGRATION_GUIDE.md

@@ -6,7 +6,7 @@ This guide explains how to integrate Model Context Protocol (MCP) servers with t
 
 The framework provides built-in support for MCP servers, allowing you to:
 
-- **Register MCP servers** via STDIO or HTTP transport
+- **Register MCP servers** via STDIO, HTTP, Unix socket, or SSE transport
 - **Auto-discover tools** from registered servers
 - **Use MCP tools** seamlessly in your agents
 - **Manage multiple MCP servers** simultaneously
@@ -104,6 +104,48 @@ runner.register_mcp_server(
 - `url`: Base URL of the MCP server
 - `headers`: HTTP headers to include (optional)
 
+### Unix Socket Transport
+
+Best for same-host inter-process communication with lower overhead than TCP:
+
+```python
+runner.register_mcp_server(
+    name="local-ipc-tools",
+    transport="unix",
+    url="http://localhost",
+    socket_path="/tmp/mcp_server.sock",
+    headers={
+        "Authorization": "Bearer token"
+    }
+)
+```
+
+**Configuration:**
+
+- `url`: Base URL for HTTP requests over the socket (required, e.g., `"http://localhost"`)
+- `socket_path`: Absolute path to the Unix socket file (required, e.g., `"/tmp/mcp_server.sock"`)
+- `headers`: HTTP headers to include (optional)
+
+### SSE Transport
+
+Best for real-time, event-driven connections using the MCP SDK's SSE client:
+
+```python
+runner.register_mcp_server(
+    name="streaming-tools",
+    transport="sse",
+    url="http://localhost:8000/sse",
+    headers={
+        "Authorization": "Bearer token"
+    }
+)
+```
+
+**Configuration:**
+
+- `url`: SSE endpoint URL (required, e.g., `"http://localhost:8000/sse"`)
+- `headers`: HTTP headers for the SSE connection (optional)
+
 ## Using MCP Tools in Agents
 
 Once registered, MCP tools are available just like any other tool:
@@ -119,7 +161,7 @@ builder = WorkflowBuilder()
 builder.add_node(
     node_id="researcher",
     name="Web Researcher",
-    node_type="llm_tool_use",
+    node_type="event_loop",
     system_prompt="Research the topic using web_search",
     tools=["web_search"],  # Tool from tools MCP server
     input_keys=["topic"],
@@ -137,7 +179,7 @@ Tools from MCP servers can be referenced in your agent.json just like built-in t
     {
       "id": "searcher",
       "name": "Web Searcher",
-      "node_type": "llm_tool_use",
+      "node_type": "event_loop",
       "system_prompt": "Search for information about {topic}",
       "tools": ["web_search", "web_scrape"],
       "input_keys": ["topic"],
@@ -258,7 +300,32 @@ runner.register_mcp_server(
 )
 ```
 
-### 3. Handle Cleanup
+### 3. Use Unix Socket for Same-Host IPC
+
+When both the agent and MCP server run on the same machine, Unix sockets avoid TCP overhead:
+
+```python
+runner.register_mcp_server(
+    name="fast-local-tools",
+    transport="unix",
+    url="http://localhost",
+    socket_path="/tmp/mcp_server.sock"
+)
+```
+
+### 4. Use SSE for Streaming and Real-Time Tools
+
+SSE transport maintains a persistent connection, ideal for event-driven servers:
+
+```python
+runner.register_mcp_server(
+    name="realtime-tools",
+    transport="sse",
+    url="http://realtime-server:8000/sse"
+)
+```
+
+### 5. Handle Cleanup
 
 Always clean up MCP connections when done:
 
@@ -280,7 +347,7 @@ async with AgentRunner.load("exports/my-agent") as runner:
     # Automatic cleanup
 ```
 
-### 4. Tool Name Conflicts
+### 6. Tool Name Conflicts
 
 If multiple MCP servers provide tools with the same name, the last registered server wins. To avoid conflicts:
 
@@ -315,6 +382,24 @@ If HTTP transport fails:
 2. Check firewall settings
 3. Verify the URL and port are correct
 
+### Unix Socket Not Connecting
+
+If Unix socket transport fails:
+
+1. Verify the socket file exists: `ls -la /tmp/mcp_server.sock`
+2. Check file permissions on the socket
+3. Ensure no other process has locked the socket
+4. Verify the `url` field is set (e.g., `"http://localhost"`)
+
+### SSE Connection Issues
+
+If SSE transport fails:
+
+1. Verify the server supports SSE at the given URL
+2. Check that the `mcp` Python package is installed (`pip install mcp`)
+3. Ensure the SSE endpoint is accessible: `curl http://localhost:8000/sse`
+4. Check for firewall or proxy issues blocking long-lived connections
+
 ## Example: Full Agent with MCP Tools
 
 Here's a complete example of an agent that uses MCP tools:

core/MCP_SERVER_GUIDE.md

@@ -1,17 +1,16 @@
-# MCP Server Guide - Agent Builder
+# MCP Server Guide - Agent Building Tools
 
-This guide covers the MCP (Model Context Protocol) server for building goal-driven agents.
+> **Note:** The standalone `agent-builder` MCP server (`framework.mcp.agent_builder_server`) has been replaced. Agent building is now done via the `coder-tools` server's `initialize_and_build_agent` tool, with underlying logic in `tools/coder_tools_server.py`.
+
+This guide covers the MCP tools available for building goal-driven agents.
 
 ## Setup
 
 ### Quick Setup
 
 ```bash
-# Using the setup script (recommended)
-python setup_mcp.py
-
-# Or using bash
-./setup_mcp.sh
+# Run the quickstart script (recommended)
+./quickstart.sh
 ```
 
 ### Manual Configuration
@@ -21,10 +20,10 @@ Add to your MCP client configuration (e.g., Claude Desktop):
 ```json
 {
   "mcpServers": {
-    "agent-builder": {
-      "command": "python",
-      "args": ["-m", "framework.mcp.agent_builder_server"],
-      "cwd": "/path/to/goal-agent"
+    "coder-tools": {
+      "command": "uv",
+      "args": ["run", "coder_tools_server.py", "--stdio"],
+      "cwd": "/path/to/hive/tools"
     }
   }
 }
@@ -103,31 +102,20 @@ Add a processing node to the agent graph.
 - `node_id` (string, required): Unique node identifier
 - `name` (string, required): Human-readable name
 - `description` (string, required): What this node does
-- `node_type` (string, required): One of: `llm_generate`, `llm_tool_use`, `router`, `function`
+- `node_type` (string, required): Must be `event_loop` (the only valid type)
 - `input_keys` (string, required): JSON array of input variable names
 - `output_keys` (string, required): JSON array of output variable names
-- `system_prompt` (string, optional): System prompt for LLM nodes
-- `tools` (string, optional): JSON array of tool names for tool_use nodes
-- `routes` (string, optional): JSON object of route mappings for router nodes
+- `system_prompt` (string, optional): System prompt for the LLM
+- `tools` (string, optional): JSON array of tool names
+- `client_facing` (boolean, optional): Set to true for human-in-the-loop interaction
 
-**Node Types:**
+**Node Type:**
 
-1. **llm_generate**: Uses LLM to generate output from inputs
-   - Requires: `system_prompt`
-   - Tools: Not used
-
-2. **llm_tool_use**: Uses LLM with tools to accomplish tasks
-   - Requires: `system_prompt`, `tools`
-   - Tools: Array of tool names (e.g., `["web_search", "web_fetch"]`)
-
-3. **router**: LLM-powered routing to different paths
-   - Requires: `system_prompt`, `routes`
-   - Routes: Object mapping route names to target node IDs
-   - Example: `{"pass": "success_node", "fail": "retry_node"}`
-
-4. **function**: Executes a pre-defined function
-   - System prompt describes the function behavior
-   - No LLM calls, pure computation
+**event_loop**: LLM-powered node with self-correction loop
+- Requires: `system_prompt`
+- Optional: `tools` (array of tool names, e.g., `["web_search", "web_fetch"]`)
+- Optional: `client_facing` (set to true for HITL / user interaction)
+- Supports: iterative refinement, judge-based evaluation, tool use, streaming
 
 **Example:**
 ```json
@@ -135,7 +123,7 @@ Add a processing node to the agent graph.
   "node_id": "search_sources",
   "name": "Search Sources",
   "description": "Searches for relevant sources on the topic",
-  "node_type": "llm_tool_use",
+  "node_type": "event_loop",
   "input_keys": "[\"topic\", \"search_queries\"]",
   "output_keys": "[\"sources\", \"source_count\"]",
   "system_prompt": "Search for sources using the provided queries...",
@@ -198,7 +186,7 @@ Export the validated graph as an agent specification.
 
 **What it does:**
 1. Validates the graph
-2. Auto-generates missing edges from router routes
+2. Validates edge connectivity
 3. Writes files to disk:
    - `exports/{agent-name}/agent.json` - Full agent specification
    - `exports/{agent-name}/README.md` - Auto-generated documentation
@@ -252,47 +240,6 @@ Test the complete agent graph with sample inputs.
 
 ---
 
-### Evaluation Rules
-
-#### `add_evaluation_rule`
-Add a rule for the HybridJudge to evaluate node outputs.
-
-**Parameters:**
-- `rule_id` (string, required): Unique rule identifier
-- `description` (string, required): What this rule checks
-- `condition` (string, required): Python expression to evaluate
-- `action` (string, required): Action to take: `accept`, `retry`, `escalate`
-- `priority` (integer, optional): Rule priority (default: 0)
-- `feedback_template` (string, optional): Feedback message template
-
-**Condition Examples:**
-- `'result.get("success") == True'` - Check for success flag
-- `'result.get("error_type") == "timeout"'` - Check error type
-- `'len(result.get("data", [])) > 0'` - Check for non-empty data
-
-**Example:**
-```json
-{
-  "rule_id": "timeout_retry",
-  "description": "Retry on timeout errors",
-  "condition": "result.get('error_type') == 'timeout'",
-  "action": "retry",
-  "priority": 10,
-  "feedback_template": "Timeout occurred, retrying..."
-}
-```
-
-#### `list_evaluation_rules`
-List all configured evaluation rules.
-
-#### `remove_evaluation_rule`
-Remove an evaluation rule.
-
-**Parameters:**
-- `rule_id` (string, required): Rule to remove
-
----
-
 ## Example Workflow
 
 Here's a complete workflow for building a research agent:
@@ -320,7 +267,7 @@ add_node(
     node_id="planner",
     name="Research Planner",
     description="Creates research strategy",
-    node_type="llm_generate",
+    node_type="event_loop",
     input_keys='["topic"]',
     output_keys='["strategy", "queries"]',
     system_prompt="Analyze topic and create research plan..."
@@ -330,7 +277,7 @@ add_node(
     node_id="searcher",
     name="Search Sources",
     description="Find relevant sources",
-    node_type="llm_tool_use",
+    node_type="event_loop",
     input_keys='["queries"]',
     output_keys='["sources"]',
     system_prompt="Search for sources...",
@@ -359,10 +306,9 @@ The exported agent will be saved to `exports/research-agent/`.
 
 1. **Start with the goal**: Define clear success criteria before building nodes
 2. **Test nodes individually**: Use `test_node` to verify each node works
-3. **Use router nodes for branching**: Don't create edges manually for routers - define routes and they'll be auto-generated
-4. **Add evaluation rules**: Help the judge evaluate outputs deterministically
-5. **Validate early, validate often**: Run `validate_graph` after adding nodes/edges
-6. **Check exports**: Review the generated README.md to verify your agent structure
+3. **Use conditional edges for branching**: Define condition_expr on edges for decision points
+4. **Validate early, validate often**: Run `validate_graph` after adding nodes/edges
+5. **Check exports**: Review the generated README.md to verify your agent structure
 
 ---
 

core/README.md

@@ -14,69 +14,14 @@ Framework provides a runtime framework that captures **decisions**, not just act
 ## Installation
 
 ```bash
-pip install -e .
+uv pip install -e .
 ```
 
-## MCP Server Setup
+## Agent Building
 
-The framework includes an MCP (Model Context Protocol) server for building agents. To set up the MCP server:
+Agent scaffolding is handled by the `coder-tools` MCP server (in `tools/coder_tools_server.py`), which provides the `initialize_and_build_agent` tool and related utilities. The package generation logic lives directly in `tools/coder_tools_server.py`.
 
-### Automated Setup
-
-**Using bash (Linux/macOS):**
-```bash
-./setup_mcp.sh
-```
-
-**Using Python (cross-platform):**
-```bash
-python setup_mcp.py
-```
-
-The setup script will:
-1. Install the framework package
-2. Install MCP dependencies (mcp, fastmcp)
-3. Create/verify `.mcp.json` configuration
-4. Test the MCP server module
-
-### Manual Setup
-
-If you prefer manual setup:
-
-```bash
-# Install framework
-pip install -e .
-
-# Install MCP dependencies
-pip install mcp fastmcp
-
-# Test the server
-python -m framework.mcp.agent_builder_server
-```
-
-### Using with MCP Clients
-
-To use the agent builder with Claude Desktop or other MCP clients, add this to your MCP client configuration:
-
-```json
-{
-  "mcpServers": {
-    "agent-builder": {
-      "command": "python",
-      "args": ["-m", "framework.mcp.agent_builder_server"],
-      "cwd": "/path/to/goal-agent"
-    }
-  }
-}
-```
-
-The MCP server provides tools for:
-- Creating agent building sessions
-- Defining goals with success criteria
-- Adding nodes (llm_generate, llm_tool_use, router, function)
-- Connecting nodes with edges
-- Validating and exporting agent graphs
-- Testing nodes and full agent graphs
+See the [Getting Started Guide](../docs/getting-started.md) for building agents.
 
 ## Quick Start
 
@@ -85,14 +30,14 @@ The MCP server provides tools for:
 Run an LLM-powered calculator:
 
 ```bash
-# Single calculation
-python -m framework calculate "2 + 3 * 4"
+# Run an exported agent
+uv run python -m framework run exports/calculator --input '{"expression": "2 + 3 * 4"}'
 
-# Interactive mode
-python -m framework interactive
+# Interactive shell session
+uv run python -m framework shell exports/calculator
 
-# Analyze runs with Builder
-python -m framework analyze calculator
+# Show agent info
+uv run python -m framework info exports/calculator
 ```
 
 ### Using the Runtime
@@ -136,16 +81,16 @@ Tests are generated using MCP tools (`generate_constraint_tests`, `generate_succ
 
 ```bash
 # Run tests against an agent
-python -m framework test-run <agent_path> --goal <goal_id> --parallel 4
+uv run python -m framework test-run <agent_path> --goal <goal_id> --parallel 4
 
 # Debug failed tests
-python -m framework test-debug <agent_path> <test_name>
+uv run python -m framework test-debug <agent_path> <test_name>
 
-# List tests for a goal
-python -m framework test-list <goal_id>
+# List tests for an agent
+uv run python -m framework test-list <agent_path>
 ```
 
-For detailed testing workflows, see the [testing-agent skill](../.claude/skills/testing-agent/SKILL.md).
+For detailed testing workflows, see [developer-guide.md](../docs/developer-guide.md).
 
 ### Analyzing Agent Behavior with Builder
 

core/antigravity_auth.py

@@ -0,0 +1,583 @@
+#!/usr/bin/env python3
+"""Antigravity authentication CLI.
+
+Implements OAuth2 flow for Google's Antigravity Code Assist gateway.
+Credentials are stored in ~/.hive/antigravity-accounts.json.
+
+Usage:
+    python -m antigravity_auth auth account add
+    python -m antigravity_auth auth account list
+    python -m antigravity_auth auth account remove <email>
+"""
+
+from __future__ import annotations
+
+import argparse
+import json
+import logging
+import os
+import secrets
+import socket
+import sys
+import time
+import urllib.parse
+import urllib.request
+import webbrowser
+from http.server import BaseHTTPRequestHandler, HTTPServer
+from pathlib import Path
+from typing import Any
+
+logging.basicConfig(level=logging.INFO, format="%(message)s")
+logger = logging.getLogger(__name__)
+
+# OAuth endpoints
+_OAUTH_AUTH_URL = "https://accounts.google.com/o/oauth2/v2/auth"
+_OAUTH_TOKEN_URL = "https://oauth2.googleapis.com/token"
+
+# Scopes for Antigravity/Cloud Code Assist
+_OAUTH_SCOPES = [
+    "https://www.googleapis.com/auth/cloud-platform",
+    "https://www.googleapis.com/auth/userinfo.email",
+    "https://www.googleapis.com/auth/userinfo.profile",
+]
+
+# Credentials file path in ~/.hive/
+_ACCOUNTS_FILE = Path.home() / ".hive" / "antigravity-accounts.json"
+
+# Default project ID
+_DEFAULT_PROJECT_ID = "rising-fact-p41fc"
+_DEFAULT_REDIRECT_PORT = 51121
+
+# OAuth credentials fetched from the opencode-antigravity-auth project.
+# This project reverse-engineered and published the public OAuth credentials
+# for Google's Antigravity/Cloud Code Assist API.
+# Source: https://github.com/NoeFabris/opencode-antigravity-auth
+_CREDENTIALS_URL = (
+    "https://raw.githubusercontent.com/NoeFabris/opencode-antigravity-auth/dev/src/constants.ts"
+)
+
+# Cached credentials fetched from public source
+_cached_client_id: str | None = None
+_cached_client_secret: str | None = None
+
+
+def _fetch_credentials_from_public_source() -> tuple[str | None, str | None]:
+    """Fetch OAuth client ID and secret from the public npm package source on GitHub."""
+    global _cached_client_id, _cached_client_secret
+    if _cached_client_id and _cached_client_secret:
+        return _cached_client_id, _cached_client_secret
+
+    try:
+        req = urllib.request.Request(
+            _CREDENTIALS_URL, headers={"User-Agent": "Hive-Antigravity-Auth/1.0"}
+        )
+        with urllib.request.urlopen(req, timeout=10) as resp:
+            content = resp.read().decode("utf-8")
+            import re
+
+            id_match = re.search(r'ANTIGRAVITY_CLIENT_ID\s*=\s*"([^"]+)"', content)
+            secret_match = re.search(r'ANTIGRAVITY_CLIENT_SECRET\s*=\s*"([^"]+)"', content)
+            if id_match:
+                _cached_client_id = id_match.group(1)
+            if secret_match:
+                _cached_client_secret = secret_match.group(1)
+            return _cached_client_id, _cached_client_secret
+    except Exception as e:
+        logger.debug(f"Failed to fetch credentials from public source: {e}")
+    return None, None
+
+
+def get_client_id() -> str:
+    """Get OAuth client ID from env, config, or public source."""
+    env_id = os.environ.get("ANTIGRAVITY_CLIENT_ID")
+    if env_id:
+        return env_id
+
+    # Try hive config
+    hive_cfg = Path.home() / ".hive" / "configuration.json"
+    if hive_cfg.exists():
+        try:
+            with open(hive_cfg) as f:
+                cfg = json.load(f)
+                cfg_id = cfg.get("llm", {}).get("antigravity_client_id")
+                if cfg_id:
+                    return cfg_id
+        except Exception:
+            pass
+
+    # Fetch from public source
+    client_id, _ = _fetch_credentials_from_public_source()
+    if client_id:
+        return client_id
+
+    raise RuntimeError("Could not obtain Antigravity OAuth client ID")
+
+
+def get_client_secret() -> str | None:
+    """Get OAuth client secret from env, config, or public source."""
+    secret = os.environ.get("ANTIGRAVITY_CLIENT_SECRET")
+    if secret:
+        return secret
+
+    # Try to read from hive config
+    hive_cfg = Path.home() / ".hive" / "configuration.json"
+    if hive_cfg.exists():
+        try:
+            with open(hive_cfg) as f:
+                cfg = json.load(f)
+                secret = cfg.get("llm", {}).get("antigravity_client_secret")
+                if secret:
+                    return secret
+        except Exception:
+            pass
+
+    # Fetch from public source (npm package on GitHub)
+    _, secret = _fetch_credentials_from_public_source()
+    return secret
+
+
+def find_free_port() -> int:
+    """Find an available local port."""
+    with socket.socket(socket.AF_INET, socket.SOCK_STREAM) as s:
+        s.bind(("", 0))
+        s.listen(1)
+        return s.getsockname()[1]
+
+
+class OAuthCallbackHandler(BaseHTTPRequestHandler):
+    """Handle OAuth callback from browser."""
+
+    auth_code: str | None = None
+    state: str | None = None
+    error: str | None = None
+
+    def log_message(self, format: str, *args: Any) -> None:
+        pass  # Suppress default logging
+
+    def do_GET(self) -> None:
+        parsed = urllib.parse.urlparse(self.path)
+
+        if parsed.path == "/oauth-callback":
+            query = urllib.parse.parse_qs(parsed.query)
+
+            if "error" in query:
+                self.error = query["error"][0]
+                self._send_response("Authentication failed. You can close this window.")
+                return
+
+            if "code" in query and "state" in query:
+                OAuthCallbackHandler.auth_code = query["code"][0]
+                OAuthCallbackHandler.state = query["state"][0]
+                self._send_response(
+                    "Authentication successful! You can close this window "
+                    "and return to the terminal."
+                )
+                return
+
+        self._send_response("Waiting for authentication...")
+
+    def _send_response(self, message: str) -> None:
+        self.send_response(200)
+        self.send_header("Content-Type", "text/html")
+        self.end_headers()
+        html = f"""<!DOCTYPE html>
+<html>
+<head><title>Antigravity Auth</title></head>
+<body style="font-family: system-ui; display: flex; align-items: center;
+      justify-content: center; height: 100vh; margin: 0; background: #1a1a2e;
+      color: #eee;">
+    <div style="text-align: center;">
+        <h2>{message}</h2>
+    </div>
+</body>
+</html>"""
+        self.wfile.write(html.encode())
+
+
+def wait_for_callback(port: int, timeout: int = 300) -> tuple[str | None, str | None, str | None]:
+    """Start local server and wait for OAuth callback."""
+    server = HTTPServer(("localhost", port), OAuthCallbackHandler)
+    server.timeout = 1
+
+    start = time.time()
+    while time.time() - start < timeout:
+        if OAuthCallbackHandler.auth_code:
+            return (
+                OAuthCallbackHandler.auth_code,
+                OAuthCallbackHandler.state,
+                OAuthCallbackHandler.error,
+            )
+        server.handle_request()
+
+    return None, None, "timeout"
+
+
+def exchange_code_for_tokens(
+    code: str, redirect_uri: str, client_id: str, client_secret: str | None
+) -> dict[str, Any] | None:
+    """Exchange authorization code for tokens."""
+    data = {
+        "code": code,
+        "client_id": client_id,
+        "redirect_uri": redirect_uri,
+        "grant_type": "authorization_code",
+    }
+    if client_secret:
+        data["client_secret"] = client_secret
+
+    body = urllib.parse.urlencode(data).encode()
+
+    req = urllib.request.Request(
+        _OAUTH_TOKEN_URL,
+        data=body,
+        headers={"Content-Type": "application/x-www-form-urlencoded"},
+        method="POST",
+    )
+
+    try:
+        with urllib.request.urlopen(req, timeout=30) as resp:
+            return json.loads(resp.read())
+    except Exception as e:
+        logger.error(f"Token exchange failed: {e}")
+        return None
+
+
+def get_user_email(access_token: str) -> str | None:
+    """Get user email from Google API."""
+    req = urllib.request.Request(
+        "https://www.googleapis.com/oauth2/v2/userinfo",
+        headers={"Authorization": f"Bearer {access_token}"},
+    )
+    try:
+        with urllib.request.urlopen(req, timeout=10) as resp:
+            data = json.loads(resp.read())
+            return data.get("email")
+    except Exception:
+        return None
+
+
+def load_accounts() -> dict[str, Any]:
+    """Load existing accounts from file."""
+    if not _ACCOUNTS_FILE.exists():
+        return {"schemaVersion": 4, "accounts": []}
+    try:
+        with open(_ACCOUNTS_FILE) as f:
+            return json.load(f)
+    except Exception:
+        return {"schemaVersion": 4, "accounts": []}
+
+
+def save_accounts(data: dict[str, Any]) -> None:
+    """Save accounts to file."""
+    _ACCOUNTS_FILE.parent.mkdir(parents=True, exist_ok=True)
+    with open(_ACCOUNTS_FILE, "w") as f:
+        json.dump(data, f, indent=2)
+    logger.info(f"Saved credentials to {_ACCOUNTS_FILE}")
+
+
+def validate_credentials(access_token: str, project_id: str = _DEFAULT_PROJECT_ID) -> bool:
+    """Test if credentials work by making a simple API call to Antigravity.
+
+    Returns True if credentials are valid, False otherwise.
+    """
+    endpoint = "https://daily-cloudcode-pa.sandbox.googleapis.com"
+    body = {
+        "project": project_id,
+        "model": "gemini-3-flash",
+        "request": {
+            "contents": [{"role": "user", "parts": [{"text": "hi"}]}],
+            "generationConfig": {"maxOutputTokens": 10},
+        },
+        "requestType": "agent",
+        "userAgent": "antigravity",
+        "requestId": "validation-test",
+    }
+    headers = {
+        "Authorization": f"Bearer {access_token}",
+        "Content-Type": "application/json",
+        "User-Agent": (
+            "Mozilla/5.0 (Macintosh; Intel Mac OS X 10_15_7) "
+            "AppleWebKit/537.36 (KHTML, like Gecko) Antigravity/1.18.3"
+        ),
+        "X-Goog-Api-Client": "google-cloud-sdk vscode_cloudshelleditor/0.1",
+    }
+
+    try:
+        req = urllib.request.Request(
+            f"{endpoint}/v1internal:generateContent",
+            data=json.dumps(body).encode("utf-8"),
+            headers=headers,
+            method="POST",
+        )
+        with urllib.request.urlopen(req, timeout=30) as resp:
+            json.loads(resp.read())
+            return True
+    except Exception:
+        return False
+
+
+def refresh_access_token(
+    refresh_token: str, client_id: str, client_secret: str | None
+) -> dict | None:
+    """Refresh the access token using the refresh token."""
+    data = {
+        "grant_type": "refresh_token",
+        "refresh_token": refresh_token,
+        "client_id": client_id,
+    }
+    if client_secret:
+        data["client_secret"] = client_secret
+
+    body = urllib.parse.urlencode(data).encode()
+    req = urllib.request.Request(
+        _OAUTH_TOKEN_URL,
+        data=body,
+        headers={"Content-Type": "application/x-www-form-urlencoded"},
+        method="POST",
+    )
+    try:
+        with urllib.request.urlopen(req, timeout=30) as resp:
+            return json.loads(resp.read())
+    except Exception as e:
+        logger.debug(f"Token refresh failed: {e}")
+        return None
+
+
+def cmd_account_add(args: argparse.Namespace) -> int:
+    """Add a new Antigravity account via OAuth2.
+
+    First checks if valid credentials already exist. If so, validates them
+    and skips OAuth if they work. Otherwise, proceeds with OAuth flow.
+    """
+    client_id = get_client_id()
+    client_secret = get_client_secret()
+
+    # Check if credentials already exist
+    accounts_data = load_accounts()
+    accounts = accounts_data.get("accounts", [])
+
+    if accounts:
+        account = next((a for a in accounts if a.get("enabled", True) is not False), accounts[0])
+        access_token = account.get("access")
+        refresh_token_str = account.get("refresh", "")
+        refresh_token = refresh_token_str.split("|")[0] if refresh_token_str else None
+        project_id = (
+            refresh_token_str.split("|")[1] if "|" in refresh_token_str else _DEFAULT_PROJECT_ID
+        )
+        email = account.get("email", "unknown")
+        expires_ms = account.get("expires", 0)
+        expires_at = expires_ms / 1000.0 if expires_ms else 0.0
+
+        # Check if token is expired or near expiry
+        if access_token and expires_at and time.time() < expires_at - 60:
+            # Token still valid, test it
+            logger.info(f"Found existing credentials for: {email}")
+            logger.info("Validating existing credentials...")
+            if validate_credentials(access_token, project_id):
+                logger.info("✓ Credentials valid! Skipping OAuth.")
+                return 0
+            else:
+                logger.info("Credentials failed validation, refreshing...")
+        elif refresh_token:
+            logger.info(f"Found expired credentials for: {email}")
+            logger.info("Attempting token refresh...")
+
+            tokens = refresh_access_token(refresh_token, client_id, client_secret)
+            if tokens:
+                new_access = tokens.get("access_token")
+                expires_in = tokens.get("expires_in", 3600)
+                if new_access:
+                    # Update the account
+                    account["access"] = new_access
+                    account["expires"] = int((time.time() + expires_in) * 1000)
+                    accounts_data["last_refresh"] = time.strftime(
+                        "%Y-%m-%dT%H:%M:%SZ", time.gmtime()
+                    )
+                    save_accounts(accounts_data)
+
+                    # Validate the refreshed token
+                    logger.info("Validating refreshed credentials...")
+                    if validate_credentials(new_access, project_id):
+                        logger.info("✓ Credentials refreshed and validated!")
+                        return 0
+                    else:
+                        logger.info("Refreshed token failed validation, proceeding with OAuth...")
+            else:
+                logger.info("Token refresh failed, proceeding with OAuth...")
+
+    # No valid credentials, proceed with OAuth
+    if not client_secret:
+        logger.warning(
+            "No client secret configured. Token refresh may fail.\n"
+            "Set ANTIGRAVITY_CLIENT_SECRET env var or add "
+            "'antigravity_client_secret' to ~/.hive/configuration.json"
+        )
+
+    # Use fixed port and path matching Google's expected OAuth redirect URI
+    port = _DEFAULT_REDIRECT_PORT
+    redirect_uri = f"http://localhost:{port}/oauth-callback"
+
+    # Generate state for CSRF protection
+    state = secrets.token_urlsafe(16)
+
+    # Build authorization URL
+    params = {
+        "client_id": client_id,
+        "redirect_uri": redirect_uri,
+        "response_type": "code",
+        "scope": " ".join(_OAUTH_SCOPES),
+        "state": state,
+        "access_type": "offline",
+        "prompt": "consent",
+    }
+    auth_url = f"{_OAUTH_AUTH_URL}?{urllib.parse.urlencode(params)}"
+
+    logger.info("Opening browser for authentication...")
+    logger.info(f"If the browser doesn't open, visit: {auth_url}\n")
+
+    # Open browser
+    webbrowser.open(auth_url)
+
+    # Wait for callback
+    logger.info(f"Listening for callback on port {port}...")
+    code, received_state, error = wait_for_callback(port)
+
+    if error:
+        logger.error(f"Authentication failed: {error}")
+        return 1
+
+    if not code:
+        logger.error("No authorization code received")
+        return 1
+
+    if received_state != state:
+        logger.error("State mismatch - possible CSRF attack")
+        return 1
+
+    # Exchange code for tokens
+    logger.info("Exchanging authorization code for tokens...")
+    tokens = exchange_code_for_tokens(code, redirect_uri, client_id, client_secret)
+
+    if not tokens:
+        return 1
+
+    access_token = tokens.get("access_token")
+    refresh_token = tokens.get("refresh_token")
+    expires_in = tokens.get("expires_in", 3600)
+
+    if not access_token:
+        logger.error("No access token in response")
+        return 1
+
+    # Get user email
+    email = get_user_email(access_token)
+    if email:
+        logger.info(f"Authenticated as: {email}")
+
+    # Load existing accounts and add/update
+    accounts_data = load_accounts()
+    accounts = accounts_data.get("accounts", [])
+
+    # Build new account entry (V4 schema)
+    expires_ms = int((time.time() + expires_in) * 1000)
+    refresh_entry = f"{refresh_token}|{_DEFAULT_PROJECT_ID}"
+
+    new_account = {
+        "access": access_token,
+        "refresh": refresh_entry,
+        "expires": expires_ms,
+        "email": email,
+        "enabled": True,
+    }
+
+    # Update existing account or add new one
+    existing_idx = next((i for i, a in enumerate(accounts) if a.get("email") == email), None)
+    if existing_idx is not None:
+        accounts[existing_idx] = new_account
+        logger.info(f"Updated existing account: {email}")
+    else:
+        accounts.append(new_account)
+        logger.info(f"Added new account: {email}")
+
+    accounts_data["accounts"] = accounts
+    accounts_data["schemaVersion"] = 4
+    accounts_data["last_refresh"] = time.strftime("%Y-%m-%dT%H:%M:%SZ", time.gmtime())
+
+    save_accounts(accounts_data)
+    logger.info("\n✓ Authentication complete!")
+    return 0
+
+
+def cmd_account_list(args: argparse.Namespace) -> int:
+    """List all stored accounts."""
+    data = load_accounts()
+    accounts = data.get("accounts", [])
+
+    if not accounts:
+        logger.info("No accounts configured.")
+        logger.info("Run 'antigravity auth account add' to add one.")
+        return 0
+
+    logger.info("Configured accounts:\n")
+    for i, account in enumerate(accounts, 1):
+        email = account.get("email", "unknown")
+        enabled = "enabled" if account.get("enabled", True) else "disabled"
+        logger.info(f"  {i}. {email} ({enabled})")
+
+    return 0
+
+
+def cmd_account_remove(args: argparse.Namespace) -> int:
+    """Remove an account by email."""
+    email = args.email
+    data = load_accounts()
+    accounts = data.get("accounts", [])
+
+    original_len = len(accounts)
+    accounts = [a for a in accounts if a.get("email") != email]
+
+    if len(accounts) == original_len:
+        logger.error(f"No account found with email: {email}")
+        return 1
+
+    data["accounts"] = accounts
+    save_accounts(data)
+    logger.info(f"Removed account: {email}")
+    return 0
+
+
+def main() -> int:
+    parser = argparse.ArgumentParser(
+        description="Antigravity authentication CLI",
+        formatter_class=argparse.RawDescriptionHelpFormatter,
+    )
+    subparsers = parser.add_subparsers(dest="command", help="Commands")
+
+    # auth account add
+    auth_parser = subparsers.add_parser("auth", help="Authentication commands")
+    auth_subparsers = auth_parser.add_subparsers(dest="auth_command")
+
+    account_parser = auth_subparsers.add_parser("account", help="Account management")
+    account_subparsers = account_parser.add_subparsers(dest="account_command")
+
+    add_parser = account_subparsers.add_parser("add", help="Add a new account via OAuth2")
+    add_parser.set_defaults(func=cmd_account_add)
+
+    list_parser = account_subparsers.add_parser("list", help="List configured accounts")
+    list_parser.set_defaults(func=cmd_account_list)
+
+    remove_parser = account_subparsers.add_parser("remove", help="Remove an account")
+    remove_parser.add_argument("email", help="Email of account to remove")
+    remove_parser.set_defaults(func=cmd_account_remove)
+
+    args = parser.parse_args()
+
+    if hasattr(args, "func"):
+        return args.func(args)
+
+    parser.print_help()
+    return 0
+
+
+if __name__ == "__main__":
+    sys.exit(main())

core/codex_oauth.py

@@ -0,0 +1,441 @@
+"""OpenAI Codex OAuth PKCE login flow.
+
+Runs the full browser-based OAuth flow so users can authenticate with their
+ChatGPT Plus/Pro subscription without needing the Codex CLI installed.
+
+Usage (from quickstart.sh):
+    uv run python codex_oauth.py
+
+Exit codes:
+    0 - success (credentials saved to ~/.codex/auth.json)
+    1 - failure (user cancelled, timeout, or token exchange error)
+"""
+
+import base64
+import hashlib
+import http.server
+import json
+import os
+import platform
+import queue
+import secrets
+import subprocess
+import sys
+import threading
+import time
+import urllib.error
+import urllib.parse
+import urllib.request
+from datetime import UTC, datetime
+from pathlib import Path
+from typing import TextIO
+
+# OAuth constants (from the Codex CLI binary)
+CLIENT_ID = "app_EMoamEEZ73f0CkXaXp7hrann"
+AUTHORIZE_URL = "https://auth.openai.com/oauth/authorize"
+TOKEN_URL = "https://auth.openai.com/oauth/token"
+REDIRECT_URI = "http://localhost:1455/auth/callback"
+SCOPE = "openid profile email offline_access"
+CALLBACK_PORT = 1455
+
+# Where to save credentials (same location the Codex CLI uses)
+CODEX_AUTH_FILE = Path.home() / ".codex" / "auth.json"
+
+# JWT claim path for account_id
+JWT_CLAIM_PATH = "https://api.openai.com/auth"
+
+
+def _base64url(data: bytes) -> str:
+    return base64.urlsafe_b64encode(data).rstrip(b"=").decode("ascii")
+
+
+def generate_pkce() -> tuple[str, str]:
+    """Generate PKCE code_verifier and code_challenge (S256)."""
+    verifier_bytes = secrets.token_bytes(32)
+    verifier = _base64url(verifier_bytes)
+    challenge = _base64url(hashlib.sha256(verifier.encode("ascii")).digest())
+    return verifier, challenge
+
+
+def build_authorize_url(state: str, challenge: str) -> str:
+    """Build the OpenAI OAuth authorize URL with PKCE."""
+    params = urllib.parse.urlencode(
+        {
+            "response_type": "code",
+            "client_id": CLIENT_ID,
+            "redirect_uri": REDIRECT_URI,
+            "scope": SCOPE,
+            "code_challenge": challenge,
+            "code_challenge_method": "S256",
+            "state": state,
+            "id_token_add_organizations": "true",
+            "codex_cli_simplified_flow": "true",
+            "originator": "hive",
+        }
+    )
+    return f"{AUTHORIZE_URL}?{params}"
+
+
+def exchange_code_for_tokens(code: str, verifier: str) -> dict | None:
+    """Exchange the authorization code for tokens."""
+    data = urllib.parse.urlencode(
+        {
+            "grant_type": "authorization_code",
+            "client_id": CLIENT_ID,
+            "code": code,
+            "code_verifier": verifier,
+            "redirect_uri": REDIRECT_URI,
+        }
+    ).encode("utf-8")
+
+    req = urllib.request.Request(
+        TOKEN_URL,
+        data=data,
+        headers={"Content-Type": "application/x-www-form-urlencoded"},
+        method="POST",
+    )
+
+    try:
+        with urllib.request.urlopen(req, timeout=15) as resp:
+            token_data = json.loads(resp.read())
+    except (urllib.error.URLError, json.JSONDecodeError, TimeoutError, OSError) as exc:
+        print(f"\033[0;31mToken exchange failed: {exc}\033[0m", file=sys.stderr)
+        return None
+
+    if not token_data.get("access_token") or not token_data.get("refresh_token"):
+        print("\033[0;31mToken response missing required fields\033[0m", file=sys.stderr)
+        return None
+
+    return token_data
+
+
+def decode_jwt_payload(token: str) -> dict | None:
+    """Decode the payload of a JWT (no signature verification)."""
+    try:
+        parts = token.split(".")
+        if len(parts) != 3:
+            return None
+        payload = parts[1]
+        # Add padding
+        padding = 4 - len(payload) % 4
+        if padding != 4:
+            payload += "=" * padding
+        decoded = base64.urlsafe_b64decode(payload)
+        return json.loads(decoded)
+    except Exception:
+        return None
+
+
+def get_account_id(access_token: str) -> str | None:
+    """Extract the ChatGPT account_id from the access token JWT."""
+    payload = decode_jwt_payload(access_token)
+    if not payload:
+        return None
+    auth = payload.get(JWT_CLAIM_PATH)
+    if isinstance(auth, dict):
+        account_id = auth.get("chatgpt_account_id")
+        if isinstance(account_id, str) and account_id:
+            return account_id
+    return None
+
+
+def save_credentials(token_data: dict, account_id: str) -> None:
+    """Save credentials to ~/.codex/auth.json in the same format the Codex CLI uses."""
+    auth_data = {
+        "tokens": {
+            "access_token": token_data["access_token"],
+            "refresh_token": token_data["refresh_token"],
+            "account_id": account_id,
+        },
+        "auth_mode": "chatgpt",
+        "last_refresh": datetime.now(UTC).isoformat(),
+    }
+    if "id_token" in token_data:
+        auth_data["tokens"]["id_token"] = token_data["id_token"]
+
+    CODEX_AUTH_FILE.parent.mkdir(parents=True, exist_ok=True, mode=0o700)
+    fd = os.open(CODEX_AUTH_FILE, os.O_WRONLY | os.O_CREAT | os.O_TRUNC, 0o600)
+    with os.fdopen(fd, "w") as f:
+        json.dump(auth_data, f, indent=2)
+
+
+def open_browser(url: str) -> bool:
+    """Open the URL in the user's default browser."""
+    system = platform.system()
+    try:
+        devnull = subprocess.DEVNULL
+        if system == "Darwin":
+            subprocess.Popen(["open", url], stdout=devnull, stderr=devnull)
+        elif system == "Windows":
+            os.startfile(url)  # type: ignore[attr-defined]
+        else:
+            subprocess.Popen(["xdg-open", url], stdout=devnull, stderr=devnull)
+        return True
+    except (AttributeError, OSError):
+        return False
+
+
+class OAuthCallbackHandler(http.server.BaseHTTPRequestHandler):
+    """HTTP handler that captures the OAuth callback."""
+
+    auth_code: str | None = None
+    received_state: str | None = None
+
+    def do_GET(self) -> None:
+        parsed = urllib.parse.urlparse(self.path)
+        if parsed.path != "/auth/callback":
+            self.send_response(404)
+            self.end_headers()
+            self.wfile.write(b"Not found")
+            return
+
+        params = urllib.parse.parse_qs(parsed.query)
+        code = params.get("code", [None])[0]
+        state = params.get("state", [None])[0]
+
+        if not code:
+            self.send_response(400)
+            self.end_headers()
+            self.wfile.write(b"Missing authorization code")
+            return
+
+        OAuthCallbackHandler.auth_code = code
+        OAuthCallbackHandler.received_state = state
+
+        self.send_response(200)
+        self.send_header("Content-Type", "text/html; charset=utf-8")
+        self.end_headers()
+        self.wfile.write(
+            b"<!doctype html><html><head><meta charset='utf-8'/></head>"
+            b"<body><h2>Authentication successful</h2>"
+            b"<p>Return to your terminal to continue.</p></body></html>"
+        )
+
+    def log_message(self, format: str, *args: object) -> None:
+        # Suppress request logging
+        pass
+
+
+def wait_for_callback(state: str, timeout_secs: int = 120) -> str | None:
+    """Start a local HTTP server and wait for the OAuth callback.
+
+    Returns the authorization code on success, None on timeout.
+    """
+    OAuthCallbackHandler.auth_code = None
+    OAuthCallbackHandler.received_state = None
+
+    server = http.server.HTTPServer(("127.0.0.1", CALLBACK_PORT), OAuthCallbackHandler)
+    server.timeout = 1
+
+    deadline = time.time() + timeout_secs
+    server_thread = threading.Thread(target=_serve_until_done, args=(server, deadline, state))
+    server_thread.daemon = True
+    server_thread.start()
+    server_thread.join(timeout=timeout_secs + 2)
+
+    server.server_close()
+
+    if OAuthCallbackHandler.auth_code and OAuthCallbackHandler.received_state == state:
+        return OAuthCallbackHandler.auth_code
+    return None
+
+
+def _serve_until_done(server: http.server.HTTPServer, deadline: float, state: str) -> None:
+    while time.time() < deadline:
+        server.handle_request()
+        if OAuthCallbackHandler.auth_code and OAuthCallbackHandler.received_state == state:
+            return
+
+
+def parse_manual_input(value: str, expected_state: str) -> str | None:
+    """Parse user-pasted redirect URL or auth code."""
+    value = value.strip()
+    if not value:
+        return None
+    try:
+        parsed = urllib.parse.urlparse(value)
+        params = urllib.parse.parse_qs(parsed.query)
+        code = params.get("code", [None])[0]
+        state = params.get("state", [None])[0]
+        if state and state != expected_state:
+            return None
+        return code
+    except Exception:
+        pass
+    # Maybe it's just the raw code
+    if len(value) > 10 and " " not in value:
+        return value
+    return None
+
+
+def _read_manual_input_lines(
+    manual_inputs: queue.Queue[str],
+    stop_event: threading.Event,
+    stdin: TextIO | None = None,
+) -> None:
+    stream = sys.stdin if stdin is None else stdin
+
+    while not stop_event.is_set():
+        try:
+            manual = stream.readline()
+        except (EOFError, OSError):
+            return
+
+        if not manual:
+            return
+
+        if manual.strip():
+            manual_inputs.put(manual)
+
+
+def wait_for_code_from_callback_or_stdin(
+    expected_state: str,
+    callback_result: list[str | None],
+    callback_done: threading.Event,
+    timeout_secs: float = 120,
+    poll_interval: float = 0.1,
+    stdin: TextIO | None = None,
+) -> str | None:
+    manual_inputs: queue.Queue[str] = queue.Queue()
+    stop_event = threading.Event()
+
+    # Read stdin on a daemon thread so manual paste works on platforms where
+    # select() cannot poll console handles, including Windows terminals.
+    threading.Thread(
+        target=_read_manual_input_lines,
+        args=(manual_inputs, stop_event, stdin),
+        daemon=True,
+    ).start()
+
+    deadline = time.time() + timeout_secs
+    try:
+        while time.time() < deadline:
+            if callback_result[0]:
+                return callback_result[0]
+
+            while True:
+                try:
+                    manual = manual_inputs.get_nowait()
+                except queue.Empty:
+                    break
+
+                code = parse_manual_input(manual, expected_state)
+                if code:
+                    return code
+
+            if callback_done.is_set():
+                return callback_result[0]
+
+            time.sleep(poll_interval)
+
+        return callback_result[0]
+    finally:
+        stop_event.set()
+
+
+def main() -> int:
+    # Generate PKCE and state
+    verifier, challenge = generate_pkce()
+    state = secrets.token_hex(16)
+
+    # Build URL
+    auth_url = build_authorize_url(state, challenge)
+
+    print()
+    print("\033[1mOpenAI Codex OAuth Login\033[0m")
+    print()
+
+    # Try to start the local callback server first
+    try:
+        server_available = True
+        # Quick test that port is free
+        import socket
+
+        sock = socket.socket(socket.AF_INET, socket.SOCK_STREAM)
+        sock.settimeout(1)
+        result = sock.connect_ex(("127.0.0.1", CALLBACK_PORT))
+        sock.close()
+        if result == 0:
+            print(f"\033[1;33mPort {CALLBACK_PORT} is in use. Using manual paste mode.\033[0m")
+            server_available = False
+    except Exception:
+        server_available = True
+
+    # Open browser
+    browser_opened = open_browser(auth_url)
+    if browser_opened:
+        print("  Browser opened for OpenAI sign-in...")
+    else:
+        print("  Could not open browser automatically.")
+
+    print()
+    print("  If the browser didn't open, visit this URL:")
+    print(f"  \033[0;36m{auth_url}\033[0m")
+    print()
+
+    code = None
+
+    if server_available:
+        print("  Waiting for authentication (up to 2 minutes)...")
+        print("  \033[2mOr paste the redirect URL below if the callback didn't work:\033[0m")
+        print()
+
+        # Start callback server in background
+        callback_result: list[str | None] = [None]
+        callback_done = threading.Event()
+
+        def run_server() -> None:
+            try:
+                callback_result[0] = wait_for_callback(state, timeout_secs=120)
+            finally:
+                callback_done.set()
+
+        server_thread = threading.Thread(target=run_server)
+        server_thread.daemon = True
+        server_thread.start()
+
+        try:
+            code = wait_for_code_from_callback_or_stdin(
+                state,
+                callback_result,
+                callback_done,
+                timeout_secs=120,
+            )
+        except KeyboardInterrupt:
+            print("\n\033[0;31mCancelled.\033[0m")
+            return 1
+    else:
+        # Manual paste mode
+        try:
+            manual = input("  Paste the redirect URL: ").strip()
+            code = parse_manual_input(manual, state)
+        except (KeyboardInterrupt, EOFError):
+            print("\n\033[0;31mCancelled.\033[0m")
+            return 1
+
+    if not code:
+        print("\n\033[0;31mAuthentication timed out or failed.\033[0m")
+        return 1
+
+    # Exchange code for tokens
+    print()
+    print("  Exchanging authorization code for tokens...")
+    token_data = exchange_code_for_tokens(code, verifier)
+    if not token_data:
+        return 1
+
+    # Extract account_id from JWT
+    account_id = get_account_id(token_data["access_token"])
+    if not account_id:
+        print("\033[0;31mFailed to extract account ID from token.\033[0m", file=sys.stderr)
+        return 1
+
+    # Save credentials
+    save_credentials(token_data, account_id)
+    print("  \033[0;32mAuthentication successful!\033[0m")
+    print(f"  Credentials saved to {CODEX_AUTH_FILE}")
+    return 0
+
+
+if __name__ == "__main__":
+    sys.exit(main())

core/demos/event_loop_wss_demo.py

@@ -1,779 +0,0 @@
-#!/usr/bin/env python3
-"""
-EventLoopNode WebSocket Demo
-
-Real LLM, real FileConversationStore, real EventBus.
-Streams EventLoopNode execution to a browser via WebSocket.
-
-Usage:
-    cd /home/timothy/oss/hive/core
-    python demos/event_loop_wss_demo.py
-
-    Then open http://localhost:8765 in your browser.
-"""
-
-import asyncio
-import json
-import logging
-import sys
-import tempfile
-from http import HTTPStatus
-from pathlib import Path
-
-import httpx
-import websockets
-from bs4 import BeautifulSoup
-from websockets.http11 import Request, Response
-
-# Add core, tools, and hive root to path
-_CORE_DIR = Path(__file__).resolve().parent.parent
-_HIVE_DIR = _CORE_DIR.parent
-sys.path.insert(0, str(_CORE_DIR))  # framework.*
-sys.path.insert(0, str(_HIVE_DIR / "tools" / "src"))  # aden_tools.*
-sys.path.insert(0, str(_HIVE_DIR))  # core.framework.* (for aden_tools imports)
-
-import os  # noqa: E402
-
-from aden_tools.credentials import CREDENTIAL_SPECS, CredentialStoreAdapter  # noqa: E402
-from core.framework.credentials import CredentialStore  # noqa: E402
-
-from framework.credentials.storage import (  # noqa: E402
-    CompositeStorage,
-    EncryptedFileStorage,
-    EnvVarStorage,
-)
-from framework.graph.event_loop_node import EventLoopNode, JudgeVerdict, LoopConfig  # noqa: E402
-from framework.graph.node import NodeContext, NodeSpec, SharedMemory  # noqa: E402
-from framework.llm.litellm import LiteLLMProvider  # noqa: E402
-from framework.llm.provider import Tool  # noqa: E402
-from framework.runner.tool_registry import ToolRegistry  # noqa: E402
-from framework.runtime.core import Runtime  # noqa: E402
-from framework.runtime.event_bus import EventBus, EventType  # noqa: E402
-from framework.storage.conversation_store import FileConversationStore  # noqa: E402
-
-logging.basicConfig(level=logging.INFO, format="%(asctime)s %(name)s %(message)s")
-logger = logging.getLogger("demo")
-
-# -------------------------------------------------------------------------
-# Persistent state (shared across WebSocket connections)
-# -------------------------------------------------------------------------
-
-STORE_DIR = Path(tempfile.mkdtemp(prefix="hive_demo_"))
-STORE = FileConversationStore(STORE_DIR / "conversation")
-RUNTIME = Runtime(STORE_DIR / "runtime")
-LLM = LiteLLMProvider(model="claude-sonnet-4-5-20250929")
-
-# -------------------------------------------------------------------------
-# Tool Registry — real tools via ToolRegistry (same pattern as GraphExecutor)
-# -------------------------------------------------------------------------
-
-TOOL_REGISTRY = ToolRegistry()
-
-# Credential store: Aden sync (OAuth2 tokens) + encrypted files + env var fallback
-_env_mapping = {name: spec.env_var for name, spec in CREDENTIAL_SPECS.items()}
-_local_storage = CompositeStorage(
-    primary=EncryptedFileStorage(),
-    fallbacks=[EnvVarStorage(env_mapping=_env_mapping)],
-)
-
-if os.environ.get("ADEN_API_KEY"):
-    try:
-        from framework.credentials.aden import (  # noqa: E402
-            AdenCachedStorage,
-            AdenClientConfig,
-            AdenCredentialClient,
-            AdenSyncProvider,
-        )
-
-        _client = AdenCredentialClient(AdenClientConfig(base_url="https://api.adenhq.com"))
-        _provider = AdenSyncProvider(client=_client)
-        _storage = AdenCachedStorage(
-            local_storage=_local_storage,
-            aden_provider=_provider,
-        )
-        _cred_store = CredentialStore(storage=_storage, providers=[_provider], auto_refresh=True)
-        _synced = _provider.sync_all(_cred_store)
-        logger.info("Synced %d credentials from Aden", _synced)
-    except Exception as e:
-        logger.warning("Aden sync unavailable: %s", e)
-        _cred_store = CredentialStore(storage=_local_storage)
-else:
-    logger.info("ADEN_API_KEY not set, using local credential storage")
-    _cred_store = CredentialStore(storage=_local_storage)
-
-CREDENTIALS = CredentialStoreAdapter(_cred_store)
-
-# Debug: log which credentials resolved
-for _name in ["brave_search", "hubspot", "anthropic"]:
-    _val = CREDENTIALS.get(_name)
-    if _val:
-        logger.debug("credential %s: OK (len=%d)", _name, len(_val))
-    else:
-        logger.debug("credential %s: not found", _name)
-
-# --- web_search (Brave Search API) ---
-
-TOOL_REGISTRY.register(
-    name="web_search",
-    tool=Tool(
-        name="web_search",
-        description=(
-            "Search the web for current information. "
-            "Returns titles, URLs, and snippets from search results."
-        ),
-        parameters={
-            "type": "object",
-            "properties": {
-                "query": {
-                    "type": "string",
-                    "description": "The search query (1-500 characters)",
-                },
-                "num_results": {
-                    "type": "integer",
-                    "description": "Number of results to return (1-20, default 10)",
-                },
-            },
-            "required": ["query"],
-        },
-    ),
-    executor=lambda inputs: _exec_web_search(inputs),
-)
-
-
-def _exec_web_search(inputs: dict) -> dict:
-    api_key = CREDENTIALS.get("brave_search")
-    if not api_key:
-        return {"error": "brave_search credential not configured"}
-    query = inputs.get("query", "")
-    num_results = min(inputs.get("num_results", 10), 20)
-    resp = httpx.get(
-        "https://api.search.brave.com/res/v1/web/search",
-        params={"q": query, "count": num_results},
-        headers={"X-Subscription-Token": api_key, "Accept": "application/json"},
-        timeout=30.0,
-    )
-    if resp.status_code != 200:
-        return {"error": f"Brave API HTTP {resp.status_code}"}
-    data = resp.json()
-    results = [
-        {
-            "title": item.get("title", ""),
-            "url": item.get("url", ""),
-            "snippet": item.get("description", ""),
-        }
-        for item in data.get("web", {}).get("results", [])[:num_results]
-    ]
-    return {"query": query, "results": results, "total": len(results)}
-
-
-# --- web_scrape (httpx + BeautifulSoup, no playwright for sync compat) ---
-
-TOOL_REGISTRY.register(
-    name="web_scrape",
-    tool=Tool(
-        name="web_scrape",
-        description=(
-            "Scrape and extract text content from a webpage URL. "
-            "Returns the page title and main text content."
-        ),
-        parameters={
-            "type": "object",
-            "properties": {
-                "url": {
-                    "type": "string",
-                    "description": "URL of the webpage to scrape",
-                },
-                "max_length": {
-                    "type": "integer",
-                    "description": "Maximum text length (default 50000)",
-                },
-            },
-            "required": ["url"],
-        },
-    ),
-    executor=lambda inputs: _exec_web_scrape(inputs),
-)
-
-_SCRAPE_HEADERS = {
-    "User-Agent": (
-        "Mozilla/5.0 (Windows NT 10.0; Win64; x64) "
-        "AppleWebKit/537.36 (KHTML, like Gecko) "
-        "Chrome/131.0.0.0 Safari/537.36"
-    ),
-    "Accept": "text/html,application/xhtml+xml",
-}
-
-
-def _exec_web_scrape(inputs: dict) -> dict:
-    url = inputs.get("url", "")
-    max_length = max(1000, min(inputs.get("max_length", 50000), 500000))
-    if not url.startswith(("http://", "https://")):
-        url = "https://" + url
-    try:
-        resp = httpx.get(url, timeout=30.0, follow_redirects=True, headers=_SCRAPE_HEADERS)
-        if resp.status_code != 200:
-            return {"error": f"HTTP {resp.status_code}"}
-        soup = BeautifulSoup(resp.text, "html.parser")
-        for tag in soup(["script", "style", "nav", "footer", "header", "aside", "noscript"]):
-            tag.decompose()
-        title = soup.title.get_text(strip=True) if soup.title else ""
-        main = (
-            soup.find("article")
-            or soup.find("main")
-            or soup.find(attrs={"role": "main"})
-            or soup.find("body")
-        )
-        text = main.get_text(separator=" ", strip=True) if main else ""
-        text = " ".join(text.split())
-        if len(text) > max_length:
-            text = text[:max_length] + "..."
-        return {"url": url, "title": title, "content": text, "length": len(text)}
-    except httpx.TimeoutException:
-        return {"error": "Request timed out"}
-    except Exception as e:
-        return {"error": f"Scrape failed: {e}"}
-
-
-# --- HubSpot CRM tools (optional, requires HUBSPOT_ACCESS_TOKEN) ---
-
-_HUBSPOT_API = "https://api.hubapi.com"
-
-
-def _hubspot_headers() -> dict | None:
-    token = CREDENTIALS.get("hubspot")
-    if token:
-        logger.debug("HubSpot token: %s...%s (len=%d)", token[:8], token[-4:], len(token))
-    else:
-        logger.debug("HubSpot token: not found")
-    if not token:
-        return None
-    return {
-        "Authorization": f"Bearer {token}",
-        "Content-Type": "application/json",
-        "Accept": "application/json",
-    }
-
-
-def _exec_hubspot_search(inputs: dict) -> dict:
-    headers = _hubspot_headers()
-    if not headers:
-        return {"error": "HUBSPOT_ACCESS_TOKEN not set"}
-    object_type = inputs.get("object_type", "contacts")
-    query = inputs.get("query", "")
-    limit = min(inputs.get("limit", 10), 100)
-    body: dict = {"limit": limit}
-    if query:
-        body["query"] = query
-    try:
-        resp = httpx.post(
-            f"{_HUBSPOT_API}/crm/v3/objects/{object_type}/search",
-            headers=headers,
-            json=body,
-            timeout=30.0,
-        )
-        if resp.status_code != 200:
-            return {"error": f"HubSpot API HTTP {resp.status_code}: {resp.text[:200]}"}
-        return resp.json()
-    except httpx.TimeoutException:
-        return {"error": "Request timed out"}
-    except Exception as e:
-        return {"error": f"HubSpot error: {e}"}
-
-
-TOOL_REGISTRY.register(
-    name="hubspot_search",
-    tool=Tool(
-        name="hubspot_search",
-        description=(
-            "Search HubSpot CRM objects (contacts, companies, or deals). "
-            "Returns matching records with their properties."
-        ),
-        parameters={
-            "type": "object",
-            "properties": {
-                "object_type": {
-                    "type": "string",
-                    "description": "CRM object type: 'contacts', 'companies', or 'deals'",
-                },
-                "query": {
-                    "type": "string",
-                    "description": "Search query (name, email, domain, etc.)",
-                },
-                "limit": {
-                    "type": "integer",
-                    "description": "Max results (1-100, default 10)",
-                },
-            },
-            "required": ["object_type"],
-        },
-    ),
-    executor=lambda inputs: _exec_hubspot_search(inputs),
-)
-
-logger.info(
-    "ToolRegistry loaded: %s",
-    ", ".join(TOOL_REGISTRY.get_registered_names()),
-)
-
-
-# -------------------------------------------------------------------------
-# ChatJudge — keeps the event loop alive between user messages
-# -------------------------------------------------------------------------
-
-
-class ChatJudge:
-    """Judge that blocks between user messages, keeping the loop alive.
-
-    After the LLM finishes responding, the judge awaits a signal indicating
-    a new user message has been injected, then returns RETRY to continue.
-    """
-
-    def __init__(self, on_ready=None):
-        self._message_ready = asyncio.Event()
-        self._shutdown = False
-        self._on_ready = on_ready  # async callback fired when waiting for input
-
-    async def evaluate(self, context: dict) -> JudgeVerdict:
-        # Notify client that the LLM is done — ready for next input
-        if self._on_ready:
-            await self._on_ready()
-
-        # Block until next user message (or shutdown)
-        self._message_ready.clear()
-        await self._message_ready.wait()
-
-        if self._shutdown:
-            return JudgeVerdict(action="ACCEPT")
-
-        return JudgeVerdict(action="RETRY")
-
-    def signal_message(self):
-        """Unblock the judge — a new user message has been injected."""
-        self._message_ready.set()
-
-    def signal_shutdown(self):
-        """Unblock the judge and let the loop exit cleanly."""
-        self._shutdown = True
-        self._message_ready.set()
-
-
-# -------------------------------------------------------------------------
-# HTML page (embedded)
-# -------------------------------------------------------------------------
-
-HTML_PAGE = (  # noqa: E501
-    """<!DOCTYPE html>
-<html lang="en">
-<head>
-<meta charset="utf-8">
-<meta name="viewport" content="width=device-width, initial-scale=1">
-<title>EventLoopNode Live Demo</title>
-<style>
-  * { box-sizing: border-box; margin: 0; padding: 0; }
-  body {
-    font-family: 'SF Mono', 'Fira Code', monospace;
-    background: #0d1117; color: #c9d1d9;
-    height: 100vh; display: flex; flex-direction: column;
-  }
-  header {
-    background: #161b22; padding: 12px 20px;
-    border-bottom: 1px solid #30363d;
-    display: flex; align-items: center; gap: 16px;
-  }
-  header h1 { font-size: 16px; color: #58a6ff; font-weight: 600; }
-  .status {
-    font-size: 12px; padding: 3px 10px; border-radius: 12px;
-    background: #21262d; color: #8b949e;
-  }
-  .status.running { background: #1a4b2e; color: #3fb950; }
-  .status.done { background: #1a3a5c; color: #58a6ff; }
-  .status.error { background: #4b1a1a; color: #f85149; }
-  .chat { flex: 1; overflow-y: auto; padding: 16px; }
-  .msg {
-    margin: 8px 0; padding: 10px 14px; border-radius: 8px;
-    line-height: 1.6; white-space: pre-wrap; word-wrap: break-word;
-  }
-  .msg.user { background: #1a3a5c; color: #58a6ff; }
-  .msg.assistant { background: #161b22; color: #c9d1d9; }
-  .msg.event {
-    background: transparent; color: #8b949e; font-size: 11px;
-    padding: 4px 14px; border-left: 3px solid #30363d;
-  }
-  .msg.event.loop { border-left-color: #58a6ff; }
-  .msg.event.tool { border-left-color: #d29922; }
-  .msg.event.stall { border-left-color: #f85149; }
-  .input-bar {
-    padding: 12px 16px; background: #161b22;
-    border-top: 1px solid #30363d; display: flex; gap: 8px;
-  }
-  .input-bar input {
-    flex: 1; background: #0d1117; border: 1px solid #30363d;
-    color: #c9d1d9; padding: 8px 12px; border-radius: 6px;
-    font-family: inherit; font-size: 14px; outline: none;
-  }
-  .input-bar input:focus { border-color: #58a6ff; }
-  .input-bar button {
-    background: #238636; color: #fff; border: none;
-    padding: 8px 20px; border-radius: 6px; cursor: pointer;
-    font-family: inherit; font-weight: 600;
-  }
-  .input-bar button:hover { background: #2ea043; }
-  .input-bar button:disabled {
-    background: #21262d; color: #484f58; cursor: not-allowed;
-  }
-  .input-bar button.clear { background: #da3633; }
-  .input-bar button.clear:hover { background: #f85149; }
-</style>
-</head>
-<body>
-  <header>
-    <h1>EventLoopNode Live</h1>
-    <span id="status" class="status">Idle</span>
-    <span id="iter" class="status" style="display:none">Step 0</span>
-  </header>
-  <div id="chat" class="chat"></div>
-  <div class="input-bar">
-    <input id="input" type="text"
-           placeholder="Ask anything..." autofocus />
-    <button id="go" onclick="run()">Send</button>
-    <button class="clear"
-            onclick="clearConversation()">Clear</button>
-  </div>
-
-<script>
-let ws = null;
-let currentAssistantEl = null;
-let iterCount = 0;
-const chat = document.getElementById('chat');
-const status = document.getElementById('status');
-const iterEl = document.getElementById('iter');
-const goBtn = document.getElementById('go');
-const inputEl = document.getElementById('input');
-
-inputEl.addEventListener('keydown', e => {
-  if (e.key === 'Enter') run();
-});
-
-function setStatus(text, cls) {
-  status.textContent = text;
-  status.className = 'status ' + cls;
-}
-
-function addMsg(text, cls) {
-  const el = document.createElement('div');
-  el.className = 'msg ' + cls;
-  el.textContent = text;
-  chat.appendChild(el);
-  chat.scrollTop = chat.scrollHeight;
-  return el;
-}
-
-function connect() {
-  ws = new WebSocket('ws://' + location.host + '/ws');
-  ws.onopen = () => {
-    setStatus('Ready', 'done');
-    goBtn.disabled = false;
-  };
-  ws.onmessage = handleEvent;
-  ws.onerror = () => { setStatus('Error', 'error'); };
-  ws.onclose = () => {
-    setStatus('Reconnecting...', '');
-    goBtn.disabled = true;
-    setTimeout(connect, 2000);
-  };
-}
-
-function handleEvent(msg) {
-  const evt = JSON.parse(msg.data);
-
-  if (evt.type === 'llm_text_delta') {
-    if (currentAssistantEl) {
-      currentAssistantEl.textContent += evt.content;
-      chat.scrollTop = chat.scrollHeight;
-    }
-  }
-  else if (evt.type === 'ready') {
-    setStatus('Ready', 'done');
-    if (currentAssistantEl && !currentAssistantEl.textContent)
-      currentAssistantEl.remove();
-    goBtn.disabled = false;
-  }
-  else if (evt.type === 'node_loop_iteration') {
-    iterCount = evt.iteration || (iterCount + 1);
-    iterEl.textContent = 'Step ' + iterCount;
-    iterEl.style.display = '';
-  }
-  else if (evt.type === 'tool_call_started') {
-    var info = evt.tool_name + '('
-      + JSON.stringify(evt.tool_input).slice(0, 120) + ')';
-    addMsg('TOOL  ' + info, 'event tool');
-  }
-  else if (evt.type === 'tool_call_completed') {
-    var preview = (evt.result || '').slice(0, 200);
-    var cls = evt.is_error ? 'stall' : 'tool';
-    addMsg('RESULT  ' + evt.tool_name + ': ' + preview,
-           'event ' + cls);
-    currentAssistantEl = addMsg('', 'assistant');
-  }
-  else if (evt.type === 'result') {
-    setStatus('Session ended', evt.success ? 'done' : 'error');
-    if (evt.error) addMsg('ERROR  ' + evt.error, 'event stall');
-    if (currentAssistantEl && !currentAssistantEl.textContent)
-      currentAssistantEl.remove();
-    goBtn.disabled = false;
-  }
-  else if (evt.type === 'node_stalled') {
-    addMsg('STALLED  ' + evt.reason, 'event stall');
-  }
-  else if (evt.type === 'cleared') {
-    chat.innerHTML = '';
-    iterCount = 0;
-    iterEl.textContent = 'Step 0';
-    iterEl.style.display = 'none';
-    setStatus('Ready', 'done');
-    goBtn.disabled = false;
-  }
-}
-
-function run() {
-  const text = inputEl.value.trim();
-  if (!text || !ws || ws.readyState !== 1) return;
-  addMsg(text, 'user');
-  currentAssistantEl = addMsg('', 'assistant');
-  inputEl.value = '';
-  setStatus('Running', 'running');
-  goBtn.disabled = true;
-  ws.send(JSON.stringify({ topic: text }));
-}
-
-function clearConversation() {
-  if (ws && ws.readyState === 1) {
-    ws.send(JSON.stringify({ command: 'clear' }));
-  }
-}
-
-connect();
-</script>
-</body>
-</html>"""
-)
-
-
-# -------------------------------------------------------------------------
-# WebSocket handler
-# -------------------------------------------------------------------------
-
-
-async def handle_ws(websocket):
-    """Persistent WebSocket: long-lived EventLoopNode kept alive by ChatJudge."""
-    global STORE
-
-    # -- Event forwarding (WebSocket ← EventBus) ----------------------------
-    bus = EventBus()
-
-    async def forward_event(event):
-        try:
-            payload = {"type": event.type.value, **event.data}
-            if event.node_id:
-                payload["node_id"] = event.node_id
-            await websocket.send(json.dumps(payload))
-        except Exception:
-            pass
-
-    bus.subscribe(
-        event_types=[
-            EventType.NODE_LOOP_STARTED,
-            EventType.NODE_LOOP_ITERATION,
-            EventType.NODE_LOOP_COMPLETED,
-            EventType.LLM_TEXT_DELTA,
-            EventType.TOOL_CALL_STARTED,
-            EventType.TOOL_CALL_COMPLETED,
-            EventType.NODE_STALLED,
-        ],
-        handler=forward_event,
-    )
-
-    # -- Ready callback (tells browser the LLM is done, waiting for input) --
-    async def send_ready():
-        try:
-            await websocket.send(json.dumps({"type": "ready"}))
-        except Exception:
-            pass
-
-    # -- Per-connection state -----------------------------------------------
-    judge = ChatJudge(on_ready=send_ready)
-    node = None
-    loop_task = None
-
-    tools = list(TOOL_REGISTRY.get_tools().values())
-    tool_executor = TOOL_REGISTRY.get_executor()
-
-    node_spec = NodeSpec(
-        id="assistant",
-        name="Chat Assistant",
-        description="A conversational assistant that remembers context across messages",
-        node_type="event_loop",
-        system_prompt=(
-            "You are a helpful assistant with access to tools. "
-            "You can search the web, scrape webpages, and query HubSpot CRM. "
-            "Use tools when the user asks for current information or external data. "
-            "You have full conversation history, so you can reference previous messages."
-        ),
-    )
-
-    async def start_loop(first_message: str):
-        """Create an EventLoopNode and run it as a background task."""
-        nonlocal node, loop_task
-
-        memory = SharedMemory()
-        ctx = NodeContext(
-            runtime=RUNTIME,
-            node_id="assistant",
-            node_spec=node_spec,
-            memory=memory,
-            input_data={},
-            llm=LLM,
-            available_tools=tools,
-        )
-        node = EventLoopNode(
-            event_bus=bus,
-            judge=judge,
-            config=LoopConfig(max_iterations=10_000, max_history_tokens=32_000),
-            conversation_store=STORE,
-            tool_executor=tool_executor,
-        )
-        await node.inject_event(first_message)
-
-        async def _run():
-            try:
-                result = await node.execute(ctx)
-                try:
-                    await websocket.send(
-                        json.dumps(
-                            {
-                                "type": "result",
-                                "success": result.success,
-                                "output": result.output,
-                                "error": result.error,
-                                "tokens": result.tokens_used,
-                            }
-                        )
-                    )
-                except Exception:
-                    pass
-                logger.info(f"Loop ended: success={result.success}, tokens={result.tokens_used}")
-            except websockets.exceptions.ConnectionClosed:
-                logger.info("Loop stopped: WebSocket closed")
-            except Exception as e:
-                logger.exception("Loop error")
-                try:
-                    await websocket.send(
-                        json.dumps(
-                            {
-                                "type": "result",
-                                "success": False,
-                                "error": str(e),
-                                "output": {},
-                            }
-                        )
-                    )
-                except Exception:
-                    pass
-
-        loop_task = asyncio.create_task(_run())
-
-    async def stop_loop():
-        """Signal the judge and wait for the loop task to finish."""
-        nonlocal node, loop_task
-        if loop_task and not loop_task.done():
-            judge.signal_shutdown()
-            try:
-                await asyncio.wait_for(loop_task, timeout=5.0)
-            except (TimeoutError, asyncio.CancelledError):
-                loop_task.cancel()
-        node = None
-        loop_task = None
-
-    # -- Message loop (runs for the lifetime of this WebSocket) -------------
-    try:
-        async for raw in websocket:
-            try:
-                msg = json.loads(raw)
-            except Exception:
-                continue
-
-            # Clear command
-            if msg.get("command") == "clear":
-                import shutil
-
-                await stop_loop()
-                await STORE.close()
-                conv_dir = STORE_DIR / "conversation"
-                if conv_dir.exists():
-                    shutil.rmtree(conv_dir)
-                STORE = FileConversationStore(conv_dir)
-                # Reset judge for next session
-                judge = ChatJudge(on_ready=send_ready)
-                await websocket.send(json.dumps({"type": "cleared"}))
-                logger.info("Conversation cleared")
-                continue
-
-            topic = msg.get("topic", "")
-            if not topic:
-                continue
-
-            if node is None:
-                # First message — spin up the loop
-                logger.info(f"Starting persistent loop: {topic}")
-                await start_loop(topic)
-            else:
-                # Subsequent message — inject and unblock the judge
-                logger.info(f"Injecting message: {topic}")
-                await node.inject_event(topic)
-                judge.signal_message()
-
-    except websockets.exceptions.ConnectionClosed:
-        pass
-    finally:
-        await stop_loop()
-        logger.info("WebSocket closed, loop stopped")
-
-
-# -------------------------------------------------------------------------
-# HTTP handler for serving the HTML page
-# -------------------------------------------------------------------------
-
-
-async def process_request(connection, request: Request):
-    """Serve HTML on GET /, upgrade to WebSocket on /ws."""
-    if request.path == "/ws":
-        return None  # let websockets handle the upgrade
-    # Serve the HTML page for any other path
-    return Response(
-        HTTPStatus.OK,
-        "OK",
-        websockets.Headers({"Content-Type": "text/html; charset=utf-8"}),
-        HTML_PAGE.encode(),
-    )
-
-
-# -------------------------------------------------------------------------
-# Main
-# -------------------------------------------------------------------------
-
-
-async def main():
-    port = 8765
-    async with websockets.serve(
-        handle_ws,
-        "0.0.0.0",
-        port,
-        process_request=process_request,
-    ):
-        logger.info(f"Demo running at http://localhost:{port}")
-        logger.info("Open in your browser and enter a topic to research.")
-        await asyncio.Future()  # run forever
-
-
-if __name__ == "__main__":
-    asyncio.run(main())

core/demos/handoff_demo.py

@@ -1,930 +0,0 @@
-#!/usr/bin/env python3
-"""
-Two-Node ContextHandoff Demo
-
-Demonstrates ContextHandoff between two EventLoopNode instances:
-  Node A (Researcher) → ContextHandoff → Node B (Analyst)
-
-Real LLM, real FileConversationStore, real EventBus.
-Streams both nodes to a browser via WebSocket.
-
-Usage:
-    cd /home/timothy/oss/hive/core
-    python demos/handoff_demo.py
-
-    Then open http://localhost:8766 in your browser.
-"""
-
-import asyncio
-import json
-import logging
-import sys
-import tempfile
-from http import HTTPStatus
-from pathlib import Path
-
-import httpx
-import websockets
-from bs4 import BeautifulSoup
-from websockets.http11 import Request, Response
-
-# Add core, tools, and hive root to path
-_CORE_DIR = Path(__file__).resolve().parent.parent
-_HIVE_DIR = _CORE_DIR.parent
-sys.path.insert(0, str(_CORE_DIR))  # framework.*
-sys.path.insert(0, str(_HIVE_DIR / "tools" / "src"))  # aden_tools.*
-sys.path.insert(0, str(_HIVE_DIR))  # core.framework.* (for aden_tools imports)
-
-from aden_tools.credentials import CREDENTIAL_SPECS, CredentialStoreAdapter  # noqa: E402
-from core.framework.credentials import CredentialStore  # noqa: E402
-
-from framework.credentials.storage import (  # noqa: E402
-    CompositeStorage,
-    EncryptedFileStorage,
-    EnvVarStorage,
-)
-from framework.graph.context_handoff import ContextHandoff  # noqa: E402
-from framework.graph.conversation import NodeConversation  # noqa: E402
-from framework.graph.event_loop_node import EventLoopNode, LoopConfig  # noqa: E402
-from framework.graph.node import NodeContext, NodeSpec, SharedMemory  # noqa: E402
-from framework.llm.litellm import LiteLLMProvider  # noqa: E402
-from framework.llm.provider import Tool  # noqa: E402
-from framework.runner.tool_registry import ToolRegistry  # noqa: E402
-from framework.runtime.core import Runtime  # noqa: E402
-from framework.runtime.event_bus import EventBus, EventType  # noqa: E402
-from framework.storage.conversation_store import FileConversationStore  # noqa: E402
-
-logging.basicConfig(level=logging.INFO, format="%(asctime)s %(name)s %(message)s")
-logger = logging.getLogger("handoff_demo")
-
-# -------------------------------------------------------------------------
-# Persistent state
-# -------------------------------------------------------------------------
-
-STORE_DIR = Path(tempfile.mkdtemp(prefix="hive_handoff_"))
-RUNTIME = Runtime(STORE_DIR / "runtime")
-LLM = LiteLLMProvider(model="claude-sonnet-4-5-20250929")
-
-# -------------------------------------------------------------------------
-# Credentials
-# -------------------------------------------------------------------------
-
-# Composite credential store: encrypted files (primary) + env vars (fallback)
-_env_mapping = {name: spec.env_var for name, spec in CREDENTIAL_SPECS.items()}
-_composite = CompositeStorage(
-    primary=EncryptedFileStorage(),
-    fallbacks=[EnvVarStorage(env_mapping=_env_mapping)],
-)
-CREDENTIALS = CredentialStoreAdapter(CredentialStore(storage=_composite))
-
-for _name in ["brave_search", "hubspot"]:
-    _val = CREDENTIALS.get(_name)
-    if _val:
-        logger.debug("credential %s: OK (len=%d)", _name, len(_val))
-    else:
-        logger.debug("credential %s: not found", _name)
-
-# -------------------------------------------------------------------------
-# Tool Registry — web_search + web_scrape for Node A (Researcher)
-# -------------------------------------------------------------------------
-
-TOOL_REGISTRY = ToolRegistry()
-
-
-def _exec_web_search(inputs: dict) -> dict:
-    api_key = CREDENTIALS.get("brave_search")
-    if not api_key:
-        return {"error": "brave_search credential not configured"}
-    query = inputs.get("query", "")
-    num_results = min(inputs.get("num_results", 10), 20)
-    resp = httpx.get(
-        "https://api.search.brave.com/res/v1/web/search",
-        params={"q": query, "count": num_results},
-        headers={
-            "X-Subscription-Token": api_key,
-            "Accept": "application/json",
-        },
-        timeout=30.0,
-    )
-    if resp.status_code != 200:
-        return {"error": f"Brave API HTTP {resp.status_code}"}
-    data = resp.json()
-    results = [
-        {
-            "title": item.get("title", ""),
-            "url": item.get("url", ""),
-            "snippet": item.get("description", ""),
-        }
-        for item in data.get("web", {}).get("results", [])[:num_results]
-    ]
-    return {"query": query, "results": results, "total": len(results)}
-
-
-TOOL_REGISTRY.register(
-    name="web_search",
-    tool=Tool(
-        name="web_search",
-        description=(
-            "Search the web for current information. "
-            "Returns titles, URLs, and snippets from search results."
-        ),
-        parameters={
-            "type": "object",
-            "properties": {
-                "query": {
-                    "type": "string",
-                    "description": "The search query (1-500 characters)",
-                },
-                "num_results": {
-                    "type": "integer",
-                    "description": "Number of results (1-20, default 10)",
-                },
-            },
-            "required": ["query"],
-        },
-    ),
-    executor=lambda inputs: _exec_web_search(inputs),
-)
-
-_SCRAPE_HEADERS = {
-    "User-Agent": (
-        "Mozilla/5.0 (Windows NT 10.0; Win64; x64) "
-        "AppleWebKit/537.36 (KHTML, like Gecko) "
-        "Chrome/131.0.0.0 Safari/537.36"
-    ),
-    "Accept": "text/html,application/xhtml+xml",
-}
-
-
-def _exec_web_scrape(inputs: dict) -> dict:
-    url = inputs.get("url", "")
-    max_length = max(1000, min(inputs.get("max_length", 50000), 500000))
-    if not url.startswith(("http://", "https://")):
-        url = "https://" + url
-    try:
-        resp = httpx.get(
-            url,
-            timeout=30.0,
-            follow_redirects=True,
-            headers=_SCRAPE_HEADERS,
-        )
-        if resp.status_code != 200:
-            return {"error": f"HTTP {resp.status_code}"}
-        soup = BeautifulSoup(resp.text, "html.parser")
-        for tag in soup(["script", "style", "nav", "footer", "header", "aside", "noscript"]):
-            tag.decompose()
-        title = soup.title.get_text(strip=True) if soup.title else ""
-        main = (
-            soup.find("article")
-            or soup.find("main")
-            or soup.find(attrs={"role": "main"})
-            or soup.find("body")
-        )
-        text = main.get_text(separator=" ", strip=True) if main else ""
-        text = " ".join(text.split())
-        if len(text) > max_length:
-            text = text[:max_length] + "..."
-        return {
-            "url": url,
-            "title": title,
-            "content": text,
-            "length": len(text),
-        }
-    except httpx.TimeoutException:
-        return {"error": "Request timed out"}
-    except Exception as e:
-        return {"error": f"Scrape failed: {e}"}
-
-
-TOOL_REGISTRY.register(
-    name="web_scrape",
-    tool=Tool(
-        name="web_scrape",
-        description=(
-            "Scrape and extract text content from a webpage URL. "
-            "Returns the page title and main text content."
-        ),
-        parameters={
-            "type": "object",
-            "properties": {
-                "url": {
-                    "type": "string",
-                    "description": "URL of the webpage to scrape",
-                },
-                "max_length": {
-                    "type": "integer",
-                    "description": "Maximum text length (default 50000)",
-                },
-            },
-            "required": ["url"],
-        },
-    ),
-    executor=lambda inputs: _exec_web_scrape(inputs),
-)
-
-logger.info(
-    "ToolRegistry loaded: %s",
-    ", ".join(TOOL_REGISTRY.get_registered_names()),
-)
-
-# -------------------------------------------------------------------------
-# Node Specs
-# -------------------------------------------------------------------------
-
-RESEARCHER_SPEC = NodeSpec(
-    id="researcher",
-    name="Researcher",
-    description="Researches a topic using web search and scraping tools",
-    node_type="event_loop",
-    input_keys=["topic"],
-    output_keys=["research_summary"],
-    system_prompt=(
-        "You are a thorough research assistant. Your job is to research "
-        "the given topic using the web_search and web_scrape tools.\n\n"
-        "1. Search for relevant information on the topic\n"
-        "2. Scrape 1-2 of the most promising URLs for details\n"
-        "3. Synthesize your findings into a comprehensive summary\n"
-        "4. Use set_output with key='research_summary' to save your "
-        "findings\n\n"
-        "Be thorough but efficient. Aim for 2-4 search/scrape calls, "
-        "then summarize and set_output."
-    ),
-)
-
-ANALYST_SPEC = NodeSpec(
-    id="analyst",
-    name="Analyst",
-    description="Analyzes research findings and provides insights",
-    node_type="event_loop",
-    input_keys=["context"],
-    output_keys=["analysis"],
-    system_prompt=(
-        "You are a strategic analyst. You receive research findings from "
-        "a previous researcher and must:\n\n"
-        "1. Identify key themes and patterns\n"
-        "2. Assess the reliability and significance of the findings\n"
-        "3. Provide actionable insights and recommendations\n"
-        "4. Use set_output with key='analysis' to save your analysis\n\n"
-        "Be concise but insightful. Focus on what matters most."
-    ),
-)
-
-
-# -------------------------------------------------------------------------
-# HTML page
-# -------------------------------------------------------------------------
-
-HTML_PAGE = (  # noqa: E501
-    """<!DOCTYPE html>
-<html lang="en">
-<head>
-<meta charset="utf-8">
-<meta name="viewport" content="width=device-width, initial-scale=1">
-<title>ContextHandoff Demo</title>
-<style>
-  * {
-    box-sizing: border-box;
-    margin: 0;
-    padding: 0;
-  }
-  body {
-    font-family: 'SF Mono', 'Fira Code', monospace;
-    background: #0d1117;
-    color: #c9d1d9;
-    height: 100vh;
-    display: flex;
-    flex-direction: column;
-  }
-  header {
-    background: #161b22;
-    padding: 12px 20px;
-    border-bottom: 1px solid #30363d;
-    display: flex;
-    align-items: center;
-    gap: 16px;
-  }
-  header h1 {
-    font-size: 16px;
-    color: #58a6ff;
-    font-weight: 600;
-  }
-  .badge {
-    font-size: 12px;
-    padding: 3px 10px;
-    border-radius: 12px;
-    background: #21262d;
-    color: #8b949e;
-  }
-  .badge.researcher {
-    background: #1a3a5c;
-    color: #58a6ff;
-  }
-  .badge.analyst {
-    background: #1a4b2e;
-    color: #3fb950;
-  }
-  .badge.handoff {
-    background: #3d1f00;
-    color: #d29922;
-  }
-  .badge.done {
-    background: #21262d;
-    color: #8b949e;
-  }
-  .badge.error {
-    background: #4b1a1a;
-    color: #f85149;
-  }
-  .chat {
-    flex: 1;
-    overflow-y: auto;
-    padding: 16px;
-  }
-  .msg {
-    margin: 8px 0;
-    padding: 10px 14px;
-    border-radius: 8px;
-    line-height: 1.6;
-    white-space: pre-wrap;
-    word-wrap: break-word;
-  }
-  .msg.user {
-    background: #1a3a5c;
-    color: #58a6ff;
-  }
-  .msg.assistant {
-    background: #161b22;
-    color: #c9d1d9;
-  }
-  .msg.assistant.analyst-msg {
-    border-left: 3px solid #3fb950;
-  }
-  .msg.event {
-    background: transparent;
-    color: #8b949e;
-    font-size: 11px;
-    padding: 4px 14px;
-    border-left: 3px solid #30363d;
-  }
-  .msg.event.loop {
-    border-left-color: #58a6ff;
-  }
-  .msg.event.tool {
-    border-left-color: #d29922;
-  }
-  .msg.event.stall {
-    border-left-color: #f85149;
-  }
-  .handoff-banner {
-    margin: 16px 0;
-    padding: 16px;
-    background: #1c1200;
-    border: 1px solid #d29922;
-    border-radius: 8px;
-    text-align: center;
-  }
-  .handoff-banner h3 {
-    color: #d29922;
-    font-size: 14px;
-    margin-bottom: 8px;
-  }
-  .handoff-banner p, .result-banner p {
-    color: #8b949e;
-    font-size: 12px;
-    line-height: 1.5;
-    max-height: 200px;
-    overflow-y: auto;
-    white-space: pre-wrap;
-    text-align: left;
-  }
-  .result-banner {
-    margin: 16px 0;
-    padding: 16px;
-    background: #0a2614;
-    border: 1px solid #3fb950;
-    border-radius: 8px;
-  }
-  .result-banner h3 {
-    color: #3fb950;
-    font-size: 14px;
-    margin-bottom: 8px;
-    text-align: center;
-  }
-  .result-banner .label {
-    color: #58a6ff;
-    font-size: 11px;
-    font-weight: 600;
-    margin-top: 10px;
-    margin-bottom: 2px;
-  }
-  .result-banner .tokens {
-    color: #484f58;
-    font-size: 11px;
-    text-align: center;
-    margin-top: 10px;
-  }
-  .input-bar {
-    padding: 12px 16px;
-    background: #161b22;
-    border-top: 1px solid #30363d;
-    display: flex;
-    gap: 8px;
-  }
-  .input-bar input {
-    flex: 1;
-    background: #0d1117;
-    border: 1px solid #30363d;
-    color: #c9d1d9;
-    padding: 8px 12px;
-    border-radius: 6px;
-    font-family: inherit;
-    font-size: 14px;
-    outline: none;
-  }
-  .input-bar input:focus {
-    border-color: #58a6ff;
-  }
-  .input-bar button {
-    background: #238636;
-    color: #fff;
-    border: none;
-    padding: 8px 20px;
-    border-radius: 6px;
-    cursor: pointer;
-    font-family: inherit;
-    font-weight: 600;
-  }
-  .input-bar button:hover {
-    background: #2ea043;
-  }
-  .input-bar button:disabled {
-    background: #21262d;
-    color: #484f58;
-    cursor: not-allowed;
-  }
-</style>
-</head>
-<body>
-  <header>
-    <h1>ContextHandoff Demo</h1>
-    <span id="phase" class="badge">Idle</span>
-    <span id="iter" class="badge" style="display:none">Step 0</span>
-  </header>
-  <div id="chat" class="chat"></div>
-  <div class="input-bar">
-    <input id="input" type="text"
-           placeholder="Enter a research topic..." autofocus />
-    <button id="go" onclick="run()">Research</button>
-  </div>
-
-<script>
-let ws = null;
-let currentAssistantEl = null;
-let iterCount = 0;
-let currentPhase = 'idle';
-const chat = document.getElementById('chat');
-const phase = document.getElementById('phase');
-const iterEl = document.getElementById('iter');
-const goBtn = document.getElementById('go');
-const inputEl = document.getElementById('input');
-
-inputEl.addEventListener('keydown', e => {
-  if (e.key === 'Enter') run();
-});
-
-function setPhase(text, cls) {
-  phase.textContent = text;
-  phase.className = 'badge ' + cls;
-  currentPhase = cls;
-}
-
-function addMsg(text, cls) {
-  const el = document.createElement('div');
-  el.className = 'msg ' + cls;
-  el.textContent = text;
-  chat.appendChild(el);
-  chat.scrollTop = chat.scrollHeight;
-  return el;
-}
-
-function addHandoffBanner(summary) {
-  const banner = document.createElement('div');
-  banner.className = 'handoff-banner';
-  const h3 = document.createElement('h3');
-  h3.textContent = 'Context Handoff: Researcher -> Analyst';
-  const p = document.createElement('p');
-  p.textContent = summary || 'Passing research context...';
-  banner.appendChild(h3);
-  banner.appendChild(p);
-  chat.appendChild(banner);
-  chat.scrollTop = chat.scrollHeight;
-}
-
-function addResultBanner(researcher, analyst, tokens) {
-  const banner = document.createElement('div');
-  banner.className = 'result-banner';
-  const h3 = document.createElement('h3');
-  h3.textContent = 'Pipeline Complete';
-  banner.appendChild(h3);
-
-  if (researcher && researcher.research_summary) {
-    const lbl = document.createElement('div');
-    lbl.className = 'label';
-    lbl.textContent = 'RESEARCH SUMMARY';
-    banner.appendChild(lbl);
-    const p = document.createElement('p');
-    p.textContent = researcher.research_summary;
-    banner.appendChild(p);
-  }
-
-  if (analyst && analyst.analysis) {
-    const lbl = document.createElement('div');
-    lbl.className = 'label';
-    lbl.textContent = 'ANALYSIS';
-    lbl.style.color = '#3fb950';
-    banner.appendChild(lbl);
-    const p = document.createElement('p');
-    p.textContent = analyst.analysis;
-    banner.appendChild(p);
-  }
-
-  if (tokens) {
-    const t = document.createElement('div');
-    t.className = 'tokens';
-    t.textContent = 'Total tokens: ' + tokens.toLocaleString();
-    banner.appendChild(t);
-  }
-
-  chat.appendChild(banner);
-  chat.scrollTop = chat.scrollHeight;
-}
-
-function connect() {
-  ws = new WebSocket('ws://' + location.host + '/ws');
-  ws.onopen = () => {
-    setPhase('Ready', 'done');
-    goBtn.disabled = false;
-  };
-  ws.onmessage = handleEvent;
-  ws.onerror = () => { setPhase('Error', 'error'); };
-  ws.onclose = () => {
-    setPhase('Reconnecting...', '');
-    goBtn.disabled = true;
-    setTimeout(connect, 2000);
-  };
-}
-
-function handleEvent(msg) {
-  const evt = JSON.parse(msg.data);
-
-  if (evt.type === 'phase') {
-    if (evt.phase === 'researcher') {
-      setPhase('Researcher', 'researcher');
-    } else if (evt.phase === 'handoff') {
-      setPhase('Handoff', 'handoff');
-    } else if (evt.phase === 'analyst') {
-      setPhase('Analyst', 'analyst');
-    }
-    iterCount = 0;
-    iterEl.style.display = 'none';
-  }
-  else if (evt.type === 'llm_text_delta') {
-    if (currentAssistantEl) {
-      currentAssistantEl.textContent += evt.content;
-      chat.scrollTop = chat.scrollHeight;
-    }
-  }
-  else if (evt.type === 'node_loop_iteration') {
-    iterCount = evt.iteration || (iterCount + 1);
-    iterEl.textContent = 'Step ' + iterCount;
-    iterEl.style.display = '';
-  }
-  else if (evt.type === 'tool_call_started') {
-    var info = evt.tool_name + '('
-      + JSON.stringify(evt.tool_input).slice(0, 120) + ')';
-    addMsg('TOOL  ' + info, 'event tool');
-  }
-  else if (evt.type === 'tool_call_completed') {
-    var preview = (evt.result || '').slice(0, 200);
-    var cls = evt.is_error ? 'stall' : 'tool';
-    addMsg(
-      'RESULT  ' + evt.tool_name + ': ' + preview,
-      'event ' + cls
-    );
-    var assistCls = currentPhase === 'analyst'
-      ? 'assistant analyst-msg' : 'assistant';
-    currentAssistantEl = addMsg('', assistCls);
-  }
-  else if (evt.type === 'handoff_context') {
-    addHandoffBanner(evt.summary);
-    var assistCls = 'assistant analyst-msg';
-    currentAssistantEl = addMsg('', assistCls);
-  }
-  else if (evt.type === 'node_result') {
-    if (evt.node_id === 'researcher') {
-      if (currentAssistantEl
-          && !currentAssistantEl.textContent) {
-        currentAssistantEl.remove();
-      }
-    }
-  }
-  else if (evt.type === 'done') {
-    setPhase('Done', 'done');
-    iterEl.style.display = 'none';
-    if (currentAssistantEl
-        && !currentAssistantEl.textContent) {
-      currentAssistantEl.remove();
-    }
-    currentAssistantEl = null;
-    addResultBanner(
-      evt.researcher, evt.analyst, evt.total_tokens
-    );
-    goBtn.disabled = false;
-    inputEl.placeholder = 'Enter another topic...';
-  }
-  else if (evt.type === 'error') {
-    setPhase('Error', 'error');
-    addMsg('ERROR  ' + evt.message, 'event stall');
-    goBtn.disabled = false;
-  }
-  else if (evt.type === 'node_stalled') {
-    addMsg('STALLED  ' + evt.reason, 'event stall');
-  }
-}
-
-function run() {
-  const text = inputEl.value.trim();
-  if (!text || !ws || ws.readyState !== 1) return;
-  chat.innerHTML = '';
-  addMsg(text, 'user');
-  currentAssistantEl = addMsg('', 'assistant');
-  inputEl.value = '';
-  goBtn.disabled = true;
-  ws.send(JSON.stringify({ topic: text }));
-}
-
-connect();
-</script>
-</body>
-</html>"""
-)
-
-
-# -------------------------------------------------------------------------
-# WebSocket handler — sequential Node A → Handoff → Node B
-# -------------------------------------------------------------------------
-
-
-async def handle_ws(websocket):
-    """Run the two-node handoff pipeline per user message."""
-    try:
-        async for raw in websocket:
-            try:
-                msg = json.loads(raw)
-            except Exception:
-                continue
-
-            topic = msg.get("topic", "")
-            if not topic:
-                continue
-
-            logger.info(f"Starting handoff pipeline for: {topic}")
-
-            try:
-                await _run_pipeline(websocket, topic)
-            except websockets.exceptions.ConnectionClosed:
-                logger.info("WebSocket closed during pipeline")
-                return
-            except Exception as e:
-                logger.exception("Pipeline error")
-                try:
-                    await websocket.send(json.dumps({"type": "error", "message": str(e)}))
-                except Exception:
-                    pass
-
-    except websockets.exceptions.ConnectionClosed:
-        pass
-
-
-async def _run_pipeline(websocket, topic: str):
-    """Execute: Node A (research) → ContextHandoff → Node B (analysis)."""
-    import shutil
-
-    # Fresh stores for each run
-    run_dir = Path(tempfile.mkdtemp(prefix="hive_run_", dir=STORE_DIR))
-    store_a = FileConversationStore(run_dir / "node_a")
-    store_b = FileConversationStore(run_dir / "node_b")
-
-    # Shared event bus
-    bus = EventBus()
-
-    async def forward_event(event):
-        try:
-            payload = {"type": event.type.value, **event.data}
-            if event.node_id:
-                payload["node_id"] = event.node_id
-            await websocket.send(json.dumps(payload))
-        except Exception:
-            pass
-
-    bus.subscribe(
-        event_types=[
-            EventType.NODE_LOOP_STARTED,
-            EventType.NODE_LOOP_ITERATION,
-            EventType.NODE_LOOP_COMPLETED,
-            EventType.LLM_TEXT_DELTA,
-            EventType.TOOL_CALL_STARTED,
-            EventType.TOOL_CALL_COMPLETED,
-            EventType.NODE_STALLED,
-        ],
-        handler=forward_event,
-    )
-
-    tools = list(TOOL_REGISTRY.get_tools().values())
-    tool_executor = TOOL_REGISTRY.get_executor()
-
-    # ---- Phase 1: Researcher ------------------------------------------------
-    await websocket.send(json.dumps({"type": "phase", "phase": "researcher"}))
-
-    node_a = EventLoopNode(
-        event_bus=bus,
-        judge=None,  # implicit judge: accept when output_keys filled
-        config=LoopConfig(
-            max_iterations=20,
-            max_tool_calls_per_turn=10,
-            max_history_tokens=32_000,
-        ),
-        conversation_store=store_a,
-        tool_executor=tool_executor,
-    )
-
-    ctx_a = NodeContext(
-        runtime=RUNTIME,
-        node_id="researcher",
-        node_spec=RESEARCHER_SPEC,
-        memory=SharedMemory(),
-        input_data={"topic": topic},
-        llm=LLM,
-        available_tools=tools,
-    )
-
-    result_a = await node_a.execute(ctx_a)
-    logger.info(
-        "Researcher done: success=%s, tokens=%s",
-        result_a.success,
-        result_a.tokens_used,
-    )
-
-    await websocket.send(
-        json.dumps(
-            {
-                "type": "node_result",
-                "node_id": "researcher",
-                "success": result_a.success,
-                "output": result_a.output,
-            }
-        )
-    )
-
-    if not result_a.success:
-        await websocket.send(
-            json.dumps(
-                {
-                    "type": "error",
-                    "message": f"Researcher failed: {result_a.error}",
-                }
-            )
-        )
-        return
-
-    # ---- Phase 2: Context Handoff -------------------------------------------
-    await websocket.send(json.dumps({"type": "phase", "phase": "handoff"}))
-
-    # Restore the researcher's conversation from store
-    conversation_a = await NodeConversation.restore(store_a)
-    if conversation_a is None:
-        await websocket.send(
-            json.dumps(
-                {
-                    "type": "error",
-                    "message": "Failed to restore researcher conversation",
-                }
-            )
-        )
-        return
-
-    handoff_engine = ContextHandoff(llm=LLM)
-    handoff_context = handoff_engine.summarize_conversation(
-        conversation=conversation_a,
-        node_id="researcher",
-        output_keys=["research_summary"],
-    )
-
-    formatted_handoff = ContextHandoff.format_as_input(handoff_context)
-    logger.info(
-        "Handoff: %d turns, ~%d tokens, keys=%s",
-        handoff_context.turn_count,
-        handoff_context.total_tokens_used,
-        list(handoff_context.key_outputs.keys()),
-    )
-
-    # Send handoff context to browser
-    await websocket.send(
-        json.dumps(
-            {
-                "type": "handoff_context",
-                "summary": handoff_context.summary[:500],
-                "turn_count": handoff_context.turn_count,
-                "tokens": handoff_context.total_tokens_used,
-                "key_outputs": handoff_context.key_outputs,
-            }
-        )
-    )
-
-    # ---- Phase 3: Analyst ---------------------------------------------------
-    await websocket.send(json.dumps({"type": "phase", "phase": "analyst"}))
-
-    node_b = EventLoopNode(
-        event_bus=bus,
-        judge=None,  # implicit judge
-        config=LoopConfig(
-            max_iterations=10,
-            max_tool_calls_per_turn=5,
-            max_history_tokens=32_000,
-        ),
-        conversation_store=store_b,
-    )
-
-    ctx_b = NodeContext(
-        runtime=RUNTIME,
-        node_id="analyst",
-        node_spec=ANALYST_SPEC,
-        memory=SharedMemory(),
-        input_data={"context": formatted_handoff},
-        llm=LLM,
-        available_tools=[],
-    )
-
-    result_b = await node_b.execute(ctx_b)
-    logger.info(
-        "Analyst done: success=%s, tokens=%s",
-        result_b.success,
-        result_b.tokens_used,
-    )
-
-    # ---- Done ---------------------------------------------------------------
-    await websocket.send(
-        json.dumps(
-            {
-                "type": "done",
-                "researcher": result_a.output,
-                "analyst": result_b.output,
-                "total_tokens": ((result_a.tokens_used or 0) + (result_b.tokens_used or 0)),
-            }
-        )
-    )
-
-    # Clean up temp stores
-    try:
-        shutil.rmtree(run_dir)
-    except Exception:
-        pass
-
-
-# -------------------------------------------------------------------------
-# HTTP handler
-# -------------------------------------------------------------------------
-
-
-async def process_request(connection, request: Request):
-    """Serve HTML on GET /, upgrade to WebSocket on /ws."""
-    if request.path == "/ws":
-        return None
-    return Response(
-        HTTPStatus.OK,
-        "OK",
-        websockets.Headers({"Content-Type": "text/html; charset=utf-8"}),
-        HTML_PAGE.encode(),
-    )
-
-
-# -------------------------------------------------------------------------
-# Main
-# -------------------------------------------------------------------------
-
-
-async def main():
-    port = 8766
-    async with websockets.serve(
-        handle_ws,
-        "0.0.0.0",
-        port,
-        process_request=process_request,
-    ):
-        logger.info(f"Handoff demo at http://localhost:{port}")
-        logger.info("Enter a research topic to start the pipeline.")
-        await asyncio.Future()
-
-
-if __name__ == "__main__":
-    asyncio.run(main())

core/examples/manual_agent.py

@@ -4,34 +4,45 @@ Minimal Manual Agent Example
 This example demonstrates how to build and run an agent programmatically
 without using the Claude Code CLI or external LLM APIs.
 
-It uses 'function' nodes to define logic in pure Python, making it perfect
-for understanding the core runtime loop:
+It uses custom NodeProtocol implementations to define logic in pure Python,
+making it perfect for understanding the core runtime loop:
 Setup -> Graph definition -> Execution -> Result
 
 Run with:
-    PYTHONPATH=core python core/examples/manual_agent.py
+    uv run python core/examples/manual_agent.py
 """
 
 import asyncio
 
 from framework.graph import EdgeCondition, EdgeSpec, Goal, GraphSpec, NodeSpec
 from framework.graph.executor import GraphExecutor
+from framework.graph.node import NodeContext, NodeProtocol, NodeResult
 from framework.runtime.core import Runtime
 
 
-# 1. Define Node Logic (Pure Python Functions)
-def greet(name: str) -> str:
+# 1. Define Node Logic (Custom NodeProtocol implementations)
+class GreeterNode(NodeProtocol):
     """Generate a simple greeting."""
-    return f"Hello, {name}!"
+
+    async def execute(self, ctx: NodeContext) -> NodeResult:
+        name = ctx.input_data.get("name", "World")
+        greeting = f"Hello, {name}!"
+        ctx.buffer.write("greeting", greeting)
+        return NodeResult(success=True, output={"greeting": greeting})
 
 
-def uppercase(greeting: str) -> str:
+class UppercaserNode(NodeProtocol):
     """Convert text to uppercase."""
-    return greeting.upper()
+
+    async def execute(self, ctx: NodeContext) -> NodeResult:
+        greeting = ctx.input_data.get("greeting") or ctx.buffer.read("greeting") or ""
+        result = greeting.upper()
+        ctx.buffer.write("final_greeting", result)
+        return NodeResult(success=True, output={"final_greeting": result})
 
 
 async def main():
-    print("🚀 Setting up Manual Agent...")
+    print("Setting up Manual Agent...")
 
     # 2. Define the Goal
     # Every agent needs a goal with success criteria
@@ -55,8 +66,7 @@ async def main():
         id="greeter",
         name="Greeter",
         description="Generates a simple greeting",
-        node_type="function",
-        function="greet",  # Matches the registered function name
+        node_type="event_loop",
         input_keys=["name"],
         output_keys=["greeting"],
     )
@@ -65,8 +75,7 @@ async def main():
         id="uppercaser",
         name="Uppercaser",
         description="Converts greeting to uppercase",
-        node_type="function",
-        function="uppercase",
+        node_type="event_loop",
         input_keys=["greeting"],
         output_keys=["final_greeting"],
     )
@@ -98,23 +107,23 @@ async def main():
     runtime = Runtime(storage_path=Path("./agent_logs"))
     executor = GraphExecutor(runtime=runtime)
 
-    # 7. Register Function Implementations
-    # Connect string names in NodeSpecs to actual Python functions
-    executor.register_function("greeter", greet)
-    executor.register_function("uppercaser", uppercase)
+    # 7. Register Node Implementations
+    # Connect node IDs in the graph to actual Python implementations
+    executor.register_node("greeter", GreeterNode())
+    executor.register_node("uppercaser", UppercaserNode())
 
     # 8. Execute Agent
-    print("▶ Executing agent with input: name='Alice'...")
+    print("Executing agent with input: name='Alice'...")
 
     result = await executor.execute(graph=graph, goal=goal, input_data={"name": "Alice"})
 
     # 9. Verify Results
     if result.success:
-        print("\n✅ Success!")
+        print("\nSuccess!")
         print(f"Path taken: {' -> '.join(result.path)}")
         print(f"Final output: {result.output.get('final_greeting')}")
     else:
-        print(f"\n❌ Failed: {result.error}")
+        print(f"\nFailed: {result.error}")
 
 
 if __name__ == "__main__":

core/examples/mcp_integration_example.py

@@ -79,7 +79,7 @@ async def example_3_config_file():
     # Copy example config (in practice, you'd place this in your agent folder)
     import shutil
 
-    shutil.copy("examples/mcp_servers.json", test_agent_path / "mcp_servers.json")
+    shutil.copy(Path(__file__).parent / "mcp_servers.json", test_agent_path / "mcp_servers.json")
 
     # Load agent - MCP servers will be auto-discovered
     runner = AgentRunner.load(test_agent_path)
@@ -95,81 +95,6 @@ async def example_3_config_file():
     (test_agent_path / "mcp_servers.json").unlink()
 
 
-async def example_4_custom_agent_with_mcp_tools():
-    """Example 4: Build custom agent that uses MCP tools"""
-    print("\n=== Example 4: Custom Agent with MCP Tools ===\n")
-
-    from framework.builder.workflow import GraphBuilder
-
-    # Create a workflow builder
-    builder = GraphBuilder()
-
-    # Define goal
-    builder.set_goal(
-        goal_id="web-researcher",
-        name="Web Research Agent",
-        description="Search the web and summarize findings",
-    )
-
-    # Add success criteria
-    builder.add_success_criterion(
-        "search-results", "Successfully retrieve at least 3 web search results"
-    )
-    builder.add_success_criterion("summary", "Provide a clear, concise summary of the findings")
-
-    # Add nodes that will use MCP tools
-    builder.add_node(
-        node_id="web-searcher",
-        name="Web Search",
-        description="Search the web for information",
-        node_type="llm_tool_use",
-        system_prompt="Search for {query} and return the top results. Use the web_search tool.",
-        tools=["web_search"],  # This tool comes from tools MCP server
-        input_keys=["query"],
-        output_keys=["search_results"],
-    )
-
-    builder.add_node(
-        node_id="summarizer",
-        name="Summarize Results",
-        description="Summarize the search results",
-        node_type="llm_generate",
-        system_prompt="Summarize the following search results in 2-3 sentences: {search_results}",
-        input_keys=["search_results"],
-        output_keys=["summary"],
-    )
-
-    # Connect nodes
-    builder.add_edge("web-searcher", "summarizer")
-
-    # Set entry point
-    builder.set_entry("web-searcher")
-    builder.set_terminal("summarizer")
-
-    # Export the agent
-    export_path = Path("exports/web-research-agent")
-    export_path.mkdir(parents=True, exist_ok=True)
-    builder.export(export_path)
-
-    # Load and register MCP server
-    runner = AgentRunner.load(export_path)
-    runner.register_mcp_server(
-        name="tools",
-        transport="stdio",
-        command="python",
-        args=["-m", "aden_tools.mcp_server", "--stdio"],
-        cwd="../tools",
-    )
-
-    # Run the agent
-    result = await runner.run({"query": "latest AI breakthroughs 2026"})
-
-    print(f"\nAgent completed with result:\n{result}")
-
-    # Cleanup
-    runner.cleanup()
-
-
 async def main():
     """Run all examples"""
     print("=" * 60)

core/examples/mcp_servers.json

@@ -4,8 +4,8 @@
       "name": "tools",
       "description": "Aden tools including web search, file operations, and PDF reading",
       "transport": "stdio",
-      "command": "python",
-      "args": ["mcp_server.py", "--stdio"],
+      "command": "uv",
+      "args": ["run", "python", "mcp_server.py", "--stdio"],
       "cwd": "../tools",
       "env": {
         "BRAVE_SEARCH_API_KEY": "${BRAVE_SEARCH_API_KEY}"

core/framework/__init__.py

@@ -22,9 +22,8 @@ The framework includes a Goal-Based Testing system (Goal → Agent → Eval):
 See `framework.testing` for details.
 """
 
-from framework.builder.query import BuilderQuery
 from framework.llm import AnthropicProvider, LLMProvider
-from framework.runner import AgentOrchestrator, AgentRunner
+from framework.runner import AgentRunner
 from framework.runtime.core import Runtime
 from framework.schemas.decision import Decision, DecisionEvaluation, Option, Outcome
 from framework.schemas.run import Problem, Run, RunSummary
@@ -51,14 +50,11 @@ __all__ = [
     "Problem",
     # Runtime
     "Runtime",
-    # Builder
-    "BuilderQuery",
     # LLM
     "LLMProvider",
     "AnthropicProvider",
     # Runner
     "AgentRunner",
-    "AgentOrchestrator",
     # Testing
     "Test",
     "TestResult",

core/framework/agents/__init__.py

@@ -0,0 +1,13 @@
+"""Framework-provided agents."""
+
+from pathlib import Path
+
+FRAMEWORK_AGENTS_DIR = Path(__file__).parent
+
+
+def list_framework_agents() -> list[Path]:
+    """List all framework agent directories."""
+    return sorted(
+        [p for p in FRAMEWORK_AGENTS_DIR.iterdir() if p.is_dir() and (p / "agent.py").exists()],
+        key=lambda p: p.name,
+    )

core/framework/agents/credential_tester/__init__.py

@@ -0,0 +1,55 @@
+"""
+Credential Tester — verify credentials (Aden OAuth + local API keys) via live API calls.
+
+Interactive agent that lists all testable accounts, lets the user pick one,
+loads the provider's tools, and runs a chat session to test the credential.
+"""
+
+from .agent import (
+    CredentialTesterAgent,
+    _list_aden_accounts,
+    _list_env_fallback_accounts,
+    _list_local_accounts,
+    configure_for_account,
+    conversation_mode,
+    edges,
+    entry_node,
+    entry_points,
+    get_tools_for_provider,
+    goal,
+    identity_prompt,
+    list_connected_accounts,
+    loop_config,
+    nodes,
+    pause_nodes,
+    requires_account_selection,
+    skip_credential_validation,
+    terminal_nodes,
+)
+from .config import default_config
+
+__version__ = "1.0.0"
+
+__all__ = [
+    "CredentialTesterAgent",
+    "configure_for_account",
+    "conversation_mode",
+    "default_config",
+    "edges",
+    "entry_node",
+    "entry_points",
+    "get_tools_for_provider",
+    "goal",
+    "identity_prompt",
+    "list_connected_accounts",
+    "loop_config",
+    "nodes",
+    "pause_nodes",
+    "requires_account_selection",
+    "skip_credential_validation",
+    "terminal_nodes",
+    # Internal list helpers (exposed for testing)
+    "_list_aden_accounts",
+    "_list_local_accounts",
+    "_list_env_fallback_accounts",
+]

core/framework/agents/credential_tester/__main__.py

@@ -0,0 +1,111 @@
+"""CLI entry point for Credential Tester agent."""
+
+import asyncio
+
+import click
+
+from .agent import CredentialTesterAgent
+
+
+def setup_logging(verbose=False, debug=False):
+    from framework.observability import configure_logging
+
+    if debug:
+        configure_logging(level="DEBUG")
+    elif verbose:
+        configure_logging(level="INFO")
+    else:
+        configure_logging(level="WARNING")
+
+
+def pick_account(agent: CredentialTesterAgent) -> dict | None:
+    """Interactive account picker. Returns selected account dict or None."""
+    accounts = agent.list_accounts()
+    if not accounts:
+        click.echo("No connected accounts found.")
+        click.echo("Set ADEN_API_KEY and connect accounts at https://app.adenhq.com")
+        return None
+
+    click.echo("\nConnected accounts:\n")
+    for i, acct in enumerate(accounts, 1):
+        provider = acct.get("provider", "?")
+        alias = acct.get("alias", "?")
+        identity = acct.get("identity", {})
+        detail_parts = [f"{k}: {v}" for k, v in identity.items() if v]
+        detail = f"  ({', '.join(detail_parts)})" if detail_parts else ""
+        click.echo(f"  {i}. {provider}/{alias}{detail}")
+
+    click.echo()
+    while True:
+        choice = click.prompt("Pick an account to test", type=int, default=1)
+        if 1 <= choice <= len(accounts):
+            return accounts[choice - 1]
+        click.echo(f"Invalid choice. Enter 1-{len(accounts)}.")
+
+
+@click.group()
+@click.version_option(version="1.0.0")
+def cli():
+    """Credential Tester — verify synced credentials via live API calls."""
+    pass
+
+
+@cli.command()
+@click.option("--verbose", "-v", is_flag=True)
+@click.option("--debug", is_flag=True)
+def shell(verbose, debug):
+    """Interactive CLI session to test a credential."""
+    setup_logging(verbose=verbose, debug=debug)
+    asyncio.run(_interactive_shell(verbose))
+
+
+async def _interactive_shell(verbose=False):
+    agent = CredentialTesterAgent()
+    account = pick_account(agent)
+    if account is None:
+        return
+
+    agent.select_account(account)
+    provider = account.get("provider", "?")
+    alias = account.get("alias", "?")
+
+    click.echo(f"\nTesting {provider}/{alias}")
+    click.echo("Type your requests or 'quit' to exit.\n")
+
+    await agent.start()
+
+    try:
+        result = await agent._agent_runtime.trigger_and_wait(
+            entry_point_id="start",
+            input_data={},
+        )
+        if result:
+            click.echo(f"\nSession ended: {'success' if result.success else result.error}")
+    except KeyboardInterrupt:
+        click.echo("\nGoodbye!")
+    finally:
+        await agent.stop()
+
+
+@cli.command(name="list")
+def list_accounts():
+    """List all connected accounts."""
+    agent = CredentialTesterAgent()
+    accounts = agent.list_accounts()
+
+    if not accounts:
+        click.echo("No connected accounts found.")
+        return
+
+    click.echo("\nConnected accounts:\n")
+    for acct in accounts:
+        provider = acct.get("provider", "?")
+        alias = acct.get("alias", "?")
+        identity = acct.get("identity", {})
+        detail_parts = [f"{k}: {v}" for k, v in identity.items() if v]
+        detail = f"  ({', '.join(detail_parts)})" if detail_parts else ""
+        click.echo(f"  {provider}/{alias}{detail}")
+
+
+if __name__ == "__main__":
+    cli()

core/framework/agents/credential_tester/agent.py

@@ -0,0 +1,659 @@
+"""Credential Tester agent — verify credentials via live API calls.
+
+Supports both Aden OAuth2-synced accounts AND locally-stored API key accounts.
+Aden accounts use account="alias" routing; local accounts inject the key into
+the session environment so tools read it without an account= parameter.
+
+When loaded via AgentRunner.load() (TUI picker, ``hive run``), the module-level
+``nodes`` / ``edges`` variables provide a static graph.  The TUI detects
+``requires_account_selection`` and shows an account picker *before* starting
+the agent.  ``configure_for_account()`` then scopes the node's tools to the
+selected provider.
+
+When used directly (``CredentialTesterAgent``), the graph is built dynamically
+after the user picks an account programmatically.
+"""
+
+from __future__ import annotations
+
+import logging
+from pathlib import Path
+from typing import TYPE_CHECKING
+
+from framework.config import get_max_context_tokens
+from framework.graph import Goal, NodeSpec, SuccessCriterion
+from framework.graph.checkpoint_config import CheckpointConfig
+from framework.graph.edge import GraphSpec
+from framework.graph.executor import ExecutionResult
+from framework.llm import LiteLLMProvider
+from framework.runner.mcp_registry import MCPRegistry
+from framework.runner.tool_registry import ToolRegistry
+from framework.runtime.agent_runtime import AgentRuntime, create_agent_runtime
+from framework.runtime.execution_stream import EntryPointSpec
+
+from .config import default_config
+from .nodes import build_tester_node
+
+logger = logging.getLogger(__name__)
+
+if TYPE_CHECKING:
+    from framework.runner import AgentRunner
+
+logger = logging.getLogger(__name__)
+
+# ---------------------------------------------------------------------------
+# Goal
+# ---------------------------------------------------------------------------
+
+goal = Goal(
+    id="credential-tester",
+    name="Credential Tester",
+    description="Verify that a credential can make real API calls.",
+    success_criteria=[
+        SuccessCriterion(
+            id="api-call-success",
+            description="At least one API call succeeds using the credential",
+            metric="api_call_success",
+            target="true",
+            weight=1.0,
+        ),
+    ],
+    constraints=[],
+)
+
+# ---------------------------------------------------------------------------
+# Helpers
+# ---------------------------------------------------------------------------
+
+
+def get_tools_for_provider(provider_name: str) -> list[str]:
+    """Collect tool names for a credential by credential_id OR credential_group.
+
+    Matches on both ``credential_id`` (e.g. "google" → Gmail tools) and
+    ``credential_group`` (e.g. "google_custom_search" → all google search tools).
+    """
+    from aden_tools.credentials import CREDENTIAL_SPECS
+
+    tools: list[str] = []
+    for spec in CREDENTIAL_SPECS.values():
+        if spec.credential_id == provider_name or spec.credential_group == provider_name:
+            tools.extend(spec.tools)
+    return sorted(set(tools))
+
+
+def _list_aden_accounts() -> list[dict]:
+    """List active accounts from the Aden platform (requires ADEN_API_KEY)."""
+    import os
+
+    api_key = os.environ.get("ADEN_API_KEY")
+    if not api_key:
+        return []
+
+    try:
+        from framework.credentials.aden.client import AdenClientConfig, AdenCredentialClient
+
+        client = AdenCredentialClient(
+            AdenClientConfig(
+                base_url=os.environ.get("ADEN_API_URL", "https://api.adenhq.com"),
+            )
+        )
+        try:
+            integrations = client.list_integrations()
+        finally:
+            client.close()
+
+        return [
+            {
+                "provider": c.provider,
+                "alias": c.alias,
+                "identity": {"email": c.email} if c.email else {},
+                "integration_id": c.integration_id,
+                "source": "aden",
+            }
+            for c in integrations
+            if c.status == "active"
+        ]
+    except (ImportError, OSError) as exc:
+        logger.debug("Could not list Aden accounts: %s", exc)
+        return []
+    except Exception:
+        logger.warning("Unexpected error listing Aden accounts", exc_info=True)
+        return []
+
+
+def _list_local_accounts() -> list[dict]:
+    """List named local API key accounts from LocalCredentialRegistry."""
+    try:
+        from framework.credentials.local.registry import LocalCredentialRegistry
+
+        return [
+            info.to_account_dict() for info in LocalCredentialRegistry.default().list_accounts()
+        ]
+    except ImportError as exc:
+        logger.debug("Local credential registry unavailable: %s", exc)
+        return []
+    except Exception:
+        logger.warning("Unexpected error listing local accounts", exc_info=True)
+        return []
+
+
+def _list_env_fallback_accounts() -> list[dict]:
+    """Surface configured-but-unregistered credentials as testable entries.
+
+    Detects credentials available via env vars OR stored in the encrypted
+    store in the old flat format (e.g. ``brave_search`` with no alias).
+    These are users who haven't yet run ``save_account()`` but have a working key.
+    Shows with alias="default" and status="unknown".
+    """
+    import os
+
+    from aden_tools.credentials import CREDENTIAL_SPECS
+
+    # Collect IDs in encrypted store (includes old flat entries like "brave_search")
+    try:
+        from framework.credentials.storage import EncryptedFileStorage
+
+        encrypted_ids: set[str] = set(EncryptedFileStorage().list_all())
+    except (ImportError, OSError) as exc:
+        logger.debug("Could not read encrypted store: %s", exc)
+        encrypted_ids = set()
+    except Exception:
+        logger.warning("Unexpected error reading encrypted store", exc_info=True)
+        encrypted_ids = set()
+
+    def _is_configured(cred_name: str, spec) -> bool:
+        # 1. Env var present
+        if os.environ.get(spec.env_var):
+            return True
+        # 2. Old flat encrypted entry (no slash — new entries have {x}/{y})
+        if cred_name in encrypted_ids:
+            return True
+        return False
+
+    seen_groups: set[str] = set()
+    accounts: list[dict] = []
+
+    for cred_name, spec in CREDENTIAL_SPECS.items():
+        if not spec.direct_api_key_supported or not spec.tools:
+            continue
+
+        if spec.credential_group:
+            if spec.credential_group in seen_groups:
+                continue
+            group_available = all(
+                _is_configured(n, s)
+                for n, s in CREDENTIAL_SPECS.items()
+                if s.credential_group == spec.credential_group
+            )
+            if not group_available:
+                continue
+            seen_groups.add(spec.credential_group)
+            provider = spec.credential_group
+        else:
+            if not _is_configured(cred_name, spec):
+                continue
+            provider = cred_name
+
+        accounts.append(
+            {
+                "provider": provider,
+                "alias": "default",
+                "identity": {},
+                "integration_id": None,
+                "source": "local",
+                "status": "unknown",
+            }
+        )
+
+    return accounts
+
+
+def list_connected_accounts() -> list[dict]:
+    """List all testable accounts: Aden-synced + named local + env-var fallbacks."""
+    aden = _list_aden_accounts()
+    local = _list_local_accounts()
+
+    # Show env-var fallbacks only for credentials not already in the named registry
+    local_providers = {a["provider"] for a in local}
+    env_fallbacks = [
+        a for a in _list_env_fallback_accounts() if a["provider"] not in local_providers
+    ]
+
+    return aden + local + env_fallbacks
+
+
+# ---------------------------------------------------------------------------
+# Module-level hooks (read by AgentRunner.load / TUI)
+# ---------------------------------------------------------------------------
+
+skip_credential_validation = True
+"""Don't validate credentials at load time — we don't know which provider yet."""
+
+requires_account_selection = True
+"""Signal TUI to show account picker before starting the agent."""
+
+
+def configure_for_account(runner: AgentRunner, account: dict) -> None:
+    """Scope the tester node's tools to the selected provider.
+
+    Handles both Aden accounts (account= routing) and local accounts
+    (session-level env var injection, no account= parameter in prompt).
+    """
+    provider = account["provider"]
+    source = account.get("source", "aden")
+    alias = account.get("alias", "unknown")
+    identity = account.get("identity", {})
+    tools = get_tools_for_provider(provider)
+
+    if source == "aden":
+        tools.append("get_account_info")
+        email = identity.get("email", "")
+        detail = f" (email: {email})" if email else ""
+        _configure_aden_node(runner, provider, alias, detail, tools)
+    else:
+        status = account.get("status", "unknown")
+        _activate_local_account(provider, alias)
+        _configure_local_node(runner, provider, alias, identity, tools, status)
+
+
+def _activate_local_account(credential_id: str, alias: str) -> None:
+    """Inject a named local account's key into the session environment.
+
+    Handles three cases:
+    1. Named account in LocalCredentialRegistry (new format: {credential_id}/{alias})
+    2. Old flat credential in EncryptedFileStorage (id == credential_id, no alias)
+    3. Env var already set — skip injection (nothing to do)
+    """
+    import os
+
+    from aden_tools.credentials import CREDENTIAL_SPECS
+
+    # Collect specs for this credential (handles grouped credentials too)
+    group_specs = [
+        (cred_name, spec)
+        for cred_name, spec in CREDENTIAL_SPECS.items()
+        if spec.credential_group == credential_id
+        or spec.credential_id == credential_id
+        or cred_name == credential_id
+    ]
+    # Deduplicate — credential_id and credential_group may both match the same spec
+    seen_env_vars: set[str] = set()
+
+    try:
+        from framework.credentials.local.registry import LocalCredentialRegistry
+        from framework.credentials.storage import EncryptedFileStorage
+
+        registry = LocalCredentialRegistry.default()
+        flat_storage = EncryptedFileStorage()
+
+        for _cred_name, spec in group_specs:
+            if spec.env_var in seen_env_vars:
+                continue
+            # If env var is already set, nothing to do for this one
+            if os.environ.get(spec.env_var):
+                seen_env_vars.add(spec.env_var)
+                continue
+
+            seen_env_vars.add(spec.env_var)
+
+            # Determine key name based on spec
+            key_name = "api_key"
+            if spec.credential_group and "cse" in spec.env_var.lower():
+                key_name = "cse_id"
+
+            key: str | None = None
+
+            # 1. Try named account in registry (new format)
+            if alias != "default":
+                key = registry.get_key(credential_id, alias, key_name)
+            else:
+                # For "default" alias, check registry first, then fall back to flat store
+                key = registry.get_key(credential_id, "default", key_name)
+
+            # 2. Fall back to old flat encrypted entry (id == credential_id, no alias)
+            if key is None:
+                flat_cred = flat_storage.load(credential_id)
+                if flat_cred is not None:
+                    key = flat_cred.get_key(key_name) or flat_cred.get_default_key()
+
+            if key:
+                os.environ[spec.env_var] = key
+    except (ImportError, KeyError, OSError) as exc:
+        logger.debug("Could not inject credentials: %s", exc)
+    except Exception:
+        logger.warning("Unexpected error injecting credentials", exc_info=True)
+
+
+def _configure_aden_node(
+    runner: AgentRunner,
+    provider: str,
+    alias: str,
+    detail: str,
+    tools: list[str],
+) -> None:
+    for node in runner.graph.nodes:
+        if node.id == "tester":
+            node.tools = sorted(set(tools))
+            node.system_prompt = f"""\
+You are a credential tester for the account: {provider}/{alias}{detail}
+
+# Instructions
+
+1. Suggest a simple read-only API call to verify the credential works \
+(e.g. list messages, list channels, list contacts).
+2. Execute the call when the user agrees.
+3. Report the result: success (with sample data) or failure (with error).
+4. Let the user request additional API calls to further test the credential.
+
+# Account routing
+
+IMPORTANT: Always pass `account="{alias}"` when calling any tool. \
+This routes the API call to the correct credential. Never use the email \
+or any other identifier — always use the alias exactly as shown.
+
+# Rules
+
+- Start with read-only operations (list, get) before write operations.
+- Always confirm with the user before performing write operations.
+- If a call fails, report the exact error — this helps diagnose credential issues.
+- Be concise. No emojis.
+"""
+            break
+
+    runner.intro_message = (
+        f"Testing {provider}/{alias}{detail} — "
+        f"{len(tools)} tools loaded. "
+        "I'll suggest a read-only API call to verify the credential works."
+    )
+
+
+def _configure_local_node(
+    runner: AgentRunner,
+    provider: str,
+    alias: str,
+    identity: dict,
+    tools: list[str],
+    status: str,
+) -> None:
+    identity_parts = [f"{k}: {v}" for k, v in identity.items() if v]
+    detail = f" ({', '.join(identity_parts)})" if identity_parts else ""
+    status_note = " [key not yet validated]" if status == "unknown" else ""
+
+    for node in runner.graph.nodes:
+        if node.id == "tester":
+            node.tools = sorted(set(tools))
+            node.system_prompt = f"""\
+You are a credential tester for the local API key: {provider}/{alias}{detail}{status_note}
+
+# Instructions
+
+1. Suggest a simple test call to verify the credential works \
+(e.g. search for "test", list items, get profile info).
+2. Execute the call when the user agrees.
+3. Report the result: success (with sample data) or failure (with error).
+4. Let the user request additional API calls to further test the credential.
+
+# Rules
+
+- Do NOT pass an `account` parameter — this credential is injected \
+directly into the session environment and tools read it automatically.
+- Start with read-only operations before write operations.
+- Always confirm with the user before performing write operations.
+- If a call fails, report the exact error — this helps diagnose credential issues.
+- Be concise. No emojis.
+"""
+            break
+
+    runner.intro_message = (
+        f"Testing {provider}/{alias}{detail} — "
+        f"{len(tools)} tools loaded. "
+        "I'll suggest a test API call to verify the credential works."
+    )
+
+
+# ---------------------------------------------------------------------------
+# Module-level graph variables (read by AgentRunner.load)
+# ---------------------------------------------------------------------------
+
+nodes = [
+    NodeSpec(
+        id="tester",
+        name="Credential Tester",
+        description=(
+            "Interactive credential testing — lets the user pick an account "
+            "and verify it via API calls."
+        ),
+        node_type="event_loop",
+        client_facing=True,
+        max_node_visits=0,
+        input_keys=[],
+        output_keys=["test_result"],
+        nullable_output_keys=["test_result"],
+        tools=["get_account_info"],
+        system_prompt="""\
+You are a credential tester. Your job is to help the user verify that their \
+connected accounts and API keys can make real API calls.
+
+# Startup
+
+1. Call ``get_account_info`` to list the user's connected accounts.
+2. Present the list and ask the user which account to test.
+3. Once they pick one, note the account's **alias** (e.g. "Timothy", "work-slack").
+4. Suggest a simple read-only API call to verify the credential works \
+(e.g. list messages, list channels, list contacts).
+5. Execute the call when the user agrees.
+6. Report the result: success (with sample data) or failure (with error).
+7. Let the user request additional API calls to further test the credential.
+
+# Account routing (Aden accounts only)
+
+IMPORTANT: For Aden-synced accounts, always pass the account's **alias** as the \
+``account`` parameter when calling any tool. For local API key accounts, do NOT \
+pass an account parameter — they are pre-injected into the session.
+
+# Rules
+
+- Start with read-only operations (list, get) before write operations.
+- Always confirm with the user before performing write operations.
+- If a call fails, report the exact error — this helps diagnose credential issues.
+- Be concise. No emojis.
+""",
+    ),
+]
+
+edges = []
+
+entry_node = "tester"
+entry_points = {"start": "tester"}
+pause_nodes = []
+terminal_nodes = ["tester"]  # Tester node can terminate
+
+conversation_mode = "continuous"
+identity_prompt = (
+    "You are a credential tester that verifies connected accounts and API keys "
+    "can make real API calls."
+)
+loop_config = {
+    "max_iterations": 50,
+    "max_tool_calls_per_turn": 30,
+}
+
+# ---------------------------------------------------------------------------
+# Programmatic agent class (used by __main__.py CLI)
+# ---------------------------------------------------------------------------
+
+
+class CredentialTesterAgent:
+    """Interactive agent that tests a specific credential via API calls.
+
+    Usage:
+        agent = CredentialTesterAgent()
+        accounts = agent.list_accounts()
+        agent.select_account(accounts[0])
+        await agent.start()
+        await agent.stop()
+    """
+
+    def __init__(self, config=None):
+        self.config = config or default_config
+        self._selected_account: dict | None = None
+        self._agent_runtime: AgentRuntime | None = None
+        self._tool_registry: ToolRegistry | None = None
+        self._storage_path: Path | None = None
+
+    def list_accounts(self) -> list[dict]:
+        """List all testable accounts (Aden + local named + env-var fallbacks)."""
+        return list_connected_accounts()
+
+    def select_account(self, account: dict) -> None:
+        """Select an account to test.
+
+        Args:
+            account: Account dict from list_accounts() with
+                     provider, alias, identity, source keys.
+        """
+        self._selected_account = account
+
+    @property
+    def selected_provider(self) -> str:
+        if self._selected_account is None:
+            raise RuntimeError("No account selected. Call select_account() first.")
+        return self._selected_account["provider"]
+
+    @property
+    def selected_alias(self) -> str:
+        if self._selected_account is None:
+            raise RuntimeError("No account selected. Call select_account() first.")
+        return self._selected_account.get("alias", "unknown")
+
+    def _build_graph(self) -> GraphSpec:
+        provider = self.selected_provider
+        alias = self.selected_alias
+        source = self._selected_account.get("source", "aden")
+        identity = self._selected_account.get("identity", {})
+        tools = get_tools_for_provider(provider)
+
+        if source == "local":
+            _activate_local_account(provider, alias)
+        elif source == "aden":
+            tools.append("get_account_info")
+
+        tester_node = build_tester_node(
+            provider=provider,
+            alias=alias,
+            tools=tools,
+            identity=identity,
+            source=source,
+        )
+
+        return GraphSpec(
+            id="credential-tester-graph",
+            goal_id=goal.id,
+            version="1.0.0",
+            entry_node="tester",
+            entry_points={"start": "tester"},
+            terminal_nodes=["tester"],  # Tester node can terminate
+            pause_nodes=[],
+            nodes=[tester_node],
+            edges=[],
+            default_model=self.config.model,
+            max_tokens=self.config.max_tokens,
+            loop_config={
+                "max_iterations": 50,
+                "max_tool_calls_per_turn": 30,
+                "max_context_tokens": get_max_context_tokens(),
+            },
+            conversation_mode="continuous",
+            identity_prompt=(
+                f"You are testing the {provider}/{alias} credential. "
+                "Help the user verify it works by making real API calls."
+            ),
+        )
+
+    def _setup(self) -> None:
+        if self._selected_account is None:
+            raise RuntimeError("No account selected. Call select_account() first.")
+
+        self._storage_path = Path.home() / ".hive" / "agents" / "credential_tester"
+        self._storage_path.mkdir(parents=True, exist_ok=True)
+
+        self._tool_registry = ToolRegistry()
+
+        mcp_config_path = Path(__file__).parent / "mcp_servers.json"
+        if mcp_config_path.exists():
+            self._tool_registry.load_mcp_config(mcp_config_path)
+
+        try:
+            agent_dir = Path(__file__).parent
+            registry = MCPRegistry()
+            registry.initialize()
+            if (agent_dir / "mcp_registry.json").is_file():
+                self._tool_registry.set_mcp_registry_agent_path(agent_dir)
+            registry_configs, selection_max_tools = registry.load_agent_selection(agent_dir)
+            if registry_configs:
+                self._tool_registry.load_registry_servers(
+                    registry_configs,
+                    preserve_existing_tools=True,
+                    log_collisions=True,
+                    max_tools=selection_max_tools,
+                )
+        except Exception:
+            logger.warning("MCP registry config failed to load", exc_info=True)
+
+        extra_kwargs = getattr(self.config, "extra_kwargs", {}) or {}
+        llm = LiteLLMProvider(
+            model=self.config.model,
+            api_key=self.config.api_key,
+            api_base=self.config.api_base,
+            **extra_kwargs,
+        )
+
+        tool_executor = self._tool_registry.get_executor()
+        tools = list(self._tool_registry.get_tools().values())
+
+        graph = self._build_graph()
+
+        self._agent_runtime = create_agent_runtime(
+            graph=graph,
+            goal=goal,
+            storage_path=self._storage_path,
+            entry_points=[
+                EntryPointSpec(
+                    id="start",
+                    name="Test Credential",
+                    entry_node="tester",
+                    trigger_type="manual",
+                    isolation_level="isolated",
+                ),
+            ],
+            llm=llm,
+            tools=tools,
+            tool_executor=tool_executor,
+            checkpoint_config=CheckpointConfig(enabled=False),
+            graph_id="credential_tester",
+        )
+
+    async def start(self) -> None:
+        """Set up and start the agent runtime."""
+        if self._agent_runtime is None:
+            self._setup()
+        if not self._agent_runtime.is_running:
+            await self._agent_runtime.start()
+
+    async def stop(self) -> None:
+        """Stop the agent runtime."""
+        if self._agent_runtime and self._agent_runtime.is_running:
+            await self._agent_runtime.stop()
+        self._agent_runtime = None
+
+    async def run(self) -> ExecutionResult:
+        """Run the agent (convenience for single execution)."""
+        await self.start()
+        try:
+            result = await self._agent_runtime.trigger_and_wait(
+                entry_point_id="start",
+                input_data={},
+            )
+            return result or ExecutionResult(success=False, error="Execution timeout")
+        finally:
+            await self.stop()

core/framework/agents/credential_tester/config.py

@@ -0,0 +1,19 @@
+"""Runtime configuration for Credential Tester agent."""
+
+from dataclasses import dataclass
+
+from framework.config import RuntimeConfig
+
+
+@dataclass
+class AgentMetadata:
+    name: str = "Credential Tester"
+    version: str = "1.0.0"
+    description: str = (
+        "Test connected accounts by making real API calls. "
+        "Pick an account, verify credentials work, and explore available tools."
+    )
+
+
+metadata = AgentMetadata()
+default_config = RuntimeConfig(temperature=0.3)

core/framework/agents/credential_tester/mcp_servers.json

@@ -0,0 +1,9 @@
+{
+  "hive-tools": {
+    "transport": "stdio",
+    "command": "uv",
+    "args": ["run", "python", "mcp_server.py", "--stdio"],
+    "cwd": "../../../../tools",
+    "description": "Hive tools MCP server with provider-specific tools"
+  }
+}

core/framework/agents/credential_tester/nodes/__init__.py

@@ -0,0 +1,85 @@
+"""Node definitions for Credential Tester agent."""
+
+from framework.graph import NodeSpec
+
+
+def build_tester_node(
+    provider: str,
+    alias: str,
+    tools: list[str],
+    identity: dict[str, str],
+    source: str = "aden",
+) -> NodeSpec:
+    """Build the tester node dynamically for the selected account.
+
+    Args:
+        provider: Provider / credential name (e.g. "google", "brave_search").
+        alias: User-set alias (e.g. "Timothy", "work").
+        tools: Tool names available for this provider.
+        identity: Identity dict (email, workspace, etc.) for context.
+        source: "aden" or "local" — controls routing instructions in the prompt.
+    """
+    detail_parts = [f"{k}: {v}" for k, v in identity.items() if v]
+    detail = f" ({', '.join(detail_parts)})" if detail_parts else ""
+
+    if source == "aden":
+        routing_section = f"""\
+# Account routing
+
+IMPORTANT: Always pass `account="{alias}"` when calling any tool. \
+This routes the API call to the correct credential. Never use the email \
+or any other identifier — always use the alias exactly as shown.
+"""
+    else:
+        routing_section = """\
+# Credential routing
+
+This is a local API key credential — do NOT pass an `account` parameter. \
+The key is pre-injected into the session environment and tools read it automatically.
+"""
+
+    account_label = "account" if source == "aden" else "local API key"
+
+    return NodeSpec(
+        id="tester",
+        name="Credential Tester",
+        description=(
+            f"Interactive testing node for {provider}/{alias}. "
+            f"Has access to all {provider} tools to verify the credential works."
+        ),
+        node_type="event_loop",
+        client_facing=True,
+        max_node_visits=0,
+        input_keys=[],
+        output_keys=["test_result"],
+        nullable_output_keys=["test_result"],
+        tools=tools,
+        system_prompt=f"""\
+You are a credential tester for the {account_label}: {provider}/{alias}{detail}
+
+Your job is to help the user verify that this credential works by making \
+real API calls using the available tools.
+
+{routing_section}
+# Instructions
+
+1. Start by greeting the user and confirming which account you're testing.
+2. Suggest a simple, safe, read-only API call to verify the credential works \
+(e.g. list messages, list channels, list contacts, search for "test").
+3. Execute the call when the user agrees.
+4. Report the result clearly: success (with sample data) or failure (with error).
+5. Let the user request additional API calls to further test the credential.
+
+# Available tools
+
+You have access to {len(tools)} tools for {provider}:
+{chr(10).join(f"- {t}" for t in tools)}
+
+# Rules
+
+- Start with read-only operations (list, get) before write operations (create, update, delete).
+- Always confirm with the user before performing write operations.
+- If a call fails, report the exact error — this helps diagnose credential issues.
+- Be concise. No emojis.
+""",
+    )

core/framework/agents/discovery.py

@@ -0,0 +1,209 @@
+"""Agent discovery — scan known directories and return categorised AgentEntry lists."""
+
+from __future__ import annotations
+
+import json
+from dataclasses import dataclass, field
+from pathlib import Path
+
+
+@dataclass
+class AgentEntry:
+    """Lightweight agent metadata for the picker / API discover endpoint."""
+
+    path: Path
+    name: str
+    description: str
+    category: str
+    session_count: int = 0
+    run_count: int = 0
+    node_count: int = 0
+    tool_count: int = 0
+    tags: list[str] = field(default_factory=list)
+    last_active: str | None = None
+
+
+def _get_last_active(agent_path: Path) -> str | None:
+    """Return the most recent updated_at timestamp across all sessions.
+
+    Checks both worker sessions (``~/.hive/agents/{name}/sessions/``) and
+    queen sessions (``~/.hive/queen/session/``) whose ``meta.json`` references
+    the same *agent_path*.
+    """
+    from datetime import datetime
+
+    agent_name = agent_path.name
+    latest: str | None = None
+
+    # 1. Worker sessions
+    sessions_dir = Path.home() / ".hive" / "agents" / agent_name / "sessions"
+    if sessions_dir.exists():
+        for session_dir in sessions_dir.iterdir():
+            if not session_dir.is_dir() or not session_dir.name.startswith("session_"):
+                continue
+            state_file = session_dir / "state.json"
+            if not state_file.exists():
+                continue
+            try:
+                data = json.loads(state_file.read_text(encoding="utf-8"))
+                ts = data.get("timestamps", {}).get("updated_at")
+                if ts and (latest is None or ts > latest):
+                    latest = ts
+            except Exception:
+                continue
+
+    # 2. Queen sessions
+    queen_sessions_dir = Path.home() / ".hive" / "queen" / "session"
+    if queen_sessions_dir.exists():
+        resolved = agent_path.resolve()
+        for d in queen_sessions_dir.iterdir():
+            if not d.is_dir():
+                continue
+            meta_file = d / "meta.json"
+            if not meta_file.exists():
+                continue
+            try:
+                meta = json.loads(meta_file.read_text(encoding="utf-8"))
+                stored = meta.get("agent_path")
+                if not stored or Path(stored).resolve() != resolved:
+                    continue
+                ts = datetime.fromtimestamp(d.stat().st_mtime).isoformat()
+                if latest is None or ts > latest:
+                    latest = ts
+            except Exception:
+                continue
+
+    return latest
+
+
+def _count_sessions(agent_name: str) -> int:
+    """Count session directories under ~/.hive/agents/{agent_name}/sessions/."""
+    sessions_dir = Path.home() / ".hive" / "agents" / agent_name / "sessions"
+    if not sessions_dir.exists():
+        return 0
+    return sum(1 for d in sessions_dir.iterdir() if d.is_dir() and d.name.startswith("session_"))
+
+
+def _count_runs(agent_name: str) -> int:
+    """Count unique run_ids across all sessions for an agent."""
+    sessions_dir = Path.home() / ".hive" / "agents" / agent_name / "sessions"
+    if not sessions_dir.exists():
+        return 0
+    run_ids: set[str] = set()
+    for session_dir in sessions_dir.iterdir():
+        if not session_dir.is_dir() or not session_dir.name.startswith("session_"):
+            continue
+        # runs.jsonl lives inside workspace subdirectories
+        for runs_file in session_dir.rglob("runs.jsonl"):
+            try:
+                for line in runs_file.read_text(encoding="utf-8").splitlines():
+                    line = line.strip()
+                    if not line:
+                        continue
+                    record = json.loads(line)
+                    rid = record.get("run_id")
+                    if rid:
+                        run_ids.add(rid)
+            except Exception:
+                continue
+    return len(run_ids)
+
+
+def _extract_agent_stats(agent_path: Path) -> tuple[int, int, list[str]]:
+    """Extract node count, tool count, and tags from an agent directory.
+
+    Prefers agent.py (AST-parsed) over agent.json for node/tool counts
+    since agent.json may be stale.  Tags are only available from agent.json.
+    """
+    import ast
+
+    node_count, tool_count, tags = 0, 0, []
+
+    agent_py = agent_path / "agent.py"
+    if agent_py.exists():
+        try:
+            tree = ast.parse(agent_py.read_text(encoding="utf-8"))
+            for node in ast.walk(tree):
+                if isinstance(node, ast.Assign):
+                    for target in node.targets:
+                        if isinstance(target, ast.Name) and target.id == "nodes":
+                            if isinstance(node.value, ast.List):
+                                node_count = len(node.value.elts)
+        except Exception:
+            pass
+
+    agent_json = agent_path / "agent.json"
+    if agent_json.exists():
+        try:
+            data = json.loads(agent_json.read_text(encoding="utf-8"))
+            json_nodes = data.get("graph", {}).get("nodes", []) or data.get("nodes", [])
+            if node_count == 0:
+                node_count = len(json_nodes)
+            tools: set[str] = set()
+            for n in json_nodes:
+                tools.update(n.get("tools", []))
+            tool_count = len(tools)
+            tags = data.get("agent", {}).get("tags", [])
+        except Exception:
+            pass
+
+    return node_count, tool_count, tags
+
+
+def discover_agents() -> dict[str, list[AgentEntry]]:
+    """Discover agents from all known sources grouped by category."""
+    from framework.runner.cli import (
+        _extract_python_agent_metadata,
+        _get_framework_agents_dir,
+        _is_valid_agent_dir,
+    )
+
+    groups: dict[str, list[AgentEntry]] = {}
+    sources = [
+        ("Your Agents", Path("exports")),
+        ("Framework", _get_framework_agents_dir()),
+        ("Examples", Path("examples/templates")),
+    ]
+
+    for category, base_dir in sources:
+        if not base_dir.exists():
+            continue
+        entries: list[AgentEntry] = []
+        for path in sorted(base_dir.iterdir(), key=lambda p: p.name):
+            if not _is_valid_agent_dir(path):
+                continue
+
+            name, desc = _extract_python_agent_metadata(path)
+            config_fallback_name = path.name.replace("_", " ").title()
+            used_config = name != config_fallback_name
+
+            node_count, tool_count, tags = _extract_agent_stats(path)
+            if not used_config:
+                agent_json = path / "agent.json"
+                if agent_json.exists():
+                    try:
+                        data = json.loads(agent_json.read_text(encoding="utf-8"))
+                        meta = data.get("agent", {})
+                        name = meta.get("name", name)
+                        desc = meta.get("description", desc)
+                    except Exception:
+                        pass
+
+            entries.append(
+                AgentEntry(
+                    path=path,
+                    name=name,
+                    description=desc,
+                    category=category,
+                    session_count=_count_sessions(path.name),
+                    run_count=_count_runs(path.name),
+                    node_count=node_count,
+                    tool_count=tool_count,
+                    tags=tags,
+                    last_active=_get_last_active(path),
+                )
+            )
+        if entries:
+            groups[category] = entries
+
+    return groups

core/framework/agents/queen/__init__.py

@@ -0,0 +1,21 @@
+"""
+Queen — Native agent builder for the Hive framework.
+
+Deeply understands the agent framework and produces complete Python packages
+with goals, nodes, edges, system prompts, MCP configuration, and tests
+from natural language specifications.
+"""
+
+from .agent import queen_goal, queen_graph
+from .config import AgentMetadata, RuntimeConfig, default_config, metadata
+
+__version__ = "1.0.0"
+
+__all__ = [
+    "queen_goal",
+    "queen_graph",
+    "RuntimeConfig",
+    "AgentMetadata",
+    "default_config",
+    "metadata",
+]

core/framework/agents/queen/agent.py

@@ -0,0 +1,38 @@
+"""Queen graph definition."""
+
+from framework.graph import Goal
+from framework.graph.edge import GraphSpec
+
+from .nodes import queen_node
+
+# ---------------------------------------------------------------------------
+# Queen graph — the primary persistent conversation.
+# Loaded by queen_orchestrator.create_queen(), NOT by AgentRunner.
+# ---------------------------------------------------------------------------
+
+queen_goal = Goal(
+    id="queen-manager",
+    name="Queen Manager",
+    description=(
+        "Manage the worker agent lifecycle and serve as the user's primary interactive interface."
+    ),
+    success_criteria=[],
+    constraints=[],
+)
+
+queen_graph = GraphSpec(
+    id="queen-graph",
+    goal_id=queen_goal.id,
+    version="1.0.0",
+    entry_node="queen",
+    entry_points={"start": "queen"},
+    terminal_nodes=[],
+    pause_nodes=[],
+    nodes=[queen_node],
+    edges=[],
+    conversation_mode="continuous",
+    loop_config={
+        "max_iterations": 999_999,
+        "max_tool_calls_per_turn": 30,
+    },
+)

core/framework/agents/queen/config.py

@@ -0,0 +1,51 @@
+"""Runtime configuration for Queen agent."""
+
+import json
+from dataclasses import dataclass, field
+from pathlib import Path
+
+
+def _load_preferred_model() -> str:
+    """Load preferred model from ~/.hive/configuration.json."""
+    config_path = Path.home() / ".hive" / "configuration.json"
+    if config_path.exists():
+        try:
+            with open(config_path, encoding="utf-8") as f:
+                config = json.load(f)
+            llm = config.get("llm", {})
+            if llm.get("provider") and llm.get("model"):
+                return f"{llm['provider']}/{llm['model']}"
+        except Exception:
+            pass
+    return "anthropic/claude-sonnet-4-20250514"
+
+
+@dataclass
+class RuntimeConfig:
+    model: str = field(default_factory=_load_preferred_model)
+    temperature: float = 0.7
+    max_tokens: int = 8000
+    api_key: str | None = None
+    api_base: str | None = None
+
+
+default_config = RuntimeConfig()
+
+
+@dataclass
+class AgentMetadata:
+    name: str = "Queen"
+    version: str = "1.0.0"
+    description: str = (
+        "Native coding agent that builds production-ready Hive agent packages "
+        "from natural language specifications. Deeply understands the agent framework "
+        "and produces complete Python packages with goals, nodes, edges, system prompts, "
+        "MCP configuration, and tests."
+    )
+    intro_message: str = (
+        "I'm Queen — I build Hive agents. Describe what kind of agent "
+        "you want to create and I'll design, implement, and validate it for you."
+    )
+
+
+metadata = AgentMetadata()

core/framework/agents/queen/mcp_servers.json

@@ -0,0 +1,9 @@
+{
+  "coder-tools": {
+    "transport": "stdio",
+    "command": "uv",
+    "args": ["run", "python", "coder_tools_server.py", "--stdio"],
+    "cwd": "../../../../tools",
+    "description": "Unsandboxed file system tools for code generation and validation"
+  }
+}

core/framework/agents/queen/nodes/thinking_hook.py

@@ -0,0 +1,138 @@
+"""Queen thinking hook — persona + communication style classifier.
+
+Fires once when the queen enters building mode at session start.
+Makes a single non-streaming LLM call (acting as an HR Director) to select
+the best-fit expert persona for the user's request AND classify the user's
+communication style, then returns a PersonaResult containing both.
+
+This is designed to activate the model's latent domain expertise — a CFO
+persona on a financial question, a Lawyer on a legal question, etc. — while
+also adapting the Queen's communication approach to the individual user.
+"""
+
+from __future__ import annotations
+
+import json
+import logging
+from dataclasses import dataclass
+from typing import TYPE_CHECKING
+
+if TYPE_CHECKING:
+    from framework.llm.provider import LLMProvider
+
+logger = logging.getLogger(__name__)
+
+_HR_SYSTEM_PROMPT = """\
+You are an expert HR Director and communication consultant at a world-class firm.
+A new request has arrived. You must:
+1. Identify which professional role best serves this request.
+2. Read the user's signals to determine HOW to communicate with them.
+
+For communication style, look for:
+- Technical depth: Do they use precise terms? Do they ask "how" or "what"?
+- Pace: Short messages = fast and direct. Long explanations = exploratory.
+- Tone: Are they casual ("hey, can you...") or formal ("I need a system that...")?
+
+If cross-session memory is provided, factor in what is already known about this \
+person — don't rediscover what's already understood.
+
+Reply with ONLY a valid JSON object — no markdown, no prose, no explanation:
+{"role": "<job title>", "persona": "<2-3 sentence first-person identity statement>", \
+"style": "<one of: peer-technical, mentor-guiding, consultant-structured>"}
+
+Rules:
+- Choose from any real professional role: CFO, CEO, CTO, Lawyer, Data Scientist,
+  Product Manager, Security Engineer, DevOps Engineer, Software Architect,
+  HR Director, Marketing Director, Business Analyst, UX Designer,
+  Financial Analyst, Operations Director, Legal Counsel, etc.
+- The persona statement must be written in first person ("I am..." or "I have...").
+- Select the role whose domain knowledge most directly applies to solving the request.
+- If the request is clearly about coding or building software systems, pick Software Architect.
+- "Queen" is your internal alias — do not include it in the persona.
+- For style: "peer-technical" for users who demonstrate domain expertise, \
+"mentor-guiding" for users who are learning or exploring, \
+"consultant-structured" for users who want structured, accountable delivery.
+- Default to "peer-technical" if signals are ambiguous.
+"""
+
+# Communication style directives injected into the Queen's system prompt.
+_STYLE_DIRECTIVES: dict[str, str] = {
+    "peer-technical": (
+        "## Communication Style: Peer\n\n"
+        "This person is technical. Use precise language, skip high-level "
+        "overviews they already know, and get into specifics quickly. "
+        "When they push back on a design choice, engage with the technical "
+        "argument directly."
+    ),
+    "mentor-guiding": (
+        "## Communication Style: Guide\n\n"
+        "This person is learning or exploring. Explain your reasoning as you "
+        "go — not patronizingly, but so they can follow the logic. When you "
+        "make a design choice, briefly say why. Offer to go deeper on anything."
+    ),
+    "consultant-structured": (
+        "## Communication Style: Structured\n\n"
+        "This person wants structured, accountable delivery. Lead with "
+        "summaries and options. Number your proposals. Be explicit about "
+        "trade-offs. Avoid open-ended questions — give them choices to react to."
+    ),
+}
+
+
+@dataclass
+class PersonaResult:
+    """Result of persona + style classification."""
+
+    persona_prefix: str  # e.g. "You are a CFO. I am a CFO with 20 years..."
+    style_directive: str  # e.g. "## Communication Style: Peer\n\n..."
+
+
+async def select_expert_persona(
+    user_message: str,
+    llm: LLMProvider,
+    *,
+    memory_context: str = "",
+) -> PersonaResult | None:
+    """Run the HR classifier and return a PersonaResult.
+
+    Makes a single non-streaming acomplete() call with the session LLM.
+    Returns None on any failure so the queen falls back gracefully to its
+    default character with no style directive.
+
+    Args:
+        user_message: The user's opening message for the session.
+        llm: The session LLM provider.
+        memory_context: Optional cross-session memory to inform style classification.
+
+    Returns:
+        A PersonaResult with persona_prefix and style_directive, or None on failure.
+    """
+    if not user_message.strip():
+        return None
+
+    prompt = user_message
+    if memory_context:
+        prompt = f"{user_message}\n\n{memory_context}"
+
+    try:
+        response = await llm.acomplete(
+            messages=[{"role": "user", "content": prompt}],
+            system=_HR_SYSTEM_PROMPT,
+            max_tokens=1024,
+            json_mode=True,
+        )
+        raw = response.content.strip()
+        parsed = json.loads(raw)
+        role = parsed.get("role", "").strip()
+        persona = parsed.get("persona", "").strip()
+        style_key = parsed.get("style", "peer-technical").strip()
+        if not role or not persona:
+            logger.warning("Thinking hook: empty role/persona in response: %r", raw)
+            return None
+        persona_prefix = f"You are a {role}. {persona}"
+        style_directive = _STYLE_DIRECTIVES.get(style_key, _STYLE_DIRECTIVES["peer-technical"])
+        logger.info("Thinking hook: selected persona — %s, style — %s", role, style_key)
+        return PersonaResult(persona_prefix=persona_prefix, style_directive=style_directive)
+    except Exception:
+        logger.warning("Thinking hook: persona classification failed", exc_info=True)
+        return None

core/framework/agents/queen/queen_memory.py

@@ -0,0 +1,420 @@
+"""Queen global cross-session memory.
+
+Three-tier memory architecture:
+  ~/.hive/queen/MEMORY.md                            — semantic (who, what, why)
+  ~/.hive/queen/memories/MEMORY-YYYY-MM-DD.md        — episodic (daily journals)
+  ~/.hive/queen/session/{id}/data/adapt.md           — working (session-scoped)
+
+Semantic and episodic files are injected at queen session start.
+
+Semantic memory (MEMORY.md) is updated automatically at session end via
+consolidate_queen_memory() — the queen never rewrites this herself.
+
+Episodic memory (MEMORY-date.md) can be written by the queen during a session
+via the write_to_diary tool, and is also appended to at session end by
+consolidate_queen_memory().
+"""
+
+from __future__ import annotations
+
+import asyncio
+import json
+import logging
+import traceback
+from datetime import date, datetime
+from pathlib import Path
+
+logger = logging.getLogger(__name__)
+
+
+def _queen_dir() -> Path:
+    return Path.home() / ".hive" / "queen"
+
+
+def format_memory_date(d: date) -> str:
+    """Return a cross-platform long date label without a zero-padded day."""
+    return f"{d.strftime('%B')} {d.day}, {d.year}"
+
+
+def semantic_memory_path() -> Path:
+    return _queen_dir() / "MEMORY.md"
+
+
+def episodic_memory_path(d: date | None = None) -> Path:
+    d = d or date.today()
+    return _queen_dir() / "memories" / f"MEMORY-{d.strftime('%Y-%m-%d')}.md"
+
+
+def read_semantic_memory() -> str:
+    path = semantic_memory_path()
+    return path.read_text(encoding="utf-8").strip() if path.exists() else ""
+
+
+def read_episodic_memory(d: date | None = None) -> str:
+    path = episodic_memory_path(d)
+    return path.read_text(encoding="utf-8").strip() if path.exists() else ""
+
+
+def _find_recent_episodic(lookback: int = 7) -> tuple[date, str] | None:
+    """Find the most recent non-empty episodic memory within *lookback* days."""
+    from datetime import timedelta
+
+    today = date.today()
+    for offset in range(lookback):
+        d = today - timedelta(days=offset)
+        content = read_episodic_memory(d)
+        if content:
+            return d, content
+    return None
+
+
+# Budget (in characters) for episodic memory in the system prompt.
+_EPISODIC_CHAR_BUDGET = 6_000
+
+
+def format_for_injection() -> str:
+    """Format cross-session memory for system prompt injection.
+
+    Returns an empty string if no meaningful content exists yet (e.g. first
+    session with only the seed template).
+    """
+    semantic = read_semantic_memory()
+    recent = _find_recent_episodic()
+
+    # Suppress injection if semantic is still just the seed template
+    if semantic and semantic.startswith("# My Understanding of the User\n\n*No sessions"):
+        semantic = ""
+
+    parts: list[str] = []
+    if semantic:
+        parts.append(semantic)
+
+    if recent:
+        d, content = recent
+        # Trim oversized episodic entries to keep the prompt manageable
+        if len(content) > _EPISODIC_CHAR_BUDGET:
+            content = content[:_EPISODIC_CHAR_BUDGET] + "\n\n…(truncated)"
+        today = date.today()
+        if d == today:
+            label = f"## Today — {format_memory_date(d)}"
+        else:
+            label = f"## {format_memory_date(d)}"
+        parts.append(f"{label}\n\n{content}")
+
+    if not parts:
+        return ""
+
+    body = "\n\n---\n\n".join(parts)
+    return "--- Your Cross-Session Memory ---\n\n" + body + "\n\n--- End Cross-Session Memory ---"
+
+
+_SEED_TEMPLATE = """\
+# My Understanding of the User
+
+*No sessions recorded yet.*
+
+## Who They Are
+
+## How They Communicate
+
+## What They're Trying to Achieve
+
+## What's Working
+
+## What I've Learned
+"""
+
+
+def append_episodic_entry(content: str) -> None:
+    """Append a timestamped prose entry to today's episodic memory file.
+
+    Creates the file (with a date heading) if it doesn't exist yet.
+    Used both by the queen's diary tool and by the consolidation hook.
+    """
+    ep_path = episodic_memory_path()
+    ep_path.parent.mkdir(parents=True, exist_ok=True)
+    today = date.today()
+    today_str = format_memory_date(today)
+    timestamp = datetime.now().strftime("%H:%M")
+    if not ep_path.exists():
+        header = f"# {today_str}\n\n"
+        block = f"{header}### {timestamp}\n\n{content.strip()}\n"
+    else:
+        block = f"\n\n### {timestamp}\n\n{content.strip()}\n"
+    with ep_path.open("a", encoding="utf-8") as f:
+        f.write(block)
+
+
+def seed_if_missing() -> None:
+    """Create MEMORY.md with a blank template if it doesn't exist yet."""
+    path = semantic_memory_path()
+    if path.exists():
+        return
+    path.parent.mkdir(parents=True, exist_ok=True)
+    path.write_text(_SEED_TEMPLATE, encoding="utf-8")
+
+
+# ---------------------------------------------------------------------------
+# Consolidation prompt
+# ---------------------------------------------------------------------------
+
+_SEMANTIC_SYSTEM = """\
+You maintain the persistent cross-session memory of an AI assistant called the Queen.
+Review the session notes and rewrite MEMORY.md — the Queen's durable understanding of the
+person she works with across all sessions.
+
+Write entirely in the Queen's voice — first person, reflective, honest.
+Not a log of events, but genuine understanding of who this person is over time.
+
+Rules:
+- Update and synthesise: incorporate new understanding, update facts that have changed, remove
+  details that are stale, superseded, or no longer say anything meaningful about the person.
+- Keep it as structured markdown with named sections about the PERSON, not about today.
+- Do NOT include diary sections, daily logs, or session summaries. Those belong elsewhere.
+  MEMORY.md is about who they are, what they want, what works — not what happened today.
+- Maintain a "How They Communicate" section: technical depth, preferred pace
+  (fast/exploratory/thorough), what communication approaches have worked or not,
+  tone preferences. Update based on diary reflections about communication.
+  This section should evolve — "prefers direct answers" is useful on day 1;
+  "prefers direct answers for technical questions but wants more context when
+  discussing architecture trade-offs" is better by day 5.
+- Reference dates only when noting a lasting milestone (e.g. "since March 8th they prefer X").
+- If the session had no meaningful new information about the person,
+  return the existing text unchanged.
+- Do not add fictional details. Only reflect what is evidenced in the notes.
+- Stay concise. Prune rather than accumulate. A lean, accurate file is more useful than a
+  dense one. If something was true once but has been resolved or superseded, remove it.
+- Output only the raw markdown content of MEMORY.md. No preamble, no code fences.
+"""
+
+_DIARY_SYSTEM = """\
+You maintain the daily episodic diary of an AI assistant called the Queen.
+You receive: (1) today's existing diary so far, and (2) notes from the latest session.
+
+Rewrite the complete diary for today as a single unified narrative —
+first person, reflective, honest.
+Merge and deduplicate: if the same story (e.g. a research agent stalling) recurred several times,
+describe it once with appropriate weight rather than retelling it. Weave in new developments from
+the session notes. Preserve important milestones, emotional texture, and session path references.
+Preserve reflections about communication effectiveness — these are important inputs for the
+Queen's evolving understanding of the user. A reflection like "they responded much better when
+I led with the recommendation instead of listing options" is as important as
+"we built a Gmail agent."
+
+If today's diary is empty, write the initial entry based on the session notes alone.
+
+Output only the full diary prose — no date heading, no timestamp headers,
+no preamble, no code fences.
+"""
+
+
+def read_session_context(session_dir: Path, max_messages: int = 80) -> str:
+    """Extract a readable transcript from conversation parts + adapt.md.
+
+    Reads the last ``max_messages`` conversation parts and the session's
+    adapt.md (working memory). Tool results are omitted — only user and
+    assistant turns (with tool-call names noted) are included.
+    """
+    parts: list[str] = []
+
+    # Working notes
+    adapt_path = session_dir / "data" / "adapt.md"
+    if adapt_path.exists():
+        text = adapt_path.read_text(encoding="utf-8").strip()
+        if text:
+            parts.append(f"## Session Working Notes (adapt.md)\n\n{text}")
+
+    # Conversation transcript
+    parts_dir = session_dir / "conversations" / "parts"
+    if parts_dir.exists():
+        part_files = sorted(parts_dir.glob("*.json"))[-max_messages:]
+        lines: list[str] = []
+        for pf in part_files:
+            try:
+                data = json.loads(pf.read_text(encoding="utf-8"))
+                role = data.get("role", "")
+                content = str(data.get("content", "")).strip()
+                tool_calls = data.get("tool_calls") or []
+                if role == "tool":
+                    continue  # skip verbose tool results
+                if role == "assistant" and tool_calls and not content:
+                    names = [tc.get("function", {}).get("name", "?") for tc in tool_calls]
+                    lines.append(f"[queen calls: {', '.join(names)}]")
+                elif content:
+                    label = "user" if role == "user" else "queen"
+                    lines.append(f"[{label}]: {content[:600]}")
+            except (KeyError, TypeError) as exc:
+                logger.debug("Skipping malformed conversation message: %s", exc)
+                continue
+            except Exception:
+                logger.warning("Unexpected error parsing conversation message", exc_info=True)
+                continue
+        if lines:
+            parts.append("## Conversation\n\n" + "\n".join(lines))
+
+    return "\n\n".join(parts)
+
+
+# ---------------------------------------------------------------------------
+# Context compaction (binary-split LLM summarisation)
+# ---------------------------------------------------------------------------
+
+# If the raw session context exceeds this many characters, compact it first
+# before sending to the consolidation LLM. ~200 k chars ≈ 50 k tokens.
+_CTX_COMPACT_CHAR_LIMIT = 200_000
+_CTX_COMPACT_MAX_DEPTH = 8
+
+_COMPACT_SYSTEM = (
+    "Summarise this conversation segment. Preserve: user goals, key decisions, "
+    "what was built or changed, emotional tone, and important outcomes. "
+    "Write concisely in third person past tense. Omit routine tool invocations "
+    "unless the result matters."
+)
+
+
+async def _compact_context(text: str, llm: object, *, _depth: int = 0) -> str:
+    """Binary-split and LLM-summarise *text* until it fits within the char limit.
+
+    Mirrors the recursive binary-splitting strategy used by the main agent
+    compaction pipeline (EventLoopNode._llm_compact).
+    """
+    if len(text) <= _CTX_COMPACT_CHAR_LIMIT or _depth >= _CTX_COMPACT_MAX_DEPTH:
+        return text
+
+    # Split near the midpoint on a line boundary so we don't cut mid-message
+    mid = len(text) // 2
+    split_at = text.rfind("\n", 0, mid) + 1
+    if split_at <= 0:
+        split_at = mid
+
+    half1, half2 = text[:split_at], text[split_at:]
+
+    async def _summarise(chunk: str) -> str:
+        try:
+            resp = await llm.acomplete(
+                messages=[{"role": "user", "content": chunk}],
+                system=_COMPACT_SYSTEM,
+                max_tokens=2048,
+            )
+            return resp.content.strip()
+        except Exception:
+            logger.warning(
+                "queen_memory: context compaction LLM call failed (depth=%d), truncating",
+                _depth,
+            )
+            return chunk[: _CTX_COMPACT_CHAR_LIMIT // 4]
+
+    s1, s2 = await asyncio.gather(_summarise(half1), _summarise(half2))
+    combined = s1 + "\n\n" + s2
+    if len(combined) > _CTX_COMPACT_CHAR_LIMIT:
+        return await _compact_context(combined, llm, _depth=_depth + 1)
+    return combined
+
+
+async def consolidate_queen_memory(
+    session_id: str,
+    session_dir: Path,
+    llm: object,
+) -> None:
+    """Update MEMORY.md and append a diary entry based on the current session.
+
+    Reads conversation parts and adapt.md from session_dir. Called
+    periodically in the background and once at session end. Failures are
+    logged and silently swallowed so they never block teardown.
+
+    Args:
+        session_id: The session ID (used for the adapt.md path reference).
+        session_dir: Path to the session directory (~/.hive/queen/session/{id}).
+        llm: LLMProvider instance (must support acomplete()).
+    """
+    try:
+        session_context = read_session_context(session_dir)
+        if not session_context:
+            logger.debug("queen_memory: no session context, skipping consolidation")
+            return
+
+        logger.info("queen_memory: consolidating memory for session %s ...", session_id)
+
+        # If the transcript is very large, compact it with recursive binary LLM
+        # summarisation before sending to the consolidation model.
+        if len(session_context) > _CTX_COMPACT_CHAR_LIMIT:
+            logger.info(
+                "queen_memory: session context is %d chars — compacting first",
+                len(session_context),
+            )
+            session_context = await _compact_context(session_context, llm)
+            logger.info("queen_memory: compacted to %d chars", len(session_context))
+
+        existing_semantic = read_semantic_memory()
+        today_journal = read_episodic_memory()
+        today = date.today()
+        today_str = format_memory_date(today)
+        adapt_path = session_dir / "data" / "adapt.md"
+
+        user_msg = (
+            f"## Existing Semantic Memory (MEMORY.md)\n\n"
+            f"{existing_semantic or '(none yet)'}\n\n"
+            f"## Today's Diary So Far ({today_str})\n\n"
+            f"{today_journal or '(none yet)'}\n\n"
+            f"{session_context}\n\n"
+            f"## Session Reference\n\n"
+            f"Session ID: {session_id}\n"
+            f"Session path: {adapt_path}\n"
+        )
+
+        logger.debug(
+            "queen_memory: calling LLM (%d chars of context, ~%d tokens est.)",
+            len(user_msg),
+            len(user_msg) // 4,
+        )
+
+        from framework.agents.queen.config import default_config
+
+        semantic_resp, diary_resp = await asyncio.gather(
+            llm.acomplete(
+                messages=[{"role": "user", "content": user_msg}],
+                system=_SEMANTIC_SYSTEM,
+                max_tokens=default_config.max_tokens,
+            ),
+            llm.acomplete(
+                messages=[{"role": "user", "content": user_msg}],
+                system=_DIARY_SYSTEM,
+                max_tokens=default_config.max_tokens,
+            ),
+        )
+
+        new_semantic = semantic_resp.content.strip()
+        diary_entry = diary_resp.content.strip()
+
+        if new_semantic:
+            path = semantic_memory_path()
+            path.parent.mkdir(parents=True, exist_ok=True)
+            path.write_text(new_semantic, encoding="utf-8")
+            logger.info("queen_memory: semantic memory updated (%d chars)", len(new_semantic))
+
+        if diary_entry:
+            # Rewrite today's episodic file in-place — the LLM has merged and
+            # deduplicated the full day's content, so we replace rather than append.
+            ep_path = episodic_memory_path()
+            ep_path.parent.mkdir(parents=True, exist_ok=True)
+            heading = f"# {today_str}"
+            ep_path.write_text(f"{heading}\n\n{diary_entry}\n", encoding="utf-8")
+            logger.info(
+                "queen_memory: episodic diary rewritten for %s (%d chars)",
+                today_str,
+                len(diary_entry),
+            )
+
+    except Exception:
+        tb = traceback.format_exc()
+        logger.exception("queen_memory: consolidation failed")
+        # Write to file so the cause is findable regardless of log verbosity.
+        error_path = _queen_dir() / "consolidation_error.txt"
+        try:
+            error_path.parent.mkdir(parents=True, exist_ok=True)
+            error_path.write_text(
+                f"session: {session_id}\ntime: {datetime.now().isoformat()}\n\n{tb}",
+                encoding="utf-8",
+            )
+        except OSError:
+            pass  # Cannot write error file; original exception already logged

core/framework/agents/queen/queen_memory_v2.py

@@ -0,0 +1,553 @@
+"""Shared memory helpers for queen/worker recall and reflection.
+
+Each memory is an individual ``.md`` file in ``~/.hive/queen/memories/``
+with optional YAML frontmatter (name, type, description).  Frontmatter
+is a convention enforced by prompt instructions — parsing is lenient and
+malformed files degrade gracefully (appear in scans with ``None`` metadata).
+
+Cursor-based incremental processing tracks which conversation messages
+have already been processed by the reflection agent.
+"""
+
+from __future__ import annotations
+
+import json
+import logging
+import re
+import shutil
+import time
+from dataclasses import dataclass, field
+from datetime import date
+from pathlib import Path
+from typing import Any
+
+logger = logging.getLogger(__name__)
+
+# ---------------------------------------------------------------------------
+# Constants
+# ---------------------------------------------------------------------------
+
+MEMORY_TYPES: tuple[str, ...] = ("goal", "environment", "technique", "reference", "diary")
+GLOBAL_MEMORY_CATEGORIES: tuple[str, ...] = ("profile", "preference", "environment", "feedback")
+
+_HIVE_QUEEN_DIR = Path.home() / ".hive" / "queen"
+# Legacy shared v2 root.  Colony memory now lives under queen sessions.
+MEMORY_DIR: Path = _HIVE_QUEEN_DIR / "memories"
+
+MAX_FILES: int = 200
+MAX_FILE_SIZE_BYTES: int = 4096  # 4 KB hard limit per memory file
+
+# How many lines of a memory file to read for header scanning.
+_HEADER_LINE_LIMIT: int = 30
+_MIGRATION_MARKER = ".migrated-from-shared-memory"
+_GLOBAL_MEMORY_CODE_PATTERN = re.compile(
+    r"(/Users/|~/.hive|\.py\b|\.ts\b|\.tsx\b|\.js\b|"
+    r"\b(graph|node|runtime|session|execution|worker|queen|subagent|checkpoint|flowchart)\b)",
+    re.IGNORECASE,
+)
+
+# Frontmatter example provided to the reflection agent via prompt.
+MEMORY_FRONTMATTER_EXAMPLE: list[str] = [
+    "```markdown",
+    "---",
+    "name: {{memory name}}",
+    (
+        "description: {{one-line description — used to decide "
+        "relevance in future conversations, so be specific}}"
+    ),
+    f"type: {{{{{', '.join(MEMORY_TYPES)}}}}}",
+    "---",
+    "",
+    (
+        "{{memory content — for feedback/project types, "
+        "structure as: rule/fact, then **Why:** "
+        "and **How to apply:** lines}}"
+    ),
+    "```",
+]
+
+
+def colony_memory_dir(colony_id: str) -> Path:
+    """Return the colony memory directory for a queen session."""
+    return _HIVE_QUEEN_DIR / "session" / colony_id / "memory" / "colony"
+
+
+def global_memory_dir() -> Path:
+    """Return the queen-global memory directory."""
+    return _HIVE_QUEEN_DIR / "global_memory"
+
+
+# ---------------------------------------------------------------------------
+# Frontmatter parsing (lenient)
+# ---------------------------------------------------------------------------
+
+_FRONTMATTER_RE = re.compile(r"^---\s*\n(.*?)\n---\s*\n?", re.DOTALL)
+
+
+def parse_frontmatter(text: str) -> dict[str, str]:
+    """Extract YAML-ish frontmatter from *text*.
+
+    Returns a dict of key-value pairs.  Never raises — returns ``{}`` on
+    any parse failure.  Values are stripped strings; no nested structures.
+    """
+    m = _FRONTMATTER_RE.match(text)
+    if not m:
+        return {}
+    result: dict[str, str] = {}
+    for line in m.group(1).splitlines():
+        line = line.strip()
+        if not line or line.startswith("#"):
+            continue
+        colon = line.find(":")
+        if colon < 1:
+            continue
+        key = line[:colon].strip().lower()
+        val = line[colon + 1 :].strip()
+        if val:
+            result[key] = val
+    return result
+
+
+def parse_memory_type(raw: str | None) -> str | None:
+    """Validate *raw* against supported memory categories."""
+    if raw is None:
+        return None
+    normalized = raw.strip().lower()
+    allowed = set(MEMORY_TYPES) | set(GLOBAL_MEMORY_CATEGORIES)
+    return normalized if normalized in allowed else None
+
+
+def parse_global_memory_category(raw: str | None) -> str | None:
+    """Validate *raw* against ``GLOBAL_MEMORY_CATEGORIES``."""
+    if raw is None:
+        return None
+    normalized = raw.strip().lower()
+    return normalized if normalized in GLOBAL_MEMORY_CATEGORIES else None
+
+
+# ---------------------------------------------------------------------------
+# MemoryFile dataclass
+# ---------------------------------------------------------------------------
+
+
+@dataclass
+class MemoryFile:
+    """Parsed representation of a single memory file on disk."""
+
+    filename: str
+    path: Path
+    # Frontmatter fields — all nullable (lenient parsing).
+    name: str | None = None
+    type: str | None = None
+    description: str | None = None
+    # First N lines of the file (for manifest / header scanning).
+    header_lines: list[str] = field(default_factory=list)
+    # Filesystem modification time (seconds since epoch).
+    mtime: float = 0.0
+
+    @classmethod
+    def from_path(cls, path: Path) -> MemoryFile:
+        """Read a memory file and leniently parse its frontmatter."""
+        try:
+            text = path.read_text(encoding="utf-8")
+        except OSError:
+            return cls(filename=path.name, path=path)
+
+        fm = parse_frontmatter(text)
+        lines = text.splitlines()[:_HEADER_LINE_LIMIT]
+
+        try:
+            mtime = path.stat().st_mtime
+        except OSError:
+            mtime = 0.0
+
+        return cls(
+            filename=path.name,
+            path=path,
+            name=fm.get("name"),
+            type=parse_memory_type(fm.get("type")),
+            description=fm.get("description"),
+            header_lines=lines,
+            mtime=mtime,
+        )
+
+
+# ---------------------------------------------------------------------------
+# Scanning
+# ---------------------------------------------------------------------------
+
+
+def scan_memory_files(memory_dir: Path | None = None) -> list[MemoryFile]:
+    """Scan *memory_dir* for ``.md`` files, returning up to ``MAX_FILES``.
+
+    Files are sorted by modification time (newest first).  Dotfiles and
+    subdirectories are ignored.
+    """
+    d = memory_dir or MEMORY_DIR
+    if not d.is_dir():
+        return []
+
+    md_files = sorted(
+        (f for f in d.glob("*.md") if f.is_file() and not f.name.startswith(".")),
+        key=lambda p: p.stat().st_mtime,
+        reverse=True,
+    )
+
+    return [MemoryFile.from_path(f) for f in md_files[:MAX_FILES]]
+
+
+def slugify_memory_name(raw: str) -> str:
+    """Create a filesystem-safe slug for a memory filename."""
+    slug = re.sub(r"[^a-z0-9]+", "-", raw.strip().lower()).strip("-")
+    return slug or "memory"
+
+
+def allocate_memory_filename(
+    memory_dir: Path,
+    name: str,
+    *,
+    suffix: str = ".md",
+) -> str:
+    """Allocate a unique filename in *memory_dir* based on *name*."""
+    base = slugify_memory_name(name)
+    candidate = f"{base}{suffix}"
+    counter = 2
+    while (memory_dir / candidate).exists():
+        candidate = f"{base}-{counter}{suffix}"
+        counter += 1
+    return candidate
+
+
+def build_memory_document(
+    *,
+    name: str,
+    description: str,
+    mem_type: str,
+    body: str,
+) -> str:
+    """Build one memory file with frontmatter and body."""
+    return (
+        f"---\n"
+        f"name: {name.strip()}\n"
+        f"description: {description.strip()}\n"
+        f"type: {mem_type.strip()}\n"
+        f"---\n\n"
+        f"{body.strip()}\n"
+    )
+
+
+def diary_filename(d: date | None = None) -> str:
+    """Return the diary memory filename for date *d* (default: today)."""
+    d = d or date.today()
+    return f"MEMORY-{d.strftime('%Y-%m-%d')}.md"
+
+
+def build_diary_document(*, date_str: str, body: str) -> str:
+    """Build a diary memory file with frontmatter."""
+    return build_memory_document(
+        name=f"diary-{date_str}",
+        description=f"Daily session narrative for {date_str}",
+        mem_type="diary",
+        body=body,
+    )
+
+
+def validate_global_memory_payload(
+    *,
+    category: str,
+    description: str,
+    content: str,
+) -> str:
+    """Validate a queen-global memory save request."""
+    parsed = parse_global_memory_category(category)
+    if parsed is None:
+        raise ValueError(
+            "Invalid global memory category. Use one of: "
+            + ", ".join(GLOBAL_MEMORY_CATEGORIES)
+        )
+    if not description.strip():
+        raise ValueError("Global memory description cannot be empty.")
+    if not content.strip():
+        raise ValueError("Global memory content cannot be empty.")
+
+    probe = f"{description}\n{content}"
+    if _GLOBAL_MEMORY_CODE_PATTERN.search(probe):
+        raise ValueError(
+            "Global memory is only for durable user profile, preferences, "
+            "environment, or feedback — not task/code/runtime details."
+        )
+    return parsed
+
+
+def save_global_memory(
+    *,
+    category: str,
+    description: str,
+    content: str,
+    name: str | None = None,
+    memory_dir: Path | None = None,
+) -> tuple[str, Path]:
+    """Persist one queen-global memory entry."""
+    parsed = validate_global_memory_payload(
+        category=category,
+        description=description,
+        content=content,
+    )
+    target_dir = memory_dir or global_memory_dir()
+    target_dir.mkdir(parents=True, exist_ok=True)
+    memory_name = (name or description).strip()
+    filename = allocate_memory_filename(target_dir, memory_name)
+    doc = build_memory_document(
+        name=memory_name,
+        description=description,
+        mem_type=parsed,
+        body=content,
+    )
+    if len(doc.encode("utf-8")) > MAX_FILE_SIZE_BYTES:
+        raise ValueError(
+            f"Global memory entry exceeds the {MAX_FILE_SIZE_BYTES} byte limit."
+        )
+    path = target_dir / filename
+    path.write_text(doc, encoding="utf-8")
+    return filename, path
+
+
+# ---------------------------------------------------------------------------
+# Manifest formatting
+# ---------------------------------------------------------------------------
+
+def _age_label(mtime: float) -> str:
+    """Human-readable age string from an mtime."""
+    age_days = memory_age_days(mtime)
+    if age_days <= 0:
+        return "today"
+    if age_days == 1:
+        return "1 day ago"
+    return f"{age_days} days ago"
+
+
+def format_memory_manifest(files: list[MemoryFile]) -> str:
+    """One-line-per-file text manifest for the recall selector / reflection agent.
+
+    Format: ``[type] filename (age): description``
+    """
+    lines: list[str] = []
+    for mf in files:
+        t = mf.type or "unknown"
+        desc = mf.description or "(no description)"
+        age = _age_label(mf.mtime)
+        lines.append(f"[{t}] {mf.filename} ({age}): {desc}")
+    return "\n".join(lines)
+
+
+# ---------------------------------------------------------------------------
+# Freshness / staleness
+# ---------------------------------------------------------------------------
+
+_SECONDS_PER_DAY = 86_400
+
+
+def memory_age_days(mtime: float) -> int:
+    """Return the age of a memory file in whole days."""
+    if mtime <= 0:
+        return 0
+    return int((time.time() - mtime) / _SECONDS_PER_DAY)
+
+
+def memory_freshness_text(mtime: float) -> str:
+    """Return a staleness warning for injection, or empty string if fresh."""
+    d = memory_age_days(mtime)
+    if d <= 1:
+        return ""
+    return (
+        f"This memory is {d} days old. "
+        "Memories are point-in-time observations, not live state — "
+        "claims about code behavior or file:line citations may be outdated. "
+        "Verify against current code before asserting as fact."
+    )
+
+
+# ---------------------------------------------------------------------------
+# Cursor-based incremental processing
+# ---------------------------------------------------------------------------
+
+
+async def read_conversation_parts(session_dir: Path) -> list[dict[str, Any]]:
+    """Read all conversation parts for a session using FileConversationStore.
+
+    Returns a list of raw message dicts in sequence order.
+    """
+    from framework.storage.conversation_store import FileConversationStore
+
+    store = FileConversationStore(session_dir / "conversations")
+    return await store.read_parts()
+
+
+# ---------------------------------------------------------------------------
+# Initialisation and legacy migration
+# ---------------------------------------------------------------------------
+
+
+def init_memory_dir(
+    memory_dir: Path | None = None,
+    *,
+    migrate_legacy: bool = False,
+) -> None:
+    """Create the memory directory if missing.
+
+    When ``migrate_legacy`` is true, migrate both v1 memory files and the
+    previous shared v2 queen memory store into this directory.
+    """
+    d = memory_dir or MEMORY_DIR
+    first_run = not d.exists()
+    d.mkdir(parents=True, exist_ok=True)
+    if migrate_legacy:
+        migrate_legacy_memories(d)
+        migrate_shared_v2_memories(d)
+    elif first_run and d == MEMORY_DIR:
+        migrate_legacy_memories(d)
+
+
+def migrate_legacy_memories(memory_dir: Path | None = None) -> None:
+    """Convert old MEMORY.md + MEMORY-YYYY-MM-DD.md files to individual memory files.
+
+    Originals are moved to ``{memory_dir}/.legacy/``.
+    """
+    d = memory_dir or MEMORY_DIR
+    queen_dir = _HIVE_QUEEN_DIR
+    legacy_archive = d / ".legacy"
+
+    migrated_any = False
+
+    # --- Semantic memory (MEMORY.md) ---
+    semantic = queen_dir / "MEMORY.md"
+    if semantic.exists():
+        content = semantic.read_text(encoding="utf-8").strip()
+        # Skip the blank seed template.
+        if content and not content.startswith("# My Understanding of the User\n\n*No sessions"):
+            _write_migration_file(
+                d,
+                filename="legacy-semantic-memory.md",
+                name="legacy-semantic-memory",
+                mem_type="reference",
+                description="Migrated semantic memory from previous memory system",
+                body=content,
+            )
+            migrated_any = True
+        # Archive original.
+        legacy_archive.mkdir(parents=True, exist_ok=True)
+        semantic.rename(legacy_archive / "MEMORY.md")
+
+    # --- Episodic memories (MEMORY-YYYY-MM-DD.md) ---
+    old_memories_dir = queen_dir / "memories"
+    if old_memories_dir.is_dir():
+        for ep_file in sorted(old_memories_dir.glob("MEMORY-*.md")):
+            content = ep_file.read_text(encoding="utf-8").strip()
+            if not content:
+                continue
+            date_part = ep_file.stem.replace("MEMORY-", "")
+            slug = f"legacy-diary-{date_part}.md"
+            _write_migration_file(
+                d,
+                filename=slug,
+                name=f"legacy-diary-{date_part}",
+                mem_type="diary",
+                description=f"Migrated diary entry from {date_part}",
+                body=content,
+            )
+            migrated_any = True
+            # Archive original.
+            legacy_archive.mkdir(parents=True, exist_ok=True)
+            ep_file.rename(legacy_archive / ep_file.name)
+
+    if migrated_any:
+        logger.info("queen_memory_v2: migrated legacy memory files to %s", d)
+
+
+def migrate_shared_v2_memories(
+    memory_dir: Path | None = None,
+    *,
+    source_dir: Path | None = None,
+) -> None:
+    """Move shared queen v2 memory files into a colony directory once."""
+    d = memory_dir or MEMORY_DIR
+    d.mkdir(parents=True, exist_ok=True)
+    src = source_dir or MEMORY_DIR
+    if d.resolve() == src.resolve():
+        return
+
+    marker = d / _MIGRATION_MARKER
+    if marker.exists():
+        return
+
+    if not src.is_dir():
+        return
+
+    md_files = sorted(
+        f for f in src.glob("*.md")
+        if f.is_file() and not f.name.startswith(".")
+    )
+    if not md_files:
+        marker.write_text("no shared memories found\n", encoding="utf-8")
+        return
+
+    archive = src / ".legacy_colony_migration"
+    archive.mkdir(parents=True, exist_ok=True)
+    migrated_any = False
+
+    for src_file in md_files:
+        target = d / src_file.name
+        if not target.exists():
+            try:
+                shutil.copy2(src_file, target)
+                migrated_any = True
+            except OSError:
+                logger.debug("shared memory migration copy failed for %s", src_file, exc_info=True)
+                continue
+
+        archived = archive / src_file.name
+        counter = 2
+        while archived.exists():
+            archived = archive / f"{src_file.stem}-{counter}{src_file.suffix}"
+            counter += 1
+        try:
+            src_file.rename(archived)
+        except OSError:
+            logger.debug("shared memory migration archive failed for %s", src_file, exc_info=True)
+
+    if migrated_any:
+        logger.info("queen_memory_v2: migrated shared queen memories to %s", d)
+    marker.write_text(
+        f"migrated_at={int(time.time())}\nsource={src}\n",
+        encoding="utf-8",
+    )
+
+
+def _write_migration_file(
+    memory_dir: Path,
+    filename: str,
+    name: str,
+    mem_type: str,
+    description: str,
+    body: str,
+) -> None:
+    """Write a single migrated memory file with frontmatter."""
+    # Truncate body to respect file size limit (leave room for frontmatter).
+    header = (
+        f"---\n"
+        f"name: {name}\n"
+        f"description: {description}\n"
+        f"type: {mem_type}\n"
+        f"---\n\n"
+    )
+    max_body = MAX_FILE_SIZE_BYTES - len(header.encode("utf-8"))
+    if len(body.encode("utf-8")) > max_body:
+        # Rough truncation — cut at character level then trim to last newline.
+        body = body[: max_body - 20]
+        nl = body.rfind("\n")
+        if nl > 0:
+            body = body[:nl]
+        body += "\n\n...(truncated during migration)"
+
+    path = memory_dir / filename
+    path.write_text(header + body + "\n", encoding="utf-8")

core/framework/agents/queen/recall_selector.py

@@ -0,0 +1,236 @@
+"""Recall selector — pre-turn memory selection for queen and worker memory.
+
+Before each conversation turn the system:
+  1. Scans the memory directory for ``.md`` files (cap: 200).
+  2. Reads headers (frontmatter + first 30 lines).
+  3. Uses a single LLM call with structured JSON output to pick the ~5
+     most relevant memories.
+  4. Injects them into context with staleness warnings for older ones.
+
+The selector only sees the user's query string — no full conversation
+context.  This keeps it cheap and fast.  Errors are caught and return
+``[]`` so the main conversation is never blocked.
+"""
+
+from __future__ import annotations
+
+import json
+import logging
+from pathlib import Path
+from typing import Any
+
+from framework.agents.queen.queen_memory_v2 import (
+    MEMORY_DIR,
+    format_memory_manifest,
+    memory_freshness_text,
+    scan_memory_files,
+)
+
+logger = logging.getLogger(__name__)
+
+# ---------------------------------------------------------------------------
+# Structured output schema
+# ---------------------------------------------------------------------------
+
+RECALL_SCHEMA: dict[str, Any] = {
+    "type": "json_schema",
+    "json_schema": {
+        "name": "memory_selection",
+        "strict": True,
+        "schema": {
+            "type": "object",
+            "properties": {
+                "selected_memories": {
+                    "type": "array",
+                    "items": {"type": "string"},
+                },
+            },
+            "required": ["selected_memories"],
+            "additionalProperties": False,
+        },
+    },
+}
+
+# ---------------------------------------------------------------------------
+# System prompt
+# ---------------------------------------------------------------------------
+
+SELECT_MEMORIES_SYSTEM_PROMPT = """\
+You are selecting memories that will be useful to the Queen agent as it \
+processes a user's query.
+
+You will be given the user's query and a list of available memory files \
+with their filenames and descriptions.
+
+Return a JSON object with a single key "selected_memories" containing a \
+list of filenames for the memories that will clearly be useful as the \
+Queen processes the user's query (up to 5).
+
+Only include memories that you are certain will be helpful based on their \
+name and description.
+- If you are unsure if a memory will be useful in processing the user's \
+query, then do not include it in your list.  Be selective and discerning.
+- If there are no memories in the list that would clearly be useful, \
+return an empty list.
+- If a list of recently-used tools is provided, do not select memories \
+that are usage reference or API documentation for those tools (the Queen \
+is already exercising them).  Still select warnings or gotchas about them.
+"""
+
+# ---------------------------------------------------------------------------
+# Core functions
+# ---------------------------------------------------------------------------
+
+
+async def select_memories(
+    query: str,
+    llm: Any,
+    memory_dir: Path | None = None,
+    active_tools: list[str] | None = None,
+    *,
+    max_results: int = 5,
+) -> list[str]:
+    """Select up to 5 relevant memory filenames for *query*.
+
+    Returns a list of filenames.  Best-effort: on any error returns ``[]``.
+    """
+    mem_dir = memory_dir or MEMORY_DIR
+    files = scan_memory_files(mem_dir)
+    if not files:
+        logger.debug("recall: no memory files found, skipping selection")
+        return []
+
+    logger.debug("recall: selecting from %d memory files for query: %.80s", len(files), query)
+    manifest = format_memory_manifest(files)
+
+    user_msg_parts = [f"## User query\n\n{query}\n\n## Available memories\n\n{manifest}"]
+    if active_tools:
+        user_msg_parts.append(f"\n\n## Recently-used tools\n\n{', '.join(active_tools)}")
+
+    user_msg = "".join(user_msg_parts)
+
+    try:
+        resp = await llm.acomplete(
+            messages=[{"role": "user", "content": user_msg}],
+            system=SELECT_MEMORIES_SYSTEM_PROMPT,
+            max_tokens=512,
+            response_format=RECALL_SCHEMA,
+        )
+        data = json.loads(resp.content)
+        selected = data.get("selected_memories", [])
+        # Validate: only return filenames that actually exist.
+        valid_names = {f.filename for f in files}
+        result = [s for s in selected if s in valid_names][:max_results]
+        logger.debug("recall: selected %d memories: %s", len(result), result)
+        return result
+    except Exception:
+        logger.debug("recall: memory selection failed, returning []", exc_info=True)
+        return []
+
+
+def format_recall_injection(
+    filenames: list[str],
+    memory_dir: Path | None = None,
+    *,
+    heading: str = "Selected Memories",
+) -> str:
+    """Read selected memory files and format for system prompt injection.
+
+    Prepends a staleness warning for memories older than 1 day.
+    """
+    mem_dir = memory_dir or MEMORY_DIR
+    if not filenames:
+        return ""
+
+    blocks: list[str] = []
+    for fname in filenames:
+        path = mem_dir / fname
+        if not path.is_file():
+            continue
+        try:
+            content = path.read_text(encoding="utf-8").strip()
+        except OSError:
+            continue
+
+        try:
+            mtime = path.stat().st_mtime
+        except OSError:
+            mtime = 0.0
+
+        freshness = memory_freshness_text(mtime)
+        header = f"### {fname}"
+        if freshness:
+            header += f"\n\n> {freshness}"
+        blocks.append(f"{header}\n\n{content}")
+
+    if not blocks:
+        return ""
+
+    body = "\n\n---\n\n".join(blocks)
+    logger.debug("recall: injecting %d memory blocks into context", len(blocks))
+    return f"--- {heading} ---\n\n{body}\n\n--- End {heading} ---"
+
+
+# ---------------------------------------------------------------------------
+# Cache update (called after each queen turn)
+# ---------------------------------------------------------------------------
+
+
+async def update_recall_cache(
+    session_dir: Path,
+    llm: Any,
+    phase_state: Any | None = None,
+    memory_dir: Path | None = None,
+    *,
+    cache_setter: Any = None,
+    heading: str = "Selected Memories",
+    active_tools: list[str] | None = None,
+) -> None:
+    """Update the recall cache on *phase_state* for the next turn.
+
+    Reads the latest user message from conversation parts to use as the
+    query for memory selection.
+    """
+    mem_dir = memory_dir or MEMORY_DIR
+
+    # Extract latest user message as the query.
+    query = _extract_latest_user_query(session_dir)
+    if not query:
+        logger.debug("recall: no user query found, skipping cache update")
+        return
+    logger.debug("recall: updating cache for query: %.80s", query)
+
+    try:
+        selected = await select_memories(
+            query,
+            llm,
+            mem_dir,
+            active_tools=active_tools,
+        )
+        injection = format_recall_injection(selected, mem_dir, heading=heading)
+        if cache_setter is not None:
+            cache_setter(injection)
+        elif phase_state is not None:
+            phase_state._cached_recall_block = injection
+    except Exception:
+        logger.debug("recall: cache update failed", exc_info=True)
+
+
+def _extract_latest_user_query(session_dir: Path) -> str:
+    """Read the most recent user message from conversation parts."""
+    parts_dir = session_dir / "conversations" / "parts"
+    if not parts_dir.is_dir():
+        return ""
+
+    part_files = sorted(parts_dir.glob("*.json"), reverse=True)
+    for f in part_files[:20]:  # Look back at most 20 messages.
+        try:
+            data = json.loads(f.read_text(encoding="utf-8"))
+            if data.get("role") == "user":
+                content = str(data.get("content", "")).strip()
+                if content:
+                    # Truncate very long queries.
+                    return content[:1000] if len(content) > 1000 else content
+        except (json.JSONDecodeError, OSError):
+            continue
+    return ""

core/framework/agents/queen/reference/anti_patterns.md

@@ -0,0 +1,35 @@
+# Common Mistakes When Building Hive Agents
+
+## Critical Errors
+1. **Using tools that don't exist** — Always verify tools via `list_agent_tools()` before designing. Common hallucinations: `csv_read`, `csv_write`, `file_upload`, `database_query`, `bulk_fetch_emails`.
+2. **Wrong mcp_servers.json format** — Flat dict (no `"mcpServers"` wrapper). `cwd` must be `"../../tools"`. `command` must be `"uv"` with args `["run", "python", ...]`.
+3. **Missing module-level exports in `__init__.py`** — The runner reads `goal`, `nodes`, `edges`, `entry_node`, `entry_points`, `terminal_nodes`, `conversation_mode`, `identity_prompt`, `loop_config` via `getattr()`. ALL module-level variables from agent.py must be re-exported in `__init__.py`.
+
+## Value Errors
+4. **Fabricating tools** — Always verify via `list_agent_tools()` before designing and `validate_agent_package()` after building.
+
+## Design Errors
+5. **Adding framework gating for LLM behavior** — Don't add output rollback or premature rejection. Fix with better prompts or custom judges.
+6. **Calling set_output in same turn as tool calls** — Call set_output in a SEPARATE turn.
+
+## File Template Errors
+7. **Wrong import paths** — Use `from framework.graph import ...`, NOT `from core.framework.graph import ...`.
+8. **Missing storage path** — Agent class must set `self._storage_path = Path.home() / ".hive" / "agents" / "agent_name"`.
+9. **Missing mcp_servers.json** — Without this, the agent has no tools at runtime.
+10. **Bare `python` command** — Use `"command": "uv"` with args `["run", "python", ...]`.
+
+## Testing Errors
+11. **Using `runner.run()` on forever-alive agents** — `runner.run()` hangs forever because forever-alive agents have no terminal node. Write structural tests instead: validate graph structure, verify node specs, test `AgentRunner.load()` succeeds (no API key needed).
+12. **Stale tests after restructuring** — When changing nodes/edges, update tests to match. Tests referencing old node names will fail.
+13. **Running integration tests without API keys** — Use `pytest.skip()` when credentials are missing.
+14. **Forgetting sys.path setup in conftest.py** — Tests need `exports/` and `core/` on sys.path.
+
+## GCU Errors
+15. **Manually wiring browser tools on event_loop nodes** — Use `node_type="gcu"` which auto-includes browser tools. Do NOT manually list browser tool names.
+16. **Using GCU nodes as regular graph nodes** — GCU nodes are subagents only. They must ONLY appear in `sub_agents=["gcu-node-id"]` and be invoked via `delegate_to_sub_agent()`. Never connect via edges or use as entry/terminal nodes.
+17. **Reusing the same GCU node ID for parallel tasks** — Each concurrent browser task needs a distinct GCU node ID (e.g. `gcu-site-a`, `gcu-site-b`). Two `delegate_to_sub_agent` calls with the same `agent_id` share a browser profile and will interfere with each other's pages.
+18. **Passing `profile=` in GCU tool calls** — Profile isolation for parallel subagents is automatic. The framework injects a unique profile per subagent via an asyncio `ContextVar`. Hardcoding `profile="default"` in a GCU system prompt breaks this isolation.
+
+## Worker Agent Errors
+19. **Adding client-facing intake node to workers** — The queen owns intake. Workers should start with an autonomous processing node. Route worker review/approval through queen escalation instead of direct worker HITL.
+20. **Putting `escalate` or `set_output` in NodeSpec `tools=[]`** — These are synthetic framework tools, auto-injected at runtime. Only list MCP tools from `list_agent_tools()`.

core/framework/agents/queen/reference/file_templates.md

@@ -0,0 +1,569 @@
+# Agent File Templates
+
+Complete code templates for each file in a Hive agent package.
+
+## config.py
+
+```python
+"""Runtime configuration."""
+
+import json
+from dataclasses import dataclass, field
+from pathlib import Path
+
+
+def _load_preferred_model() -> str:
+    """Load preferred model from ~/.hive/configuration.json."""
+    config_path = Path.home() / ".hive" / "configuration.json"
+    if config_path.exists():
+        try:
+            with open(config_path) as f:
+                config = json.load(f)
+            llm = config.get("llm", {})
+            if llm.get("provider") and llm.get("model"):
+                return f"{llm['provider']}/{llm['model']}"
+        except Exception:
+            pass
+    return "anthropic/claude-sonnet-4-20250514"
+
+
+@dataclass
+class RuntimeConfig:
+    model: str = field(default_factory=_load_preferred_model)
+    temperature: float = 0.7
+    max_tokens: int = 40000
+    api_key: str | None = None
+    api_base: str | None = None
+
+
+default_config = RuntimeConfig()
+
+
+@dataclass
+class AgentMetadata:
+    name: str = "My Agent Name"
+    version: str = "1.0.0"
+    description: str = "What this agent does."
+    intro_message: str = "Welcome! What would you like me to do?"
+
+
+metadata = AgentMetadata()
+```
+
+## nodes/__init__.py
+
+```python
+"""Node definitions for My Agent."""
+
+from framework.graph import NodeSpec
+
+# Node 1: Process (autonomous entry node)
+# The queen handles intake and passes structured input via
+# run_agent_with_input(task). NO client-facing intake node.
+# The queen defines input_keys at build time and fills them at run time.
+process_node = NodeSpec(
+    id="process",
+    name="Process",
+    description="Execute the task using available tools",
+    node_type="event_loop",
+    max_node_visits=0,  # Unlimited for forever-alive
+    input_keys=["user_request", "feedback"],
+    output_keys=["results"],
+    nullable_output_keys=["feedback"],  # Only on feedback edge
+    success_criteria="Results are complete and accurate.",
+    system_prompt="""\
+You are a processing agent. Your task is in memory under "user_request". \
+If "feedback" is present, this is a revision — address the feedback.
+
+Work in phases:
+1. Use tools to gather/process data
+2. Analyze results
+3. Call set_output in a SEPARATE turn:
+   - set_output("results", "structured results")
+""",
+    tools=["web_search", "web_scrape", "save_data", "load_data", "list_data_files"],
+)
+
+# Node 2: Handoff (autonomous)
+handoff_node = NodeSpec(
+    id="handoff",
+    name="Handoff",
+    description="Prepare worker results for queen review",
+    node_type="event_loop",
+    client_facing=False,
+    max_node_visits=0,
+    input_keys=["results", "user_request"],
+    output_keys=["next_action", "feedback", "worker_summary"],
+    nullable_output_keys=["feedback", "worker_summary"],
+    success_criteria="Results are packaged for queen decision-making.",
+    system_prompt="""\
+Do NOT talk to the user directly. The queen is the only user interface.
+
+If blocked by tool failures, missing credentials, or unclear constraints, call:
+- escalate(reason, context)
+Then set:
+- set_output("next_action", "escalated")
+- set_output("feedback", "what help is needed")
+
+Otherwise summarize findings for queen and set:
+- set_output("worker_summary", "short summary for queen")
+- set_output("next_action", "done") or set_output("next_action", "revise")
+- set_output("feedback", "what to revise") only when revising
+""",
+    tools=[],
+)
+
+__all__ = ["process_node", "handoff_node"]
+```
+
+## agent.py
+
+```python
+"""Agent graph construction for My Agent."""
+
+from pathlib import Path
+
+from framework.graph import EdgeSpec, EdgeCondition, Goal, SuccessCriterion, Constraint
+from framework.graph.edge import GraphSpec
+from framework.graph.executor import ExecutionResult
+from framework.graph.checkpoint_config import CheckpointConfig
+from framework.llm import LiteLLMProvider
+from framework.runner.tool_registry import ToolRegistry
+from framework.runtime.agent_runtime import AgentRuntime, create_agent_runtime
+from framework.runtime.execution_stream import EntryPointSpec
+
+from .config import default_config, metadata
+from .nodes import process_node, handoff_node
+
+# Goal definition
+goal = Goal(
+    id="my-agent-goal",
+    name="My Agent Goal",
+    description="What this agent achieves.",
+    success_criteria=[
+        SuccessCriterion(id="sc-1", description="...", metric="...", target="...", weight=0.5),
+        SuccessCriterion(id="sc-2", description="...", metric="...", target="...", weight=0.5),
+    ],
+    constraints=[
+        Constraint(id="c-1", description="...", constraint_type="hard", category="quality"),
+    ],
+)
+
+# Node list
+nodes = [process_node, handoff_node]
+
+# Edge definitions
+edges = [
+    EdgeSpec(id="process-to-handoff", source="process", target="handoff",
+             condition=EdgeCondition.ON_SUCCESS, priority=1),
+    # Feedback loop — revise results
+    EdgeSpec(id="handoff-to-process", source="handoff", target="process",
+             condition=EdgeCondition.CONDITIONAL,
+             condition_expr="str(next_action).lower() == 'revise'", priority=2),
+    # Escalation loop — queen injects guidance and worker retries
+    EdgeSpec(id="handoff-escalated", source="handoff", target="process",
+             condition=EdgeCondition.CONDITIONAL,
+             condition_expr="str(next_action).lower() == 'escalated'", priority=3),
+    # Loop back for next task after queen decision
+    EdgeSpec(id="handoff-done", source="handoff", target="process",
+             condition=EdgeCondition.CONDITIONAL,
+             condition_expr="str(next_action).lower() == 'done'", priority=1),
+]
+
+# Graph configuration — entry is the autonomous process node
+# The queen handles intake and passes the task via run_agent_with_input(task)
+entry_node = "process"
+entry_points = {"start": "process"}
+pause_nodes = []
+terminal_nodes = []  # Forever-alive
+
+# Module-level vars read by AgentRunner.load()
+conversation_mode = "continuous"
+identity_prompt = "You are a helpful agent."
+loop_config = {"max_iterations": 100, "max_tool_calls_per_turn": 20, "max_context_tokens": 32000}
+
+
+class MyAgent:
+    def __init__(self, config=None):
+        self.config = config or default_config
+        self.goal = goal
+        self.nodes = nodes
+        self.edges = edges
+        self.entry_node = entry_node  # "process" — autonomous entry
+        self.entry_points = entry_points
+        self.pause_nodes = pause_nodes
+        self.terminal_nodes = terminal_nodes
+        self._graph = None
+        self._agent_runtime = None
+        self._tool_registry = None
+        self._storage_path = None
+
+    def _build_graph(self):
+        return GraphSpec(
+            id="my-agent-graph",
+            goal_id=self.goal.id,
+            version="1.0.0",
+            entry_node=self.entry_node,
+            entry_points=self.entry_points,
+            terminal_nodes=self.terminal_nodes,
+            pause_nodes=self.pause_nodes,
+            nodes=self.nodes,
+            edges=self.edges,
+            default_model=self.config.model,
+            max_tokens=self.config.max_tokens,
+            loop_config=loop_config,
+            conversation_mode=conversation_mode,
+            identity_prompt=identity_prompt,
+        )
+
+    def _setup(self):
+        self._storage_path = Path.home() / ".hive" / "agents" / "my_agent"
+        self._storage_path.mkdir(parents=True, exist_ok=True)
+        self._tool_registry = ToolRegistry()
+        mcp_config = Path(__file__).parent / "mcp_servers.json"
+        if mcp_config.exists():
+            self._tool_registry.load_mcp_config(mcp_config)
+        llm = LiteLLMProvider(model=self.config.model, api_key=self.config.api_key, api_base=self.config.api_base)
+        tools = list(self._tool_registry.get_tools().values())
+        tool_executor = self._tool_registry.get_executor()
+        self._graph = self._build_graph()
+        self._agent_runtime = create_agent_runtime(
+            graph=self._graph, goal=self.goal, storage_path=self._storage_path,
+            entry_points=[EntryPointSpec(id="default", name="Default", entry_node=self.entry_node,
+                                         trigger_type="manual", isolation_level="shared")],
+            llm=llm, tools=tools, tool_executor=tool_executor,
+            checkpoint_config=CheckpointConfig(enabled=True, checkpoint_on_node_complete=True,
+                                                checkpoint_max_age_days=7, async_checkpoint=True),
+        )
+
+    async def start(self):
+        if self._agent_runtime is None:
+            self._setup()
+        if not self._agent_runtime.is_running:
+            await self._agent_runtime.start()
+
+    async def stop(self):
+        if self._agent_runtime and self._agent_runtime.is_running:
+            await self._agent_runtime.stop()
+        self._agent_runtime = None
+
+    async def trigger_and_wait(self, entry_point="default", input_data=None, timeout=None, session_state=None):
+        if self._agent_runtime is None:
+            raise RuntimeError("Agent not started. Call start() first.")
+        return await self._agent_runtime.trigger_and_wait(
+            entry_point_id=entry_point, input_data=input_data or {}, session_state=session_state)
+
+    async def run(self, context, session_state=None):
+        await self.start()
+        try:
+            result = await self.trigger_and_wait("default", context, session_state=session_state)
+            return result or ExecutionResult(success=False, error="Execution timeout")
+        finally:
+            await self.stop()
+
+    def info(self):
+        return {
+            "name": metadata.name, "version": metadata.version, "description": metadata.description,
+            "goal": {"name": self.goal.name, "description": self.goal.description},
+            "nodes": [n.id for n in self.nodes], "edges": [e.id for e in self.edges],
+            "entry_node": self.entry_node, "entry_points": self.entry_points,
+            "terminal_nodes": self.terminal_nodes,
+            "client_facing_nodes": [n.id for n in self.nodes if n.client_facing],
+        }
+
+    def validate(self):
+        """Validate graph wiring and entry-point contract."""
+        errors, warnings = [], []
+        node_ids = {n.id for n in self.nodes}
+        for e in self.edges:
+            if e.source not in node_ids:
+                errors.append(f"Edge {e.id}: source '{e.source}' not found")
+            if e.target not in node_ids:
+                errors.append(f"Edge {e.id}: target '{e.target}' not found")
+        if self.entry_node not in node_ids:
+            errors.append(f"Entry node '{self.entry_node}' not found")
+        for t in self.terminal_nodes:
+            if t not in node_ids:
+                errors.append(f"Terminal node '{t}' not found")
+
+        if not isinstance(self.entry_points, dict):
+            errors.append(
+                "Invalid entry_points: expected dict[str, str] like "
+                "{'start': '<entry-node-id>'}. "
+                f"Got {type(self.entry_points).__name__}. "
+                "Fix agent.py: set entry_points = {'start': '<entry-node-id>'}."
+            )
+        else:
+            if "start" not in self.entry_points:
+                errors.append(
+                    "entry_points must include 'start' mapped to entry_node. "
+                    "Example: {'start': '<entry-node-id>'}."
+                )
+            else:
+                start_node = self.entry_points.get("start")
+                if start_node != self.entry_node:
+                    errors.append(
+                        f"entry_points['start'] points to '{start_node}' "
+                        f"but entry_node is '{self.entry_node}'. Keep these aligned."
+                    )
+
+            for ep_id, nid in self.entry_points.items():
+                if not isinstance(ep_id, str):
+                    errors.append(
+                        f"Invalid entry_points key {ep_id!r} "
+                        f"({type(ep_id).__name__}). Entry point names must be strings."
+                    )
+                    continue
+                if not isinstance(nid, str):
+                    errors.append(
+                        f"Invalid entry_points['{ep_id}']={nid!r} "
+                        f"({type(nid).__name__}). Node ids must be strings."
+                    )
+                    continue
+                if nid not in node_ids:
+                    errors.append(
+                        f"Entry point '{ep_id}' references unknown node '{nid}'. "
+                        f"Known nodes: {sorted(node_ids)}"
+                    )
+
+        return {"valid": len(errors) == 0, "errors": errors, "warnings": warnings}
+
+
+default_agent = MyAgent()
+```
+
+## triggers.json — Timer and Webhook Triggers
+
+When an agent needs timers, webhooks, or event-driven triggers, create a
+`triggers.json` file in the agent's directory (alongside `agent.py`).
+The queen loads these at session start and the user can manage them via
+the `set_trigger` / `remove_trigger` tools at runtime.
+
+```json
+[
+  {
+    "id": "daily-check",
+    "name": "Daily Check",
+    "trigger_type": "timer",
+    "trigger_config": {"cron": "0 9 * * *"},
+    "task": "Run the daily check process"
+  },
+  {
+    "id": "scheduled-check",
+    "name": "Scheduled Check",
+    "trigger_type": "timer",
+    "trigger_config": {"interval_minutes": 20},
+    "task": "Run the scheduled check"
+  },
+  {
+    "id": "webhook-event",
+    "name": "Webhook Event Handler",
+    "trigger_type": "webhook",
+    "trigger_config": {"event_types": ["webhook_received"]},
+    "task": "Process incoming webhook event"
+  }
+]
+```
+
+**Key rules for triggers.json:**
+- Valid trigger_types: `timer`, `webhook`
+- Timer trigger_config (cron): `{"cron": "0 9 * * *"}` — standard 5-field cron expression
+- Timer trigger_config (interval): `{"interval_minutes": float}`
+- Each trigger must have a unique `id`
+- The `task` field describes what the worker should do when the trigger fires
+- Triggers are persisted back to `triggers.json` when modified via queen tools
+
+## __init__.py
+
+**CRITICAL:** The runner imports the package (`__init__.py`) and reads ALL module-level
+variables via `getattr()`. Every variable defined in `agent.py` that the runner needs
+MUST be re-exported here. Missing exports cause silent failures (variables default to
+`None` or `{}`), leading to "must define goal, nodes, edges" errors or graph validation
+failures like "node X is unreachable".
+
+```python
+"""My Agent — description."""
+
+from .agent import (
+    MyAgent,
+    default_agent,
+    goal,
+    nodes,
+    edges,
+    entry_node,
+    entry_points,
+    pause_nodes,
+    terminal_nodes,
+    conversation_mode,
+    identity_prompt,
+    loop_config,
+)
+from .config import default_config, metadata
+
+__all__ = [
+    "MyAgent",
+    "default_agent",
+    "goal",
+    "nodes",
+    "edges",
+    "entry_node",
+    "entry_points",
+    "pause_nodes",
+    "terminal_nodes",
+    "conversation_mode",
+    "identity_prompt",
+    "loop_config",
+    "default_config",
+    "metadata",
+]
+```
+
+## __main__.py
+
+```python
+"""CLI entry point for My Agent."""
+
+import asyncio, json, logging, sys
+import click
+from .agent import default_agent, MyAgent
+
+
+def setup_logging(verbose=False, debug=False):
+    if debug: level, fmt = logging.DEBUG, "%(asctime)s %(name)s: %(message)s"
+    elif verbose: level, fmt = logging.INFO, "%(message)s"
+    else: level, fmt = logging.WARNING, "%(levelname)s: %(message)s"
+    logging.basicConfig(level=level, format=fmt, stream=sys.stderr)
+
+
+@click.group()
+@click.version_option(version="1.0.0")
+def cli():
+    """My Agent — description."""
+    pass
+
+
+@cli.command()
+@click.option("--topic", "-t", required=True)
+@click.option("--verbose", "-v", is_flag=True)
+def run(topic, verbose):
+    """Execute the agent."""
+    setup_logging(verbose=verbose)
+    result = asyncio.run(default_agent.run({"topic": topic}))
+    click.echo(json.dumps({"success": result.success, "output": result.output}, indent=2, default=str))
+    sys.exit(0 if result.success else 1)
+
+
+@cli.command()
+def tui():
+    """Launch TUI dashboard."""
+    from pathlib import Path
+    from framework.tui.app import AdenTUI
+    from framework.llm import LiteLLMProvider
+    from framework.runner.tool_registry import ToolRegistry
+    from framework.runtime.agent_runtime import create_agent_runtime
+    from framework.runtime.execution_stream import EntryPointSpec
+
+    async def run_tui():
+        agent = MyAgent()
+        agent._tool_registry = ToolRegistry()
+        storage = Path.home() / ".hive" / "agents" / "my_agent"
+        storage.mkdir(parents=True, exist_ok=True)
+        mcp_cfg = Path(__file__).parent / "mcp_servers.json"
+        if mcp_cfg.exists(): agent._tool_registry.load_mcp_config(mcp_cfg)
+        llm = LiteLLMProvider(model=agent.config.model, api_key=agent.config.api_key, api_base=agent.config.api_base)
+        runtime = create_agent_runtime(
+            graph=agent._build_graph(), goal=agent.goal, storage_path=storage,
+            entry_points=[EntryPointSpec(id="start", name="Start", entry_node="process", trigger_type="manual", isolation_level="isolated")],
+            llm=llm, tools=list(agent._tool_registry.get_tools().values()), tool_executor=agent._tool_registry.get_executor())
+        await runtime.start()
+        try:
+            app = AdenTUI(runtime)
+            await app.run_async()
+        finally:
+            await runtime.stop()
+    asyncio.run(run_tui())
+
+
+@cli.command()
+def info():
+    """Show agent info."""
+    data = default_agent.info()
+    click.echo(f"Agent: {data['name']}\nVersion: {data['version']}\nDescription: {data['description']}")
+    click.echo(f"Nodes: {', '.join(data['nodes'])}\nClient-facing: {', '.join(data['client_facing_nodes'])}")
+
+
+@cli.command()
+def validate():
+    """Validate agent structure."""
+    v = default_agent.validate()
+    if v["valid"]: click.echo("Agent is valid")
+    else:
+        click.echo("Errors:")
+        for e in v["errors"]: click.echo(f"  {e}")
+    sys.exit(0 if v["valid"] else 1)
+
+
+if __name__ == "__main__":
+    cli()
+```
+
+## mcp_servers.json
+
+> **Auto-generated.** `initialize_and_build_agent` creates this file with hive-tools
+> as the default. Only edit manually to add additional MCP servers.
+
+```json
+{
+  "hive-tools": {
+    "transport": "stdio",
+    "command": "uv",
+    "args": ["run", "python", "mcp_server.py", "--stdio"],
+    "cwd": "../../tools",
+    "description": "Hive tools MCP server"
+  }
+}
+```
+
+**CRITICAL FORMAT RULES:**
+- NO `"mcpServers"` wrapper (flat dict, not nested)
+- `cwd` MUST be `"../../tools"` (relative from `exports/AGENT_NAME/` to `tools/`)
+- `command` MUST be `"uv"` with `"args": ["run", "python", ...]` (NOT bare `"python"`)
+
+## tests/conftest.py
+
+```python
+"""Test fixtures."""
+
+import sys
+from pathlib import Path
+
+import pytest
+
+_repo_root = Path(__file__).resolve().parents[3]
+for _p in ["exports", "core"]:
+    _path = str(_repo_root / _p)
+    if _path not in sys.path:
+        sys.path.insert(0, _path)
+
+AGENT_PATH = str(Path(__file__).resolve().parents[1])
+
+
+@pytest.fixture(scope="session")
+def agent_module():
+    """Import the agent package for structural validation."""
+    import importlib
+    return importlib.import_module(Path(AGENT_PATH).name)
+
+
+@pytest.fixture(scope="session")
+def runner_loaded():
+    """Load the agent through AgentRunner (structural only, no LLM needed)."""
+    from framework.runner.runner import AgentRunner
+    return AgentRunner.load(AGENT_PATH)
+```
+
+## entry_points Format
+
+MUST be: `{"start": "first-node-id"}`
+NOT: `{"first-node-id": ["input_keys"]}` (WRONG)
+NOT: `{"first-node-id"}` (WRONG — this is a set)

core/framework/agents/queen/reference/framework_guide.md

@@ -0,0 +1,306 @@
+# Hive Agent Framework — Condensed Reference
+
+## Architecture
+
+Agents are Python packages in `exports/`:
+```
+exports/my_agent/
+├── __init__.py          # MUST re-export ALL module-level vars from agent.py
+├── __main__.py          # CLI (run, tui, info, validate, shell)
+├── agent.py             # Graph construction (goal, edges, agent class)
+├── config.py            # Runtime config
+├── nodes/__init__.py    # Node definitions (NodeSpec)
+├── mcp_servers.json     # MCP tool server config
+└── tests/               # pytest tests
+```
+
+## Agent Loading Contract
+
+`AgentRunner.load()` imports the package (`__init__.py`) and reads these
+module-level variables via `getattr()`:
+
+| Variable | Required | Default if missing | Consequence |
+|----------|----------|--------------------|-------------|
+| `goal` | YES | `None` | **FATAL** — "must define goal, nodes, edges" |
+| `nodes` | YES | `None` | **FATAL** — same error |
+| `edges` | YES | `None` | **FATAL** — same error |
+| `entry_node` | no | `nodes[0].id` | Probably wrong node |
+| `entry_points` | no | `{}` | **Nodes unreachable** — validation fails |
+| `terminal_nodes` | **YES** | `[]` | **FATAL** — graph must have at least one terminal node |
+| `pause_nodes` | no | `[]` | OK |
+| `conversation_mode` | no | not passed | Isolated mode (no context carryover) |
+| `identity_prompt` | no | not passed | No agent-level identity |
+| `loop_config` | no | `{}` | No iteration limits |
+| `triggers.json` (file) | no | not present | No triggers (timers, webhooks) |
+
+**CRITICAL:** `__init__.py` MUST import and re-export ALL of these from
+`agent.py`. Missing exports silently fall back to defaults, causing
+hard-to-debug failures.
+
+**Why `default_agent.validate()` is NOT sufficient:**
+`validate()` checks the agent CLASS's internal graph (self.nodes, self.edges).
+These are always correct because the constructor references agent.py's module
+vars directly. But `AgentRunner.load()` reads from the PACKAGE (`__init__.py`),
+not the class. So `validate()` passes while `AgentRunner.load()` fails.
+Always test with `AgentRunner.load("exports/{name}")` — this is the same
+code path the TUI and `hive run` use.
+
+## Goal
+
+Defines success criteria and constraints:
+```python
+goal = Goal(
+    id="kebab-case-id",
+    name="Display Name",
+    description="What the agent does",
+    success_criteria=[
+        SuccessCriterion(id="sc-id", description="...", metric="...", target="...", weight=0.25),
+    ],
+    constraints=[
+        Constraint(id="c-id", description="...", constraint_type="hard", category="quality"),
+    ],
+)
+```
+- 3-5 success criteria, weights sum to 1.0
+- 1-5 constraints (hard/soft, categories: quality, accuracy, interaction, functional)
+
+## NodeSpec Fields
+
+| Field | Type | Default | Description |
+|-------|------|---------|-------------|
+| id | str | required | kebab-case identifier |
+| name | str | required | Display name |
+| description | str | required | What the node does |
+| node_type | str | required | `"event_loop"` or `"gcu"` (browser automation — see GCU Guide appendix) |
+| input_keys | list[str] | required | Memory keys this node reads |
+| output_keys | list[str] | required | Memory keys this node writes via set_output |
+| system_prompt | str | "" | LLM instructions |
+| tools | list[str] | [] | Tool names from MCP servers |
+| client_facing | bool | False | Deprecated compatibility field. Queen interactivity is implicit; workers should escalate instead |
+| nullable_output_keys | list[str] | [] | Keys that may remain unset |
+| max_node_visits | int | 0 | 0=unlimited (default); >1 for one-shot feedback loops |
+| max_retries | int | 3 | Retries on failure |
+| success_criteria | str | "" | Natural language for judge evaluation |
+
+## EdgeSpec Fields
+
+| Field | Type | Description |
+|-------|------|-------------|
+| id | str | kebab-case identifier |
+| source | str | Source node ID |
+| target | str | Target node ID |
+| condition | EdgeCondition | ON_SUCCESS, ON_FAILURE, ALWAYS, CONDITIONAL |
+| condition_expr | str | Python expression evaluated against memory (for CONDITIONAL) |
+| priority | int | Positive=forward (evaluated first), negative=feedback (loop-back) |
+
+## Key Patterns
+
+### STEP 1/STEP 2 (Client-Facing Nodes)
+```
+**STEP 1 — Respond to the user (text only, NO tool calls):**
+[Present information, ask questions]
+
+**STEP 2 — After the user responds, call set_output:**
+- set_output("key", "value based on user response")
+```
+This prevents premature set_output before user interaction.
+
+### Fewer, Richer Nodes (CRITICAL)
+
+**Hard limit: 3-6 nodes for most agents.** Never exceed 6 unless the user
+explicitly requests a complex multi-phase pipeline.
+
+Each node boundary serializes outputs to the shared buffer and **destroys** all
+in-context information: tool call results, intermediate reasoning, conversation
+history. A research node that searches, fetches, and analyzes in ONE node keeps
+all source material in its conversation context. Split across 3 nodes, each
+downstream node only sees the serialized summary string.
+
+**Decision framework — merge unless ANY of these apply:**
+1. **Client-facing boundary** — Autonomous and client-facing work MUST be
+   separate nodes (different interaction models)
+2. **Disjoint tool sets** — If tools are fundamentally different (e.g., web
+   search vs database), separate nodes make sense
+3. **Parallel execution** — Fan-out branches must be separate nodes
+
+**Red flags that you have too many nodes:**
+- A node with 0 tools (pure LLM reasoning) → merge into predecessor/successor
+- A node that sets only 1 trivial output → collapse into predecessor
+- Multiple consecutive autonomous nodes → combine into one rich node
+- A "report" node that presents analysis → merge into the client-facing node
+- A "confirm" or "schedule" node that doesn't call any external service → remove
+
+**Typical agent structure (2 nodes):**
+```
+process (autonomous) ←→ review (queen-mediated)
+```
+The queen owns intake — she gathers requirements from the user, then
+passes structured input via `run_agent_with_input(task)`. When building
+the agent, design the entry node's `input_keys` to match what the queen
+will provide at run time. Worker agents should NOT have a client-facing
+intake node. Mid-execution review/approval should happen through queen
+escalation rather than direct worker HITL.
+
+For simpler agents, just 1 autonomous node:
+```
+process (autonomous) — loops back to itself
+```
+
+### nullable_output_keys
+For inputs that only arrive on certain edges:
+```python
+research_node = NodeSpec(
+    input_keys=["brief", "feedback"],
+    nullable_output_keys=["feedback"],  # Only present on feedback edge
+    max_node_visits=3,
+)
+```
+
+### Mutually Exclusive Outputs
+For routing decisions:
+```python
+review_node = NodeSpec(
+    output_keys=["approved", "feedback"],
+    nullable_output_keys=["approved", "feedback"],  # Node sets one or the other
+)
+```
+
+### Continuous Loop Pattern
+Mark the primary event_loop node as terminal: `terminal_nodes=["process"]`.
+The node has `output_keys` and can complete when the agent finishes its work.
+Use `conversation_mode="continuous"` to preserve context across transitions.
+
+### set_output
+- Synthetic tool injected by framework
+- Call separately from real tool calls (separate turn)
+- `set_output("key", "value")` stores to the shared buffer
+
+## Edge Conditions
+
+| Condition | When |
+|-----------|------|
+| ON_SUCCESS | Node completed successfully |
+| ON_FAILURE | Node failed |
+| ALWAYS | Unconditional |
+| CONDITIONAL | condition_expr evaluates to True against memory |
+
+condition_expr examples:
+- `"needs_more_research == True"`
+- `"str(next_action).lower() == 'new_agent'"`
+- `"feedback is not None"`
+
+## Graph Lifecycle
+
+| Pattern | terminal_nodes | When |
+|---------|---------------|------|
+| **Continuous loop** | `["node-with-output-keys"]` | **DEFAULT for all agents** |
+| Linear | `["last-node"]` | One-shot/batch agents |
+
+**Every graph must have at least one terminal node.** Terminal nodes
+define where execution ends. For interactive agents that loop continuously,
+mark the primary event_loop node as terminal (it has `output_keys` and can
+complete at any point). The framework default for `max_node_visits` is 0
+(unbounded), so nodes work correctly in continuous loops without explicit
+override. Only set `max_node_visits > 0` in one-shot agents with feedback loops.
+Every node must have at least one outgoing edge — no dead ends.
+
+## Continuous Conversation Mode
+
+`conversation_mode` has ONLY two valid states:
+- `"continuous"` — recommended for interactive agents
+- Omit entirely — isolated per-node conversations (each node starts fresh)
+
+**INVALID values** (do NOT use): `"client_facing"`, `"interactive"`,
+`"adaptive"`, `"shared"`. These do not exist in the framework.
+
+When `conversation_mode="continuous"`:
+- Same conversation thread carries across node transitions
+- Layered system prompts: identity (agent-level) + narrative + focus (per-node)
+- Transition markers inserted at boundaries
+- Compaction happens opportunistically at phase transitions
+
+## loop_config
+
+Only three valid keys:
+```python
+loop_config = {
+    "max_iterations": 100,          # Max LLM turns per node visit
+    "max_tool_calls_per_turn": 20,  # Max tool calls per LLM response
+    "max_context_tokens": 32000,    # Triggers conversation compaction
+}
+```
+**INVALID keys** (do NOT use): `"strategy"`, `"mode"`, `"timeout"`,
+`"temperature"`. These are silently ignored or cause errors.
+
+## Data Tools (Spillover)
+
+For large data that exceeds context:
+- `save_data(filename, data)` — Write to session data dir
+- `load_data(filename, offset, limit)` — Read with pagination
+- `list_data_files()` — List files
+- `serve_file_to_user(filename, label)` — Clickable file:// URI
+
+`data_dir` is auto-injected by framework — LLM never sees it.
+
+## Fan-Out / Fan-In
+
+Multiple ON_SUCCESS edges from same source → parallel execution via asyncio.gather().
+- Parallel nodes must have disjoint output_keys
+- Only one branch may have client_facing nodes
+- Fan-in node gets all outputs in the shared buffer
+
+## Judge System
+
+- **Implicit** (default): ACCEPTs when LLM finishes with no tool calls and all required outputs set
+- **SchemaJudge**: Validates against Pydantic model
+- **Custom**: Implement `evaluate(context) -> JudgeVerdict`
+
+Judge is the SOLE acceptance mechanism — no ad-hoc framework gating.
+
+## Triggers (Timers, Webhooks)
+
+For agents that react to external events, create a `triggers.json` file
+in the agent's export directory:
+
+```json
+[
+  {
+    "id": "daily-check",
+    "name": "Daily Check",
+    "trigger_type": "timer",
+    "trigger_config": {"cron": "0 9 * * *"},
+    "task": "Run the daily check process"
+  }
+]
+```
+
+### Key Fields
+- `trigger_type`: `"timer"` or `"webhook"`
+- `trigger_config`: `{"cron": "0 9 * * *"}` or `{"interval_minutes": 20}`
+- `task`: describes what the worker should do when the trigger fires
+- Triggers can also be created/removed at runtime via `set_trigger` / `remove_trigger` queen tools
+
+## Tool Discovery
+
+Do NOT rely on a static tool list — it will be outdated. Always call
+`list_agent_tools()` with NO arguments first to see ALL available tools.
+Only use `group=` or `output_schema=` as follow-up calls after seeing the
+full list.
+
+```
+list_agent_tools()                            # ALWAYS call this first
+list_agent_tools(group="gmail", output_schema="full")  # then drill into a category
+list_agent_tools("exports/my_agent/mcp_servers.json")  # specific agent's tools
+```
+
+After building, run `validate_agent_package("{name}")` to check everything at once.
+
+Common tool categories (verify via list_agent_tools):
+- **Web**: search, scrape, PDF
+- **Data**: save/load/append/list data files, serve to user
+- **File**: view, write, replace, diff, list, grep
+- **Communication**: email, gmail, slack, telegram
+- **CRM**: hubspot, apollo, calcom
+- **GitHub**: stargazers, user profiles, repos
+- **Vision**: image analysis
+- **Time**: current time

core/framework/agents/queen/reference/gcu_guide.md

@@ -0,0 +1,158 @@
+# GCU Browser Automation Guide
+
+## When to Use GCU Nodes
+
+Use `node_type="gcu"` when:
+- The user's workflow requires **navigating real websites** (scraping, form-filling, social media interaction, testing web UIs)
+- The task involves **dynamic/JS-rendered pages** that `web_scrape` cannot handle (SPAs, infinite scroll, login-gated content)
+- The agent needs to **interact with a website** — clicking, typing, scrolling, selecting, uploading files
+
+Do NOT use GCU for:
+- Static content that `web_scrape` handles fine
+- API-accessible data (use the API directly)
+- PDF/file processing
+- Anything that doesn't require a browser UI
+
+## What GCU Nodes Are
+
+- `node_type="gcu"` — a declarative enhancement over `event_loop`
+- Framework auto-prepends browser best-practices system prompt
+- Framework auto-includes all 31 browser tools from `gcu-tools` MCP server
+- Same underlying `EventLoopNode` class — no new imports needed
+- `tools=[]` is correct — tools are auto-populated at runtime
+
+## GCU Architecture Pattern  
+
+GCU nodes are **subagents** — invoked via `delegate_to_sub_agent()`, not connected via edges.
+
+- Primary nodes (`event_loop`, client-facing) orchestrate; GCU nodes do browser work
+- Parent node declares `sub_agents=["gcu-node-id"]` and calls `delegate_to_sub_agent(agent_id="gcu-node-id", task="...")`
+- GCU nodes set `max_node_visits=1` (single execution per delegation), `client_facing=False`
+- GCU nodes use `output_keys=["result"]` and return structured JSON via `set_output("result", ...)`
+
+## GCU Node Definition Template
+
+```python
+gcu_browser_node = NodeSpec(
+    id="gcu-browser-worker",
+    name="Browser Worker",
+    description="Browser subagent that does X.",
+    node_type="gcu",
+    client_facing=False,
+    max_node_visits=1,
+    input_keys=[],
+    output_keys=["result"],
+    tools=[],  # Auto-populated with all browser tools
+    system_prompt="""\
+You are a browser agent. Your job: [specific task].
+
+## Workflow
+1. browser_start (only if no browser is running yet)
+2. browser_open(url=TARGET_URL) — note the returned targetId
+3. browser_snapshot to read the page
+4. [task-specific steps]
+5. set_output("result", JSON)
+
+## Output format
+set_output("result", JSON) with:
+- [field]: [type and description]
+""",
+)
+```
+
+## Parent Node Template (orchestrating GCU subagents)
+
+```python
+orchestrator_node = NodeSpec(
+    id="orchestrator",
+    ...
+    node_type="event_loop",
+    sub_agents=["gcu-browser-worker"],
+    system_prompt="""\
+...
+delegate_to_sub_agent(
+    agent_id="gcu-browser-worker",
+    task="Navigate to [URL]. Do [specific task]. Return JSON with [fields]."
+)
+...
+""",
+    tools=[],  # Orchestrator doesn't need browser tools
+)
+```
+
+## mcp_servers.json with GCU
+
+```json
+{
+  "hive-tools": { ... },
+  "gcu-tools": {
+    "transport": "stdio",
+    "command": "uv",
+    "args": ["run", "python", "-m", "gcu.server", "--stdio"],
+    "cwd": "../../tools",
+    "description": "GCU tools for browser automation"
+  }
+}
+```
+
+Note: `gcu-tools` is auto-added if any node uses `node_type="gcu"`, but including it explicitly is fine.
+
+## GCU System Prompt Best Practices
+
+Key rules to bake into GCU node prompts:
+
+- Prefer `browser_snapshot` over `browser_get_text("body")` — compact accessibility tree vs 100KB+ raw HTML
+- Always `browser_wait` after navigation
+- Use large scroll amounts (~2000-5000) for lazy-loaded content
+- For spillover files, use `run_command` with grep, not `read_file`
+- If auth wall detected, report immediately — don't attempt login
+- Keep tool calls per turn ≤10
+- Tab isolation: when browser is already running, use `browser_open(background=true)` and pass `target_id` to every call
+
+## Multiple Concurrent GCU Subagents
+
+When a task can be parallelized across multiple sites or profiles, declare a distinct GCU
+node for each and invoke them all in the same LLM turn.  The framework batches all
+`delegate_to_sub_agent` calls made in one turn and runs them with `asyncio.gather`, so
+they execute concurrently — not sequentially.
+
+**Each GCU subagent automatically gets its own isolated browser context** — no `profile=`
+argument is needed in tool calls.  The framework derives a unique profile from the subagent's
+node ID and instance counter and injects it via an asyncio `ContextVar` before the subagent
+runs.
+
+### Example: three sites in parallel
+
+```python
+# Three distinct GCU nodes
+gcu_site_a = NodeSpec(id="gcu-site-a", node_type="gcu", ...)
+gcu_site_b = NodeSpec(id="gcu-site-b", node_type="gcu", ...)
+gcu_site_c = NodeSpec(id="gcu-site-c", node_type="gcu", ...)
+
+orchestrator = NodeSpec(
+    id="orchestrator",
+    node_type="event_loop",
+    sub_agents=["gcu-site-a", "gcu-site-b", "gcu-site-c"],
+    system_prompt="""\
+Call all three subagents in a single response to run them in parallel:
+  delegate_to_sub_agent(agent_id="gcu-site-a", task="Scrape prices from site A")
+  delegate_to_sub_agent(agent_id="gcu-site-b", task="Scrape prices from site B")
+  delegate_to_sub_agent(agent_id="gcu-site-c", task="Scrape prices from site C")
+""",
+)
+```
+
+**Rules:**
+- Use distinct node IDs for each concurrent task — sharing an ID shares the browser context.
+- The GCU node prompts do not need to mention `profile=`; isolation is automatic.
+- Cleanup is automatic at session end, but GCU nodes can call `browser_stop()` explicitly
+  if they want to release resources mid-run.
+
+## GCU Anti-Patterns
+
+- Using `browser_screenshot` to read text (use `browser_snapshot` instead; screenshots are for visual context only)
+- Re-navigating after scrolling (resets scroll position)
+- Attempting login on auth walls
+- Forgetting `target_id` in multi-tab scenarios
+- Putting browser tools directly on `event_loop` nodes instead of using GCU subagent pattern
+- Making GCU nodes `client_facing=True` (they should be autonomous subagents)

core/framework/agents/queen/reflection_agent.py

@@ -0,0 +1,783 @@
+"""Reflect agent — background memory extraction for queen and worker memory.
+
+A lightweight side agent that runs after each queen LLM turn.  It
+inspects recent conversation messages (cursor-based incremental
+processing) and extracts learnings into individual memory files.
+
+Two reflection types:
+  - **Short reflection**: every queen turn. Distills learnings. Nudged
+    toward a 2-turn pattern (batch reads → batch writes).
+  - **Long reflection**: every 5 short reflections, on CONTEXT_COMPACTED,
+    and at session end.  Organises, deduplicates, trims holistically.
+
+The agent has restricted tool access: it can only read/write/delete
+memory files in ``~/.hive/queen/memories/`` and list them.
+
+Concurrency: an ``asyncio.Lock`` prevents overlapping runs.  If a
+trigger fires while a reflection is already active the event is skipped
+(cursor hasn't advanced, so messages will be reconsidered next time).
+"""
+
+from __future__ import annotations
+
+import asyncio
+import json
+import logging
+import re
+import traceback
+from datetime import datetime
+from pathlib import Path
+from typing import Any
+
+from framework.agents.queen.queen_memory_v2 import (
+    MAX_FILE_SIZE_BYTES,
+    MAX_FILES,
+    MEMORY_DIR,
+    MEMORY_FRONTMATTER_EXAMPLE,
+    MEMORY_TYPES,
+    build_diary_document,
+    diary_filename,
+    format_memory_manifest,
+    parse_frontmatter,
+    read_conversation_parts,
+    scan_memory_files,
+)
+from framework.llm.provider import LLMResponse, Tool
+
+logger = logging.getLogger(__name__)
+
+# ---------------------------------------------------------------------------
+# Reflection tool definitions (internal — not in queen's main registry)
+# ---------------------------------------------------------------------------
+
+_REFLECTION_TOOLS: list[Tool] = [
+    Tool(
+        name="list_memory_files",
+        description=(
+            "List all memory files with their type, name, age, and description. "
+            "Returns a text manifest — one line per file."
+        ),
+        parameters={
+            "type": "object",
+            "properties": {},
+            "additionalProperties": False,
+        },
+    ),
+    Tool(
+        name="read_memory_file",
+        description="Read the full content of a memory file by filename.",
+        parameters={
+            "type": "object",
+            "properties": {
+                "filename": {
+                    "type": "string",
+                    "description": "The filename (e.g. 'user-prefers-dark-mode.md').",
+                },
+            },
+            "required": ["filename"],
+            "additionalProperties": False,
+        },
+    ),
+    Tool(
+        name="write_memory_file",
+        description=(
+            "Create or overwrite a memory file.  Content should include YAML "
+            "frontmatter (name, description, type) followed by the memory body.  "
+            f"Max file size: {MAX_FILE_SIZE_BYTES} bytes.  Max files: {MAX_FILES}."
+        ),
+        parameters={
+            "type": "object",
+            "properties": {
+                "filename": {
+                    "type": "string",
+                    "description": "Filename ending in .md (e.g. 'user-prefers-dark-mode.md').",
+                },
+                "content": {
+                    "type": "string",
+                    "description": "Full file content including frontmatter.",
+                },
+            },
+            "required": ["filename", "content"],
+            "additionalProperties": False,
+        },
+    ),
+    Tool(
+        name="delete_memory_file",
+        description=(
+            "Delete a memory file by filename.  Use during long "
+            "reflection to prune stale or redundant memories."
+        ),
+        parameters={
+            "type": "object",
+            "properties": {
+                "filename": {
+                    "type": "string",
+                    "description": "The filename to delete.",
+                },
+            },
+            "required": ["filename"],
+            "additionalProperties": False,
+        },
+    ),
+]
+
+
+def _safe_memory_path(filename: str, memory_dir: Path) -> Path:
+    """Resolve *filename* inside *memory_dir*, raising if it escapes."""
+    if not filename or filename.strip() != filename:
+        raise ValueError(f"Invalid filename: {filename!r}")
+    if "/" in filename or "\\" in filename or ".." in filename:
+        raise ValueError(f"Invalid filename: path components not allowed: {filename!r}")
+    candidate = (memory_dir / filename).resolve()
+    root = memory_dir.resolve()
+    if not candidate.is_relative_to(root):
+        raise ValueError(f"Path escapes memory directory: {filename!r}")
+    return candidate
+
+
+# Memory types that workers are NOT allowed to write.
+_WORKER_BLOCKED_TYPES: frozenset[str] = frozenset(
+    {"environment", "technique", "reference", "diary", "goal"}
+)
+
+
+def _inject_last_modified_by(content: str, caller: str) -> str:
+    """Inject or update ``last_modified_by`` in frontmatter."""
+    m = re.match(r"^---\s*\n(.*?)\n---", content, re.DOTALL)
+    if not m:
+        return content
+    fm_body = m.group(1)
+    # Remove existing last_modified_by line if present.
+    fm_lines = [
+        ln for ln in fm_body.splitlines()
+        if not ln.strip().lower().startswith("last_modified_by")
+    ]
+    fm_lines.append(f"last_modified_by: {caller}")
+    new_fm = "\n".join(fm_lines)
+    return f"---\n{new_fm}\n---{content[m.end():]}"
+
+
+def _execute_tool(name: str, args: dict[str, Any], memory_dir: Path, caller: str) -> str:
+    """Execute a reflection tool synchronously.  Returns the result string."""
+    if name == "list_memory_files":
+        files = scan_memory_files(memory_dir)
+        logger.debug("reflect: tool list_memory_files → %d files", len(files))
+        if not files:
+            return "(no memory files yet)"
+        return format_memory_manifest(files)
+
+    if name == "read_memory_file":
+        filename = args.get("filename", "")
+        try:
+            path = _safe_memory_path(filename, memory_dir)
+        except ValueError as exc:
+            return f"ERROR: {exc}"
+        if not path.exists() or not path.is_file():
+            return f"ERROR: File not found: {filename}"
+        try:
+            return path.read_text(encoding="utf-8")
+        except OSError as e:
+            return f"ERROR: {e}"
+
+    if name == "write_memory_file":
+        filename = args.get("filename", "")
+        content = args.get("content", "")
+        if not filename.endswith(".md"):
+            return "ERROR: Filename must end with .md"
+        # Enforce caller-based type restrictions.
+        fm = parse_frontmatter(content)
+        mem_type = (fm.get("type") or "").strip().lower()
+        if caller == "worker" and mem_type in _WORKER_BLOCKED_TYPES:
+            return (
+                f"ERROR: Workers cannot write memory type '{mem_type}'. "
+                f"Blocked types for workers: {', '.join(sorted(_WORKER_BLOCKED_TYPES))}."
+            )
+        # Inject last_modified_by into frontmatter.
+        content = _inject_last_modified_by(content, caller)
+        # Enforce file size limit.
+        if len(content.encode("utf-8")) > MAX_FILE_SIZE_BYTES:
+            return f"ERROR: Content exceeds {MAX_FILE_SIZE_BYTES} byte limit."
+        # Enforce file cap (only for new files).
+        try:
+            path = _safe_memory_path(filename, memory_dir)
+        except ValueError as exc:
+            return f"ERROR: {exc}"
+        if not path.exists():
+            existing = list(memory_dir.glob("*.md"))
+            if len(existing) >= MAX_FILES:
+                return f"ERROR: File cap reached ({MAX_FILES}).  Delete a file first."
+        memory_dir.mkdir(parents=True, exist_ok=True)
+        path.write_text(content, encoding="utf-8")
+        logger.debug("reflect: tool write_memory_file [%s] → %s (%d chars)", caller, filename, len(content))
+        return f"Wrote {filename} ({len(content)} chars)."
+
+    if name == "delete_memory_file":
+        filename = args.get("filename", "")
+        try:
+            path = _safe_memory_path(filename, memory_dir)
+        except ValueError as exc:
+            return f"ERROR: {exc}"
+        if not path.exists():
+            return f"ERROR: File not found: {filename}"
+        path.unlink()
+        logger.debug("reflect: tool delete_memory_file [%s] → %s", caller, filename)
+        return f"Deleted {filename}."
+
+    return f"ERROR: Unknown tool: {name}"
+
+
+# ---------------------------------------------------------------------------
+# Mini event loop
+# ---------------------------------------------------------------------------
+
+_MAX_TURNS = 5
+
+
+async def _reflection_loop(
+    llm: Any,
+    system: str,
+    user_msg: str,
+    memory_dir: Path,
+    caller: str,
+    max_turns: int = _MAX_TURNS,
+) -> tuple[bool, list[str], str]:
+    """Run a mini tool-use loop: LLM → tool calls → repeat.
+
+    Hard cap of *max_turns* iterations.  Prompt nudges the LLM toward a
+    2-turn pattern (batch reads in turn 1, batch writes in turn 2).
+
+    Returns a tuple of (success, changed_files, last_text) where *success*
+    is ``True`` if the loop completed without LLM errors, *changed_files*
+    lists filenames that were written or deleted, and *last_text* is the
+    final assistant text (useful as a skip-reason when no files changed).
+    """
+    messages: list[dict[str, Any]] = [{"role": "user", "content": user_msg}]
+    changed_files: list[str] = []
+    last_text: str = ""
+    logger.debug("reflect: starting loop (caller=%s, max %d turns)", caller, max_turns)
+
+    for _turn in range(max_turns):
+        # Log what we're sending to the LLM.
+        user_content = messages[-1].get("content", "") if messages else ""
+        preview = user_content[:300] if isinstance(user_content, str) else str(user_content)[:300]
+        logger.debug(
+            "reflect: turn %d — sending %d messages to LLM, last msg role=%s, preview=%s",
+            _turn, len(messages), messages[-1].get("role", "?") if messages else "?", preview,
+        )
+
+        try:
+            resp: LLMResponse = await llm.acomplete(
+                messages=messages,
+                system=system,
+                tools=_REFLECTION_TOOLS,
+                max_tokens=2048,
+            )
+        except Exception:
+            logger.warning("reflect: LLM call failed", exc_info=True)
+            return False, changed_files, last_text
+
+        # Build assistant message.
+        tool_calls_raw: list[dict[str, Any]] = []
+        if resp.raw_response and isinstance(resp.raw_response, dict):
+            tool_calls_raw = resp.raw_response.get("tool_calls", [])
+
+        # Log the full LLM response for debugging.
+        raw_keys = list(resp.raw_response.keys()) if isinstance(resp.raw_response, dict) else type(resp.raw_response).__name__
+        logger.debug(
+            "reflect: turn %d — LLM response: content=%r (len=%d), stop_reason=%s, "
+            "tool_calls=%d, model=%s, tokens=%d/%d, raw_keys=%s",
+            _turn, (resp.content or "")[:200], len(resp.content or ""),
+            resp.stop_reason, len(tool_calls_raw), resp.model,
+            resp.input_tokens, resp.output_tokens, raw_keys,
+        )
+        # Accumulate non-empty text across turns so we don't lose a reason
+        # given alongside tool calls on an earlier turn.
+        turn_text = resp.content or ""
+        if turn_text:
+            last_text = turn_text
+        assistant_msg: dict[str, Any] = {
+            "role": "assistant",
+            "content": turn_text,
+        }
+        if tool_calls_raw:
+            # Convert to OpenAI format for the conversation.
+            assistant_msg["tool_calls"] = [
+                {
+                    "id": tc["id"],
+                    "type": "function",
+                    "function": {
+                        "name": tc["name"],
+                        "arguments": json.dumps(tc.get("input", {})),
+                    },
+                }
+                for tc in tool_calls_raw
+            ]
+        messages.append(assistant_msg)
+
+        # No tool calls → agent is done.
+        if not tool_calls_raw:
+            logger.debug("reflect: loop done after %d turn(s) (no tool calls)", _turn + 1)
+            break
+
+        # Execute each tool call and append results.
+        logger.debug("reflect: turn %d — executing %d tool call(s): %s", _turn + 1, len(tool_calls_raw), [tc["name"] for tc in tool_calls_raw])
+        for tc in tool_calls_raw:
+            result = _execute_tool(tc["name"], tc.get("input", {}), memory_dir, caller)
+            # Track files that were written or deleted.
+            if tc["name"] in ("write_memory_file", "delete_memory_file"):
+                fname = tc.get("input", {}).get("filename", "")
+                if fname and not result.startswith("ERROR"):
+                    changed_files.append(fname)
+            messages.append({
+                "role": "tool",
+                "tool_call_id": tc["id"],
+                "content": result,
+            })
+
+    return True, changed_files, last_text
+
+
+# ---------------------------------------------------------------------------
+# System prompts
+# ---------------------------------------------------------------------------
+
+_FRONTMATTER_EXAMPLE = "\n".join(MEMORY_FRONTMATTER_EXAMPLE)
+
+_SHORT_REFLECT_SYSTEM = f"""\
+You are a reflection agent that distills learnings from a conversation into
+persistent memory files.  You run in the background after each assistant turn.
+
+Your goal: identify anything from the recent messages worth remembering across
+future sessions — user preferences, project context, techniques that worked,
+goals, environment details, reference pointers.
+
+Memory types: {', '.join(MEMORY_TYPES)}
+
+Expected format for each memory file:
+{_FRONTMATTER_EXAMPLE}
+
+Workflow (aim for 2 turns):
+  Turn 1 — call list_memory_files to see what already exists, then
+            read_memory_file for any that might need updating.
+  Turn 2 — call write_memory_file for new/updated memories.
+
+Rules:
+- Only persist information that would be useful in a *future* conversation.
+  Skip ephemeral task details, routine tool output, and anything obvious
+  from the code or git history.
+- Keep files concise.  Each file should cover ONE topic.
+- If an existing memory already covers the learning, UPDATE it rather than
+  creating a duplicate.
+- If there is nothing worth remembering from these messages, do nothing
+  (respond with a brief reason why nothing was saved — no tool calls needed).
+- IMPORTANT: Always end with a text message (no tool calls) summarising what
+  you did or why you skipped.  Never end on an empty response.
+- File names should be kebab-case slugs ending in .md.
+- Include a specific, search-friendly description in the frontmatter.
+- Do NOT exceed {MAX_FILE_SIZE_BYTES} bytes per file or {MAX_FILES} total files.
+"""
+
+_LONG_REFLECT_SYSTEM = f"""\
+You are a reflection agent performing a periodic housekeeping pass over the
+memory directory.  Your job is to organise, deduplicate, and trim noise from
+the accumulated memory files.
+
+Memory types: {', '.join(MEMORY_TYPES)}
+
+Expected format for each memory file:
+{_FRONTMATTER_EXAMPLE}
+
+Workflow:
+  1. list_memory_files to get the full manifest.
+  2. read_memory_file for files that look redundant, stale, or overlapping.
+  3. Merge duplicates, delete stale entries, consolidate related memories.
+  4. Ensure descriptions are specific and search-friendly.
+  5. Enforce limits: max {MAX_FILES} files, max {MAX_FILE_SIZE_BYTES} bytes each.
+
+Rules:
+- Prefer merging over deleting — combine related memories into one file.
+- Remove memories that are no longer relevant or are superseded.
+- Keep the total collection lean and high-signal.
+- Do NOT invent new information — only reorganise what exists.
+- Do NOT delete or merge MEMORY-*.md diary files. These are daily narratives
+  managed by a separate process. You may read them for context but should not
+  modify them.
+"""
+
+_DIARY_SYSTEM = """\
+You maintain a daily diary entry for an AI colony session. You receive:
+(1) Today's existing diary content (may be empty if this is the first entry).
+(2) A transcript of recent conversation messages.
+
+Write a cohesive 3-8 sentence narrative about what happened in this session today.
+Cover: what the user asked for, what was accomplished, key decisions or obstacles,
+and current status.
+
+Rules:
+- If an existing diary is provided, rewrite it as a unified narrative incorporating
+  the new developments. Merge and deduplicate — do not simply append.
+- Keep the total narrative under 3000 characters.
+- Focus on the story arc of the day, not individual tool calls or code details.
+- If the recent messages contain nothing substantive (greetings, routine
+  confirmations), return the existing diary text unchanged.
+- Output only the diary prose. No headings, no timestamps, no code fences, no
+  frontmatter.
+"""
+
+
+# ---------------------------------------------------------------------------
+# Short & long reflection entry points
+# ---------------------------------------------------------------------------
+
+
+async def run_short_reflection(
+    session_dir: Path,
+    llm: Any,
+    memory_dir: Path | None = None,
+    *,
+    caller: str,
+) -> None:
+    """Run a short reflection: extract learnings from conversation."""
+    mem_dir = memory_dir or MEMORY_DIR
+
+    messages = await read_conversation_parts(session_dir)
+    if not messages:
+        logger.debug("reflect: short [%s] — no conversation parts", caller)
+        return
+
+    logger.debug("reflect: short [%s] — %d conversation parts", caller, len(messages))
+
+    # Build a readable transcript from recent messages.
+    transcript_lines: list[str] = []
+    for msg in messages[-50:]:
+        role = msg.get("role", "")
+        content = str(msg.get("content", "")).strip()
+        if role == "tool":
+            continue  # Skip verbose tool results.
+        if not content:
+            continue
+        label = "user" if role == "user" else "assistant"
+        if len(content) > 800:
+            content = content[:800] + "…"
+        transcript_lines.append(f"[{label}]: {content}")
+
+    if not transcript_lines:
+        return
+
+    transcript = "\n".join(transcript_lines)
+    user_msg = (
+        f"## Recent conversation ({len(messages)} messages total)\n\n"
+        f"{transcript}\n\n"
+        f"Timestamp: {datetime.now().isoformat(timespec='minutes')}"
+    )
+
+    _, changed, reason = await _reflection_loop(
+        llm, _SHORT_REFLECT_SYSTEM, user_msg, mem_dir, caller=caller,
+    )
+    if changed:
+        logger.debug("reflect: short reflection done [%s], changed files: %s", caller, changed)
+    else:
+        logger.debug("reflect: short reflection done [%s], no changes — %s", caller, reason or "no reason given")
+
+
+async def run_long_reflection(
+    llm: Any,
+    memory_dir: Path | None = None,
+    *,
+    caller: str,
+) -> None:
+    """Run a long reflection: organise and deduplicate all memories."""
+    mem_dir = memory_dir or MEMORY_DIR
+    files = scan_memory_files(mem_dir)
+
+    if not files:
+        logger.debug("reflect: long [%s] — no memory files to organise", caller)
+        return
+
+    logger.debug("reflect: long [%s] — organising %d memory files", caller, len(files))
+    manifest = format_memory_manifest(files)
+    user_msg = (
+        f"## Current memory manifest ({len(files)} files)\n\n"
+        f"{manifest}\n\n"
+        f"Timestamp: {datetime.now().isoformat(timespec='minutes')}"
+    )
+
+    _, changed, reason = await _reflection_loop(
+        llm, _LONG_REFLECT_SYSTEM, user_msg, mem_dir, caller=caller,
+    )
+    if changed:
+        logger.debug("reflect: long reflection done [%s] (%d files), changed files: %s", caller, len(files), changed)
+    else:
+        logger.debug("reflect: long reflection done [%s] (%d files), no changes — %s", caller, len(files), reason or "no reason given")
+
+
+async def run_diary_update(
+    session_dir: Path,
+    llm: Any,
+    memory_dir: Path | None = None,
+) -> None:
+    """Update today's diary file with a narrative of recent activity."""
+    mem_dir = memory_dir or MEMORY_DIR
+
+    fname = diary_filename()
+    diary_path = mem_dir / fname
+    today_str = datetime.now().strftime("%Y-%m-%d")
+
+    # Read existing diary body (strip frontmatter).
+    existing_body = ""
+    if diary_path.exists():
+        try:
+            raw = diary_path.read_text(encoding="utf-8")
+            m = re.match(r"^---\s*\n.*?\n---\s*\n?", raw, re.DOTALL)
+            existing_body = raw[m.end() :].strip() if m else raw.strip()
+        except OSError:
+            pass
+
+    # Read all conversation messages for context.
+    messages = await read_conversation_parts(session_dir)
+    transcript_lines: list[str] = []
+    for msg in messages[-40:]:
+        role = msg.get("role", "")
+        content = str(msg.get("content", "")).strip()
+        if role == "tool" or not content:
+            continue
+        label = "user" if role == "user" else "assistant"
+        if len(content) > 600:
+            content = content[:600] + "..."
+        transcript_lines.append(f"[{label}]: {content}")
+
+    if not transcript_lines:
+        return
+
+    transcript = "\n".join(transcript_lines)
+    user_msg = (
+        f"## Today's Diary So Far\n\n"
+        f"{existing_body or '(no entries yet)'}\n\n"
+        f"## Recent Conversation\n\n"
+        f"{transcript}\n\n"
+        f"Date: {today_str}"
+    )
+
+    try:
+        from framework.agents.queen.config import default_config
+
+        resp = await llm.acomplete(
+            messages=[{"role": "user", "content": user_msg}],
+            system=_DIARY_SYSTEM,
+            max_tokens=min(default_config.max_tokens, 1024),
+        )
+        new_body = (resp.content or "").strip()
+        if not new_body:
+            return
+
+        doc = build_diary_document(date_str=today_str, body=new_body)
+        if len(doc.encode("utf-8")) > MAX_FILE_SIZE_BYTES:
+            new_body = new_body[:2800]
+            doc = build_diary_document(date_str=today_str, body=new_body)
+
+        mem_dir.mkdir(parents=True, exist_ok=True)
+        diary_path.write_text(doc, encoding="utf-8")
+        logger.debug("diary: updated %s (%d chars)", fname, len(doc))
+    except Exception:
+        logger.warning("diary: update failed", exc_info=True)
+
+
+# ---------------------------------------------------------------------------
+# Event-bus integration
+# ---------------------------------------------------------------------------
+
+# Run a long reflection every N short reflections.
+_LONG_REFLECT_INTERVAL = 5
+
+
+async def subscribe_reflection_triggers(
+    event_bus: Any,
+    session_dir: Path,
+    llm: Any,
+    memory_dir: Path | None = None,
+    phase_state: Any = None,
+) -> list[str]:
+    """Subscribe to queen turn events and return subscription IDs.
+
+    Call this once during queen setup.  Returns a list of event-bus
+    subscription IDs for cleanup during session teardown.
+    """
+    from framework.runtime.event_bus import EventType
+
+    mem_dir = memory_dir or MEMORY_DIR
+    _lock = asyncio.Lock()
+    _short_count = 0
+
+    async def _on_turn_complete(event: Any) -> None:
+        nonlocal _short_count
+
+        # Only process queen turns.
+        if getattr(event, "stream_id", None) != "queen":
+            return
+
+        _short_count += 1
+
+        # Decide whether to reflect: only when the LLM turn ended without
+        # tool calls (a conversational response) OR every _LONG_REFLECT_INTERVAL turns.
+        event_data = getattr(event, "data", {}) or {}
+        stop_reason = event_data.get("stop_reason", "")
+        is_tool_turn = stop_reason in ("tool_use", "tool_calls")
+        is_interval = _short_count % _LONG_REFLECT_INTERVAL == 0
+
+        if is_tool_turn and not is_interval:
+            logger.debug(
+                "reflect: skipping turn %d (stop_reason=%s, next reflect at %d)",
+                _short_count, stop_reason,
+                (_short_count // _LONG_REFLECT_INTERVAL + 1) * _LONG_REFLECT_INTERVAL,
+            )
+            return
+
+        if _lock.locked():
+            logger.debug("reflect: skipping — reflection already in progress")
+            return
+
+        async with _lock:
+            try:
+                logger.debug("reflect: turn complete — count %d/%d (stop_reason=%s)", _short_count, _LONG_REFLECT_INTERVAL, stop_reason)
+                if is_interval:
+                    await run_short_reflection(session_dir, llm, mem_dir, caller="queen")
+                    await run_long_reflection(llm, mem_dir, caller="queen")
+                else:
+                    await run_short_reflection(session_dir, llm, mem_dir, caller="queen")
+            except Exception:
+                logger.warning("reflect: reflection failed", exc_info=True)
+                _write_error("short/long reflection")
+
+            # Update daily diary after reflection.
+            try:
+                await run_diary_update(session_dir, llm, mem_dir)
+            except Exception:
+                logger.warning("reflect: diary update failed", exc_info=True)
+
+            # Update recall cache after reflection completes, guaranteeing
+            # recall sees the current turn's extracted memories.
+            if phase_state is not None:
+                try:
+                    from framework.agents.queen.recall_selector import update_recall_cache
+                    await update_recall_cache(
+                        session_dir,
+                        llm,
+                        cache_setter=lambda block: (
+                            setattr(phase_state, "_cached_colony_recall_block", block),
+                            setattr(phase_state, "_cached_recall_block", block),
+                        ),
+                        memory_dir=mem_dir,
+                        heading="Colony Memories",
+                    )
+                    await update_recall_cache(
+                        session_dir,
+                        llm,
+                        cache_setter=lambda block: setattr(
+                            phase_state, "_cached_global_recall_block", block
+                        ),
+                        memory_dir=getattr(phase_state, "global_memory_dir", None),
+                        heading="Global Memories",
+                    )
+                except Exception:
+                    logger.debug("recall: cache update failed", exc_info=True)
+
+    async def _on_compaction(event: Any) -> None:
+        if getattr(event, "stream_id", None) != "queen":
+            return
+
+        if _lock.locked():
+            return
+
+        async with _lock:
+            try:
+                await run_long_reflection(llm, mem_dir, caller="queen")
+            except Exception:
+                logger.warning("reflect: compaction-triggered reflection failed", exc_info=True)
+                _write_error("compaction reflection")
+
+    sub_ids: list[str] = []
+
+    sub1 = event_bus.subscribe(
+        event_types=[EventType.LLM_TURN_COMPLETE],
+        handler=_on_turn_complete,
+    )
+    sub_ids.append(sub1)
+
+    sub2 = event_bus.subscribe(
+        event_types=[EventType.CONTEXT_COMPACTED],
+        handler=_on_compaction,
+    )
+    sub_ids.append(sub2)
+
+    return sub_ids
+
+
+async def subscribe_worker_memory_triggers(
+    event_bus: Any,
+    llm: Any,
+    *,
+    worker_sessions_dir: Path,
+    colony_memory_dir: Path,
+    recall_cache: dict[str, str],
+) -> list[str]:
+    """Subscribe colony memory lifecycle events for worker runs.
+
+    Short reflection is now handled synchronously at node handoff in
+    ``WorkerAgent._reflect_colony_memory()``.  This function only manages:
+    - Recall cache initialisation on execution start
+    - Final long reflection + cleanup on execution end
+    """
+    from framework.runtime.event_bus import EventType
+
+    _terminal_lock = asyncio.Lock()
+
+    def _is_worker_event(event: Any) -> bool:
+        return bool(
+            getattr(event, "execution_id", None)
+            and getattr(event, "stream_id", None) not in ("queen", "judge")
+        )
+
+    async def _on_execution_started(event: Any) -> None:
+        if not _is_worker_event(event):
+            return
+        if event.execution_id is not None:
+            recall_cache[event.execution_id] = ""
+
+    async def _on_execution_terminal(event: Any) -> None:
+        if not _is_worker_event(event):
+            return
+        execution_id = event.execution_id
+        if execution_id is None:
+            return
+        async with _terminal_lock:
+            try:
+                await run_long_reflection(llm, colony_memory_dir, caller="worker")
+            except Exception:
+                logger.warning("reflect: worker final reflection failed", exc_info=True)
+                _write_error("worker final reflection")
+            finally:
+                recall_cache.pop(execution_id, None)
+
+    return [
+        event_bus.subscribe(
+            event_types=[EventType.EXECUTION_STARTED],
+            handler=_on_execution_started,
+        ),
+        event_bus.subscribe(
+            event_types=[EventType.EXECUTION_COMPLETED, EventType.EXECUTION_FAILED],
+            handler=_on_execution_terminal,
+        ),
+    ]
+
+
+def _write_error(context: str) -> None:
+    """Best-effort write of the last traceback to an error file."""
+    try:
+        error_path = MEMORY_DIR / ".reflection_error.txt"
+        error_path.parent.mkdir(parents=True, exist_ok=True)
+        error_path.write_text(
+            f"context: {context}\ntime: {datetime.now().isoformat()}\n\n{traceback.format_exc()}",
+            encoding="utf-8",
+        )
+    except OSError:
+        pass

core/framework/agents/queen/tests/conftest.py

@@ -0,0 +1,31 @@
+"""Test fixtures for Queen agent."""
+
+import sys
+from pathlib import Path
+
+import pytest
+import pytest_asyncio
+
+_repo_root = Path(__file__).resolve().parents[3]
+for _p in ["exports", "core"]:
+    _path = str(_repo_root / _p)
+    if _path not in sys.path:
+        sys.path.insert(0, _path)
+
+AGENT_PATH = str(Path(__file__).resolve().parents[1])
+
+
+@pytest.fixture(scope="session")
+def mock_mode():
+    return True
+
+
+@pytest_asyncio.fixture(scope="session")
+async def runner(tmp_path_factory, mock_mode):
+    from framework.runner.runner import AgentRunner
+
+    storage = tmp_path_factory.mktemp("agent_storage")
+    r = AgentRunner.load(AGENT_PATH, mock_mode=mock_mode, storage_path=storage)
+    r._setup()
+    yield r
+    await r.cleanup_async()

core/framework/builder/__init__.py

@@ -1,21 +0,0 @@
-"""Builder interface for analyzing and building agents."""
-
-from framework.builder.query import BuilderQuery
-from framework.builder.workflow import (
-    BuildPhase,
-    BuildSession,
-    GraphBuilder,
-    TestCase,
-    TestResult,
-    ValidationResult,
-)
-
-__all__ = [
-    "BuilderQuery",
-    "GraphBuilder",
-    "BuildSession",
-    "BuildPhase",
-    "ValidationResult",
-    "TestCase",
-    "TestResult",
-]

core/framework/builder/query.py

@@ -1,501 +0,0 @@
-"""
-Builder Query Interface - How I (Builder) analyze agent runs.
-
-This is designed around the questions I need to answer:
-1. What happened? (summaries, narratives)
-2. Why did it fail? (failure analysis, decision traces)
-3. What patterns emerge? (across runs, across nodes)
-4. What should we change? (suggestions)
-"""
-
-from collections import defaultdict
-from pathlib import Path
-from typing import Any
-
-from framework.schemas.decision import Decision
-from framework.schemas.run import Run, RunStatus, RunSummary
-from framework.storage.backend import FileStorage
-
-
-class FailureAnalysis:
-    """Structured analysis of why a run failed."""
-
-    def __init__(
-        self,
-        run_id: str,
-        failure_point: str,
-        root_cause: str,
-        decision_chain: list[str],
-        problems: list[str],
-        suggestions: list[str],
-    ):
-        self.run_id = run_id
-        self.failure_point = failure_point
-        self.root_cause = root_cause
-        self.decision_chain = decision_chain
-        self.problems = problems
-        self.suggestions = suggestions
-
-    def to_dict(self) -> dict[str, Any]:
-        return {
-            "run_id": self.run_id,
-            "failure_point": self.failure_point,
-            "root_cause": self.root_cause,
-            "decision_chain": self.decision_chain,
-            "problems": self.problems,
-            "suggestions": self.suggestions,
-        }
-
-    def __str__(self) -> str:
-        lines = [
-            f"=== Failure Analysis for {self.run_id} ===",
-            "",
-            f"Failure Point: {self.failure_point}",
-            f"Root Cause: {self.root_cause}",
-            "",
-            "Decision Chain Leading to Failure:",
-        ]
-        for i, dec in enumerate(self.decision_chain, 1):
-            lines.append(f"  {i}. {dec}")
-
-        if self.problems:
-            lines.append("")
-            lines.append("Reported Problems:")
-            for prob in self.problems:
-                lines.append(f"  - {prob}")
-
-        if self.suggestions:
-            lines.append("")
-            lines.append("Suggestions:")
-            for sug in self.suggestions:
-                lines.append(f"  → {sug}")
-
-        return "\n".join(lines)
-
-
-class PatternAnalysis:
-    """Patterns detected across multiple runs."""
-
-    def __init__(
-        self,
-        goal_id: str,
-        run_count: int,
-        success_rate: float,
-        common_failures: list[tuple[str, int]],
-        problematic_nodes: list[tuple[str, float]],
-        decision_patterns: dict[str, Any],
-    ):
-        self.goal_id = goal_id
-        self.run_count = run_count
-        self.success_rate = success_rate
-        self.common_failures = common_failures
-        self.problematic_nodes = problematic_nodes
-        self.decision_patterns = decision_patterns
-
-    def to_dict(self) -> dict[str, Any]:
-        return {
-            "goal_id": self.goal_id,
-            "run_count": self.run_count,
-            "success_rate": self.success_rate,
-            "common_failures": self.common_failures,
-            "problematic_nodes": self.problematic_nodes,
-            "decision_patterns": self.decision_patterns,
-        }
-
-    def __str__(self) -> str:
-        lines = [
-            f"=== Pattern Analysis for Goal {self.goal_id} ===",
-            "",
-            f"Runs Analyzed: {self.run_count}",
-            f"Success Rate: {self.success_rate:.1%}",
-        ]
-
-        if self.common_failures:
-            lines.append("")
-            lines.append("Common Failures:")
-            for failure, count in self.common_failures:
-                lines.append(f"  - {failure} ({count} occurrences)")
-
-        if self.problematic_nodes:
-            lines.append("")
-            lines.append("Problematic Nodes (failure rate):")
-            for node, rate in self.problematic_nodes:
-                lines.append(f"  - {node}: {rate:.1%} failure rate")
-
-        return "\n".join(lines)
-
-
-class BuilderQuery:
-    """
-    The interface I (Builder) use to understand what agents are doing.
-
-    This is optimized for the questions I need to answer when analyzing
-    agent behavior and deciding what to improve.
-    """
-
-    def __init__(self, storage_path: str | Path):
-        self.storage = FileStorage(storage_path)
-
-    # === WHAT HAPPENED? ===
-
-    def get_run_summary(self, run_id: str) -> RunSummary | None:
-        """Get a quick summary of a run."""
-        return self.storage.load_summary(run_id)
-
-    def get_full_run(self, run_id: str) -> Run | None:
-        """Get the complete run with all decisions."""
-        return self.storage.load_run(run_id)
-
-    def list_runs_for_goal(self, goal_id: str) -> list[RunSummary]:
-        """Get summaries of all runs for a goal."""
-        run_ids = self.storage.get_runs_by_goal(goal_id)
-        summaries = []
-        for run_id in run_ids:
-            summary = self.storage.load_summary(run_id)
-            if summary:
-                summaries.append(summary)
-        return summaries
-
-    def get_recent_failures(self, limit: int = 10) -> list[RunSummary]:
-        """Get recent failed runs."""
-        run_ids = self.storage.get_runs_by_status(RunStatus.FAILED)
-        summaries = []
-        for run_id in run_ids[:limit]:
-            summary = self.storage.load_summary(run_id)
-            if summary:
-                summaries.append(summary)
-        return summaries
-
-    # === WHY DID IT FAIL? ===
-
-    def analyze_failure(self, run_id: str) -> FailureAnalysis | None:
-        """
-        Deep analysis of why a run failed.
-
-        This is my primary tool for understanding what went wrong.
-        """
-        run = self.storage.load_run(run_id)
-        if run is None or run.status != RunStatus.FAILED:
-            return None
-
-        # Find the first failed decision
-        failed_decisions = [d for d in run.decisions if not d.was_successful]
-        if not failed_decisions:
-            failure_point = "Unknown - no decision marked as failed"
-            root_cause = "Run failed but all decisions succeeded (external cause?)"
-        else:
-            first_failure = failed_decisions[0]
-            failure_point = first_failure.summary_for_builder()
-            root_cause = first_failure.outcome.error if first_failure.outcome else "Unknown"
-
-        # Build the decision chain leading to failure
-        decision_chain = []
-        for d in run.decisions:
-            decision_chain.append(d.summary_for_builder())
-            if not d.was_successful:
-                break
-
-        # Extract problems
-        problems = [f"[{p.severity}] {p.description}" for p in run.problems]
-
-        # Generate suggestions based on the failure
-        suggestions = self._generate_suggestions(run, failed_decisions)
-
-        return FailureAnalysis(
-            run_id=run_id,
-            failure_point=failure_point,
-            root_cause=root_cause,
-            decision_chain=decision_chain,
-            problems=problems,
-            suggestions=suggestions,
-        )
-
-    def get_decision_trace(self, run_id: str) -> list[str]:
-        """Get a readable trace of all decisions in a run."""
-        run = self.storage.load_run(run_id)
-        if run is None:
-            return []
-        return [d.summary_for_builder() for d in run.decisions]
-
-    # === WHAT PATTERNS EMERGE? ===
-
-    def find_patterns(self, goal_id: str) -> PatternAnalysis | None:
-        """
-        Find patterns across runs for a goal.
-
-        This helps me understand systemic issues vs one-off failures.
-        """
-        run_ids = self.storage.get_runs_by_goal(goal_id)
-        if not run_ids:
-            return None
-
-        runs = []
-        for run_id in run_ids:
-            run = self.storage.load_run(run_id)
-            if run:
-                runs.append(run)
-
-        if not runs:
-            return None
-
-        # Calculate success rate
-        completed = [r for r in runs if r.status == RunStatus.COMPLETED]
-        success_rate = len(completed) / len(runs) if runs else 0.0
-
-        # Find common failures
-        failure_counts: dict[str, int] = defaultdict(int)
-        for run in runs:
-            for decision in run.decisions:
-                if not decision.was_successful and decision.outcome:
-                    error = decision.outcome.error or "Unknown error"
-                    failure_counts[error] += 1
-
-        common_failures = sorted(failure_counts.items(), key=lambda x: x[1], reverse=True)[:5]
-
-        # Find problematic nodes
-        node_stats: dict[str, dict[str, int]] = defaultdict(lambda: {"total": 0, "failed": 0})
-        for run in runs:
-            for decision in run.decisions:
-                node_stats[decision.node_id]["total"] += 1
-                if not decision.was_successful:
-                    node_stats[decision.node_id]["failed"] += 1
-
-        problematic_nodes = []
-        for node_id, stats in node_stats.items():
-            if stats["total"] > 0:
-                failure_rate = stats["failed"] / stats["total"]
-                if failure_rate > 0.1:  # More than 10% failure rate
-                    problematic_nodes.append((node_id, failure_rate))
-
-        problematic_nodes.sort(key=lambda x: x[1], reverse=True)
-
-        # Decision patterns
-        decision_patterns = self._analyze_decision_patterns(runs)
-
-        return PatternAnalysis(
-            goal_id=goal_id,
-            run_count=len(runs),
-            success_rate=success_rate,
-            common_failures=common_failures,
-            problematic_nodes=problematic_nodes,
-            decision_patterns=decision_patterns,
-        )
-
-    def compare_runs(self, run_id_1: str, run_id_2: str) -> dict[str, Any]:
-        """Compare two runs to understand what differed."""
-        run1 = self.storage.load_run(run_id_1)
-        run2 = self.storage.load_run(run_id_2)
-
-        if run1 is None or run2 is None:
-            return {"error": "One or both runs not found"}
-
-        return {
-            "run_1": {
-                "id": run1.id,
-                "status": run1.status.value,
-                "decisions": len(run1.decisions),
-                "success_rate": run1.metrics.success_rate,
-            },
-            "run_2": {
-                "id": run2.id,
-                "status": run2.status.value,
-                "decisions": len(run2.decisions),
-                "success_rate": run2.metrics.success_rate,
-            },
-            "differences": self._find_differences(run1, run2),
-        }
-
-    # === WHAT SHOULD WE CHANGE? ===
-
-    def suggest_improvements(self, goal_id: str) -> list[dict[str, Any]]:
-        """
-        Generate improvement suggestions based on run analysis.
-
-        This is what I use to propose changes to the human engineer.
-        """
-        patterns = self.find_patterns(goal_id)
-        if patterns is None:
-            return []
-
-        suggestions = []
-
-        # Suggestion: Fix problematic nodes
-        for node_id, failure_rate in patterns.problematic_nodes:
-            suggestions.append(
-                {
-                    "type": "node_improvement",
-                    "target": node_id,
-                    "reason": f"Node has {failure_rate:.1%} failure rate",
-                    "recommendation": (
-                        f"Review and improve node '{node_id}' - "
-                        "high failure rate suggests prompt or tool issues"
-                    ),
-                    "priority": "high" if failure_rate > 0.3 else "medium",
-                }
-            )
-
-        # Suggestion: Address common failures
-        for failure, count in patterns.common_failures:
-            if count >= 2:
-                suggestions.append(
-                    {
-                        "type": "error_handling",
-                        "target": failure,
-                        "reason": f"Error occurred {count} times",
-                        "recommendation": f"Add handling for: {failure}",
-                        "priority": "high" if count >= 5 else "medium",
-                    }
-                )
-
-        # Suggestion: Overall success rate
-        if patterns.success_rate < 0.8:
-            suggestions.append(
-                {
-                    "type": "architecture",
-                    "target": goal_id,
-                    "reason": f"Goal success rate is only {patterns.success_rate:.1%}",
-                    "recommendation": (
-                        "Consider restructuring the agent graph or improving goal definition"
-                    ),
-                    "priority": "high",
-                }
-            )
-
-        return suggestions
-
-    def get_node_performance(self, node_id: str) -> dict[str, Any]:
-        """Get performance metrics for a specific node across all runs."""
-        run_ids = self.storage.get_runs_by_node(node_id)
-
-        total_decisions = 0
-        successful_decisions = 0
-        total_latency = 0
-        total_tokens = 0
-        decision_types: dict[str, int] = defaultdict(int)
-
-        for run_id in run_ids:
-            run = self.storage.load_run(run_id)
-            if run:
-                for decision in run.decisions:
-                    if decision.node_id == node_id:
-                        total_decisions += 1
-                        if decision.was_successful:
-                            successful_decisions += 1
-                        if decision.outcome:
-                            total_latency += decision.outcome.latency_ms
-                            total_tokens += decision.outcome.tokens_used
-                        decision_types[decision.decision_type.value] += 1
-
-        return {
-            "node_id": node_id,
-            "total_decisions": total_decisions,
-            "success_rate": successful_decisions / total_decisions if total_decisions > 0 else 0,
-            "avg_latency_ms": total_latency / total_decisions if total_decisions > 0 else 0,
-            "total_tokens": total_tokens,
-            "decision_type_distribution": dict(decision_types),
-        }
-
-    # === PRIVATE HELPERS ===
-
-    def _generate_suggestions(
-        self,
-        run: Run,
-        failed_decisions: list[Decision],
-    ) -> list[str]:
-        """Generate suggestions based on failure analysis."""
-        suggestions = []
-
-        for decision in failed_decisions:
-            # Check if there were alternatives
-            if len(decision.options) > 1:
-                chosen = decision.chosen_option
-                alternatives = [o for o in decision.options if o.id != decision.chosen_option_id]
-                if alternatives:
-                    alt_desc = alternatives[0].description
-                    chosen_desc = chosen.description if chosen else "unknown"
-                    suggestions.append(
-                        f"Consider alternative: '{alt_desc}' instead of '{chosen_desc}'"
-                    )
-
-            # Check for missing context
-            if not decision.input_context:
-                suggestions.append(
-                    f"Decision '{decision.intent}' had no input context - "
-                    "ensure relevant data is passed"
-                )
-
-            # Check for constraint issues
-            if decision.active_constraints:
-                constraints = ", ".join(decision.active_constraints)
-                suggestions.append(f"Review constraints: {constraints} - may be too restrictive")
-
-        # Check for reported problems with suggestions
-        for problem in run.problems:
-            if problem.suggested_fix:
-                suggestions.append(problem.suggested_fix)
-
-        return suggestions
-
-    def _analyze_decision_patterns(self, runs: list[Run]) -> dict[str, Any]:
-        """Analyze decision patterns across runs."""
-        type_counts: dict[str, int] = defaultdict(int)
-        option_counts: dict[str, dict[str, int]] = defaultdict(lambda: defaultdict(int))
-
-        for run in runs:
-            for decision in run.decisions:
-                type_counts[decision.decision_type.value] += 1
-
-                # Track which options are chosen for similar intents
-                intent_key = decision.intent[:50]  # Truncate for grouping
-                if decision.chosen_option:
-                    option_counts[intent_key][decision.chosen_option.description] += 1
-
-        # Find most common choices per intent
-        common_choices = {}
-        for intent, choices in option_counts.items():
-            if choices:
-                most_common = max(choices.items(), key=lambda x: x[1])
-                common_choices[intent] = {
-                    "choice": most_common[0],
-                    "count": most_common[1],
-                    "alternatives": len(choices) - 1,
-                }
-
-        return {
-            "decision_type_distribution": dict(type_counts),
-            "common_choices": common_choices,
-        }
-
-    def _find_differences(self, run1: Run, run2: Run) -> list[str]:
-        """Find key differences between two runs."""
-        differences = []
-
-        # Status difference
-        if run1.status != run2.status:
-            differences.append(f"Status: {run1.status.value} vs {run2.status.value}")
-
-        # Decision count difference
-        if len(run1.decisions) != len(run2.decisions):
-            differences.append(f"Decision count: {len(run1.decisions)} vs {len(run2.decisions)}")
-
-        # Find first divergence point
-        for i, (d1, d2) in enumerate(zip(run1.decisions, run2.decisions, strict=False)):
-            if d1.chosen_option_id != d2.chosen_option_id:
-                differences.append(
-                    f"Diverged at decision {i}: "
-                    f"chose '{d1.chosen_option_id}' vs '{d2.chosen_option_id}'"
-                )
-                break
-
-        # Node differences
-        nodes1 = set(run1.metrics.nodes_executed)
-        nodes2 = set(run2.metrics.nodes_executed)
-        if nodes1 != nodes2:
-            only_1 = nodes1 - nodes2
-            only_2 = nodes2 - nodes1
-            if only_1:
-                differences.append(f"Nodes only in run 1: {only_1}")
-            if only_2:
-                differences.append(f"Nodes only in run 2: {only_2}")
-
-        return differences

core/framework/builder/workflow.py

@@ -1,807 +0,0 @@
-"""
-GraphBuilder Workflow - Enforced incremental building with HITL approval.
-
-The build process:
-1. Define Goal → APPROVE
-2. Add Node → VALIDATE → TEST → APPROVE
-3. Add Edge → VALIDATE → TEST → APPROVE
-4. Repeat until graph is complete
-5. Final integration test → APPROVE
-6. Export
-
-Each step requires validation and human approval before proceeding.
-You cannot skip steps or bypass validation.
-"""
-
-from collections.abc import Callable
-from datetime import datetime
-from enum import Enum
-from pathlib import Path
-from typing import Any
-
-from pydantic import BaseModel, Field
-
-from framework.graph.edge import EdgeCondition, EdgeSpec, GraphSpec
-from framework.graph.goal import Goal
-from framework.graph.node import NodeSpec
-
-
-class BuildPhase(str, Enum):
-    """Current phase of the build process."""
-
-    INIT = "init"  # Just started
-    GOAL_DRAFT = "goal_draft"  # Drafting goal
-    GOAL_APPROVED = "goal_approved"  # Goal approved
-    ADDING_NODES = "adding_nodes"  # Adding nodes
-    ADDING_EDGES = "adding_edges"  # Adding edges
-    TESTING = "testing"  # Running tests
-    APPROVED = "approved"  # Fully approved
-    EXPORTED = "exported"  # Exported to file
-
-
-class ValidationResult(BaseModel):
-    """Result of a validation check."""
-
-    valid: bool
-    errors: list[str] = Field(default_factory=list)
-    warnings: list[str] = Field(default_factory=list)
-    suggestions: list[str] = Field(default_factory=list)
-
-
-class TestCase(BaseModel):
-    """A test case for validating agent behavior."""
-
-    id: str
-    description: str
-    input: dict[str, Any]
-    expected_output: Any = None  # None means just check it doesn't error
-    expected_contains: str | None = None
-
-
-class TestResult(BaseModel):
-    """Result of running a test case."""
-
-    test_id: str
-    passed: bool
-    actual_output: Any = None
-    error: str | None = None
-    execution_path: list[str] = Field(default_factory=list)
-
-
-class BuildSession(BaseModel):
-    """
-    Persistent build session state.
-
-    Saved after each approved step so you can resume later.
-    """
-
-    id: str
-    name: str
-    phase: BuildPhase = BuildPhase.INIT
-    created_at: datetime = Field(default_factory=datetime.now)
-    updated_at: datetime = Field(default_factory=datetime.now)
-
-    # The artifacts being built
-    goal: Goal | None = None
-    nodes: list[NodeSpec] = Field(default_factory=list)
-    edges: list[EdgeSpec] = Field(default_factory=list)
-
-    # Test cases
-    test_cases: list[TestCase] = Field(default_factory=list)
-    test_results: list[TestResult] = Field(default_factory=list)
-
-    # Approval history
-    approvals: list[dict[str, Any]] = Field(default_factory=list)
-
-    # Tools (stored as dicts for serialization)
-    tools: list[dict[str, Any]] = Field(default_factory=list)
-
-    model_config = {"extra": "allow"}
-
-
-class GraphBuilder:
-    """
-    Enforced incremental graph building with HITL approval.
-
-    Usage:
-        builder = GraphBuilder("my-agent")
-
-        # Step 1: Define and approve goal
-        builder.set_goal(goal)
-        builder.validate()  # Must pass
-        builder.approve("Goal looks good")  # Human approval required
-
-        # Step 2: Add nodes one by one
-        builder.add_node(node_spec)
-        builder.validate()  # Must pass
-        builder.test(test_case)  # Must pass
-        builder.approve("Node works")
-
-        # Step 3: Add edges
-        builder.add_edge(edge_spec)
-        builder.validate()
-        builder.approve("Edge correct")
-
-        # Step 4: Final approval
-        builder.run_all_tests()
-        builder.final_approve("Ready for production")
-
-        # Step 5: Export
-        graph = builder.export()
-    """
-
-    def __init__(
-        self,
-        name: str,
-        storage_path: Path | str | None = None,
-        session_id: str | None = None,
-    ):
-        self.storage_path = Path(storage_path) if storage_path else Path.home() / ".core" / "builds"
-        self.storage_path.mkdir(parents=True, exist_ok=True)
-
-        if session_id:
-            self.session = self._load_session(session_id)
-        else:
-            self.session = BuildSession(
-                id=f"build_{datetime.now().strftime('%Y%m%d_%H%M%S')}",
-                name=name,
-            )
-
-        self._pending_validation: ValidationResult | None = None
-
-    # =========================================================================
-    # PHASE 1: GOAL
-    # =========================================================================
-
-    def set_goal(self, goal: Goal) -> ValidationResult:
-        """
-        Set the goal for this agent.
-
-        Returns validation result. Must call approve() after validation passes.
-        """
-        self._require_phase([BuildPhase.INIT, BuildPhase.GOAL_DRAFT])
-
-        self.session.goal = goal
-        self.session.phase = BuildPhase.GOAL_DRAFT
-
-        validation = self._validate_goal(goal)
-        self._pending_validation = validation
-        self._save_session()
-
-        return validation
-
-    def _validate_goal(self, goal: Goal) -> ValidationResult:
-        """Validate a goal definition."""
-        errors = []
-        warnings = []
-        suggestions = []
-
-        if not goal.id:
-            errors.append("Goal must have an id")
-        if not goal.name:
-            errors.append("Goal must have a name")
-        if not goal.description:
-            errors.append("Goal must have a description")
-
-        if not goal.success_criteria:
-            errors.append("Goal must have at least one success criterion")
-        else:
-            for sc in goal.success_criteria:
-                if not sc.description:
-                    errors.append(f"Success criterion '{sc.id}' needs a description")
-
-        if not goal.constraints:
-            warnings.append("Consider adding constraints to define boundaries")
-
-        if not goal.required_capabilities:
-            suggestions.append("Specify required_capabilities (e.g., ['llm', 'tools'])")
-
-        return ValidationResult(
-            valid=len(errors) == 0,
-            errors=errors,
-            warnings=warnings,
-            suggestions=suggestions,
-        )
-
-    # =========================================================================
-    # PHASE 2: NODES
-    # =========================================================================
-
-    def add_node(self, node: NodeSpec) -> ValidationResult:
-        """
-        Add a node to the graph.
-
-        Returns validation result. Must call approve() after validation passes.
-        """
-        self._require_phase([BuildPhase.GOAL_APPROVED, BuildPhase.ADDING_NODES])
-
-        # Check for duplicate
-        if any(n.id == node.id for n in self.session.nodes):
-            return ValidationResult(
-                valid=False,
-                errors=[f"Node with id '{node.id}' already exists"],
-            )
-
-        self.session.nodes.append(node)
-        self.session.phase = BuildPhase.ADDING_NODES
-
-        validation = self._validate_node(node)
-        self._pending_validation = validation
-        self._save_session()
-
-        return validation
-
-    def _validate_node(self, node: NodeSpec) -> ValidationResult:
-        """Validate a node definition."""
-        errors = []
-        warnings = []
-        suggestions = []
-
-        if not node.id:
-            errors.append("Node must have an id")
-        if not node.name:
-            errors.append("Node must have a name")
-        if not node.description:
-            warnings.append(f"Node '{node.id}' should have a description")
-
-        # Type-specific validation
-        if node.node_type == "llm_tool_use":
-            if not node.tools:
-                errors.append(f"LLM tool node '{node.id}' must specify tools")
-            if not node.system_prompt:
-                warnings.append(f"LLM node '{node.id}' should have a system_prompt")
-
-        if node.node_type == "router":
-            if not node.routes:
-                errors.append(f"Router node '{node.id}' must specify routes")
-
-        if node.node_type == "function":
-            if not node.function:
-                errors.append(f"Function node '{node.id}' must specify function name")
-
-        # Check input/output keys
-        if not node.input_keys:
-            suggestions.append(f"Consider specifying input_keys for '{node.id}'")
-        if not node.output_keys:
-            suggestions.append(f"Consider specifying output_keys for '{node.id}'")
-
-        return ValidationResult(
-            valid=len(errors) == 0,
-            errors=errors,
-            warnings=warnings,
-            suggestions=suggestions,
-        )
-
-    def update_node(self, node_id: str, **updates) -> ValidationResult:
-        """Update an existing node."""
-        self._require_phase([BuildPhase.ADDING_NODES])
-
-        for i, node in enumerate(self.session.nodes):
-            if node.id == node_id:
-                node_dict = node.model_dump()
-                node_dict.update(updates)
-                updated_node = NodeSpec(**node_dict)
-                self.session.nodes[i] = updated_node
-
-                validation = self._validate_node(updated_node)
-                self._pending_validation = validation
-                self._save_session()
-                return validation
-
-        return ValidationResult(valid=False, errors=[f"Node '{node_id}' not found"])
-
-    def remove_node(self, node_id: str) -> ValidationResult:
-        """Remove a node (only if no edges reference it)."""
-        self._require_phase([BuildPhase.ADDING_NODES])
-
-        # Check for edge references
-        for edge in self.session.edges:
-            if edge.source == node_id or edge.target == node_id:
-                return ValidationResult(
-                    valid=False,
-                    errors=[f"Cannot remove node '{node_id}': referenced by edge '{edge.id}'"],
-                )
-
-        self.session.nodes = [n for n in self.session.nodes if n.id != node_id]
-        self._save_session()
-
-        return ValidationResult(valid=True)
-
-    # =========================================================================
-    # PHASE 3: EDGES
-    # =========================================================================
-
-    def add_edge(self, edge: EdgeSpec) -> ValidationResult:
-        """
-        Add an edge to the graph.
-
-        Returns validation result. Must call approve() after validation passes.
-        """
-        self._require_phase([BuildPhase.ADDING_NODES, BuildPhase.ADDING_EDGES])
-
-        # Check for duplicate
-        if any(e.id == edge.id for e in self.session.edges):
-            return ValidationResult(
-                valid=False,
-                errors=[f"Edge with id '{edge.id}' already exists"],
-            )
-
-        self.session.edges.append(edge)
-        self.session.phase = BuildPhase.ADDING_EDGES
-
-        validation = self._validate_edge(edge)
-        self._pending_validation = validation
-        self._save_session()
-
-        return validation
-
-    def _validate_edge(self, edge: EdgeSpec) -> ValidationResult:
-        """Validate an edge definition."""
-        errors = []
-        warnings = []
-
-        if not edge.id:
-            errors.append("Edge must have an id")
-
-        # Check source exists
-        if not any(n.id == edge.source for n in self.session.nodes):
-            errors.append(f"Edge source '{edge.source}' not found in nodes")
-
-        # Check target exists
-        if not any(n.id == edge.target for n in self.session.nodes):
-            errors.append(f"Edge target '{edge.target}' not found in nodes")
-
-        # Warn about conditional edges without expressions
-        if edge.condition == EdgeCondition.CONDITIONAL and not edge.condition_expr:
-            warnings.append(f"Conditional edge '{edge.id}' has no condition_expr")
-
-        return ValidationResult(
-            valid=len(errors) == 0,
-            errors=errors,
-            warnings=warnings,
-        )
-
-    # =========================================================================
-    # VALIDATION & TESTING
-    # =========================================================================
-
-    def validate(self) -> ValidationResult:
-        """Validate the entire current graph state."""
-        errors = []
-        warnings = []
-
-        # Must have a goal
-        if not self.session.goal:
-            errors.append("No goal defined")
-            return ValidationResult(valid=False, errors=errors)
-
-        # Must have at least one node
-        if not self.session.nodes:
-            errors.append("No nodes defined")
-
-        # Check for entry node
-        entry_candidates = []
-        for node in self.session.nodes:
-            # A node is an entry candidate if no edges point to it
-            if not any(e.target == node.id for e in self.session.edges):
-                entry_candidates.append(node.id)
-
-        if len(entry_candidates) == 0 and self.session.nodes:
-            errors.append("No entry node found (all nodes have incoming edges)")
-        elif len(entry_candidates) > 1:
-            warnings.append(f"Multiple entry candidates: {entry_candidates}. Specify one.")
-
-        # Check for terminal nodes
-        terminal_candidates = []
-        for node in self.session.nodes:
-            if not any(e.source == node.id for e in self.session.edges):
-                terminal_candidates.append(node.id)
-
-        if not terminal_candidates and self.session.nodes:
-            warnings.append("No terminal nodes found (all nodes have outgoing edges)")
-
-        # Check reachability
-        if entry_candidates and self.session.nodes:
-            reachable = self._compute_reachable(entry_candidates[0])
-            unreachable = [n.id for n in self.session.nodes if n.id not in reachable]
-            if unreachable:
-                errors.append(f"Unreachable nodes: {unreachable}")
-
-        validation = ValidationResult(
-            valid=len(errors) == 0,
-            errors=errors,
-            warnings=warnings,
-        )
-        self._pending_validation = validation
-        return validation
-
-    def _compute_reachable(self, start: str) -> set[str]:
-        """Compute all nodes reachable from start."""
-        reachable = set()
-        to_visit = [start]
-
-        while to_visit:
-            current = to_visit.pop()
-            if current in reachable:
-                continue
-            reachable.add(current)
-
-            for edge in self.session.edges:
-                if edge.source == current:
-                    to_visit.append(edge.target)
-
-            # Also follow router routes
-            for node in self.session.nodes:
-                if node.id == current and node.routes:
-                    for target in node.routes.values():
-                        to_visit.append(target)
-
-        return reachable
-
-    def add_test(self, test: TestCase) -> None:
-        """Add a test case."""
-        self.session.test_cases.append(test)
-        self._save_session()
-
-    def run_test(
-        self,
-        test: TestCase,
-        executor_factory: Callable,
-    ) -> TestResult:
-        """
-        Run a single test case.
-
-        executor_factory should return a configured GraphExecutor.
-        """
-        self._require_phase([BuildPhase.ADDING_NODES, BuildPhase.ADDING_EDGES, BuildPhase.TESTING])
-        self.session.phase = BuildPhase.TESTING
-
-        try:
-            # Build temporary graph for testing
-            graph = self._build_graph()
-            executor = executor_factory()
-
-            # Run the test
-            import asyncio
-
-            result = asyncio.run(
-                executor.execute(
-                    graph=graph,
-                    goal=self.session.goal,
-                    input_data=test.input,
-                )
-            )
-
-            # Check result
-            passed = result.success
-            if test.expected_output is not None:
-                passed = passed and (result.output.get("result") == test.expected_output)
-            if test.expected_contains:
-                output_str = str(result.output)
-                passed = passed and (test.expected_contains in output_str)
-
-            test_result = TestResult(
-                test_id=test.id,
-                passed=passed,
-                actual_output=result.output,
-                execution_path=result.path,
-            )
-
-        except Exception as e:
-            test_result = TestResult(
-                test_id=test.id,
-                passed=False,
-                error=str(e),
-            )
-
-        self.session.test_results.append(test_result)
-        self._save_session()
-
-        return test_result
-
-    def run_all_tests(self, executor_factory: Callable) -> list[TestResult]:
-        """Run all test cases."""
-        results = []
-        for test in self.session.test_cases:
-            result = self.run_test(test, executor_factory)
-            results.append(result)
-        return results
-
-    # =========================================================================
-    # APPROVAL
-    # =========================================================================
-
-    def approve(self, comment: str) -> bool:
-        """
-        Approve the current pending change.
-
-        Must have a passing validation to approve.
-        Returns True if approved, False if validation failed.
-        """
-        if self._pending_validation is None:
-            raise RuntimeError("Nothing to approve. Run validation first.")
-
-        if not self._pending_validation.valid:
-            return False
-
-        self.session.approvals.append(
-            {
-                "phase": self.session.phase.value,
-                "comment": comment,
-                "timestamp": datetime.now().isoformat(),
-                "validation": self._pending_validation.model_dump(),
-            }
-        )
-
-        # Advance phase if appropriate
-        if self.session.phase == BuildPhase.GOAL_DRAFT:
-            self.session.phase = BuildPhase.GOAL_APPROVED
-
-        self._pending_validation = None
-        self._save_session()
-
-        return True
-
-    def final_approve(self, comment: str) -> bool:
-        """
-        Final approval for the complete graph.
-
-        Requires all tests to pass.
-        """
-        # Run final validation
-        validation = self.validate()
-        if not validation.valid:
-            self._pending_validation = validation
-            return False
-
-        # Check test results
-        if self.session.test_cases:
-            failed_tests = [t for t in self.session.test_results if not t.passed]
-            if failed_tests:
-                self._pending_validation = ValidationResult(
-                    valid=False,
-                    errors=[f"Failed tests: {[t.test_id for t in failed_tests]}"],
-                )
-                return False
-
-        self.session.phase = BuildPhase.APPROVED
-        self.session.approvals.append(
-            {
-                "phase": "final",
-                "comment": comment,
-                "timestamp": datetime.now().isoformat(),
-            }
-        )
-
-        self._save_session()
-        return True
-
-    # =========================================================================
-    # EXPORT
-    # =========================================================================
-
-    def export(self) -> GraphSpec:
-        """
-        Export the approved graph.
-
-        Requires final approval.
-        """
-        self._require_phase([BuildPhase.APPROVED])
-
-        graph = self._build_graph()
-
-        self.session.phase = BuildPhase.EXPORTED
-        self._save_session()
-
-        return graph
-
-    def _build_graph(self) -> GraphSpec:
-        """Build a GraphSpec from current session."""
-        # Determine entry node
-        entry_node = None
-        for node in self.session.nodes:
-            if not any(e.target == node.id for e in self.session.edges):
-                entry_node = node.id
-                break
-
-        # Determine terminal nodes
-        terminal_nodes = []
-        for node in self.session.nodes:
-            if not any(e.source == node.id for e in self.session.edges):
-                terminal_nodes.append(node.id)
-
-        # Collect all memory keys
-        memory_keys = set()
-        for node in self.session.nodes:
-            memory_keys.update(node.input_keys)
-            memory_keys.update(node.output_keys)
-
-        return GraphSpec(
-            id=f"{self.session.name}-graph",
-            goal_id=self.session.goal.id if self.session.goal else "",
-            entry_node=entry_node or "",
-            terminal_nodes=terminal_nodes,
-            nodes=self.session.nodes,
-            edges=self.session.edges,
-            memory_keys=list(memory_keys),
-        )
-
-    def export_to_file(self, path: Path | str) -> None:
-        """Export the graph to a Python file."""
-        self._require_phase([BuildPhase.APPROVED, BuildPhase.EXPORTED])
-
-        graph = self._build_graph()
-
-        # Generate Python code
-        code = self._generate_code(graph)
-
-        Path(path).write_text(code)
-        self.session.phase = BuildPhase.EXPORTED
-        self._save_session()
-
-    def _generate_code(self, graph: GraphSpec) -> str:
-        """Generate Python code for the graph."""
-        lines = [
-            '"""',
-            f"Generated agent: {self.session.name}",
-            f"Generated at: {datetime.now().isoformat()}",
-            '"""',
-            "",
-            "from framework.graph import (",
-            "    Goal, SuccessCriterion, Constraint,",
-            "    NodeSpec, EdgeSpec, EdgeCondition,",
-            ")",
-            "from framework.graph.edge import GraphSpec",
-            "from framework.graph.goal import GoalStatus",
-            "",
-            "",
-            "# Goal",
-        ]
-
-        if self.session.goal:
-            goal_json = self.session.goal.model_dump_json(indent=4)
-            lines.append("GOAL = Goal.model_validate_json('''")
-            lines.append(goal_json)
-            lines.append("''')")
-        else:
-            lines.append("GOAL = None")
-
-        lines.extend(
-            [
-                "",
-                "",
-                "# Nodes",
-                "NODES = [",
-            ]
-        )
-
-        for node in self.session.nodes:
-            node_json = node.model_dump_json(indent=4)
-            lines.append("    NodeSpec.model_validate_json('''")
-            lines.append(node_json)
-            lines.append("    '''),")
-
-        lines.extend(
-            [
-                "]",
-                "",
-                "",
-                "# Edges",
-                "EDGES = [",
-            ]
-        )
-
-        for edge in self.session.edges:
-            edge_json = edge.model_dump_json(indent=4)
-            lines.append("    EdgeSpec.model_validate_json('''")
-            lines.append(edge_json)
-            lines.append("    '''),")
-
-        lines.extend(
-            [
-                "]",
-                "",
-                "",
-                "# Graph",
-            ]
-        )
-
-        graph_json = graph.model_dump_json(indent=4)
-        lines.append("GRAPH = GraphSpec.model_validate_json('''")
-        lines.append(graph_json)
-        lines.append("''')")
-
-        return "\n".join(lines)
-
-    # =========================================================================
-    # SESSION MANAGEMENT
-    # =========================================================================
-
-    def _require_phase(self, allowed: list[BuildPhase]) -> None:
-        """Ensure we're in an allowed phase."""
-        if self.session.phase not in allowed:
-            raise RuntimeError(
-                f"Cannot perform this action in phase '{self.session.phase.value}'. "
-                f"Allowed phases: {[p.value for p in allowed]}"
-            )
-
-    def _save_session(self) -> None:
-        """Save session to disk."""
-        self.session.updated_at = datetime.now()
-        path = self.storage_path / f"{self.session.id}.json"
-        path.write_text(self.session.model_dump_json(indent=2))
-
-    def _load_session(self, session_id: str) -> BuildSession:
-        """Load session from disk."""
-        path = self.storage_path / f"{session_id}.json"
-        if not path.exists():
-            raise FileNotFoundError(f"Session not found: {session_id}")
-        return BuildSession.model_validate_json(path.read_text())
-
-    @classmethod
-    def list_sessions(cls, storage_path: Path | str | None = None) -> list[str]:
-        """List all saved sessions."""
-        path = Path(storage_path) if storage_path else Path.home() / ".core" / "builds"
-        if not path.exists():
-            return []
-        return [f.stem for f in path.glob("*.json")]
-
-    # =========================================================================
-    # STATUS
-    # =========================================================================
-
-    def status(self) -> dict[str, Any]:
-        """Get current build status."""
-        return {
-            "session_id": self.session.id,
-            "name": self.session.name,
-            "phase": self.session.phase.value,
-            "goal": self.session.goal.name if self.session.goal else None,
-            "nodes": len(self.session.nodes),
-            "edges": len(self.session.edges),
-            "tests": len(self.session.test_cases),
-            "tests_passed": sum(1 for t in self.session.test_results if t.passed),
-            "approvals": len(self.session.approvals),
-            "pending_validation": self._pending_validation.model_dump()
-            if self._pending_validation
-            else None,
-        }
-
-    def show(self) -> str:
-        """Show current graph as text."""
-        lines = [
-            f"=== Build: {self.session.name} ===",
-            f"Phase: {self.session.phase.value}",
-            "",
-        ]
-
-        if self.session.goal:
-            lines.extend(
-                [
-                    f"Goal: {self.session.goal.name}",
-                    f"  {self.session.goal.description}",
-                    "",
-                ]
-            )
-
-        if self.session.nodes:
-            lines.append("Nodes:")
-            for node in self.session.nodes:
-                lines.append(f"  [{node.id}] {node.name} ({node.node_type})")
-            lines.append("")
-
-        if self.session.edges:
-            lines.append("Edges:")
-            for edge in self.session.edges:
-                lines.append(f"  {edge.source} --{edge.condition.value}--> {edge.target}")
-            lines.append("")
-
-        if self._pending_validation:
-            lines.append("Pending Validation:")
-            lines.append(f"  Valid: {self._pending_validation.valid}")
-            for err in self._pending_validation.errors:
-                lines.append(f"  ERROR: {err}")
-            for warn in self._pending_validation.warnings:
-                lines.append(f"  WARN: {warn}")
-
-        return "\n".join(lines)

core/framework/cli.py

@@ -6,14 +6,13 @@ Usage:
     hive info exports/my-agent
     hive validate exports/my-agent
     hive list exports/
-    hive dispatch exports/ --input '{"key": "value"}'
     hive shell exports/my-agent
 
 Testing commands:
     hive test-run <agent_path> --goal <goal_id>
-    hive test-debug <goal_id> <test_id>
-    hive test-list <goal_id>
-    hive test-stats <goal_id>
+    hive test-debug <agent_path> <test_name>
+    hive test-list <agent_path>
+    hive test-stats <agent_path>
 """
 
 import argparse
@@ -44,11 +43,25 @@ def _configure_paths():
         if exports_str not in sys.path:
             sys.path.insert(0, exports_str)
 
+    # Add examples/templates/ to sys.path so template agents are importable
+    templates_dir = project_root / "examples" / "templates"
+    if templates_dir.is_dir():
+        templates_str = str(templates_dir)
+        if templates_str not in sys.path:
+            sys.path.insert(0, templates_str)
+
     # Ensure core/ is also in sys.path (for non-editable-install scenarios)
     core_str = str(project_root / "core")
     if (project_root / "core").is_dir() and core_str not in sys.path:
         sys.path.insert(0, core_str)
 
+    # Add core/framework/agents/ so framework agents are importable as top-level packages
+    framework_agents_dir = project_root / "core" / "framework" / "agents"
+    if framework_agents_dir.is_dir():
+        fa_str = str(framework_agents_dir)
+        if fa_str not in sys.path:
+            sys.path.insert(0, fa_str)
+
 
 def main():
     _configure_paths()
@@ -65,7 +78,7 @@ def main():
 
     subparsers = parser.add_subparsers(dest="command", required=True)
 
-    # Register runner commands (run, info, validate, list, dispatch, shell)
+    # Register runner commands (run, info, validate, list, shell)
     from framework.runner.cli import register_commands
 
     register_commands(subparsers)
@@ -75,6 +88,21 @@ def main():
 
     register_testing_commands(subparsers)
 
+    # Register skill commands (skill list, skill trust, ...)
+    from framework.skills.cli import register_skill_commands
+
+    register_skill_commands(subparsers)
+
+    # Register debugger commands (debugger)
+    from framework.debugger.cli import register_debugger_commands
+
+    register_debugger_commands(subparsers)
+
+    # Register MCP registry commands (mcp install, mcp add, ...)
+    from framework.runner.mcp_registry_cli import register_mcp_commands
+
+    register_mcp_commands(subparsers)
+
     args = parser.parse_args()
 
     if hasattr(args, "func"):