fix: internal tag choice of diction

* test(event_bus): add comprehensive unit tests for EventBus

- Add 38 tests covering all EventBus functionality
- Test subscription management (subscribe/unsubscribe)
- Test event publishing and delivery to subscribers
- Test filtering by stream, node, and execution
- Test concurrency (handler errors, semaphore limits)
- Test history operations (get_history, get_stats)
- Test wait_for async waiting with timeout
- Test convenience publishers (emit_* methods)
- Test AgentEvent dataclass and EventType enum

Fixes #4782

* test(event_bus): add comprehensive unit tests for EventBus

- Add 41 tests covering all EventBus functionality
- Test subscription management (subscribe/unsubscribe)
- Test event publishing and delivery to subscribers
- Test filtering by stream, node, execution, and graph_id
- Test concurrency (handler errors, semaphore limits)
- Test history operations (get_history, get_stats)
- Test wait_for async waiting with timeout
- Test convenience publishers including new emit_tool_doom_loop
  and emit_escalation_requested methods
- Test AgentEvent dataclass with graph_id field
- Test EventType enum including NODE_TOOL_DOOM_LOOP and ESCALATION_REQUESTED

Fixes #4782

* test(event_bus): add tests for new worker monitoring events

Add tests for newly added emit methods:
- emit_worker_escalation_ticket (judge → queen escalation)
- emit_queen_intervention_requested (queen → operator escalation)
- emit_llm_turn_complete (LLM turn metadata)
- emit_node_action_plan (node planning)

Update test_key_event_types_exist to include:
- WORKER_ESCALATION_TICKET, QUEEN_INTERVENTION_REQUESTED
- LLM_TURN_COMPLETE, NODE_ACTION_PLAN
- WORKER_LOADED, CREDENTIALS_REQUIRED

Total: 45 tests (up from 41)

* test(event_bus): add tests for run_id, subagent_report, and new event types

- Add tests for run_id field in AgentEvent.to_dict()
- Add test for emit_subagent_report convenience method
- Update test_key_event_types_exist with new event types
- Total: 48 tests

* fix(test): remove tests for deleted EventBus methods and fix enum names

Remove test_emit_worker_escalation_ticket and
test_emit_queen_intervention_requested (methods were removed in recent
refactors). Fix WORKER_LOADED -> WORKER_GRAPH_LOADED in enum assertions.

* style: add future annotations import and fix empty assertion in test_emit_node_action_plan

.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 +1,113 @@
+#!/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
+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("\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 +1,96 @@
+#!/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
+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 +1,152 @@
+#!/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
+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 +1,180 @@
+#!/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
+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 +1,253 @@
+#!/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
+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 +1,190 @@
+#!/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
+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!</p>'
+                            + '<button id="home-btn">Home Action</button>',
+                        about: '<h1>About Page</h1><p>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 +1,267 @@
+#!/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(
+            "  ⚠ NOT IMPLEMENTED: selector param ignored"
+            f" (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 +1,333 @@
+#!/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/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)

.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

@@ -2,14 +2,22 @@ name: Bounty completed
 description: Awards points and notifies Discord when a bounty PR is merged
 
 on:
-  pull_request:
+  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.pull_request.merged == true &&
-      contains(join(github.event.pull_request.labels.*.name, ','), 'bounty:')
+      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:
@@ -32,6 +40,8 @@ jobs:
           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: ${{ github.event.pull_request.number }}
+          PR_NUMBER: ${{ inputs.pr_number || github.event.pull_request.number }}

.github/workflows/ci.yml

@@ -63,7 +63,7 @@ jobs:
         working-directory: core
         run: |
           uv sync
-          uv run pytest tests/ -v
+          uv run pytest tests/ -v --ignore=tests/dummy_agents
 
   test-tools:
     name: Test Tools (${{ matrix.os }})

.github/workflows/link-discord.yml

@@ -1,126 +0,0 @@
-name: Link Discord account
-description: Auto-creates a PR to add contributor to contributors.yml when a link-discord issue is opened
-
-on:
-  issues:
-    types: [opened]
-
-jobs:
-  link-discord:
-    if: contains(github.event.issue.labels.*.name, 'link-discord')
-    runs-on: ubuntu-latest
-    timeout-minutes: 2
-    permissions:
-      contents: write
-      issues: write
-      pull-requests: write
-
-    steps:
-      - name: Checkout repository
-        uses: actions/checkout@v4
-
-      - name: Parse issue and update contributors.yml
-        uses: actions/github-script@v7
-        with:
-          script: |
-            const fs = require('fs');
-
-            const issue = context.payload.issue;
-            const githubUsername = issue.user.login;
-
-            // Parse the issue body for form fields
-            const body = issue.body || '';
-
-            // Extract Discord ID — look for the numeric value after the "Discord User ID" heading
-            const discordMatch = body.match(/### Discord User ID\s*\n\s*(\d{17,20})/);
-            if (!discordMatch) {
-              await github.rest.issues.createComment({
-                ...context.repo,
-                issue_number: issue.number,
-                body: `Could not find a valid Discord ID in the issue body. Please make sure you entered a numeric ID (17-20 digits), not a username.\n\nExample: \`123456789012345678\``
-              });
-              await github.rest.issues.update({
-                ...context.repo,
-                issue_number: issue.number,
-                state: 'closed',
-                state_reason: 'not_planned'
-              });
-              return;
-            }
-            const discordId = discordMatch[1];
-
-            // Extract display name (optional)
-            const nameMatch = body.match(/### Display Name \(optional\)\s*\n\s*(.+)/);
-            const displayName = nameMatch ? nameMatch[1].trim() : '';
-
-            // Check if user already exists
-            const yml = fs.readFileSync('contributors.yml', 'utf-8');
-            if (yml.includes(`github: ${githubUsername}`)) {
-              await github.rest.issues.createComment({
-                ...context.repo,
-                issue_number: issue.number,
-                body: `@${githubUsername} is already in \`contributors.yml\`. If you need to update your Discord ID, please edit the file directly via PR.`
-              });
-              await github.rest.issues.update({
-                ...context.repo,
-                issue_number: issue.number,
-                state: 'closed',
-                state_reason: 'completed'
-              });
-              return;
-            }
-
-            // Append entry to contributors.yml
-            let entry = `  - github: ${githubUsername}\n    discord: "${discordId}"`;
-            if (displayName && displayName !== '_No response_') {
-              entry += `\n    name: ${displayName}`;
-            }
-            entry += '\n';
-
-            const updated = yml.trimEnd() + '\n' + entry;
-            fs.writeFileSync('contributors.yml', updated);
-
-            // Set outputs for commit step
-            core.exportVariable('GITHUB_USERNAME', githubUsername);
-            core.exportVariable('DISCORD_ID', discordId);
-            core.exportVariable('ISSUE_NUMBER', issue.number.toString());
-
-      - name: Create PR
-        run: |
-          # Check if there are changes
-          if git diff --quiet contributors.yml; then
-            echo "No changes to contributors.yml"
-            exit 0
-          fi
-
-          BRANCH="docs/link-discord-${GITHUB_USERNAME}"
-          git config user.name "github-actions[bot]"
-          git config user.email "41898282+github-actions[bot]@users.noreply.github.com"
-          git checkout -b "$BRANCH"
-          git add contributors.yml
-          git commit -m "docs: link @${GITHUB_USERNAME} to Discord"
-          git push origin "$BRANCH"
-
-          gh pr create \
-            --title "docs: link @${GITHUB_USERNAME} to Discord" \
-            --body "Adds @${GITHUB_USERNAME} (Discord \`${DISCORD_ID}\`) to \`contributors.yml\` for bounty XP tracking.
-
-          Closes #${ISSUE_NUMBER}" \
-            --base main \
-            --head "$BRANCH" \
-            --label "link-discord"
-        env:
-          GH_TOKEN: ${{ secrets.GITHUB_TOKEN }}
-
-      - name: Notify on issue
-        uses: actions/github-script@v7
-        with:
-          script: |
-            const username = process.env.GITHUB_USERNAME;
-            const issueNumber = parseInt(process.env.ISSUE_NUMBER);
-
-            await github.rest.issues.createComment({
-              ...context.repo,
-              issue_number: issueNumber,
-              body: `A PR has been created to link your account. A maintainer will merge it shortly — once merged, you'll receive XP and Discord pings when your bounty PRs are merged.`
-            });

.github/workflows/weekly-leaderboard.yml

@@ -35,6 +35,8 @@ jobs:
           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
@@ -66,11 +70,10 @@ tmp/
 temp/
 
 exports/*
+exports.old*
+artifacts/*
 
 .claude/settings.local.json
-.claude/skills/ship-it/
-
-.venv
 
 docs/github-issues/*
 core/tests/*dumps/*
@@ -78,3 +81,4 @@ core/tests/*dumps/*
 screenshots/*
 
 .gemini/*
+.coverage

.hive/browser-logs/browser-2026-04-04.jsonl

@@ -0,0 +1,9 @@
+{"type": "connection", "event": "connect", "ts": "2026-04-04T01:10:38.245667+00:00", "profile": "default"}
+{"type": "connection", "event": "hello", "details": {"version": "1.0"}, "ts": "2026-04-04T01:10:38.247207+00:00", "profile": "default"}
+{"type": "connection", "event": "disconnect", "ts": "2026-04-04T01:11:57.148273+00:00", "profile": "default"}
+{"type": "connection", "event": "connect", "ts": "2026-04-04T01:12:09.162378+00:00", "profile": "default"}
+{"type": "connection", "event": "hello", "details": {"version": "1.0"}, "ts": "2026-04-04T01:12:09.163899+00:00", "profile": "default"}
+{"type": "connection", "event": "disconnect", "ts": "2026-04-04T01:15:12.826042+00:00", "profile": "default"}
+{"type": "connection", "event": "connect", "ts": "2026-04-04T01:15:30.842533+00:00", "profile": "default"}
+{"type": "connection", "event": "hello", "details": {"version": "1.0"}, "ts": "2026-04-04T01:15:30.845025+00:00", "profile": "default"}
+{"type": "tool_call", "tool": "browser_stop", "params": {"profile": "gcu-browser-worker:3"}, "result": {"ok": true, "status": "not_running", "profile": "gcu-browser-worker:3"}, "ok": true, "duration_ms": 0.01, "ts": "2026-04-04T01:29:04.294954+00:00", "profile": "default"}

CHANGELOG.md

@@ -1,17 +1,149 @@
 # Release Notes
 
+## v0.7.1
+
+**Release Date:** March 13, 2026
+**Tag:** v0.7.1
+
+### Chrome-Native Browser Control
+
+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.
+
+---
+
+### Highlights
+
+#### System Chrome via CDP
+
+The entire GCU browser stack has been rewritten:
+
+- **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
+
+#### Per-Subagent Browser Isolation
+
+Each GCU subagent gets its own Chrome user-data directory, preventing cookie/session cross-contamination:
+
+- Unique browser profiles injected per subagent
+- Profiles cleaned up after top-level GCU node execution
+- Tab origin and age metadata tracked per subagent
+
+#### 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
+### 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
+### Highlights
 
-### Hive Coder -- The Agent That Builds Agents
+#### 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.
 
@@ -30,7 +162,7 @@ The Coder ships with:
 - **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
+#### 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:
 
@@ -44,7 +176,7 @@ await runtime.add_graph("exports/deep_research_agent")
 
 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
+#### TUI Revamp
 
 The Terminal UI gets a ground-up rebuild with five major additions:
 
@@ -54,7 +186,7 @@ The Terminal UI gets a ground-up rebuild with five major additions:
 - **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
+#### Provider-Agnostic LLM Support
 
 Hive is no longer Anthropic-only. v0.5.1 adds first-class support for:
 
@@ -66,9 +198,9 @@ The quickstart script auto-detects Claude Code subscriptions and ZAI Code instal
 
 ---
 
-## What's New
+### What's New
 
-### Architecture & Runtime
+#### 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)
@@ -79,7 +211,7 @@ The quickstart script auto-detects Claude Code subscriptions and ZAI Code instal
 - **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
+#### 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)
@@ -89,7 +221,7 @@ The quickstart script auto-detects Claude Code subscriptions and ZAI Code instal
 - **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
+#### New Tool Integrations
 
 | Tool                   | Description                                                                                                                                                            | Contributor        |
 | ---------------------- | ---------------------------------------------------------------------------------------------------------------------------------------------------------------------- | ------------------ |
@@ -99,7 +231,7 @@ The quickstart script auto-detects Claude Code subscriptions and ZAI Code instal
 | **Google Docs**        | Document creation, reading, and editing with OAuth credential support                                                                                                  | @haliaeetusvocifer |
 | **Gmail enhancements** | Expanded mail operations for inbox management                                                                                                                          | @bryanadenhq       |
 
-### Infrastructure
+#### 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)
@@ -112,7 +244,7 @@ The quickstart script auto-detects Claude Code subscriptions and ZAI Code instal
 
 ---
 
-## Bug Fixes
+### 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)
@@ -125,13 +257,13 @@ The quickstart script auto-detects Claude Code subscriptions and ZAI Code instal
 - Fix email agent version conflicts (@RichardTang-Aden)
 - Fix coder tool timeouts (120s for tests, 300s cap for commands)
 
-## Documentation
+### Documentation
 
 - Clarify installation and prevent root pip install misuse (@paarths-collab)
 
 ---
 
-## Agent Updates
+### 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)
@@ -141,7 +273,7 @@ The quickstart script auto-detects Claude Code subscriptions and ZAI Code instal
 
 ---
 
-## Breaking Changes
+### 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"`)
@@ -150,7 +282,7 @@ The quickstart script auto-detects Claude Code subscriptions and ZAI Code instal
 
 ---
 
-## Community Contributors
+### Community Contributors
 
 A huge thank you to everyone who contributed to this release:
 
@@ -165,14 +297,14 @@ A huge thank you to everyone who contributed to this release:
 
 ---
 
-## Upgrading
+### Upgrading
 
 ```bash
 git pull origin main
 uv sync
 ```
 
-### Migration Guide
+#### Migration Guide
 
 If your agents use deprecated node types, update them:
 
@@ -196,12 +328,3 @@ hive code
 # Or from TUI -- press Ctrl+E to escalate
 hive tui
 ```
-
----
-
-## What's Next
-
-- **Agent-to-agent communication** -- one agent's output triggers another agent's entry point
-- **Cost visibility** -- detailed runtime log of LLM costs per node and per session
-- **Persistent webhook subscriptions** -- survive agent restarts without re-registering
-- **Remote agent deployment** -- run agents as long-lived services with HTTP APIs

Makefile

@@ -1,27 +1,34 @@
-.PHONY: lint format check test install-hooks help frontend-install frontend-dev frontend-build
+.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 and formatter (with auto-fix)
-	cd core && ruff check --fix .
-	cd tools && ruff check --fix .
-	cd core && ruff format .
-	cd tools && ruff format .
+	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 (core + tools, excludes live)
-	cd core && uv run python -m pytest tests/ -v
+	cd core && uv run python -m pytest tests/ -v --ignore=tests/dummy_agents
 	cd tools && uv run python -m pytest -v
 
 test-tools: ## Run tool tests only (mocked, no credentials needed)
@@ -31,7 +38,7 @@ 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 core && uv run python -m pytest tests/ -v --ignore=tests/dummy_agents
 	cd tools && uv run python -m pytest -v
 	cd tools && uv run python -m pytest -m live -s -o "addopts=" --log-cli-level=INFO
 
@@ -46,4 +53,4 @@ frontend-dev: ## Start frontend dev server
 	cd core/frontend && npm run dev
 
 frontend-build: ## Build frontend for production
-	cd core/frontend && npm run build
+	cd core/frontend && npm run build

README.md

@@ -23,11 +23,12 @@
 </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/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" />
@@ -35,37 +36,42 @@
   <img src="https://img.shields.io/badge/Google_Gemini-supported-4285F4?style=flat-square&logo=google" alt="Gemini" />
 </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 autonomous, reliable, self-improving AI agents without hardcoding workflows. Define your goal through conversation with hive coding agent(queen), 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.
 
-[![Hive Demo](https://img.youtube.com/vi/XDOG9fOaLjU/maxresdefault.jpg)](https://www.youtube.com/watch?v=XDOG9fOaLjU)
+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.
+
+https://github.com/user-attachments/assets/bf10edc3-06ba-48b6-98ba-d069b15fb69d
+
 
 ## Who Is Hive For?
 
-Hive is designed for developers and teams who want to build **production-grade AI agents** without manually wiring complex workflows.
+Hive is the multi-agent harness layer for teams moving AI agents from prototype to production. Single agents like Openclaw and Cowork can finish personal jobs pretty well but lack the rigor to fulfil business processes. 
 
 Hive is a good fit if you:
 
 - Want AI agents that **execute real business processes**, not demos
-- Need **fast or high volume agent execution** over open workflow
+- 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 environments**
+- 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 you need:
+Use Hive when the bottleneck is no longer the model but the harness around it:
 
-- Long-running, autonomous agents
-- Strong guardrails, process, and controls
-- Continuous improvement based on failures
-- Multi-agent coordination
-- A framework that evolves with your goals
+- 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
 
@@ -73,7 +79,7 @@ Use Hive when you need:
 - **[Self-Hosting Guide](https://docs.adenhq.com/getting-started/quickstart)** - Deploy Hive on your infrastructure
 - **[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/adenhq/hive/issues)** - Bug reports and feature requests
+- **[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
@@ -84,7 +90,7 @@ Use Hive when you need:
 - 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`
 
-> **Note for Windows Users:** It is strongly recommended to use **WSL (Windows Subsystem for Linux)** or **Git Bash** to run this framework. Some core automation scripts may not execute correctly in standard Command Prompt or PowerShell.
+> **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
 
@@ -98,9 +104,11 @@ Use Hive when you need:
 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:
@@ -108,18 +116,16 @@ This sets up:
 - **framework** - Core agent runtime and graph executor (in `core/.venv`)
 - **aden_tools** - MCP tools for agent capabilities (in `tools/.venv`)
 - **credential store** - Encrypted API key storage (`~/.hive/credentials`)
-- **LLM provider** - Interactive default model configuration
+- **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.
 
-<img width="2500" height="1214" alt="home-screen" src="https://github.com/user-attachments/assets/134d897f-5e75-4874-b00b-e0505f6b45c4" />
-
 ### Build Your First Agent
 
-Type the agent you want to build in the home input box
+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.
 
 <img width="2500" height="1214" alt="Image" src="https://github.com/user-attachments/assets/1ce19141-a78b-46f5-8d64-dbf987e048f4" />
 
@@ -131,7 +137,7 @@ Click "Try a sample agent" and check the templates. You can run a template direc
 
 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.
 
-<img width="2500" height="1214" alt="Image" src="https://github.com/user-attachments/assets/71c38206-2ad5-49aa-bde8-6698d0bc55f5" />
+<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
 
@@ -140,22 +146,21 @@ Now you can run an agent by selecting the agent (either an existing agent or exa
 - **[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 shared memory, local RLM memory, monitoring, tools, and LLM access out of the box
+- **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
-- **Production-Ready** - Self-hostable, built for scale and reliability
 
 ## Integration
 
 <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 is designed to support various types of LLMs, including hosted and local models through LiteLLM-compatible providers.
+- **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 Aden
+## Why Hive
 
-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.
+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
@@ -189,17 +194,6 @@ flowchart LR
     style V6 fill:#fff,stroke:#ed8c00,stroke-width:1px,color:#cc5d00
 ```
 
-### The Hive Advantage
-
-| Traditional Frameworks     | Hive                                   |
-| -------------------------- | -------------------------------------- |
-| Hardcode agent workflows   | Describe goals in natural language     |
-| Manual graph definition    | Auto-generated agent graphs            |
-| 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](docs/key_concepts/goals_outcome.md)** → Describe what you want to achieve in plain English
@@ -378,7 +372,7 @@ This project is licensed under the Apache License 2.0 - see the [LICENSE](LICENS
 
 **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. We recommend using Claude, GLM and Gemini as they have the best performance.
+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?**
 
@@ -386,16 +380,12 @@ 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](docs/key_concepts/evolution.md), 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: 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](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.
@@ -420,6 +410,16 @@ Visit [docs.adenhq.com](https://docs.adenhq.com/) for complete guides, API refer
 
 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.
 
+## Star History
+
+<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>
+
 ---
 
 <p align="center">

commit_message.txt

@@ -1,31 +0,0 @@
-perf: reduce subprocess spawning in quickstart scripts (#4427)
-
-## Problem
-Windows process creation (CreateProcess) is 10-100x slower than Linux fork/exec.
-The quickstart scripts were spawning 4+ separate `uv run python -c "import X"` 
-processes to verify imports, adding ~600ms overhead on Windows.
-
-## Solution
-Consolidated all import checks into a single batch script that checks multiple 
-modules in one subprocess call, reducing spawn overhead by ~75%.
-
-## Changes
-- **New**: `scripts/check_requirements.py` - Batched import checker
-- **New**: `scripts/test_check_requirements.py` - Test suite  
-- **New**: `scripts/benchmark_quickstart.ps1` - Performance benchmark tool
-- **Modified**: `quickstart.ps1` - Updated import verification (2 sections)
-- **Modified**: `quickstart.sh` - Updated import verification
-
-## Performance Impact
-**Benchmark results on Windows:**
-- Before: ~19.8 seconds for import checks
-- After: ~4.9 seconds for import checks
-- **Improvement: 14.9 seconds saved (75.2% faster)**
-
-## Testing
-- ✅ All functional tests pass (`scripts/test_check_requirements.py`)
-- ✅ Quickstart scripts work correctly on Windows
-- ✅ Error handling verified (invalid imports reported correctly)
-- ✅ Performance benchmark confirms 75%+ improvement
-
-Fixes #4427

contributors.yml

@@ -1,27 +0,0 @@
-# Identity mapping: GitHub username -> Discord ID
-#
-# This file links GitHub accounts to Discord accounts for the
-# Integration Bounty Program. When a bounty PR is merged, the
-# GitHub Action uses this file to ping the contributor on Discord.
-#
-# HOW TO ADD YOURSELF:
-# Open a "Link Discord Account" issue:
-# https://github.com/aden-hive/hive/issues/new?template=link-discord.yml
-# A GitHub Action will automatically add your entry here.
-#
-# To find your Discord ID:
-# 1. Open Discord Settings > Advanced > Enable Developer Mode
-# 2. Right-click your name > Copy User ID
-#
-# Format:
-#   - github: your-github-username
-#     discord: "your-discord-id"  # quotes required (it's a number)
-#     name: Your Display Name      # optional
-
-contributors:
-  # - github: example-user
-  #   discord: "123456789012345678"
-  #   name: Example User
-  - github: TimothyZhang7
-    discord: "408460790061072384"
-    name: Timothy@Aden

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:
@@ -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/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

@@ -17,6 +17,7 @@ import http.server
 import json
 import os
 import platform
+import queue
 import secrets
 import subprocess
 import sys
@@ -27,6 +28,7 @@ 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"
@@ -165,11 +167,11 @@ def open_browser(url: str) -> bool:
         if system == "Darwin":
             subprocess.Popen(["open", url], stdout=devnull, stderr=devnull)
         elif system == "Windows":
-            subprocess.Popen(["cmd", "/c", "start", url], stdout=devnull, stderr=devnull)
+            os.startfile(url)  # type: ignore[attr-defined]
         else:
             subprocess.Popen(["xdg-open", url], stdout=devnull, stderr=devnull)
         return True
-    except OSError:
+    except (AttributeError, OSError):
         return False
 
 
@@ -266,6 +268,71 @@ def parse_manual_input(value: str, expected_state: str) -> str | None:
     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()
@@ -315,41 +382,28 @@ def main() -> int:
 
         # Start callback server in background
         callback_result: list[str | None] = [None]
+        callback_done = threading.Event()
 
         def run_server() -> None:
-            callback_result[0] = wait_for_callback(state, timeout_secs=120)
+            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()
 
-        # Also accept manual input in parallel
-        # We poll for both the server result and stdin
         try:
-            import select
-
-            while server_thread.is_alive():
-                # Check if stdin has data (non-blocking on unix)
-                if hasattr(select, "select"):
-                    ready, _, _ = select.select([sys.stdin], [], [], 0.5)
-                    if ready:
-                        manual = sys.stdin.readline()
-                        if manual.strip():
-                            code = parse_manual_input(manual, state)
-                            if code:
-                                break
-                else:
-                    time.sleep(0.5)
-
-                if callback_result[0]:
-                    code = callback_result[0]
-                    break
-        except (KeyboardInterrupt, EOFError):
+            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
-
-        if not code:
-            code = callback_result[0]
     else:
         # Manual paste mode
         try:

core/demos/event_loop_wss_demo.py

@@ -1,740 +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, 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()),
-)
-
-
-# -------------------------------------------------------------------------
-# 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 with client_facing blocking."""
-    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,
-    )
-
-    # -- Per-connection state -----------------------------------------------
-    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",
-        client_facing=True,
-        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."
-        ),
-    )
-
-    # -- Ready callback: subscribe to CLIENT_INPUT_REQUESTED on the bus ---
-    async def on_input_requested(event):
-        try:
-            await websocket.send(json.dumps({"type": "ready"}))
-        except Exception:
-            pass
-
-    bus.subscribe(
-        event_types=[EventType.CLIENT_INPUT_REQUESTED],
-        handler=on_input_requested,
-    )
-
-    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,
-            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 node and wait for the loop task to finish."""
-        nonlocal node, loop_task
-        if loop_task and not loop_task.done():
-            if node:
-                node.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)
-                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 into the running loop
-                logger.info(f"Injecting message: {topic}")
-                await node.inject_event(topic)
-
-    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=30,
-            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=30,
-            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

@@ -27,7 +27,7 @@ class GreeterNode(NodeProtocol):
     async def execute(self, ctx: NodeContext) -> NodeResult:
         name = ctx.input_data.get("name", "World")
         greeting = f"Hello, {name}!"
-        ctx.memory.write("greeting", greeting)
+        ctx.buffer.write("greeting", greeting)
         return NodeResult(success=True, output={"greeting": greeting})
 
 
@@ -35,9 +35,9 @@ class UppercaserNode(NodeProtocol):
     """Convert text to uppercase."""
 
     async def execute(self, ctx: NodeContext) -> NodeResult:
-        greeting = ctx.input_data.get("greeting") or ctx.memory.read("greeting") or ""
+        greeting = ctx.input_data.get("greeting") or ctx.buffer.read("greeting") or ""
         result = greeting.upper()
-        ctx.memory.write("final_greeting", result)
+        ctx.buffer.write("final_greeting", result)
         return NodeResult(success=True, output={"final_greeting": result})
 
 

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)

core/framework/__init__.py

@@ -1,67 +1,23 @@
-"""
-Aden Hive Framework: A goal-driven agent runtime optimized for Builder observability.
+"""Hive Agent Framework.
 
-The runtime is designed around DECISIONS, not just actions. Every significant
-choice the agent makes is captured with:
-- What it was trying to do (intent)
-- What options it considered
-- What it chose and why
-- What happened as a result
-- Whether that was good or bad (evaluated post-hoc)
-
-This gives the Builder LLM the information it needs to improve agent behavior.
-
-## Testing Framework
-
-The framework includes a Goal-Based Testing system (Goal → Agent → Eval):
-- Generate tests from Goal success_criteria and constraints
-- Mandatory user approval before tests are stored
-- Parallel test execution with error categorization
-- Debug tools with fix suggestions
-
-See `framework.testing` for details.
+Core classes:
+    AgentHost      -- hosts agents, manages entry points and pipeline
+    Orchestrator   -- routes between nodes in a graph
+    AgentLoop      -- the LLM + tool execution loop (one per node)
+    AgentLoader    -- loads agent.json from disk, builds pipeline
+    DecisionTracker -- records decisions for post-hoc analysis
 """
 
-from framework.llm import AnthropicProvider, LLMProvider
-from framework.runner import AgentOrchestrator, AgentRunner
-from framework.runtime.core import Runtime
-from framework.schemas.decision import Decision, DecisionEvaluation, Option, Outcome
-from framework.schemas.run import Problem, Run, RunSummary
-
-# Testing framework
-from framework.testing import (
-    ApprovalStatus,
-    DebugTool,
-    ErrorCategory,
-    Test,
-    TestResult,
-    TestStorage,
-    TestSuiteResult,
-)
+from framework.agent_loop import AgentLoop
+from framework.host import AgentHost
+from framework.loader import AgentLoader
+from framework.orchestrator import Orchestrator
+from framework.tracker import DecisionTracker
 
 __all__ = [
-    # Schemas
-    "Decision",
-    "Option",
-    "Outcome",
-    "DecisionEvaluation",
-    "Run",
-    "RunSummary",
-    "Problem",
-    # Runtime
-    "Runtime",
-    # LLM
-    "LLMProvider",
-    "AnthropicProvider",
-    # Runner
-    "AgentRunner",
-    "AgentOrchestrator",
-    # Testing
-    "Test",
-    "TestResult",
-    "TestSuiteResult",
-    "TestStorage",
-    "ApprovalStatus",
-    "ErrorCategory",
-    "DebugTool",
+    "AgentHost",
+    "AgentLoader",
+    "AgentLoop",
+    "DecisionTracker",
+    "Orchestrator",
 ]

core/framework/agent_loop/__init__.py

@@ -0,0 +1,32 @@
+"""Agent loop -- the core agent execution primitive."""
+
+from framework.agent_loop.conversation import (  # noqa: F401
+    ConversationStore,
+    Message,
+    NodeConversation,
+)
+
+# Lazy import to avoid circular dependency with graph/event_loop/
+# (graph/event_loop/* imports framework.graph.conversation which is a shim
+# pointing here, which would trigger agent_loop.py loading, which imports
+# graph/event_loop/* again)
+
+
+def __getattr__(name: str):
+    if name in ("AgentLoop", "JudgeProtocol", "JudgeVerdict", "LoopConfig", "OutputAccumulator"):
+        from framework.agent_loop.agent_loop import (
+            AgentLoop,
+            JudgeProtocol,
+            JudgeVerdict,
+            LoopConfig,
+            OutputAccumulator,
+        )
+        _exports = {
+            "AgentLoop": AgentLoop,
+            "JudgeProtocol": JudgeProtocol,
+            "JudgeVerdict": JudgeVerdict,
+            "LoopConfig": LoopConfig,
+            "OutputAccumulator": OutputAccumulator,
+        }
+        return _exports[name]
+    raise AttributeError(f"module {__name__!r} has no attribute {name!r}")

core/framework/agent_loop/conversation.py

@@ -8,6 +8,13 @@ from dataclasses import dataclass
 from pathlib import Path
 from typing import Any, Literal, Protocol, runtime_checkable
 
+LEGACY_RUN_ID = "__legacy_run__"
+
+
+def is_legacy_run_id(run_id: str | None) -> bool:
+    """True when run_id represents pre-migration (no run boundary) data."""
+    return run_id is None or run_id == LEGACY_RUN_ID
+
 
 @dataclass
 class Message:
@@ -33,10 +40,22 @@ class Message:
     is_transition_marker: bool = False
     # True when this message is real human input (from /chat), not a system prompt
     is_client_input: bool = False
+    # Optional image content blocks (e.g. from browser_screenshot)
+    image_content: list[dict[str, Any]] | None = None
+    # True when message contains an activated skill body (AS-10: never prune)
+    is_skill_content: bool = False
+    # Logical worker run identifier for shared-session persistence
+    run_id: str | None = None
 
     def to_llm_dict(self) -> dict[str, Any]:
         """Convert to OpenAI-format message dict."""
         if self.role == "user":
+            if self.image_content:
+                blocks: list[dict[str, Any]] = []
+                if self.content:
+                    blocks.append({"type": "text", "text": self.content})
+                blocks.extend(self.image_content)
+                return {"role": "user", "content": blocks}
             return {"role": "user", "content": self.content}
 
         if self.role == "assistant":
@@ -47,6 +66,15 @@ class Message:
 
         # role == "tool"
         content = f"ERROR: {self.content}" if self.is_error else self.content
+        if self.image_content:
+            # Multimodal tool result: text + image content blocks
+            blocks: list[dict[str, Any]] = [{"type": "text", "text": content}]
+            blocks.extend(self.image_content)
+            return {
+                "role": "tool",
+                "tool_call_id": self.tool_use_id,
+                "content": blocks,
+            }
         return {
             "role": "tool",
             "tool_call_id": self.tool_use_id,
@@ -72,6 +100,10 @@ class Message:
             d["is_transition_marker"] = self.is_transition_marker
         if self.is_client_input:
             d["is_client_input"] = self.is_client_input
+        if self.image_content is not None:
+            d["image_content"] = self.image_content
+        if self.run_id is not None:
+            d["run_id"] = self.run_id
         return d
 
     @classmethod
@@ -87,9 +119,41 @@ class Message:
             phase_id=data.get("phase_id"),
             is_transition_marker=data.get("is_transition_marker", False),
             is_client_input=data.get("is_client_input", False),
+            image_content=data.get("image_content"),
+            run_id=data.get("run_id"),
         )
 
 
+def _normalize_cursor(cursor: dict[str, Any] | None) -> dict[str, Any]:
+    """Normalize legacy and run-scoped cursor formats into one flat shape."""
+    return dict(cursor) if cursor else {}
+
+
+def get_cursor_next_seq(cursor: dict[str, Any] | None) -> int | None:
+    next_seq = (cursor or {}).get("next_seq")
+    return next_seq if isinstance(next_seq, int) else None
+
+
+def update_cursor_next_seq(cursor: dict[str, Any] | None, next_seq: int) -> dict[str, Any]:
+    updated = dict(cursor or {})
+    updated["next_seq"] = next_seq
+    return updated
+
+
+def get_run_cursor(cursor: dict[str, Any] | None, run_id: str | None) -> dict[str, Any] | None:
+    return dict(cursor) if cursor else None
+
+
+def update_run_cursor(
+    cursor: dict[str, Any] | None,
+    run_id: str | None,
+    values: dict[str, Any],
+) -> dict[str, Any]:
+    updated = dict(cursor or {})
+    updated.update(values)
+    return updated
+
+
 def _extract_spillover_filename(content: str) -> str | None:
     """Extract spillover filename from a tool result annotation.
 
@@ -239,7 +303,7 @@ class ConversationStore(Protocol):
 
     async def read_cursor(self) -> dict[str, Any] | None: ...
 
-    async def delete_parts_before(self, seq: int) -> None: ...
+    async def delete_parts_before(self, seq: int, run_id: str | None = None) -> None: ...
 
     async def close(self) -> None: ...
 
@@ -260,7 +324,7 @@ def _try_extract_key(content: str, key: str) -> str | None:
     3. Colon format: ``key: value``.
     4. Equals format: ``key = value``.
     """
-    from framework.graph.node import find_json_object
+    from framework.orchestrator.node import find_json_object
 
     # 1. Whole message is JSON
     try:
@@ -307,13 +371,14 @@ class NodeConversation:
     def __init__(
         self,
         system_prompt: str = "",
-        max_history_tokens: int = 32000,
+        max_context_tokens: int = 32000,
         compaction_threshold: float = 0.8,
         output_keys: list[str] | None = None,
         store: ConversationStore | None = None,
+        run_id: str | None = None,
     ) -> None:
         self._system_prompt = system_prompt
-        self._max_history_tokens = max_history_tokens
+        self._max_context_tokens = max_context_tokens
         self._compaction_threshold = compaction_threshold
         self._output_keys = output_keys
         self._store = store
@@ -322,6 +387,7 @@ class NodeConversation:
         self._meta_persisted: bool = False
         self._last_api_input_tokens: int | None = None
         self._current_phase: str | None = None
+        self._run_id: str | None = run_id
 
     # --- Properties --------------------------------------------------------
 
@@ -373,17 +439,23 @@ class NodeConversation:
         *,
         is_transition_marker: bool = False,
         is_client_input: bool = False,
+        image_content: list[dict[str, Any]] | None = None,
     ) -> Message:
         msg = Message(
             seq=self._next_seq,
             role="user",
             content=content,
             phase_id=self._current_phase,
+            run_id=self._run_id,
             is_transition_marker=is_transition_marker,
             is_client_input=is_client_input,
+            image_content=image_content,
         )
         self._messages.append(msg)
         self._next_seq += 1
+        # Invalidate stale API token count so estimate_tokens() uses
+        # the char-based heuristic which reflects the new message.
+        self._last_api_input_tokens = None
         await self._persist(msg)
         return msg
 
@@ -398,9 +470,11 @@ class NodeConversation:
             content=content,
             tool_calls=tool_calls,
             phase_id=self._current_phase,
+            run_id=self._run_id,
         )
         self._messages.append(msg)
         self._next_seq += 1
+        self._last_api_input_tokens = None
         await self._persist(msg)
         return msg
 
@@ -409,6 +483,8 @@ class NodeConversation:
         tool_use_id: str,
         content: str,
         is_error: bool = False,
+        image_content: list[dict[str, Any]] | None = None,
+        is_skill_content: bool = False,
     ) -> Message:
         msg = Message(
             seq=self._next_seq,
@@ -417,9 +493,13 @@ class NodeConversation:
             tool_use_id=tool_use_id,
             is_error=is_error,
             phase_id=self._current_phase,
+            image_content=image_content,
+            is_skill_content=is_skill_content,
+            run_id=self._run_id,
         )
         self._messages.append(msg)
         self._next_seq += 1
+        self._last_api_input_tokens = None
         await self._persist(msg)
         return msg
 
@@ -500,12 +580,15 @@ class NodeConversation:
 
         Uses actual API input token count when available (set via
         :meth:`update_token_count`), otherwise falls back to a
-        ``total_chars / 4`` heuristic that includes both message content
-        AND tool_call argument sizes.
+        character-based heuristic that includes message content, tool_call
+        arguments, and image blocks.  The heuristic applies a 4/3 safety
+        margin to avoid under-counting (inspired by Claude Code's compact
+        service).
         """
         if self._last_api_input_tokens is not None:
             return self._last_api_input_tokens
         total_chars = 0
+        image_tokens = 0
         for m in self._messages:
             total_chars += len(m.content)
             if m.tool_calls:
@@ -513,7 +596,11 @@ class NodeConversation:
                     func = tc.get("function", {})
                     total_chars += len(func.get("arguments", ""))
                     total_chars += len(func.get("name", ""))
-        return total_chars // 4
+            if m.image_content:
+                # Images/documents have a fixed token cost per block
+                image_tokens += len(m.image_content) * 2000
+        # Apply 4/3 safety margin to character-based estimate
+        return (total_chars * 4) // (3 * 4) + image_tokens
 
     def update_token_count(self, actual_input_tokens: int) -> None:
         """Store actual API input token count for more accurate compaction.
@@ -525,16 +612,16 @@ class NodeConversation:
         self._last_api_input_tokens = actual_input_tokens
 
     def usage_ratio(self) -> float:
-        """Current token usage as a fraction of *max_history_tokens*.
+        """Current token usage as a fraction of *max_context_tokens*.
 
-        Returns 0.0 when ``max_history_tokens`` is zero (unlimited).
+        Returns 0.0 when ``max_context_tokens`` is zero (unlimited).
         """
-        if self._max_history_tokens <= 0:
+        if self._max_context_tokens <= 0:
             return 0.0
-        return self.estimate_tokens() / self._max_history_tokens
+        return self.estimate_tokens() / self._max_context_tokens
 
     def needs_compaction(self) -> bool:
-        return self.estimate_tokens() >= self._max_history_tokens * self._compaction_threshold
+        return self.estimate_tokens() >= self._max_context_tokens * self._compaction_threshold
 
     # --- Output-key extraction ---------------------------------------------
 
@@ -610,8 +697,15 @@ class NodeConversation:
                 continue
             if msg.is_error:
                 continue  # never prune errors
+            if msg.is_skill_content:
+                continue  # never prune activated skill instructions (AS-10)
             if msg.content.startswith("[Pruned tool result"):
                 continue  # already pruned
+            # Tiny results (set_output acks, confirmations) — pruning
+            # saves negligible space but makes the LLM think the call
+            # failed, causing costly retries.
+            if len(msg.content) < 100:
+                continue
 
             # Phase-aware: protect current phase messages
             if self._current_phase and msg.phase_id == self._current_phase:
@@ -653,6 +747,7 @@ class NodeConversation:
                 is_error=msg.is_error,
                 phase_id=msg.phase_id,
                 is_transition_marker=msg.is_transition_marker,
+                run_id=msg.run_id,
             )
             count += 1
 
@@ -729,14 +824,14 @@ class NodeConversation:
             summary_seq = self._next_seq
             self._next_seq += 1
 
-        summary_msg = Message(seq=summary_seq, role="user", content=summary)
+        summary_msg = Message(seq=summary_seq, role="user", content=summary, run_id=self._run_id)
 
         # Persist
         if self._store:
             delete_before = recent_messages[0].seq if recent_messages else self._next_seq
             await self._store.delete_parts_before(delete_before)
             await self._store.write_part(summary_msg.seq, summary_msg.to_storage_dict())
-            await self._store.write_cursor({"next_seq": self._next_seq})
+            await self._write_next_seq()
 
         self._messages = [summary_msg] + recent_messages
         self._last_api_input_tokens = None  # reset; next LLM call will recalibrate
@@ -794,6 +889,15 @@ class NodeConversation:
         freeform_lines: list[str] = []
         collapsed_msgs: list[Message] = []
 
+        # Collect all tool_use IDs present in old messages so we can detect
+        # orphaned tool results whose parent assistant message was already
+        # compacted away (API invariant protection).
+        old_tc_ids: set[str] = set()
+        for msg in old_messages:
+            if msg.tool_calls:
+                for tc in msg.tool_calls:
+                    old_tc_ids.add(tc.get("id", ""))
+
         if aggressive:
             # Aggressive: only keep set_output tool pairs and error results.
             # Everything else is collapsed into a tool-call history summary.
@@ -815,9 +919,17 @@ class NodeConversation:
                 else:
                     collapsible_tc_ids |= tc_ids
 
+            # Skill content and transition markers are always protected
+            for msg in old_messages:
+                if msg.role == "tool" and msg.is_skill_content and msg.tool_use_id:
+                    protected_tc_ids.add(msg.tool_use_id)
+
             # Second pass: classify all messages
             for msg in old_messages:
-                if msg.role == "tool":
+                if msg.is_transition_marker:
+                    # Transition markers are always kept (phase boundaries)
+                    kept_structural.append(msg)
+                elif msg.role == "tool":
                     tc_id = msg.tool_use_id or ""
                     if tc_id in protected_tc_ids:
                         kept_structural.append(msg)
@@ -826,6 +938,12 @@ class NodeConversation:
                         kept_structural.append(msg)
                         # Protect the parent assistant message too
                         protected_tc_ids.add(tc_id)
+                    elif msg.is_skill_content:
+                        kept_structural.append(msg)
+                    elif tc_id and tc_id not in old_tc_ids:
+                        # Orphaned tool result — parent tool_use not in old msgs.
+                        # Keep it to maintain API invariants.
+                        kept_structural.append(msg)
                     else:
                         collapsed_msgs.append(msg)
                 elif msg.role == "assistant" and msg.tool_calls:
@@ -842,6 +960,7 @@ class NodeConversation:
                                 is_error=msg.is_error,
                                 phase_id=msg.phase_id,
                                 is_transition_marker=msg.is_transition_marker,
+                                run_id=msg.run_id,
                             )
                         )
                     else:
@@ -856,7 +975,10 @@ class NodeConversation:
         else:
             # Standard mode: keep all tool call pairs as structural
             for msg in old_messages:
-                if msg.role == "tool":
+                if msg.is_transition_marker:
+                    # Transition markers are always kept (phase boundaries)
+                    kept_structural.append(msg)
+                elif msg.role == "tool":
                     kept_structural.append(msg)
                 elif msg.role == "assistant" and msg.tool_calls:
                     compact_tcs = _compact_tool_calls(msg.tool_calls)
@@ -869,6 +991,7 @@ class NodeConversation:
                             is_error=msg.is_error,
                             phase_id=msg.phase_id,
                             is_transition_marker=msg.is_transition_marker,
+                            run_id=msg.run_id,
                         )
                     )
                 else:
@@ -901,8 +1024,7 @@ class NodeConversation:
             full_path = str((spill_path / conv_filename).resolve())
             ref_parts.append(
                 f"[Previous conversation saved to '{full_path}'. "
-                f"Use load_data('{conv_filename}'), read_file('{full_path}'), "
-                f"or run_command('cat \"{full_path}\"') to review if needed.]"
+                f"Use load_data('{conv_filename}') to review if needed.]"
             )
         elif not collapsed_msgs:
             ref_parts.append("[Previous freeform messages compacted.]")
@@ -927,7 +1049,7 @@ class NodeConversation:
             ref_seq = self._next_seq
             self._next_seq += 1
 
-        ref_msg = Message(seq=ref_seq, role="user", content=ref_content)
+        ref_msg = Message(seq=ref_seq, role="user", content=ref_content, run_id=self._run_id)
 
         # Persist: delete old messages from store, write reference + kept structural.
         # In aggressive mode, collapsed messages may be interspersed with kept
@@ -941,7 +1063,7 @@ class NodeConversation:
             # Write kept structural messages (they may have been modified)
             for msg in kept_structural:
                 await self._store.write_part(msg.seq, msg.to_storage_dict())
-            await self._store.write_cursor({"next_seq": self._next_seq})
+            await self._write_next_seq()
 
         # Reassemble: reference + kept structural (in original order) + recent
         self._messages = [ref_msg] + kept_structural + recent_messages
@@ -978,7 +1100,7 @@ class NodeConversation:
         """Remove all messages, keep system prompt, preserve ``_next_seq``."""
         if self._store:
             await self._store.delete_parts_before(self._next_seq)
-            await self._store.write_cursor({"next_seq": self._next_seq})
+            await self._write_next_seq()
         self._messages.clear()
         self._last_api_input_tokens = None
 
@@ -1020,22 +1142,32 @@ class NodeConversation:
         if not self._meta_persisted:
             await self._persist_meta()
         await self._store.write_part(message.seq, message.to_storage_dict())
-        await self._store.write_cursor({"next_seq": self._next_seq})
+        await self._write_next_seq()
 
     async def _persist_meta(self) -> None:
-        """Lazily write conversation metadata to the store (called once)."""
+        """Lazily write conversation metadata to the store (called once).
+
+        When ``self._run_id`` is set, metadata is written flat for backward
+        compatibility (run-scoped isolation has been reverted).
+        """
         if self._store is None:
             return
-        await self._store.write_meta(
-            {
-                "system_prompt": self._system_prompt,
-                "max_history_tokens": self._max_history_tokens,
-                "compaction_threshold": self._compaction_threshold,
-                "output_keys": self._output_keys,
-            }
-        )
+        run_meta = {
+            "system_prompt": self._system_prompt,
+            "max_context_tokens": self._max_context_tokens,
+            "compaction_threshold": self._compaction_threshold,
+            "output_keys": self._output_keys,
+        }
+        await self._store.write_meta(run_meta)
         self._meta_persisted = True
 
+    async def _write_next_seq(self) -> None:
+        if self._store is None:
+            return
+        cursor = await self._store.read_cursor() or {}
+        cursor["next_seq"] = self._next_seq
+        await self._store.write_cursor(cursor)
+
     # --- Restore -----------------------------------------------------------
 
     @classmethod
@@ -1043,6 +1175,7 @@ class NodeConversation:
         cls,
         store: ConversationStore,
         phase_id: str | None = None,
+        run_id: str | None = None,
     ) -> NodeConversation | None:
         """Reconstruct a NodeConversation from a store.
 
@@ -1052,6 +1185,9 @@ class NodeConversation:
                 Used in isolated mode so a node only sees its own
                 messages in the shared flat store.  In continuous mode
                 pass ``None`` to load all parts.
+            run_id: If set, only load parts matching this run_id.
+                Ensures intentional restarts (new run_id) start fresh
+                while crash recovery (same run_id) resumes correctly.
 
         Returns ``None`` if the store contains no metadata (i.e. the
         conversation was never persisted).
@@ -1062,21 +1198,27 @@ class NodeConversation:
 
         conv = cls(
             system_prompt=meta.get("system_prompt", ""),
-            max_history_tokens=meta.get("max_history_tokens", 32000),
+            max_context_tokens=meta.get("max_context_tokens", 32000),
             compaction_threshold=meta.get("compaction_threshold", 0.8),
             output_keys=meta.get("output_keys"),
             store=store,
+            run_id=run_id,
         )
         conv._meta_persisted = True
 
         parts = await store.read_parts()
         if phase_id:
             parts = [p for p in parts if p.get("phase_id") == phase_id]
+        # Filter by run_id so intentional restarts (new run_id) start fresh
+        # while crash recovery (same run_id) loads prior parts.
+        if run_id and not is_legacy_run_id(run_id):
+            parts = [p for p in parts if p.get("run_id") == run_id]
         conv._messages = [Message.from_storage_dict(p) for p in parts]
 
         cursor = await store.read_cursor()
-        if cursor:
-            conv._next_seq = cursor["next_seq"]
+        next_seq = get_cursor_next_seq(cursor)
+        if next_seq is not None:
+            conv._next_seq = next_seq
         elif conv._messages:
             conv._next_seq = conv._messages[-1].seq + 1
 

core/framework/agent_loop/internals/__init__.py

@@ -0,0 +1,7 @@
+"""Agent loop internals -- compaction, judge, tools, subagent execution.
+
+Re-exports from legacy locations for the new import path.
+"""
+
+from framework.agent_loop.internals.compaction import *  # noqa: F401, F403
+from framework.agent_loop.internals.synthetic_tools import *  # noqa: F401, F403

core/framework/agent_loop/internals/compaction.py

@@ -0,0 +1,866 @@
+"""Conversation compaction pipeline.
+
+Implements the multi-level compaction strategy:
+0. Microcompaction (count-based tool result clearing — cheapest)
+1. Prune old tool results (token-budget based)
+2. Structure-preserving compaction (spillover)
+3. LLM summary compaction (with recursive splitting)
+4. Emergency deterministic summary (no LLM)
+"""
+
+from __future__ import annotations
+
+import json
+import logging
+import os
+import re
+import time
+from datetime import UTC, datetime
+from pathlib import Path
+from typing import Any
+
+from framework.agent_loop.conversation import Message, NodeConversation
+from framework.agent_loop.internals.event_publishing import publish_context_usage
+from framework.agent_loop.internals.types import LoopConfig, OutputAccumulator
+from framework.orchestrator.node import NodeContext
+from framework.host.event_bus import EventBus
+
+logger = logging.getLogger(__name__)
+
+# Limits for LLM compaction
+LLM_COMPACT_CHAR_LIMIT: int = 240_000
+LLM_COMPACT_MAX_DEPTH: int = 10
+
+# Microcompaction: tools whose results can be safely cleared
+COMPACTABLE_TOOLS: frozenset[str] = frozenset(
+    {
+        "read_file",
+        "run_command",
+        "web_search",
+        "web_fetch",
+        "grep_search",
+        "glob_search",
+        "write_file",
+        "edit_file",
+        "browser_screenshot",
+        "list_directory",
+    }
+)
+
+# Keep at most this many compactable tool results; clear older ones
+MICROCOMPACT_KEEP_RECENT: int = 8
+
+# Circuit-breaker: stop auto-compacting after this many consecutive failures
+MAX_CONSECUTIVE_FAILURES: int = 3
+
+# Track consecutive compaction failures per conversation (module-level)
+_failure_counts: dict[int, int] = {}
+
+# Track last compaction time per conversation for recompaction detection
+_last_compact_times: dict[int, float] = {}
+
+
+def microcompact(
+    conversation: NodeConversation,
+    *,
+    keep_recent: int = MICROCOMPACT_KEEP_RECENT,
+) -> int:
+    """Clear old compactable tool results by count, keeping only the most recent.
+
+    This is the cheapest possible compaction — no LLM call, no structural
+    changes, just replaces old tool result content with a short placeholder.
+    Inspired by Claude Code's cached-microcompact strategy.
+
+    Returns the number of tool results cleared.
+    """
+    # Collect indices of compactable tool results (newest first)
+    compactable_indices: list[int] = []
+    messages = conversation.messages
+    for i in range(len(messages) - 1, -1, -1):
+        msg = messages[i]
+        if msg.role != "tool" or msg.is_error or msg.is_skill_content:
+            continue
+        if msg.content.startswith(("[Pruned tool result", "[Old tool result")):
+            continue
+        if len(msg.content) < 100:
+            continue
+
+        # Check if the tool that produced this result is compactable
+        tool_name = _find_tool_name_for_result(messages, msg)
+        if tool_name and tool_name in COMPACTABLE_TOOLS:
+            compactable_indices.append(i)
+
+    # Keep the most recent N, clear the rest
+    to_clear = compactable_indices[keep_recent:]
+    if not to_clear:
+        return 0
+
+    cleared = 0
+    for i in to_clear:
+        msg = messages[i]
+        spillover = _extract_spillover_filename_inline(msg.content)
+        orig_len = len(msg.content)
+        if spillover:
+            placeholder = (
+                f"[Old tool result cleared: {orig_len} chars. "
+                f"Full data in '{spillover}'. "
+                f"Use load_data('{spillover}') to retrieve.]"
+            )
+        else:
+            placeholder = f"[Old tool result cleared: {orig_len} chars.]"
+
+        # Mutate in-place (microcompact is synchronous, no store writes)
+        conversation._messages[i] = Message(
+            seq=msg.seq,
+            role=msg.role,
+            content=placeholder,
+            tool_use_id=msg.tool_use_id,
+            tool_calls=msg.tool_calls,
+            is_error=msg.is_error,
+            phase_id=msg.phase_id,
+            is_transition_marker=msg.is_transition_marker,
+        )
+        cleared += 1
+
+    if cleared > 0:
+        # Invalidate cached token count
+        conversation._last_api_input_tokens = None
+
+    return cleared
+
+
+def _find_tool_name_for_result(messages: list[Message], tool_msg: Message) -> str | None:
+    """Find the tool name from the assistant message that triggered this tool result."""
+    if not tool_msg.tool_use_id:
+        return None
+    for msg in messages:
+        if msg.tool_calls:
+            for tc in msg.tool_calls:
+                if tc.get("id") == tool_msg.tool_use_id:
+                    return tc.get("function", {}).get("name")
+    return None
+
+
+def _extract_spillover_filename_inline(content: str) -> str | None:
+    """Quick inline check for spillover filename in tool result content."""
+    match = re.search(r"saved to '([^']+)'", content, re.IGNORECASE)
+    return match.group(1) if match else None
+
+
+async def compact(
+    ctx: NodeContext,
+    conversation: NodeConversation,
+    accumulator: OutputAccumulator | None,
+    *,
+    config: LoopConfig,
+    event_bus: EventBus | None,
+    char_limit: int = LLM_COMPACT_CHAR_LIMIT,
+    max_depth: int = LLM_COMPACT_MAX_DEPTH,
+) -> None:
+    """Run the full compaction pipeline if conversation needs compaction.
+
+    Pipeline stages (in order, short-circuits when budget is restored):
+    0. Microcompaction (count-based tool result clearing — cheapest)
+    1. Prune old tool results (token-budget based)
+    2. Structure-preserving compaction (free, no LLM)
+    3. LLM summary compaction (recursive split if too large)
+    4. Emergency deterministic summary (fallback)
+    """
+    conv_id = id(conversation)
+
+    # Circuit breaker: stop auto-compacting after repeated failures
+    if _failure_counts.get(conv_id, 0) >= MAX_CONSECUTIVE_FAILURES:
+        logger.warning(
+            "Circuit breaker: skipping compaction after %d consecutive failures",
+            _failure_counts[conv_id],
+        )
+        return
+
+    # Recompaction detection
+    now = time.monotonic()
+    last_time = _last_compact_times.get(conv_id)
+    if last_time is not None and (now - last_time) < 30:
+        logger.warning(
+            "Recompaction chain detected: only %.1fs since last compaction",
+            now - last_time,
+        )
+
+    ratio_before = conversation.usage_ratio()
+    phase_grad = getattr(ctx, "continuous_mode", False)
+    pre_inventory: list[dict[str, Any]] | None = None
+
+    if ratio_before >= 1.0:
+        pre_inventory = build_message_inventory(conversation)
+
+    # --- Step 0: Microcompaction (count-based, cheapest) ---
+    mc_cleared = microcompact(conversation)
+    if mc_cleared > 0:
+        logger.info(
+            "Microcompact cleared %d old tool results: %.0f%% -> %.0f%%",
+            mc_cleared,
+            ratio_before * 100,
+            conversation.usage_ratio() * 100,
+        )
+    if not conversation.needs_compaction():
+        _record_success(conv_id, now)
+        await log_compaction(
+            ctx,
+            conversation,
+            ratio_before,
+            event_bus,
+            pre_inventory=pre_inventory,
+        )
+        return
+
+    # --- Step 1: Prune old tool results (free, fast) ---
+    protect = max(2000, config.max_context_tokens // 12)
+    pruned = await conversation.prune_old_tool_results(
+        protect_tokens=protect,
+        min_prune_tokens=max(1000, protect // 3),
+    )
+    if pruned > 0:
+        logger.info(
+            "Pruned %d old tool results: %.0f%% -> %.0f%%",
+            pruned,
+            ratio_before * 100,
+            conversation.usage_ratio() * 100,
+        )
+    if not conversation.needs_compaction():
+        _record_success(conv_id, now)
+        await log_compaction(
+            ctx,
+            conversation,
+            ratio_before,
+            event_bus,
+            pre_inventory=pre_inventory,
+        )
+        return
+
+    # --- Step 2: Standard structure-preserving compaction (free, no LLM) ---
+    spill_dir = config.spillover_dir
+    if spill_dir:
+        await conversation.compact_preserving_structure(
+            spillover_dir=spill_dir,
+            keep_recent=4,
+            phase_graduated=phase_grad,
+        )
+    if not conversation.needs_compaction():
+        _record_success(conv_id, now)
+        await log_compaction(
+            ctx,
+            conversation,
+            ratio_before,
+            event_bus,
+            pre_inventory=pre_inventory,
+        )
+        return
+
+    # --- Step 3: LLM summary compaction ---
+    if ctx.llm is not None:
+        logger.info(
+            "LLM summary compaction triggered (%.0f%% usage)",
+            conversation.usage_ratio() * 100,
+        )
+        try:
+            summary = await llm_compact(
+                ctx,
+                list(conversation.messages),
+                accumulator,
+                char_limit=char_limit,
+                max_depth=max_depth,
+                max_context_tokens=config.max_context_tokens,
+            )
+            await conversation.compact(
+                summary,
+                keep_recent=2,
+                phase_graduated=phase_grad,
+            )
+        except Exception as e:
+            logger.warning("LLM compaction failed: %s", e)
+            _failure_counts[conv_id] = _failure_counts.get(conv_id, 0) + 1
+
+    if not conversation.needs_compaction():
+        _record_success(conv_id, now)
+        await log_compaction(
+            ctx,
+            conversation,
+            ratio_before,
+            event_bus,
+            pre_inventory=pre_inventory,
+        )
+        return
+
+    # --- Step 4: Emergency deterministic summary (LLM failed/unavailable) ---
+    logger.warning(
+        "Emergency compaction (%.0f%% usage)",
+        conversation.usage_ratio() * 100,
+    )
+    summary = build_emergency_summary(ctx, accumulator, conversation, config)
+    await conversation.compact(
+        summary,
+        keep_recent=1,
+        phase_graduated=phase_grad,
+    )
+    _record_success(conv_id, now)
+    await log_compaction(
+        ctx,
+        conversation,
+        ratio_before,
+        event_bus,
+        pre_inventory=pre_inventory,
+    )
+
+
+def _record_success(conv_id: int, timestamp: float) -> None:
+    """Reset failure counter and record compaction time on success."""
+    _failure_counts.pop(conv_id, None)
+    _last_compact_times[conv_id] = timestamp
+
+
+# --- LLM compaction with binary-search splitting ----------------------
+
+
+def strip_images_from_messages(messages: list[Message]) -> list[Message]:
+    """Strip image_content from messages before LLM summarisation.
+
+    Images/documents are replaced with ``[image]`` markers so the summary
+    notes they existed without wasting tokens sending binary data to the
+    compaction LLM.  Returns a new list (original messages are not mutated).
+    """
+    stripped: list[Message] = []
+    for msg in messages:
+        if msg.image_content:
+            n_images = len(msg.image_content)
+            marker = " ".join("[image]" for _ in range(n_images))
+            content = f"{msg.content}\n{marker}" if msg.content else marker
+            stripped.append(
+                Message(
+                    seq=msg.seq,
+                    role=msg.role,
+                    content=content,
+                    tool_use_id=msg.tool_use_id,
+                    tool_calls=msg.tool_calls,
+                    is_error=msg.is_error,
+                    phase_id=msg.phase_id,
+                    is_transition_marker=msg.is_transition_marker,
+                    image_content=None,  # stripped
+                )
+            )
+        else:
+            stripped.append(msg)
+    return stripped
+
+
+async def llm_compact(
+    ctx: NodeContext,
+    messages: list,
+    accumulator: OutputAccumulator | None = None,
+    _depth: int = 0,
+    *,
+    char_limit: int = LLM_COMPACT_CHAR_LIMIT,
+    max_depth: int = LLM_COMPACT_MAX_DEPTH,
+    max_context_tokens: int = 128_000,
+) -> str:
+    """Summarise *messages* with LLM, splitting recursively if too large.
+
+    If the formatted text exceeds ``LLM_COMPACT_CHAR_LIMIT`` or the LLM
+    rejects the call with a context-length error, the messages are split
+    in half and each half is summarised independently.  Tool history is
+    appended once at the top-level call (``_depth == 0``).
+    """
+    from framework.agent_loop.conversation import extract_tool_call_history
+    from framework.agent_loop.internals.tool_result_handler import is_context_too_large_error
+
+    if _depth > max_depth:
+        raise RuntimeError(f"LLM compaction recursion limit ({max_depth})")
+
+    # Strip images before summarisation to avoid wasting tokens
+    if _depth == 0:
+        messages = strip_images_from_messages(messages)
+
+    formatted = format_messages_for_summary(messages)
+
+    # Proactive split: avoid wasting an API call on oversized input
+    if len(formatted) > char_limit and len(messages) > 1:
+        summary = await _llm_compact_split(
+            ctx,
+            messages,
+            accumulator,
+            _depth,
+            char_limit=char_limit,
+            max_depth=max_depth,
+            max_context_tokens=max_context_tokens,
+        )
+    else:
+        prompt = build_llm_compaction_prompt(
+            ctx,
+            accumulator,
+            formatted,
+            max_context_tokens=max_context_tokens,
+        )
+        summary_budget = max(1024, max_context_tokens // 2)
+        try:
+            response = await ctx.llm.acomplete(
+                messages=[{"role": "user", "content": prompt}],
+                system=(
+                    "You are a conversation compactor for an AI agent. "
+                    "Write a detailed summary that allows the agent to "
+                    "continue its work. Preserve user-stated rules, "
+                    "constraints, and account/identity preferences verbatim."
+                ),
+                max_tokens=summary_budget,
+            )
+            summary = response.content
+        except Exception as e:
+            if is_context_too_large_error(e) and len(messages) > 1:
+                logger.info(
+                    "LLM context too large (depth=%d, msgs=%d) — splitting",
+                    _depth,
+                    len(messages),
+                )
+                summary = await _llm_compact_split(
+                    ctx,
+                    messages,
+                    accumulator,
+                    _depth,
+                    char_limit=char_limit,
+                    max_depth=max_depth,
+                    max_context_tokens=max_context_tokens,
+                )
+            else:
+                raise
+
+    # Append tool history at top level only
+    if _depth == 0:
+        tool_history = extract_tool_call_history(messages)
+        if tool_history and "TOOLS ALREADY CALLED" not in summary:
+            summary += "\n\n" + tool_history
+
+    return summary
+
+
+async def _llm_compact_split(
+    ctx: NodeContext,
+    messages: list,
+    accumulator: OutputAccumulator | None,
+    _depth: int,
+    *,
+    char_limit: int = LLM_COMPACT_CHAR_LIMIT,
+    max_depth: int = LLM_COMPACT_MAX_DEPTH,
+    max_context_tokens: int = 128_000,
+) -> str:
+    """Split messages in half and summarise each half independently."""
+    mid = max(1, len(messages) // 2)
+    s1 = await llm_compact(
+        ctx,
+        messages[:mid],
+        None,
+        _depth + 1,
+        char_limit=char_limit,
+        max_depth=max_depth,
+        max_context_tokens=max_context_tokens,
+    )
+    s2 = await llm_compact(
+        ctx,
+        messages[mid:],
+        accumulator,
+        _depth + 1,
+        char_limit=char_limit,
+        max_depth=max_depth,
+        max_context_tokens=max_context_tokens,
+    )
+    return s1 + "\n\n" + s2
+
+
+# --- Compaction helpers ------------------------------------------------
+
+
+def format_messages_for_summary(messages: list) -> str:
+    """Format messages as text for LLM summarisation."""
+    lines: list[str] = []
+    for m in messages:
+        if m.role == "tool":
+            content = m.content[:500]
+            if len(m.content) > 500:
+                content += "..."
+            lines.append(f"[tool result]: {content}")
+        elif m.role == "assistant" and m.tool_calls:
+            names = [tc.get("function", {}).get("name", "?") for tc in m.tool_calls]
+            text = m.content[:200] if m.content else ""
+            lines.append(f"[assistant (calls: {', '.join(names)})]: {text}")
+        else:
+            lines.append(f"[{m.role}]: {m.content}")
+    return "\n\n".join(lines)
+
+
+def build_llm_compaction_prompt(
+    ctx: NodeContext,
+    accumulator: OutputAccumulator | None,
+    formatted_messages: str,
+    *,
+    max_context_tokens: int = 128_000,
+) -> str:
+    """Build prompt for LLM compaction targeting 50% of token budget.
+
+    Uses a structured section format inspired by Claude Code's compact
+    service.  Each section focuses on a different aspect of the conversation
+    so the summariser produces consistently useful, well-organised output.
+    """
+    spec = ctx.node_spec
+    ctx_lines = [f"NODE: {spec.name} (id={spec.id})"]
+    if spec.description:
+        ctx_lines.append(f"PURPOSE: {spec.description}")
+    if spec.success_criteria:
+        ctx_lines.append(f"SUCCESS CRITERIA: {spec.success_criteria}")
+
+    if accumulator:
+        acc = accumulator.to_dict()
+        done = {k: v for k, v in acc.items() if v is not None}
+        todo = [k for k, v in acc.items() if v is None]
+        if done:
+            ctx_lines.append(
+                "OUTPUTS ALREADY SET:\n"
+                + "\n".join(f"  {k}: {str(v)[:150]}" for k, v in done.items())
+            )
+        if todo:
+            ctx_lines.append(f"OUTPUTS STILL NEEDED: {', '.join(todo)}")
+    elif spec.output_keys:
+        ctx_lines.append(f"OUTPUTS STILL NEEDED: {', '.join(spec.output_keys)}")
+
+    target_tokens = max_context_tokens // 2
+    target_chars = target_tokens * 4
+    node_ctx = "\n".join(ctx_lines)
+
+    return (
+        "You are compacting an AI agent's conversation history. "
+        "The agent is still working and needs to continue.\n\n"
+        f"AGENT CONTEXT:\n{node_ctx}\n\n"
+        f"CONVERSATION MESSAGES:\n{formatted_messages}\n\n"
+        "INSTRUCTIONS:\n"
+        f"Write a summary of approximately {target_chars} characters "
+        f"(~{target_tokens} tokens).\n\n"
+        "Organise the summary into these sections (omit empty ones):\n\n"
+        "1. **Primary Request and Intent** — What the user originally asked "
+        "for and the high-level goal the agent is working toward.\n"
+        "2. **Key Technical Concepts** — Important domain-specific terms, "
+        "patterns, or architectural decisions established in the conversation.\n"
+        "3. **Files and Code Sections** — Specific files read/written/edited "
+        "with brief descriptions of changes. Include short code snippets only "
+        "when they capture critical logic.\n"
+        "4. **Errors and Fixes** — Problems encountered and how they were "
+        "resolved. Include root causes so the agent doesn't repeat them.\n"
+        "5. **Problem Solving Efforts** — Approaches tried, dead ends hit, "
+        "and reasoning behind the current strategy.\n"
+        "6. **User Messages** — Preserve ALL user-stated rules, constraints, "
+        "identity preferences, and account details verbatim.\n"
+        "7. **Pending Tasks** — Work remaining, outputs still needed, and "
+        "any blockers.\n"
+        "8. **Current Work** — The most recent action taken and the immediate "
+        "next step the agent should perform. This section is the most important "
+        "for seamless resumption.\n\n"
+        "Additional rules:\n"
+        "- Be detailed enough that the agent can resume without re-doing work.\n"
+        "- Preserve key decisions made and results obtained.\n"
+        "- When in doubt, keep information rather than discard it.\n"
+    )
+
+
+def build_message_inventory(conversation: NodeConversation) -> list[dict[str, Any]]:
+    """Build a per-message size inventory for debug logging."""
+    inventory: list[dict[str, Any]] = []
+    for message in conversation.messages:
+        content_chars = len(message.content)
+        tool_call_args_chars = 0
+        tool_name = None
+        if message.tool_calls:
+            for tool_call in message.tool_calls:
+                args = tool_call.get("function", {}).get("arguments", "")
+                tool_call_args_chars += (
+                    len(args) if isinstance(args, str) else len(json.dumps(args))
+                )
+            names = [
+                tool_call.get("function", {}).get("name", "?") for tool_call in message.tool_calls
+            ]
+            tool_name = ", ".join(names)
+        elif message.role == "tool" and message.tool_use_id:
+            for previous in conversation.messages:
+                if previous.tool_calls:
+                    for tool_call in previous.tool_calls:
+                        if tool_call.get("id") == message.tool_use_id:
+                            tool_name = tool_call.get("function", {}).get("name", "?")
+                            break
+                if tool_name:
+                    break
+        entry: dict[str, Any] = {
+            "seq": message.seq,
+            "role": message.role,
+            "content_chars": content_chars,
+        }
+        if tool_call_args_chars:
+            entry["tool_call_args_chars"] = tool_call_args_chars
+        if tool_name:
+            entry["tool"] = tool_name
+        if message.is_error:
+            entry["is_error"] = True
+        if message.phase_id:
+            entry["phase"] = message.phase_id
+        if content_chars > 2000:
+            entry["preview"] = message.content[:200] + "…"
+        inventory.append(entry)
+    return inventory
+
+
+def write_compaction_debug_log(
+    ctx: NodeContext,
+    before_pct: int,
+    after_pct: int,
+    level: str,
+    inventory: list[dict[str, Any]] | None,
+) -> None:
+    """Write detailed compaction analysis to ~/.hive/compaction_log/."""
+    log_dir = Path.home() / ".hive" / "compaction_log"
+    log_dir.mkdir(parents=True, exist_ok=True)
+
+    ts = datetime.now(UTC).strftime("%Y%m%dT%H%M%S_%f")
+    node_label = ctx.node_id.replace("/", "_")
+    log_path = log_dir / f"{ts}_{node_label}.md"
+
+    lines: list[str] = [
+        f"# Compaction Debug — {ctx.node_id}",
+        f"**Time:** {datetime.now(UTC).isoformat()}",
+        f"**Node:** {ctx.node_spec.name} (`{ctx.node_id}`)",
+    ]
+    if ctx.stream_id:
+        lines.append(f"**Stream:** {ctx.stream_id}")
+    lines.append(f"**Level:** {level}")
+    lines.append(f"**Usage:** {before_pct}% → {after_pct}%")
+    lines.append("")
+
+    if inventory:
+        total_chars = sum(
+            entry.get("content_chars", 0) + entry.get("tool_call_args_chars", 0)
+            for entry in inventory
+        )
+        lines.append(
+            "## Pre-Compaction Message Inventory "
+            f"({len(inventory)} messages, {total_chars:,} total chars)"
+        )
+        lines.append("")
+        ranked = sorted(
+            inventory,
+            key=lambda entry: entry.get("content_chars", 0) + entry.get("tool_call_args_chars", 0),
+            reverse=True,
+        )
+        lines.append("| # | seq | role | tool | chars | % of total | flags |")
+        lines.append("|---|-----|------|------|------:|------------|-------|")
+        for i, entry in enumerate(ranked, 1):
+            chars = entry.get("content_chars", 0) + entry.get("tool_call_args_chars", 0)
+            pct = (chars / total_chars * 100) if total_chars else 0
+            tool = entry.get("tool", "")
+            flags: list[str] = []
+            if entry.get("is_error"):
+                flags.append("error")
+            if entry.get("phase"):
+                flags.append(f"phase={entry['phase']}")
+            lines.append(
+                f"| {i} | {entry['seq']} | {entry['role']} | {tool} "
+                f"| {chars:,} | {pct:.1f}% | {', '.join(flags)} |"
+            )
+
+        large = [entry for entry in ranked if entry.get("preview")]
+        if large:
+            lines.append("")
+            lines.append("### Large message previews")
+            for entry in large:
+                lines.append(
+                    f"\n**seq={entry['seq']}** ({entry['role']}, {entry.get('tool', '')}):"
+                )
+                lines.append(f"```\n{entry['preview']}\n```")
+    lines.append("")
+
+    try:
+        log_path.write_text("\n".join(lines), encoding="utf-8")
+        logger.debug("Compaction debug log written to %s", log_path)
+    except OSError:
+        logger.debug("Failed to write compaction debug log to %s", log_path)
+
+
+async def log_compaction(
+    ctx: NodeContext,
+    conversation: NodeConversation,
+    ratio_before: float,
+    event_bus: EventBus | None,
+    *,
+    pre_inventory: list[dict[str, Any]] | None = None,
+) -> None:
+    """Log compaction result to runtime logger and event bus."""
+    ratio_after = conversation.usage_ratio()
+    before_pct = round(ratio_before * 100)
+    after_pct = round(ratio_after * 100)
+
+    # Determine label from what happened
+    if after_pct >= before_pct - 1:
+        level = "prune_only"
+    elif ratio_after <= 0.6:
+        level = "llm"
+    else:
+        level = "structural"
+
+    logger.info(
+        "Compaction complete (%s): %d%% -> %d%%",
+        level,
+        before_pct,
+        after_pct,
+    )
+
+    if ctx.runtime_logger:
+        ctx.runtime_logger.log_step(
+            node_id=ctx.node_id,
+            node_type="event_loop",
+            step_index=-1,
+            llm_text=f"Context compacted ({level}): {before_pct}% \u2192 {after_pct}%",
+            verdict="COMPACTION",
+            verdict_feedback=f"level={level} before={before_pct}% after={after_pct}%",
+        )
+
+    if event_bus:
+        from framework.host.event_bus import AgentEvent, EventType
+
+        event_data: dict[str, Any] = {
+            "level": level,
+            "usage_before": before_pct,
+            "usage_after": after_pct,
+        }
+        if pre_inventory is not None:
+            event_data["message_inventory"] = pre_inventory
+        await event_bus.publish(
+            AgentEvent(
+                type=EventType.CONTEXT_COMPACTED,
+                stream_id=ctx.stream_id or ctx.node_id,
+                node_id=ctx.node_id,
+                data=event_data,
+            )
+        )
+
+    await publish_context_usage(event_bus, ctx, conversation, "post_compaction")
+
+    if os.environ.get("HIVE_COMPACTION_DEBUG"):
+        write_compaction_debug_log(ctx, before_pct, after_pct, level, pre_inventory)
+
+
+def build_emergency_summary(
+    ctx: NodeContext,
+    accumulator: OutputAccumulator | None = None,
+    conversation: NodeConversation | None = None,
+    config: LoopConfig | None = None,
+) -> str:
+    """Build a structured emergency compaction summary.
+
+    Unlike normal/aggressive compaction which uses an LLM summary,
+    emergency compaction cannot afford an LLM call (context is already
+    way over budget).  Instead, build a deterministic summary from the
+    node's known state so the LLM can continue working after
+    compaction without losing track of its task and inputs.
+    """
+    parts = [
+        "EMERGENCY COMPACTION — previous conversation was too large "
+        "and has been replaced with this summary.\n"
+    ]
+
+    # 1. Node identity
+    spec = ctx.node_spec
+    parts.append(f"NODE: {spec.name} (id={spec.id})")
+    if spec.description:
+        parts.append(f"PURPOSE: {spec.description}")
+
+    # 2. Inputs the node received
+    input_lines = []
+    for key in spec.input_keys:
+        value = ctx.input_data.get(key) or ctx.buffer.read(key)
+        if value is not None:
+            # Truncate long values but keep them recognisable
+            v_str = str(value)
+            if len(v_str) > 200:
+                v_str = v_str[:200] + "…"
+            input_lines.append(f"  {key}: {v_str}")
+    if input_lines:
+        parts.append("INPUTS:\n" + "\n".join(input_lines))
+
+    # 3. Output accumulator state (what's been set so far)
+    if accumulator:
+        acc_state = accumulator.to_dict()
+        set_keys = {k: v for k, v in acc_state.items() if v is not None}
+        missing = [k for k, v in acc_state.items() if v is None]
+        if set_keys:
+            lines = [f"  {k}: {str(v)[:150]}" for k, v in set_keys.items()]
+            parts.append("OUTPUTS ALREADY SET:\n" + "\n".join(lines))
+        if missing:
+            parts.append(f"OUTPUTS STILL NEEDED: {', '.join(missing)}")
+    elif spec.output_keys:
+        parts.append(f"OUTPUTS STILL NEEDED: {', '.join(spec.output_keys)}")
+
+    # 4. Available tools reminder
+    if spec.tools:
+        parts.append(f"AVAILABLE TOOLS: {', '.join(spec.tools)}")
+
+    # 5. Spillover files — list actual files so the LLM can load
+    # them immediately instead of having to call list_data_files first.
+    spillover_dir = config.spillover_dir if config else None
+    if spillover_dir:
+        try:
+            from pathlib import Path
+
+            data_dir = Path(spillover_dir)
+            if data_dir.is_dir():
+                all_files = sorted(f.name for f in data_dir.iterdir() if f.is_file())
+                # Separate conversation history files from regular data files
+                conv_files = [f for f in all_files if re.match(r"conversation_\d+\.md$", f)]
+                data_files = [f for f in all_files if f not in conv_files]
+
+                if conv_files:
+                    conv_list = "\n".join(
+                        f"  - {f}  (full path: {data_dir / f})" for f in conv_files
+                    )
+                    parts.append(
+                        "CONVERSATION HISTORY (freeform messages saved during compaction — "
+                        "use load_data('<filename>') to review earlier dialogue):\n" + conv_list
+                    )
+                if data_files:
+                    file_list = "\n".join(
+                        f"  - {f}  (full path: {data_dir / f})" for f in data_files[:30]
+                    )
+                    parts.append("DATA FILES (use load_data('<filename>') to read):\n" + file_list)
+                if not all_files:
+                    parts.append(
+                        "NOTE: Large tool results may have been saved to files. "
+                        "Use list_directory to check the data directory."
+                    )
+        except Exception:
+            parts.append(
+                "NOTE: Large tool results were saved to files. "
+                "Use read_file(path='<path>') to read them."
+            )
+
+    # 6. Tool call history (prevent re-calling tools)
+    if conversation is not None:
+        tool_history = _extract_tool_call_history(conversation)
+        if tool_history:
+            parts.append(tool_history)
+
+    parts.append(
+        "\nContinue working towards setting the remaining outputs. "
+        "Use your tools and the inputs above."
+    )
+    return "\n\n".join(parts)
+
+
+def _extract_tool_call_history(conversation: NodeConversation) -> str:
+    """Extract tool call history from conversation messages.
+
+    This is the instance-level variant that operates on a NodeConversation
+    directly (vs. the module-level extract_tool_call_history in conversation.py
+    which works on raw message lists).
+    """
+    from framework.agent_loop.conversation import extract_tool_call_history
+
+    return extract_tool_call_history(list(conversation.messages))

core/framework/agent_loop/internals/cursor_persistence.py

@@ -0,0 +1,258 @@
+"""Cursor persistence, queue draining, and pause detection.
+
+Handles the checkpoint/resume cycle: restoring state from a previous
+conversation store, writing cursor data, and managing injection/trigger
+queues between iterations.
+"""
+
+from __future__ import annotations
+
+import asyncio
+import json
+import logging
+from collections.abc import Awaitable, Callable
+from dataclasses import dataclass
+from typing import Any
+
+from framework.agent_loop.conversation import ConversationStore, NodeConversation
+from framework.agent_loop.internals.types import LoopConfig, OutputAccumulator, TriggerEvent
+from framework.orchestrator.node import NodeContext
+from framework.llm.capabilities import supports_image_tool_results
+
+logger = logging.getLogger(__name__)
+
+
+@dataclass
+class RestoredState:
+    """State recovered from a previous checkpoint."""
+
+    conversation: NodeConversation
+    accumulator: OutputAccumulator
+    start_iteration: int
+    recent_responses: list[str]
+    recent_tool_fingerprints: list[list[tuple[str, str]]]
+    pending_input: dict[str, Any] | None
+
+
+async def restore(
+    conversation_store: ConversationStore | None,
+    ctx: NodeContext,
+    config: LoopConfig,
+) -> RestoredState | None:
+    """Attempt to restore from a previous checkpoint.
+
+    Returns a ``RestoredState`` with conversation, accumulator, iteration
+    counter, and stall/doom-loop detection state — everything needed to
+    resume exactly where execution stopped.
+    """
+    if conversation_store is None:
+        return None
+
+    # In isolated mode, filter parts by phase_id so the node only sees
+    # its own messages in the shared flat conversation store.  In
+    # continuous mode (or when _restore is called for timer-resume)
+    # load all parts — the full conversation threads across nodes.
+    _is_continuous = getattr(ctx, "continuous_mode", False)
+    phase_filter = None if _is_continuous else ctx.node_id
+    conversation = await NodeConversation.restore(
+        conversation_store,
+        phase_id=phase_filter,
+        run_id=ctx.effective_run_id,
+    )
+    if conversation is None:
+        return None
+
+    # If run_id filtering removed all messages, this is an intentional
+    # restart (new run), not a crash recovery.  Return None so the caller
+    # falls through to the fresh-conversation path.
+    if conversation.message_count == 0:
+        return None
+
+    accumulator = await OutputAccumulator.restore(conversation_store, run_id=ctx.effective_run_id)
+    accumulator.spillover_dir = config.spillover_dir
+    accumulator.max_value_chars = config.max_output_value_chars
+
+    cursor = await conversation_store.read_cursor() or {}
+    start_iteration = cursor.get("iteration", 0) + 1
+
+    # Restore stall/doom-loop detection state
+    recent_responses: list[str] = cursor.get("recent_responses", [])
+    raw_fps = cursor.get("recent_tool_fingerprints", [])
+    recent_tool_fingerprints: list[list[tuple[str, str]]] = [
+        [tuple(pair) for pair in fps]  # type: ignore[misc]
+        for fps in raw_fps
+    ]
+    pending_input = cursor.get("pending_input")
+    if not isinstance(pending_input, dict):
+        pending_input = None
+
+    logger.info(
+        f"Restored event loop: iteration={start_iteration}, "
+        f"messages={conversation.message_count}, "
+        f"outputs={list(accumulator.values.keys())}, "
+        f"stall_window={len(recent_responses)}, "
+        f"doom_window={len(recent_tool_fingerprints)}"
+    )
+    return RestoredState(
+        conversation=conversation,
+        accumulator=accumulator,
+        start_iteration=start_iteration,
+        recent_responses=recent_responses,
+        recent_tool_fingerprints=recent_tool_fingerprints,
+        pending_input=pending_input,
+    )
+
+
+async def write_cursor(
+    conversation_store: ConversationStore | None,
+    ctx: NodeContext,
+    conversation: NodeConversation,
+    accumulator: OutputAccumulator,
+    iteration: int,
+    *,
+    recent_responses: list[str] | None = None,
+    recent_tool_fingerprints: list[list[tuple[str, str]]] | None = None,
+    pending_input: dict[str, Any] | None = None,
+) -> None:
+    """Write checkpoint cursor for crash recovery.
+
+    Persists iteration counter, accumulator outputs, and stall/doom-loop
+    detection state so that resume picks up exactly where execution stopped.
+    """
+    if conversation_store:
+        cursor = await conversation_store.read_cursor() or {}
+        cursor.update(
+            {
+                "iteration": iteration,
+                "node_id": ctx.node_id,
+                "outputs": accumulator.to_dict(),
+            }
+        )
+        # Persist stall/doom-loop detection state for reliable resume
+        if recent_responses is not None:
+            cursor["recent_responses"] = recent_responses
+        if recent_tool_fingerprints is not None:
+            # Convert list[list[tuple]] → list[list[list]] for JSON
+            cursor["recent_tool_fingerprints"] = [
+                [list(pair) for pair in fps] for fps in recent_tool_fingerprints
+            ]
+        # Persist blocked-input state so restored runs re-block instead of
+        # manufacturing a synthetic continuation turn.
+        cursor["pending_input"] = pending_input
+        await conversation_store.write_cursor(cursor)
+
+
+async def drain_injection_queue(
+    queue: asyncio.Queue,
+    conversation: NodeConversation,
+    *,
+    ctx: NodeContext,
+    describe_images_as_text_fn: (
+        Callable[[list[dict[str, Any]]], Awaitable[str | None]] | None
+    ) = None,
+) -> int:
+    """Drain all pending injected events as user messages. Returns count."""
+    count = 0
+    logger.debug(
+        "[drain_injection_queue] Starting to drain queue, initial queue size: %s",
+        queue.qsize() if hasattr(queue, "qsize") else "unknown",
+    )
+    while not queue.empty():
+        try:
+            content, is_client_input, image_content = queue.get_nowait()
+            logger.info(
+                "[drain] injected message (client_input=%s, images=%d): %s",
+                is_client_input,
+                len(image_content) if image_content else 0,
+                content[:200] if content else "(empty)",
+            )
+            if image_content and ctx.llm and not supports_image_tool_results(ctx.llm.model):
+                logger.info(
+                    "Model '%s' does not support images; attempting vision fallback",
+                    ctx.llm.model,
+                )
+                if describe_images_as_text_fn is not None:
+                    description = await describe_images_as_text_fn(image_content)
+                    if description:
+                        content = f"{content}\n\n{description}" if content else description
+                        logger.info("[drain] image described as text via vision fallback")
+                    else:
+                        logger.info("[drain] no vision fallback available; images dropped")
+                image_content = None
+            # Real user input is stored as-is; external events get a prefix
+            if is_client_input:
+                await conversation.add_user_message(
+                    content,
+                    is_client_input=True,
+                    image_content=image_content,
+                )
+            else:
+                await conversation.add_user_message(f"[External event]: {content}")
+            count += 1
+        except asyncio.QueueEmpty:
+            break
+    return count
+
+
+async def drain_trigger_queue(
+    queue: asyncio.Queue,
+    conversation: NodeConversation,
+) -> int:
+    """Drain all pending trigger events as a single batched user message.
+
+    Multiple triggers are merged so the LLM sees them atomically and can
+    reason about all pending triggers before acting.
+    """
+    triggers: list[TriggerEvent] = []
+    while not queue.empty():
+        try:
+            triggers.append(queue.get_nowait())
+        except asyncio.QueueEmpty:
+            break
+
+    if not triggers:
+        return 0
+
+    parts: list[str] = []
+    for t in triggers:
+        task = t.payload.get("task", "")
+        task_line = f"\nTask: {task}" if task else ""
+        payload_str = json.dumps(t.payload, default=str)
+        parts.append(f"[TRIGGER: {t.trigger_type}/{t.source_id}]{task_line}\n{payload_str}")
+
+    combined = "\n\n".join(parts)
+    logger.info("[drain] %d trigger(s): %s", len(triggers), combined[:200])
+    await conversation.add_user_message(combined)
+    return len(triggers)
+
+
+async def check_pause(
+    ctx: NodeContext,
+    conversation: NodeConversation,
+    iteration: int,
+) -> bool:
+    """
+    Check if pause has been requested. Returns True if paused.
+
+    Note: This check happens BEFORE starting iteration N, after completing N-1.
+    If paused, the node exits having completed {iteration} iterations (0 to iteration-1).
+    """
+    # Check executor-level pause event (for /pause command, Ctrl+Z)
+    if ctx.pause_event and ctx.pause_event.is_set():
+        completed = iteration  # 0-indexed: iteration=3 means 3 iterations completed (0,1,2)
+        logger.info(f"⏸ Pausing after {completed} iteration(s) completed (executor-level)")
+        return True
+
+    # Check context-level pause flags (legacy/alternative methods)
+    pause_requested = ctx.input_data.get("pause_requested", False)
+    if not pause_requested:
+        try:
+            pause_requested = ctx.buffer.read("pause_requested") or False
+        except (PermissionError, KeyError):
+            pause_requested = False
+    if pause_requested:
+        completed = iteration
+        logger.info(f"⏸ Pausing after {completed} iteration(s) completed (context-level)")
+        return True
+
+    return False

core/framework/agent_loop/internals/event_publishing.py

@@ -0,0 +1,360 @@
+"""EventBus publishing helpers for the event loop.
+
+Thin wrappers around EventBus.emit_*() calls that check for bus existence
+before publishing.  Extracted to reduce noise in the main orchestrator.
+"""
+
+from __future__ import annotations
+
+import logging
+import time
+
+from framework.agent_loop.conversation import NodeConversation
+from framework.agent_loop.internals.types import HookContext
+from framework.orchestrator.node import NodeContext
+from framework.host.event_bus import EventBus
+
+logger = logging.getLogger(__name__)
+
+
+async def publish_loop_started(
+    event_bus: EventBus | None,
+    stream_id: str,
+    node_id: str,
+    max_iterations: int,
+    execution_id: str = "",
+) -> None:
+    if event_bus:
+        await event_bus.emit_node_loop_started(
+            stream_id=stream_id,
+            node_id=node_id,
+            max_iterations=max_iterations,
+            execution_id=execution_id,
+        )
+
+
+async def generate_action_plan(
+    event_bus: EventBus | None,
+    ctx: NodeContext,
+    stream_id: str,
+    node_id: str,
+    execution_id: str,
+) -> None:
+    """Generate a brief action plan via LLM and emit it as an SSE event.
+
+    Runs as a fire-and-forget task so it never blocks the main loop.
+    """
+    try:
+        system_prompt = ctx.node_spec.system_prompt or ""
+        # Trim to keep the prompt small
+        prompt_summary = system_prompt[:500]
+        if len(system_prompt) > 500:
+            prompt_summary += "..."
+
+        tool_names = [t.name for t in ctx.available_tools]
+        output_keys = ctx.node_spec.output_keys or []
+
+        prompt = (
+            f'You are about to work on a task as node "{node_id}".\n\n'
+            f"System prompt:\n{prompt_summary}\n\n"
+            f"Tools available: {tool_names}\n"
+            f"Required outputs: {output_keys}\n\n"
+            f"Write a brief action plan (2-5 bullet points) describing "
+            f"what you will do to complete this task. Be specific and concise.\n"
+            f"Return ONLY the plan text, no preamble."
+        )
+
+        response = await ctx.llm.acomplete(
+            messages=[{"role": "user", "content": prompt}],
+            max_tokens=1024,
+        )
+
+        plan = response.content.strip()
+        if plan and event_bus:
+            await event_bus.emit_node_action_plan(
+                stream_id=stream_id,
+                node_id=node_id,
+                plan=plan,
+                execution_id=execution_id,
+            )
+    except Exception as e:
+        logger.warning("Action plan generation failed for node '%s': %s", node_id, e)
+
+
+async def publish_iteration(
+    event_bus: EventBus | None,
+    stream_id: str,
+    node_id: str,
+    iteration: int,
+    execution_id: str = "",
+    extra_data: dict | None = None,
+) -> None:
+    if event_bus:
+        await event_bus.emit_node_loop_iteration(
+            stream_id=stream_id,
+            node_id=node_id,
+            iteration=iteration,
+            execution_id=execution_id,
+            extra_data=extra_data,
+        )
+
+
+async def publish_llm_turn_complete(
+    event_bus: EventBus | None,
+    stream_id: str,
+    node_id: str,
+    stop_reason: str,
+    model: str,
+    input_tokens: int,
+    output_tokens: int,
+    cached_tokens: int = 0,
+    execution_id: str = "",
+    iteration: int | None = None,
+) -> None:
+    if event_bus:
+        await event_bus.emit_llm_turn_complete(
+            stream_id=stream_id,
+            node_id=node_id,
+            stop_reason=stop_reason,
+            model=model,
+            input_tokens=input_tokens,
+            output_tokens=output_tokens,
+            cached_tokens=cached_tokens,
+            execution_id=execution_id,
+            iteration=iteration,
+        )
+
+
+def log_skip_judge(
+    ctx: NodeContext,
+    node_id: str,
+    iteration: int,
+    feedback: str,
+    tool_calls: list[dict],
+    llm_text: str,
+    turn_tokens: dict[str, int],
+    iter_start: float,
+) -> None:
+    """Log a CONTINUE step that skips judge evaluation (e.g., waiting for input)."""
+    if ctx.runtime_logger:
+        ctx.runtime_logger.log_step(
+            node_id=node_id,
+            node_type="event_loop",
+            step_index=iteration,
+            verdict="CONTINUE",
+            verdict_feedback=feedback,
+            tool_calls=tool_calls,
+            llm_text=llm_text,
+            input_tokens=turn_tokens.get("input", 0),
+            output_tokens=turn_tokens.get("output", 0),
+            latency_ms=int((time.time() - iter_start) * 1000),
+        )
+
+
+async def publish_loop_completed(
+    event_bus: EventBus | None,
+    stream_id: str,
+    node_id: str,
+    iterations: int,
+    execution_id: str = "",
+) -> None:
+    if event_bus:
+        await event_bus.emit_node_loop_completed(
+            stream_id=stream_id,
+            node_id=node_id,
+            iterations=iterations,
+            execution_id=execution_id,
+        )
+
+
+async def publish_context_usage(
+    event_bus: EventBus | None,
+    ctx: NodeContext,
+    conversation: NodeConversation,
+    trigger: str,
+) -> None:
+    """Emit a CONTEXT_USAGE_UPDATED event with current context window state."""
+    if not event_bus:
+        return
+
+    from framework.host.event_bus import AgentEvent, EventType
+
+    estimated = conversation.estimate_tokens()
+    max_tokens = conversation._max_context_tokens
+    ratio = estimated / max_tokens if max_tokens > 0 else 0.0
+    await event_bus.publish(
+        AgentEvent(
+            type=EventType.CONTEXT_USAGE_UPDATED,
+            stream_id=ctx.stream_id or ctx.node_id,
+            node_id=ctx.node_id,
+            data={
+                "usage_ratio": round(ratio, 4),
+                "usage_pct": round(ratio * 100),
+                "message_count": conversation.message_count,
+                "estimated_tokens": estimated,
+                "max_context_tokens": max_tokens,
+                "trigger": trigger,
+            },
+        )
+    )
+
+
+async def publish_stalled(
+    event_bus: EventBus | None,
+    stream_id: str,
+    node_id: str,
+    execution_id: str = "",
+) -> None:
+    if event_bus:
+        await event_bus.emit_node_stalled(
+            stream_id=stream_id,
+            node_id=node_id,
+            reason="Consecutive similar responses detected",
+            execution_id=execution_id,
+        )
+
+
+async def publish_text_delta(
+    event_bus: EventBus | None,
+    stream_id: str,
+    node_id: str,
+    content: str,
+    snapshot: str,
+    ctx: NodeContext,
+    execution_id: str = "",
+    iteration: int | None = None,
+    inner_turn: int = 0,
+) -> None:
+    if event_bus:
+        if ctx.emits_client_io:
+            await event_bus.emit_client_output_delta(
+                stream_id=stream_id,
+                node_id=node_id,
+                content=content,
+                snapshot=snapshot,
+                execution_id=execution_id,
+                iteration=iteration,
+                inner_turn=inner_turn,
+            )
+        else:
+            await event_bus.emit_llm_text_delta(
+                stream_id=stream_id,
+                node_id=node_id,
+                content=content,
+                snapshot=snapshot,
+                execution_id=execution_id,
+                inner_turn=inner_turn,
+            )
+
+
+async def publish_tool_started(
+    event_bus: EventBus | None,
+    stream_id: str,
+    node_id: str,
+    tool_use_id: str,
+    tool_name: str,
+    tool_input: dict,
+    execution_id: str = "",
+) -> None:
+    if event_bus:
+        await event_bus.emit_tool_call_started(
+            stream_id=stream_id,
+            node_id=node_id,
+            tool_use_id=tool_use_id,
+            tool_name=tool_name,
+            tool_input=tool_input,
+            execution_id=execution_id,
+        )
+
+
+async def publish_tool_completed(
+    event_bus: EventBus | None,
+    stream_id: str,
+    node_id: str,
+    tool_use_id: str,
+    tool_name: str,
+    result: str,
+    is_error: bool,
+    execution_id: str = "",
+) -> None:
+    if event_bus:
+        await event_bus.emit_tool_call_completed(
+            stream_id=stream_id,
+            node_id=node_id,
+            tool_use_id=tool_use_id,
+            tool_name=tool_name,
+            result=result,
+            is_error=is_error,
+            execution_id=execution_id,
+        )
+
+
+async def publish_judge_verdict(
+    event_bus: EventBus | None,
+    stream_id: str,
+    node_id: str,
+    action: str,
+    feedback: str = "",
+    judge_type: str = "implicit",
+    iteration: int = 0,
+    execution_id: str = "",
+) -> None:
+    if event_bus:
+        await event_bus.emit_judge_verdict(
+            stream_id=stream_id,
+            node_id=node_id,
+            action=action,
+            feedback=feedback,
+            judge_type=judge_type,
+            iteration=iteration,
+            execution_id=execution_id,
+        )
+
+
+async def publish_output_key_set(
+    event_bus: EventBus | None,
+    stream_id: str,
+    node_id: str,
+    key: str,
+    execution_id: str = "",
+) -> None:
+    if event_bus:
+        await event_bus.emit_output_key_set(
+            stream_id=stream_id, node_id=node_id, key=key, execution_id=execution_id
+        )
+
+
+async def run_hooks(
+    hooks_config: dict[str, list],
+    event: str,
+    conversation: NodeConversation,
+    trigger: str | None = None,
+) -> None:
+    """Run all registered hooks for *event*, applying their results.
+
+    Each hook receives a HookContext and may return a HookResult that:
+    - replaces the system prompt (result.system_prompt)
+    - injects an extra user message (result.inject)
+    Hooks run in registration order; each sees the prompt as left by the
+    previous hook.
+    """
+    hook_list = hooks_config.get(event, [])
+    if not hook_list:
+        return
+    for hook in hook_list:
+        ctx = HookContext(
+            event=event,
+            trigger=trigger,
+            system_prompt=conversation.system_prompt,
+        )
+        try:
+            result = await hook(ctx)
+        except Exception:
+            logger.warning("Hook '%s' raised an exception", event, exc_info=True)
+            continue
+        if result is None:
+            continue
+        if result.system_prompt:
+            conversation.update_system_prompt(result.system_prompt)
+        if result.inject:
+            await conversation.add_user_message(result.inject)

core/framework/agent_loop/internals/judge_pipeline.py

@@ -0,0 +1,175 @@
+"""Judge evaluation pipeline for the event loop."""
+
+from __future__ import annotations
+
+import logging
+from collections.abc import Callable
+
+from framework.agent_loop.conversation import NodeConversation
+from framework.agent_loop.internals.types import JudgeProtocol, JudgeVerdict, OutputAccumulator
+from framework.orchestrator.node import NodeContext
+
+logger = logging.getLogger(__name__)
+
+
+class SubagentJudge:
+    """Judge for subagent execution."""
+
+    def __init__(self, task: str, max_iterations: int = 10):
+        self._task = task
+        self._max_iterations = max_iterations
+
+    async def evaluate(self, context: dict[str, object]) -> JudgeVerdict:
+        missing = context.get("missing_keys", [])
+        if not isinstance(missing, list) or not missing:
+            return JudgeVerdict(action="ACCEPT", feedback="")
+
+        iteration = context.get("iteration", 0)
+        if not isinstance(iteration, int):
+            iteration = 0
+        remaining = self._max_iterations - iteration - 1
+
+        if remaining <= 3:
+            urgency = (
+                f"URGENT: Only {remaining} iterations left. "
+                f"Stop all other work and call set_output NOW for: {missing}"
+            )
+        elif remaining <= self._max_iterations // 2:
+            urgency = (
+                f"WARNING: {remaining} iterations remaining. "
+                f"You must call set_output for: {missing}"
+            )
+        else:
+            urgency = f"Missing output keys: {missing}. Use set_output to provide them."
+
+        return JudgeVerdict(action="RETRY", feedback=f"Your task: {self._task}\n{urgency}")
+
+
+async def judge_turn(
+    *,
+    mark_complete_flag: bool,
+    judge: JudgeProtocol | None,
+    ctx: NodeContext,
+    conversation: NodeConversation,
+    accumulator: OutputAccumulator,
+    assistant_text: str,
+    tool_results: list[dict[str, object]],
+    iteration: int,
+    get_missing_output_keys_fn: Callable[
+        [OutputAccumulator, list[str] | None, list[str] | None],
+        list[str],
+    ],
+    max_context_tokens: int,
+) -> JudgeVerdict:
+    """Evaluate the current state using judge or implicit logic.
+
+    Evaluation levels (in order):
+      0. Short-circuits: mark_complete, skip_judge, tool-continue.
+      1. Custom judge (JudgeProtocol) — full authority when set.
+      2. Implicit judge — output-key check + optional conversation-aware
+         quality gate (when ``success_criteria`` is defined).
+
+    Returns a JudgeVerdict.  ``feedback=None`` means no real evaluation
+    happened (skip_judge, tool-continue); the caller must not inject a
+    feedback message.  Any non-None feedback (including ``""``) means a
+    real evaluation occurred and will be logged into the conversation.
+    """
+    # --- Level 0: short-circuits (no evaluation) -----------------------
+
+    if mark_complete_flag:
+        return JudgeVerdict(action="ACCEPT")
+
+    if ctx.node_spec.skip_judge:
+        return JudgeVerdict(action="RETRY")  # feedback=None → not logged
+
+    # --- Level 1: custom judge -----------------------------------------
+
+    if judge is not None:
+        context = {
+            "assistant_text": assistant_text,
+            "tool_calls": tool_results,
+            "output_accumulator": accumulator.to_dict(),
+            "accumulator": accumulator,
+            "iteration": iteration,
+            "conversation_summary": conversation.export_summary(),
+            "output_keys": ctx.node_spec.output_keys,
+            "missing_keys": get_missing_output_keys_fn(
+                accumulator, ctx.node_spec.output_keys, ctx.node_spec.nullable_output_keys
+            ),
+        }
+        verdict = await judge.evaluate(context)
+        # Ensure evaluated RETRY always carries feedback for logging.
+        if verdict.action == "RETRY" and not verdict.feedback:
+            return JudgeVerdict(action="RETRY", feedback="Custom judge returned RETRY.")
+        return verdict
+
+    # --- Level 2: implicit judge ---------------------------------------
+
+    # Real tool calls were made — let the agent keep working.
+    if tool_results:
+        return JudgeVerdict(action="RETRY")  # feedback=None → not logged
+
+    missing = get_missing_output_keys_fn(
+        accumulator, ctx.node_spec.output_keys, ctx.node_spec.nullable_output_keys
+    )
+
+    if missing:
+        return JudgeVerdict(
+            action="RETRY",
+            feedback=(
+                f"Task incomplete. Required outputs not yet produced: {missing}. "
+                f"Follow your system prompt instructions to complete the work."
+            ),
+        )
+
+    # All output keys present — run safety checks before accepting.
+
+    output_keys = ctx.node_spec.output_keys or []
+    nullable_keys = set(ctx.node_spec.nullable_output_keys or [])
+
+    # All-nullable with nothing set → node produced nothing useful.
+    all_nullable = output_keys and nullable_keys >= set(output_keys)
+    none_set = not any(accumulator.get(k) is not None for k in output_keys)
+    if all_nullable and none_set:
+        return JudgeVerdict(
+            action="RETRY",
+            feedback=(
+                f"No output keys have been set yet. "
+                f"Use set_output to set at least one of: {output_keys}"
+            ),
+        )
+
+    # Queen with no output keys → continuous interaction node.
+    # Inject tool-use pressure instead of auto-accepting.
+    if not output_keys and ctx.supports_direct_user_io:
+        return JudgeVerdict(
+            action="RETRY",
+            feedback=(
+                "STOP describing what you will do. "
+                "You have FULL access to all tools — file creation, "
+                "shell commands, MCP tools — and you CAN call them "
+                "directly in your response. Respond ONLY with tool "
+                "calls, no prose. Execute the task now."
+            ),
+        )
+
+    # Level 2b: conversation-aware quality check (if success_criteria set)
+    if ctx.node_spec.success_criteria and ctx.llm:
+        from framework.orchestrator.conversation_judge import evaluate_phase_completion
+
+        verdict = await evaluate_phase_completion(
+            llm=ctx.llm,
+            conversation=conversation,
+            phase_name=ctx.node_spec.name,
+            phase_description=ctx.node_spec.description,
+            success_criteria=ctx.node_spec.success_criteria,
+            accumulator_state=accumulator.to_dict(),
+            max_context_tokens=max_context_tokens,
+        )
+        if verdict.action != "ACCEPT":
+            return JudgeVerdict(
+                action=verdict.action,
+                feedback=verdict.feedback or "Phase criteria not met.",
+            )
+
+    return JudgeVerdict(action="ACCEPT", feedback="")

core/framework/agent_loop/internals/stall_detector.py

@@ -0,0 +1,106 @@
+"""Stall and doom-loop detection for the event loop.
+
+Pure functions with no class dependencies — safe to call from any context.
+"""
+
+from __future__ import annotations
+
+import json
+
+
+def ngram_similarity(s1: str, s2: str, n: int = 2) -> float:
+    """Jaccard similarity of n-gram sets.
+
+    Returns 0.0-1.0, where 1.0 is exact match.
+    Fast: O(len(s) + len(s2)) using set operations.
+    """
+
+    def _ngrams(s: str) -> set[str]:
+        return {s[i : i + n] for i in range(len(s) - n + 1) if s.strip()}
+
+    if not s1 or not s2:
+        return 0.0
+
+    ngrams1, ngrams2 = _ngrams(s1.lower()), _ngrams(s2.lower())
+    if not ngrams1 or not ngrams2:
+        return 0.0
+
+    intersection = len(ngrams1 & ngrams2)
+    union = len(ngrams1 | ngrams2)
+    return intersection / union if union else 0.0
+
+
+def is_stalled(
+    recent_responses: list[str],
+    threshold: int,
+    similarity_threshold: float,
+) -> bool:
+    """Detect stall using n-gram similarity.
+
+    Detects when ALL N consecutive responses are mutually similar
+    (>= threshold).  A single dissimilar response resets the signal.
+    This catches phrases like "I'm still stuck" vs "I'm stuck"
+    without false-positives on "attempt 1" vs "attempt 2".
+    """
+    if len(recent_responses) < threshold:
+        return False
+    if not recent_responses[0]:
+        return False
+
+    # Every consecutive pair must be similar
+    for i in range(1, len(recent_responses)):
+        if ngram_similarity(recent_responses[i], recent_responses[i - 1]) < similarity_threshold:
+            return False
+    return True
+
+
+def fingerprint_tool_calls(
+    tool_results: list[dict],
+) -> list[tuple[str, str]]:
+    """Create deterministic fingerprints for a turn's tool calls.
+
+    Each fingerprint is (tool_name, canonical_args_json).  Order-sensitive
+    so [search("a"), fetch("b")] != [fetch("b"), search("a")].
+    """
+    fingerprints = []
+    for tr in tool_results:
+        name = tr.get("tool_name", "")
+        args = tr.get("tool_input", {})
+        try:
+            canonical = json.dumps(args, sort_keys=True, default=str)
+        except (TypeError, ValueError):
+            canonical = str(args)
+        fingerprints.append((name, canonical))
+    return fingerprints
+
+
+def is_tool_doom_loop(
+    recent_tool_fingerprints: list[list[tuple[str, str]]],
+    threshold: int,
+    enabled: bool = True,
+) -> tuple[bool, str]:
+    """Detect doom loop via exact fingerprint match.
+
+    Detects when N consecutive turns invoke the same tools with
+    identical (canonicalized) arguments.  Different arguments mean
+    different work, so only exact matches count.
+
+    Returns (is_doom_loop, description).
+    """
+    if not enabled:
+        return False, ""
+    if len(recent_tool_fingerprints) < threshold:
+        return False, ""
+    first = recent_tool_fingerprints[0]
+    if not first:
+        return False, ""
+
+    # All turns in the window must match the first exactly
+    if all(fp == first for fp in recent_tool_fingerprints[1:]):
+        tool_names = [name for name, _ in first]
+        desc = (
+            f"Doom loop detected: {len(recent_tool_fingerprints)} "
+            f"identical consecutive tool calls ({', '.join(tool_names)})"
+        )
+        return True, desc
+    return False, ""

core/framework/agent_loop/internals/synthetic_tools.py

@@ -0,0 +1,257 @@
+"""Synthetic tool builders for the event loop.
+
+Factory functions that create ``Tool`` definitions for framework-level
+synthetic tools (set_output, ask_user, escalate, delegate, report_to_parent).
+Also includes the ``handle_set_output`` validation logic.
+
+All functions are pure — they receive explicit parameters and return
+``Tool`` or ``ToolResult`` objects with no side effects.
+"""
+
+from __future__ import annotations
+
+from typing import Any
+
+from framework.llm.provider import Tool, ToolResult
+
+
+def build_ask_user_tool() -> Tool:
+    """Build the synthetic ask_user tool for explicit user-input requests.
+
+    The queen calls ask_user() when it needs to pause and wait
+    for user input.  Text-only turns WITHOUT ask_user flow through without
+    blocking, allowing progress updates and summaries to stream freely.
+    """
+    return Tool(
+        name="ask_user",
+        description=(
+            "You MUST call this tool whenever you need the user's response. "
+            "Always call it after greeting the user, asking a question, or "
+            "requesting approval. Do NOT call it for status updates or "
+            "summaries that don't require a response. "
+            "Always include 2-3 predefined options. The UI automatically "
+            "appends an 'Other' free-text input after your options, so NEVER "
+            "include catch-all options like 'Custom idea', 'Something else', "
+            "'Other', or 'None of the above' — the UI handles that. "
+            "When the question primarily needs a typed answer but you must "
+            "include options, make one option signal that typing is expected "
+            "(e.g. 'I\\'ll type my response'). This helps users discover the "
+            "free-text input. "
+            "The ONLY exception: omit options when the question demands a "
+            "free-form answer the user must type out (e.g. 'Describe your "
+            "agent idea', 'Paste the error message'). "
+            '{"question": "What would you like to do?", "options": '
+            '["Build a new agent", "Modify existing agent", "Run tests"]} '
+            "Free-form example: "
+            '{"question": "Describe the agent you want to build."}'
+        ),
+        parameters={
+            "type": "object",
+            "properties": {
+                "question": {
+                    "type": "string",
+                    "description": "The question or prompt shown to the user.",
+                },
+                "options": {
+                    "type": "array",
+                    "items": {"type": "string"},
+                    "description": (
+                        "2-3 specific predefined choices. Include in most cases. "
+                        'Example: ["Option A", "Option B", "Option C"]. '
+                        "The UI always appends an 'Other' free-text input, so "
+                        "do NOT include catch-alls like 'Custom idea' or 'Other'. "
+                        "Omit ONLY when the user must type a free-form answer."
+                    ),
+                    "minItems": 2,
+                    "maxItems": 3,
+                },
+            },
+            "required": ["question"],
+        },
+    )
+
+
+def build_ask_user_multiple_tool() -> Tool:
+    """Build the synthetic ask_user_multiple tool for batched questions.
+
+    Queen-only tool that presents multiple questions at once so the user
+    can answer them all in a single interaction rather than one at a time.
+    """
+    return Tool(
+        name="ask_user_multiple",
+        description=(
+            "Ask the user multiple questions at once. Use this instead of "
+            "ask_user when you have 2 or more questions to ask in the same "
+            "turn — it lets the user answer everything in one go rather than "
+            "going back and forth. Each question can have its own predefined "
+            "options (2-3 choices) or be free-form. The UI renders all "
+            "questions together with a single Submit button. "
+            "ALWAYS prefer this over ask_user when you have multiple things "
+            "to clarify. "
+            "IMPORTANT: Do NOT repeat the questions in your text response — "
+            "the widget renders them. Keep your text to a brief intro only. "
+            '{"questions": ['
+            '  {"id": "scope", "prompt": "What scope?", "options": ["Full", "Partial"]},'
+            '  {"id": "format", "prompt": "Output format?", "options": ["PDF", "CSV", "JSON"]},'
+            '  {"id": "details", "prompt": "Any special requirements?"}'
+            "]}"
+        ),
+        parameters={
+            "type": "object",
+            "properties": {
+                "questions": {
+                    "type": "array",
+                    "items": {
+                        "type": "object",
+                        "properties": {
+                            "id": {
+                                "type": "string",
+                                "description": (
+                                    "Short identifier for this question (used in the response)."
+                                ),
+                            },
+                            "prompt": {
+                                "type": "string",
+                                "description": "The question text shown to the user.",
+                            },
+                            "options": {
+                                "type": "array",
+                                "items": {"type": "string"},
+                                "description": (
+                                    "2-3 predefined choices. The UI appends an "
+                                    "'Other' free-text input automatically. "
+                                    "Omit only when the user must type a free-form answer."
+                                ),
+                                "minItems": 2,
+                                "maxItems": 3,
+                            },
+                        },
+                        "required": ["id", "prompt"],
+                    },
+                    "minItems": 2,
+                    "maxItems": 8,
+                    "description": "List of questions to present to the user.",
+                },
+            },
+            "required": ["questions"],
+        },
+    )
+
+
+def build_set_output_tool(output_keys: list[str] | None) -> Tool | None:
+    """Build the synthetic set_output tool for explicit output declaration."""
+    if not output_keys:
+        return None
+    return Tool(
+        name="set_output",
+        description=(
+            "Set an output value for this node. Call once per output key. "
+            "Use this for brief notes, counts, status, and file references — "
+            "NOT for large data payloads. When a tool result was saved to a "
+            "data file, pass the filename as the value "
+            "(e.g. 'google_sheets_get_values_1.txt') so the next phase can "
+            "load the full data. Values exceeding ~2000 characters are "
+            "auto-saved to data files. "
+            f"Valid keys: {output_keys}"
+        ),
+        parameters={
+            "type": "object",
+            "properties": {
+                "key": {
+                    "type": "string",
+                    "description": f"Output key. Must be one of: {output_keys}",
+                    "enum": output_keys,
+                },
+                "value": {
+                    "type": "string",
+                    "description": (
+                        "The output value — a brief note, count, status, "
+                        "or data filename reference."
+                    ),
+                },
+            },
+            "required": ["key", "value"],
+        },
+    )
+
+
+def build_escalate_tool() -> Tool:
+    """Build the synthetic escalate tool for worker -> queen handoff."""
+    return Tool(
+        name="escalate",
+        description=(
+            "Escalate to the queen when requesting user input, "
+            "blocked by errors, missing "
+            "credentials, or ambiguous constraints that require supervisor "
+            "guidance. Include a concise reason and optional context. "
+            "The node will pause until the queen injects guidance."
+        ),
+        parameters={
+            "type": "object",
+            "properties": {
+                "reason": {
+                    "type": "string",
+                    "description": (
+                        "Short reason for escalation (e.g. 'Tool repeatedly failing')."
+                    ),
+                },
+                "context": {
+                    "type": "string",
+                    "description": "Optional diagnostic details for the queen.",
+                },
+            },
+            "required": ["reason"],
+        },
+    )
+
+def handle_set_output(
+    tool_input: dict[str, Any],
+    output_keys: list[str] | None,
+) -> ToolResult:
+    """Handle set_output tool call. Returns ToolResult (sync)."""
+    import logging
+    import re
+
+    logger = logging.getLogger(__name__)
+
+    key = tool_input.get("key", "")
+    value = tool_input.get("value", "")
+    valid_keys = output_keys or []
+
+    # Recover from truncated JSON (max_tokens hit mid-argument).
+    # The _raw key is set by litellm when json.loads fails.
+    if not key and "_raw" in tool_input:
+        raw = tool_input["_raw"]
+        key_match = re.search(r'"key"\s*:\s*"(\w+)"', raw)
+        if key_match:
+            key = key_match.group(1)
+        val_match = re.search(r'"value"\s*:\s*"', raw)
+        if val_match:
+            start = val_match.end()
+            value = raw[start:].rstrip()
+            for suffix in ('"}\n', '"}', '"'):
+                if value.endswith(suffix):
+                    value = value[: -len(suffix)]
+                    break
+        if key:
+            logger.warning(
+                "Recovered set_output args from truncated JSON: key=%s, value_len=%d",
+                key,
+                len(value),
+            )
+            # Re-inject so the caller sees proper key/value
+            tool_input["key"] = key
+            tool_input["value"] = value
+
+    if key not in valid_keys:
+        return ToolResult(
+            tool_use_id="",
+            content=f"Invalid output key '{key}'. Valid keys: {valid_keys}",
+            is_error=True,
+        )
+
+    return ToolResult(
+        tool_use_id="",
+        content=f"Output '{key}' set successfully.",
+        is_error=False,
+    )

core/framework/agent_loop/internals/tool_result_handler.py

@@ -0,0 +1,499 @@
+"""Tool result handling: truncation, spillover, JSON preview, and execution.
+
+Manages tool result size limits, file spillover for large results, and
+smart JSON previews.  Also includes transient error classification and
+the context-window-exceeded error detector.
+"""
+
+from __future__ import annotations
+
+import asyncio
+import contextvars
+import json
+import logging
+import re
+from pathlib import Path
+from typing import Any
+
+from framework.llm.provider import ToolResult, ToolUse
+from framework.llm.stream_events import ToolCallEvent
+
+logger = logging.getLogger(__name__)
+
+# Pattern for detecting context-window-exceeded errors across LLM providers.
+_CONTEXT_TOO_LARGE_RE = re.compile(
+    r"context.{0,20}(length|window|limit|size)|"
+    r"too.{0,10}(long|large|many.{0,10}tokens)|"
+    r"(exceed|exceeds|exceeded).{0,30}(limit|window|context|tokens)|"
+    r"maximum.{0,20}token|prompt.{0,20}too.{0,10}long",
+    re.IGNORECASE,
+)
+
+
+def is_context_too_large_error(exc: BaseException) -> bool:
+    """Detect whether an exception indicates the LLM input was too large."""
+    cls = type(exc).__name__
+    if "ContextWindow" in cls:
+        return True
+    return bool(_CONTEXT_TOO_LARGE_RE.search(str(exc)))
+
+
+def is_transient_error(exc: BaseException) -> bool:
+    """Classify whether an exception is transient (retryable) vs permanent.
+
+    Transient: network errors, rate limits, server errors, timeouts.
+    Permanent: auth errors, bad requests, context window exceeded.
+    """
+    try:
+        from litellm.exceptions import (
+            APIConnectionError,
+            BadGatewayError,
+            InternalServerError,
+            RateLimitError,
+            ServiceUnavailableError,
+        )
+
+        transient_types: tuple[type[BaseException], ...] = (
+            RateLimitError,
+            APIConnectionError,
+            InternalServerError,
+            BadGatewayError,
+            ServiceUnavailableError,
+            TimeoutError,
+            ConnectionError,
+            OSError,
+        )
+    except ImportError:
+        transient_types = (TimeoutError, ConnectionError, OSError)
+
+    if isinstance(exc, transient_types):
+        return True
+
+    # RuntimeError from StreamErrorEvent with "Stream error:" prefix
+    if isinstance(exc, RuntimeError):
+        error_str = str(exc).lower()
+        transient_keywords = [
+            "rate limit",
+            "429",
+            "timeout",
+            "connection",
+            "internal server",
+            "502",
+            "503",
+            "504",
+            "service unavailable",
+            "bad gateway",
+            "overloaded",
+            "failed to parse tool call",
+        ]
+        return any(kw in error_str for kw in transient_keywords)
+
+    return False
+
+
+def extract_json_metadata(parsed: Any, *, _depth: int = 0, _max_depth: int = 3) -> str:
+    """Return a concise structural summary of parsed JSON.
+
+    Reports key names, value types, and — crucially — array lengths so
+    the LLM knows how much data exists beyond the preview.
+
+    Returns an empty string for simple scalars.
+    """
+    if _depth >= _max_depth:
+        if isinstance(parsed, dict):
+            return f"dict with {len(parsed)} keys"
+        if isinstance(parsed, list):
+            return f"list of {len(parsed)} items"
+        return type(parsed).__name__
+
+    if isinstance(parsed, dict):
+        if not parsed:
+            return "empty dict"
+        lines: list[str] = []
+        indent = "  " * (_depth + 1)
+        for key, value in list(parsed.items())[:20]:
+            if isinstance(value, list):
+                line = f'{indent}"{key}": list of {len(value)} items'
+                if value:
+                    first = value[0]
+                    if isinstance(first, dict):
+                        sample_keys = list(first.keys())[:10]
+                        line += f" (each item: dict with keys {sample_keys})"
+                    elif isinstance(first, list):
+                        line += f" (each item: list of {len(first)} elements)"
+                lines.append(line)
+            elif isinstance(value, dict):
+                child = extract_json_metadata(value, _depth=_depth + 1, _max_depth=_max_depth)
+                lines.append(f'{indent}"{key}": {child}')
+            else:
+                lines.append(f'{indent}"{key}": {type(value).__name__}')
+        if len(parsed) > 20:
+            lines.append(f"{indent}... and {len(parsed) - 20} more keys")
+        return "\n".join(lines)
+
+    if isinstance(parsed, list):
+        if not parsed:
+            return "empty list"
+        desc = f"list of {len(parsed)} items"
+        first = parsed[0]
+        if isinstance(first, dict):
+            sample_keys = list(first.keys())[:10]
+            desc += f" (each item: dict with keys {sample_keys})"
+        elif isinstance(first, list):
+            desc += f" (each item: list of {len(first)} elements)"
+        return desc
+
+    return ""
+
+
+def build_json_preview(parsed: Any, *, max_chars: int = 5000) -> str | None:
+    """Build a smart preview of parsed JSON, truncating large arrays.
+
+    Shows first 3 + last 1 items of large arrays with explicit count
+    markers so the LLM cannot mistake the preview for the full dataset.
+
+    Returns ``None`` if no truncation was needed (no large arrays).
+    """
+    _LARGE_ARRAY_THRESHOLD = 10
+
+    def _truncate_arrays(obj: Any) -> tuple[Any, bool]:
+        """Return (truncated_copy, was_truncated)."""
+        if isinstance(obj, list) and len(obj) > _LARGE_ARRAY_THRESHOLD:
+            n = len(obj)
+            head = obj[:3]
+            tail = obj[-1:]
+            marker = f"... ({n - 4} more items omitted, {n} total) ..."
+            return head + [marker] + tail, True
+        if isinstance(obj, dict):
+            changed = False
+            out: dict[str, Any] = {}
+            for k, v in obj.items():
+                new_v, did = _truncate_arrays(v)
+                out[k] = new_v
+                changed = changed or did
+            return (out, True) if changed else (obj, False)
+        return obj, False
+
+    preview_obj, was_truncated = _truncate_arrays(parsed)
+    if not was_truncated:
+        return None  # No large arrays — caller should use raw slicing
+
+    try:
+        result = json.dumps(preview_obj, indent=2, ensure_ascii=False)
+    except (TypeError, ValueError):
+        return None
+
+    if len(result) > max_chars:
+        # Even 3+1 items too big — try just 1 item
+        def _minimal_arrays(obj: Any) -> Any:
+            if isinstance(obj, list) and len(obj) > _LARGE_ARRAY_THRESHOLD:
+                n = len(obj)
+                return obj[:1] + [f"... ({n - 1} more items omitted, {n} total) ..."]
+            if isinstance(obj, dict):
+                return {k: _minimal_arrays(v) for k, v in obj.items()}
+            return obj
+
+        preview_obj = _minimal_arrays(parsed)
+        try:
+            result = json.dumps(preview_obj, indent=2, ensure_ascii=False)
+        except (TypeError, ValueError):
+            return None
+        if len(result) > max_chars:
+            result = result[:max_chars] + "…"
+
+    return result
+
+
+def truncate_tool_result(
+    result: ToolResult,
+    tool_name: str,
+    *,
+    max_tool_result_chars: int,
+    spillover_dir: str | None,
+    next_spill_filename_fn: Any,  # Callable[[str], str]
+) -> ToolResult:
+    """Persist tool result to file and optionally truncate for context.
+
+    When *spillover_dir* is configured, EVERY non-error tool result is
+    saved to a file (short filename like ``web_search_1.txt``).  A
+    ``[Saved to '...']`` annotation is appended so the reference
+    survives pruning and compaction.
+
+    - Small results (≤ limit): full content kept + file annotation
+    - Large results (> limit): preview + file reference
+    - Errors: pass through unchanged
+    - read_file/load_data results: truncate with pagination hint (no re-spill)
+    """
+    limit = max_tool_result_chars
+
+    # Errors always pass through unchanged
+    if result.is_error:
+        return result
+
+    # read_file/load_data reads FROM spilled files — never re-spill (circular).
+    # Just truncate with a pagination hint if the result is too large.
+    if tool_name in ("load_data", "read_file"):
+        if limit <= 0 or len(result.content) <= limit:
+            return result  # Small result — pass through as-is
+        # Large result — truncate with smart preview
+        PREVIEW_CAP = min(5000, max(limit - 500, limit // 2))
+
+        metadata_str = ""
+        smart_preview: str | None = None
+        try:
+            parsed_ld = json.loads(result.content)
+            metadata_str = extract_json_metadata(parsed_ld)
+            smart_preview = build_json_preview(parsed_ld, max_chars=PREVIEW_CAP)
+        except (json.JSONDecodeError, TypeError, ValueError):
+            pass
+
+        if smart_preview is not None:
+            preview_block = smart_preview
+        else:
+            preview_block = result.content[:PREVIEW_CAP] + "…"
+
+        header = (
+            f"[{tool_name} result: {len(result.content):,} chars — "
+            f"too large for context. Use offset_bytes/limit_bytes "
+            f"parameters to read smaller chunks.]"
+        )
+        if metadata_str:
+            header += f"\n\nData structure:\n{metadata_str}"
+        header += (
+            "\n\nWARNING: This is an INCOMPLETE preview. Do NOT draw conclusions or counts from it."
+        )
+
+        truncated = f"{header}\n\nPreview (small sample only):\n{preview_block}"
+        logger.info(
+            "%s result truncated: %d → %d chars (use offset/limit to paginate)",
+            tool_name,
+            len(result.content),
+            len(truncated),
+        )
+        return ToolResult(
+            tool_use_id=result.tool_use_id,
+            content=truncated,
+            is_error=False,
+            image_content=result.image_content,
+            is_skill_content=result.is_skill_content,
+        )
+
+    spill_dir = spillover_dir
+    if spill_dir:
+        spill_path = Path(spill_dir)
+        spill_path.mkdir(parents=True, exist_ok=True)
+        filename = next_spill_filename_fn(tool_name)
+
+        # Pretty-print JSON content so read_file's line-based
+        # pagination works correctly.
+        write_content = result.content
+        parsed_json: Any = None  # track for metadata extraction
+        try:
+            parsed_json = json.loads(result.content)
+            write_content = json.dumps(parsed_json, indent=2, ensure_ascii=False)
+        except (json.JSONDecodeError, TypeError, ValueError):
+            pass  # Not JSON — write as-is
+
+        file_path = spill_path / filename
+        file_path.write_text(write_content, encoding="utf-8")
+        # Use absolute path so parent agents can find files from subagents
+        abs_path = str(file_path.resolve())
+
+        if limit > 0 and len(result.content) > limit:
+            # Large result: build a small, metadata-rich preview so the
+            # LLM cannot mistake it for the complete dataset.
+            PREVIEW_CAP = 5000
+
+            # Extract structural metadata (array lengths, key names)
+            metadata_str = ""
+            smart_preview: str | None = None
+            if parsed_json is not None:
+                metadata_str = extract_json_metadata(parsed_json)
+                smart_preview = build_json_preview(parsed_json, max_chars=PREVIEW_CAP)
+
+            if smart_preview is not None:
+                preview_block = smart_preview
+            else:
+                preview_block = result.content[:PREVIEW_CAP] + "…"
+
+            # Assemble header with structural info + warning
+            header = (
+                f"[Result from {tool_name}: {len(result.content):,} chars — "
+                f"too large for context, saved to '{abs_path}'.]\n"
+            )
+            if metadata_str:
+                header += f"\nData structure:\n{metadata_str}"
+            header += (
+                f"\n\nWARNING: The preview below is INCOMPLETE. "
+                f"Do NOT draw conclusions or counts from it. "
+                f"Use read_file(path='{abs_path}') to read the "
+                f"full data before analysis."
+            )
+
+            content = f"{header}\n\nPreview (small sample only):\n{preview_block}"
+            logger.info(
+                "Tool result spilled to file: %s (%d chars → %s)",
+                tool_name,
+                len(result.content),
+                abs_path,
+            )
+        else:
+            # Small result: keep full content + annotation with absolute path
+            content = f"{result.content}\n\n[Saved to '{abs_path}']"
+            logger.info(
+                "Tool result saved to file: %s (%d chars → %s)",
+                tool_name,
+                len(result.content),
+                filename,
+            )
+
+        return ToolResult(
+            tool_use_id=result.tool_use_id,
+            content=content,
+            is_error=False,
+            image_content=result.image_content,
+            is_skill_content=result.is_skill_content,
+        )
+
+    # No spillover_dir — truncate in-place if needed
+    if limit > 0 and len(result.content) > limit:
+        PREVIEW_CAP = min(5000, max(limit - 500, limit // 2))
+
+        metadata_str = ""
+        smart_preview: str | None = None
+        try:
+            parsed_inline = json.loads(result.content)
+            metadata_str = extract_json_metadata(parsed_inline)
+            smart_preview = build_json_preview(parsed_inline, max_chars=PREVIEW_CAP)
+        except (json.JSONDecodeError, TypeError, ValueError):
+            pass
+
+        if smart_preview is not None:
+            preview_block = smart_preview
+        else:
+            preview_block = result.content[:PREVIEW_CAP] + "…"
+
+        header = (
+            f"[Result from {tool_name}: {len(result.content):,} chars — "
+            f"truncated to fit context budget.]"
+        )
+        if metadata_str:
+            header += f"\n\nData structure:\n{metadata_str}"
+        header += (
+            "\n\nWARNING: This is an INCOMPLETE preview. "
+            "Do NOT draw conclusions or counts from the preview alone."
+        )
+
+        truncated = f"{header}\n\n{preview_block}"
+        logger.info(
+            "Tool result truncated in-place: %s (%d → %d chars)",
+            tool_name,
+            len(result.content),
+            len(truncated),
+        )
+        return ToolResult(
+            tool_use_id=result.tool_use_id,
+            content=truncated,
+            is_error=False,
+            image_content=result.image_content,
+            is_skill_content=result.is_skill_content,
+        )
+
+    return result
+
+
+async def execute_tool(
+    tool_executor: Any,  # Callable[[ToolUse], ToolResult | Awaitable[ToolResult]] | None
+    tc: ToolCallEvent,
+    timeout: float,
+    skill_dirs: list[str] | None = None,
+) -> ToolResult:
+    """Execute a tool call, handling both sync and async executors.
+
+    Applies ``tool_call_timeout_seconds`` to prevent hung MCP servers
+    from blocking the event loop indefinitely.  The initial executor
+    call is offloaded to a thread pool so that sync executors don't
+    freeze the event loop.
+    """
+    if tool_executor is None:
+        return ToolResult(
+            tool_use_id=tc.tool_use_id,
+            content=f"No tool executor configured for '{tc.tool_name}'",
+            is_error=True,
+        )
+
+    skill_dirs = skill_dirs or []
+    skill_read_tools = {"view_file", "load_data", "read_file"}
+    if tc.tool_name in skill_read_tools and skill_dirs:
+        raw_path = tc.tool_input.get("path", "")
+        if raw_path:
+            resolved = Path(raw_path).resolve(strict=False)
+            resolved_roots = [Path(skill_dir).resolve(strict=False) for skill_dir in skill_dirs]
+            if any(resolved.is_relative_to(root) for root in resolved_roots):
+                try:
+                    content = resolved.read_text(encoding="utf-8")
+                except Exception as exc:
+                    return ToolResult(
+                        tool_use_id=tc.tool_use_id,
+                        content=f"Could not read skill resource '{raw_path}': {exc}",
+                        is_error=True,
+                    )
+                return ToolResult(
+                    tool_use_id=tc.tool_use_id,
+                    content=content,
+                    is_skill_content=resolved.name == "SKILL.md",
+                )
+
+    tool_use = ToolUse(id=tc.tool_use_id, name=tc.tool_name, input=tc.tool_input)
+
+    async def _run() -> ToolResult:
+        # Offload the executor call to a thread.  Sync MCP executors
+        # block on future.result() — running in a thread keeps the
+        # event loop free so asyncio.wait_for can fire the timeout.
+        # Copy the current context so contextvars (e.g. data_dir from
+        # execution context) propagate into the worker thread.
+        loop = asyncio.get_running_loop()
+        ctx = contextvars.copy_context()
+        result = await loop.run_in_executor(None, ctx.run, tool_executor, tool_use)
+        # Async executors return a coroutine — await it on the loop
+        if asyncio.iscoroutine(result) or asyncio.isfuture(result):
+            result = await result
+        return result
+
+    try:
+        if timeout > 0:
+            result = await asyncio.wait_for(_run(), timeout=timeout)
+        else:
+            result = await _run()
+    except TimeoutError:
+        logger.warning("Tool '%s' timed out after %.0fs", tc.tool_name, timeout)
+        return ToolResult(
+            tool_use_id=tc.tool_use_id,
+            content=(
+                f"Tool '{tc.tool_name}' timed out after {timeout:.0f}s. "
+                "The operation took too long and was cancelled. "
+                "Try a simpler request or a different approach."
+            ),
+            is_error=True,
+        )
+    return result
+
+
+def restore_spill_counter(spillover_dir: str | None) -> int:
+    """Scan spillover_dir for existing spill files and return the max counter.
+
+    Returns the highest spill number found (or 0 if none).
+    """
+    if not spillover_dir:
+        return 0
+    spill_path = Path(spillover_dir)
+    if not spill_path.is_dir():
+        return 0
+    max_n = 0
+    for f in spill_path.iterdir():
+        if not f.is_file():
+            continue
+        m = re.search(r"_(\d+)\.txt$", f.name)
+        if m:
+            max_n = max(max_n, int(m.group(1)))
+    return max_n

core/framework/agent_loop/internals/types.py

@@ -0,0 +1,205 @@
+"""Shared types and state containers for the event loop package."""
+
+from __future__ import annotations
+
+import json
+import logging
+import time
+from dataclasses import dataclass, field
+from pathlib import Path
+from typing import Any, Literal, Protocol, runtime_checkable
+
+from framework.agent_loop.conversation import (
+    ConversationStore,
+)
+
+logger = logging.getLogger(__name__)
+
+
+@dataclass
+class TriggerEvent:
+    """A framework-level trigger signal (timer tick or webhook hit)."""
+
+    trigger_type: str
+    source_id: str
+    payload: dict[str, Any] = field(default_factory=dict)
+    timestamp: float = field(default_factory=time.time)
+
+
+@dataclass
+class JudgeVerdict:
+    """Result of judge evaluation for the event loop."""
+
+    action: Literal["ACCEPT", "RETRY", "ESCALATE"]
+    # None  = no evaluation happened (skip_judge, tool-continue); not logged.
+    # ""    = evaluated but no feedback; logged with default text.
+    # "..." = evaluated with feedback; logged as-is.
+    feedback: str | None = None
+
+
+@runtime_checkable
+class JudgeProtocol(Protocol):
+    """Protocol for event-loop judges."""
+
+    async def evaluate(self, context: dict[str, Any]) -> JudgeVerdict: ...
+
+
+@dataclass
+class LoopConfig:
+    """Configuration for the event loop."""
+
+    max_iterations: int = 50
+    max_tool_calls_per_turn: int = 30
+    judge_every_n_turns: int = 1
+    stall_detection_threshold: int = 3
+    stall_similarity_threshold: float = 0.85
+    max_context_tokens: int = 32_000
+    store_prefix: str = ""
+
+    # Overflow margin for max_tool_calls_per_turn. Tool calls are only
+    # discarded when the count exceeds max_tool_calls_per_turn * (1 + margin).
+    tool_call_overflow_margin: float = 0.5
+
+    # Tool result context management.
+    max_tool_result_chars: int = 30_000
+    spillover_dir: str | None = None
+
+    # set_output value spilling.
+    max_output_value_chars: int = 2_000
+
+    # Stream retry.
+    max_stream_retries: int = 5
+    stream_retry_backoff_base: float = 2.0
+    stream_retry_max_delay: float = 60.0
+
+    # Tool doom loop detection.
+    tool_doom_loop_threshold: int = 3
+
+    # Client-facing auto-block grace period.
+    cf_grace_turns: int = 1
+    # Worker auto-escalation: text-only turns before escalating to queen.
+    worker_escalation_grace_turns: int = 1
+    tool_doom_loop_enabled: bool = True
+
+    # Per-tool-call timeout.
+    tool_call_timeout_seconds: float = 60.0
+
+    # Subagent delegation timeout (wall-clock max).
+    subagent_timeout_seconds: float = 3600.0
+
+    # Subagent inactivity timeout - only timeout if no activity for this duration.
+    # This resets whenever the subagent makes progress (tool calls, LLM responses).
+    # Set to 0 to use only the wall-clock timeout.
+    subagent_inactivity_timeout_seconds: float = 300.0
+
+    # Lifecycle hooks.
+    hooks: dict[str, list] | None = None
+
+    def __post_init__(self) -> None:
+        if self.hooks is None:
+            object.__setattr__(self, "hooks", {})
+
+
+@dataclass
+class HookContext:
+    """Context passed to every lifecycle hook."""
+
+    event: str
+    trigger: str | None
+    system_prompt: str
+
+
+@dataclass
+class HookResult:
+    """What a hook may return to modify node state."""
+
+    system_prompt: str | None = None
+    inject: str | None = None
+
+
+@dataclass
+class OutputAccumulator:
+    """Accumulates output key-value pairs with optional write-through persistence."""
+
+    values: dict[str, Any] = field(default_factory=dict)
+    store: ConversationStore | None = None
+    spillover_dir: str | None = None
+    max_value_chars: int = 0
+    run_id: str | None = None
+
+    async def set(self, key: str, value: Any) -> None:
+        """Set a key-value pair, auto-spilling large values to files."""
+        value = self._auto_spill(key, value)
+        self.values[key] = value
+        if self.store:
+            cursor = await self.store.read_cursor() or {}
+            outputs = cursor.get("outputs", {})
+            outputs[key] = value
+            cursor["outputs"] = outputs
+            await self.store.write_cursor(cursor)
+
+    def _auto_spill(self, key: str, value: Any) -> Any:
+        """Save large values to a file and return a reference string."""
+        if self.max_value_chars <= 0 or not self.spillover_dir:
+            return value
+
+        val_str = json.dumps(value, ensure_ascii=False) if not isinstance(value, str) else value
+        if len(val_str) <= self.max_value_chars:
+            return value
+
+        spill_path = Path(self.spillover_dir)
+        spill_path.mkdir(parents=True, exist_ok=True)
+        ext = ".json" if isinstance(value, (dict, list)) else ".txt"
+        filename = f"output_{key}{ext}"
+        write_content = (
+            json.dumps(value, indent=2, ensure_ascii=False)
+            if isinstance(value, (dict, list))
+            else str(value)
+        )
+        file_path = spill_path / filename
+        file_path.write_text(write_content, encoding="utf-8")
+        file_size = file_path.stat().st_size
+        logger.info(
+            "set_output value auto-spilled: key=%s, %d chars -> %s (%d bytes)",
+            key,
+            len(val_str),
+            filename,
+            file_size,
+        )
+        # Use absolute path so parent agents can find files from subagents
+        abs_path = str(file_path.resolve())
+        return (
+            f"[Saved to '{abs_path}' ({file_size:,} bytes). "
+            f"Use read_file(path='{abs_path}') "
+            f"to access full data.]"
+        )
+
+    def get(self, key: str) -> Any | None:
+        return self.values.get(key)
+
+    def to_dict(self) -> dict[str, Any]:
+        return dict(self.values)
+
+    def has_all_keys(self, required: list[str]) -> bool:
+        return all(key in self.values and self.values[key] is not None for key in required)
+
+    @classmethod
+    async def restore(
+        cls,
+        store: ConversationStore,
+        run_id: str | None = None,
+    ) -> OutputAccumulator:
+        cursor = await store.read_cursor()
+        values = cursor.get("outputs", {}) if cursor else {}
+        return cls(values=values, store=store, run_id=run_id)
+
+
+__all__ = [
+    "HookContext",
+    "HookResult",
+    "JudgeProtocol",
+    "JudgeVerdict",
+    "LoopConfig",
+    "OutputAccumulator",
+    "TriggerEvent",
+]

core/framework/agents/__init__.py

@@ -8,6 +8,14 @@ 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()],
+        [
+            p
+            for p in FRAMEWORK_AGENTS_DIR.iterdir()
+            if p.is_dir()
+            and (
+                (p / "agent.json").exists()
+                or (p / "agent.py").exists()
+            )
+        ],
         key=lambda p: p.name,
     )

core/framework/agents/credential_tester/__main__.py

@@ -1,8 +1,6 @@
 """CLI entry point for Credential Tester agent."""
 
 import asyncio
-import logging
-import sys
 
 import click
 
@@ -10,13 +8,14 @@ from .agent import CredentialTesterAgent
 
 
 def setup_logging(verbose=False, debug=False):
+    from framework.observability import configure_logging
+
     if debug:
-        level, fmt = logging.DEBUG, "%(asctime)s %(name)s: %(message)s"
+        configure_logging(level="DEBUG")
     elif verbose:
-        level, fmt = logging.INFO, "%(message)s"
+        configure_logging(level="INFO")
     else:
-        level, fmt = logging.WARNING, "%(levelname)s: %(message)s"
-    logging.basicConfig(level=level, format=fmt, stream=sys.stderr)
+        configure_logging(level="WARNING")
 
 
 def pick_account(agent: CredentialTesterAgent) -> dict | None:

core/framework/agents/credential_tester/agent.py

@@ -16,23 +16,30 @@ after the user picks an account programmatically.
 
 from __future__ import annotations
 
+import logging
 from pathlib import Path
 from typing import TYPE_CHECKING
 
-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.config import get_max_context_tokens
+from framework.orchestrator import Goal, NodeSpec, SuccessCriterion
+from framework.orchestrator.checkpoint_config import CheckpointConfig
+from framework.orchestrator.edge import GraphSpec
+from framework.orchestrator.orchestrator import ExecutionResult
 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 framework.loader.mcp_registry import MCPRegistry
+from framework.loader.tool_registry import ToolRegistry
+from framework.host.agent_host import AgentHost
+from framework.host.execution_manager 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
+    from framework.loader import AgentLoader
+
+logger = logging.getLogger(__name__)
 
 # ---------------------------------------------------------------------------
 # Goal
@@ -106,7 +113,11 @@ def _list_aden_accounts() -> list[dict]:
             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 []
 
 
@@ -118,7 +129,11 @@ def _list_local_accounts() -> list[dict]:
         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 []
 
 
@@ -139,7 +154,11 @@ def _list_env_fallback_accounts() -> list[dict]:
         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:
@@ -214,7 +233,7 @@ requires_account_selection = True
 """Signal TUI to show account picker before starting the agent."""
 
 
-def configure_for_account(runner: AgentRunner, account: dict) -> None:
+def configure_for_account(runner: AgentLoader, account: dict) -> None:
     """Scope the tester node's tools to the selected provider.
 
     Handles both Aden accounts (account= routing) and local accounts
@@ -299,12 +318,14 @@ def _activate_local_account(credential_id: str, alias: str) -> None:
 
             if key:
                 os.environ[spec.env_var] = key
+    except (ImportError, KeyError, OSError) as exc:
+        logger.debug("Could not inject credentials: %s", exc)
     except Exception:
-        pass
+        logger.warning("Unexpected error injecting credentials", exc_info=True)
 
 
 def _configure_aden_node(
-    runner: AgentRunner,
+    runner: AgentLoader,
     provider: str,
     alias: str,
     detail: str,
@@ -347,7 +368,7 @@ or any other identifier — always use the alias exactly as shown.
 
 
 def _configure_local_node(
-    runner: AgentRunner,
+    runner: AgentLoader,
     provider: str,
     alias: str,
     identity: dict,
@@ -455,7 +476,6 @@ identity_prompt = (
 loop_config = {
     "max_iterations": 50,
     "max_tool_calls_per_turn": 30,
-    "max_history_tokens": 32000,
 }
 
 # ---------------------------------------------------------------------------
@@ -477,7 +497,7 @@ class CredentialTesterAgent:
     def __init__(self, config=None):
         self.config = config or default_config
         self._selected_account: dict | None = None
-        self._agent_runtime: AgentRuntime | None = None
+        self._agent_runtime: AgentHost | None = None
         self._tool_registry: ToolRegistry | None = None
         self._storage_path: Path | None = None
 
@@ -541,7 +561,7 @@ class CredentialTesterAgent:
             loop_config={
                 "max_iterations": 50,
                 "max_tool_calls_per_turn": 30,
-                "max_history_tokens": 32000,
+                "max_context_tokens": get_max_context_tokens(),
             },
             conversation_mode="continuous",
             identity_prompt=(
@@ -563,6 +583,23 @@ class CredentialTesterAgent:
         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,
@@ -576,7 +613,7 @@ class CredentialTesterAgent:
 
         graph = self._build_graph()
 
-        self._agent_runtime = create_agent_runtime(
+        self._agent_runtime = AgentHost(
             graph=graph,
             goal=goal,
             storage_path=self._storage_path,

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

@@ -1,6 +1,6 @@
 """Node definitions for Credential Tester agent."""
 
-from framework.graph import NodeSpec
+from framework.orchestrator import NodeSpec
 
 
 def build_tester_node(

core/framework/agents/discovery.py

@@ -16,31 +16,70 @@ class AgentEntry:
     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_name: str) -> str | None:
-    """Return the most recent updated_at timestamp across all sessions."""
-    sessions_dir = Path.home() / ".hive" / "agents" / agent_name / "sessions"
-    if not sessions_dir.exists():
-        return 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/agents/queens/default/sessions/``) whose
+    ``meta.json`` references the same *agent_path*.
+    """
+    from datetime import datetime
+
+    agent_name = agent_path.name
     latest: str | None = None
-    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
+
+    # 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 (scan all queen identity directories)
+    from framework.config import QUEENS_DIR
+
+    if QUEENS_DIR.exists():
+        resolved = agent_path.resolve()
+        for queen_dir in QUEENS_DIR.iterdir():
+            if not queen_dir.is_dir():
+                continue
+            sessions_dir = queen_dir / "sessions"
+            if not sessions_dir.exists():
+                continue
+            for d in 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
 
 
@@ -52,16 +91,61 @@ def _count_sessions(agent_name: str) -> int:
     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.
+    Checks agent.json (declarative) first, then agent.py (legacy).
     """
     import ast
 
     node_count, tool_count, tags = 0, 0, []
 
+    # Declarative JSON agents (preferred)
+    agent_json = agent_path / "agent.json"
+    if agent_json.exists():
+        try:
+            data = json.loads(agent_json.read_text(encoding="utf-8"))
+            if isinstance(data, dict):
+                json_nodes = data.get("nodes", [])
+                node_count = len(json_nodes)
+                tools: set[str] = set()
+                for n in json_nodes:
+                    node_tools = n.get("tools", {})
+                    if isinstance(node_tools, dict):
+                        tools.update(node_tools.get("allowed", []))
+                    elif isinstance(node_tools, list):
+                        tools.update(node_tools)
+                tool_count = len(tools)
+                return node_count, tool_count, tags
+        except Exception:
+            pass
+
+    # Legacy: agent.py (AST-parsed)
     agent_py = agent_path / "agent.py"
     if agent_py.exists():
         try:
@@ -75,39 +159,31 @@ def _extract_agent_stats(agent_path: Path) -> tuple[int, int, list[str]]:
         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("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 (
+    from framework.loader.cli import (
         _extract_python_agent_metadata,
         _get_framework_agents_dir,
         _is_valid_agent_dir,
     )
 
+    from framework.config import COLONIES_DIR
+
     groups: dict[str, list[AgentEntry]] = {}
     sources = [
-        ("Your Agents", Path("exports")),
+        ("Your Agents", COLONIES_DIR),
+        ("Your Agents", Path("exports")),  # compat fallback
         ("Framework", _get_framework_agents_dir()),
         ("Examples", Path("examples/templates")),
     ]
 
+    # Track seen agent directory names to avoid duplicates when the same
+    # agent exists in both colonies/ and exports/ (colonies takes priority).
+    _seen_agent_names: set[str] = set()
+
     for category, base_dir in sources:
         if not base_dir.exists():
             continue
@@ -115,6 +191,9 @@ def discover_agents() -> dict[str, list[AgentEntry]]:
         for path in sorted(base_dir.iterdir(), key=lambda p: p.name):
             if not _is_valid_agent_dir(path):
                 continue
+            if path.name in _seen_agent_names:
+                continue
+            _seen_agent_names.add(path.name)
 
             name, desc = _extract_python_agent_metadata(path)
             config_fallback_name = path.name.replace("_", " ").title()
@@ -122,13 +201,19 @@ def discover_agents() -> dict[str, list[AgentEntry]]:
 
             node_count, tool_count, tags = _extract_agent_stats(path)
             if not used_config:
-                agent_json = path / "agent.json"
-                if agent_json.exists():
+                # Try agent.json (declarative) for metadata
+                agent_json_path = path / "agent.json"
+                if agent_json_path.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)
+                        data = json.loads(
+                            agent_json_path.read_text(encoding="utf-8"),
+                        )
+                        if isinstance(data, dict):
+                            raw_name = data.get("name", name)
+                            if "-" in raw_name and " " not in raw_name:
+                                raw_name = raw_name.replace("-", " ").title()
+                            name = raw_name
+                            desc = data.get("description", desc)
                     except Exception:
                         pass
 
@@ -139,13 +224,16 @@ def discover_agents() -> dict[str, list[AgentEntry]]:
                     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.name),
+                    last_active=_get_last_active(path),
                 )
             )
         if entries:
-            groups[category] = entries
+            existing = groups.get(category, [])
+            existing.extend(entries)
+            groups[category] = existing
 
     return groups

core/framework/agents/queen/__init__.py

@@ -1,19 +1,13 @@
-"""
-Queen — Native agent builder for the Hive framework.
+"""Queen -- the 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 .agent import queen_goal, queen_loop_config
 from .config import AgentMetadata, RuntimeConfig, default_config, metadata
 
 __version__ = "1.0.0"
 
 __all__ = [
     "queen_goal",
-    "queen_graph",
+    "queen_loop_config",
     "RuntimeConfig",
     "AgentMetadata",
     "default_config",

core/framework/agents/queen/agent.py

@@ -1,40 +1,29 @@
-"""Queen graph definition."""
+"""Queen agent definition.
 
-from framework.graph import Goal
-from framework.graph.edge import GraphSpec
+The queen is a single AgentLoop -- no graph, no orchestrator.
+Loaded by queen_orchestrator.create_queen().
+"""
+
+from framework.orchestrator.goal import Goal
 
 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. Triage health escalations from the judge."
+        "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,
-        "max_history_tokens": 32000,
-    },
-)
+# Loop config -- used by queen_orchestrator to build LoopConfig
+queen_loop_config = {
+    "max_iterations": 999_999,
+    "max_tool_calls_per_turn": 30,
+    "max_context_tokens": 180_000,
+}
+
+__all__ = ["queen_goal", "queen_loop_config", "queen_node"]

core/framework/agents/queen/mcp_registry.json

@@ -0,0 +1,3 @@
+{
+  "include": ["gcu-tools"]
+}

core/framework/agents/queen/mcp_servers.json

@@ -5,5 +5,12 @@
     "args": ["run", "python", "coder_tools_server.py", "--stdio"],
     "cwd": "../../../../tools",
     "description": "Unsandboxed file system tools for code generation and validation"
+  },
+  "gcu-tools": {
+    "transport": "stdio",
+    "command": "uv",
+    "args": ["run", "python", "-m", "gcu.server", "--stdio", "--capabilities", "browser"],
+    "cwd": "../../../../tools",
+    "description": "Browser automation tools (Playwright-based)"
   }
 }

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

@@ -1,80 +0,0 @@
-"""Queen thinking hook — HR persona 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, then returns a persona
-prefix string that replaces the queen's default "Solution Architect" identity.
-
-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.
-"""
-
-from __future__ import annotations
-
-import json
-import logging
-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 talent consultant at a world-class firm.
-A new request has arrived and you must identify which professional's expertise
-would produce the highest-quality response.
-
-Reply with ONLY a valid JSON object — no markdown, no prose, no explanation:
-{"role": "<job title>", "persona": "<2-3 sentence first-person identity statement>"}
-
-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.
-"""
-
-
-async def select_expert_persona(user_message: str, llm: LLMProvider) -> str:
-    """Run the HR classifier and return a persona prefix string.
-
-    Makes a single non-streaming acomplete() call with the session LLM.
-    Returns an empty string on any failure so the queen falls back
-    gracefully to its default "Solution Architect" identity.
-
-    Args:
-        user_message: The user's opening message for the session.
-        llm: The session LLM provider.
-
-    Returns:
-        A persona prefix like "You are a CFO. I am a CFO with 20 years..."
-        or "" on failure.
-    """
-    if not user_message.strip():
-        return ""
-
-    try:
-        response = await llm.acomplete(
-            messages=[{"role": "user", "content": user_message}],
-            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()
-        if not role or not persona:
-            logger.warning("Thinking hook: empty role/persona in response: %r", raw)
-            return ""
-        result = f"You are a {role}. {persona}"
-        logger.info("Thinking hook: selected persona — %s", role)
-        return result
-    except Exception:
-        logger.warning("Thinking hook: persona classification failed", exc_info=True)
-        return ""

core/framework/agents/queen/queen_memory.py

@@ -1,371 +0,0 @@
-"""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 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 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()
-    episodic = read_episodic_memory()
-
-    # 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 episodic:
-        today_str = date.today().strftime("%B %-d, %Y")
-        parts.append(f"## Today — {today_str}\n\n{episodic}")
-
-    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
-
-## 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_str = date.today().strftime("%B %-d, %Y")
-    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.
-- 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.
-
-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 Exception:
-                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_str = date.today().strftime("%B %-d, %Y")
-        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 Exception:
-            pass

core/framework/agents/queen/queen_memory_v2.py

@@ -0,0 +1,235 @@
+"""Queen global memory helpers.
+
+Memory hierarchy::
+
+    ~/.hive/memories/
+        global/              # shared across all queens and colonies
+        colonies/{name}/     # colony-scoped memories
+        agents/queens/{name}/ # queen-specific memories
+        agents/{name}/       # per-worker-agent memories
+
+Each memory is an individual ``.md`` file with optional YAML frontmatter
+(name, type, description).
+"""
+
+from __future__ import annotations
+
+import logging
+import re
+from dataclasses import dataclass, field
+from pathlib import Path
+
+logger = logging.getLogger(__name__)
+
+# ---------------------------------------------------------------------------
+# Constants
+# ---------------------------------------------------------------------------
+
+GLOBAL_MEMORY_CATEGORIES: tuple[str, ...] = ("profile", "preference", "environment", "feedback")
+
+from framework.config import MEMORIES_DIR
+
+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
+
+
+def global_memory_dir() -> Path:
+    """Return the global memory directory (shared across all queens/colonies)."""
+    return MEMORIES_DIR / "global"
+
+
+def colony_memory_dir(colony_name: str) -> Path:
+    """Return the memory directory for a named colony."""
+    return MEMORIES_DIR / "colonies" / colony_name
+
+
+def queen_memory_dir(queen_name: str = "default") -> Path:
+    """Return the memory directory for a named queen."""
+    return MEMORIES_DIR / "agents" / "queens" / queen_name
+
+
+def agent_memory_dir(agent_name: str) -> Path:
+    """Return the memory directory for a worker agent."""
+    return MEMORIES_DIR / "agents" / agent_name
+
+
+# ---------------------------------------------------------------------------
+# 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_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_global_memory_category(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 global_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"
+    )
+
+
+# ---------------------------------------------------------------------------
+# Manifest formatting
+# ---------------------------------------------------------------------------
+
+
+def format_memory_manifest(files: list[MemoryFile]) -> str:
+    """One-line-per-file text manifest.
+
+    Format: ``[type] filename: description``
+    """
+    lines: list[str] = []
+    for mf in files:
+        t = mf.type or "unknown"
+        desc = mf.description or "(no description)"
+        lines.append(f"[{t}] {mf.filename}: {desc}")
+    return "\n".join(lines)
+
+
+# ---------------------------------------------------------------------------
+# Initialisation
+# ---------------------------------------------------------------------------
+
+
+def init_memory_dir(memory_dir: Path | None = None) -> None:
+    """Create the memory directory if missing."""
+    d = memory_dir or global_memory_dir()
+    d.mkdir(parents=True, exist_ok=True)

core/framework/agents/queen/recall_selector.py

@@ -0,0 +1,141 @@
+"""Recall selector — pre-turn global memory selection for the queen.
+
+Before each conversation turn the system:
+  1. Scans the global 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 the system prompt.
+
+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 (
+    format_memory_manifest,
+    global_memory_dir,
+    scan_memory_files,
+)
+
+logger = logging.getLogger(__name__)
+
+# ---------------------------------------------------------------------------
+# Structured output schema
+# ---------------------------------------------------------------------------
+
+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.
+"""
+
+# ---------------------------------------------------------------------------
+# Core functions
+# ---------------------------------------------------------------------------
+
+
+async def select_memories(
+    query: str,
+    llm: Any,
+    memory_dir: Path | 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 global_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 memories for query: %.100s", len(files), query)
+    manifest = format_memory_manifest(files)
+    user_msg = f"## User query\n\n{query}\n\n## Available memories\n\n{manifest}"
+
+    try:
+        resp = await llm.acomplete(
+            messages=[{"role": "user", "content": user_msg}],
+            system=SELECT_MEMORIES_SYSTEM_PROMPT,
+            max_tokens=1024,
+            response_format={"type": "json_object"},
+        )
+        raw = (resp.content or "").strip()
+        if not raw:
+            logger.warning(
+                "recall: LLM returned empty response (model=%s, stop=%s)",
+                resp.model,
+                resp.stop_reason,
+            )
+            return []
+        # Some models wrap JSON in markdown fences or add preamble text.
+        # Try to extract the JSON object if raw parse fails.
+        try:
+            data = json.loads(raw)
+        except json.JSONDecodeError:
+            import re
+
+            m = re.search(r"\{.*\}", raw, re.DOTALL)
+            if m:
+                data = json.loads(m.group())
+            else:
+                logger.warning("recall: LLM returned non-JSON: %.200s", raw)
+                return []
+        selected = data.get("selected_memories", [])
+        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 as exc:
+        logger.warning("recall: memory selection failed (%s), returning []", exc)
+        return []
+
+
+def format_recall_injection(
+    filenames: list[str],
+    memory_dir: Path | None = None,
+) -> str:
+    """Read selected memory files and format for system prompt injection."""
+    mem_dir = memory_dir or global_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
+        blocks.append(f"### {fname}\n\n{content}")
+
+    if not blocks:
+        return ""
+
+    body = "\n\n---\n\n".join(blocks)
+    return f"--- Global Memories ---\n\n{body}\n\n--- End Global Memories ---"

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

@@ -25,9 +25,8 @@
 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.
+15. **Manually wiring browser tools on event_loop nodes** — Browser nodes use tools: {policy: "all"} to get all browser tools.
 
 ## Worker Agent Errors
-17. **Adding client-facing intake node to workers** — The queen owns intake. Workers should start with an autonomous processing node. Client-facing nodes in workers are for mid-execution review/approval only.
-18. **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()`.
+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

@@ -180,7 +180,7 @@ 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_history_tokens": 32000}
+loop_config = {"max_iterations": 100, "max_tool_calls_per_turn": 20, "max_context_tokens": 32000}
 
 
 class MyAgent:
@@ -332,81 +332,46 @@ class MyAgent:
 default_agent = MyAgent()
 ```
 
-## agent.py — Async Entry Points Variant
+## triggers.json — Timer and Webhook Triggers
 
-When an agent needs timers, webhooks, or event-driven triggers, add
-`async_entry_points` and optionally `runtime_config` as module-level variables.
-These are IN ADDITION to the standard variables above.
+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.
 
-```python
-# Additional imports for async entry points
-from framework.graph.edge import GraphSpec, AsyncEntryPointSpec
-from framework.runtime.agent_runtime import (
-    AgentRuntime, AgentRuntimeConfig, create_agent_runtime,
-)
-
-# ... (goal, nodes, edges, entry_node, entry_points, etc. as above) ...
-
-# Async entry points — event-driven triggers
-async_entry_points = [
-    # Timer with cron: daily at 9am
-    AsyncEntryPointSpec(
-        id="daily-check",
-        name="Daily Check",
-        entry_node="process-node",
-        trigger_type="timer",
-        trigger_config={"cron": "0 9 * * *"},
-        isolation_level="shared",
-        max_concurrent=1,
-    ),
-    # Timer with fixed interval: every 20 minutes
-    AsyncEntryPointSpec(
-        id="scheduled-check",
-        name="Scheduled Check",
-        entry_node="process-node",
-        trigger_type="timer",
-        trigger_config={"interval_minutes": 20, "run_immediately": False},
-        isolation_level="shared",
-        max_concurrent=1,
-    ),
-    # Event: reacts to webhook events
-    AsyncEntryPointSpec(
-        id="webhook-event",
-        name="Webhook Event Handler",
-        entry_node="process-node",
-        trigger_type="event",
-        trigger_config={"event_types": ["webhook_received"]},
-        isolation_level="shared",
-        max_concurrent=10,
-    ),
+```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"
+  }
 ]
-
-# Webhook server config (only needed if using webhooks)
-runtime_config = AgentRuntimeConfig(
-    webhook_host="127.0.0.1",
-    webhook_port=8080,
-    webhook_routes=[
-        {
-            "source_id": "my-source",
-            "path": "/webhooks/my-source",
-            "methods": ["POST"],
-        },
-    ],
-)
 ```
 
-**Key rules for async entry points:**
-- `async_entry_points` is a list of `AsyncEntryPointSpec` (NOT `EntryPointSpec`)
-- `runtime_config` is `AgentRuntimeConfig` (NOT `RuntimeConfig` from config.py)
-- Valid trigger_types: `timer`, `event`, `webhook`, `manual`, `api`
-- Valid isolation_levels: `isolated`, `shared`, `synchronized`
+**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, "run_immediately": bool}`
-- Event trigger_config: `{"event_types": ["webhook_received"], "filter_stream": "...", "filter_node": "..."}`
-- Use `isolation_level="shared"` for async entry points that need to read
-  the primary session's memory (e.g., user-configured rules)
-- The `_build_graph()` method passes `async_entry_points` to GraphSpec
-- Reference: `exports/gmail_inbox_guardian/agent.py`
+- 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
 
@@ -453,21 +418,6 @@ __all__ = [
 ]
 ```
 
-**If the agent uses async entry points**, also import and export:
-```python
-from .agent import (
-    ...,
-    async_entry_points,
-    runtime_config,  # Only if using webhooks
-)
-
-__all__ = [
-    ...,
-    "async_entry_points",
-    "runtime_config",
-]
-```
-
 ## __main__.py
 
 ```python

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

@@ -0,0 +1,227 @@
+# Declarative Agent File Templates
+
+Agents are defined as a single `agent.yaml` file. No Python code needed.
+The runner loads this file directly -- no `agent.py`, `config.py`, or
+`nodes/__init__.py` required.
+
+## agent.yaml -- Complete Agent Definition
+
+```yaml
+name: my-agent
+version: 1.0.0
+description: What this agent does.
+
+metadata:
+  intro_message: Welcome! What would you like me to do?
+
+# Template variables -- substituted into system_prompt and identity_prompt
+# via {{variable_name}} syntax.  Use this for config values that appear
+# in prompts (spreadsheet IDs, API endpoints, account names, etc.)
+variables:
+  spreadsheet_id: "1ZVxWDL..."
+  sheet_name: "contacts"
+
+goal:
+  description: What this agent achieves.
+  success_criteria:
+    - "First success criterion"
+    - "Second success criterion"
+  constraints:
+    - "Hard constraint the agent must respect"
+
+identity_prompt: |
+  You are a helpful agent.
+
+conversation_mode: continuous   # always "continuous" for Hive agents
+
+loop_config:
+  max_iterations: 100
+  max_tool_calls_per_turn: 30
+  max_context_tokens: 32000
+
+# MCP servers to connect (resolved by name from ~/.hive/mcp_registry/)
+mcp_servers:
+  - name: hive-tools
+  - name: gcu-tools
+
+nodes:
+  # 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.
+  - id: process
+    name: Process
+    description: Execute the task using available tools
+    max_node_visits: 0   # 0 = unlimited (forever-alive agents)
+    input_keys: [user_request, feedback]
+    output_keys: [results]
+    nullable_output_keys: [feedback]
+    tools:
+      policy: explicit
+      allowed: [web_search, web_scrape, save_data, load_data, list_data_files]
+    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.
+
+      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")
+
+  # Node 2: Handoff (autonomous)
+  - id: handoff
+    name: Handoff
+    description: Prepare worker results for queen review
+    max_node_visits: 0
+    input_keys: [results, user_request]
+    output_keys: [next_action, feedback, worker_summary]
+    nullable_output_keys: [feedback, worker_summary]
+    tools:
+      policy: none   # handoff nodes don't need tools
+    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, call escalate(reason, context) then set:
+      - set_output("next_action", "escalated")
+      - set_output("feedback", "what help is needed")
+
+      Otherwise summarize and set:
+      - set_output("worker_summary", "short summary for queen")
+      - set_output("next_action", "done") or "revise"
+      - set_output("feedback", "what to revise") only when revising
+
+edges:
+  - from_node: process
+    to_node: handoff
+  # Feedback loop
+  - from_node: handoff
+    to_node: process
+    condition: conditional
+    condition_expr: "str(next_action).lower() == 'revise'"
+    priority: 2
+  # Escalation loop
+  - from_node: handoff
+    to_node: process
+    condition: conditional
+    condition_expr: "str(next_action).lower() == 'escalated'"
+    priority: 3
+  # Loop back for next task
+  - from_node: handoff
+    to_node: process
+    condition: conditional
+    condition_expr: "str(next_action).lower() == 'done'"
+
+entry_node: process
+terminal_nodes: []   # [] = forever-alive
+```
+
+## Key differences from Python templates
+
+| Before (Python)                     | After (YAML)                           |
+|-------------------------------------|----------------------------------------|
+| `agent.py` (250 lines boilerplate)  | Not needed                             |
+| `config.py` (dataclass + metadata)  | `variables:` + `metadata:` in YAML     |
+| `nodes/__init__.py` (NodeSpec calls)| `nodes:` list in YAML                  |
+| `__init__.py`, `__main__.py`        | Not needed                             |
+| f-string config injection           | `{{variable_name}}` templates          |
+| `mcp_servers.json` (separate file)  | `mcp_servers:` in YAML (or keep file)  |
+
+## Node types
+
+| Type         | Description                           | Tools                    |
+|--------------|---------------------------------------|--------------------------|
+| `event_loop` | LLM-driven orchestration (default)    | Explicit list or `none`  |
+| `gcu`        | Browser automation via GCU tools      | `policy: all` (auto)     |
+
+## Tool access policies
+
+```yaml
+# Explicit list (recommended for most nodes)
+tools:
+  policy: explicit
+  allowed: [web_search, save_data]
+
+# All tools (for browser automation nodes)
+tools:
+  policy: all
+
+# No tools (for handoff/summary nodes)
+tools:
+  policy: none
+```
+
+## Edge conditions
+
+| Condition     | When to use                                           |
+|---------------|-------------------------------------------------------|
+| `on_success`  | Default. Next node after current succeeds.            |
+| `on_failure`  | Fallback path when current node fails.                |
+| `always`      | Always traverse regardless of outcome.                |
+| `conditional` | Evaluate `condition_expr` against shared memory keys. |
+| `llm_decide`  | Let the LLM decide at runtime.                        |
+
+## Template variables
+
+Use `{{variable_name}}` in `system_prompt` and `identity_prompt`.
+Variables are defined in the top-level `variables:` map.
+
+```yaml
+variables:
+  spreadsheet_id: "1ZVxWDL..."
+  api_endpoint: "https://api.example.com"
+
+nodes:
+  - id: start
+    system_prompt: |
+      Connect to spreadsheet: {{spreadsheet_id}}
+      API endpoint: {{api_endpoint}}
+```
+
+## Entry points
+
+Default is a single manual entry point. For timer/scheduled triggers:
+
+```yaml
+entry_points:
+  - id: default
+    trigger_type: manual
+  - id: daily-check
+    trigger_type: timer
+    trigger_config:
+      interval_minutes: 30
+```
+
+## mcp_servers.json -- Still Supported
+
+The `mcp_servers.json` file is still loaded automatically if present alongside
+`agent.yaml`.  You can also inline servers in the YAML:
+
+```yaml
+mcp_servers:
+  - name: hive-tools
+  - name: gcu-tools
+```
+
+Both approaches work. The JSON file takes precedence for backward compatibility.
+
+## Migration from Python agents
+
+Run the migration tool to convert existing agents:
+
+```bash
+uv run python -m framework.tools.migrate_agent exports/my_agent
+```
+
+This generates `agent.yaml` from the existing `agent.py` + `nodes/` + `config.py`.
+The original files are left untouched. Once verified, you can delete the Python files.
+
+## Files after migration
+
+```
+my_agent/
+  agent.yaml           # The only required file
+  mcp_servers.json     # Optional (can inline in YAML)
+  flowchart.json       # Optional (auto-generated)
+```

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

@@ -1,322 +1,193 @@
-# Hive Agent Framework — Condensed Reference
+# Hive Agent Framework -- Condensed Reference
 
 ## Architecture
 
-Agents are Python packages in `exports/`:
+Agents are declarative JSON configs 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.json          # The entire agent definition
+  mcp_servers.json    # MCP tool server config (optional, prefer registry refs)
 ```
 
-## Agent Loading Contract
+No Python files. No `__init__.py`, `__main__.py`, `config.py`, or `nodes/`.
 
-`AgentRunner.load()` imports the package (`__init__.py`) and reads these
-module-level variables via `getattr()`:
+## Agent Loading
 
-| 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 |
-| `async_entry_points` | no | `[]` | No async triggers (timers, webhooks, events) |
-| `runtime_config` | no | `None` | No webhook server |
+`AgentLoader.load()` reads `agent.json` and builds the execution graph.
+If `agent.py` exists (legacy), it's loaded as a Python module instead.
 
-**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.
+## agent.json Schema
 
-**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"),
-    ],
-)
+```json
+{
+  "name": "my-agent",
+  "version": "1.0.0",
+  "description": "What this agent does",
+  "goal": {
+    "description": "What to achieve",
+    "success_criteria": ["criterion 1", "criterion 2"],
+    "constraints": ["constraint 1"]
+  },
+  "identity_prompt": "You are a helpful agent.",
+  "conversation_mode": "continuous",
+  "loop_config": {
+    "max_iterations": 100,
+    "max_tool_calls_per_turn": 30,
+    "max_context_tokens": 32000
+  },
+  "mcp_servers": [
+    {"name": "hive-tools"},
+    {"name": "gcu-tools"}
+  ],
+  "variables": {
+    "spreadsheet_id": "1ZVx..."
+  },
+  "nodes": [...],
+  "edges": [...],
+  "entry_node": "process",
+  "terminal_nodes": []
+}
 ```
-- 3-5 success criteria, weights sum to 1.0
-- 1-5 constraints (hard/soft, categories: quality, accuracy, interaction, functional)
 
-## NodeSpec Fields
+## Template Variables
+
+Use `{{variable_name}}` in `system_prompt` and `identity_prompt`. Variables
+are defined in the top-level `variables` object:
+
+```json
+{
+  "variables": {"sheet_id": "1ZVx..."},
+  "nodes": [{
+    "id": "start",
+    "system_prompt": "Use sheet: {{sheet_id}}"
+  }]
+}
+```
+
+## Node Fields
 
 | Field | Type | Default | Description |
 |-------|------|---------|-------------|
 | id | str | required | kebab-case identifier |
-| name | str | required | Display name |
+| name | str | id | 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 |
+| node_type | str | "event_loop" | `"event_loop"` |
+| input_keys | list | [] | Memory keys this node reads |
+| output_keys | list | [] | 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 | If True, streams to user and blocks for input |
-| 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 |
+| tools | object | {} | Tool access policy (see below) |
+| nullable_output_keys | list | [] | Keys that may remain unset |
+| max_node_visits | int | 1 | 0=unlimited (for forever-alive agents) |
 | success_criteria | str | "" | Natural language for judge evaluation |
+| client_facing | bool | false | Whether output is shown to user |
 
-## EdgeSpec Fields
+## Tool Access Policies
+
+Each node declares its tools via a policy object:
+
+```json
+{"tools": {"policy": "explicit", "allowed": ["web_search", "save_data"]}}
+{"tools": {"policy": "all"}}
+{"tools": {"policy": "none"}}
+```
+
+- `explicit` (default): only named tools. Empty `allowed` = zero tools.
+- `all`: all tools from registry (e.g. for browser automation nodes).
+- `none`: no tools (for handoff/summary nodes).
+
+## Edge 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) |
+| from_node | str | Source node ID |
+| to_node | str | Target node ID |
+| condition | str | `on_success`, `on_failure`, `always`, `conditional` |
+| condition_expr | str | Python expression for conditional routing |
+| priority | int | Higher = evaluated first |
+
+condition_expr examples:
+- `"needs_more_research == True"`
+- `"str(next_action).lower() == 'revise'"`
 
 ## 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.
+**Hard limit: 3-6 nodes for most agents.** Each node boundary serializes
+outputs and destroys in-context information. Merge unless:
+1. Client-facing boundary (different interaction models)
+2. Disjoint tool sets
+3. Parallel execution (fan-out branches)
 
-Each node boundary serializes outputs to shared memory 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):**
+**Typical structure (2 nodes):**
 ```
-process (autonomous) ←→ review (client-facing)
-```
-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. Client-facing nodes are for mid-execution review/approval only.
-
-For simpler agents, just 1 autonomous node:
-```
-process (autonomous) — loops back to itself
+process (autonomous) <-> review (queen-mediated)
 ```
 
-### 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.
+The queen owns intake. Worker agents should NOT have a client-facing intake
+node. Mid-execution review should happen through queen escalation.
 
 ### set_output
 - Synthetic tool injected by framework
 - Call separately from real tool calls (separate turn)
-- `set_output("key", "value")` stores to shared memory
+- `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
+### Graph Lifecycle
 
 | Pattern | terminal_nodes | When |
 |---------|---------------|------|
-| **Continuous loop** | `["node-with-output-keys"]` | **DEFAULT for all agents** |
+| 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.
+Every graph must have at least one terminal node.
 
-## Continuous Conversation Mode
+### 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)
+- `"continuous"` -- recommended (context carries across node transitions)
+- Omit entirely -- isolated per-node conversations
 
-**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
+**INVALID values:** `"client_facing"`, `"interactive"`, `"shared"`.
 
 ## 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_history_tokens": 32000,    # Triggers conversation compaction
+```json
+{
+  "max_iterations": 100,
+  "max_tool_calls_per_turn": 20,
+  "max_context_tokens": 32000
 }
 ```
-**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
+- `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.
+`data_dir` is auto-injected by framework.
 
 ## 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 shared memory
+Multiple `on_success` edges from same source = parallel execution.
+Parallel nodes must have disjoint output_keys.
 
 ## 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.
-
-## Async Entry Points (Webhooks, Timers, Events)
-
-For agents that react to external events, use `AsyncEntryPointSpec`:
-
-```python
-from framework.graph.edge import AsyncEntryPointSpec
-from framework.runtime.agent_runtime import AgentRuntimeConfig
-
-# Timer trigger (cron or interval)
-async_entry_points = [
-    AsyncEntryPointSpec(
-        id="daily-check",
-        name="Daily Check",
-        entry_node="process",
-        trigger_type="timer",
-        trigger_config={"cron": "0 9 * * *"},  # daily at 9am
-        isolation_level="shared",
-    )
-]
-
-# Webhook server (optional)
-runtime_config = AgentRuntimeConfig(
-    webhook_host="127.0.0.1",
-    webhook_port=8080,
-    webhook_routes=[{"source_id": "gmail", "path": "/webhooks/gmail", "methods": ["POST"]}],
-)
-```
-
-### Key Fields
-- `trigger_type`: `"timer"`, `"event"`, `"webhook"`, `"manual"`
-- `trigger_config`: `{"cron": "0 9 * * *"}` or `{"interval_minutes": 20}`
-- `isolation_level`: `"shared"` (recommended), `"isolated"`, `"synchronized"`
-- `event_types`: For event triggers, e.g., `["webhook_received"]`
-
-### Exports Required
-Both `async_entry_points` and `runtime_config` must be exported from `__init__.py`.
-
-See `exports/gmail_inbox_guardian/agent.py` for complete example.
 
 ## 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.
+Always call `list_agent_tools()` first to see available tools.
+Do NOT rely on a static tool 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
+list_agent_tools()                                      # full summary
+list_agent_tools(group="gmail", output_schema="full")   # drill into category
 ```
 
-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
+After building, run `validate_agent_package("{name}")` to check everything.

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

@@ -1,119 +1,53 @@
-# GCU Browser Automation Guide
+# Browser Automation Guide
 
-## When to Use GCU Nodes
+## When to Use Browser 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
+Use browser nodes (with `tools: {policy: "all"}`) when:
+- The task requires interacting with web pages (clicking, typing, navigating)
+- No API is available for the target service
+- The user is already logged in to the target site
 
-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 Browser Nodes Are
 
-## What GCU Nodes Are
+- Regular `event_loop` nodes with browser tools from gcu-tools MCP server
+- Set `tools: {policy: "all"}` to give access to all browser tools
+- Wire into the graph with edges like any other node
+- No special node_type needed
 
-- `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
+## Available Browser Tools
 
-## GCU Architecture Pattern  
+All tools are prefixed with `browser_`:
+- `browser_start`, `browser_open` -- launch/navigate
+- `browser_click`, `browser_fill`, `browser_type` -- interact
+- `browser_snapshot` -- read page content (preferred over screenshot)
+- `browser_screenshot` -- visual capture
+- `browser_scroll`, `browser_wait` -- navigation helpers
+- `browser_evaluate` -- run JavaScript
 
-GCU nodes are **subagents** — invoked via `delegate_to_sub_agent()`, not connected via edges.
+## System Prompt Tips for Browser Nodes
 
-- 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]
-""",
-)
+```
+1. Use browser_snapshot() to read page content (NOT browser_get_text)
+2. Use browser_wait(seconds=2-3) after navigation for page load
+3. If you hit an auth wall, call set_output with an error and move on
+4. Keep tool calls per turn <= 10 for reliability
 ```
 
-## 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
+## Example
 
 ```json
 {
-  "hive-tools": { ... },
-  "gcu-tools": {
-    "transport": "stdio",
-    "command": "uv",
-    "args": ["run", "python", "-m", "gcu.server", "--stdio"],
-    "cwd": "../../tools",
-    "description": "GCU tools for browser automation"
-  }
+  "id": "scan-profiles",
+  "name": "Scan LinkedIn Profiles",
+  "description": "Navigate LinkedIn search results and collect profile data",
+  "tools": {"policy": "all"},
+  "input_keys": ["search_url"],
+  "output_keys": ["profiles"],
+  "system_prompt": "Navigate to the search URL, paginate through results..."
 }
 ```
 
-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
-
-## GCU Anti-Patterns
-
-- Using `browser_screenshot` to read text (use `browser_snapshot`)
-- 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)
+Connected via regular edges:
+```
+search-setup -> scan-profiles -> process-results
+```

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

@@ -1,63 +0,0 @@
-# Queen Memory — File System Structure
-
-```
-~/.hive/
-├── queen/
-│   ├── MEMORY.md                          ← Semantic memory
-│   ├── memories/
-│   │   ├── MEMORY-2026-03-09.md           ← Episodic memory (today)
-│   │   ├── MEMORY-2026-03-08.md
-│   │   └── ...
-│   └── session/
-│       └── {session_id}/                  ← One dir per session (or resumed-from session)
-│           ├── conversations/
-│           │   ├── parts/
-│           │   │   ├── 00001.json         ← One file per message (role, content, tool_calls)
-│           │   │   ├── 00002.json
-│           │   │   └── ...
-│           │   └── spillover/
-│           │       ├── conversation_1.md  ← Compacted old conversation segments
-│           │       ├── conversation_2.md
-│           │       └── ...
-│           └── data/
-│               ├── adapt.md              ← Working memory (session-scoped)
-│               ├── web_search_1.txt      ← Spillover: large tool results
-│               ├── web_search_2.txt
-│               └── ...
-```
-
----
-
-## The three memory tiers
-
-| File | Tier | Written by | Read at |
-|---|---|---|---|
-| `MEMORY.md` | Semantic | Consolidation LLM (auto, post-session) | Session start (injected into system prompt) |
-| `memories/MEMORY-YYYY-MM-DD.md` | Episodic | Queen via `write_to_diary` tool + consolidation LLM | Session start (today's file injected) |
-| `data/adapt.md` | Working | Queen via `update_session_notes` tool | Every turn (inlined in system prompt) |
-
----
-
-## Session directory naming
-
-The session directory name is **`queen_resume_from`** when a cold-restore resumes an existing
-session, otherwise the new **`session_id`**. This means resumed sessions accumulate all messages
-in the original directory rather than fragmenting across multiple folders.
-
----
-
-## Consolidation
-
-`consolidate_queen_memory()` runs every **5 minutes** in the background and once more at session
-end. It reads:
-
-1. `conversations/parts/*.json` — full message history (user + assistant turns; tool results skipped)
-2. `data/adapt.md` — current working notes
-
-It then makes two LLM writes:
-
-- Rewrites `MEMORY.md` in place (semantic memory — queen never touches this herself)
-- Appends a timestamped prose entry to today's `memories/MEMORY-YYYY-MM-DD.md`
-
-If the combined transcript exceeds ~200 K characters it is recursively binary-compacted via the
-LLM before being sent to the consolidation model (mirrors `EventLoopNode._llm_compact`).

core/framework/agents/queen/reflection_agent.py

@@ -0,0 +1,594 @@
+"""Reflection agent — background global memory extraction for the queen.
+
+A lightweight side agent that runs after each queen LLM turn.  It inspects
+recent conversation messages and extracts durable user knowledge into
+individual memory files in ``~/.hive/memories/global/``.
+
+Two reflection types:
+  - **Short reflection**: after conversational queen turns.  Distills
+    learnings about the user (profile, preferences, environment, feedback).
+  - **Long reflection**: every 5 short reflections and on CONTEXT_COMPACTED.
+    Organises, deduplicates, trims the global memory directory.
+
+Concurrency: an ``asyncio.Lock`` prevents overlapping runs.  If a trigger
+fires while a reflection is already active the event is skipped.
+
+All reflections are fire-and-forget (spawned via ``asyncio.create_task``)
+so they never block the queen's event loop.
+"""
+
+from __future__ import annotations
+
+import asyncio
+import json
+import logging
+import traceback
+from datetime import datetime
+from pathlib import Path
+from typing import Any
+
+from framework.agents.queen.queen_memory_v2 import (
+    GLOBAL_MEMORY_CATEGORIES,
+    MAX_FILE_SIZE_BYTES,
+    MAX_FILES,
+    format_memory_manifest,
+    global_memory_dir,
+    parse_frontmatter,
+    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, 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
+
+
+def _execute_tool(name: str, args: dict[str, Any], memory_dir: Path) -> 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 global memory type restrictions.
+        fm = parse_frontmatter(content)
+        mem_type = (fm.get("type") or "").strip().lower()
+        if mem_type and mem_type not in GLOBAL_MEMORY_CATEGORIES:
+            return (
+                f"ERROR: Invalid memory type '{mem_type}'. "
+                f"Allowed types: {', '.join(GLOBAL_MEMORY_CATEGORIES)}."
+            )
+        # 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 (%d chars)", 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", 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,
+    max_turns: int = _MAX_TURNS,
+) -> tuple[bool, list[str], str]:
+    """Run a mini tool-use loop: LLM → tool calls → repeat.
+
+    Returns (success, changed_files, last_text).
+    """
+    messages: list[dict[str, Any]] = [{"role": "user", "content": user_msg}]
+    changed_files: list[str] = []
+    last_text: str = ""
+
+    for _turn in range(max_turns):
+        logger.info("reflect: loop turn %d/%d (msgs=%d)", _turn + 1, max_turns, len(messages))
+        try:
+            resp: LLMResponse = await llm.acomplete(
+                messages=messages,
+                system=system,
+                tools=_REFLECTION_TOOLS,
+                max_tokens=2048,
+            )
+        except asyncio.CancelledError:
+            logger.warning("reflect: LLM call cancelled (task cancelled)")
+            return False, changed_files, last_text
+        except Exception:
+            logger.warning("reflect: LLM call failed", exc_info=True)
+            return False, changed_files, last_text
+
+        # Extract tool calls from litellm/OpenAI response object.
+        tool_calls_raw: list[dict[str, Any]] = []
+        raw = resp.raw_response
+        if raw is not None:
+            # litellm returns a ModelResponse object; tool calls live on
+            # choices[0].message.tool_calls as a list of ChatCompletionMessageToolCall.
+            try:
+                msg_obj = raw.choices[0].message
+                if hasattr(msg_obj, "tool_calls") and msg_obj.tool_calls:
+                    for tc in msg_obj.tool_calls:
+                        fn = tc.function
+                        try:
+                            args = json.loads(fn.arguments) if fn.arguments else {}
+                        except (json.JSONDecodeError, TypeError):
+                            args = {}
+                        tool_calls_raw.append(
+                            {
+                                "id": tc.id,
+                                "name": fn.name,
+                                "input": args,
+                            }
+                        )
+            except (AttributeError, IndexError):
+                pass
+
+        logger.info(
+            "reflect: LLM responded, text=%d chars, tool_calls=%d",
+            len(resp.content or ""),
+            len(tool_calls_raw),
+        )
+
+        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:
+            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)
+
+        if not tool_calls_raw:
+            break
+
+        for tc in tool_calls_raw:
+            result = _execute_tool(tc["name"], tc.get("input", {}), memory_dir)
+            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
+# ---------------------------------------------------------------------------
+
+_CATEGORIES_STR = ", ".join(GLOBAL_MEMORY_CATEGORIES)
+
+_SHORT_REFLECT_SYSTEM = f"""\
+You are a reflection agent that distills durable knowledge about the USER
+into persistent global memory files.  You run in the background after each
+assistant turn.
+
+Your goal: identify anything from the recent messages worth remembering
+about the user across ALL future sessions — their profile, preferences,
+environment setup, or feedback on assistant behavior.
+
+Memory categories: {_CATEGORIES_STR}
+
+Expected format for each memory file:
+```markdown
+---
+name: {{{{memory name}}}}
+description: {{{{one-line description — specific and search-friendly}}}}
+type: {{{{{_CATEGORIES_STR}}}}}
+---
+
+{{{{memory content}}}}
+```
+
+Workflow (aim for 2 turns):
+  Turn 1 — call list_memory_files to see what exists, then read_memory_file
+            for any that might need updating.
+  Turn 2 — call write_memory_file for new/updated memories.
+
+Rules:
+- ONLY persist durable knowledge about the USER — who they are, how they
+  like to work, their tech environment, their feedback on your behavior.
+- Do NOT store task-specific details, code patterns, file paths, or
+  ephemeral session state.
+- 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, do nothing (respond with a brief
+  reason — no tool calls needed).
+- File names should be kebab-case slugs ending in .md.
+- 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
+global memory directory.  Your job is to organise, deduplicate, and trim
+noise from the accumulated memory files.
+
+Memory categories: {_CATEGORIES_STR}
+
+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.
+"""
+
+
+# ---------------------------------------------------------------------------
+# Short & long reflection entry points
+# ---------------------------------------------------------------------------
+
+
+async def _read_conversation_parts(session_dir: Path) -> list[dict[str, Any]]:
+    """Read conversation parts from the queen session directory."""
+    from framework.storage.conversation_store import FileConversationStore
+
+    store = FileConversationStore(session_dir / "conversations")
+    return await store.read_parts()
+
+
+async def run_short_reflection(
+    session_dir: Path,
+    llm: Any,
+    memory_dir: Path | None = None,
+) -> None:
+    """Run a short reflection: extract user knowledge from conversation."""
+    logger.info("reflect: starting short reflection for %s", session_dir)
+    mem_dir = memory_dir or global_memory_dir()
+
+    messages = await _read_conversation_parts(session_dir)
+    if not messages:
+        logger.info("reflect: no conversation parts found in %s, skipping", session_dir)
+        return
+
+    transcript_lines: list[str] = []
+    for msg in messages[-50:]:
+        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) > 800:
+            content = content[:800] + "…"
+        transcript_lines.append(f"[{label}]: {content}")
+
+    if not transcript_lines:
+        logger.info("reflect: no transcript lines after filtering, skipping")
+        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)
+    if changed:
+        logger.info("reflect: short reflection done, changed files: %s", changed)
+    else:
+        logger.info("reflect: short reflection done, no changes — %s", reason or "no reason")
+
+
+async def run_long_reflection(
+    llm: Any,
+    memory_dir: Path | None = None,
+) -> None:
+    """Run a long reflection: organise and deduplicate all global memories."""
+    logger.debug("reflect: starting long reflection")
+    mem_dir = memory_dir or global_memory_dir()
+    files = scan_memory_files(mem_dir)
+
+    if not files:
+        logger.debug("reflect: no memory files, skipping long reflection")
+        return
+
+    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)
+    if changed:
+        logger.debug("reflect: long reflection done (%d files), changed: %s", len(files), changed)
+    else:
+        logger.debug(
+            "reflect: long reflection done (%d files), no changes — %s",
+            len(files),
+            reason or "no reason",
+        )
+
+
+async def run_shutdown_reflection(
+    session_dir: Path,
+    llm: Any,
+    memory_dir: Path | None = None,
+) -> None:
+    """Run a final short reflection on session shutdown.
+
+    Called during session teardown so recent conversation insights are
+    persisted before the session is destroyed.
+    """
+    logger.info("reflect: running shutdown reflection for %s", session_dir)
+    mem_dir = memory_dir or global_memory_dir()
+    try:
+        await run_short_reflection(session_dir, llm, mem_dir)
+        logger.info("reflect: shutdown reflection completed for %s", session_dir)
+    except asyncio.CancelledError:
+        logger.warning("reflect: shutdown reflection cancelled for %s", session_dir)
+    except Exception:
+        logger.warning("reflect: shutdown reflection failed", exc_info=True)
+        _write_error("shutdown reflection")
+
+
+# ---------------------------------------------------------------------------
+# Event-bus integration
+# ---------------------------------------------------------------------------
+
+_LONG_REFLECT_INTERVAL = 5
+
+
+async def subscribe_reflection_triggers(
+    event_bus: Any,
+    session_dir: Path,
+    llm: Any,
+    memory_dir: Path | None = 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.host.event_bus import EventType
+
+    mem_dir = memory_dir or global_memory_dir()
+    _lock = asyncio.Lock()
+    _short_count = 0
+    _background_tasks: set[asyncio.Task] = set()
+
+    async def _do_turn_reflect(is_interval: bool, count: int) -> None:
+        async with _lock:
+            try:
+                if is_interval:
+                    await run_short_reflection(session_dir, llm, mem_dir)
+                    await run_long_reflection(llm, mem_dir)
+                else:
+                    await run_short_reflection(session_dir, llm, mem_dir)
+            except Exception:
+                logger.warning("reflect: reflection failed", exc_info=True)
+                _write_error("short/long reflection")
+
+    async def _do_compaction_reflect() -> None:
+        async with _lock:
+            try:
+                await run_long_reflection(llm, mem_dir)
+            except Exception:
+                logger.warning("reflect: compaction-triggered reflection failed", exc_info=True)
+                _write_error("compaction reflection")
+
+    def _fire_and_forget(coro: Any) -> None:
+        """Spawn a background task and prevent GC before it finishes."""
+        task = asyncio.create_task(coro)
+        _background_tasks.add(task)
+        task.add_done_callback(_background_tasks.discard)
+
+    async def _on_turn_complete(event: Any) -> None:
+        nonlocal _short_count
+
+        if getattr(event, "stream_id", None) != "queen":
+            return
+
+        _short_count += 1
+
+        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 tool turn (count=%d)", _short_count)
+            return
+
+        if _lock.locked():
+            logger.debug("reflect: skipping, already running (count=%d)", _short_count)
+            return
+
+        logger.debug(
+            "reflect: triggered (count=%d, interval=%s, stop_reason=%s)",
+            _short_count,
+            is_interval,
+            stop_reason,
+        )
+        _fire_and_forget(_do_turn_reflect(is_interval, _short_count))
+
+    async def _on_compaction(event: Any) -> None:
+        if getattr(event, "stream_id", None) != "queen":
+            return
+        if _lock.locked():
+            logger.debug("reflect: skipping compaction trigger, already running")
+            return
+        logger.debug("reflect: compaction triggered long reflection")
+        _fire_and_forget(_do_compaction_reflect())
+
+    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
+
+
+def _write_error(context: str) -> None:
+    """Best-effort write of the last traceback to an error file."""
+    try:
+        error_path = global_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

@@ -22,10 +22,10 @@ def mock_mode():
 
 @pytest_asyncio.fixture(scope="session")
 async def runner(tmp_path_factory, mock_mode):
-    from framework.runner.runner import AgentRunner
+    from framework.loader.agent_loader import AgentLoader
 
     storage = tmp_path_factory.mktemp("agent_storage")
-    r = AgentRunner.load(AGENT_PATH, mock_mode=mock_mode, storage_path=storage)
+    r = AgentLoader.load(AGENT_PATH, mock_mode=mock_mode, storage_path=storage)
     r._setup()
     yield r
     await r.cleanup_async()

core/framework/agents/queen/ticket_receiver.py

@@ -1,27 +0,0 @@
-"""Queen's ticket receiver entry point.
-
-When the Worker Health Judge emits a WORKER_ESCALATION_TICKET event on the
-shared EventBus, this entry point fires and routes to the ``ticket_triage``
-node, where the Queen deliberates and decides whether to notify the operator.
-
-Isolation level is ``isolated`` — the queen's triage memory is kept separate
-from the worker's shared memory. Each ticket triage runs in its own context.
-"""
-
-from __future__ import annotations
-
-from framework.graph.edge import AsyncEntryPointSpec
-
-TICKET_RECEIVER_ENTRY_POINT = AsyncEntryPointSpec(
-    id="ticket_receiver",
-    name="Worker Escalation Ticket Receiver",
-    entry_node="ticket_triage",
-    trigger_type="event",
-    trigger_config={
-        "event_types": ["worker_escalation_ticket"],
-        # Do not fire on our own graph's events (prevents loops if queen
-        # somehow emits a worker_escalation_ticket for herself)
-        "exclude_own_graph": True,
-    },
-    isolation_level="isolated",
-)

core/framework/cli.py

@@ -6,7 +6,6 @@ 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:
@@ -79,8 +78,8 @@ def main():
 
     subparsers = parser.add_subparsers(dest="command", required=True)
 
-    # Register runner commands (run, info, validate, list, dispatch, shell)
-    from framework.runner.cli import register_commands
+    # Register runner commands (run, info, validate, list, shell)
+    from framework.loader.cli import register_commands
 
     register_commands(subparsers)
 
@@ -89,6 +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.loader.mcp_registry_cli import register_mcp_commands
+
+    register_mcp_commands(subparsers)
+
     args = parser.parse_args()
 
     if hasattr(args, "func"):

core/framework/config.py

@@ -12,13 +12,51 @@ from dataclasses import dataclass, field
 from pathlib import Path
 from typing import Any
 
-from framework.graph.edge import DEFAULT_MAX_TOKENS
+from framework.orchestrator.edge import DEFAULT_MAX_TOKENS
+
+# ---------------------------------------------------------------------------
+# Hive home directory structure
+# ---------------------------------------------------------------------------
+
+HIVE_HOME = Path.home() / ".hive"
+QUEENS_DIR = HIVE_HOME / "agents" / "queens"
+COLONIES_DIR = HIVE_HOME / "colonies"
+MEMORIES_DIR = HIVE_HOME / "memories"
+
+
+def queen_dir(queen_name: str = "default") -> Path:
+    """Return the storage directory for a named queen agent."""
+    return QUEENS_DIR / queen_name
+
+
+def colony_dir(colony_name: str) -> Path:
+    """Return the directory for a named colony."""
+    return COLONIES_DIR / colony_name
+
+
+def memory_dir(scope: str, name: str | None = None) -> Path:
+    """Return memory dir for a scope.
+
+    Examples::
+
+        memory_dir("global")                  -> ~/.hive/memories/global
+        memory_dir("colonies", "my_agent")    -> ~/.hive/memories/colonies/my_agent
+        memory_dir("agents/queens", "default")-> ~/.hive/memories/agents/queens/default
+        memory_dir("agents", "worker_name")   -> ~/.hive/memories/agents/worker_name
+    """
+    base = MEMORIES_DIR / scope
+    return base / name if name else base
+
 
 # ---------------------------------------------------------------------------
 # Low-level config file access
 # ---------------------------------------------------------------------------
 
-HIVE_CONFIG_FILE = Path.home() / ".hive" / "configuration.json"
+HIVE_CONFIG_FILE = HIVE_HOME / "configuration.json"
+
+# Hive LLM router endpoint (Anthropic-compatible).
+# litellm's Anthropic handler appends /v1/messages, so this is just the base host.
+HIVE_LLM_ENDPOINT = "https://api.adenhq.com"
 logger = logging.getLogger(__name__)
 
 
@@ -38,6 +76,48 @@ def get_hive_config() -> dict[str, Any]:
         return {}
 
 
+# ---------------------------------------------------------------------------
+# Credential store helpers (for BYOK keys)
+# ---------------------------------------------------------------------------
+
+# Provider name → credential store ID mapping
+_PROVIDER_CRED_MAP: dict[str, str] = {
+    "anthropic": "anthropic",
+    "openai": "openai",
+    "gemini": "gemini",
+    "google": "gemini",
+    "minimax": "minimax",
+    "groq": "groq",
+    "cerebras": "cerebras",
+    "openrouter": "openrouter",
+    "mistral": "mistral",
+    "together": "together",
+    "together_ai": "together",
+    "deepseek": "deepseek",
+    "kimi": "kimi",
+    "hive": "hive",
+}
+
+
+def _get_api_key_from_credential_store(provider: str) -> str | None:
+    """Look up a BYOK API key from the encrypted credential store.
+
+    Returns None if no key is found or the credential store is unavailable.
+    """
+    if not os.environ.get("HIVE_CREDENTIAL_KEY"):
+        return None
+    cred_id = _PROVIDER_CRED_MAP.get(provider.lower())
+    if not cred_id:
+        return None
+    try:
+        from framework.credentials import CredentialStore
+
+        store = CredentialStore.with_encrypted_storage()
+        return store.get(cred_id)
+    except Exception:
+        return None
+
+
 # ---------------------------------------------------------------------------
 # Derived helpers
 # ---------------------------------------------------------------------------
@@ -47,31 +127,213 @@ def get_preferred_model() -> str:
     """Return the user's preferred LLM model string (e.g. 'anthropic/claude-sonnet-4-20250514')."""
     llm = get_hive_config().get("llm", {})
     if llm.get("provider") and llm.get("model"):
-        return f"{llm['provider']}/{llm['model']}"
+        provider = str(llm["provider"])
+        model = str(llm["model"]).strip()
+        # OpenRouter quickstart stores raw model IDs; tolerate pasted "openrouter/<id>" too.
+        if provider.lower() == "openrouter" and model.lower().startswith("openrouter/"):
+            model = model[len("openrouter/") :]
+        if model:
+            return f"{provider}/{model}"
     return "anthropic/claude-sonnet-4-20250514"
 
 
+def get_preferred_worker_model() -> str | None:
+    """Return the user's preferred worker LLM model, or None if not configured.
+
+    Reads from the ``worker_llm`` section of ~/.hive/configuration.json.
+    Returns None when no worker-specific model is set, so callers can
+    fall back to the default (queen) model via ``get_preferred_model()``.
+    """
+    worker_llm = get_hive_config().get("worker_llm", {})
+    if worker_llm.get("provider") and worker_llm.get("model"):
+        provider = str(worker_llm["provider"])
+        model = str(worker_llm["model"]).strip()
+        if provider.lower() == "openrouter" and model.lower().startswith("openrouter/"):
+            model = model[len("openrouter/") :]
+        if model:
+            return f"{provider}/{model}"
+    return None
+
+
+def get_worker_api_key() -> str | None:
+    """Return the API key for the worker LLM, falling back to the default key."""
+    worker_llm = get_hive_config().get("worker_llm", {})
+    if not worker_llm:
+        return get_api_key()
+
+    # Worker-specific subscription / env var
+    if worker_llm.get("use_claude_code_subscription"):
+        try:
+            from framework.loader.agent_loader import get_claude_code_token
+
+            token = get_claude_code_token()
+            if token:
+                return token
+        except ImportError:
+            pass
+
+    if worker_llm.get("use_codex_subscription"):
+        try:
+            from framework.loader.agent_loader import get_codex_token
+
+            token = get_codex_token()
+            if token:
+                return token
+        except ImportError:
+            pass
+
+    if worker_llm.get("use_kimi_code_subscription"):
+        try:
+            from framework.loader.agent_loader import get_kimi_code_token
+
+            token = get_kimi_code_token()
+            if token:
+                return token
+        except ImportError:
+            pass
+
+    if worker_llm.get("use_antigravity_subscription"):
+        try:
+            from framework.loader.agent_loader import get_antigravity_token
+
+            token = get_antigravity_token()
+            if token:
+                return token
+        except ImportError:
+            pass
+
+    api_key_env_var = worker_llm.get("api_key_env_var")
+    if api_key_env_var:
+        return os.environ.get(api_key_env_var)
+
+    # Fall back to default key
+    return get_api_key()
+
+
+def get_worker_api_base() -> str | None:
+    """Return the api_base for the worker LLM, falling back to the default."""
+    worker_llm = get_hive_config().get("worker_llm", {})
+    if not worker_llm:
+        return get_api_base()
+
+    if worker_llm.get("use_codex_subscription"):
+        return "https://chatgpt.com/backend-api/codex"
+    if worker_llm.get("use_kimi_code_subscription"):
+        return "https://api.kimi.com/coding"
+    if worker_llm.get("use_antigravity_subscription"):
+        # Antigravity uses AntigravityProvider directly — no api_base needed.
+        return None
+    if worker_llm.get("api_base"):
+        return worker_llm["api_base"]
+    if str(worker_llm.get("provider", "")).lower() == "openrouter":
+        return OPENROUTER_API_BASE
+    return None
+
+
+def get_worker_llm_extra_kwargs() -> dict[str, Any]:
+    """Return extra kwargs for the worker LLM provider."""
+    worker_llm = get_hive_config().get("worker_llm", {})
+    if not worker_llm:
+        return get_llm_extra_kwargs()
+
+    if worker_llm.get("use_claude_code_subscription"):
+        api_key = get_worker_api_key()
+        if api_key:
+            return {
+                "extra_headers": {"authorization": f"Bearer {api_key}"},
+            }
+    if worker_llm.get("use_codex_subscription"):
+        api_key = get_worker_api_key()
+        if api_key:
+            headers: dict[str, str] = {
+                "Authorization": f"Bearer {api_key}",
+                "User-Agent": "CodexBar",
+            }
+            try:
+                from framework.loader.agent_loader import get_codex_account_id
+
+                account_id = get_codex_account_id()
+                if account_id:
+                    headers["ChatGPT-Account-Id"] = account_id
+            except ImportError:
+                pass
+            return {
+                "extra_headers": headers,
+                "store": False,
+                "allowed_openai_params": ["store"],
+            }
+    if worker_llm.get("provider") == "ollama":
+        return {"num_ctx": worker_llm.get("num_ctx", 16384)}
+    return {}
+
+
+def get_worker_max_tokens() -> int:
+    """Return max_tokens for the worker LLM, falling back to default."""
+    worker_llm = get_hive_config().get("worker_llm", {})
+    if worker_llm and "max_tokens" in worker_llm:
+        return worker_llm["max_tokens"]
+    return get_max_tokens()
+
+
+def get_worker_max_context_tokens() -> int:
+    """Return max_context_tokens for the worker LLM, falling back to default."""
+    worker_llm = get_hive_config().get("worker_llm", {})
+    if worker_llm and "max_context_tokens" in worker_llm:
+        return worker_llm["max_context_tokens"]
+    return get_max_context_tokens()
+
+
 def get_max_tokens() -> int:
     """Return the configured max_tokens, falling back to DEFAULT_MAX_TOKENS."""
     return get_hive_config().get("llm", {}).get("max_tokens", DEFAULT_MAX_TOKENS)
 
 
+DEFAULT_MAX_CONTEXT_TOKENS = 32_000
+OPENROUTER_API_BASE = "https://openrouter.ai/api/v1"
+
+
+def get_max_context_tokens() -> int:
+    """Return the configured max_context_tokens, falling back to DEFAULT_MAX_CONTEXT_TOKENS."""
+    return get_hive_config().get("llm", {}).get("max_context_tokens", DEFAULT_MAX_CONTEXT_TOKENS)
+
+
+def get_api_keys() -> list[str] | None:
+    """Return a list of API keys if ``api_keys`` is configured, else ``None``.
+
+    This supports key-pool rotation: configure multiple keys in
+    ``~/.hive/configuration.json`` under ``llm.api_keys`` and the
+    :class:`~framework.llm.key_pool.KeyPool` will rotate through them.
+    """
+    llm = get_hive_config().get("llm", {})
+    keys = llm.get("api_keys")
+    if keys and isinstance(keys, list) and len(keys) > 0:
+        return [k for k in keys if k]  # filter empties
+    return None
+
+
 def get_api_key() -> str | None:
     """Return the API key, supporting env var, Claude Code subscription, Codex, and ZAI Code.
 
     Priority:
+    0. Explicit key pool (``api_keys`` list) -- returns first key for
+       single-key callers; full pool available via :func:`get_api_keys`.
     1. Claude Code subscription (``use_claude_code_subscription: true``)
        reads the OAuth token from ``~/.claude/.credentials.json``.
     2. Codex subscription (``use_codex_subscription: true``)
        reads the OAuth token from macOS Keychain or ``~/.codex/auth.json``.
     3. Environment variable named in ``api_key_env_var``.
     """
+    # If an explicit key pool is configured, use the first key.
+    pool_keys = get_api_keys()
+    if pool_keys:
+        return pool_keys[0]
+
     llm = get_hive_config().get("llm", {})
 
     # Claude Code subscription: read OAuth token directly
     if llm.get("use_claude_code_subscription"):
         try:
-            from framework.runner.runner import get_claude_code_token
+            from framework.loader.agent_loader import get_claude_code_token
 
             token = get_claude_code_token()
             if token:
@@ -82,7 +344,7 @@ def get_api_key() -> str | None:
     # Codex subscription: read OAuth token from Keychain / auth.json
     if llm.get("use_codex_subscription"):
         try:
-            from framework.runner.runner import get_codex_token
+            from framework.loader.agent_loader import get_codex_token
 
             token = get_codex_token()
             if token:
@@ -93,7 +355,7 @@ def get_api_key() -> str | None:
     # Kimi Code subscription: read API key from ~/.kimi/config.toml
     if llm.get("use_kimi_code_subscription"):
         try:
-            from framework.runner.runner import get_kimi_code_token
+            from framework.loader.agent_loader import get_kimi_code_token
 
             token = get_kimi_code_token()
             if token:
@@ -101,11 +363,106 @@ def get_api_key() -> str | None:
         except ImportError:
             pass
 
+    # Antigravity subscription: read OAuth token from accounts JSON
+    if llm.get("use_antigravity_subscription"):
+        try:
+            from framework.loader.agent_loader import get_antigravity_token
+
+            token = get_antigravity_token()
+            if token:
+                return token
+        except ImportError:
+            pass
+
     # Standard env-var path (covers ZAI Code and all API-key providers)
     api_key_env_var = llm.get("api_key_env_var")
     if api_key_env_var:
-        return os.environ.get(api_key_env_var)
-    return None
+        key = os.environ.get(api_key_env_var)
+        if key:
+            return key
+
+    # Credential store fallback — BYOK keys stored via the UI
+    return _get_api_key_from_credential_store(llm.get("provider", ""))
+
+
+# OAuth credentials for Antigravity are 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
+_ANTIGRAVITY_CREDENTIALS_URL = (
+    "https://raw.githubusercontent.com/NoeFabris/opencode-antigravity-auth/dev/src/constants.ts"
+)
+_antigravity_credentials_cache: tuple[str | None, str | None] = (None, None)
+
+
+def _fetch_antigravity_credentials() -> tuple[str | None, str | None]:
+    """Fetch OAuth client ID and secret from the public npm package source on GitHub."""
+    global _antigravity_credentials_cache
+    if _antigravity_credentials_cache[0] and _antigravity_credentials_cache[1]:
+        return _antigravity_credentials_cache
+
+    import re
+    import urllib.request
+
+    try:
+        req = urllib.request.Request(
+            _ANTIGRAVITY_CREDENTIALS_URL, headers={"User-Agent": "Hive/1.0"}
+        )
+        with urllib.request.urlopen(req, timeout=10) as resp:
+            content = resp.read().decode("utf-8")
+            id_match = re.search(r'ANTIGRAVITY_CLIENT_ID\s*=\s*"([^"]+)"', content)
+            secret_match = re.search(r'ANTIGRAVITY_CLIENT_SECRET\s*=\s*"([^"]+)"', content)
+            client_id = id_match.group(1) if id_match else None
+            client_secret = secret_match.group(1) if secret_match else None
+            if client_id and client_secret:
+                _antigravity_credentials_cache = (client_id, client_secret)
+            return client_id, client_secret
+    except Exception as e:
+        logger.debug("Failed to fetch Antigravity credentials from public source: %s", e)
+    return None, None
+
+
+def get_antigravity_client_id() -> str:
+    """Return the Antigravity OAuth application client ID.
+
+    Checked in order:
+    1. ``ANTIGRAVITY_CLIENT_ID`` environment variable
+    2. ``llm.antigravity_client_id`` in ~/.hive/configuration.json
+    3. Fetch from public source (opencode-antigravity-auth project on GitHub)
+    """
+    env = os.environ.get("ANTIGRAVITY_CLIENT_ID")
+    if env:
+        return env
+    cfg_val = get_hive_config().get("llm", {}).get("antigravity_client_id")
+    if cfg_val:
+        return cfg_val
+    # Fetch from public source
+    client_id, _ = _fetch_antigravity_credentials()
+    if client_id:
+        return client_id
+    raise RuntimeError("Could not obtain Antigravity OAuth client ID")
+
+
+def get_antigravity_client_secret() -> str | None:
+    """Return the Antigravity OAuth client secret.
+
+    Checked in order:
+    1. ``ANTIGRAVITY_CLIENT_SECRET`` environment variable
+    2. ``llm.antigravity_client_secret`` in ~/.hive/configuration.json
+    3. Fetch from public source (opencode-antigravity-auth project on GitHub)
+
+    Returns None when not found — token refresh will be skipped and
+    the caller must use whatever access token is already available.
+    """
+    env = os.environ.get("ANTIGRAVITY_CLIENT_SECRET")
+    if env:
+        return env
+    cfg_val = get_hive_config().get("llm", {}).get("antigravity_client_secret") or None
+    if cfg_val:
+        return cfg_val
+    # Fetch from public source
+    _, secret = _fetch_antigravity_credentials()
+    return secret
 
 
 def get_gcu_enabled() -> bool:
@@ -113,6 +470,14 @@ def get_gcu_enabled() -> bool:
     return get_hive_config().get("gcu_enabled", True)
 
 
+def get_gcu_viewport_scale() -> float:
+    """Return GCU viewport scale factor (0.1-1.0), default 0.8."""
+    scale = get_hive_config().get("gcu_viewport_scale", 0.8)
+    if isinstance(scale, (int, float)) and 0.1 <= scale <= 1.0:
+        return float(scale)
+    return 0.8
+
+
 def get_api_base() -> str | None:
     """Return the api_base URL for OpenAI-compatible endpoints, if configured."""
     llm = get_hive_config().get("llm", {})
@@ -122,7 +487,14 @@ def get_api_base() -> str | None:
     if llm.get("use_kimi_code_subscription"):
         # Kimi Code uses an Anthropic-compatible endpoint (no /v1 suffix).
         return "https://api.kimi.com/coding"
-    return llm.get("api_base")
+    if llm.get("use_antigravity_subscription"):
+        # Antigravity uses AntigravityProvider directly — no api_base needed.
+        return None
+    if llm.get("api_base"):
+        return llm["api_base"]
+    if str(llm.get("provider", "")).lower() == "openrouter":
+        return OPENROUTER_API_BASE
+    return None
 
 
 def get_llm_extra_kwargs() -> dict[str, Any]:
@@ -151,7 +523,7 @@ def get_llm_extra_kwargs() -> dict[str, Any]:
                 "User-Agent": "CodexBar",
             }
             try:
-                from framework.runner.runner import get_codex_account_id
+                from framework.loader.agent_loader import get_codex_account_id
 
                 account_id = get_codex_account_id()
                 if account_id:
@@ -163,6 +535,11 @@ def get_llm_extra_kwargs() -> dict[str, Any]:
                 "store": False,
                 "allowed_openai_params": ["store"],
             }
+    if llm.get("provider") == "ollama":
+        # Pass num_ctx to Ollama so it doesn't silently truncate the ~9.5k Queen prompt.
+        # Ollama's default num_ctx is only 2048. We set it to 16384 here so LiteLLM
+        # passes it through as a provider-specific option.
+        return {"num_ctx": llm.get("num_ctx", 16384)}
     return {}
 
 
@@ -178,6 +555,7 @@ class RuntimeConfig:
     model: str = field(default_factory=get_preferred_model)
     temperature: float = 0.7
     max_tokens: int = field(default_factory=get_max_tokens)
+    max_context_tokens: int = field(default_factory=get_max_context_tokens)
     api_key: str | None = field(default_factory=get_api_key)
     api_base: str | None = field(default_factory=get_api_base)
     extra_kwargs: dict[str, Any] = field(default_factory=get_llm_extra_kwargs)

core/framework/credentials/key_storage.py

@@ -142,13 +142,17 @@ def save_aden_api_key(key: str) -> None:
     os.environ[ADEN_ENV_VAR] = key
 
 
-def delete_aden_api_key() -> None:
-    """Remove ADEN_API_KEY from the encrypted store and ``os.environ``."""
+def delete_aden_api_key() -> bool:
+    """Remove ADEN_API_KEY from the encrypted store and ``os.environ``.
+
+    Returns True if the key existed and was deleted, False otherwise.
+    """
+    deleted = False
     try:
         from .storage import EncryptedFileStorage
 
         storage = EncryptedFileStorage()
-        storage.delete(ADEN_CREDENTIAL_ID)
+        deleted = storage.delete(ADEN_CREDENTIAL_ID)
     except (FileNotFoundError, PermissionError) as e:
         logger.debug("Could not delete %s from encrypted store: %s", ADEN_CREDENTIAL_ID, e)
     except Exception:
@@ -157,8 +161,8 @@ def delete_aden_api_key() -> None:
             ADEN_CREDENTIAL_ID,
             exc_info=True,
         )
-
     os.environ.pop(ADEN_ENV_VAR, None)
+    return deleted
 
 
 # ---------------------------------------------------------------------------

core/framework/credentials/setup.py

@@ -27,6 +27,7 @@ from __future__ import annotations
 
 import getpass
 import json
+import logging
 import os
 import sys
 from collections.abc import Callable
@@ -35,7 +36,9 @@ from pathlib import Path
 from typing import TYPE_CHECKING, Any
 
 if TYPE_CHECKING:
-    from framework.graph import NodeSpec
+    from framework.orchestrator import NodeSpec
+
+logger = logging.getLogger(__name__)
 
 
 # ANSI colors for terminal output
@@ -365,8 +368,11 @@ class CredentialSetupSession:
         self._print("")
         try:
             api_key = self.password_fn(f"Paste your {cred.env_var}: ").strip()
+        except (EOFError, OSError) as exc:
+            logger.debug("Password input unavailable, falling back to plain input: %s", exc)
+            api_key = self._input(f"Paste your {cred.env_var}: ").strip()
         except Exception:
-            # Fallback to regular input if password input fails
+            logger.warning("Unexpected error reading password input", exc_info=True)
             api_key = self._input(f"Paste your {cred.env_var}: ").strip()
 
         if not api_key:
@@ -403,7 +409,11 @@ class CredentialSetupSession:
 
             try:
                 aden_key = self.password_fn("Paste your ADEN_API_KEY: ").strip()
+            except (EOFError, OSError) as exc:
+                logger.debug("Password input unavailable for ADEN_API_KEY: %s", exc)
+                aden_key = self._input("Paste your ADEN_API_KEY: ").strip()
             except Exception:
+                logger.warning("Unexpected error reading ADEN_API_KEY input", exc_info=True)
                 aden_key = self._input("Paste your ADEN_API_KEY: ").strip()
 
             if not aden_key:
@@ -433,8 +443,10 @@ class CredentialSetupSession:
                     value = store.get_key(cred_id, cred.credential_key)
                     if value:
                         os.environ[cred.env_var] = value
+                except (KeyError, OSError) as exc:
+                    logger.debug("Could not export credential to env: %s", exc)
                 except Exception:
-                    pass
+                    logger.warning("Unexpected error exporting credential to env", exc_info=True)
                 return True
             else:
                 self._print(
@@ -457,9 +469,12 @@ class CredentialSetupSession:
                 "message": result.message,
                 "details": result.details,
             }
-        except Exception:
+        except ImportError:
             # No health checker available
             return None
+        except Exception:
+            logger.warning("Health check failed for %s", cred.credential_name, exc_info=True)
+            return None
 
     def _store_credential(self, cred: MissingCredential, value: str) -> None:
         """Store credential in encrypted store and export to env."""
@@ -518,7 +533,9 @@ class CredentialSetupSession:
 
 
 def load_agent_nodes(agent_path: str | Path) -> list:
-    """Load NodeSpec list from an agent's agent.py or agent.json.
+    """Load NodeSpec list from an agent directory.
+
+    Checks agent.json (declarative) first, then agent.py (legacy).
 
     Args:
         agent_path: Path to agent directory.
@@ -527,16 +544,28 @@ def load_agent_nodes(agent_path: str | Path) -> list:
         List of NodeSpec objects (empty list if agent can't be loaded).
     """
     agent_path = Path(agent_path)
+    agent_json_file = agent_path / "agent.json"
     agent_py = agent_path / "agent.py"
-    agent_json = agent_path / "agent.json"
 
-    if agent_py.exists():
+    if agent_json_file.exists():
+        return _load_nodes_from_json_declarative(agent_json_file)
+    elif agent_py.exists():
         return _load_nodes_from_python_agent(agent_path)
-    elif agent_json.exists():
-        return _load_nodes_from_json_agent(agent_json)
     return []
 
 
+def _load_nodes_from_json_declarative(agent_json: Path) -> list:
+    """Load nodes from a declarative JSON agent."""
+    try:
+        from framework.loader.agent_loader import load_agent_config
+
+        data = json.loads(agent_json.read_text(encoding="utf-8"))
+        graph, _ = load_agent_config(data)
+        return list(graph.nodes)
+    except Exception:
+        return []
+
+
 def _load_nodes_from_python_agent(agent_path: Path) -> list:
     """Load nodes from a Python-based agent."""
     import importlib.util
@@ -561,7 +590,11 @@ def _load_nodes_from_python_agent(agent_path: Path) -> list:
         sys.modules[spec.name] = module
         spec.loader.exec_module(module)
         return getattr(module, "nodes", [])
+    except (ImportError, OSError) as exc:
+        logger.debug("Could not load agent module: %s", exc)
+        return []
     except Exception:
+        logger.warning("Unexpected error loading agent module", exc_info=True)
         return []
 
 
@@ -571,7 +604,7 @@ def _load_nodes_from_json_agent(agent_json: Path) -> list:
         with open(agent_json, encoding="utf-8-sig") as f:
             data = json.load(f)
 
-        from framework.graph import NodeSpec
+        from framework.orchestrator import NodeSpec
 
         nodes_data = data.get("graph", {}).get("nodes", [])
         nodes = []
@@ -588,7 +621,11 @@ def _load_nodes_from_json_agent(agent_json: Path) -> list:
                 )
             )
         return nodes
+    except (json.JSONDecodeError, KeyError, OSError) as exc:
+        logger.debug("Could not load JSON agent: %s", exc)
+        return []
     except Exception:
+        logger.warning("Unexpected error loading JSON agent", exc_info=True)
         return []
 
 

core/framework/credentials/validation.py

@@ -51,6 +51,16 @@ def ensure_credential_key_env() -> None:
                     if found and value:
                         os.environ[var_name] = value
                         logger.debug("Loaded %s from shell config", var_name)
+        # Also load the currently configured LLM env var even if it's not in CREDENTIAL_SPECS.
+        # This keeps quickstart-written keys available to fresh processes on Unix shells.
+        from framework.config import get_hive_config
+
+        llm_env_var = str(get_hive_config().get("llm", {}).get("api_key_env_var", "")).strip()
+        if llm_env_var and not os.environ.get(llm_env_var):
+            found, value = check_env_var_in_shell_config(llm_env_var)
+            if found and value:
+                os.environ[llm_env_var] = value
+                logger.debug("Loaded configured LLM env var %s from shell config", llm_env_var)
     except ImportError:
         pass
 

core/framework/debugger/cli.py

@@ -0,0 +1,76 @@
+"""CLI command for the LLM debug log viewer."""
+
+import argparse
+import subprocess
+import sys
+from pathlib import Path
+
+_SCRIPT = Path(__file__).resolve().parents[3] / "scripts" / "llm_debug_log_visualizer.py"
+
+
+def register_debugger_commands(subparsers: argparse._SubParsersAction) -> None:
+    """Register the ``hive debugger`` command."""
+    parser = subparsers.add_parser(
+        "debugger",
+        help="Open the LLM debug log viewer",
+        description=(
+            "Start a local server that lets you browse LLM debug sessions "
+            "recorded in ~/.hive/llm_logs. Sessions are loaded on demand so "
+            "the browser stays responsive."
+        ),
+    )
+    parser.add_argument(
+        "--session",
+        help="Execution ID to select initially.",
+    )
+    parser.add_argument(
+        "--port",
+        type=int,
+        default=0,
+        help="Port for the local server (0 = auto-pick a free port).",
+    )
+    parser.add_argument(
+        "--logs-dir",
+        help="Directory containing JSONL log files (default: ~/.hive/llm_logs).",
+    )
+    parser.add_argument(
+        "--limit-files",
+        type=int,
+        default=None,
+        help="Maximum number of newest log files to scan (default: 200).",
+    )
+    parser.add_argument(
+        "--output",
+        help="Write a static HTML file instead of starting a server.",
+    )
+    parser.add_argument(
+        "--no-open",
+        action="store_true",
+        help="Start the server but do not open a browser.",
+    )
+    parser.add_argument(
+        "--include-tests",
+        action="store_true",
+        help="Show test/mock sessions (hidden by default).",
+    )
+    parser.set_defaults(func=cmd_debugger)
+
+
+def cmd_debugger(args: argparse.Namespace) -> int:
+    """Launch the LLM debug log visualizer."""
+    cmd: list[str] = [sys.executable, str(_SCRIPT)]
+    if args.session:
+        cmd += ["--session", args.session]
+    if args.port:
+        cmd += ["--port", str(args.port)]
+    if args.logs_dir:
+        cmd += ["--logs-dir", args.logs_dir]
+    if args.limit_files is not None:
+        cmd += ["--limit-files", str(args.limit_files)]
+    if args.output:
+        cmd += ["--output", args.output]
+    if args.no_open:
+        cmd.append("--no-open")
+    if args.include_tests:
+        cmd.append("--include-tests")
+    return subprocess.call(cmd)

core/framework/graph/__init__.py

@@ -1,59 +0,0 @@
-"""Graph structures: Goals, Nodes, Edges, and Execution."""
-
-from framework.graph.client_io import (
-    ActiveNodeClientIO,
-    ClientIOGateway,
-    InertNodeClientIO,
-    NodeClientIO,
-)
-from framework.graph.context_handoff import ContextHandoff, HandoffContext
-from framework.graph.conversation import ConversationStore, Message, NodeConversation
-from framework.graph.edge import DEFAULT_MAX_TOKENS, EdgeCondition, EdgeSpec, GraphSpec
-from framework.graph.event_loop_node import (
-    EventLoopNode,
-    JudgeProtocol,
-    JudgeVerdict,
-    LoopConfig,
-    OutputAccumulator,
-)
-from framework.graph.executor import GraphExecutor
-from framework.graph.goal import Constraint, Goal, GoalStatus, SuccessCriterion
-from framework.graph.node import NodeContext, NodeProtocol, NodeResult, NodeSpec
-
-__all__ = [
-    # Goal
-    "Goal",
-    "SuccessCriterion",
-    "Constraint",
-    "GoalStatus",
-    # Node
-    "NodeSpec",
-    "NodeContext",
-    "NodeResult",
-    "NodeProtocol",
-    # Edge
-    "EdgeSpec",
-    "EdgeCondition",
-    "GraphSpec",
-    "DEFAULT_MAX_TOKENS",
-    # Executor
-    "GraphExecutor",
-    # Conversation
-    "NodeConversation",
-    "ConversationStore",
-    "Message",
-    # Event Loop
-    "EventLoopNode",
-    "LoopConfig",
-    "OutputAccumulator",
-    "JudgeProtocol",
-    "JudgeVerdict",
-    # Context Handoff
-    "ContextHandoff",
-    "HandoffContext",
-    # Client I/O
-    "NodeClientIO",
-    "ActiveNodeClientIO",
-    "InertNodeClientIO",
-    "ClientIOGateway",
-]

core/framework/graph/gcu.py

@@ -1,86 +0,0 @@
-"""GCU (browser automation) node type constants.
-
-A ``gcu`` node is an ``event_loop`` node with two automatic enhancements:
-1. A canonical browser best-practices system prompt is prepended.
-2. All tools from the GCU MCP server are auto-included.
-
-No new ``NodeProtocol`` subclass — the ``gcu`` type is purely a declarative
-signal processed by the runner and executor at setup time.
-"""
-
-# ---------------------------------------------------------------------------
-# MCP server identity
-# ---------------------------------------------------------------------------
-
-GCU_SERVER_NAME = "gcu-tools"
-"""Name used to identify the GCU MCP server in ``mcp_servers.json``."""
-
-GCU_MCP_SERVER_CONFIG: dict = {
-    "name": GCU_SERVER_NAME,
-    "transport": "stdio",
-    "command": "uv",
-    "args": ["run", "python", "-m", "gcu.server", "--stdio"],
-    "cwd": "../../tools",
-    "description": "GCU tools for browser automation",
-}
-"""Default stdio config for the GCU MCP server (relative to exports/<agent>/)."""
-
-# ---------------------------------------------------------------------------
-# Browser best-practices system prompt
-# ---------------------------------------------------------------------------
-
-GCU_BROWSER_SYSTEM_PROMPT = """\
-# Browser Automation Best Practices
-
-Follow these rules for reliable, efficient browser interaction.
-
-## Reading Pages
-- ALWAYS prefer `browser_snapshot` over `browser_get_text("body")`
-  — it returns a compact ~1-5 KB accessibility tree vs 100+ KB of raw HTML.
-- Use `browser_snapshot_aria` when you need full ARIA properties
-  for detailed element inspection.
-- Do NOT use `browser_screenshot` for reading text content
-  — it produces huge base64 images with no searchable text.
-- Only fall back to `browser_get_text` for extracting specific
-  small elements by CSS selector.
-
-## Navigation & Waiting
-- Always call `browser_wait` after navigation actions
-  (`browser_open`, `browser_navigate`, `browser_click` on links)
-  to let the page load.
-- NEVER re-navigate to the same URL after scrolling
-  — this resets your scroll position and loses loaded content.
-
-## Scrolling
-- Use large scroll amounts ~2000 when loading more content
-  — sites like twitter and linkedin have lazy loading for paging.
-- After scrolling, take a new `browser_snapshot` to see updated content.
-
-## Error Recovery
-- If a tool fails, retry once with the same approach.
-- If it fails a second time, STOP retrying and switch approach.
-- If `browser_snapshot` fails → try `browser_get_text` with a
-  specific small selector as fallback.
-- If `browser_open` fails or page seems stale → `browser_stop`,
-  then `browser_start`, then retry.
-
-## Tab Management
-- Use `browser_tabs` to list open tabs when managing multiple pages.
-- Pass `target_id` to tools when operating on a specific tab.
-- Open background tabs with `browser_open(url=..., background=true)`
-  to avoid losing your current context.
-- Close tabs you no longer need with `browser_close` to free resources.
-
-## Login & Auth Walls
-- If you see a "Log in" or "Sign up" prompt instead of expected
-  content, report the auth wall immediately — do NOT attempt to log in.
-- Check for cookie consent banners and dismiss them if they block content.
-
-## Efficiency
-- Minimize tool calls — combine actions where possible.
-- When a snapshot result is saved to a spillover file, use
-  `run_command` with grep to extract specific data rather than
-  re-reading the full file.
-- Call `set_output` in the same turn as your last browser action
-  when possible — don't waste a turn.
-"""

core/framework/graph/prompt_composer.py

@@ -1,306 +0,0 @@
-"""Prompt composition for continuous agent mode.
-
-Composes the three-layer system prompt (onion model) and generates
-transition markers inserted into the conversation at phase boundaries.
-
-Layer 1 — Identity (static, defined at agent level, never changes):
-  "You are a thorough research agent. You prefer clarity over jargon..."
-
-Layer 2 — Narrative (auto-generated from conversation/memory state):
-  "We've finished scoping the project. The user wants to focus on..."
-
-Layer 3 — Focus (per-node system_prompt, reframed as focus directive):
-  "Your current attention: synthesize findings into a report..."
-"""
-
-from __future__ import annotations
-
-import logging
-from datetime import datetime
-from pathlib import Path
-from typing import TYPE_CHECKING, Any
-
-if TYPE_CHECKING:
-    from framework.graph.edge import GraphSpec
-    from framework.graph.node import NodeSpec, SharedMemory
-
-logger = logging.getLogger(__name__)
-
-
-def _with_datetime(prompt: str) -> str:
-    """Append current datetime with local timezone to a system prompt."""
-    local = datetime.now().astimezone()
-    stamp = f"Current date and time: {local.strftime('%Y-%m-%d %H:%M %Z (UTC%z)')}"
-    return f"{prompt}\n\n{stamp}" if prompt else stamp
-
-
-def build_accounts_prompt(
-    accounts: list[dict[str, Any]],
-    tool_provider_map: dict[str, str] | None = None,
-    node_tool_names: list[str] | None = None,
-) -> str:
-    """Build a prompt section describing connected accounts.
-
-    When tool_provider_map is provided, produces structured output grouped
-    by provider with tool mapping, so the LLM knows which ``account`` value
-    to pass to which tool.
-
-    When node_tool_names is also provided, filters to only show providers
-    whose tools overlap with the node's tool list.
-
-    Args:
-        accounts: List of account info dicts from
-            CredentialStoreAdapter.get_all_account_info().
-        tool_provider_map: Mapping of tool_name -> provider_name
-            (e.g. {"gmail_list_messages": "google"}).
-        node_tool_names: Tool names available to the current node.
-            When provided, only providers with matching tools are shown.
-
-    Returns:
-        Formatted accounts block, or empty string if no accounts.
-    """
-    if not accounts:
-        return ""
-
-    # Flat format (backward compat) when no tool mapping provided
-    if tool_provider_map is None:
-        lines = [
-            "Connected accounts (use the alias as the `account` parameter "
-            "when calling tools to target a specific account):"
-        ]
-        for acct in accounts:
-            provider = acct.get("provider", "unknown")
-            alias = acct.get("alias", "unknown")
-            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 ""
-            lines.append(f"- {provider}/{alias}{detail}")
-        return "\n".join(lines)
-
-    # --- Structured format: group by provider with tool mapping ---
-
-    # Invert tool_provider_map to provider -> [tools]
-    provider_tools: dict[str, list[str]] = {}
-    for tool_name, provider in tool_provider_map.items():
-        provider_tools.setdefault(provider, []).append(tool_name)
-
-    # Filter to relevant providers based on node tools
-    node_tool_set = set(node_tool_names) if node_tool_names else None
-
-    # Group accounts by provider
-    provider_accounts: dict[str, list[dict[str, Any]]] = {}
-    for acct in accounts:
-        provider = acct.get("provider", "unknown")
-        provider_accounts.setdefault(provider, []).append(acct)
-
-    sections: list[str] = ["Connected accounts:"]
-
-    for provider, acct_list in provider_accounts.items():
-        tools_for_provider = sorted(provider_tools.get(provider, []))
-
-        # If node tools specified, only show providers with overlapping tools
-        if node_tool_set is not None:
-            relevant_tools = [t for t in tools_for_provider if t in node_tool_set]
-            if not relevant_tools:
-                continue
-            tools_for_provider = relevant_tools
-
-        # Local-only providers: tools read from env vars, no account= routing
-        all_local = all(a.get("source") == "local" for a in acct_list)
-
-        # Provider header with tools
-        display_name = provider.replace("_", " ").title()
-        if tools_for_provider and not all_local:
-            tools_str = ", ".join(tools_for_provider)
-            sections.append(f'\n{display_name} (use account="<alias>" with: {tools_str}):')
-        elif tools_for_provider and all_local:
-            tools_str = ", ".join(tools_for_provider)
-            sections.append(f"\n{display_name} (tools: {tools_str}):")
-        else:
-            sections.append(f"\n{display_name}:")
-
-        # Account entries
-        for acct in acct_list:
-            alias = acct.get("alias", "unknown")
-            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 ""
-            source_tag = " [local]" if acct.get("source") == "local" else ""
-            sections.append(f"  - {provider}/{alias}{detail}{source_tag}")
-
-    # If filtering removed all providers, return empty
-    if len(sections) <= 1:
-        return ""
-
-    return "\n".join(sections)
-
-
-def compose_system_prompt(
-    identity_prompt: str | None,
-    focus_prompt: str | None,
-    narrative: str | None = None,
-    accounts_prompt: str | None = None,
-) -> str:
-    """Compose the three-layer system prompt.
-
-    Args:
-        identity_prompt: Layer 1 — static agent identity (from GraphSpec).
-        focus_prompt: Layer 3 — per-node focus directive (from NodeSpec.system_prompt).
-        narrative: Layer 2 — auto-generated from conversation state.
-        accounts_prompt: Connected accounts block (sits between identity and narrative).
-
-    Returns:
-        Composed system prompt with all layers present, plus current datetime.
-    """
-    parts: list[str] = []
-
-    # Layer 1: Identity (always first, anchors the personality)
-    if identity_prompt:
-        parts.append(identity_prompt)
-
-    # Accounts (semi-static, deployment-specific)
-    if accounts_prompt:
-        parts.append(f"\n{accounts_prompt}")
-
-    # Layer 2: Narrative (what's happened so far)
-    if narrative:
-        parts.append(f"\n--- Context (what has happened so far) ---\n{narrative}")
-
-    # Layer 3: Focus (current phase directive)
-    if focus_prompt:
-        parts.append(f"\n--- Current Focus ---\n{focus_prompt}")
-
-    return _with_datetime("\n".join(parts) if parts else "")
-
-
-def build_narrative(
-    memory: SharedMemory,
-    execution_path: list[str],
-    graph: GraphSpec,
-) -> str:
-    """Build Layer 2 (narrative) from structured state.
-
-    Deterministic — no LLM call. Reads SharedMemory and execution path
-    to describe what has happened so far. Cheap and fast.
-
-    Args:
-        memory: Current shared memory state.
-        execution_path: List of node IDs visited so far.
-        graph: Graph spec (for node names/descriptions).
-
-    Returns:
-        Narrative string describing the session state.
-    """
-    parts: list[str] = []
-
-    # Describe execution path
-    if execution_path:
-        phase_descriptions: list[str] = []
-        for node_id in execution_path:
-            node_spec = graph.get_node(node_id)
-            if node_spec:
-                phase_descriptions.append(f"- {node_spec.name}: {node_spec.description}")
-            else:
-                phase_descriptions.append(f"- {node_id}")
-        parts.append("Phases completed:\n" + "\n".join(phase_descriptions))
-
-    # Describe key memory values (skip very long values)
-    all_memory = memory.read_all()
-    if all_memory:
-        memory_lines: list[str] = []
-        for key, value in all_memory.items():
-            if value is None:
-                continue
-            val_str = str(value)
-            if len(val_str) > 200:
-                val_str = val_str[:200] + "..."
-            memory_lines.append(f"- {key}: {val_str}")
-        if memory_lines:
-            parts.append("Current state:\n" + "\n".join(memory_lines))
-
-    return "\n\n".join(parts) if parts else ""
-
-
-def build_transition_marker(
-    previous_node: NodeSpec,
-    next_node: NodeSpec,
-    memory: SharedMemory,
-    cumulative_tool_names: list[str],
-    data_dir: Path | str | None = None,
-    adapt_content: str | None = None,
-) -> str:
-    """Build a 'State of the World' transition marker.
-
-    Inserted into the conversation as a user message at phase boundaries.
-    Gives the LLM full situational awareness: what happened, what's stored,
-    what tools are available, and what to focus on next.
-
-    Args:
-        previous_node: NodeSpec of the phase just completed.
-        next_node: NodeSpec of the phase about to start.
-        memory: Current shared memory state.
-        cumulative_tool_names: All tools available (cumulative set).
-        data_dir: Path to spillover data directory.
-        adapt_content: Agent working memory (adapt.md) content.
-
-    Returns:
-        Transition marker message text.
-    """
-    sections: list[str] = []
-
-    # Header
-    sections.append(f"--- PHASE TRANSITION: {previous_node.name} → {next_node.name} ---")
-
-    # What just completed
-    sections.append(f"\nCompleted: {previous_node.name}")
-    sections.append(f"  {previous_node.description}")
-
-    # Outputs in memory
-    all_memory = memory.read_all()
-    if all_memory:
-        memory_lines: list[str] = []
-        for key, value in all_memory.items():
-            if value is None:
-                continue
-            val_str = str(value)
-            if len(val_str) > 300:
-                val_str = val_str[:300] + "..."
-            memory_lines.append(f"  {key}: {val_str}")
-        if memory_lines:
-            sections.append("\nOutputs available:\n" + "\n".join(memory_lines))
-
-    # Files in data directory
-    if data_dir:
-        data_path = Path(data_dir)
-        if data_path.exists():
-            files = sorted(data_path.iterdir())
-            if files:
-                file_lines = [
-                    f"  {f.name} ({f.stat().st_size:,} bytes)" for f in files if f.is_file()
-                ]
-                if file_lines:
-                    sections.append(
-                        "\nData files (use read_file to access):\n" + "\n".join(file_lines)
-                    )
-
-    # Agent working memory
-    if adapt_content:
-        sections.append(f"\n--- Agent Memory ---\n{adapt_content}")
-
-    # Available tools
-    if cumulative_tool_names:
-        sections.append("\nAvailable tools: " + ", ".join(sorted(cumulative_tool_names)))
-
-    # Next phase
-    sections.append(f"\nNow entering: {next_node.name}")
-    sections.append(f"  {next_node.description}")
-
-    # Reflection prompt (engineered metacognition)
-    sections.append(
-        "\nBefore proceeding, briefly reflect: what went well in the "
-        "previous phase? Are there any gaps or surprises worth noting?"
-    )
-
-    sections.append("\n--- END TRANSITION ---")
-
-    return "\n".join(sections)

core/framework/host/__init__.py

@@ -0,0 +1,11 @@
+"""Host layer -- how agents are triggered and hosted."""
+
+from framework.host.agent_host import (  # noqa: F401
+    AgentHost,
+    AgentRuntimeConfig,
+)
+from framework.host.event_bus import AgentEvent, EventBus, EventType  # noqa: F401
+from framework.host.execution_manager import (  # noqa: F401
+    EntryPointSpec,
+    ExecutionManager,
+)

core/framework/host/event_bus.py

@@ -94,17 +94,18 @@ class EventType(StrEnum):
     TOOL_CALL_STARTED = "tool_call_started"
     TOOL_CALL_COMPLETED = "tool_call_completed"
 
-    # Client I/O (client_facing=True nodes only)
+    # Queen/user interaction events
     CLIENT_OUTPUT_DELTA = "client_output_delta"
     CLIENT_INPUT_REQUESTED = "client_input_requested"
+    CLIENT_INPUT_RECEIVED = "client_input_received"
 
-    # Internal node observability (client_facing=False nodes)
+    # Internal node observability
     NODE_INTERNAL_OUTPUT = "node_internal_output"
     NODE_INPUT_BLOCKED = "node_input_blocked"
     NODE_STALLED = "node_stalled"
     NODE_TOOL_DOOM_LOOP = "node_tool_doom_loop"
 
-    # Judge decisions
+    # Judge decisions (implicit judge in event loop nodes)
     JUDGE_VERDICT = "judge_verdict"
 
     # Output tracking
@@ -114,8 +115,13 @@ class EventType(StrEnum):
     NODE_RETRY = "node_retry"
     EDGE_TRAVERSED = "edge_traversed"
 
+    # Worker agent lifecycle (event-driven graph execution)
+    WORKER_COMPLETED = "worker_completed"
+    WORKER_FAILED = "worker_failed"
+
     # Context management
     CONTEXT_COMPACTED = "context_compacted"
+    CONTEXT_USAGE_UPDATED = "context_usage_updated"
 
     # External triggers
     WEBHOOK_RECEIVED = "webhook_received"
@@ -126,26 +132,36 @@ class EventType(StrEnum):
     # Escalation (agent requests handoff to queen)
     ESCALATION_REQUESTED = "escalation_requested"
 
-    # Worker health monitoring (judge → queen → operator)
-    WORKER_ESCALATION_TICKET = "worker_escalation_ticket"
-    QUEEN_INTERVENTION_REQUESTED = "queen_intervention_requested"
-
     # Execution resurrection (auto-restart on non-fatal failure)
     EXECUTION_RESURRECTED = "execution_resurrected"
 
-    # Worker lifecycle (session manager → frontend)
-    WORKER_LOADED = "worker_loaded"
+    # Graph lifecycle (session manager → frontend)
+    WORKER_GRAPH_LOADED = "worker_graph_loaded"
     CREDENTIALS_REQUIRED = "credentials_required"
 
+    # Draft graph (planning phase — lightweight graph preview)
+    DRAFT_GRAPH_UPDATED = "draft_graph_updated"
+
+    # Flowchart map updated (after reconciliation with runtime graph)
+    FLOWCHART_MAP_UPDATED = "flowchart_map_updated"
+
     # Queen phase changes (building <-> staging <-> running)
     QUEEN_PHASE_CHANGED = "queen_phase_changed"
 
-    # Queen thinking hook — persona selected for the current building session
-    QUEEN_PERSONA_SELECTED = "queen_persona_selected"
+    # Queen identity — which queen profile was selected for this session
+    QUEEN_IDENTITY_SELECTED = "queen_identity_selected"
 
     # Subagent reports (one-way progress updates from sub-agents)
     SUBAGENT_REPORT = "subagent_report"
 
+    # Trigger lifecycle (queen-level triggers / heartbeats)
+    TRIGGER_AVAILABLE = "trigger_available"
+    TRIGGER_ACTIVATED = "trigger_activated"
+    TRIGGER_DEACTIVATED = "trigger_deactivated"
+    TRIGGER_FIRED = "trigger_fired"
+    TRIGGER_REMOVED = "trigger_removed"
+    TRIGGER_UPDATED = "trigger_updated"
+
 
 @dataclass
 class AgentEvent:
@@ -159,10 +175,11 @@ class AgentEvent:
     timestamp: datetime = field(default_factory=datetime.now)
     correlation_id: str | None = None  # For tracking related events
     graph_id: str | None = None  # Which graph emitted this event (multi-graph sessions)
+    run_id: str | None = None  # Unique ID per trigger() invocation — used for run dividers
 
     def to_dict(self) -> dict:
         """Convert to dictionary for serialization."""
-        return {
+        d = {
             "type": self.type.value,
             "stream_id": self.stream_id,
             "node_id": self.node_id,
@@ -172,6 +189,9 @@ class AgentEvent:
             "correlation_id": self.correlation_id,
             "graph_id": self.graph_id,
         }
+        if self.run_id is not None:
+            d["run_id"] = self.run_id
+        return d
 
 
 # Type for event handlers
@@ -240,6 +260,128 @@ class EventBus:
         self._semaphore = asyncio.Semaphore(max_concurrent_handlers)
         self._subscription_counter = 0
         self._lock = asyncio.Lock()
+        # Per-session persistent event log (always-on, survives restarts)
+        self._session_log: IO[str] | None = None
+        self._session_log_iteration_offset: int = 0
+        # Accumulator for client_output_delta snapshots — flushed on llm_turn_complete.
+        # Key: (stream_id, node_id, execution_id, iteration, inner_turn) → latest AgentEvent
+        self._pending_output_snapshots: dict[tuple, AgentEvent] = {}
+
+    def set_session_log(self, path: Path, *, iteration_offset: int = 0) -> None:
+        """Enable per-session event persistence to a JSONL file.
+
+        Called once when the queen starts so that all events survive server
+        restarts and can be replayed to reconstruct the frontend state.
+
+        ``iteration_offset`` is added to the ``iteration`` field in logged
+        events so that cold-resumed sessions produce monotonically increasing
+        iteration values — preventing frontend message ID collisions between
+        the original run and resumed runs.
+        """
+        if self._session_log is not None:
+            try:
+                self._session_log.close()
+            except Exception:
+                pass
+        path.parent.mkdir(parents=True, exist_ok=True)
+        self._session_log = open(path, "a", encoding="utf-8")  # noqa: SIM115
+        self._session_log_iteration_offset = iteration_offset
+        logger.info("Session event log → %s (iteration_offset=%d)", path, iteration_offset)
+
+    def close_session_log(self) -> None:
+        """Close the per-session event log file."""
+        # Flush any pending output snapshots before closing
+        self._flush_pending_snapshots()
+        if self._session_log is not None:
+            try:
+                self._session_log.close()
+            except Exception:
+                pass
+            self._session_log = None
+
+    # Event types that are high-frequency streaming deltas — accumulated rather
+    # than written individually to the session log.
+    _STREAMING_DELTA_TYPES = frozenset(
+        {
+            EventType.CLIENT_OUTPUT_DELTA,
+            EventType.LLM_TEXT_DELTA,
+            EventType.LLM_REASONING_DELTA,
+        }
+    )
+
+    def _write_session_log_event(self, event: AgentEvent) -> None:
+        """Write an event to the per-session log with streaming coalescing.
+
+        Streaming deltas (client_output_delta, llm_text_delta) are accumulated
+        in memory.  When llm_turn_complete fires, any pending snapshots for that
+        (stream_id, node_id, execution_id) are flushed as single consolidated
+        events before the turn-complete event itself is written.
+
+        Note: iteration offset is already applied in publish() before this is
+        called, so events here already have correct iteration values.
+        """
+        if self._session_log is None:
+            return
+
+        if event.type in self._STREAMING_DELTA_TYPES:
+            # Accumulate — keep only the latest event (which carries the full snapshot)
+            key = (
+                event.stream_id,
+                event.node_id,
+                event.execution_id,
+                event.data.get("iteration"),
+                event.data.get("inner_turn", 0),
+            )
+            self._pending_output_snapshots[key] = event
+            return
+
+        # On turn-complete, flush accumulated snapshots for this stream first
+        if event.type == EventType.LLM_TURN_COMPLETE:
+            self._flush_pending_snapshots(
+                stream_id=event.stream_id,
+                node_id=event.node_id,
+                execution_id=event.execution_id,
+            )
+
+        line = json.dumps(event.to_dict(), default=str)
+        self._session_log.write(line + "\n")
+        self._session_log.flush()
+
+    def _flush_pending_snapshots(
+        self,
+        stream_id: str | None = None,
+        node_id: str | None = None,
+        execution_id: str | None = None,
+    ) -> None:
+        """Flush accumulated streaming snapshots to the session log.
+
+        When called with filters, only matching entries are flushed.
+        When called without filters (e.g. on close), everything is flushed.
+        """
+        if self._session_log is None or not self._pending_output_snapshots:
+            return
+
+        to_flush: list[tuple] = []
+        for key, _evt in self._pending_output_snapshots.items():
+            if stream_id is not None:
+                k_stream, k_node, k_exec, _, _ = key
+                if k_stream != stream_id or k_node != node_id or k_exec != execution_id:
+                    continue
+            to_flush.append(key)
+
+        for key in to_flush:
+            evt = self._pending_output_snapshots.pop(key)
+            try:
+                line = json.dumps(evt.to_dict(), default=str)
+                self._session_log.write(line + "\n")
+            except Exception:
+                pass
+
+        if to_flush:
+            try:
+                self._session_log.flush()
+            except Exception:
+                pass
 
     def subscribe(
         self,
@@ -305,6 +447,19 @@ class EventBus:
         Args:
             event: Event to publish
         """
+        # Apply iteration offset at the source so ALL consumers (SSE subscribers,
+        # event history, session log) see the same monotonically increasing
+        # iteration values.  Without this, live SSE would use raw iterations
+        # while events.jsonl would use offset iterations, causing ID collisions
+        # on the frontend when replaying after cold resume.
+        if (
+            self._session_log_iteration_offset
+            and isinstance(event.data, dict)
+            and "iteration" in event.data
+        ):
+            offset = self._session_log_iteration_offset
+            event.data = {**event.data, "iteration": event.data["iteration"] + offset}
+
         # Add to history
         async with self._lock:
             self._event_history.append(event)
@@ -325,6 +480,15 @@ class EventBus:
                 except Exception:
                     pass  # never break event delivery
 
+        # Per-session persistent log (always-on when set_session_log was called).
+        # Streaming deltas are coalesced: client_output_delta and llm_text_delta
+        # are accumulated and flushed as a single snapshot event on llm_turn_complete.
+        if self._session_log is not None:
+            try:
+                self._write_session_log_event(event)
+            except Exception:
+                pass  # never break event delivery
+
         # Find matching subscriptions
         matching_handlers: list[EventHandler] = []
 
@@ -371,8 +535,8 @@ class EventBus:
             async with self._semaphore:
                 try:
                     await handler(event)
-                except Exception as e:
-                    logger.error(f"Handler error for {event.type}: {e}")
+                except Exception:
+                    logger.exception(f"Handler error for {event.type}")
 
         # Run all handlers concurrently
         await asyncio.gather(*[run_handler(h) for h in handlers], return_exceptions=True)
@@ -385,6 +549,7 @@ class EventBus:
         execution_id: str,
         input_data: dict[str, Any] | None = None,
         correlation_id: str | None = None,
+        run_id: str | None = None,
     ) -> None:
         """Emit execution started event."""
         await self.publish(
@@ -394,6 +559,7 @@ class EventBus:
                 execution_id=execution_id,
                 data={"input": input_data or {}},
                 correlation_id=correlation_id,
+                run_id=run_id,
             )
         )
 
@@ -403,6 +569,7 @@ class EventBus:
         execution_id: str,
         output: dict[str, Any] | None = None,
         correlation_id: str | None = None,
+        run_id: str | None = None,
     ) -> None:
         """Emit execution completed event."""
         await self.publish(
@@ -412,6 +579,7 @@ class EventBus:
                 execution_id=execution_id,
                 data={"output": output or {}},
                 correlation_id=correlation_id,
+                run_id=run_id,
             )
         )
 
@@ -421,6 +589,7 @@ class EventBus:
         execution_id: str,
         error: str,
         correlation_id: str | None = None,
+        run_id: str | None = None,
     ) -> None:
         """Emit execution failed event."""
         await self.publish(
@@ -430,6 +599,7 @@ class EventBus:
                 execution_id=execution_id,
                 data={"error": error},
                 correlation_id=correlation_id,
+                run_id=run_id,
             )
         )
 
@@ -521,15 +691,19 @@ class EventBus:
         node_id: str,
         iteration: int,
         execution_id: str | None = None,
+        extra_data: dict[str, Any] | None = None,
     ) -> None:
         """Emit node loop iteration event."""
+        data: dict[str, Any] = {"iteration": iteration}
+        if extra_data:
+            data.update(extra_data)
         await self.publish(
             AgentEvent(
                 type=EventType.NODE_LOOP_ITERATION,
                 stream_id=stream_id,
                 node_id=node_id,
                 execution_id=execution_id,
-                data={"iteration": iteration},
+                data=data,
             )
         )
 
@@ -578,6 +752,7 @@ class EventBus:
         content: str,
         snapshot: str,
         execution_id: str | None = None,
+        inner_turn: int = 0,
     ) -> None:
         """Emit LLM text delta event."""
         await self.publish(
@@ -586,7 +761,7 @@ class EventBus:
                 stream_id=stream_id,
                 node_id=node_id,
                 execution_id=execution_id,
-                data={"content": content, "snapshot": snapshot},
+                data={"content": content, "snapshot": snapshot, "inner_turn": inner_turn},
             )
         )
 
@@ -616,6 +791,7 @@ class EventBus:
         model: str,
         input_tokens: int,
         output_tokens: int,
+        cached_tokens: int = 0,
         execution_id: str | None = None,
         iteration: int | None = None,
     ) -> None:
@@ -625,6 +801,7 @@ class EventBus:
             "model": model,
             "input_tokens": input_tokens,
             "output_tokens": output_tokens,
+            "cached_tokens": cached_tokens,
         }
         if iteration is not None:
             data["iteration"] = iteration
@@ -700,9 +877,10 @@ class EventBus:
         snapshot: str,
         execution_id: str | None = None,
         iteration: int | None = None,
+        inner_turn: int = 0,
     ) -> None:
-        """Emit client output delta event (client_facing=True nodes)."""
-        data: dict = {"content": content, "snapshot": snapshot}
+        """Emit user-facing output delta for interactive queen turns."""
+        data: dict = {"content": content, "snapshot": snapshot, "inner_turn": inner_turn}
         if iteration is not None:
             data["iteration"] = iteration
         await self.publish(
@@ -722,16 +900,23 @@ class EventBus:
         prompt: str = "",
         execution_id: str | None = None,
         options: list[str] | None = None,
+        questions: list[dict] | None = None,
     ) -> None:
-        """Emit client input requested event (client_facing=True nodes).
+        """Emit a user-input request for interactive queen turns.
 
         Args:
             options: Optional predefined choices for the user (1-3 items).
-                     The frontend appends an "Other" free-text option automatically.
+                     The frontend appends an "Other" free-text option
+                     automatically.
+            questions: Optional list of question dicts for multi-question
+                       batches (from ask_user_multiple). Each dict has id,
+                       prompt, and optional options.
         """
         data: dict[str, Any] = {"prompt": prompt}
         if options:
             data["options"] = options
+        if questions:
+            data["questions"] = questions
         await self.publish(
             AgentEvent(
                 type=EventType.CLIENT_INPUT_REQUESTED,
@@ -751,7 +936,7 @@ class EventBus:
         content: str,
         execution_id: str | None = None,
     ) -> None:
-        """Emit node internal output event (client_facing=False nodes)."""
+        """Emit node internal output for non-user-facing execution."""
         await self.publish(
             AgentEvent(
                 type=EventType.NODE_INTERNAL_OUTPUT,
@@ -909,6 +1094,54 @@ class EventBus:
             )
         )
 
+    async def emit_worker_completed(
+        self,
+        stream_id: str,
+        node_id: str,
+        worker_id: str,
+        success: bool,
+        output: dict[str, Any],
+        activations: list[dict[str, Any]] | None = None,
+        execution_id: str | None = None,
+        **extra_data: Any,
+    ) -> None:
+        """Emit worker completed event with outgoing activations."""
+        data: dict[str, Any] = {
+            "worker_id": worker_id,
+            "success": success,
+            "output": output,
+            "activations": activations or [],
+            **extra_data,
+        }
+        await self.publish(
+            AgentEvent(
+                type=EventType.WORKER_COMPLETED,
+                stream_id=stream_id,
+                node_id=node_id,
+                execution_id=execution_id,
+                data=data,
+            )
+        )
+
+    async def emit_worker_failed(
+        self,
+        stream_id: str,
+        node_id: str,
+        worker_id: str,
+        error: str,
+        execution_id: str | None = None,
+    ) -> None:
+        """Emit worker failed event."""
+        await self.publish(
+            AgentEvent(
+                type=EventType.WORKER_FAILED,
+                stream_id=stream_id,
+                node_id=node_id,
+                execution_id=execution_id,
+                data={"worker_id": worker_id, "error": error},
+            )
+        )
+
     async def emit_execution_paused(
         self,
         stream_id: str,
@@ -987,52 +1220,6 @@ class EventBus:
             )
         )
 
-    async def emit_worker_escalation_ticket(
-        self,
-        stream_id: str,
-        node_id: str,
-        ticket: dict,
-        execution_id: str | None = None,
-    ) -> None:
-        """Emitted by health judge when worker shows a degradation pattern."""
-        await self.publish(
-            AgentEvent(
-                type=EventType.WORKER_ESCALATION_TICKET,
-                stream_id=stream_id,
-                node_id=node_id,
-                execution_id=execution_id,
-                data={"ticket": ticket},
-            )
-        )
-
-    async def emit_queen_intervention_requested(
-        self,
-        stream_id: str,
-        node_id: str,
-        ticket_id: str,
-        analysis: str,
-        severity: str,
-        queen_graph_id: str,
-        queen_stream_id: str,
-        execution_id: str | None = None,
-    ) -> None:
-        """Emitted by queen when she decides the operator should be involved."""
-        await self.publish(
-            AgentEvent(
-                type=EventType.QUEEN_INTERVENTION_REQUESTED,
-                stream_id=stream_id,
-                node_id=node_id,
-                execution_id=execution_id,
-                data={
-                    "ticket_id": ticket_id,
-                    "analysis": analysis,
-                    "severity": severity,
-                    "queen_graph_id": queen_graph_id,
-                    "queen_stream_id": queen_stream_id,
-                },
-            )
-        )
-
     async def emit_subagent_report(
         self,
         stream_id: str,

core/framework/host/execution_manager.py

@@ -18,18 +18,18 @@ from dataclasses import dataclass, field
 from datetime import datetime
 from typing import TYPE_CHECKING, Any
 
-from framework.graph.checkpoint_config import CheckpointConfig
-from framework.graph.executor import ExecutionResult, GraphExecutor
-from framework.runtime.event_bus import EventBus
-from framework.runtime.shared_state import IsolationLevel, SharedStateManager
-from framework.runtime.stream_runtime import StreamRuntime, StreamRuntimeAdapter
+from framework.orchestrator.checkpoint_config import CheckpointConfig
+from framework.orchestrator.orchestrator import ExecutionResult, Orchestrator
+from framework.host.event_bus import EventBus
+from framework.host.shared_state import IsolationLevel, SharedBufferManager
+from framework.host.stream_runtime import StreamDecisionTracker, StreamRuntimeAdapter
 
 if TYPE_CHECKING:
-    from framework.graph.edge import GraphSpec
-    from framework.graph.goal import Goal
+    from framework.orchestrator.edge import GraphSpec
+    from framework.orchestrator.goal import Goal
     from framework.llm.provider import LLMProvider, Tool
-    from framework.runtime.event_bus import AgentEvent
-    from framework.runtime.outcome_aggregator import OutcomeAggregator
+    from framework.host.event_bus import AgentEvent
+    from framework.host.outcome_aggregator import OutcomeAggregator
     from framework.storage.concurrent import ConcurrentStorage
     from framework.storage.session_store import SessionStore
 
@@ -127,12 +127,13 @@ class ExecutionContext:
     input_data: dict[str, Any]
     isolation_level: IsolationLevel
     session_state: dict[str, Any] | None = None  # For resuming from pause
+    run_id: str | None = None  # Unique ID per trigger() invocation
     started_at: datetime = field(default_factory=datetime.now)
     completed_at: datetime | None = None
     status: str = "pending"  # pending, running, completed, failed, paused
 
 
-class ExecutionStream:
+class ExecutionManager:
     """
     Manages concurrent executions for a single entry point.
 
@@ -169,7 +170,7 @@ class ExecutionStream:
         entry_spec: EntryPointSpec,
         graph: "GraphSpec",
         goal: "Goal",
-        state_manager: SharedStateManager,
+        state_manager: SharedBufferManager,
         storage: "ConcurrentStorage",
         outcome_aggregator: "OutcomeAggregator",
         event_bus: "EventBus | None" = None,
@@ -185,6 +186,12 @@ class ExecutionStream:
         accounts_prompt: str = "",
         accounts_data: list[dict] | None = None,
         tool_provider_map: dict[str, str] | None = None,
+        skills_catalog_prompt: str = "",
+        protocols_prompt: str = "",
+        skill_dirs: list[str] | None = None,
+        context_warn_ratio: float | None = None,
+        batch_init_nudge: str | None = None,
+        dynamic_memory_provider_factory: Callable[[str], Callable[[], str] | None] | None = None,
     ):
         """
         Initialize execution stream.
@@ -208,6 +215,11 @@ class ExecutionStream:
             accounts_prompt: Connected accounts block for system prompt injection
             accounts_data: Raw account data for per-node prompt generation
             tool_provider_map: Tool name to provider name mapping for account routing
+            skills_catalog_prompt: Available skills catalog for system prompt
+            protocols_prompt: Default skill operational protocols for system prompt
+            skill_dirs: Skill base directories for Tier 3 resource access
+            context_warn_ratio: Token usage ratio to trigger DS-13 preservation warning
+            batch_init_nudge: System prompt nudge for DS-12 batch auto-detection
         """
         self.stream_id = stream_id
         self.entry_spec = entry_spec
@@ -229,9 +241,28 @@ class ExecutionStream:
         self._accounts_prompt = accounts_prompt
         self._accounts_data = accounts_data
         self._tool_provider_map = tool_provider_map
+        self._skills_catalog_prompt = skills_catalog_prompt
+        self._protocols_prompt = protocols_prompt
+        self._skill_dirs: list[str] = skill_dirs or []
+        self._context_warn_ratio: float | None = context_warn_ratio
+        self._batch_init_nudge: str | None = batch_init_nudge
+        self._dynamic_memory_provider_factory = dynamic_memory_provider_factory
+
+        _es_logger = logging.getLogger(__name__)
+        if protocols_prompt:
+            _es_logger.info(
+                "ExecutionStream[%s] received protocols_prompt (%d chars)",
+                stream_id,
+                len(protocols_prompt),
+            )
+        else:
+            _es_logger.warning(
+                "ExecutionStream[%s] received EMPTY protocols_prompt",
+                stream_id,
+            )
 
         # Create stream-scoped runtime
-        self._runtime = StreamRuntime(
+        self._runtime = StreamDecisionTracker(
             stream_id=stream_id,
             storage=storage,
             outcome_aggregator=outcome_aggregator,
@@ -240,7 +271,7 @@ class ExecutionStream:
         # Execution tracking
         self._active_executions: dict[str, ExecutionContext] = {}
         self._execution_tasks: dict[str, asyncio.Task] = {}
-        self._active_executors: dict[str, GraphExecutor] = {}
+        self._active_executors: dict[str, Orchestrator] = {}
         self._cancel_reasons: dict[str, str] = {}
         self._execution_results: OrderedDict[str, ExecutionResult] = OrderedDict()
         self._execution_result_times: dict[str, float] = {}
@@ -270,7 +301,7 @@ class ExecutionStream:
 
         # Emit stream started event
         if self._scoped_event_bus:
-            from framework.runtime.event_bus import AgentEvent, EventType
+            from framework.host.event_bus import AgentEvent, EventType
 
             await self._scoped_event_bus.publish(
                 AgentEvent(
@@ -328,7 +359,7 @@ class ExecutionStream:
 
         Each entry is ``{"node_id": ..., "execution_id": ...}``.
         The currently executing node is placed first so that
-        ``inject_worker_message`` targets the active node, not a stale one.
+        ``inject_message`` targets the active node, not a stale one.
         """
         injectable: list[dict[str, str]] = []
         current_first: list[dict[str, str]] = []
@@ -395,7 +426,7 @@ class ExecutionStream:
 
         # Emit stream stopped event
         if self._scoped_event_bus:
-            from framework.runtime.event_bus import AgentEvent, EventType
+            from framework.host.event_bus import AgentEvent, EventType
 
             await self._scoped_event_bus.publish(
                 AgentEvent(
@@ -410,6 +441,7 @@ class ExecutionStream:
         content: str,
         *,
         is_client_input: bool = False,
+        image_content: list[dict[str, Any]] | None = None,
     ) -> bool:
         """Inject user input into a running client-facing EventLoopNode.
 
@@ -421,7 +453,33 @@ class ExecutionStream:
         for executor in self._active_executors.values():
             node = executor.node_registry.get(node_id)
             if node is not None and hasattr(node, "inject_event"):
-                await node.inject_event(content, is_client_input=is_client_input)
+                await node.inject_event(
+                    content, is_client_input=is_client_input, image_content=image_content
+                )
+                return True
+        return False
+
+    async def inject_trigger(
+        self,
+        node_id: str,
+        trigger: Any,
+    ) -> bool:
+        """Inject a trigger event into a running queen EventLoopNode.
+
+        Searches active executors for a node matching ``node_id`` and calls
+        its ``inject_trigger()`` method to wake the queen.
+
+        Args:
+            node_id: The queen EventLoopNode ID.
+            trigger: A ``TriggerEvent`` instance (typed as Any to avoid
+                circular imports with graph layer).
+
+        Returns True if the trigger was delivered, False otherwise.
+        """
+        for executor in self._active_executors.values():
+            node = executor.node_registry.get(node_id)
+            if node is not None and hasattr(node, "inject_trigger"):
+                await node.inject_trigger(trigger)
                 return True
         return False
 
@@ -430,6 +488,7 @@ class ExecutionStream:
         input_data: dict[str, Any],
         correlation_id: str | None = None,
         session_state: dict[str, Any] | None = None,
+        run_id: str | None = None,
     ) -> str:
         """
         Queue an execution and return its ID.
@@ -440,6 +499,7 @@ class ExecutionStream:
             input_data: Input data for this execution
             correlation_id: Optional ID to correlate related executions
             session_state: Optional session state to resume from (with paused_at, memory)
+            run_id: Unique ID for this trigger invocation (for run dividers)
 
         Returns:
             Execution ID for tracking
@@ -492,6 +552,14 @@ class ExecutionStream:
             correlation_id = execution_id
 
         # Create execution context
+        effective_run_id = None
+        if session_state:
+            existing_run_id = session_state.get("run_id")
+            if isinstance(existing_run_id, str) and existing_run_id:
+                effective_run_id = existing_run_id
+        if effective_run_id is None:
+            effective_run_id = run_id
+
         ctx = ExecutionContext(
             id=execution_id,
             correlation_id=correlation_id,
@@ -500,6 +568,7 @@ class ExecutionStream:
             input_data=input_data,
             isolation_level=self.entry_spec.get_isolation_level(),
             session_state=session_state,
+            run_id=effective_run_id,
         )
 
         async with self._lock:
@@ -575,10 +644,12 @@ class ExecutionStream:
                         execution_id=execution_id,
                         input_data=ctx.input_data,
                         correlation_id=ctx.correlation_id,
+                        run_id=ctx.run_id,
                     )
+                self._write_run_event(execution_id, ctx.run_id, "run_started")
 
                 # Create execution-scoped memory
-                self._state_manager.create_memory(
+                self._state_manager.create_buffer(
                     execution_id=execution_id,
                     stream_id=self.stream_id,
                     isolation=ctx.isolation_level,
@@ -597,7 +668,7 @@ class ExecutionStream:
                 # Create per-execution runtime logger
                 runtime_logger = None
                 if self._runtime_log_store:
-                    from framework.runtime.runtime_logger import RuntimeLogger
+                    from framework.tracker.runtime_logger import RuntimeLogger
 
                     runtime_logger = RuntimeLogger(
                         store=self._runtime_log_store, agent_id=self.graph.id
@@ -626,12 +697,7 @@ class ExecutionStream:
                 # forward so the next attempt resumes at the failed node.
                 while True:
                     # Create executor for this execution.
-                    # Each execution gets its own storage under sessions/{exec_id}/
-                    # so conversations, spillover, and data files are all scoped
-                    # to this execution.  The executor sets data_dir via execution
-                    # context (contextvars) so data tools and spillover share the
-                    # same session-scoped directory.
-                    executor = GraphExecutor(
+                    executor = Orchestrator(
                         runtime=runtime_adapter,
                         llm=self._llm,
                         tools=self._tools,
@@ -639,12 +705,23 @@ class ExecutionStream:
                         event_bus=self._scoped_event_bus,
                         stream_id=self.stream_id,
                         execution_id=execution_id,
+                        run_id=ctx.run_id or "",
                         storage_path=exec_storage,
                         runtime_logger=runtime_logger,
                         loop_config=self.graph.loop_config,
                         accounts_prompt=self._accounts_prompt,
                         accounts_data=self._accounts_data,
                         tool_provider_map=self._tool_provider_map,
+                        skills_catalog_prompt=self._skills_catalog_prompt,
+                        protocols_prompt=self._protocols_prompt,
+                        skill_dirs=self._skill_dirs,
+                        context_warn_ratio=self._context_warn_ratio,
+                        batch_init_nudge=self._batch_init_nudge,
+                        dynamic_memory_provider=(
+                            self._dynamic_memory_provider_factory(execution_id)
+                            if self._dynamic_memory_provider_factory is not None
+                            else None
+                        ),
                     )
                     # Track executor so inject_input() can reach EventLoopNode instances
                     self._active_executors[execution_id] = executor
@@ -681,7 +758,7 @@ class ExecutionStream:
 
                         # Emit resurrection event
                         if self._scoped_event_bus:
-                            from framework.runtime.event_bus import AgentEvent, EventType
+                            from framework.host.event_bus import AgentEvent, EventType
 
                             await self._scoped_event_bus.publish(
                                 AgentEvent(
@@ -740,6 +817,7 @@ class ExecutionStream:
                             execution_id=execution_id,
                             output=result.output,
                             correlation_id=ctx.correlation_id,
+                            run_id=ctx.run_id,
                         )
                     elif result.paused_at:
                         # The executor returns paused_at on CancelledError but
@@ -757,8 +835,22 @@ class ExecutionStream:
                             execution_id=execution_id,
                             error=result.error or "Unknown error",
                             correlation_id=ctx.correlation_id,
+                            run_id=ctx.run_id,
                         )
 
+                # Write run event for historical restoration
+                if result.success:
+                    self._write_run_event(execution_id, ctx.run_id, "run_completed")
+                elif result.paused_at:
+                    self._write_run_event(execution_id, ctx.run_id, "run_paused")
+                else:
+                    self._write_run_event(
+                        execution_id,
+                        ctx.run_id,
+                        "run_failed",
+                        {"error": result.error or "Unknown error"},
+                    )
+
                 logger.debug(f"Execution {execution_id} completed: success={result.success}")
 
             except asyncio.CancelledError:
@@ -818,8 +910,10 @@ class ExecutionStream:
                             execution_id=execution_id,
                             error=cancel_reason,
                             correlation_id=ctx.correlation_id,
+                            run_id=ctx.run_id,
                         )
 
+                self._write_run_event(execution_id, ctx.run_id, "run_cancelled")
                 # Don't re-raise - we've handled it and saved state
 
             except Exception as e:
@@ -856,7 +950,9 @@ class ExecutionStream:
                         execution_id=execution_id,
                         error=str(e),
                         correlation_id=ctx.correlation_id,
+                        run_id=ctx.run_id,
                     )
+                self._write_run_event(execution_id, ctx.run_id, "run_failed", {"error": str(e)})
 
             finally:
                 # Clean up state
@@ -872,6 +968,39 @@ class ExecutionStream:
                     self._completion_events.pop(execution_id, None)
                     self._execution_tasks.pop(execution_id, None)
 
+    def _write_run_event(
+        self,
+        execution_id: str,
+        run_id: str | None,
+        event: str,
+        extra: dict[str, Any] | None = None,
+    ) -> None:
+        """Append a run lifecycle event to runs.jsonl for historical restoration."""
+        if not self._session_store or not run_id:
+            return
+        import json as _json
+
+        try:
+            session_dir = self._session_store.get_session_path(execution_id)
+        except ValueError:
+            return
+        runs_file = session_dir / "runs.jsonl"
+        now = datetime.now()
+        record = {
+            "run_id": run_id,
+            "event": event,
+            "timestamp": now.isoformat(),
+            "created_at": now.timestamp(),
+        }
+        if extra:
+            record.update(extra)
+        try:
+            runs_file.parent.mkdir(parents=True, exist_ok=True)
+            with open(runs_file, "a", encoding="utf-8") as f:
+                f.write(_json.dumps(record) + "\n")
+        except OSError:
+            pass  # Non-critical — don't break execution
+
     async def _write_session_state(
         self,
         execution_id: str,
@@ -926,6 +1055,7 @@ class ExecutionStream:
                     agent_id=self.graph.id,
                     entry_point=self.entry_spec.id,
                 )
+                state.current_run_id = ctx.run_id
             else:
                 # Create initial state — when resuming, preserve the previous
                 # execution's progress so crashes don't lose track of state.
@@ -956,8 +1086,9 @@ class ExecutionStream:
                         updated_at=now,
                     ),
                     progress=progress,
-                    memory=ss.get("memory", {}),
+                    data_buffer=ss.get("data_buffer", ss.get("memory", {})),
                     input_data=ctx.input_data,
+                    current_run_id=ctx.run_id,
                 )
 
             # Handle error case
@@ -978,12 +1109,12 @@ class ExecutionStream:
     def _create_modified_graph(self) -> "GraphSpec":
         """Create a graph with the entry point overridden.
 
-        Preserves the original graph's entry_points and async_entry_points
-        so that validation correctly considers ALL entry nodes reachable.
+        Preserves the original graph's entry_points so that validation
+        correctly considers ALL entry nodes reachable.
         Each stream only executes from its own entry_node, but the full
         graph must validate with all entry points accounted for.
         """
-        from framework.graph.edge import GraphSpec
+        from framework.orchestrator.edge import GraphSpec
 
         # Merge entry points: this stream's entry + original graph's primary
         # entry + any other entry points. This ensures all nodes are
@@ -1004,7 +1135,6 @@ class ExecutionStream:
             version=self.graph.version,
             entry_node=self.entry_spec.entry_node,  # Use our entry point
             entry_points=merged_entry_points,
-            async_entry_points=self.graph.async_entry_points,
             terminal_nodes=self.graph.terminal_nodes,
             pause_nodes=self.graph.pause_nodes,
             nodes=self.graph.nodes,
@@ -1081,9 +1211,21 @@ class ExecutionStream:
             task.cancel()
             # Wait briefly for the task to finish. Don't block indefinitely —
             # the task may be stuck in a long LLM API call that doesn't
-            # respond to cancellation quickly. The cancellation is already
-            # requested; the task will clean up in the background.
+            # respond to cancellation quickly.
             done, _ = await asyncio.wait({task}, timeout=5.0)
+            if not done:
+                # Task didn't finish within timeout — clean up bookkeeping now
+                # so the session doesn't think it still has running executions.
+                # The task will continue winding down in the background and its
+                # finally block will harmlessly pop already-removed keys.
+                logger.warning(
+                    "Execution %s did not finish within cancel timeout; force-cleaning bookkeeping",
+                    execution_id,
+                )
+                async with self._lock:
+                    self._active_executions.pop(execution_id, None)
+                    self._execution_tasks.pop(execution_id, None)
+                self._active_executors.pop(execution_id, None)
             return True
         return False
 

core/framework/host/outcome_aggregator.py

@@ -14,8 +14,8 @@ from typing import TYPE_CHECKING, Any
 from framework.schemas.decision import Decision, Outcome
 
 if TYPE_CHECKING:
-    from framework.graph.goal import Goal
-    from framework.runtime.event_bus import EventBus
+    from framework.orchestrator.goal import Goal
+    from framework.host.event_bus import EventBus
 
 logger = logging.getLogger(__name__)
 

core/framework/host/shared_state.py

@@ -1,10 +1,10 @@
 """
-Shared State Manager - Manages state across concurrent executions.
+Shared Buffer Manager - Manages state across concurrent executions.
 
 Provides different isolation levels:
-- ISOLATED: Each execution has its own memory copy
-- SHARED: All executions read/write same memory (eventual consistency)
-- SYNCHRONIZED: Shared memory with write locks (strong consistency)
+- ISOLATED: Each execution has its own state copy
+- SHARED: All executions read/write same state (eventual consistency)
+- SYNCHRONIZED: Shared state with write locks (strong consistency)
 """
 
 import asyncio
@@ -46,7 +46,7 @@ class StateChange:
     timestamp: float = field(default_factory=time.time)
 
 
-class SharedStateManager:
+class SharedBufferManager:
     """
     Manages shared state across concurrent executions.
 
@@ -61,18 +61,18 @@ class SharedStateManager:
     - SYNCHRONIZED: Like SHARED but with write locks
 
     Example:
-        manager = SharedStateManager()
+        manager = SharedBufferManager()
 
-        # Create memory for an execution
-        memory = manager.create_memory(
+        # Create buffer for an execution
+        buf = manager.create_buffer(
             execution_id="exec_123",
             stream_id="webhook",
             isolation=IsolationLevel.SHARED,
         )
 
-        # Read/write through the memory
-        await memory.write("customer_id", "cust_456", scope=StateScope.STREAM)
-        value = await memory.read("customer_id")
+        # Read/write through the buffer
+        await buf.write("customer_id", "cust_456", scope=StateScope.STREAM)
+        value = await buf.read("customer_id")
     """
 
     def __init__(self):
@@ -93,14 +93,14 @@ class SharedStateManager:
         # Version tracking
         self._version = 0
 
-    def create_memory(
+    def create_buffer(
         self,
         execution_id: str,
         stream_id: str,
         isolation: IsolationLevel,
-    ) -> "StreamMemory":
+    ) -> "StreamBuffer":
         """
-        Create a memory instance for an execution.
+        Create a buffer instance for an execution.
 
         Args:
             execution_id: Unique execution identifier
@@ -108,7 +108,7 @@ class SharedStateManager:
             isolation: Isolation level for this execution
 
         Returns:
-            StreamMemory instance for reading/writing state
+            StreamBuffer instance for reading/writing state
         """
         # Initialize execution state
         if execution_id not in self._execution_state:
@@ -119,7 +119,7 @@ class SharedStateManager:
             self._stream_state[stream_id] = {}
             self._stream_locks[stream_id] = asyncio.Lock()
 
-        return StreamMemory(
+        return StreamBuffer(
             manager=self,
             execution_id=execution_id,
             stream_id=stream_id,
@@ -343,17 +343,17 @@ class SharedStateManager:
         return self._change_history[-limit:]
 
 
-class StreamMemory:
+class StreamBuffer:
     """
-    Memory interface for a single execution.
+    Buffer interface for a single execution.
 
     Provides scoped access to shared state with proper isolation.
-    Compatible with the existing SharedMemory interface where possible.
+    Compatible with the existing DataBuffer interface where possible.
     """
 
     def __init__(
         self,
-        manager: SharedStateManager,
+        manager: SharedBufferManager,
         execution_id: str,
         stream_id: str,
         isolation: IsolationLevel,
@@ -371,13 +371,13 @@ class StreamMemory:
         self,
         read_keys: list[str],
         write_keys: list[str],
-    ) -> "StreamMemory":
+    ) -> "StreamBuffer":
         """
         Create a scoped view with read/write permissions.
 
-        Compatible with existing SharedMemory.with_permissions().
+        Compatible with existing DataBuffer.with_permissions().
         """
-        scoped = StreamMemory(
+        scoped = StreamBuffer(
             manager=self._manager,
             execution_id=self._execution_id,
             stream_id=self._stream_id,
@@ -434,7 +434,7 @@ class StreamMemory:
 
         return all_state
 
-    # === SYNC API (for backward compatibility with SharedMemory) ===
+    # === SYNC API (for backward compatibility with DataBuffer) ===
 
     def read_sync(self, key: str) -> Any:
         """

core/framework/host/stream_runtime.py

@@ -18,12 +18,12 @@ from framework.schemas.run import Run, RunStatus
 from framework.storage.concurrent import ConcurrentStorage
 
 if TYPE_CHECKING:
-    from framework.runtime.outcome_aggregator import OutcomeAggregator
+    from framework.host.outcome_aggregator import OutcomeAggregator
 
 logger = logging.getLogger(__name__)
 
 
-class StreamRuntime:
+class StreamDecisionTracker:
     """
     Thread-safe runtime for a single execution stream.
 
@@ -431,7 +431,7 @@ class StreamRuntimeAdapter:
     by providing the same API as Runtime but routing to a specific execution.
     """
 
-    def __init__(self, stream_runtime: StreamRuntime, execution_id: str):
+    def __init__(self, stream_runtime: StreamDecisionTracker, execution_id: str):
         """
         Create adapter for a specific execution.
 

core/framework/host/triggers.py

@@ -0,0 +1,22 @@
+"""Trigger definitions for queen-level heartbeats (timers, webhooks)."""
+
+from __future__ import annotations
+
+from dataclasses import dataclass, field
+from typing import Any
+
+
+@dataclass
+class TriggerDefinition:
+    """A registered trigger that can be activated on the queen runtime.
+
+    Trigger *definitions* come from the worker's ``triggers.json``.
+    Activation state is per-session (persisted in ``SessionState.active_triggers``).
+    """
+
+    id: str
+    trigger_type: str  # "timer" | "webhook"
+    trigger_config: dict[str, Any] = field(default_factory=dict)
+    description: str = ""
+    task: str = ""
+    active: bool = False

core/framework/host/webhook_server.py

@@ -13,7 +13,7 @@ from dataclasses import dataclass
 
 from aiohttp import web
 
-from framework.runtime.event_bus import EventBus
+from framework.host.event_bus import EventBus
 
 logger = logging.getLogger(__name__)
 

core/framework/llm/antigravity.py

@@ -0,0 +1,706 @@
+"""Antigravity (Google internal Cloud Code Assist) LLM provider.
+
+Antigravity is Google's unified gateway API that routes requests to Gemini,
+Claude, and GPT-OSS models through a single Gemini-style interface.  It is
+NOT the public ``generativelanguage.googleapis.com`` API.
+
+Authentication uses Google OAuth2.  Token refresh is done directly with the
+OAuth client secret — no local proxy required.
+
+Credential sources (checked in order):
+  1. ``~/.hive/antigravity-accounts.json`` (native OAuth implementation)
+  2. Antigravity IDE SQLite state DB (macOS / Linux)
+"""
+
+from __future__ import annotations
+
+import json
+import logging
+import re
+import time
+import uuid
+from collections.abc import AsyncIterator, Callable, Iterator
+from pathlib import Path
+from typing import Any
+
+from framework.llm.provider import LLMProvider, LLMResponse, Tool
+from framework.llm.stream_events import (
+    FinishEvent,
+    StreamErrorEvent,
+    StreamEvent,
+    TextDeltaEvent,
+    TextEndEvent,
+    ToolCallEvent,
+)
+
+logger = logging.getLogger(__name__)
+
+# ---------------------------------------------------------------------------
+# Constants
+# ---------------------------------------------------------------------------
+
+_TOKEN_URL = "https://oauth2.googleapis.com/token"
+
+# Fallback order: daily sandbox → autopush sandbox → production
+_ENDPOINTS = [
+    "https://daily-cloudcode-pa.sandbox.googleapis.com",
+    "https://autopush-cloudcode-pa.sandbox.googleapis.com",
+    "https://cloudcode-pa.googleapis.com",
+]
+_DEFAULT_PROJECT_ID = "rising-fact-p41fc"
+_TOKEN_REFRESH_BUFFER_SECS = 60
+
+# Credentials file in ~/.hive/ (native implementation)
+_ACCOUNTS_FILE = Path.home() / ".hive" / "antigravity-accounts.json"
+_IDE_STATE_DB_MAC = (
+    Path.home()
+    / "Library"
+    / "Application Support"
+    / "Antigravity"
+    / "User"
+    / "globalStorage"
+    / "state.vscdb"
+)
+_IDE_STATE_DB_LINUX = (
+    Path.home() / ".config" / "Antigravity" / "User" / "globalStorage" / "state.vscdb"
+)
+_IDE_STATE_DB_KEY = "antigravityUnifiedStateSync.oauthToken"
+
+_BASE_HEADERS: dict[str, str] = {
+    # Mimic the Antigravity Electron app so the API accepts the request.
+    "User-Agent": (
+        "Mozilla/5.0 (Macintosh; Intel Mac OS X 10_15_7) AppleWebKit/537.36 "
+        "(KHTML, like Gecko) Antigravity/1.18.3 Chrome/138.0.7204.235 "
+        "Electron/37.3.1 Safari/537.36"
+    ),
+    "X-Goog-Api-Client": "google-cloud-sdk vscode_cloudshelleditor/0.1",
+    "Client-Metadata": '{"ideType":"ANTIGRAVITY","platform":"MACOS","pluginType":"GEMINI"}',
+}
+
+
+# ---------------------------------------------------------------------------
+# Credential loading helpers
+# ---------------------------------------------------------------------------
+
+
+def _load_from_json_file() -> tuple[str | None, str | None, str, float]:
+    """Read credentials from JSON accounts file.
+
+    Reads from ~/.hive/antigravity-accounts.json.
+
+    Returns ``(access_token | None, refresh_token | None, project_id, expires_at)``.
+    ``expires_at`` is a Unix timestamp (seconds); 0.0 means unknown.
+    """
+    if not _ACCOUNTS_FILE.exists():
+        return None, None, _DEFAULT_PROJECT_ID, 0.0
+    try:
+        with open(_ACCOUNTS_FILE, encoding="utf-8") as fh:
+            data = json.load(fh)
+    except (OSError, json.JSONDecodeError) as exc:
+        logger.debug("Failed to read Antigravity accounts file: %s", exc)
+        return None, None, _DEFAULT_PROJECT_ID, 0.0
+
+    accounts = data.get("accounts", [])
+    if not accounts:
+        return None, None, _DEFAULT_PROJECT_ID, 0.0
+
+    account = next((a for a in accounts if a.get("enabled", True) is not False), accounts[0])
+    schema_version = data.get("schemaVersion", 1)
+
+    if schema_version >= 4:
+        # V4 schema: refresh = "refreshToken|projectId[|managedProjectId]"
+        refresh_str = account.get("refresh", "")
+        parts = refresh_str.split("|") if refresh_str else []
+        refresh_token: str | None = parts[0] if parts else None
+        project_id = parts[1] if len(parts) >= 2 and parts[1] else _DEFAULT_PROJECT_ID
+
+        access_token: str | None = account.get("access")
+        expires_ms: int = account.get("expires", 0)
+        expires_at = float(expires_ms) / 1000.0 if expires_ms else 0.0
+
+        # Treat near-expiry tokens as absent so _ensure_token() triggers a refresh.
+        if access_token and expires_at and time.time() >= expires_at - _TOKEN_REFRESH_BUFFER_SECS:
+            access_token = None
+            expires_at = 0.0
+
+        return access_token, refresh_token, project_id, expires_at
+    else:
+        # V1–V3 schema: plain accessToken / refreshToken fields
+        access_token = account.get("accessToken")
+        refresh_token = account.get("refreshToken")
+        # Estimate expiry from last_refresh + 1 h
+        last_refresh_str: str | None = data.get("last_refresh")
+        expires_at = 0.0
+        if last_refresh_str:
+            try:
+                from datetime import datetime  # noqa: PLC0415
+
+                ts = datetime.fromisoformat(last_refresh_str.replace("Z", "+00:00")).timestamp()
+                expires_at = ts + 3600.0
+                if time.time() >= expires_at - _TOKEN_REFRESH_BUFFER_SECS:
+                    access_token = None
+            except (ValueError, TypeError):
+                pass
+        return access_token, refresh_token, _DEFAULT_PROJECT_ID, expires_at
+
+
+def _load_from_ide_db() -> tuple[str | None, str | None, float]:
+    """Extract ``(access_token, refresh_token, expires_at)`` from the IDE SQLite DB."""
+    import base64  # noqa: PLC0415
+    import sqlite3  # noqa: PLC0415
+
+    for db_path in (_IDE_STATE_DB_MAC, _IDE_STATE_DB_LINUX):
+        if not db_path.exists():
+            continue
+        try:
+            con = sqlite3.connect(f"file:{db_path}?mode=ro", uri=True)
+            try:
+                row = con.execute(
+                    "SELECT value FROM ItemTable WHERE key = ?",
+                    (_IDE_STATE_DB_KEY,),
+                ).fetchone()
+            finally:
+                con.close()
+            if not row:
+                continue
+
+            blob = base64.b64decode(row[0])
+            candidates = re.findall(rb"[A-Za-z0-9+/=_\-]{40,}", blob)
+            access_token: str | None = None
+            refresh_token: str | None = None
+            for candidate in candidates:
+                try:
+                    padded = candidate + b"=" * (-len(candidate) % 4)
+                    inner = base64.urlsafe_b64decode(padded)
+                except Exception:
+                    continue
+                if not access_token:
+                    m = re.search(rb"ya29\.[A-Za-z0-9_\-\.]+", inner)
+                    if m:
+                        access_token = m.group(0).decode("ascii")
+                if not refresh_token:
+                    m = re.search(rb"1//[A-Za-z0-9_\-\.]+", inner)
+                    if m:
+                        refresh_token = m.group(0).decode("ascii")
+                if access_token and refresh_token:
+                    break
+
+            if access_token:
+                # Estimate expiry from DB mtime (IDE refreshes while running)
+                mtime = db_path.stat().st_mtime
+                expires_at = mtime + 3600.0
+                return access_token, refresh_token, expires_at
+        except Exception as exc:
+            logger.debug("Failed to read Antigravity IDE state DB: %s", exc)
+            continue
+    return None, None, 0.0
+
+
+def _do_token_refresh(refresh_token: str) -> tuple[str, float] | None:
+    """POST to Google OAuth endpoint and return ``(new_access_token, expires_at)``.
+
+    The client secret is sourced via ``get_antigravity_client_secret()`` (env var,
+    config file, or npm package fallback). When unavailable the refresh is attempted
+    without it — Google will reject it for web-app clients, but the npm fallback in
+    ``get_antigravity_client_secret()`` should ensure the secret is found at runtime.
+
+    Returns None when the HTTP request fails.
+    """
+    from framework.config import get_antigravity_client_secret  # noqa: PLC0415
+
+    client_secret = get_antigravity_client_secret()
+    if not client_secret:
+        logger.debug(
+            "Antigravity client secret not configured — attempting refresh without it. "
+            "Set ANTIGRAVITY_CLIENT_SECRET or run quickstart to configure."
+        )
+
+    import urllib.error  # noqa: PLC0415
+    import urllib.parse  # noqa: PLC0415
+    import urllib.request  # noqa: PLC0415
+
+    from framework.config import get_antigravity_client_id  # noqa: PLC0415
+
+    params: dict[str, str] = {
+        "grant_type": "refresh_token",
+        "refresh_token": refresh_token,
+        "client_id": get_antigravity_client_id(),
+    }
+    if client_secret:
+        params["client_secret"] = client_secret
+    body = urllib.parse.urlencode(params).encode("utf-8")
+
+    req = urllib.request.Request(
+        _TOKEN_URL,
+        data=body,
+        headers={"Content-Type": "application/x-www-form-urlencoded"},
+        method="POST",
+    )
+    try:
+        with urllib.request.urlopen(req, timeout=15) as resp:  # noqa: S310
+            payload = json.loads(resp.read())
+        access_token: str = payload["access_token"]
+        expires_in: int = payload.get("expires_in", 3600)
+        logger.debug("Antigravity token refreshed successfully")
+        return access_token, time.time() + expires_in
+    except Exception as exc:
+        logger.debug("Antigravity token refresh failed: %s", exc)
+        return None
+
+
+# ---------------------------------------------------------------------------
+# Message conversion helpers
+# ---------------------------------------------------------------------------
+
+
+def _clean_tool_name(name: str) -> str:
+    """Sanitize a tool name for the Antigravity function-calling schema."""
+    name = re.sub(r"[/\s]", "_", name)
+    if name and not (name[0].isalpha() or name[0] == "_"):
+        name = "_" + name
+    return name[:64]
+
+
+def _to_gemini_contents(
+    messages: list[dict[str, Any]],
+    thought_sigs: dict[str, str] | None = None,
+) -> list[dict[str, Any]]:
+    """Convert OpenAI-format messages to Gemini-style ``contents`` array."""
+    # Pre-build a map tool_call_id → function_name from assistant messages.
+    # Tool result messages (role="tool") only carry tool_call_id, not the name,
+    # but Gemini requires functionResponse.name to match the functionCall.name.
+    tc_id_to_name: dict[str, str] = {}
+    for msg in messages:
+        if msg.get("role") == "assistant":
+            for tc in msg.get("tool_calls") or []:
+                tc_id = tc.get("id")
+                fn_name = tc.get("function", {}).get("name", "")
+                if tc_id and fn_name:
+                    tc_id_to_name[tc_id] = fn_name
+
+    contents: list[dict[str, Any]] = []
+    # Consecutive tool-result messages must be batched into one user turn.
+    pending_tool_parts: list[dict[str, Any]] = []
+
+    def _flush_tool_results() -> None:
+        if pending_tool_parts:
+            contents.append({"role": "user", "parts": list(pending_tool_parts)})
+            pending_tool_parts.clear()
+
+    for msg in messages:
+        role = msg.get("role", "user")
+        content = msg.get("content")
+
+        if role == "system":
+            continue  # Handled via systemInstruction, not in contents.
+
+        if role == "tool":
+            # OpenAI tool result → Gemini functionResponse part.
+            result_str = content if isinstance(content, str) else str(content or "")
+            tc_id = msg.get("tool_call_id", "")
+            # Look up function name from the pre-built map; fall back to msg.name.
+            fn_name = tc_id_to_name.get(tc_id) or msg.get("name", "")
+            pending_tool_parts.append(
+                {
+                    "functionResponse": {
+                        "name": fn_name,
+                        "id": tc_id,
+                        "response": {"content": result_str},
+                    }
+                }
+            )
+            continue
+
+        _flush_tool_results()
+
+        gemini_role = "model" if role == "assistant" else "user"
+        parts: list[dict[str, Any]] = []
+
+        if isinstance(content, str) and content:
+            parts.append({"text": content})
+        elif isinstance(content, list):
+            for block in content:
+                if not isinstance(block, dict):
+                    continue
+                if block.get("type") == "text":
+                    text = block.get("text", "")
+                    if text:
+                        parts.append({"text": text})
+                # Other block types (image_url etc.) skipped.
+
+        # Assistant messages may carry OpenAI-style tool_calls.
+        for tc in msg.get("tool_calls") or []:
+            fn = tc.get("function", {})
+            try:
+                args = json.loads(fn.get("arguments", "{}") or "{}")
+            except (json.JSONDecodeError, TypeError):
+                args = {}
+            tc_id = tc.get("id", str(uuid.uuid4()))
+            fc_part: dict[str, Any] = {
+                "functionCall": {
+                    "name": fn.get("name", ""),
+                    "args": args,
+                    "id": tc_id,
+                }
+            }
+            if thought_sigs:
+                sig = thought_sigs.get(tc_id, "")
+                if sig:
+                    fc_part["thoughtSignature"] = sig  # part-level, not inside functionCall
+            parts.append(fc_part)
+
+        if parts:
+            contents.append({"role": gemini_role, "parts": parts})
+
+    _flush_tool_results()
+
+    # Gemini requires the first turn to be a user turn.  Drop any leading
+    # model messages so the API doesn't reject with a 400.
+    while contents and contents[0].get("role") == "model":
+        contents.pop(0)
+
+    return contents
+
+
+# ---------------------------------------------------------------------------
+# Response parsing helpers
+# ---------------------------------------------------------------------------
+
+
+def _map_finish_reason(reason: str) -> str:
+    return {"STOP": "stop", "MAX_TOKENS": "max_tokens", "OTHER": "tool_use"}.get(
+        (reason or "").upper(), "stop"
+    )
+
+
+def _parse_complete_response(raw: dict[str, Any], model: str) -> LLMResponse:
+    """Parse a non-streaming Antigravity response dict → LLMResponse."""
+    payload: dict[str, Any] = raw.get("response", raw)
+    candidates: list[dict[str, Any]] = payload.get("candidates", [])
+    usage: dict[str, Any] = payload.get("usageMetadata", {})
+
+    text_parts: list[str] = []
+    if candidates:
+        for part in candidates[0].get("content", {}).get("parts", []):
+            if "text" in part and not part.get("thought"):
+                text_parts.append(part["text"])
+
+    return LLMResponse(
+        content="".join(text_parts),
+        model=payload.get("modelVersion", model),
+        input_tokens=usage.get("promptTokenCount", 0),
+        output_tokens=usage.get("candidatesTokenCount", 0),
+        stop_reason=_map_finish_reason(candidates[0].get("finishReason", "") if candidates else ""),
+        raw_response=raw,
+    )
+
+
+def _parse_sse_stream(
+    response: Any,
+    model: str,
+    on_thought_signature: Callable[[str, str], None] | None = None,
+) -> Iterator[StreamEvent]:
+    """Parse Antigravity SSE response line-by-line → StreamEvents.
+
+    Each SSE line looks like::
+
+        data: {"response": {"candidates": [...], "usageMetadata": {...}}, "traceId": "..."}
+    """
+    accumulated = ""
+    input_tokens = 0
+    output_tokens = 0
+    finish_reason = ""
+
+    for raw_line in response:
+        line: str = raw_line.decode("utf-8", errors="replace").rstrip("\r\n")
+        if not line.startswith("data:"):
+            continue
+        data_str = line[5:].strip()
+        if not data_str or data_str == "[DONE]":
+            continue
+        try:
+            data: dict[str, Any] = json.loads(data_str)
+        except json.JSONDecodeError:
+            continue
+
+        # The outer envelope is {"response": {...}, "traceId": "..."}.
+        payload: dict[str, Any] = data.get("response", data)
+
+        usage = payload.get("usageMetadata", {})
+        if usage:
+            input_tokens = usage.get("promptTokenCount", input_tokens)
+            output_tokens = usage.get("candidatesTokenCount", output_tokens)
+
+        for candidate in payload.get("candidates", []):
+            fr = candidate.get("finishReason", "")
+            if fr:
+                finish_reason = fr
+
+            for part in candidate.get("content", {}).get("parts", []):
+                if "text" in part and not part.get("thought"):
+                    delta: str = part["text"]
+                    accumulated += delta
+                    yield TextDeltaEvent(content=delta, snapshot=accumulated)
+                elif "functionCall" in part:
+                    fc: dict[str, Any] = part["functionCall"]
+                    tool_use_id = fc.get("id") or str(uuid.uuid4())
+                    thought_sig = part.get("thoughtSignature", "")  # sibling of functionCall
+                    if thought_sig and on_thought_signature:
+                        on_thought_signature(tool_use_id, thought_sig)
+                    args = fc.get("args", {})
+                    if isinstance(args, str):
+                        try:
+                            args = json.loads(args)
+                        except json.JSONDecodeError:
+                            args = {}
+                    yield ToolCallEvent(
+                        tool_use_id=tool_use_id,
+                        tool_name=fc.get("name", ""),
+                        tool_input=args,
+                    )
+
+    if accumulated:
+        yield TextEndEvent(full_text=accumulated)
+    yield FinishEvent(
+        stop_reason=_map_finish_reason(finish_reason),
+        input_tokens=input_tokens,
+        output_tokens=output_tokens,
+        model=model,
+    )
+
+
+# ---------------------------------------------------------------------------
+# Provider
+# ---------------------------------------------------------------------------
+
+
+class AntigravityProvider(LLMProvider):
+    """LLM provider for Google's internal Antigravity Code Assist gateway.
+
+    No local proxy required.  Handles OAuth token refresh, Gemini-format
+    request/response conversion, and SSE streaming directly.
+    """
+
+    def __init__(self, model: str = "gemini-3-flash") -> None:
+        # Strip any provider prefix ("openai/gemini-3-flash" → "gemini-3-flash").
+        if "/" in model:
+            model = model.split("/", 1)[1]
+        self.model = model
+
+        self._access_token: str | None = None
+        self._refresh_token: str | None = None
+        self._project_id: str = _DEFAULT_PROJECT_ID
+        self._token_expires_at: float = 0.0
+        self._thought_sigs: dict[str, str] = {}  # tool_use_id → thoughtSignature
+
+        self._init_credentials()
+
+    # --- Credential management -------------------------------------------- #
+
+    def _init_credentials(self) -> None:
+        """Load credentials from the best available source."""
+        access, refresh, project_id, expires_at = _load_from_json_file()
+        if refresh:
+            self._refresh_token = refresh
+            self._project_id = project_id
+            self._access_token = access
+            self._token_expires_at = expires_at
+            return
+
+        # Fall back to IDE state DB.
+        access, refresh, expires_at = _load_from_ide_db()
+        if access:
+            self._access_token = access
+            self._refresh_token = refresh
+            self._token_expires_at = expires_at
+
+    def has_credentials(self) -> bool:
+        """Return True if any credential is available."""
+        return bool(self._access_token or self._refresh_token)
+
+    def _ensure_token(self) -> str:
+        """Return a valid access token, refreshing via OAuth if needed."""
+        if (
+            self._access_token
+            and self._token_expires_at
+            and time.time() < self._token_expires_at - _TOKEN_REFRESH_BUFFER_SECS
+        ):
+            return self._access_token
+
+        if self._refresh_token:
+            result = _do_token_refresh(self._refresh_token)
+            if result:
+                self._access_token, self._token_expires_at = result
+                return self._access_token
+
+        if self._access_token:
+            logger.warning("Using potentially stale Antigravity access token")
+            return self._access_token
+
+        raise RuntimeError(
+            "No valid Antigravity credentials. "
+            "Run: uv run python core/antigravity_auth.py auth account add"
+        )
+
+    # --- Request building -------------------------------------------------- #
+
+    def _build_body(
+        self,
+        messages: list[dict[str, Any]],
+        system: str,
+        tools: list[Tool] | None,
+        max_tokens: int,
+    ) -> dict[str, Any]:
+        contents = _to_gemini_contents(messages, self._thought_sigs)
+        inner: dict[str, Any] = {
+            "contents": contents,
+            "generationConfig": {"maxOutputTokens": max_tokens},
+        }
+        if system:
+            inner["systemInstruction"] = {"parts": [{"text": system}]}
+        if tools:
+            inner["tools"] = [
+                {
+                    "functionDeclarations": [
+                        {
+                            "name": _clean_tool_name(t.name),
+                            "description": t.description,
+                            "parameters": t.parameters
+                            or {
+                                "type": "object",
+                                "properties": {},
+                            },
+                        }
+                        for t in tools
+                    ]
+                }
+            ]
+        return {
+            "project": self._project_id,
+            "model": self.model,
+            "request": inner,
+            "requestType": "agent",
+            "userAgent": "antigravity",
+            "requestId": f"agent-{uuid.uuid4()}",
+        }
+
+    # --- HTTP transport ---------------------------------------------------- #
+
+    def _post(self, body: dict[str, Any], *, streaming: bool) -> Any:
+        """POST to the Antigravity endpoint, falling back through the endpoint list."""
+        import urllib.error  # noqa: PLC0415
+        import urllib.request  # noqa: PLC0415
+
+        token = self._ensure_token()
+        body_bytes = json.dumps(body).encode("utf-8")
+        path = (
+            "/v1internal:streamGenerateContent?alt=sse"
+            if streaming
+            else "/v1internal:generateContent"
+        )
+        headers = {
+            **_BASE_HEADERS,
+            "Authorization": f"Bearer {token}",
+            "Content-Type": "application/json",
+        }
+        if streaming:
+            headers["Accept"] = "text/event-stream"
+
+        last_exc: Exception | None = None
+        for base_url in _ENDPOINTS:
+            url = f"{base_url}{path}"
+            req = urllib.request.Request(url, data=body_bytes, headers=headers, method="POST")
+            try:
+                return urllib.request.urlopen(req, timeout=120)  # noqa: S310
+            except urllib.error.HTTPError as exc:
+                if exc.code in (401, 403) and self._refresh_token:
+                    # Token rejected — refresh once and retry this endpoint.
+                    result = _do_token_refresh(self._refresh_token)
+                    if result:
+                        self._access_token, self._token_expires_at = result
+                        headers["Authorization"] = f"Bearer {self._access_token}"
+                        req2 = urllib.request.Request(
+                            url, data=body_bytes, headers=headers, method="POST"
+                        )
+                        try:
+                            return urllib.request.urlopen(req2, timeout=120)  # noqa: S310
+                        except urllib.error.HTTPError as exc2:
+                            last_exc = exc2
+                            continue
+                    last_exc = exc
+                    continue
+                elif exc.code >= 500:
+                    last_exc = exc
+                    continue
+                # Include the API response body in the exception for easier debugging.
+                try:
+                    err_body = exc.read().decode("utf-8", errors="replace")
+                except Exception:
+                    err_body = "(unreadable)"
+                raise RuntimeError(f"Antigravity HTTP {exc.code} from {url}: {err_body}") from exc
+            except (urllib.error.URLError, OSError) as exc:
+                last_exc = exc
+                continue
+
+        raise RuntimeError(
+            f"All Antigravity endpoints failed. Last error: {last_exc}"
+        ) from last_exc
+
+    # --- LLMProvider interface --------------------------------------------- #
+
+    def complete(
+        self,
+        messages: list[dict[str, Any]],
+        system: str = "",
+        tools: list[Tool] | None = None,
+        max_tokens: int = 1024,
+        response_format: dict[str, Any] | None = None,
+        json_mode: bool = False,
+        max_retries: int | None = None,
+    ) -> LLMResponse:
+        if json_mode:
+            suffix = "\n\nPlease respond with a valid JSON object."
+            system = (system + suffix) if system else suffix.strip()
+
+        body = self._build_body(messages, system, tools, max_tokens)
+        resp = self._post(body, streaming=False)
+        return _parse_complete_response(json.loads(resp.read()), self.model)
+
+    async def stream(
+        self,
+        messages: list[dict[str, Any]],
+        system: str = "",
+        tools: list[Tool] | None = None,
+        max_tokens: int = 4096,
+    ) -> AsyncIterator[StreamEvent]:
+        import asyncio  # noqa: PLC0415
+        import concurrent.futures  # noqa: PLC0415
+
+        loop = asyncio.get_running_loop()
+        queue: asyncio.Queue[StreamEvent | None] = asyncio.Queue()
+
+        def _blocking_work() -> None:
+            try:
+                body = self._build_body(messages, system, tools, max_tokens)
+                http_resp = self._post(body, streaming=True)
+                for event in _parse_sse_stream(
+                    http_resp, self.model, self._thought_sigs.__setitem__
+                ):
+                    loop.call_soon_threadsafe(queue.put_nowait, event)
+            except Exception as exc:
+                logger.error("Antigravity stream error: %s", exc)
+                loop.call_soon_threadsafe(queue.put_nowait, StreamErrorEvent(error=str(exc)))
+            finally:
+                loop.call_soon_threadsafe(queue.put_nowait, None)  # sentinel
+
+        executor = concurrent.futures.ThreadPoolExecutor(max_workers=1)
+        fut = loop.run_in_executor(executor, _blocking_work)
+        try:
+            while True:
+                event = await queue.get()
+                if event is None:
+                    break
+                yield event
+        finally:
+            await fut
+            executor.shutdown(wait=False)

core/framework/llm/capabilities.py

@@ -0,0 +1,106 @@
+"""Model capability checks for LLM providers.
+
+Vision support rules are derived from official vendor documentation:
+- ZAI (z.ai): docs.z.ai/guides/vlm — GLM-4.6V variants are vision; GLM-5/4.6/4.7 are text-only
+- MiniMax: platform.minimax.io/docs — minimax-vl-01 is vision; M2.x are text-only
+- DeepSeek: api-docs.deepseek.com — deepseek-vl2 is vision; chat/reasoner are text-only
+- Cerebras: inference-docs.cerebras.ai — no vision models at all
+- Groq: console.groq.com/docs/vision — vision capable; treat as supported by default
+- Ollama/LM Studio/vLLM/llama.cpp: local runners denied by default; model names
+  don't reliably indicate vision support, so users must configure explicitly
+"""
+
+from __future__ import annotations
+
+
+def _model_name(model: str) -> str:
+    """Return the bare model name after stripping any 'provider/' prefix."""
+    if "/" in model:
+        return model.split("/", 1)[1]
+    return model
+
+
+# Step 1: explicit vision allow-list — these always support images regardless
+# of what the provider-level rules say.  Checked first so that e.g. glm-4.6v
+# is allowed even though glm-4.6 is denied.
+_VISION_ALLOW_BARE_PREFIXES: tuple[str, ...] = (
+    # ZAI/GLM vision models (docs.z.ai/guides/vlm)
+    "glm-4v",  # GLM-4V series (legacy)
+    "glm-4.6v",  # GLM-4.6V, GLM-4.6V-flash, GLM-4.6V-flashx
+    # DeepSeek vision models
+    "deepseek-vl",  # deepseek-vl2, deepseek-vl2-small, deepseek-vl2-tiny
+    # MiniMax vision model
+    "minimax-vl",  # minimax-vl-01
+)
+
+# Step 2: provider-level deny — every model from this provider is text-only.
+_TEXT_ONLY_PROVIDER_PREFIXES: tuple[str, ...] = (
+    # Cerebras: inference-docs.cerebras.ai lists only text models
+    "cerebras/",
+    # Local runners: model names don't reliably indicate vision support
+    "ollama/",
+    "ollama_chat/",
+    "lm_studio/",
+    "vllm/",
+    "llamacpp/",
+)
+
+# Step 3: per-model deny — text-only models within otherwise mixed providers.
+# Matched against the bare model name (provider prefix stripped, lower-cased).
+# The vision allow-list above is checked first, so vision variants of the same
+# family are already handled before these deny patterns are reached.
+_TEXT_ONLY_MODEL_BARE_PREFIXES: tuple[str, ...] = (
+    # --- ZAI / GLM family ---
+    # text-only: glm-5, glm-4.6, glm-4.7, glm-4.5, zai-glm-*
+    # vision:    glm-4v, glm-4.6v (caught by allow-list above)
+    "glm-5",
+    "glm-4.6",  # bare glm-4.6 is text-only; glm-4.6v is caught by allow-list
+    "glm-4.7",
+    "glm-4.5",
+    "zai-glm",
+    # --- DeepSeek ---
+    # text-only: deepseek-chat, deepseek-coder, deepseek-reasoner
+    # vision:    deepseek-vl2 (caught by allow-list above)
+    # Note: LiteLLM's deepseek handler may flatten content lists for some models;
+    # VL models are allowed through and rely on LiteLLM's native VL support.
+    "deepseek-chat",
+    "deepseek-coder",
+    "deepseek-reasoner",
+    # --- MiniMax ---
+    # text-only: minimax-m2.*, minimax-text-*, abab* (legacy)
+    # vision:    minimax-vl-01 (caught by allow-list above)
+    "minimax-m2",
+    "minimax-text",
+    "abab",
+)
+
+
+def supports_image_tool_results(model: str) -> bool:
+    """Return whether *model* can receive image content in messages.
+
+    Used to gate both user-message images and tool-result image blocks.
+
+    Logic (checked in order):
+    1. Vision allow-list  → True  (known vision model, skip all denies)
+    2. Provider deny      → False (entire provider is text-only)
+    3. Model deny         → False (specific text-only model within a mixed provider)
+    4. Default            → True  (assume capable; unknown providers and models)
+    """
+    model_lower = model.lower()
+    bare = _model_name(model_lower)
+
+    # 1. Explicit vision allow — takes priority over all denies
+    if any(bare.startswith(p) for p in _VISION_ALLOW_BARE_PREFIXES):
+        return True
+
+    # 2. Provider-level deny (all models from this provider are text-only)
+    if any(model_lower.startswith(p) for p in _TEXT_ONLY_PROVIDER_PREFIXES):
+        return False
+
+    # 3. Per-model deny (text-only variants within mixed-capability families)
+    if any(bare.startswith(p) for p in _TEXT_ONLY_MODEL_BARE_PREFIXES):
+        return False
+
+    # 5. Default: assume vision capable
+    #    Covers: OpenAI, Anthropic, Google, Mistral, Kimi, and other hosted providers
+    return True

core/framework/llm/key_pool.py

@@ -0,0 +1,101 @@
+"""Thread-safe API key pool with round-robin rotation and health tracking.
+
+When multiple API keys are configured, the pool rotates through them on each
+request.  Keys that hit rate limits are temporarily cooled-down so the next
+call automatically uses a healthy key -- no sleep required.
+"""
+
+from __future__ import annotations
+
+import logging
+import threading
+import time
+from dataclasses import dataclass
+
+logger = logging.getLogger(__name__)
+
+
+@dataclass
+class KeyHealth:
+    """Per-key health counters."""
+
+    rate_limited_until: float = 0.0  # monotonic timestamp
+    consecutive_errors: int = 0
+    total_requests: int = 0
+    total_successes: int = 0
+
+
+class KeyPool:
+    """Round-robin key pool with health tracking.
+
+    Thread-safe: all mutations protected by a lock so concurrent LLM calls
+    (e.g. parallel tool execution in EventLoopNode) don't race.
+    """
+
+    def __init__(self, keys: list[str]) -> None:
+        if not keys:
+            raise ValueError("KeyPool requires at least one key")
+        self._keys = list(keys)
+        self._index = 0
+        self._health: dict[str, KeyHealth] = {k: KeyHealth() for k in keys}
+        self._lock = threading.Lock()
+
+    @property
+    def size(self) -> int:
+        return len(self._keys)
+
+    def get_key(self) -> str:
+        """Return the next healthy key (round-robin).
+
+        If every key is currently rate-limited, returns the one whose cooldown
+        expires soonest so the caller can proceed with minimal delay.
+        """
+        with self._lock:
+            now = time.monotonic()
+            for _ in range(len(self._keys)):
+                key = self._keys[self._index]
+                self._index = (self._index + 1) % len(self._keys)
+                health = self._health[key]
+                if health.rate_limited_until <= now:
+                    health.total_requests += 1
+                    return key
+            # All rate-limited -- pick the one that expires soonest.
+            soonest = min(self._keys, key=lambda k: self._health[k].rate_limited_until)
+            self._health[soonest].total_requests += 1
+            return soonest
+
+    def mark_rate_limited(self, key: str, retry_after: float = 60.0) -> None:
+        """Mark *key* as rate-limited for *retry_after* seconds."""
+        with self._lock:
+            health = self._health.get(key)
+            if health:
+                health.rate_limited_until = time.monotonic() + retry_after
+                health.consecutive_errors += 1
+                logger.info(
+                    "[key-pool] Key ...%s rate-limited for %.0fs (errors=%d)",
+                    key[-6:],
+                    retry_after,
+                    health.consecutive_errors,
+                )
+
+    def mark_success(self, key: str) -> None:
+        """Record a successful call on *key*."""
+        with self._lock:
+            health = self._health.get(key)
+            if health:
+                health.consecutive_errors = 0
+                health.total_successes += 1
+
+    def get_stats(self) -> dict[str, dict]:
+        """Return health stats keyed by the last 6 chars of each key."""
+        with self._lock:
+            now = time.monotonic()
+            return {
+                f"...{k[-6:]}": {
+                    "healthy": self._health[k].rate_limited_until <= now,
+                    "requests": self._health[k].total_requests,
+                    "successes": self._health[k].total_successes,
+                    "consecutive_errors": self._health[k].consecutive_errors,
+                }
+                for k in self._keys
+            }

core/framework/llm/provider.py

@@ -45,6 +45,8 @@ class ToolResult:
     tool_use_id: str
     content: str
     is_error: bool = False
+    image_content: list[dict[str, Any]] | None = None
+    is_skill_content: bool = False  # AS-10: marks activated skill body, protected from pruning
 
 
 class LLMProvider(ABC):

core/framework/llm/stream_events.py

@@ -71,6 +71,7 @@ class FinishEvent:
     stop_reason: str = ""
     input_tokens: int = 0
     output_tokens: int = 0
+    cached_tokens: int = 0
     model: str = ""
 
 

core/framework/loader/__init__.py

@@ -0,0 +1,4 @@
+"""Loader layer -- agent loading from disk (JSON config, MCP, credentials)."""
+
+from framework.loader.agent_loader import AgentLoader  # noqa: F401
+from framework.loader.tool_registry import ToolRegistry  # noqa: F401