lint fixes

Sample agent, micro-fix: remove dependency of brave search

.agent/mcp_config.json

@@ -0,0 +1,9 @@
+{
+  "mcpServers": {
+    "agent-builder": {
+      "command": "uv",
+      "args": ["run", "--directory", "core", "-m", "framework.mcp.agent_builder_server"],
+      "disabled": false
+    }
+  }
+}

.agent/skills/hive

@@ -0,0 +1 @@
+../../.claude/skills/hive

.agent/skills/hive-concepts

@@ -0,0 +1 @@
+../../.claude/skills/hive-concepts

.agent/skills/hive-create

@@ -0,0 +1 @@
+../../.claude/skills/hive-create

.agent/skills/hive-credentials

@@ -0,0 +1 @@
+../../.claude/skills/hive-credentials

.agent/skills/hive-patterns

@@ -0,0 +1 @@
+../../.claude/skills/hive-patterns

.agent/skills/hive-test

@@ -0,0 +1 @@
+../../.claude/skills/hive-test

.agent/workflows/hive-concepts.md

@@ -0,0 +1,5 @@
+---
+description: hive-concepts
+---
+
+use hive-concepts skill

.agent/workflows/hive-create.md

@@ -0,0 +1,5 @@
+---
+description: hive-create
+---
+
+use hive-create skill

.agent/workflows/hive-credentials.md

@@ -0,0 +1,5 @@
+---
+description: hive-credentials
+---
+
+use hive-credentials skill

.agent/workflows/hive-patterns.md

@@ -0,0 +1,5 @@
+---
+description: hive-patterns
+---
+
+use hive-patterns skill

.agent/workflows/hive-test.md

@@ -0,0 +1,5 @@
+---
+description: hive-test
+---
+
+use hive-test skill

.agent/workflows/hive.md

@@ -0,0 +1,5 @@
+---
+description: hive
+---
+
+use hive skill

.agents/skills/hive

@@ -0,0 +1 @@
+../../.claude/skills/hive

.agents/skills/hive-concepts

@@ -0,0 +1 @@
+../../.claude/skills/hive-concepts

.agents/skills/hive-create

@@ -0,0 +1 @@
+../../.claude/skills/hive-create

.agents/skills/hive-credentials

@@ -0,0 +1 @@
+../../.claude/skills/hive-credentials

.agents/skills/hive-patterns

@@ -0,0 +1 @@
+../../.claude/skills/hive-patterns

.agents/skills/hive-test

@@ -0,0 +1 @@
+../../.claude/skills/hive-test

.claude/settings.local.json.example

@@ -0,0 +1,34 @@
+{
+  "permissions": {
+    "allow": [
+      "mcp__agent-builder__create_session",
+      "mcp__agent-builder__set_goal",
+      "mcp__agent-builder__add_node",
+      "mcp__agent-builder__add_edge",
+      "mcp__agent-builder__configure_loop",
+      "mcp__agent-builder__add_mcp_server",
+      "mcp__agent-builder__validate_graph",
+      "mcp__agent-builder__export_graph",
+      "mcp__agent-builder__load_session_by_id",
+      "Bash(git status:*)",
+      "Bash(gh run view:*)",
+      "Bash(uv run:*)",
+      "Bash(env:*)",
+      "mcp__agent-builder__test_node",
+      "mcp__agent-builder__list_mcp_tools",
+      "Bash(python -m py_compile:*)",
+      "Bash(python -m pytest:*)",
+      "Bash(source:*)",
+      "mcp__agent-builder__update_node",
+      "mcp__agent-builder__check_missing_credentials",
+      "mcp__agent-builder__list_stored_credentials",
+      "Bash(find:*)",
+      "mcp__agent-builder__run_tests",
+      "Bash(PYTHONPATH=core:exports:tools/src uv run pytest:*)",
+      "mcp__agent-builder__list_agent_sessions",
+      "mcp__agent-builder__generate_constraint_tests",
+      "mcp__agent-builder__generate_success_tests"
+    ]
+  },
+  "enabledMcpjsonServers": ["agent-builder", "tools"]
+}

.claude/skills/hive-create/SKILL.md

@@ -553,6 +553,26 @@ AskUserQuestion(questions=[{
 - condition_expr (Python expression, only if conditional)
 - priority (positive = forward, negative = feedback/loop-back)
 
+**DETERMINE the graph lifecycle.** Not every agent needs a terminal node:
+
+| Pattern | `terminal_nodes` | When to Use |
+|---------|-------------------|-------------|
+| **Linear (finish)** | `["last-node"]` | Agent completes a task and exits (batch processing, one-shot generation) |
+| **Forever-alive (loop)** | `[]` (empty) | Agent stays alive for continuous interaction (research assistant, personal assistant, monitoring) |
+
+**Forever-alive pattern:** The deep_research_agent example uses `terminal_nodes=[]`. Every leaf node has edges that loop back to earlier nodes, creating a perpetual session. The agent only stops when the user explicitly exits. This is the preferred pattern for interactive, multi-turn agents.
+
+**Key design rules for forever-alive graphs:**
+- Every node must have at least one outgoing edge (no dead ends)
+- Client-facing nodes block for user input — these are the natural "pause points"
+- The user controls when to stop, not the graph
+- Sessions accumulate memory across loops — plan for conversation compaction
+- Use `conversation_mode="continuous"` to preserve conversation history across node transitions
+- `max_iterations` should be set high (e.g., 100) since the agent is designed to run indefinitely
+- The agent will NOT enter a "completed" execution state — this is intentional, not a bug
+
+**Ask the user** which lifecycle pattern fits their agent. Default to forever-alive for interactive agents, linear for batch/one-shot tasks.
+
 **RENDER the complete graph as ASCII art.** Make it large and clear — the user needs to see and understand the full workflow at a glance.
 
 **IMPORTANT: Make the ASCII art BIG and READABLE.** Use a box-and-arrow style with generous spacing. Do NOT make it tiny or compressed. Example format:
@@ -669,6 +689,7 @@ AskUserQuestion(questions=[{
 |------|---------------|
 | `config.py` | `AgentMetadata.name` — the display name shown in TUI agent selection |
 | `config.py` | `AgentMetadata.description` — agent description |
+| `config.py` | `AgentMetadata.intro_message` — greeting shown to user when TUI loads |
 | `agent.py` | Module docstring (line 1) |
 | `agent.py` | `class OldNameAgent:` → `class NewNameAgent:` |
 | `agent.py` | `GraphSpec(id="old-name-graph")` → `GraphSpec(id="new-name-graph")` — shown in TUI status bar |
@@ -735,7 +756,7 @@ mcp__agent-builder__export_graph()
 
 **THEN write the Python package files** using the exported data. Create these files in `exports/AGENT_NAME/`:
 
-1. `config.py` - Runtime configuration with model settings
+1. `config.py` - Runtime configuration with model settings and `AgentMetadata` (including `intro_message` — the greeting shown when TUI loads)
 2. `nodes/__init__.py` - All NodeSpec definitions
 3. `agent.py` - Goal, edges, graph config, and agent class
 4. `__init__.py` - Package exports
@@ -911,6 +932,46 @@ result = await executor.execute(graph=graph, goal=goal, input_data=input_data)
 
 ---
 
+## REFERENCE: Graph Lifecycle & Conversation Memory
+
+### Terminal vs Forever-Alive Graphs
+
+Agents have two lifecycle patterns:
+
+**Linear (terminal) graphs** have `terminal_nodes=["last-node"]`. Execution ends when the terminal node completes. The session enters a "completed" state. Use for batch processing, one-shot generation, and fire-and-forget tasks.
+
+**Forever-alive graphs** have `terminal_nodes=[]` (empty). Every node has at least one outgoing edge — the graph loops indefinitely. The session **never enters a "completed" state** — this is intentional. The agent stays alive until the user explicitly exits. Use for interactive assistants, research tools, and any agent where the user drives the conversation.
+
+The deep_research_agent example demonstrates this: `report` loops back to either `research` (dig deeper) or `intake` (new topic). The agent is a persistent, interactive assistant.
+
+### Continuous Conversation Mode
+
+When `conversation_mode="continuous"` is set on the GraphSpec, the framework preserves a **single conversation thread** across all node transitions:
+
+**What the framework does automatically:**
+- **Inherits conversation**: Same message history carries forward to the next node
+- **Composes layered system prompts**: Identity (agent-level) + Narrative (auto-generated state summary) + Focus (per-node instructions)
+- **Inserts transition markers**: At each node boundary, a "State of the World" message showing completed phases, current memory, and available data files
+- **Accumulates tools**: Once a tool becomes available, it stays available in subsequent nodes
+- **Compacts opportunistically**: At phase transitions, old tool results are pruned to stay within token budget
+
+**What this means for agent builders:**
+- Nodes don't need to re-explain context — the conversation carries it forward
+- Output keys from earlier nodes are available in memory for edge conditions and later nodes
+- For forever-alive agents, conversation memory persists across the entire session lifetime
+- Plan for compaction: very long sessions will have older tool results summarized automatically
+
+**When to use continuous mode:**
+- Interactive agents with client-facing nodes (always)
+- Multi-phase workflows where context matters across phases
+- Forever-alive agents that loop indefinitely
+
+**When NOT to use continuous mode:**
+- Embarrassingly parallel fan-out nodes (each branch should be independent)
+- Stateless utility agents that process items independently
+
+---
+
 ## REFERENCE: Framework Capabilities for Qualification
 
 Use this reference during STEP 2 to give accurate, honest assessments.
@@ -943,7 +1004,7 @@ Use this reference during STEP 2 to give accurate, honest assessments.
 
 | Use Case | Why It's Problematic | Alternative |
 |----------|---------------------|-------------|
-| Long-running daemons | Framework is request-response, not persistent | External scheduler + agent |
+| Persistent background daemons (no user) | Forever-alive graphs need a user at client-facing nodes; no autonomous background polling without user | External scheduler triggering agent runs |
 | Sub-second responses | LLM latency is inherent | Traditional code, no LLM |
 | Processing millions of items | Context windows and rate limits | Batch processing + sampling |
 | Real-time streaming data | No built-in pub/sub or streaming input | Custom MCP server + agent |
@@ -978,3 +1039,6 @@ Use this reference during STEP 2 to give accurate, honest assessments.
 11. **Adding framework gating for LLM behavior** - Fix prompts or use judges, not ad-hoc code
 12. **Writing code before user approves the graph** - Always get approval on goal, nodes, and graph BEFORE writing any agent code
 13. **Wrong mcp_servers.json format** - Use flat format (no `"mcpServers"` wrapper), `cwd` must be `"../../tools"`, and `command` must be `"uv"` with args `["run", "python", ...]`
+14. **Assuming all agents need terminal nodes** - Interactive agents often work best with `terminal_nodes=[]` (forever-alive pattern). The agent never enters "completed" state — this is intentional. Only batch/one-shot agents need terminal nodes
+15. **Creating dead-end nodes in forever-alive graphs** - Every node must have at least one outgoing edge. A node with no outgoing edges will cause execution to end unexpectedly, breaking the forever-alive loop
+16. **Not using continuous conversation mode for interactive agents** - Multi-phase interactive agents should use `conversation_mode="continuous"` to preserve context across node transitions. Without it, each node starts with a blank conversation and loses all prior context

.claude/skills/hive-create/examples/deep_research_agent/agent.py

@@ -1,12 +1,15 @@
 """Agent graph construction for Deep Research Agent."""
 
+from pathlib import Path
+
 from framework.graph import EdgeSpec, EdgeCondition, Goal, SuccessCriterion, Constraint
 from framework.graph.edge import GraphSpec
-from framework.graph.executor import ExecutionResult, GraphExecutor
-from framework.runtime.event_bus import EventBus
-from framework.runtime.core import Runtime
+from framework.graph.executor import ExecutionResult
+from framework.graph.checkpoint_config import CheckpointConfig
 from framework.llm import LiteLLMProvider
 from framework.runner.tool_registry import ToolRegistry
+from framework.runtime.agent_runtime import AgentRuntime, create_agent_runtime
+from framework.runtime.execution_stream import EntryPointSpec
 
 from .config import default_config, metadata
 from .nodes import (
@@ -120,13 +123,31 @@ edges = [
         condition_expr="needs_more_research == False",
         priority=2,
     ),
+    # report -> research (user wants deeper research on current topic)
+    EdgeSpec(
+        id="report-to-research",
+        source="report",
+        target="research",
+        condition=EdgeCondition.CONDITIONAL,
+        condition_expr="str(next_action).lower() == 'more_research'",
+        priority=2,
+    ),
+    # report -> intake (user wants a new topic — default when not more_research)
+    EdgeSpec(
+        id="report-to-intake",
+        source="report",
+        target="intake",
+        condition=EdgeCondition.CONDITIONAL,
+        condition_expr="str(next_action).lower() != 'more_research'",
+        priority=1,
+    ),
 ]
 
 # Graph configuration
 entry_node = "intake"
 entry_points = {"start": "intake"}
 pause_nodes = []
-terminal_nodes = ["report"]
+terminal_nodes = []
 
 
 class DeepResearchAgent:
@@ -136,6 +157,12 @@ class DeepResearchAgent:
     Flow: intake -> research -> review -> report
                       ^           |
                       +-- feedback loop (if user wants more)
+
+    Uses AgentRuntime for proper session management:
+    - Session-scoped storage (sessions/{session_id}/)
+    - Checkpointing for resume capability
+    - Runtime logging
+    - Data folder for save_data/load_data
     """
 
     def __init__(self, config=None):
@@ -147,10 +174,10 @@ class DeepResearchAgent:
         self.entry_points = entry_points
         self.pause_nodes = pause_nodes
         self.terminal_nodes = terminal_nodes
-        self._executor: GraphExecutor | None = None
         self._graph: GraphSpec | None = None
-        self._event_bus: EventBus | None = None
+        self._agent_runtime: AgentRuntime | None = None
         self._tool_registry: ToolRegistry | None = None
+        self._storage_path: Path | None = None
 
     def _build_graph(self) -> GraphSpec:
         """Build the GraphSpec."""
@@ -171,16 +198,20 @@ class DeepResearchAgent:
                 "max_tool_calls_per_turn": 20,
                 "max_history_tokens": 32000,
             },
+            conversation_mode="continuous",
+            identity_prompt=(
+                "You are a rigorous research agent. You search for information "
+                "from diverse, authoritative sources, analyze findings critically, "
+                "and produce well-cited reports. You never fabricate information — "
+                "every claim must trace back to a source you actually retrieved."
+            ),
         )
 
-    def _setup(self, mock_mode=False) -> GraphExecutor:
-        """Set up the executor with all components."""
-        from pathlib import Path
+    def _setup(self, mock_mode=False) -> None:
+        """Set up the agent runtime with sessions, checkpoints, and logging."""
+        self._storage_path = Path.home() / ".hive" / "agents" / "deep_research_agent"
+        self._storage_path.mkdir(parents=True, exist_ok=True)
 
-        storage_path = Path.home() / ".hive" / "agents" / "deep_research_agent"
-        storage_path.mkdir(parents=True, exist_ok=True)
-
-        self._event_bus = EventBus()
         self._tool_registry = ToolRegistry()
 
         mcp_config_path = Path(__file__).parent / "mcp_servers.json"
@@ -199,47 +230,63 @@ class DeepResearchAgent:
         tools = list(self._tool_registry.get_tools().values())
 
         self._graph = self._build_graph()
-        runtime = Runtime(storage_path)
 
-        self._executor = GraphExecutor(
-            runtime=runtime,
+        checkpoint_config = CheckpointConfig(
+            enabled=True,
+            checkpoint_on_node_start=False,
+            checkpoint_on_node_complete=True,
+            checkpoint_max_age_days=7,
+            async_checkpoint=True,
+        )
+
+        entry_point_specs = [
+            EntryPointSpec(
+                id="default",
+                name="Default",
+                entry_node=self.entry_node,
+                trigger_type="manual",
+                isolation_level="shared",
+            )
+        ]
+
+        self._agent_runtime = create_agent_runtime(
+            graph=self._graph,
+            goal=self.goal,
+            storage_path=self._storage_path,
+            entry_points=entry_point_specs,
             llm=llm,
             tools=tools,
             tool_executor=tool_executor,
-            event_bus=self._event_bus,
-            storage_path=storage_path,
-            loop_config=self._graph.loop_config,
+            checkpoint_config=checkpoint_config,
         )
 
-        return self._executor
-
     async def start(self, mock_mode=False) -> None:
-        """Set up the agent (initialize executor and tools)."""
-        if self._executor is None:
+        """Set up and start the agent runtime."""
+        if self._agent_runtime is None:
             self._setup(mock_mode=mock_mode)
+        if not self._agent_runtime.is_running:
+            await self._agent_runtime.start()
 
     async def stop(self) -> None:
-        """Clean up resources."""
-        self._executor = None
-        self._event_bus = None
+        """Stop the agent runtime and clean up."""
+        if self._agent_runtime and self._agent_runtime.is_running:
+            await self._agent_runtime.stop()
+        self._agent_runtime = None
 
     async def trigger_and_wait(
         self,
-        entry_point: str,
-        input_data: dict,
+        entry_point: str = "default",
+        input_data: dict | None = None,
         timeout: float | None = None,
         session_state: dict | None = None,
     ) -> ExecutionResult | None:
         """Execute the graph and wait for completion."""
-        if self._executor is None:
+        if self._agent_runtime is None:
             raise RuntimeError("Agent not started. Call start() first.")
-        if self._graph is None:
-            raise RuntimeError("Graph not built. Call start() first.")
 
-        return await self._executor.execute(
-            graph=self._graph,
-            goal=self.goal,
-            input_data=input_data,
+        return await self._agent_runtime.trigger_and_wait(
+            entry_point_id=entry_point,
+            input_data=input_data or {},
             session_state=session_state,
         )
 
@@ -250,7 +297,7 @@ class DeepResearchAgent:
         await self.start(mock_mode=mock_mode)
         try:
             result = await self.trigger_and_wait(
-                "start", context, session_state=session_state
+                "default", context, session_state=session_state
             )
             return result or ExecutionResult(success=False, error="Execution timeout")
         finally:

.claude/skills/hive-create/examples/deep_research_agent/config.py

@@ -16,6 +16,11 @@ class AgentMetadata:
         "multi-source search, quality evaluation, and synthesis - with TUI conversation "
         "at key checkpoints for user guidance and feedback."
     )
+    intro_message: str = (
+        "Hi! I'm your deep research assistant. Tell me a topic and I'll investigate it "
+        "thoroughly — searching multiple sources, evaluating quality, and synthesizing "
+        "a comprehensive report. What would you like me to research?"
+    )
 
 
 metadata = AgentMetadata()

.claude/skills/hive-create/examples/deep_research_agent/nodes/__init__.py

@@ -10,8 +10,13 @@ intake_node = NodeSpec(
     description="Discuss the research topic with the user, clarify scope, and confirm direction",
     node_type="event_loop",
     client_facing=True,
+    max_node_visits=0,
     input_keys=["topic"],
     output_keys=["research_brief"],
+    success_criteria=(
+        "The research brief is specific and actionable: it states the topic, "
+        "the key questions to answer, the desired scope, and depth."
+    ),
     system_prompt="""\
 You are a research intake specialist. The user wants to research a topic.
 Have a brief conversation to clarify what they need.
@@ -38,10 +43,14 @@ research_node = NodeSpec(
     name="Research",
     description="Search the web, fetch source content, and compile findings",
     node_type="event_loop",
-    max_node_visits=3,
+    max_node_visits=0,
     input_keys=["research_brief", "feedback"],
     output_keys=["findings", "sources", "gaps"],
     nullable_output_keys=["feedback"],
+    success_criteria=(
+        "Findings reference at least 3 distinct sources with URLs. "
+        "Key claims are substantiated by fetched content, not generated."
+    ),
     system_prompt="""\
 You are a research agent. Given a research brief, find and analyze sources.
 
@@ -56,18 +65,26 @@ Work in phases:
    and any contradictions between sources.
 
 Important:
-- Work in batches of 3-4 tool calls at a time to manage context
+- Work in batches of 3-4 tool calls at a time — never more than 10 per turn
 - After each batch, assess whether you have enough material
 - Prefer quality over quantity — 5 good sources beat 15 thin ones
 - Track which URL each finding comes from (you'll need citations later)
+- Call set_output for each key in a SEPARATE turn (not in the same turn as other tool calls)
 
-When done, use set_output:
+When done, use set_output (one key at a time, separate turns):
 - set_output("findings", "Structured summary: key findings with source URLs for each claim. \
 Include themes, contradictions, and confidence levels.")
 - set_output("sources", [{"url": "...", "title": "...", "summary": "..."}])
 - set_output("gaps", "What aspects of the research brief are NOT well-covered yet, if any.")
 """,
-    tools=["web_search", "web_scrape", "load_data", "save_data", "list_data_files"],
+    tools=[
+        "web_search",
+        "web_scrape",
+        "load_data",
+        "save_data",
+        "append_data",
+        "list_data_files",
+    ],
 )
 
 # Node 3: Review (client-facing)
@@ -78,9 +95,13 @@ review_node = NodeSpec(
     description="Present findings to user and decide whether to research more or write the report",
     node_type="event_loop",
     client_facing=True,
-    max_node_visits=3,
+    max_node_visits=0,
     input_keys=["findings", "sources", "gaps", "research_brief"],
     output_keys=["needs_more_research", "feedback"],
+    success_criteria=(
+        "The user has been presented with findings and has explicitly indicated "
+        "whether they want more research or are ready for the report."
+    ),
     system_prompt="""\
 Present the research findings to the user clearly and concisely.
 
@@ -109,49 +130,70 @@ report_node = NodeSpec(
     description="Write a cited HTML report from the findings and present it to the user",
     node_type="event_loop",
     client_facing=True,
+    max_node_visits=0,
     input_keys=["findings", "sources", "research_brief"],
-    output_keys=["delivery_status"],
+    output_keys=["delivery_status", "next_action"],
+    success_criteria=(
+        "An HTML report has been saved, the file link has been presented to the user, "
+        "and the user has indicated what they want to do next."
+    ),
     system_prompt="""\
-Write a comprehensive research report as an HTML file and present it to the user.
+Write a research report as an HTML file and present it to the user.
 
-**STEP 1 — Write the HTML report (tool calls, NO text to user yet):**
+IMPORTANT: save_data requires TWO separate arguments: filename and data.
+Call it like: save_data(filename="report.html", data="<html>...</html>")
+Do NOT use _raw, do NOT nest arguments inside a JSON string.
 
-1. Compose a complete, self-contained HTML document with embedded CSS styling.
-   Use a clean, readable design: max-width container, pleasant typography,
-   numbered citation links, a table of contents, and a references section.
+**STEP 1 — Write and save the HTML report (tool calls, NO text to user yet):**
 
-   Report structure inside the HTML:
-   - Title & date
-   - Executive Summary (2-3 paragraphs)
-   - Table of Contents
-   - Findings (organized by theme, with [n] citation links)
-   - Analysis (synthesis, implications, areas of debate)
-   - Conclusion (key takeaways, confidence assessment)
-   - References (numbered list with clickable URLs)
+Build a clean HTML document. Keep the HTML concise — aim for clarity over length.
+Use minimal embedded CSS (a few lines of style, not a full framework).
 
-   Requirements:
-   - Every factual claim must cite its source with [n] notation
-   - Be objective — present multiple viewpoints where sources disagree
-   - Distinguish well-supported conclusions from speculation
-   - Answer the original research questions from the brief
+Report structure:
+- Title & date
+- Executive Summary (2-3 paragraphs)
+- Key Findings (organized by theme, with [n] citation links)
+- Analysis (synthesis, implications)
+- Conclusion (key takeaways)
+- References (numbered list with clickable URLs)
 
-2. Save the HTML file:
-   save_data(filename="report.html", data=<your_html>)
+Requirements:
+- Every factual claim must cite its source with [n] notation
+- Be objective — present multiple viewpoints where sources disagree
+- Answer the original research questions from the brief
 
-3. Get the clickable link:
-   serve_file_to_user(filename="report.html", label="Research Report")
+Save the HTML:
+  save_data(filename="report.html", data="<html>...</html>")
+
+Then get the clickable link:
+  serve_file_to_user(filename="report.html", label="Research Report")
+
+If save_data fails, simplify and shorten the HTML, then retry.
 
 **STEP 2 — Present the link to the user (text only, NO tool calls):**
 
 Tell the user the report is ready and include the file:// URI from
 serve_file_to_user so they can click it to open. Give a brief summary
-of what the report covers. Ask if they have questions.
+of what the report covers. Ask if they have questions or want to continue.
 
 **STEP 3 — After the user responds:**
-- Answer follow-up questions from the research material
-- When the user is satisfied: set_output("delivery_status", "completed")
+- Answer any follow-up questions from the research material
+- When the user is ready to move on, ask what they'd like to do next:
+  - Research a new topic?
+  - Dig deeper into the current topic?
+- Then call set_output:
+  - set_output("delivery_status", "completed")
+  - set_output("next_action", "new_topic")       — if they want a new topic
+  - set_output("next_action", "more_research")   — if they want deeper research
 """,
-    tools=["save_data", "serve_file_to_user", "load_data", "list_data_files"],
+    tools=[
+        "save_data",
+        "append_data",
+        "edit_data",
+        "serve_file_to_user",
+        "load_data",
+        "list_data_files",
+    ],
 )
 
 __all__ = [

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

@@ -141,6 +141,12 @@ for f in ~/.zshrc ~/.bashrc ~/.profile; do [ -f "$f" ] && grep -q 'HIVE_CREDENTI
 - **In shell config but NOT in current session** — run `source ~/.zshrc` (or `~/.bashrc`) first, then proceed
 - **Not set anywhere** — `EncryptedFileStorage` will auto-generate one. After storing, tell the user to persist it: `export HIVE_CREDENTIAL_KEY="{generated_key}"` in their shell profile
 
+> **⚠️ IMPORTANT: After adding `HIVE_CREDENTIAL_KEY` to the user's shell config, always display:**
+> ```
+> ⚠️  Environment variables were added to your shell config.
+>     Open a NEW TERMINAL for them to take effect outside this session.
+> ```
+
 #### Option 1: Aden Platform (OAuth)
 
 This is the recommended flow for supported integrations (HubSpot, etc.).
@@ -202,6 +208,12 @@ if success:
     print(f"Run: {source_cmd}")
 ```
 
+> **⚠️ IMPORTANT: After adding `ADEN_API_KEY` to the user's shell config, always display:**
+> ```
+> ⚠️  Environment variables were added to your shell config.
+>     Open a NEW TERMINAL for them to take effect outside this session.
+> ```
+
 Also save to `~/.hive/configuration.json` for the framework:
 
 ```python
@@ -460,9 +472,14 @@ result: HealthCheckResult = check_credential_health("hubspot", token_value)
 The local encrypted store requires `HIVE_CREDENTIAL_KEY` to encrypt/decrypt credentials.
 
 - If the user doesn't have one, `EncryptedFileStorage` will auto-generate one and log it
-- The user MUST persist this key (e.g., in `~/.bashrc` or a secrets manager)
+- The user MUST persist this key (e.g., in `~/.bashrc`/`~/.zshrc` or a secrets manager)
 - Without this key, stored credentials cannot be decrypted
-- This is the ONLY secret that should live in `~/.bashrc` or environment config
+
+**Shell config rule:** Only TWO keys belong in shell config (`~/.zshrc`/`~/.bashrc`):
+- `HIVE_CREDENTIAL_KEY` — encryption key for the credential store
+- `ADEN_API_KEY` — Aden platform auth key (needed before the store can sync)
+
+All other API keys (Brave, Google, HubSpot, etc.) must go in the encrypted store only. **Never offer to add them to shell config.**
 
 If `HIVE_CREDENTIAL_KEY` is not set:
 
@@ -475,6 +492,7 @@ If `HIVE_CREDENTIAL_KEY` is not set:
 - **NEVER** log, print, or echo credential values in tool output
 - **NEVER** store credentials in plaintext files, git-tracked files, or agent configs
 - **NEVER** hardcode credentials in source code
+- **NEVER** offer to save API keys to shell config (`~/.zshrc`/`~/.bashrc`) — the **only** keys that belong in shell config are `HIVE_CREDENTIAL_KEY` and `ADEN_API_KEY`. All other credentials (Brave, Google, HubSpot, GitHub, Resend, etc.) go in the encrypted store only.
 - **ALWAYS** use `SecretStr` from Pydantic when handling credential values in Python
 - **ALWAYS** use the local encrypted store (`~/.hive/credentials`) for persistence
 - **ALWAYS** run health checks before storing credentials (when possible)
@@ -490,7 +508,7 @@ All credential specs are defined in `tools/src/aden_tools/credentials/`:
 | `llm.py`          | LLM Providers | `anthropic`                                   | No             |
 | `search.py`       | Search Tools  | `brave_search`, `google_search`, `google_cse` | No             |
 | `email.py`        | Email         | `resend`                                      | No             |
-| `integrations.py` | Integrations  | `github`, `hubspot`                           | No / Yes       |
+| `integrations.py` | Integrations  | `github`, `hubspot`, `google_calendar_oauth`  | No / Yes       |
 
 **Note:** Additional LLM providers (Cerebras, Groq, OpenAI) are handled by LiteLLM via environment
 variables (`CEREBRAS_API_KEY`, `GROQ_API_KEY`, `OPENAI_API_KEY`) but are not yet in CREDENTIAL_SPECS.
@@ -601,18 +619,22 @@ All credentials are now configured:
 │                      ✅ CREDENTIALS CONFIGURED                              │
 ├─────────────────────────────────────────────────────────────────────────────┤
 │                                                                             │
+│     OPEN A NEW TERMINAL before running commands below.                      │
+│     Environment variables were saved to your shell config but               │
+│     only take effect in new terminal sessions.                              │
+│                                                                             │
 │  NEXT STEPS:                                                                │
 │                                                                             │
 │  1. RUN YOUR AGENT:                                                         │
 │                                                                             │
-│     PYTHONPATH=core:exports python -m research-agent tui                    │
+│     hive tui                                                                │
 │                                                                             │
 │  2. IF YOU ENCOUNTER ISSUES, USE THE DEBUGGER:                              │
 │                                                                             │
 │     /hive-debugger                                                          │
 │                                                                             │
 │     The debugger analyzes runtime logs, identifies retry loops, tool        │
-│     failures, stalled execution, and provides actionable fix suggestions.  │
+│     failures, stalled execution, and provides actionable fix suggestions.   │
 │                                                                             │
 └─────────────────────────────────────────────────────────────────────────────┘
 ```

.claude/skills/hive-debugger/SKILL.md

@@ -26,6 +26,17 @@ Use `/hive-debugger` when:
 
 This skill works alongside agents running in TUI mode and provides supervisor-level insights into execution behavior.
 
+### Forever-Alive Agent Awareness
+
+Some agents use `terminal_nodes=[]` (the "forever-alive" pattern), meaning they loop indefinitely and never enter a "completed" execution state. For these agents:
+- Sessions with status "in_progress" or "paused" are **normal**, not failures
+- High step counts, long durations, and many node visits are expected behavior
+- The agent stops only when the user explicitly exits — there is no graph-driven completion
+- Debug focus should be on **quality of individual node visits and iterations**, not whether the session reached a terminal state
+- Conversation memory accumulates across loops — watch for context overflow and stale data issues
+
+**How to identify forever-alive agents:** Check `agent.py` or `agent.json` for `terminal_nodes=[]` (empty list). If empty, the agent is forever-alive.
+
 ---
 
 ## Prerequisites
@@ -47,7 +58,7 @@ Before using this skill, ensure:
 **What to do:**
 
 1. **Ask the developer which agent needs debugging:**
-   - Get agent name (e.g., "twitter_outreach", "deep_research_agent")
+   - Get agent name (e.g., "deep_research_agent", "deep_research_agent")
    - Confirm the agent exists in `exports/{agent_name}/`
 
 2. **Determine agent working directory:**
@@ -66,7 +77,7 @@ Before using this skill, ensure:
 
 4. **Store context for the debugging session:**
    - agent_name
-   - agent_work_dir (e.g., `/home/user/.hive/twitter_outreach`)
+   - agent_work_dir (e.g., `/home/user/.hive/deep_research_agent`)
    - goal_id
    - success_criteria
    - constraints
@@ -74,19 +85,19 @@ Before using this skill, ensure:
 
 **Example:**
 ```
-Developer: "My twitter_outreach agent keeps failing"
+Developer: "My deep_research_agent agent keeps failing"
 
-You: "I'll help debug the twitter_outreach agent. Let me gather context..."
+You: "I'll help debug the deep_research_agent agent. Let me gather context..."
 
-[Read exports/twitter_outreach/agent.json]
+[Read exports/deep_research_agent/agent.json]
 
 Context gathered:
-- Agent: twitter_outreach
-- Goal: twitter-outreach-multi-loop
-- Working Directory: /home/user/.hive/twitter_outreach
-- Success Criteria: ["Successfully send 5 personalized outreach messages"]
-- Constraints: ["Must verify handle exists", "Must personalize message"]
-- Nodes: ["intake-collector", "profile-analyzer", "message-composer", "outreach-sender"]
+- Agent: deep_research_agent
+- Goal: deep-research
+- Working Directory: /home/user/.hive/deep_research_agent
+- Success Criteria: ["Produce a comprehensive research report with cited sources"]
+- Constraints: ["Must cite all sources", "Must cover multiple perspectives"]
+- Nodes: ["intake", "research", "analysis", "report-writer"]
 ```
 
 ---
@@ -142,6 +153,7 @@ Store the selected mode for the session.
    - Check `attention_summary.categories` for issue types
    - Note the `run_id` of problematic sessions
    - Check `status` field: "degraded", "failure", "in_progress"
+   - **For forever-alive agents:** Sessions with status "in_progress" or "paused" are normal — these agents never reach "completed". Only flag sessions with `needs_attention: true` or actual error indicators (tool failures, retry loops, missing outputs). High step counts alone do not indicate a problem.
 
 3. **Attention flag triggers to understand:**
    From runtime_logger.py, runs are flagged when:
@@ -199,13 +211,20 @@ Which run would you like to investigate?
    | **Tool Errors** | `tool_error_count > 0`, `attention_reasons` contains "tool_failures" | Tool calls failed (API errors, timeouts, auth issues) |
    | **Retry Loops** | `retry_count > 3`, `verdict_counts.RETRY > 5` | Judge repeatedly rejecting outputs |
    | **Guard Failures** | `guard_reject_count > 0` | Output validation failed (wrong types, missing keys) |
-   | **Stalled Execution** | `total_steps > 20`, `verdict_counts.CONTINUE > 10` | EventLoopNode not making progress |
+   | **Stalled Execution** | `total_steps > 20`, `verdict_counts.CONTINUE > 10` | EventLoopNode not making progress. **Caveat:** Forever-alive agents may legitimately have high step counts — check if agent is blocked at a client-facing node (normal) vs genuinely stuck in a loop |
    | **High Latency** | `latency_ms > 60000`, `avg_step_latency > 5000` | Slow tool calls or LLM responses |
    | **Client-Facing Issues** | `client_input_requested` but no `user_input_received` | Premature set_output before user input |
    | **Edge Routing Errors** | `exit_status == "no_valid_edge"`, `attention_reasons` contains "routing_issue" | No edges match current state |
    | **Memory/Context Issues** | `tokens_used > 100000`, `context_overflow_count > 0` | Conversation history too long |
    | **Constraint Violations** | Compare output against goal constraints | Agent violated goal-level rules |
 
+   **Forever-Alive Agent Caveat:** If the agent uses `terminal_nodes=[]`, sessions will never reach "completed" status. This is by design. When debugging these agents, focus on:
+   - Whether individual node visits succeed (not whether the graph "finishes")
+   - Quality of each loop iteration — are outputs improving or degrading across loops?
+   - Whether client-facing nodes are correctly blocking for user input
+   - Memory accumulation issues: stale data from previous loops, context overflow across many iterations
+   - Conversation compaction behavior: is the conversation growing unbounded?
+
 3. **Analyze each flagged node:**
    - Node ID and name
    - Exit status
@@ -224,7 +243,7 @@ Which run would you like to investigate?
 ```
 Diagnosis for session_20260206_115718_e22339c5:
 
-Problem Node: intake-collector
+Problem Node: research
 ├─ Exit Status: escalate
 ├─ Retry Count: 5 (HIGH)
 ├─ Verdict Counts: {RETRY: 5, ESCALATE: 1}
@@ -232,7 +251,7 @@ Problem Node: intake-collector
 ├─ Total Steps: 8
 └─ Categories: Missing Outputs + Retry Loops
 
-Root Issue: The intake-collector node is stuck in a retry loop because it's not setting required outputs.
+Root Issue: The research node is stuck in a retry loop because it's not setting required outputs.
 ```
 
 ---
@@ -293,25 +312,25 @@ Root Issue: The intake-collector node is stuck in a retry loop because it's not
 
 **Example Output:**
 ```
-Root Cause Analysis for intake-collector:
+Root Cause Analysis for research:
 
 Step-by-step breakdown:
 
 Step 3:
-- Tool Call: web_search(query="@RomuloNevesOf")
-- Result: Found Twitter profile information
+- Tool Call: web_search(query="latest AI regulations 2026")
+- Result: Found relevant articles and sources
 - Verdict: RETRY
-- Feedback: "Missing required output 'twitter_handles'. You found the handle but didn't call set_output."
+- Feedback: "Missing required output 'research_findings'. You found sources but didn't call set_output."
 
 Step 4:
-- Tool Call: web_search(query="@RomuloNevesOf twitter")
-- Result: Found additional Twitter information
+- Tool Call: web_search(query="AI regulation policy 2026")
+- Result: Found additional policy information
 - Verdict: RETRY
-- Feedback: "Still missing 'twitter_handles'. Use set_output to save your findings."
+- Feedback: "Still missing 'research_findings'. Use set_output to save your findings."
 
 Steps 5-7: Similar pattern continues...
 
-ROOT CAUSE: The node is successfully finding Twitter handles via web_search, but the LLM is not calling set_output to save the results. It keeps searching for more information instead of completing the task.
+ROOT CAUSE: The node is successfully finding research sources via web_search, but the LLM is not calling set_output to save the results. It keeps searching for more information instead of completing the task.
 ```
 
 ---
@@ -562,15 +581,33 @@ PYTHONPATH=core:exports python -m {agent_name} --tui
 
 ### Find Available Checkpoints:
 
-```bash
-# In TUI:
-/sessions {session_id}
+Use MCP tools to programmatically find and inspect checkpoints:
 
-# This shows all checkpoints with timestamps:
-Available Checkpoints: (3)
-  1. cp_node_complete_intake_143030
-  2. cp_node_complete_research_143115
-  3. cp_pause_research_143130
+```
+# List all sessions to find the failed one
+list_agent_sessions(agent_work_dir="~/.hive/agents/{agent_name}", status="failed")
+
+# Inspect session state
+get_agent_session_state(agent_work_dir="~/.hive/agents/{agent_name}", session_id="{session_id}")
+
+# Find clean checkpoints to resume from
+list_agent_checkpoints(agent_work_dir="~/.hive/agents/{agent_name}", session_id="{session_id}", is_clean="true")
+
+# Compare checkpoints to understand what changed
+compare_agent_checkpoints(
+    agent_work_dir="~/.hive/agents/{agent_name}",
+    session_id="{session_id}",
+    checkpoint_id_before="cp_node_complete_intake_143030",
+    checkpoint_id_after="cp_node_complete_research_143115"
+)
+
+# Inspect memory at a specific checkpoint
+get_agent_checkpoint(agent_work_dir="~/.hive/agents/{agent_name}", session_id="{session_id}", checkpoint_id="cp_node_complete_intake_143030")
+```
+
+Or in TUI:
+```bash
+/sessions {session_id}
 ```
 
 **Verification:**
@@ -653,7 +690,7 @@ Available Checkpoints: (3)
 
 **Example interaction:**
 ```
-Developer: "I applied the fix to intake-collector. How do I verify it works?"
+Developer: "I applied the fix to research. How do I verify it works?"
 
 You: "Great! Let's verify the fix with these steps:
 
@@ -665,11 +702,11 @@ You: "Great! Let's verify the fix with these steps:
    [Use query_runtime_logs to check for attention flags]
 
 3. Verify the specific node:
-   [Use query_runtime_log_details for intake-collector]
+   [Use query_runtime_log_details for research]
 
 Expected results:
 - No 'needs_attention' flags
-- intake-collector shows exit_status='success'
+- research shows exit_status='success'
 - retry_count should be 0
 
 Let me know when you've run it and I'll help check the logs!"
@@ -687,7 +724,7 @@ Let me know when you've run it and I'll help check the logs!"
 - **Example:**
   ```
   query_runtime_logs(
-      agent_work_dir="/home/user/.hive/twitter_outreach",
+      agent_work_dir="/home/user/.hive/deep_research_agent",
       status="needs_attention",
       limit=20
   )
@@ -699,7 +736,7 @@ Let me know when you've run it and I'll help check the logs!"
 - **Example:**
   ```
   query_runtime_log_details(
-      agent_work_dir="/home/user/.hive/twitter_outreach",
+      agent_work_dir="/home/user/.hive/deep_research_agent",
       run_id="session_20260206_115718_e22339c5",
       needs_attention_only=True
   )
@@ -711,9 +748,83 @@ Let me know when you've run it and I'll help check the logs!"
 - **Example:**
   ```
   query_runtime_log_raw(
-      agent_work_dir="/home/user/.hive/twitter_outreach",
+      agent_work_dir="/home/user/.hive/deep_research_agent",
       run_id="session_20260206_115718_e22339c5",
-      node_id="intake-collector"
+      node_id="research"
+  )
+  ```
+
+### Session & Checkpoint Tools
+
+**list_agent_sessions** - Browse sessions with filtering
+- **When to use:** Finding resumable sessions, identifying failed sessions, Stage 3 triage
+- **Returns:** Session list with status, timestamps, is_resumable, current_node, quality
+- **Example:**
+  ```
+  list_agent_sessions(
+      agent_work_dir="/home/user/.hive/agents/twitter_outreach",
+      status="failed",
+      limit=10
+  )
+  ```
+
+**get_agent_session_state** - Load full session state (excludes memory values)
+- **When to use:** Inspecting session progress, checking is_resumable, examining path
+- **Returns:** Full state with memory_keys/memory_size instead of memory values
+- **Example:**
+  ```
+  get_agent_session_state(
+      agent_work_dir="/home/user/.hive/agents/twitter_outreach",
+      session_id="session_20260208_143022_abc12345"
+  )
+  ```
+
+**get_agent_session_memory** - Get memory contents from a session
+- **When to use:** Stage 5 root cause analysis, inspecting produced data
+- **Returns:** All memory keys+values, or a single key's value
+- **Example:**
+  ```
+  get_agent_session_memory(
+      agent_work_dir="/home/user/.hive/agents/twitter_outreach",
+      session_id="session_20260208_143022_abc12345",
+      key="twitter_handles"
+  )
+  ```
+
+**list_agent_checkpoints** - List checkpoints for a session
+- **When to use:** Stage 6 recovery, finding clean checkpoints to resume from
+- **Returns:** Checkpoint summaries with type, node, clean status
+- **Example:**
+  ```
+  list_agent_checkpoints(
+      agent_work_dir="/home/user/.hive/agents/twitter_outreach",
+      session_id="session_20260208_143022_abc12345",
+      is_clean="true"
+  )
+  ```
+
+**get_agent_checkpoint** - Load a specific checkpoint with full state
+- **When to use:** Inspecting exact state at a checkpoint, comparing to current state
+- **Returns:** Full checkpoint: memory snapshot, execution path, metrics
+- **Example:**
+  ```
+  get_agent_checkpoint(
+      agent_work_dir="/home/user/.hive/agents/twitter_outreach",
+      session_id="session_20260208_143022_abc12345",
+      checkpoint_id="cp_node_complete_intake_143030"
+  )
+  ```
+
+**compare_agent_checkpoints** - Diff memory between two checkpoints
+- **When to use:** Understanding data flow, finding where state diverged
+- **Returns:** Memory diff (added/removed/changed keys) + execution path diff
+- **Example:**
+  ```
+  compare_agent_checkpoints(
+      agent_work_dir="/home/user/.hive/agents/twitter_outreach",
+      session_id="session_20260208_143022_abc12345",
+      checkpoint_id_before="cp_node_complete_intake_143030",
+      checkpoint_id_after="cp_node_complete_research_143115"
   )
   ```
 
@@ -739,27 +850,37 @@ Loop every 10 seconds:
   2. If found: Alert and drill into L2
 ```
 
+**Pattern 4: Session State + Checkpoint Recovery**
+```
+1. list_agent_sessions: Find failed/paused sessions
+2. get_agent_session_state: Check is_resumable, see execution path
+3. get_agent_session_memory: Inspect what data was produced
+4. list_agent_checkpoints: Find clean checkpoints before failure
+5. compare_agent_checkpoints: Understand what changed between checkpoints
+6. Recommend resume command with specific checkpoint
+```
+
 ---
 
 ## Complete Example Walkthrough
 
-### Scenario: twitter_outreach Agent Failing
+### Scenario: deep_research_agent Agent Failing
 
-**Developer says:** "My twitter_outreach agent keeps failing. It's not finding Twitter handles correctly."
+**Developer says:** "My deep_research_agent keeps failing. It's not saving research findings correctly."
 
 #### Stage 1: Setup
 ```
-You: "I'll help debug the twitter_outreach agent. Let me gather context..."
+You: "I'll help debug the deep_research_agent agent. Let me gather context..."
 
-[Read exports/twitter_outreach/agent.json]
+[Read exports/deep_research_agent/agent.json]
 
 Context:
-- Agent: twitter_outreach
-- Goal: twitter-outreach-multi-loop
-- Working Dir: ~/.hive/agents/twitter_outreach
+- Agent: deep_research_agent
+- Goal: deep-research
+- Working Dir: ~/.hive/agents/deep_research_agent
 - Success Criteria: ["Successfully send 5 personalized outreach messages"]
 - Constraints: ["Must verify handle exists", "Must personalize message"]
-- Nodes: intake-collector, profile-analyzer, message-composer, outreach-sender
+- Nodes: research, profile-analyzer, message-composer, outreach-sender
 ```
 
 #### Stage 2: Mode Selection
@@ -791,31 +912,31 @@ You: "Analyzing node-level details..."
 [Call query_runtime_log_details for the session]
 
 Results:
-- Problem Node: intake-collector
+- Problem Node: research
 - Exit Status: escalate
 - Retry Count: 5
 - Verdict Counts: {RETRY: 5, ESCALATE: 1}
 - Attention Reasons: ["high_retry_count", "missing_outputs"]
 - Total Steps: 8
 
-Diagnosis: intake-collector has retry loop + missing outputs issue
+Diagnosis: research has retry loop + missing outputs issue
 Category: Missing Outputs + Retry Loops
 ```
 
 #### Stage 5: Root Cause (L3)
 ```
-You: "Drilling into detailed logs for intake-collector..."
+You: "Drilling into detailed logs for research..."
 
-[Call query_runtime_log_raw for node intake-collector]
+[Call query_runtime_log_raw for node research]
 
 Evidence:
-Step 3: web_search("@RomuloNevesOf") → Found profile
+Step 3: web_search("latest AI regulations 2026") → Found articles
   Verdict: RETRY
-  Feedback: "Missing required output 'twitter_handles'. You found the handle but didn't call set_output."
+  Feedback: "Missing required output 'research_findings'. You found sources but didn't call set_output."
 
-Step 4: web_search("@RomuloNevesOf twitter") → Found more info
+Step 4: web_search("AI regulation policy 2026") → Found more info
   Verdict: RETRY
-  Feedback: "Still missing 'twitter_handles'. Use set_output to save your findings."
+  Feedback: "Still missing 'research_findings'. Use set_output to save your findings."
 
 Steps 5-7: Same pattern repeats...
 
@@ -826,37 +947,37 @@ ROOT CAUSE: Node successfully finds handles but LLM doesn't call set_output. It
 ```
 You: "Here's the fix for this issue:
 
-## Issue: intake-collector Not Setting Outputs
+## Issue: research Node Not Setting Outputs
 
-**Root Cause:** Node successfully finds Twitter handles but doesn't call set_output
+**Root Cause:** Node successfully finds research sources but doesn't call set_output
 
 **Category:** Missing Outputs
 
-**File to edit:** exports/twitter_outreach/nodes/intake_collector.py
+**File to edit:** exports/deep_research_agent/nodes/research.py
 
 **Changes:**
 Update the system_prompt to emphasize the output requirement:
 
 ```python
 system_prompt = """
-You are the Intake Collector for the Twitter outreach agent.
+You are the Research node for the deep research agent.
 
 Your task:
-1. Extract Twitter handle from user input (URL or @mention)
-2. Verify it's a valid Twitter handle format
-3. **CRITICAL: Call set_output(twitter_handles=[handle]) immediately after finding valid handle**
+1. Search for relevant sources on the research topic
+2. Extract key findings from each source
+3. **CRITICAL: Call set_output(research_findings=[...]) immediately after gathering sufficient findings**
 
-IMPORTANT: You MUST call set_output even if you found the handle on the first try.
-Do not continue searching once you have a valid handle - set output and finish.
+IMPORTANT: You MUST call set_output even if you found findings on the first search.
+Do not continue searching indefinitely - set output and finish once you have enough data.
 
 Output format:
 {
-  "twitter_handles": ["@handle1", "@handle2", ...]
+  "research_findings": [{"source": "...", "summary": "...", "url": "..."}, ...]
 }
 """
 ```
 
-**Why this works:** The explicit instruction to call set_output immediately prevents the LLM from continuing to search after finding the handle.
+**Why this works:** The explicit instruction to call set_output immediately prevents the LLM from continuing to search after finding sufficient sources.
 ```
 
 #### Stage 7: Verification
@@ -880,13 +1001,13 @@ You: "Checking now..."
 [Call query_runtime_log_details for the new run]
 
 Results:
-- Node: intake-collector
+- Node: research
 - Exit Status: success ✓
 - Retry Count: 0 ✓
 - Total Steps: 2 ✓
 - No attention flags ✓
 
-"Perfect! The fix worked. The intake-collector now:
+"Perfect! The fix worked. The research now:
 - Completes successfully on first try
 - No retries needed
 - Calls set_output properly
@@ -913,6 +1034,9 @@ Your agent should now work correctly!"
 3. **Don't ignore edge conditions** - Missing edges cause routing failures
 4. **Don't overlook judge configuration** - Mismatched expectations cause retry loops
 5. **Don't forget nullable_output_keys** - Optional inputs need explicit marking
+6. **Don't diagnose "in_progress" as a failure for forever-alive agents** - Agents with `terminal_nodes=[]` are designed to never enter "completed" state. This is intentional. Focus on quality of individual node visits, not session completion status
+7. **Don't ignore conversation memory issues in long-running sessions** - In continuous conversation mode, history grows across node transitions and loop iterations. Watch for context overflow (tokens_used > 100K), stale data from previous loops affecting edge conditions, and compaction failures that cause the LLM to lose important context
+8. **Don't confuse "waiting for user" with "stalled"** - Client-facing nodes in forever-alive agents block for user input by design. A session paused at a client-facing node is working correctly, not stalled
 
 ---
 

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

@@ -1,351 +1,333 @@
-# Example: Testing a YouTube Research Agent
+# Example: Iterative Testing of a Research Agent
 
-This example walks through testing a YouTube research agent that finds relevant videos based on a topic.
+This example walks through the full iterative test loop for a research agent that searches the web, reviews findings, and produces a cited report.
 
-## Prerequisites
+## Agent Structure
 
-- Agent built with hive-create skill at `exports/youtube-research/`
-- Goal defined with success criteria and constraints
-
-## Step 1: Load the Goal
-
-First, load the goal that was defined during the Goal stage:
-
-```json
-{
-    "id": "youtube-research",
-    "name": "YouTube Research Agent",
-    "description": "Find relevant YouTube videos on a given topic",
-    "success_criteria": [
-        {
-            "id": "find_videos",
-            "description": "Find 3-5 relevant videos",
-            "metric": "video_count",
-            "target": "3-5",
-            "weight": 1.0
-        },
-        {
-            "id": "relevance",
-            "description": "Videos must be relevant to the topic",
-            "metric": "relevance_score",
-            "target": ">0.8",
-            "weight": 0.8
-        }
-    ],
-    "constraints": [
-        {
-            "id": "api_limits",
-            "description": "Must not exceed YouTube API rate limits",
-            "constraint_type": "hard",
-            "category": "technical"
-        },
-        {
-            "id": "content_safety",
-            "description": "Must filter out inappropriate content",
-            "constraint_type": "hard",
-            "category": "safety"
-        }
-    ]
-}
+```
+exports/deep_research_agent/
+├── agent.py          # Goal + graph: intake → research → review → report
+├── nodes/__init__.py # Node definitions (system_prompt, input/output keys)
+├── config.py         # Model config
+├── mcp_servers.json  # Tools: web_search, web_scrape
+└── tests/            # Test files (we'll create these)
 ```
 
-## Step 2: Get Constraint Test Guidelines
+**Goal:** "Rigorous Interactive Research" — find 5+ diverse sources, cite every claim, produce a complete report.
 
-During the Goal stage (or early Eval), get test guidelines for constraints:
+---
+
+## Phase 1: Generate Tests
+
+### Read the goal
 
 ```python
-result = generate_constraint_tests(
-    goal_id="youtube-research",
-    goal_json='<goal JSON above>',
-    agent_path="exports/youtube-research"
-)
+Read(file_path="exports/deep_research_agent/agent.py")
+# Extract: goal_id="rigorous-interactive-research"
+# success_criteria: source-diversity (>=5), citation-coverage (100%), report-completeness (90%)
+# constraints: no-hallucination, source-attribution
 ```
 
-**The result contains guidelines (not generated tests):**
-- `output_file`: Where to write tests
-- `file_header`: Imports and fixtures to use
-- `test_template`: Format for test functions
-- `constraints_formatted`: The constraints to test
-- `test_guidelines`: Rules for writing tests
-
-## Step 3: Write Constraint Tests
-
-Using the guidelines, write tests directly with the Write tool:
-
-```python
-# Write constraint tests using the provided file_header and guidelines
-Write(
-    file_path="exports/youtube-research/tests/test_constraints.py",
-    content='''
-"""Constraint tests for youtube-research agent."""
-
-import os
-import pytest
-from exports.youtube_research import default_agent
-
-
-pytestmark = pytest.mark.skipif(
-    not os.environ.get("ANTHROPIC_API_KEY") and not os.environ.get("MOCK_MODE"),
-    reason="API key required for real testing."
-)
-
-
-@pytest.mark.asyncio
-async def test_constraint_api_limits_respected():
-    """Verify API rate limits are not exceeded."""
-    import time
-    mock_mode = bool(os.environ.get("MOCK_MODE"))
-
-    for i in range(10):
-        result = await default_agent.run({"topic": f"test_{i}"}, mock_mode=mock_mode)
-        time.sleep(0.1)
-
-    # Should complete without rate limit errors
-    assert "rate limit" not in str(result).lower()
-
-
-@pytest.mark.asyncio
-async def test_constraint_content_safety_filter():
-    """Verify inappropriate content is filtered."""
-    mock_mode = bool(os.environ.get("MOCK_MODE"))
-    result = await default_agent.run({"topic": "general topic"}, mock_mode=mock_mode)
-
-    for video in result.videos:
-        assert video.safe_for_work is True
-        assert video.age_restricted is False
-'''
-)
-```
-
-## Step 4: Get Success Criteria Test Guidelines
-
-After the agent is built, get success criteria test guidelines:
+### Get test guidelines
 
 ```python
 result = generate_success_tests(
-    goal_id="youtube-research",
-    goal_json='<goal JSON>',
-    node_names="search_node,filter_node,rank_node,format_node",
-    tool_names="youtube_search,video_details,channel_info",
-    agent_path="exports/youtube-research"
+    goal_id="rigorous-interactive-research",
+    goal_json='{"id": "rigorous-interactive-research", "success_criteria": [{"id": "source-diversity", "description": "Use multiple diverse sources", "target": ">=5"}, {"id": "citation-coverage", "description": "Every claim cites its source", "target": "100%"}, {"id": "report-completeness", "description": "Report answers the research questions", "target": "90%"}]}',
+    node_names="intake,research,review,report",
+    tool_names="web_search,web_scrape",
+    agent_path="exports/deep_research_agent"
 )
 ```
 
-## Step 5: Write Success Criteria Tests
-
-Using the guidelines, write success criteria tests:
+### Write tests
 
 ```python
 Write(
-    file_path="exports/youtube-research/tests/test_success_criteria.py",
-    content='''
-"""Success criteria tests for youtube-research agent."""
-
-import os
-import pytest
-from exports.youtube_research import default_agent
-
-
-pytestmark = pytest.mark.skipif(
-    not os.environ.get("ANTHROPIC_API_KEY") and not os.environ.get("MOCK_MODE"),
-    reason="API key required for real testing."
-)
-
+    file_path="exports/deep_research_agent/tests/test_success_criteria.py",
+    content=result["file_header"] + '''
 
 @pytest.mark.asyncio
-async def test_find_videos_happy_path():
-    """Test finding videos for a common topic."""
-    mock_mode = bool(os.environ.get("MOCK_MODE"))
-    result = await default_agent.run({"topic": "machine learning"}, mock_mode=mock_mode)
-
-    assert result.success
-    assert 3 <= len(result.videos) <= 5
-    assert all(v.title for v in result.videos)
-    assert all(v.video_id for v in result.videos)
-
+async def test_success_source_diversity(runner, auto_responder, mock_mode):
+    """At least 5 diverse sources are found."""
+    await auto_responder.start()
+    try:
+        result = await runner.run({"query": "impact of remote work on productivity"})
+    finally:
+        await auto_responder.stop()
+    assert result.success, f"Agent failed: {result.error}"
+    output = result.output or {}
+    sources = output.get("sources", [])
+    if isinstance(sources, list):
+        assert len(sources) >= 5, f"Expected >= 5 sources, got {len(sources)}"
 
 @pytest.mark.asyncio
-async def test_find_videos_minimum_boundary():
-    """Test at minimum threshold (3 videos)."""
-    mock_mode = bool(os.environ.get("MOCK_MODE"))
-    result = await default_agent.run({"topic": "niche topic xyz"}, mock_mode=mock_mode)
-
-    assert len(result.videos) >= 3
-
+async def test_success_citation_coverage(runner, auto_responder, mock_mode):
+    """Every factual claim in the report cites its source."""
+    await auto_responder.start()
+    try:
+        result = await runner.run({"query": "climate change effects on agriculture"})
+    finally:
+        await auto_responder.stop()
+    assert result.success, f"Agent failed: {result.error}"
+    output = result.output or {}
+    report = output.get("report", "")
+    # Check that report contains numbered references
+    assert "[1]" in str(report) or "[source" in str(report).lower(), "Report lacks citations"
 
 @pytest.mark.asyncio
-async def test_relevance_score_threshold():
-    """Test relevance scoring meets threshold."""
-    mock_mode = bool(os.environ.get("MOCK_MODE"))
-    result = await default_agent.run({"topic": "python programming"}, mock_mode=mock_mode)
-
-    for video in result.videos:
-        assert video.relevance_score > 0.8
-
+async def test_success_report_completeness(runner, auto_responder, mock_mode):
+    """Report addresses the original research question."""
+    query = "pros and cons of nuclear energy"
+    await auto_responder.start()
+    try:
+        result = await runner.run({"query": query})
+    finally:
+        await auto_responder.stop()
+    assert result.success, f"Agent failed: {result.error}"
+    output = result.output or {}
+    report = output.get("report", "")
+    assert len(str(report)) > 200, f"Report too short: {len(str(report))} chars"
 
 @pytest.mark.asyncio
-async def test_find_videos_no_results_graceful():
-    """Test graceful handling of no results."""
-    mock_mode = bool(os.environ.get("MOCK_MODE"))
-    result = await default_agent.run({"topic": "xyznonexistent123"}, mock_mode=mock_mode)
+async def test_empty_query_handling(runner, auto_responder, mock_mode):
+    """Agent handles empty input gracefully."""
+    await auto_responder.start()
+    try:
+        result = await runner.run({"query": ""})
+    finally:
+        await auto_responder.stop()
+    output = result.output or {}
+    assert not result.success or output.get("error"), "Should handle empty query"
 
-    # Should not crash, return empty or message
-    assert result.videos == [] or result.message
+@pytest.mark.asyncio
+async def test_feedback_loop_terminates(runner, auto_responder, mock_mode):
+    """Feedback loop between review and research terminates."""
+    await auto_responder.start()
+    try:
+        result = await runner.run({"query": "quantum computing basics"})
+    finally:
+        await auto_responder.stop()
+    visits = result.node_visit_counts or {}
+    for node_id, count in visits.items():
+        assert count <= 5, f"Node {node_id} visited {count} times"
 '''
 )
 ```
 
-## Step 6: Run All Tests
+---
 
-Execute all tests:
+## Phase 2: First Execution
 
 ```python
-result = run_tests(
-    goal_id="youtube-research",
-    agent_path="exports/youtube-research",
-    test_types='["all"]',
-    parallel=4
+run_tests(
+    goal_id="rigorous-interactive-research",
+    agent_path="exports/deep_research_agent",
+    fail_fast=True
 )
 ```
 
-**Results:**
-
+**Result:**
 ```json
 {
-    "goal_id": "youtube-research",
-    "overall_passed": false,
-    "summary": {
-        "total": 6,
-        "passed": 5,
-        "failed": 1,
-        "pass_rate": "83.3%"
-    },
-    "duration_ms": 4521,
-    "results": [
-        {"test_id": "test_constraint_api_001", "passed": true, "duration_ms": 1234},
-        {"test_id": "test_constraint_content_001", "passed": true, "duration_ms": 456},
-        {"test_id": "test_success_001", "passed": true, "duration_ms": 789},
-        {"test_id": "test_success_002", "passed": true, "duration_ms": 654},
-        {"test_id": "test_success_003", "passed": true, "duration_ms": 543},
-        {"test_id": "test_success_004", "passed": false, "duration_ms": 845,
-         "error_category": "IMPLEMENTATION_ERROR",
-         "error_message": "TypeError: 'NoneType' object has no attribute 'videos'"}
-    ]
+  "overall_passed": false,
+  "summary": {"total": 5, "passed": 3, "failed": 2, "pass_rate": "60.0%"},
+  "failures": [
+    {"test_name": "test_success_source_diversity", "details": "AssertionError: Expected >= 5 sources, got 2"},
+    {"test_name": "test_success_citation_coverage", "details": "AssertionError: Report lacks citations"}
+  ]
 }
 ```
 
-## Step 7: Debug the Failed Test
+---
+
+## Phase 3: Analyze (Iteration 1)
+
+### Debug the first failure
 
 ```python
-result = debug_test(
-    goal_id="youtube-research",
-    test_name="test_find_videos_no_results_graceful",
-    agent_path="exports/youtube-research"
+debug_test(
+    goal_id="rigorous-interactive-research",
+    test_name="test_success_source_diversity",
+    agent_path="exports/deep_research_agent"
+)
+# Category: ASSERTION_FAILURE — Expected >= 5 sources, got 2
+```
+
+### Find the session and inspect memory
+
+```python
+list_agent_sessions(
+    agent_work_dir="~/.hive/agents/deep_research_agent",
+    status="completed",
+    limit=1
+)
+# → session_20260209_150000_abc12345
+
+get_agent_session_memory(
+    agent_work_dir="~/.hive/agents/deep_research_agent",
+    session_id="session_20260209_150000_abc12345",
+    key="research_results"
+)
+# → Only 2 sources found. LLM stopped searching after 2 queries.
+```
+
+### Check LLM behavior in the research node
+
+```python
+query_runtime_log_raw(
+    agent_work_dir="~/.hive/agents/deep_research_agent",
+    run_id="session_20260209_150000_abc12345",
+    node_id="research"
+)
+# → LLM called web_search twice, got results, immediately called set_output.
+# → Prompt doesn't instruct it to find at least 5 sources.
+```
+
+**Root cause:** The research node's system_prompt doesn't specify minimum source requirements.
+
+---
+
+## Phase 4: Fix (Iteration 1)
+
+```python
+Read(file_path="exports/deep_research_agent/nodes/__init__.py")
+
+# Fix the research node prompt
+Edit(
+    file_path="exports/deep_research_agent/nodes/__init__.py",
+    old_string='system_prompt="Search for information on the user\'s topic using web search."',
+    new_string='system_prompt="Search for information on the user\'s topic using web search. You MUST find at least 5 diverse, authoritative sources. Use multiple different search queries with varied keywords. Do NOT call set_output until you have gathered at least 5 distinct sources from different domains."'
 )
 ```
 
-**Debug Output:**
+---
 
+## Phase 5: Recover & Resume (Iteration 1)
+
+The fix is to the `research` node. Since this was a `run_tests` execution (no checkpoints), we re-run from scratch:
+
+```python
+run_tests(
+    goal_id="rigorous-interactive-research",
+    agent_path="exports/deep_research_agent",
+    fail_fast=True
+)
+```
+
+**Result:**
 ```json
 {
-    "test_id": "test_success_004",
-    "test_name": "test_find_videos_no_results_graceful",
-    "input": {"topic": "xyznonexistent123"},
-    "expected": "Empty list or message",
-    "actual": {"error": "TypeError: 'NoneType' object has no attribute 'videos'"},
-    "passed": false,
-    "error_message": "TypeError: 'NoneType' object has no attribute 'videos'",
-    "error_category": "IMPLEMENTATION_ERROR",
-    "stack_trace": "Traceback (most recent call last):\n  File \"filter_node.py\", line 42\n    for video in result.videos:\nTypeError: 'NoneType' object has no attribute 'videos'",
-    "logs": [
-        {"timestamp": "2026-01-20T10:00:01", "node": "search_node", "level": "INFO", "msg": "Searching for: xyznonexistent123"},
-        {"timestamp": "2026-01-20T10:00:02", "node": "search_node", "level": "WARNING", "msg": "No results found"},
-        {"timestamp": "2026-01-20T10:00:02", "node": "filter_node", "level": "ERROR", "msg": "NoneType error"}
-    ],
-    "runtime_data": {
-        "execution_path": ["start", "search_node", "filter_node"],
-        "node_outputs": {
-            "search_node": null
-        }
-    },
-    "suggested_fix": "Add null check in filter_node before accessing .videos attribute",
-    "iteration_guidance": {
-        "stage": "Agent",
-        "action": "Fix the code in nodes/edges",
-        "restart_required": false,
-        "description": "The goal is correct, but filter_node doesn't handle null results from search_node."
-    }
+  "overall_passed": false,
+  "summary": {"total": 5, "passed": 4, "failed": 1, "pass_rate": "80.0%"},
+  "failures": [
+    {"test_name": "test_success_citation_coverage", "details": "AssertionError: Report lacks citations"}
+  ]
 }
 ```
 
-## Step 8: Iterate Based on Category
+Source diversity now passes. Citation coverage still fails.
 
-Since this is an **IMPLEMENTATION_ERROR**, we:
+---
 
-1. **Don't restart** the Goal → Agent → Eval flow
-2. **Fix the agent** using hive-create skill:
-   - Modify `filter_node` to handle null results
-3. **Re-run Eval** (tests only)
-
-### Fix in hive-create:
+## Phase 3: Analyze (Iteration 2)
 
 ```python
-# Update the filter_node to handle null
-add_node(
-    node_id="filter_node",
-    name="Filter Node",
-    description="Filter and rank videos",
-    node_type="function",
-    input_keys=["search_results"],
-    output_keys=["filtered_videos"],
-    system_prompt="""
-    Filter videos by relevance.
-    IMPORTANT: Handle case where search_results is None or empty.
-    Return empty list if no results.
-    """
+debug_test(
+    goal_id="rigorous-interactive-research",
+    test_name="test_success_citation_coverage",
+    agent_path="exports/deep_research_agent"
+)
+# Category: ASSERTION_FAILURE — Report lacks citations
+
+# Check what the report node produced
+list_agent_sessions(
+    agent_work_dir="~/.hive/agents/deep_research_agent",
+    status="completed",
+    limit=1
+)
+# → session_20260209_151500_def67890
+
+get_agent_session_memory(
+    agent_work_dir="~/.hive/agents/deep_research_agent",
+    session_id="session_20260209_151500_def67890",
+    key="report"
+)
+# → Report text exists but uses no numbered references.
+# → Sources are in memory but report node doesn't cite them.
+```
+
+**Root cause:** The report node's prompt doesn't instruct the LLM to include numbered citations.
+
+---
+
+## Phase 4: Fix (Iteration 2)
+
+```python
+Edit(
+    file_path="exports/deep_research_agent/nodes/__init__.py",
+    old_string='system_prompt="Write a comprehensive report based on the research findings."',
+    new_string='system_prompt="Write a comprehensive report based on the research findings. You MUST include numbered citations [1], [2], etc. for every factual claim. At the end, include a References section listing all sources with their URLs. Every claim must be traceable to a specific source."'
 )
 ```
 
-### Re-export and re-test:
+---
+
+## Phase 5: Resume (Iteration 2)
+
+The fix is to the `report` node (the last node). To demonstrate checkpoint recovery, run via CLI:
+
+```bash
+# Run via CLI to get checkpoints
+uv run hive run exports/deep_research_agent --input '{"topic": "climate change effects"}'
+
+# After it runs, find the clean checkpoint before report
+list_agent_checkpoints(
+    agent_work_dir="~/.hive/agents/deep_research_agent",
+    session_id="session_20260209_152000_ghi34567",
+    is_clean="true"
+)
+# → cp_node_complete_review_152100 (after review, before report)
+
+# Resume — skips intake, research, review entirely
+uv run hive run exports/deep_research_agent \
+  --resume-session session_20260209_152000_ghi34567 \
+  --checkpoint cp_node_complete_review_152100
+```
+
+Only the `report` node re-runs with the fixed prompt, using research data from the checkpoint.
+
+---
+
+## Phase 6: Final Verification
 
 ```python
-# Re-export the fixed agent
-export_graph(path="exports/youtube-research")
-
-# Re-run tests
-result = run_tests(
-    goal_id="youtube-research",
-    agent_path="exports/youtube-research",
-    test_types='["all"]'
+run_tests(
+    goal_id="rigorous-interactive-research",
+    agent_path="exports/deep_research_agent"
 )
 ```
 
-**Updated Results:**
-
+**Result:**
 ```json
 {
-    "goal_id": "youtube-research",
-    "overall_passed": true,
-    "summary": {
-        "total": 6,
-        "passed": 6,
-        "failed": 0,
-        "pass_rate": "100.0%"
-    }
+  "overall_passed": true,
+  "summary": {"total": 5, "passed": 5, "failed": 0, "pass_rate": "100.0%"}
 }
 ```
 
+All tests pass.
+
+---
+
 ## Summary
 
-1. **Got guidelines** for constraint tests during Goal stage
-2. **Wrote** constraint tests using Write tool
-3. **Got guidelines** for success criteria tests during Eval stage
-4. **Wrote** success criteria tests using Write tool
-5. **Ran** tests in parallel
-6. **Debugged** the one failure
-7. **Categorized** as IMPLEMENTATION_ERROR
-8. **Fixed** the agent (not the goal)
-9. **Re-ran** Eval only (didn't restart full flow)
-10. **Passed** all tests
+| Iteration | Failure | Root Cause | Fix | Recovery |
+|-----------|---------|------------|-----|----------|
+| 1 | Source diversity (2 < 5) | Research prompt too vague | Added "at least 5 sources" to prompt | Re-run (no checkpoints) |
+| 2 | No citations in report | Report prompt lacks citation instructions | Added citation requirements | Checkpoint resume (skipped 3 nodes) |
 
-The agent is now validated and ready for production use.
+**Key takeaways:**
+- Phase 3 analysis (session memory + L3 logs) identified root causes without guessing
+- Checkpoint recovery in iteration 2 saved time by skipping 3 expensive nodes
+- Final `run_tests` confirms all scenarios pass end-to-end

.codex/config.toml

@@ -0,0 +1,7 @@
+# Project-level Codex config for Hive.
+# Keep this file minimal: MCP connectivity + skill discovery.
+
+[mcp_servers.agent-builder]
+command = "uv"
+args = ["run", "--directory", "core", "-m", "framework.mcp.agent_builder_server"]
+cwd = "."

.gitignore

@@ -74,4 +74,6 @@ exports/*
 
 docs/github-issues/*
 core/tests/*dumps/*
-screenshots/*
+
+screenshots/*
+

.mcp.json

@@ -4,11 +4,6 @@
       "command": "uv",
       "args": ["run", "-m", "framework.mcp.agent_builder_server"],
       "cwd": "core"
-    },
-    "tools": {
-      "command": "uv",
-      "args": ["run", "mcp_server.py", "--stdio"],
-      "cwd": "tools"
     }
   }
 }

.opencode/mcp.json

@@ -0,0 +1,30 @@
+{
+  "mcpServers": {
+    "agent-builder": {
+      "command": "uv",
+      "args": [
+        "run",
+        "python",
+        "-m",
+        "framework.mcp.agent_builder_server"
+      ],
+      "cwd": "core",
+      "env": {
+        "PYTHONPATH": "../tools/src"
+      }
+    },
+    "tools": {
+      "command": "uv",
+      "args": [
+        "run",
+        "python",
+        "mcp_server.py",
+        "--stdio"
+      ],
+      "cwd": "tools",
+      "env": {
+        "PYTHONPATH": "src"
+      }
+    }
+  }
+}

.opencode/skills/hive

@@ -0,0 +1 @@
+../../.claude/skills/hive

.opencode/skills/hive-concepts

@@ -0,0 +1 @@
+../../.claude/skills/hive-concepts

.opencode/skills/hive-create

@@ -0,0 +1 @@
+../../.claude/skills/hive-create

.opencode/skills/hive-credentials

@@ -0,0 +1 @@
+../../.claude/skills/hive-credentials

.opencode/skills/hive-debugger

@@ -0,0 +1 @@
+../../.claude/skills/hive-debugger

.opencode/skills/hive-patterns

@@ -0,0 +1 @@
+../../.claude/skills/hive-patterns

.opencode/skills/hive-test

@@ -0,0 +1 @@
+../../.claude/skills/hive-test

.opencode/skills/triage-issue

@@ -0,0 +1 @@
+../../.claude/skills/triage-issue

CONTRIBUTING.md

@@ -49,8 +49,8 @@ You may submit PRs without prior assignment for:
    make check    # Lint and format checks (ruff check + ruff format --check on core/ and tools/)
    make test     # Core tests (cd core && pytest tests/ -v)
    ```
-6. Commit your changes following our commit conventions
-7. Push to your fork and submit a Pull Request
+8. Commit your changes following our commit conventions
+9. Push to your fork and submit a Pull Request
 
 ## Development Setup
 
@@ -99,8 +99,7 @@ docs(readme): update installation instructions
 2. Update documentation if needed
 3. Add tests for new functionality
 4. Ensure `make check` and `make test` pass
-5. Update the CHANGELOG.md if applicable
-6. Request review from maintainers
+5. Request review from maintainers
 
 ### PR Title Format
 
@@ -145,6 +144,9 @@ make test
 # Or run tests directly
 cd core && pytest tests/ -v
 
+# Run tools package tests (when contributing to tools/)
+cd tools && uv run pytest tests/ -v
+
 # Run tests for a specific agent
 PYTHONPATH=exports uv run python -m agent_name test
 ```

README.md

@@ -22,7 +22,6 @@
   <img src="https://img.shields.io/badge/MCP-102_Tools-00ADD8?style=flat-square" alt="MCP" />
 </p>
 
-
 <p align="center">
   <img src="https://img.shields.io/badge/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" />
@@ -82,7 +81,7 @@ Use Hive when you need:
 ### Prerequisites
 
 - Python 3.11+ for agent development
-- Claude Code or Cursor for utilizing agent skills
+- Claude Code, Codex CLI, or Cursor for utilizing agent skills
 
 > **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.
 
@@ -120,9 +119,41 @@ hive tui
 # Or run directly
 hive run exports/your_agent_name --input '{"key": "value"}'
 ```
+##  Coding Agent Support
+### Codex CLI
+Hive includes native support for [OpenAI Codex CLI](https://github.com/openai/codex) (v0.101.0+).
+
+1. **Config:** `.codex/config.toml` with `agent-builder` MCP server (tracked in git)
+2. **Skills:** `.agents/skills/` symlinks to Hive skills (tracked in git)
+3. **Launch:** Run `codex` in the repo root, then type `use hive`
+
+Example:
+```
+codex> use hive
+```
+
+### Opencode 
+Hive includes native support for [Opencode](https://github.com/opencode-ai/opencode).
+
+1. **Setup:** Run the quickstart script 
+2. **Launch:** Open Opencode in the project root.
+3. **Activate:** Type `/hive` in the chat to switch to the Hive Agent.
+4. **Verify:** Ask the agent *"List your tools"* to confirm the connection.
+
+The agent has access to all Hive skills and can scaffold agents, add tools, and debug workflows directly from the chat.
 
 **[📖 Complete Setup Guide](docs/environment-setup.md)** - Detailed instructions for agent development
 
+### Antigravity IDE Support
+
+Skills and MCP servers are also available in [Antigravity IDE](https://antigravity.google/) (Google's AI-powered IDE). **Easiest:** open a terminal in the hive repo folder and run (use `./` — the script is inside the repo):
+
+```bash
+./scripts/setup-antigravity-mcp.sh
+```
+
+**Important:** Always restart/refresh Antigravity IDE after running the setup script—MCP servers only load on startup. After restart, **agent-builder** and **tools** MCP servers should connect. Skills are under `.agent/skills/` (symlinks to `.claude/skills/`). See [docs/antigravity-setup.md](docs/antigravity-setup.md) for manual setup and troubleshooting.
+
 ## Features
 
 - **[Goal-Driven Development](docs/key_concepts/goals_outcome.md)** - Define objectives in natural language; the coding agent generates the agent graph and connection code to achieve them
@@ -302,6 +333,7 @@ subgraph Expansion
         j2["Cursor"]
         j3["Opencode"]
         j4["Antigravity"]
+        j5["Codex CLI"]
     end
     subgraph plat["Platform"]
         k1["JavaScript/TypeScript SDK"]

core/framework/config.py

@@ -6,6 +6,7 @@ helper functions.
 """
 
 import json
+import os
 from dataclasses import dataclass, field
 from pathlib import Path
 from typing import Any
@@ -24,7 +25,7 @@ def get_hive_config() -> dict[str, Any]:
     if not HIVE_CONFIG_FILE.exists():
         return {}
     try:
-        with open(HIVE_CONFIG_FILE) as f:
+        with open(HIVE_CONFIG_FILE, encoding="utf-8-sig") as f:
             return json.load(f)
     except (json.JSONDecodeError, OSError):
         return {}
@@ -48,6 +49,15 @@ def get_max_tokens() -> int:
     return get_hive_config().get("llm", {}).get("max_tokens", DEFAULT_MAX_TOKENS)
 
 
+def get_api_key() -> str | None:
+    """Return the API key from the environment variable specified in configuration."""
+    llm = get_hive_config().get("llm", {})
+    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
+
+
 # ---------------------------------------------------------------------------
 # RuntimeConfig – shared across agent templates
 # ---------------------------------------------------------------------------
@@ -60,5 +70,5 @@ class RuntimeConfig:
     model: str = field(default_factory=get_preferred_model)
     temperature: float = 0.7
     max_tokens: int = field(default_factory=get_max_tokens)
-    api_key: str | None = None
+    api_key: str | None = field(default_factory=get_api_key)
     api_base: str | None = None

core/framework/credentials/__init__.py

@@ -68,6 +68,7 @@ from .storage import (
 )
 from .store import CredentialStore
 from .template import TemplateResolver
+from .validation import ensure_credential_key_env, validate_agent_credentials
 
 # Aden sync components (lazy import to avoid httpx dependency when not needed)
 # Usage: from core.framework.credentials.aden import AdenSyncProvider
@@ -111,6 +112,9 @@ __all__ = [
     "CredentialRefreshError",
     "CredentialValidationError",
     "CredentialDecryptionError",
+    # Validation
+    "ensure_credential_key_env",
+    "validate_agent_credentials",
     # Aden sync (optional - requires httpx)
     "AdenSyncProvider",
     "AdenCredentialClient",

core/framework/credentials/validation.py

@@ -0,0 +1,133 @@
+"""Credential validation utilities.
+
+Provides reusable credential validation for agents, whether run through
+the AgentRunner or directly via GraphExecutor.
+"""
+
+from __future__ import annotations
+
+import logging
+import os
+
+logger = logging.getLogger(__name__)
+
+
+def ensure_credential_key_env() -> None:
+    """Load HIVE_CREDENTIAL_KEY from shell config if not already in environment.
+
+    The setup-credentials skill writes the encryption key to ~/.zshrc or ~/.bashrc.
+    If the user hasn't sourced their config in the current shell, this reads it
+    directly so the runner (and any MCP subprocesses it spawns) can unlock the
+    encrypted credential store.
+
+    Only HIVE_CREDENTIAL_KEY is loaded this way — all other secrets (API keys, etc.)
+    come from the credential store itself.
+    """
+    if os.environ.get("HIVE_CREDENTIAL_KEY"):
+        return
+
+    try:
+        from aden_tools.credentials.shell_config import check_env_var_in_shell_config
+
+        found, value = check_env_var_in_shell_config("HIVE_CREDENTIAL_KEY")
+        if found and value:
+            os.environ["HIVE_CREDENTIAL_KEY"] = value
+            logger.debug("Loaded HIVE_CREDENTIAL_KEY from shell config")
+    except ImportError:
+        pass
+
+
+def validate_agent_credentials(nodes: list) -> None:
+    """Check that required credentials are available before running an agent.
+
+    Scans node specs for required tools and node types, then checks whether
+    the corresponding credentials exist in the credential store.
+
+    Raises CredentialError with actionable guidance if any are missing.
+
+    Args:
+        nodes: List of NodeSpec objects from the agent graph.
+    """
+    required_tools: set[str] = set()
+    for node in nodes:
+        if node.tools:
+            required_tools.update(node.tools)
+    node_types: set[str] = {node.node_type for node in nodes}
+
+    try:
+        from aden_tools.credentials import CREDENTIAL_SPECS
+
+        from framework.credentials import CredentialStore
+        from framework.credentials.storage import (
+            CompositeStorage,
+            EncryptedFileStorage,
+            EnvVarStorage,
+        )
+    except ImportError:
+        return  # aden_tools not installed, skip check
+
+    # Build credential store
+    env_mapping = {
+        (spec.credential_id or name): spec.env_var for name, spec in CREDENTIAL_SPECS.items()
+    }
+    storages: list = [EnvVarStorage(env_mapping=env_mapping)]
+    if os.environ.get("HIVE_CREDENTIAL_KEY"):
+        storages.insert(0, EncryptedFileStorage())
+    if len(storages) == 1:
+        storage = storages[0]
+    else:
+        storage = CompositeStorage(primary=storages[0], fallbacks=storages[1:])
+    store = CredentialStore(storage=storage)
+
+    # Build reverse mappings
+    tool_to_cred: dict[str, str] = {}
+    node_type_to_cred: dict[str, str] = {}
+    for cred_name, spec in CREDENTIAL_SPECS.items():
+        for tool_name in spec.tools:
+            tool_to_cred[tool_name] = cred_name
+        for nt in spec.node_types:
+            node_type_to_cred[nt] = cred_name
+
+    missing: list[str] = []
+    checked: set[str] = set()
+
+    # Check tool credentials
+    for tool_name in sorted(required_tools):
+        cred_name = tool_to_cred.get(tool_name)
+        if cred_name is None or cred_name in checked:
+            continue
+        checked.add(cred_name)
+        spec = CREDENTIAL_SPECS[cred_name]
+        cred_id = spec.credential_id or cred_name
+        if spec.required and not store.is_available(cred_id):
+            affected = sorted(t for t in required_tools if t in spec.tools)
+            entry = f"  {spec.env_var} for {', '.join(affected)}"
+            if spec.help_url:
+                entry += f"\n    Get it at: {spec.help_url}"
+            missing.append(entry)
+
+    # Check node type credentials (e.g., ANTHROPIC_API_KEY for LLM nodes)
+    for nt in sorted(node_types):
+        cred_name = node_type_to_cred.get(nt)
+        if cred_name is None or cred_name in checked:
+            continue
+        checked.add(cred_name)
+        spec = CREDENTIAL_SPECS[cred_name]
+        cred_id = spec.credential_id or cred_name
+        if spec.required and not store.is_available(cred_id):
+            affected_types = sorted(t for t in node_types if t in spec.node_types)
+            entry = f"  {spec.env_var} for {', '.join(affected_types)} nodes"
+            if spec.help_url:
+                entry += f"\n    Get it at: {spec.help_url}"
+            missing.append(entry)
+
+    if missing:
+        from framework.credentials.models import CredentialError
+
+        lines = ["Missing required credentials:\n"]
+        lines.extend(missing)
+        lines.append(
+            "\nTo fix: run /hive-credentials in Claude Code."
+            "\nIf you've already set up credentials, restart your terminal to load them."
+        )
+        raise CredentialError("\n".join(lines))

core/framework/graph/conversation.py

@@ -27,6 +27,9 @@ class Message:
     tool_use_id: str | None = None
     tool_calls: list[dict[str, Any]] | None = None
     is_error: bool = False
+    # Phase-aware compaction metadata (continuous mode)
+    phase_id: str | None = None
+    is_transition_marker: bool = False
 
     def to_llm_dict(self) -> dict[str, Any]:
         """Convert to OpenAI-format message dict."""
@@ -60,6 +63,10 @@ class Message:
             d["tool_calls"] = self.tool_calls
         if self.is_error:
             d["is_error"] = self.is_error
+        if self.phase_id is not None:
+            d["phase_id"] = self.phase_id
+        if self.is_transition_marker:
+            d["is_transition_marker"] = self.is_transition_marker
         return d
 
     @classmethod
@@ -72,6 +79,8 @@ class Message:
             tool_use_id=data.get("tool_use_id"),
             tool_calls=data.get("tool_calls"),
             is_error=data.get("is_error", False),
+            phase_id=data.get("phase_id"),
+            is_transition_marker=data.get("is_transition_marker", False),
         )
 
 
@@ -188,6 +197,7 @@ class NodeConversation:
         self._next_seq: int = 0
         self._meta_persisted: bool = False
         self._last_api_input_tokens: int | None = None
+        self._current_phase: str | None = None
 
     # --- Properties --------------------------------------------------------
 
@@ -195,6 +205,33 @@ class NodeConversation:
     def system_prompt(self) -> str:
         return self._system_prompt
 
+    def update_system_prompt(self, new_prompt: str) -> None:
+        """Update the system prompt.
+
+        Used in continuous conversation mode at phase transitions to swap
+        Layer 3 (focus) while preserving the conversation history.
+        """
+        self._system_prompt = new_prompt
+
+    def set_current_phase(self, phase_id: str) -> None:
+        """Set the current phase ID. Subsequent messages will be stamped with it."""
+        self._current_phase = phase_id
+
+    async def switch_store(self, new_store: ConversationStore) -> None:
+        """Switch to a new persistence store at a phase transition.
+
+        Subsequent messages are written to *new_store*.  Meta (system
+        prompt, config) is re-persisted on the next write so the new
+        store's ``meta.json`` reflects the updated prompt.
+        """
+        self._store = new_store
+        self._meta_persisted = False
+        await new_store.write_cursor({"next_seq": self._next_seq})
+
+    @property
+    def current_phase(self) -> str | None:
+        return self._current_phase
+
     @property
     def messages(self) -> list[Message]:
         """Return a defensive copy of the message list."""
@@ -216,8 +253,19 @@ class NodeConversation:
 
     # --- Add messages ------------------------------------------------------
 
-    async def add_user_message(self, content: str) -> Message:
-        msg = Message(seq=self._next_seq, role="user", content=content)
+    async def add_user_message(
+        self,
+        content: str,
+        *,
+        is_transition_marker: bool = False,
+    ) -> Message:
+        msg = Message(
+            seq=self._next_seq,
+            role="user",
+            content=content,
+            phase_id=self._current_phase,
+            is_transition_marker=is_transition_marker,
+        )
         self._messages.append(msg)
         self._next_seq += 1
         await self._persist(msg)
@@ -233,6 +281,7 @@ class NodeConversation:
             role="assistant",
             content=content,
             tool_calls=tool_calls,
+            phase_id=self._current_phase,
         )
         self._messages.append(msg)
         self._next_seq += 1
@@ -251,6 +300,7 @@ class NodeConversation:
             content=content,
             tool_use_id=tool_use_id,
             is_error=is_error,
+            phase_id=self._current_phase,
         )
         self._messages.append(msg)
         self._next_seq += 1
@@ -380,6 +430,11 @@ class NodeConversation:
         spillover filename reference (if any). Message structure (role,
         seq, tool_use_id) stays valid for the LLM API.
 
+        Phase-aware behavior (continuous mode): when messages have ``phase_id``
+        metadata, all messages in the current phase are protected regardless of
+        token budget. Transition markers are never pruned. Older phases' tool
+        results are pruned more aggressively.
+
         Error tool results are never pruned — they prevent re-calling
         failing tools.
 
@@ -388,13 +443,18 @@ class NodeConversation:
         if not self._messages:
             return 0
 
-        # Phase 1: Walk backward, classify tool results as protected vs pruneable
+        # Walk backward, classify tool results as protected vs pruneable
         protected_tokens = 0
         pruneable: list[int] = []  # indices into self._messages
         pruneable_tokens = 0
 
         for i in range(len(self._messages) - 1, -1, -1):
             msg = self._messages[i]
+
+            # Transition markers are never pruned (any role)
+            if msg.is_transition_marker:
+                continue
+
             if msg.role != "tool":
                 continue
             if msg.is_error:
@@ -402,6 +462,10 @@ class NodeConversation:
             if msg.content.startswith("[Pruned tool result"):
                 continue  # already pruned
 
+            # Phase-aware: protect current phase messages
+            if self._current_phase and msg.phase_id == self._current_phase:
+                continue
+
             est = len(msg.content) // 4
             if protected_tokens < protect_tokens:
                 protected_tokens += est
@@ -409,11 +473,11 @@ class NodeConversation:
                 pruneable.append(i)
                 pruneable_tokens += est
 
-        # Phase 2: Only prune if enough to be worthwhile
+        # Only prune if enough to be worthwhile
         if pruneable_tokens < min_prune_tokens:
             return 0
 
-        # Phase 3: Replace content with compact placeholder
+        # Replace content with compact placeholder
         count = 0
         for i in pruneable:
             msg = self._messages[i]
@@ -436,6 +500,8 @@ class NodeConversation:
                 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,
             )
             count += 1
 
@@ -446,22 +512,38 @@ class NodeConversation:
         self._last_api_input_tokens = None
         return count
 
-    async def compact(self, summary: str, keep_recent: int = 2) -> None:
+    async def compact(
+        self,
+        summary: str,
+        keep_recent: int = 2,
+        phase_graduated: bool = False,
+    ) -> None:
         """Replace old messages with a summary, optionally keeping recent ones.
 
         Args:
             summary: Caller-provided summary text.
             keep_recent: Number of recent messages to preserve (default 2).
                          Clamped to [0, len(messages) - 1].
+            phase_graduated: When True and messages have phase_id metadata,
+                split at phase boundaries instead of using keep_recent.
+                Keeps current + previous phase intact; compacts older phases.
         """
         if not self._messages:
             return
 
-        # Clamp: must discard at least 1 message
-        keep_recent = max(0, min(keep_recent, len(self._messages) - 1))
-
         total = len(self._messages)
-        split = total - keep_recent if keep_recent > 0 else total
+
+        # Phase-graduated: find the split point based on phase boundaries.
+        # Keeps current phase + previous phase intact, compacts older phases.
+        if phase_graduated and self._current_phase:
+            split = self._find_phase_graduated_split()
+        else:
+            split = None
+
+        if split is None:
+            # Fallback: use keep_recent (non-phase or single-phase conversation)
+            keep_recent = max(0, min(keep_recent, total - 1))
+            split = total - keep_recent if keep_recent > 0 else total
 
         # Advance split past orphaned tool results at the boundary.
         # Tool-role messages reference a tool_use from the preceding
@@ -470,6 +552,10 @@ class NodeConversation:
         while split < total and self._messages[split].role == "tool":
             split += 1
 
+        # Nothing to compact
+        if split == 0:
+            return
+
         old_messages = list(self._messages[:split])
         recent_messages = list(self._messages[split:])
 
@@ -504,6 +590,33 @@ class NodeConversation:
         self._messages = [summary_msg] + recent_messages
         self._last_api_input_tokens = None  # reset; next LLM call will recalibrate
 
+    def _find_phase_graduated_split(self) -> int | None:
+        """Find split point that preserves current + previous phase.
+
+        Returns the index of the first message in the protected set,
+        or None if phase graduation doesn't apply (< 3 phases).
+        """
+        # Collect distinct phases in order of first appearance
+        phases_seen: list[str] = []
+        for msg in self._messages:
+            if msg.phase_id and msg.phase_id not in phases_seen:
+                phases_seen.append(msg.phase_id)
+
+        # Need at least 3 phases for graduation to be meaningful
+        # (current + previous are protected, older get compacted)
+        if len(phases_seen) < 3:
+            return None
+
+        # Protect: current phase + previous phase
+        protected_phases = {phases_seen[-1], phases_seen[-2]}
+
+        # Find split: first message belonging to a protected phase
+        for i, msg in enumerate(self._messages):
+            if msg.phase_id in protected_phases:
+                return i
+
+        return None
+
     async def clear(self) -> None:
         """Remove all messages, keep system prompt, preserve ``_next_seq``."""
         if self._store:

core/framework/graph/conversation_judge.py

@@ -0,0 +1,177 @@
+"""Level 2 Conversation-Aware Judge.
+
+When a node has `success_criteria` set, the implicit judge upgrades:
+after Level 0 passes (all output keys set), a fast LLM call evaluates
+whether the conversation actually meets the criteria.
+
+This prevents nodes from "checking boxes" (setting output keys) without
+doing quality work. The LLM reads the recent conversation and assesses
+whether the phase's goal was genuinely accomplished.
+"""
+
+from __future__ import annotations
+
+import logging
+from dataclasses import dataclass
+from typing import Any
+
+from framework.graph.conversation import NodeConversation
+from framework.llm.provider import LLMProvider
+
+logger = logging.getLogger(__name__)
+
+
+@dataclass
+class PhaseVerdict:
+    """Result of Level 2 conversation-aware evaluation."""
+
+    action: str  # "ACCEPT" or "RETRY"
+    confidence: float = 0.8
+    feedback: str = ""
+
+
+async def evaluate_phase_completion(
+    llm: LLMProvider,
+    conversation: NodeConversation,
+    phase_name: str,
+    phase_description: str,
+    success_criteria: str,
+    accumulator_state: dict[str, Any],
+    max_history_tokens: int = 8_196,
+) -> PhaseVerdict:
+    """Level 2 judge: read the conversation and evaluate quality.
+
+    Only called after Level 0 passes (all output keys set).
+
+    Args:
+        llm: LLM provider for evaluation
+        conversation: The current conversation to evaluate
+        phase_name: Name of the current phase/node
+        phase_description: Description of the phase
+        success_criteria: Natural-language criteria for phase completion
+        accumulator_state: Current output key values
+        max_history_tokens: Main conversation token budget (judge gets 20%)
+
+    Returns:
+        PhaseVerdict with action and optional feedback
+    """
+    # Build a compact view of the recent conversation
+    recent_messages = _extract_recent_context(conversation, max_messages=10)
+    outputs_summary = _format_outputs(accumulator_state)
+
+    system_prompt = (
+        "You are a quality judge evaluating whether a phase of work is complete. "
+        "Be concise. Evaluate based on the success criteria, not on style."
+    )
+
+    user_prompt = f"""Evaluate this phase:
+
+PHASE: {phase_name}
+DESCRIPTION: {phase_description}
+
+SUCCESS CRITERIA:
+{success_criteria}
+
+OUTPUTS SET:
+{outputs_summary}
+
+RECENT CONVERSATION:
+{recent_messages}
+
+Has this phase accomplished its goal based on the success criteria?
+
+Respond in exactly this format:
+ACTION: ACCEPT or RETRY
+CONFIDENCE: 0.X
+FEEDBACK: (reason if RETRY, empty if ACCEPT)"""
+
+    try:
+        response = llm.complete(
+            messages=[{"role": "user", "content": user_prompt}],
+            system=system_prompt,
+            max_tokens=max(1024, max_history_tokens // 5),
+            max_retries=1,
+        )
+        if not response.content or not response.content.strip():
+            logger.debug("Level 2 judge: empty response, accepting by default")
+            return PhaseVerdict(action="ACCEPT", confidence=0.5, feedback="")
+        return _parse_verdict(response.content)
+    except Exception as e:
+        logger.warning(f"Level 2 judge failed, accepting by default: {e}")
+        # On failure, don't block — Level 0 already passed
+        return PhaseVerdict(action="ACCEPT", confidence=0.5, feedback="")
+
+
+def _extract_recent_context(conversation: NodeConversation, max_messages: int = 10) -> str:
+    """Extract recent conversation messages for evaluation."""
+    messages = conversation.messages
+    recent = messages[-max_messages:] if len(messages) > max_messages else messages
+
+    parts = []
+    for msg in recent:
+        role = msg.role.upper()
+        content = msg.content or ""
+        # Truncate long tool results
+        if msg.role == "tool" and len(content) > 200:
+            content = content[:200] + "..."
+        if content.strip():
+            parts.append(f"[{role}]: {content.strip()}")
+
+    return "\n".join(parts) if parts else "(no messages)"
+
+
+def _format_outputs(accumulator_state: dict[str, Any]) -> str:
+    """Format output key values for evaluation.
+
+    Lists and dicts get structural formatting so the judge can assess
+    quantity and structure, not just a truncated stringification.
+    """
+    if not accumulator_state:
+        return "(none)"
+    parts = []
+    for key, value in accumulator_state.items():
+        if isinstance(value, list):
+            # Show count + brief per-item preview so the judge can
+            # verify quantity without the full serialization.
+            items_preview = []
+            for i, item in enumerate(value[:8]):
+                item_str = str(item)
+                if len(item_str) > 150:
+                    item_str = item_str[:150] + "..."
+                items_preview.append(f"    [{i}]: {item_str}")
+            val_str = f"list ({len(value)} items):\n" + "\n".join(items_preview)
+            if len(value) > 8:
+                val_str += f"\n    ... and {len(value) - 8} more"
+        elif isinstance(value, dict):
+            val_str = str(value)
+            if len(val_str) > 400:
+                val_str = val_str[:400] + "..."
+        else:
+            val_str = str(value)
+            if len(val_str) > 300:
+                val_str = val_str[:300] + "..."
+        parts.append(f"  {key}: {val_str}")
+    return "\n".join(parts)
+
+
+def _parse_verdict(response: str) -> PhaseVerdict:
+    """Parse LLM response into PhaseVerdict."""
+    action = "ACCEPT"
+    confidence = 0.8
+    feedback = ""
+
+    for line in response.strip().split("\n"):
+        line = line.strip()
+        if line.startswith("ACTION:"):
+            action_str = line.split(":", 1)[1].strip().upper()
+            if action_str in ("ACCEPT", "RETRY"):
+                action = action_str
+        elif line.startswith("CONFIDENCE:"):
+            try:
+                confidence = float(line.split(":", 1)[1].strip())
+            except ValueError:
+                pass
+        elif line.startswith("FEEDBACK:"):
+            feedback = line.split(":", 1)[1].strip()
+
+    return PhaseVerdict(action=action, confidence=confidence, feedback=feedback)

core/framework/graph/edge.py

@@ -21,6 +21,9 @@ allowing the LLM to evaluate whether proceeding along an edge makes sense
 given the current goal, context, and execution state.
 """
 
+import json
+import logging
+import re
 from enum import StrEnum
 from typing import Any
 
@@ -28,6 +31,8 @@ from pydantic import BaseModel, Field, model_validator
 
 from framework.graph.safe_eval import safe_eval
 
+logger = logging.getLogger(__name__)
+
 DEFAULT_MAX_TOKENS = 8192
 
 
@@ -158,9 +163,6 @@ class EdgeSpec(BaseModel):
         memory: dict[str, Any],
     ) -> bool:
         """Evaluate a conditional expression."""
-        import logging
-
-        logger = logging.getLogger(__name__)
 
         if not self.condition_expr:
             return True
@@ -217,8 +219,6 @@ class EdgeSpec(BaseModel):
         The LLM evaluates whether proceeding to the target node
         is the best next step toward achieving the goal.
         """
-        import json
-
         # Build context for LLM
         prompt = f"""You are evaluating whether to proceed along an edge in an agent workflow.
 
@@ -254,8 +254,6 @@ Respond with ONLY a JSON object:
             )
 
             # Parse response
-            import re
-
             json_match = re.search(r"\{[^{}]*\}", response.content, re.DOTALL)
             if json_match:
                 data = json.loads(json_match.group())
@@ -263,9 +261,6 @@ Respond with ONLY a JSON object:
                 reasoning = data.get("reasoning", "")
 
                 # Log the decision (using basic print for now)
-                import logging
-
-                logger = logging.getLogger(__name__)
                 logger.info(f"      🤔 LLM routing decision: {'PROCEED' if proceed else 'SKIP'}")
                 logger.info(f"         Reason: {reasoning}")
 
@@ -273,9 +268,6 @@ Respond with ONLY a JSON object:
 
         except Exception as e:
             # Fallback: proceed on success
-            import logging
-
-            logger = logging.getLogger(__name__)
             logger.warning(f"      ⚠ LLM routing failed, defaulting to on_success: {e}")
             return source_success
 
@@ -443,6 +435,25 @@ class GraphSpec(BaseModel):
         description="EventLoopNode configuration (max_iterations, max_tool_calls_per_turn, etc.)",
     )
 
+    # Conversation mode
+    conversation_mode: str = Field(
+        default="continuous",
+        description=(
+            "How conversations flow between event_loop nodes. "
+            "'continuous' (default): one conversation threads through all "
+            "event_loop nodes with cumulative tools and layered prompt composition. "
+            "'isolated': each node gets a fresh conversation."
+        ),
+    )
+    identity_prompt: str | None = Field(
+        default=None,
+        description=(
+            "Agent-level identity prompt (Layer 1 of the onion model). "
+            "In continuous mode, this is the static identity that persists "
+            "unchanged across all node transitions. In isolated mode, ignored."
+        ),
+    )
+
     # Metadata
     description: str = ""
     created_by: str = ""  # "human" or "builder_agent"

core/framework/graph/event_loop_node.py

@@ -149,7 +149,7 @@ class EventLoopNode(NodeProtocol):
     1. Try to restore from durable state (crash recovery)
     2. If no prior state, init from NodeSpec.system_prompt + input_keys
     3. Loop: drain injection queue -> stream LLM -> execute tools
-       -> if client_facing + ask_user called: block for user input
+       -> if client_facing: block for user input (see below)
        -> judge evaluates (acceptance criteria)
        (each add_* and set_output writes through to store immediately)
     4. Publish events to EventBus at each stage
@@ -157,11 +157,17 @@ class EventLoopNode(NodeProtocol):
     6. Terminate when judge returns ACCEPT, shutdown signaled, or max iterations
     7. Build output dict from OutputAccumulator
 
-    Client-facing blocking: When ``client_facing=True``, a synthetic
-    ``ask_user`` tool is injected.  The node blocks for user input ONLY
-    when the LLM explicitly calls ``ask_user()``.  Text-only turns
-    without ``ask_user`` flow through without blocking, allowing the LLM
-    to stream progress updates and summaries freely.
+    Client-facing blocking (``client_facing=True``):
+
+    - **Text-only turns** (no real tool calls, no set_output)
+      automatically block for user input.  If the LLM is talking to the
+      user (not calling tools or setting outputs), it should wait for
+      the user's response before the judge runs.
+    - **Work turns** (tool calls or set_output) flow through without
+      blocking — the LLM is making progress, not asking the user.
+    - A synthetic ``ask_user`` tool is also injected for explicit
+      blocking when the LLM wants to be deliberate about requesting
+      input (e.g. mid-tool-call).
 
     Always returns NodeResult with retryable=False semantics. The executor
     must NOT retry event loop nodes -- retry is handled internally by the
@@ -234,23 +240,48 @@ class EventLoopNode(NodeProtocol):
             return NodeResult(success=False, error=error_msg)
 
         # 2. Restore or create new conversation + accumulator
-        conversation, accumulator, start_iteration = await self._restore(ctx)
-        if conversation is None:
-            system_prompt = ctx.node_spec.system_prompt or ""
+        # Track whether we're in continuous mode (conversation threaded across nodes)
+        _is_continuous = getattr(ctx, "continuous_mode", False)
 
-            conversation = NodeConversation(
-                system_prompt=system_prompt,
-                max_history_tokens=self._config.max_history_tokens,
-                output_keys=ctx.node_spec.output_keys or None,
-                store=self._conversation_store,
+        if _is_continuous and ctx.inherited_conversation is not None:
+            # Continuous mode with inherited conversation from prior node.
+            # This takes priority over store restoration — when the graph loops
+            # back to a previously-visited node, the inherited conversation
+            # carries forward the full thread rather than restoring stale state.
+            # System prompt already updated by executor. Transition marker
+            # already inserted by executor. Fresh accumulator for this phase.
+            # Phase already set by executor via set_current_phase().
+            conversation = ctx.inherited_conversation
+            # Use cumulative output keys for compaction protection (all phases),
+            # falling back to current node's keys if not in continuous mode.
+            conversation._output_keys = (
+                ctx.cumulative_output_keys or ctx.node_spec.output_keys or None
             )
             accumulator = OutputAccumulator(store=self._conversation_store)
             start_iteration = 0
+        else:
+            # Try crash-recovery restore from store, then fall back to fresh.
+            conversation, accumulator, start_iteration = await self._restore(ctx)
+            if conversation is None:
+                # Fresh conversation: either isolated mode or first node in continuous mode.
+                system_prompt = ctx.node_spec.system_prompt or ""
 
-            # Add initial user message from input data
-            initial_message = self._build_initial_message(ctx)
-            if initial_message:
-                await conversation.add_user_message(initial_message)
+                conversation = NodeConversation(
+                    system_prompt=system_prompt,
+                    max_history_tokens=self._config.max_history_tokens,
+                    output_keys=ctx.node_spec.output_keys or None,
+                    store=self._conversation_store,
+                )
+                # Stamp phase for first node in continuous mode
+                if _is_continuous:
+                    conversation.set_current_phase(ctx.node_id)
+                accumulator = OutputAccumulator(store=self._conversation_store)
+                start_iteration = 0
+
+                # Add initial user message from input data
+                initial_message = self._build_initial_message(ctx)
+                if initial_message:
+                    await conversation.add_user_message(initial_message)
 
         # 3. Build tool list: node tools + synthetic set_output + ask_user tools
         tools = list(ctx.available_tools)
@@ -274,6 +305,7 @@ class EventLoopNode(NodeProtocol):
 
         # 5. Stall detection state
         recent_responses: list[str] = []
+        user_interaction_count = 0  # tracks how many times this node blocked for user input
 
         # 6. Main loop
         for iteration in range(start_iteration, self._config.max_iterations):
@@ -304,6 +336,7 @@ class EventLoopNode(NodeProtocol):
                     output=accumulator.to_dict(),
                     tokens_used=total_input_tokens + total_output_tokens,
                     latency_ms=latency_ms,
+                    conversation=conversation if _is_continuous else None,
                 )
 
             # 6b. Drain injection queue
@@ -428,6 +461,7 @@ class EventLoopNode(NodeProtocol):
                         output=accumulator.to_dict(),
                         tokens_used=total_input_tokens + total_output_tokens,
                         latency_ms=latency_ms,
+                        conversation=conversation if _is_continuous else None,
                     )
 
             # 6f. Stall detection
@@ -478,6 +512,7 @@ class EventLoopNode(NodeProtocol):
                     output=accumulator.to_dict(),
                     tokens_used=total_input_tokens + total_output_tokens,
                     latency_ms=latency_ms,
+                    conversation=conversation if _is_continuous else None,
                 )
 
             # 6g. Write cursor checkpoint
@@ -485,14 +520,27 @@ class EventLoopNode(NodeProtocol):
 
             # 6h. Client-facing input blocking
             #
-            # For client_facing nodes, block for user input only when the
-            # LLM explicitly called ask_user().  Text-only turns without
-            # ask_user flow through without blocking, allowing progress
-            # updates and summaries to stream freely.
+            # Two triggers:
+            # (a) Explicit ask_user() — always blocks, then falls through
+            #     to judge evaluation (6i).
+            # (b) Auto-block — a text-only turn (no real tools, no
+            #     set_output) from a client-facing node is addressed to the
+            #     user.  Block for their response, then *skip* judge so the
+            #     next LLM turn can process the reply without confusing
+            #     "missing outputs" feedback.
             #
-            # After user input, always fall through to judge evaluation
-            # (6i).  The judge handles all acceptance decisions.
-            if ctx.node_spec.client_facing and user_input_requested:
+            # Turns that include tool calls or set_output are *work*, not
+            # conversation — they flow through without blocking.
+            _cf_block = False
+            _cf_auto = False
+            if ctx.node_spec.client_facing:
+                if user_input_requested:
+                    _cf_block = True
+                elif assistant_text and not real_tool_results and not outputs_set:
+                    _cf_block = True
+                    _cf_auto = True
+
+            if _cf_block:
                 if self._shutdown:
                     await self._publish_loop_completed(stream_id, node_id, iteration + 1)
                     latency_ms = int((time.time() - start_time) * 1000)
@@ -532,9 +580,15 @@ class EventLoopNode(NodeProtocol):
                         output=accumulator.to_dict(),
                         tokens_used=total_input_tokens + total_output_tokens,
                         latency_ms=latency_ms,
+                        conversation=conversation if _is_continuous else None,
                     )
 
-                logger.info("[%s] iter=%d: blocking for user input...", node_id, iteration)
+                logger.info(
+                    "[%s] iter=%d: blocking for user input (auto=%s)...",
+                    node_id,
+                    iteration,
+                    _cf_auto,
+                )
                 got_input = await self._await_user_input(ctx)
                 logger.info("[%s] iter=%d: unblocked, got_input=%s", node_id, iteration, got_input)
                 if not got_input:
@@ -576,10 +630,34 @@ class EventLoopNode(NodeProtocol):
                         output=accumulator.to_dict(),
                         tokens_used=total_input_tokens + total_output_tokens,
                         latency_ms=latency_ms,
+                        conversation=conversation if _is_continuous else None,
                     )
 
+                user_interaction_count += 1
                 recent_responses.clear()
-                # Fall through to judge evaluation (6i)
+
+                if _cf_auto:
+                    # Auto-block: skip judge — let the LLM process the
+                    # user's response on the next turn without confusing
+                    # "missing outputs" feedback injected between the
+                    # assistant's question and the user's answer.
+                    _continue_count += 1
+                    if ctx.runtime_logger:
+                        iter_latency_ms = int((time.time() - iter_start) * 1000)
+                        ctx.runtime_logger.log_step(
+                            node_id=node_id,
+                            node_type="event_loop",
+                            step_index=iteration,
+                            verdict="CONTINUE",
+                            verdict_feedback="Auto-blocked for user input (pre-interaction)",
+                            tool_calls=logged_tool_calls,
+                            llm_text=assistant_text,
+                            input_tokens=turn_tokens.get("input", 0),
+                            output_tokens=turn_tokens.get("output", 0),
+                            latency_ms=iter_latency_ms,
+                        )
+                    continue
+                # Explicit ask_user: fall through to judge evaluation (6i)
 
             # 6i. Judge evaluation
             should_judge = (
@@ -625,6 +703,17 @@ class EventLoopNode(NodeProtocol):
                 fb_preview,
             )
 
+            # Publish judge verdict event
+            judge_type = "custom" if self._judge is not None else "implicit"
+            await self._publish_judge_verdict(
+                stream_id,
+                node_id,
+                action=verdict.action,
+                feedback=fb_preview,
+                judge_type=judge_type,
+                iteration=iteration,
+            )
+
             if verdict.action == "ACCEPT":
                 # Check for missing output keys
                 missing = self._get_missing_output_keys(
@@ -702,6 +791,7 @@ class EventLoopNode(NodeProtocol):
                     output=accumulator.to_dict(),
                     tokens_used=total_input_tokens + total_output_tokens,
                     latency_ms=latency_ms,
+                    conversation=conversation if _is_continuous else None,
                 )
 
             elif verdict.action == "ESCALATE":
@@ -746,6 +836,7 @@ class EventLoopNode(NodeProtocol):
                     output=accumulator.to_dict(),
                     tokens_used=total_input_tokens + total_output_tokens,
                     latency_ms=latency_ms,
+                    conversation=conversation if _is_continuous else None,
                 )
 
             elif verdict.action == "RETRY":
@@ -795,6 +886,7 @@ class EventLoopNode(NodeProtocol):
             output=accumulator.to_dict(),
             tokens_used=total_input_tokens + total_output_tokens,
             latency_ms=latency_ms,
+            conversation=conversation if _is_continuous else None,
         )
 
     async def inject_event(self, content: str) -> None:
@@ -819,11 +911,20 @@ class EventLoopNode(NodeProtocol):
     async def _await_user_input(self, ctx: NodeContext) -> bool:
         """Block until user input arrives or shutdown is signaled.
 
-        Called when a client_facing node explicitly calls ask_user() —
-        an intentional conversational turn boundary.
+        Called in two situations:
+        - The LLM explicitly calls ask_user().
+        - Auto-block: any text-only turn (no real tools, no set_output)
+          from a client-facing node — ensures the user sees and responds
+          before the judge runs.
 
         Returns True if input arrived, False if shutdown was signaled.
         """
+        # Clear BEFORE emitting so that synchronous handlers (e.g. the
+        # headless stdin handler) can call inject_event() during the emit
+        # and the signal won't be lost.  TUI handlers return immediately
+        # without injecting, so the wait still blocks until the user types.
+        self._input_ready.clear()
+
         if self._event_bus:
             await self._event_bus.emit_client_input_requested(
                 stream_id=ctx.node_id,
@@ -831,7 +932,6 @@ class EventLoopNode(NodeProtocol):
                 prompt="",
             )
 
-        self._input_ready.clear()
         await self._input_ready.wait()
         return not self._shutdown
 
@@ -888,6 +988,21 @@ class EventLoopNode(NodeProtocol):
                 await self._compact_tiered(ctx, conversation, accumulator)
 
             messages = conversation.to_llm_messages()
+
+            # Defensive guard: ensure messages don't end with an assistant
+            # message.  The Anthropic API rejects "assistant message prefill"
+            # (conversations must end with a user or tool message).  This can
+            # happen after compaction trims messages leaving an assistant tail,
+            # or when a conversation is inherited without a transition marker
+            # (e.g. parallel-branch execution).
+            if messages and messages[-1].get("role") == "assistant":
+                logger.info(
+                    "[%s] Messages end with assistant — injecting continuation prompt",
+                    node_id,
+                )
+                await conversation.add_user_message("[Continue working on your current task.]")
+                messages = conversation.to_llm_messages()
+
             accumulated_text = ""
             tool_calls: list[ToolCallEvent] = []
 
@@ -954,13 +1069,20 @@ class EventLoopNode(NodeProtocol):
                     user_input_requested,
                 )
 
-            # Execute tool calls — separate real tools from set_output
+            # Execute tool calls — framework tools (set_output, ask_user)
+            # run inline; real MCP tools run in parallel.
             real_tool_results: list[dict] = []
             limit_hit = False
             executed_in_batch = 0
             hard_limit = int(
                 self._config.max_tool_calls_per_turn * (1 + self._config.tool_call_overflow_margin)
             )
+
+            # Phase 1: triage — handle framework tools immediately,
+            # queue real tools for parallel execution.
+            results_by_id: dict[str, ToolResult] = {}
+            pending_real: list[ToolCallEvent] = []
+
             for tc in tool_calls:
                 tool_call_count += 1
                 if tool_call_count > hard_limit:
@@ -968,11 +1090,9 @@ class EventLoopNode(NodeProtocol):
                     break
                 executed_in_batch += 1
 
-                # Publish tool call started
                 await self._publish_tool_started(
                     stream_id, node_id, tc.tool_use_id, tc.tool_name, tc.tool_input
                 )
-
                 logger.info(
                     "[%s] tool_call: %s(%s)",
                     node_id,
@@ -989,7 +1109,7 @@ class EventLoopNode(NodeProtocol):
                         is_error=result.is_error,
                     )
                     if not result.is_error:
-                        value = tc.tool_input["value"]
+                        value = tc.tool_input.get("value", "")
                         # Parse JSON strings into native types so downstream
                         # consumers get lists/dicts instead of serialised JSON,
                         # and the hallucination validator skips non-string values.
@@ -1000,8 +1120,10 @@ class EventLoopNode(NodeProtocol):
                                     value = parsed
                             except (json.JSONDecodeError, TypeError):
                                 pass
-                        await accumulator.set(tc.tool_input["key"], value)
-                        outputs_set_this_turn.append(tc.tool_input["key"])
+                        key = tc.tool_input.get("key", "")
+                        await accumulator.set(key, value)
+                        outputs_set_this_turn.append(key)
+                        await self._publish_output_key_set(stream_id, node_id, key)
                     logged_tool_calls.append(
                         {
                             "tool_use_id": tc.tool_use_id,
@@ -1011,6 +1133,8 @@ class EventLoopNode(NodeProtocol):
                             "is_error": result.is_error,
                         }
                     )
+                    results_by_id[tc.tool_use_id] = result
+
                 elif tc.tool_name == "ask_user":
                     # --- Framework-level ask_user handling ---
                     user_input_requested = True
@@ -1019,10 +1143,55 @@ class EventLoopNode(NodeProtocol):
                         content="Waiting for user input...",
                         is_error=False,
                     )
+                    results_by_id[tc.tool_use_id] = result
+
                 else:
-                    # --- Real tool execution ---
-                    result = await self._execute_tool(tc)
-                    result = self._truncate_tool_result(result, tc.tool_name)
+                    # --- Real tool: check for truncated args, else queue ---
+                    if "_raw" in tc.tool_input:
+                        result = ToolResult(
+                            tool_use_id=tc.tool_use_id,
+                            content=(
+                                f"Tool call to '{tc.tool_name}' failed: your arguments "
+                                "were truncated (hit output token limit). "
+                                "Simplify or shorten your arguments and try again."
+                            ),
+                            is_error=True,
+                        )
+                        logger.warning(
+                            "[%s] Blocked truncated _raw tool call: %s",
+                            node_id,
+                            tc.tool_name,
+                        )
+                        results_by_id[tc.tool_use_id] = result
+                    else:
+                        pending_real.append(tc)
+
+            # Phase 2: execute real tools in parallel.
+            if pending_real:
+                raw_results = await asyncio.gather(
+                    *(self._execute_tool(tc) for tc in pending_real),
+                    return_exceptions=True,
+                )
+                for tc, raw in zip(pending_real, raw_results, strict=True):
+                    if isinstance(raw, BaseException):
+                        result = ToolResult(
+                            tool_use_id=tc.tool_use_id,
+                            content=f"Tool '{tc.tool_name}' raised: {raw}",
+                            is_error=True,
+                        )
+                    else:
+                        result = raw
+                    results_by_id[tc.tool_use_id] = self._truncate_tool_result(result, tc.tool_name)
+
+            # Phase 3: record results into conversation in original order,
+            # build logged/real lists, and publish completed events.
+            for tc in tool_calls[:executed_in_batch]:
+                result = results_by_id.get(tc.tool_use_id)
+                if result is None:
+                    continue  # shouldn't happen
+
+                # Build log entries for real tools
+                if tc.tool_name not in ("set_output", "ask_user"):
                     tool_entry = {
                         "tool_use_id": tc.tool_use_id,
                         "tool_name": tc.tool_name,
@@ -1033,15 +1202,11 @@ class EventLoopNode(NodeProtocol):
                     real_tool_results.append(tool_entry)
                     logged_tool_calls.append(tool_entry)
 
-                # Record tool result in conversation (both real and set_output
-                # go into the conversation for LLM context continuity)
                 await conversation.add_tool_result(
                     tool_use_id=tc.tool_use_id,
                     content=result.content,
                     is_error=result.is_error,
                 )
-
-                # Publish tool call completed
                 await self._publish_tool_completed(
                     stream_id,
                     node_id,
@@ -1283,6 +1448,44 @@ class EventLoopNode(NodeProtocol):
                 accumulator, ctx.node_spec.output_keys, ctx.node_spec.nullable_output_keys
             )
             if not missing:
+                # Safety check: when ALL output keys are nullable and NONE
+                # have been set, the node produced nothing useful.  Retry
+                # instead of accepting an empty result — this prevents
+                # client-facing nodes from terminating before the user
+                # ever interacts, and non-client-facing nodes from
+                # short-circuiting without doing their work.
+                output_keys = ctx.node_spec.output_keys or []
+                nullable_keys = set(ctx.node_spec.nullable_output_keys or [])
+                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}"
+                        ),
+                    )
+
+                # Level 2: conversation-aware quality check (if success_criteria set)
+                if ctx.node_spec.success_criteria and ctx.llm:
+                    from framework.graph.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_history_tokens=self._config.max_history_tokens,
+                    )
+                    if verdict.action != "ACCEPT":
+                        return JudgeVerdict(
+                            action=verdict.action,
+                            feedback=verdict.feedback or "Phase criteria not met.",
+                        )
+
                 return JudgeVerdict(action="ACCEPT")
             else:
                 return JudgeVerdict(
@@ -1308,26 +1511,43 @@ class EventLoopNode(NodeProtocol):
 
         Used in compaction summaries to prevent the LLM from re-calling
         tools it already called. Extracts:
-        - Tool call counts (e.g. "github_list_pull_requests (6x)")
+        - Tool call details: name, count, and *inputs* for key tools
+          (search queries, scrape URLs, loaded filenames)
         - Files saved via save_data
         - Outputs set via set_output
         - Errors encountered
         """
-        tool_counts: dict[str, int] = {}
+        # Per-tool: list of input summaries (one per call)
+        tool_calls_detail: dict[str, list[str]] = {}
         files_saved: list[str] = []
         outputs_set: list[str] = []
         errors: list[str] = []
 
+        # Tool-specific input extractors: return a short summary string
+        def _summarize_input(name: str, args: dict) -> str:
+            if name == "web_search":
+                return args.get("query", "")
+            if name == "web_scrape":
+                return args.get("url", "")
+            if name == "load_data":
+                return args.get("filename", "")
+            if name == "save_data":
+                return args.get("filename", "")
+            return ""
+
         for msg in conversation.messages:
             if msg.role == "assistant" and msg.tool_calls:
                 for tc in msg.tool_calls:
                     func = tc.get("function", {})
                     name = func.get("name", "unknown")
-                    tool_counts[name] = tool_counts.get(name, 0) + 1
                     try:
                         args = json.loads(func.get("arguments", "{}"))
                     except (json.JSONDecodeError, TypeError):
                         args = {}
+
+                    summary = _summarize_input(name, args)
+                    tool_calls_detail.setdefault(name, []).append(summary)
+
                     if name == "save_data" and args.get("filename"):
                         files_saved.append(args["filename"])
                     if name == "set_output" and args.get("key"):
@@ -1338,9 +1558,18 @@ class EventLoopNode(NodeProtocol):
                 errors.append(preview)
 
         parts: list[str] = []
-        if tool_counts:
-            lines = [f"  {n} ({c}x)" for n, c in tool_counts.items()]
-            parts.append("TOOLS ALREADY CALLED:\n" + "\n".join(lines[:max_entries]))
+        if tool_calls_detail:
+            lines: list[str] = []
+            for name, inputs in list(tool_calls_detail.items())[:max_entries]:
+                count = len(inputs)
+                # Include input details for tools where inputs matter
+                non_empty = [s for s in inputs if s]
+                if non_empty:
+                    detail_lines = [f"    - {s[:120]}" for s in non_empty[:8]]
+                    lines.append(f"  {name} ({count}x):\n" + "\n".join(detail_lines))
+                else:
+                    lines.append(f"  {name} ({count}x)")
+            parts.append("TOOLS ALREADY CALLED:\n" + "\n".join(lines))
         if files_saved:
             unique = list(dict.fromkeys(files_saved))
             parts.append("FILES SAVED: " + ", ".join(unique))
@@ -1429,6 +1658,30 @@ class EventLoopNode(NodeProtocol):
         if limit <= 0 or result.is_error or len(result.content) <= limit:
             return result
 
+        # load_data is the designated mechanism for reading spilled files.
+        # Don't re-spill (circular), but DO truncate with a pagination hint.
+        if tool_name == "load_data":
+            preview_chars = max(limit - 300, limit // 2)
+            preview = result.content[:preview_chars]
+            truncated = (
+                f"[load_data result: {len(result.content)} chars — "
+                f"too large for context. Use offset_bytes and limit_bytes parameters "
+                f"to read smaller chunks, e.g. "
+                f"load_data(filename=..., offset_bytes=0, limit_bytes=5000).]\n\n"
+                f"Preview:\n{preview}…"
+            )
+            logger.info(
+                "load_data result truncated: %d → %d chars "
+                "(use offset_bytes/limit_bytes to paginate)",
+                len(result.content),
+                len(truncated),
+            )
+            return ToolResult(
+                tool_use_id=result.tool_use_id,
+                content=truncated,
+                is_error=False,
+            )
+
         # Determine a preview size — leave room for the metadata wrapper
         preview_chars = max(limit - 300, limit // 2)
         preview = result.content[:preview_chars]
@@ -1517,40 +1770,53 @@ class EventLoopNode(NodeProtocol):
             )
             if not conversation.needs_compaction():
                 # Pruning freed enough — skip full compaction entirely
+                prune_before = round(ratio * 100)
+                prune_after = round(new_ratio * 100)
+                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 pruned (tool results): "
+                        f"{prune_before}% \u2192 {prune_after}%",
+                        verdict="COMPACTION",
+                        verdict_feedback=f"level=prune_only "
+                        f"before={prune_before}% after={prune_after}%",
+                    )
                 if self._event_bus:
                     from framework.runtime.event_bus import AgentEvent, EventType
 
                     await self._event_bus.publish(
                         AgentEvent(
-                            type=EventType.CUSTOM,
+                            type=EventType.CONTEXT_COMPACTED,
                             stream_id=ctx.node_id,
                             node_id=ctx.node_id,
                             data={
-                                "custom_type": "node_compaction",
-                                "node_id": ctx.node_id,
                                 "level": "prune_only",
-                                "usage_before": round(ratio * 100),
-                                "usage_after": round(new_ratio * 100),
+                                "usage_before": prune_before,
+                                "usage_after": prune_after,
                             },
                         )
                     )
                 return
             ratio = new_ratio
 
+        _phase_grad = getattr(ctx, "continuous_mode", False)
+
         if ratio >= 1.2:
             level = "emergency"
             logger.warning("Emergency compaction triggered (usage %.0f%%)", ratio * 100)
             summary = self._build_emergency_summary(ctx, accumulator, conversation)
-            await conversation.compact(summary, keep_recent=1)
+            await conversation.compact(summary, keep_recent=1, phase_graduated=_phase_grad)
         elif ratio >= 1.0:
             level = "aggressive"
             logger.info("Aggressive compaction triggered (usage %.0f%%)", ratio * 100)
             summary = await self._generate_compaction_summary(ctx, conversation)
-            await conversation.compact(summary, keep_recent=2)
+            await conversation.compact(summary, keep_recent=2, phase_graduated=_phase_grad)
         else:
             level = "normal"
             summary = await self._generate_compaction_summary(ctx, conversation)
-            await conversation.compact(summary, keep_recent=4)
+            await conversation.compact(summary, keep_recent=4, phase_graduated=_phase_grad)
 
         new_ratio = conversation.usage_ratio()
         logger.info(
@@ -1559,17 +1825,29 @@ class EventLoopNode(NodeProtocol):
             ratio * 100,
             new_ratio * 100,
         )
+
+        # Log compaction to session logs (tool_logs.jsonl)
+        before_pct = round(ratio * 100)
+        after_pct = round(new_ratio * 100)
+        if ctx.runtime_logger:
+            ctx.runtime_logger.log_step(
+                node_id=ctx.node_id,
+                node_type="event_loop",
+                step_index=-1,  # Not a regular LLM step
+                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 self._event_bus:
             from framework.runtime.event_bus import AgentEvent, EventType
 
             await self._event_bus.publish(
                 AgentEvent(
-                    type=EventType.CUSTOM,
+                    type=EventType.CONTEXT_COMPACTED,
                     stream_id=ctx.node_id,
                     node_id=ctx.node_id,
                     data={
-                        "custom_type": "node_compaction",
-                        "node_id": ctx.node_id,
                         "level": level,
                         "usage_before": round(ratio * 100),
                         "usage_after": round(new_ratio * 100),
@@ -1600,13 +1878,16 @@ class EventLoopNode(NodeProtocol):
                 f"{tool_history}"
             )
 
+        # Dynamic budget: reasoning models (o1, gpt-5-mini) spend max_tokens on
+        # internal thinking. 500 leaves nothing for the actual summary.
+        summary_budget = max(1024, self._config.max_history_tokens // 10)
         try:
             response = ctx.llm.complete(
                 messages=[{"role": "user", "content": prompt}],
                 system=(
                     "Summarize conversations concisely. Always preserve the tool history section."
                 ),
-                max_tokens=500,
+                max_tokens=summary_budget,
             )
             summary = response.content
             # Ensure tool history is present even if LLM dropped it
@@ -1674,12 +1955,28 @@ class EventLoopNode(NodeProtocol):
         if spec.tools:
             parts.append(f"AVAILABLE TOOLS: {', '.join(spec.tools)}")
 
-        # 5. Spillover files hint
+        # 5. Spillover files — list actual files so the LLM can load
+        # them immediately instead of having to call list_data_files first.
         if self._config.spillover_dir:
-            parts.append(
-                "NOTE: Large tool results were saved to files. "
-                "Use load_data(filename='<filename>') to read them."
-            )
+            try:
+                from pathlib import Path
+
+                data_dir = Path(self._config.spillover_dir)
+                if data_dir.is_dir():
+                    files = sorted(f.name for f in data_dir.iterdir() if f.is_file())
+                    if files:
+                        file_list = "\n".join(f"  - {f}" for f in files[:30])
+                        parts.append("DATA FILES (use load_data to read):\n" + file_list)
+                    else:
+                        parts.append(
+                            "NOTE: Large tool results may have been saved to files. "
+                            "Use list_data_files() to check."
+                        )
+            except Exception:
+                parts.append(
+                    "NOTE: Large tool results were saved to files. "
+                    "Use load_data(filename='<filename>') to read them."
+                )
 
         # 6. Tool call history (prevent re-calling tools)
         if conversation is not None:
@@ -1884,3 +2181,35 @@ class EventLoopNode(NodeProtocol):
                 result=result,
                 is_error=is_error,
             )
+
+    async def _publish_judge_verdict(
+        self,
+        stream_id: str,
+        node_id: str,
+        action: str,
+        feedback: str = "",
+        judge_type: str = "implicit",
+        iteration: int = 0,
+    ) -> None:
+        if self._event_bus:
+            await self._event_bus.emit_judge_verdict(
+                stream_id=stream_id,
+                node_id=node_id,
+                action=action,
+                feedback=feedback,
+                judge_type=judge_type,
+                iteration=iteration,
+            )
+
+    async def _publish_output_key_set(
+        self,
+        stream_id: str,
+        node_id: str,
+        key: str,
+    ) -> None:
+        if self._event_bus:
+            await self._event_bus.emit_output_key_set(
+                stream_id=stream_id,
+                node_id=node_id,
+                key=key,
+            )

core/framework/graph/executor.py

@@ -186,6 +186,55 @@ class GraphExecutor:
         # Pause/resume control
         self._pause_requested = asyncio.Event()
 
+    def _write_progress(
+        self,
+        current_node: str,
+        path: list[str],
+        memory: Any,
+        node_visit_counts: dict[str, int],
+    ) -> None:
+        """Update state.json with live progress at node transitions.
+
+        Reads the existing state.json (written by ExecutionStream at session
+        start) and patches the progress fields in-place.  This keeps
+        state.json as the single source of truth — readers always see
+        current progress, not stale initial values.
+
+        The write is synchronous and best-effort: never blocks execution.
+        """
+        if not self._storage_path:
+            return
+        try:
+            import json as _json
+            from datetime import datetime
+
+            state_path = self._storage_path / "state.json"
+            if state_path.exists():
+                state_data = _json.loads(state_path.read_text(encoding="utf-8"))
+            else:
+                state_data = {}
+
+            # Patch progress fields
+            progress = state_data.setdefault("progress", {})
+            progress["current_node"] = current_node
+            progress["path"] = list(path)
+            progress["node_visit_counts"] = dict(node_visit_counts)
+            progress["steps_executed"] = len(path)
+
+            # Update timestamp
+            timestamps = state_data.setdefault("timestamps", {})
+            timestamps["updated_at"] = datetime.now().isoformat()
+
+            # Persist full memory so state.json is sufficient for resume
+            # even if the process dies before the final write.
+            memory_snapshot = memory.read_all()
+            state_data["memory"] = memory_snapshot
+            state_data["memory_keys"] = list(memory_snapshot.keys())
+
+            state_path.write_text(_json.dumps(state_data, indent=2), encoding="utf-8")
+        except Exception:
+            pass  # Best-effort — never block execution
+
     def _validate_tools(self, graph: GraphSpec) -> list[str]:
         """
         Validate that all tools declared by nodes are available.
@@ -257,6 +306,13 @@ class GraphExecutor:
         # Initialize execution state
         memory = SharedMemory()
 
+        # Continuous conversation mode state
+        is_continuous = getattr(graph, "conversation_mode", "isolated") == "continuous"
+        continuous_conversation = None  # NodeConversation threaded across nodes
+        cumulative_tools: list = []  # Tools accumulate, never removed
+        cumulative_tool_names: set[str] = set()
+        cumulative_output_keys: list[str] = []  # Output keys from all visited nodes
+
         # Initialize checkpoint store if checkpointing is enabled
         checkpoint_store: CheckpointStore | None = None
         if checkpoint_config and checkpoint_config.enabled and self._storage_path:
@@ -273,13 +329,20 @@ class GraphExecutor:
                     f"{type(memory_data).__name__}, expected dict"
                 )
             else:
-                # Restore memory from previous session
+                # Restore memory from previous session.
+                # Skip validation — this data was already validated when
+                # originally written, and research text triggers false
+                # positives on the code-indicator heuristic.
                 for key, value in memory_data.items():
-                    memory.write(key, value)
+                    memory.write(key, value, validate=False)
                 self.logger.info(f"📥 Restored session state with {len(memory_data)} memory keys")
 
-        # Write new input data to memory (each key individually)
-        if input_data:
+        # Write new input data to memory (each key individually).
+        # Skip when resuming from a paused session — restored memory already
+        # contains all state including the original input, and re-writing
+        # input_data would overwrite intermediate results with stale values.
+        _is_resuming = bool(session_state and session_state.get("paused_at"))
+        if input_data and not _is_resuming:
             for key, value in input_data.items():
                 memory.write(key, value)
 
@@ -368,7 +431,7 @@ class GraphExecutor:
             # Check if resuming from paused_at (session state resume)
             paused_at = session_state.get("paused_at") if session_state else None
             node_ids = [n.id for n in graph.nodes]
-            self.logger.info(f"🔍 Debug: paused_at={paused_at}, available node IDs={node_ids}")
+            self.logger.debug(f"paused_at={paused_at}, available node IDs={node_ids}")
 
             if paused_at and graph.get_node(paused_at) is not None:
                 # Resume from paused_at node directly (works for any node, not just pause_nodes)
@@ -399,6 +462,13 @@ class GraphExecutor:
         if session_state and current_node_id != graph.entry_node:
             self.logger.info(f"🔄 Resuming from: {current_node_id}")
 
+            # Emit resume event
+            if self._event_bus:
+                await self._event_bus.emit_execution_resumed(
+                    stream_id=self._stream_id,
+                    node_id=current_node_id,
+                )
+
         # Start run
         _run_id = self.runtime.start_run(
             goal_id=goal.id,
@@ -435,6 +505,14 @@ class GraphExecutor:
                 if self._pause_requested.is_set():
                     self.logger.info("⏸ Pause detected - stopping at node boundary")
 
+                    # Emit pause event
+                    if self._event_bus:
+                        await self._event_bus.emit_execution_paused(
+                            stream_id=self._stream_id,
+                            node_id=current_node_id,
+                            reason="User requested pause (Ctrl+Z)",
+                        )
+
                     # Create session state for pause
                     saved_memory = memory.read_all()
                     pause_session_state: dict[str, Any] = {
@@ -505,6 +583,21 @@ class GraphExecutor:
 
                 path.append(current_node_id)
 
+                # Clear stale nullable outputs from previous visits.
+                # When a node is re-visited (e.g. review → process-batch → review),
+                # nullable outputs from the PREVIOUS visit linger in shared memory.
+                # This causes stale edge conditions to fire (e.g. "feedback is not None"
+                # from visit 1 triggers even when visit 2 sets "final_summary" instead).
+                # Clearing them ensures only the CURRENT visit's outputs affect routing.
+                if node_visit_counts.get(current_node_id, 0) > 1:
+                    nullable_keys = getattr(node_spec, "nullable_output_keys", None) or []
+                    for key in nullable_keys:
+                        if memory.read(key) is not None:
+                            memory.write(key, None, validate=False)
+                            self.logger.info(
+                                f"   🧹 Cleared stale nullable output '{key}' from previous visit"
+                            )
+
                 # Check if pause (HITL) before execution
                 if current_node_id in graph.pause_nodes:
                     self.logger.info(f"⏸ Paused at HITL node: {node_spec.name}")
@@ -515,6 +608,17 @@ class GraphExecutor:
                 self.logger.info(f"   Inputs: {node_spec.input_keys}")
                 self.logger.info(f"   Outputs: {node_spec.output_keys}")
 
+                # Continuous mode: accumulate tools and output keys from this node
+                if is_continuous and node_spec.tools:
+                    for t in self.tools:
+                        if t.name in node_spec.tools and t.name not in cumulative_tool_names:
+                            cumulative_tools.append(t)
+                            cumulative_tool_names.add(t.name)
+                if is_continuous and node_spec.output_keys:
+                    for k in node_spec.output_keys:
+                        if k not in cumulative_output_keys:
+                            cumulative_output_keys.append(k)
+
                 # Build context for node
                 ctx = self._build_context(
                     node_spec=node_spec,
@@ -522,6 +626,10 @@ class GraphExecutor:
                     goal=goal,
                     input_data=input_data or {},
                     max_tokens=graph.max_tokens,
+                    continuous_mode=is_continuous,
+                    inherited_conversation=continuous_conversation if is_continuous else None,
+                    override_tools=cumulative_tools if is_continuous else None,
+                    cumulative_output_keys=cumulative_output_keys if is_continuous else None,
                 )
 
                 # Log actual input data being read
@@ -689,6 +797,17 @@ class GraphExecutor:
                         self.logger.info(
                             f"   ↻ Retrying ({node_retry_counts[current_node_id]}/{max_retries})..."
                         )
+
+                        # Emit retry event
+                        if self._event_bus:
+                            await self._event_bus.emit_node_retry(
+                                stream_id=self._stream_id,
+                                node_id=current_node_id,
+                                retry_count=retry_count,
+                                max_retries=max_retries,
+                                error=result.error or "",
+                            )
+
                         _is_retry = True
                         continue
                     else:
@@ -748,6 +867,7 @@ class GraphExecutor:
                                 "memory": saved_memory,
                                 "execution_path": list(path),
                                 "node_visit_counts": dict(node_visit_counts),
+                                "resume_from": current_node_id,
                             }
 
                             return ExecutionResult(
@@ -774,11 +894,22 @@ class GraphExecutor:
                 # This must happen BEFORE determining next node, since pause nodes may have no edges
                 if node_spec.id in graph.pause_nodes:
                     self.logger.info("💾 Saving session state after pause node")
+
+                    # Emit pause event
+                    if self._event_bus:
+                        await self._event_bus.emit_execution_paused(
+                            stream_id=self._stream_id,
+                            node_id=node_spec.id,
+                            reason="HITL pause node",
+                        )
+
                     saved_memory = memory.read_all()
                     session_state_out = {
                         "paused_at": node_spec.id,
                         "resume_from": f"{node_spec.id}_resume",  # Resume key
                         "memory": saved_memory,
+                        "execution_path": list(path),
+                        "node_visit_counts": dict(node_visit_counts),
                         "next_node": None,  # Will resume from entry point
                     }
 
@@ -827,7 +958,18 @@ class GraphExecutor:
                 if result.next_node:
                     # Router explicitly set next node
                     self.logger.info(f"   → Router directing to: {result.next_node}")
+
+                    # Emit edge traversed event for router-directed edge
+                    if self._event_bus:
+                        await self._event_bus.emit_edge_traversed(
+                            stream_id=self._stream_id,
+                            source_node=current_node_id,
+                            target_node=result.next_node,
+                            edge_condition="router",
+                        )
+
                     current_node_id = result.next_node
+                    self._write_progress(current_node_id, path, memory, node_visit_counts)
                 else:
                     # Get all traversable edges for fan-out detection
                     traversable_edges = self._get_all_traversable_edges(
@@ -849,6 +991,18 @@ class GraphExecutor:
                         targets = [e.target for e in traversable_edges]
                         fan_in_node = self._find_convergence_node(graph, targets)
 
+                        # Emit edge traversed events for fan-out branches
+                        if self._event_bus:
+                            for edge in traversable_edges:
+                                await self._event_bus.emit_edge_traversed(
+                                    stream_id=self._stream_id,
+                                    source_node=current_node_id,
+                                    target_node=edge.target,
+                                    edge_condition=edge.condition.value
+                                    if hasattr(edge.condition, "value")
+                                    else str(edge.condition),
+                                )
+
                         # Execute branches in parallel
                         (
                             _branch_results,
@@ -871,6 +1025,7 @@ class GraphExecutor:
                         if fan_in_node:
                             self.logger.info(f"   ⑃ Fan-in: converging at {fan_in_node}")
                             current_node_id = fan_in_node
+                            self._write_progress(current_node_id, path, memory, node_visit_counts)
                         else:
                             # No convergence point - branches are terminal
                             self.logger.info("   → Parallel branches completed (no convergence)")
@@ -891,6 +1046,14 @@ class GraphExecutor:
                         next_spec = graph.get_node(next_node)
                         self.logger.info(f"   → Next: {next_spec.name if next_spec else next_node}")
 
+                        # Emit edge traversed event for sequential edge
+                        if self._event_bus:
+                            await self._event_bus.emit_edge_traversed(
+                                stream_id=self._stream_id,
+                                source_node=current_node_id,
+                                target_node=next_node,
+                            )
+
                         # CHECKPOINT: node_complete (after determining next node)
                         if (
                             checkpoint_store
@@ -925,6 +1088,84 @@ class GraphExecutor:
 
                         current_node_id = next_node
 
+                # Write progress snapshot at node transition
+                self._write_progress(current_node_id, path, memory, node_visit_counts)
+
+                # Continuous mode: thread conversation forward with transition marker
+                if is_continuous and result.conversation is not None:
+                    continuous_conversation = result.conversation
+
+                    # Look up the next node spec for the transition marker
+                    next_spec = graph.get_node(current_node_id)
+                    if next_spec and next_spec.node_type == "event_loop":
+                        from framework.graph.prompt_composer import (
+                            build_narrative,
+                            build_transition_marker,
+                            compose_system_prompt,
+                        )
+
+                        # Build Layer 2 (narrative) from current state
+                        narrative = build_narrative(memory, path, graph)
+
+                        # Compose new system prompt (Layer 1 + 2 + 3)
+                        new_system = compose_system_prompt(
+                            identity_prompt=getattr(graph, "identity_prompt", None),
+                            focus_prompt=next_spec.system_prompt,
+                            narrative=narrative,
+                        )
+                        continuous_conversation.update_system_prompt(new_system)
+
+                        # Switch conversation store to the next node's directory
+                        # so the transition marker and all subsequent messages are
+                        # persisted there instead of the first node's directory.
+                        if self._storage_path:
+                            from framework.storage.conversation_store import (
+                                FileConversationStore,
+                            )
+
+                            next_store_path = self._storage_path / "conversations" / next_spec.id
+                            next_store = FileConversationStore(base_path=next_store_path)
+                            await continuous_conversation.switch_store(next_store)
+
+                        # Insert transition marker into conversation
+                        data_dir = str(self._storage_path / "data") if self._storage_path else None
+                        marker = build_transition_marker(
+                            previous_node=node_spec,
+                            next_node=next_spec,
+                            memory=memory,
+                            cumulative_tool_names=sorted(cumulative_tool_names),
+                            data_dir=data_dir,
+                        )
+                        await continuous_conversation.add_user_message(
+                            marker,
+                            is_transition_marker=True,
+                        )
+
+                        # Set current phase for phase-aware compaction
+                        continuous_conversation.set_current_phase(next_spec.id)
+
+                        # Opportunistic compaction at transition:
+                        # 1. Prune old tool results (free, no LLM call)
+                        # 2. If still over 80%, do a phase-graduated compact
+                        if continuous_conversation.usage_ratio() > 0.5:
+                            await continuous_conversation.prune_old_tool_results(
+                                protect_tokens=2000,
+                            )
+                        if continuous_conversation.needs_compaction():
+                            self.logger.info(
+                                "   Phase-boundary compaction (%.0f%% usage)",
+                                continuous_conversation.usage_ratio() * 100,
+                            )
+                            summary = (
+                                f"Summary of earlier phases (before {next_spec.name}). "
+                                "See transition markers for phase details."
+                            )
+                            await continuous_conversation.compact(
+                                summary,
+                                keep_recent=4,
+                                phase_graduated=True,
+                            )
+
                 # Update input_data for next node
                 input_data = result.output
 
@@ -978,6 +1219,11 @@ class GraphExecutor:
                 had_partial_failures=len(nodes_failed) > 0,
                 execution_quality=exec_quality,
                 node_visit_counts=dict(node_visit_counts),
+                session_state={
+                    "memory": output,  # output IS memory.read_all()
+                    "execution_path": list(path),
+                    "node_visit_counts": dict(node_visit_counts),
+                },
             )
 
         except asyncio.CancelledError:
@@ -1067,6 +1313,7 @@ class GraphExecutor:
                 "memory": saved_memory,
                 "execution_path": list(path),
                 "node_visit_counts": dict(node_visit_counts),
+                "resume_from": current_node_id,
             }
 
             # Mark latest checkpoint for resume on failure
@@ -1119,12 +1366,20 @@ class GraphExecutor:
         goal: Goal,
         input_data: dict[str, Any],
         max_tokens: int = 4096,
+        continuous_mode: bool = False,
+        inherited_conversation: Any = None,
+        override_tools: list | None = None,
+        cumulative_output_keys: list[str] | None = None,
     ) -> NodeContext:
         """Build execution context for a node."""
         # Filter tools to those available to this node
-        available_tools = []
-        if node_spec.tools:
-            available_tools = [t for t in self.tools if t.name in node_spec.tools]
+        if override_tools is not None:
+            # Continuous mode: use cumulative tool set
+            available_tools = list(override_tools)
+        else:
+            available_tools = []
+            if node_spec.tools:
+                available_tools = [t for t in self.tools if t.name in node_spec.tools]
 
         # Create scoped memory view
         scoped_memory = memory.with_permissions(
@@ -1145,6 +1400,9 @@ class GraphExecutor:
             max_tokens=max_tokens,
             runtime_logger=self.runtime_logger,
             pause_event=self._pause_requested,  # Pass pause event for granular control
+            continuous_mode=continuous_mode,
+            inherited_conversation=inherited_conversation,
+            cumulative_output_keys=cumulative_output_keys or [],
         )
 
     # Valid node types - no ambiguous "llm" type allowed

core/framework/graph/node.py

@@ -238,6 +238,16 @@ class NodeSpec(BaseModel):
         description="If True, this node streams output to the end user and can request input.",
     )
 
+    # Phase completion criteria for conversation-aware judge (Level 2)
+    success_criteria: str | None = Field(
+        default=None,
+        description=(
+            "Natural-language criteria for phase completion. When set, the "
+            "implicit judge upgrades to Level 2: after output keys are satisfied, "
+            "a fast LLM evaluates whether the conversation meets these criteria."
+        ),
+    )
+
     model_config = {"extra": "allow", "arbitrary_types_allowed": True}
 
 
@@ -483,6 +493,11 @@ class NodeContext:
     # Pause control (optional) - asyncio.Event for pause requests
     pause_event: Any = None  # asyncio.Event | None
 
+    # Continuous conversation mode
+    continuous_mode: bool = False  # True when graph has conversation_mode="continuous"
+    inherited_conversation: Any = None  # NodeConversation | None (from prior node)
+    cumulative_output_keys: list[str] = field(default_factory=list)  # All output keys from path
+
 
 @dataclass
 class NodeResult:
@@ -511,6 +526,9 @@ class NodeResult:
     # Pydantic validation errors (if any)
     validation_errors: list[str] = field(default_factory=list)
 
+    # Continuous conversation mode: return conversation for threading to next node
+    conversation: Any = None  # NodeConversation | None
+
     def to_summary(self, node_spec: Any = None) -> str:
         """
         Generate a human-readable summary of this node's execution and output.
@@ -1134,7 +1152,7 @@ Keep the same JSON structure but with shorter content values.
                 decision_id=decision_id,
                 success=True,
                 result=response.content,
-                tokens_used=response.input_tokens + response.output_tokens,
+                tokens_used=total_input_tokens + total_output_tokens,
                 latency_ms=latency_ms,
             )
 
@@ -1233,7 +1251,7 @@ Keep the same JSON structure but with shorter content values.
                         success=False,
                         error=_extraction_error,
                         output={},
-                        tokens_used=response.input_tokens + response.output_tokens,
+                        tokens_used=total_input_tokens + total_output_tokens,
                         latency_ms=latency_ms,
                     )
                     # JSON extraction failed completely - still strip code blocks
@@ -1275,7 +1293,7 @@ Keep the same JSON structure but with shorter content values.
             return NodeResult(
                 success=True,
                 output=output,
-                tokens_used=response.input_tokens + response.output_tokens,
+                tokens_used=total_input_tokens + total_output_tokens,
                 latency_ms=latency_ms,
             )
 

core/framework/graph/prompt_composer.py

@@ -0,0 +1,185 @@
+"""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 pathlib import Path
+from typing import TYPE_CHECKING
+
+if TYPE_CHECKING:
+    from framework.graph.edge import GraphSpec
+    from framework.graph.node import NodeSpec, SharedMemory
+
+logger = logging.getLogger(__name__)
+
+
+def compose_system_prompt(
+    identity_prompt: str | None,
+    focus_prompt: str | None,
+    narrative: 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.
+
+    Returns:
+        Composed system prompt with all layers present.
+    """
+    parts: list[str] = []
+
+    # Layer 1: Identity (always first, anchors the personality)
+    if identity_prompt:
+        parts.append(identity_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 "\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,
+) -> 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.
+
+    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 load_data to access):\n" + "\n".join(file_lines)
+                    )
+
+    # 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/graph/safe_eval.py

@@ -145,7 +145,7 @@ class SafeEvalVisitor(ast.NodeVisitor):
 
     def visit_Attribute(self, node: ast.Attribute) -> Any:
         # value.attr
-        # STIRCT CHECK: No access to private attributes (starting with _)
+        # STRICT CHECK: No access to private attributes (starting with _)
         if node.attr.startswith("_"):
             raise ValueError(f"Access to private attribute '{node.attr}' is not allowed")
 

core/framework/llm/anthropic.py

@@ -70,6 +70,7 @@ class AnthropicProvider(LLMProvider):
         max_tokens: int = 1024,
         response_format: dict[str, Any] | None = None,
         json_mode: bool = False,
+        max_retries: int | None = None,
     ) -> LLMResponse:
         """Generate a completion from Claude (via LiteLLM)."""
         return self._provider.complete(
@@ -79,6 +80,7 @@ class AnthropicProvider(LLMProvider):
             max_tokens=max_tokens,
             response_format=response_format,
             json_mode=json_mode,
+            max_retries=max_retries,
         )
 
     def complete_with_tools(

core/framework/llm/litellm.py

@@ -150,10 +150,13 @@ class LiteLLMProvider(LLMProvider):
                 "LiteLLM is not installed. Please install it with: uv pip install litellm"
             )
 
-    def _completion_with_rate_limit_retry(self, **kwargs: Any) -> Any:
+    def _completion_with_rate_limit_retry(
+        self, max_retries: int | None = None, **kwargs: Any
+    ) -> Any:
         """Call litellm.completion with retry on 429 rate limit errors and empty responses."""
         model = kwargs.get("model", self.model)
-        for attempt in range(RATE_LIMIT_MAX_RETRIES + 1):
+        retries = max_retries if max_retries is not None else RATE_LIMIT_MAX_RETRIES
+        for attempt in range(retries + 1):
             try:
                 response = litellm.completion(**kwargs)  # type: ignore[union-attr]
 
@@ -194,9 +197,9 @@ class LiteLLMProvider(LLMProvider):
                         f"Full request dumped to: {dump_path}"
                     )
 
-                    if attempt == RATE_LIMIT_MAX_RETRIES:
+                    if attempt == retries:
                         logger.error(
-                            f"[retry] GAVE UP on {model} after {RATE_LIMIT_MAX_RETRIES + 1} "
+                            f"[retry] GAVE UP on {model} after {retries + 1} "
                             f"attempts — empty response "
                             f"(finish_reason={finish_reason}, "
                             f"choices={len(response.choices) if response.choices else 0})"
@@ -209,7 +212,7 @@ class LiteLLMProvider(LLMProvider):
                         f"choices={len(response.choices) if response.choices else 0}) — "
                         f"likely rate limited or quota exceeded. "
                         f"Retrying in {wait}s "
-                        f"(attempt {attempt + 1}/{RATE_LIMIT_MAX_RETRIES})"
+                        f"(attempt {attempt + 1}/{retries})"
                     )
                     time.sleep(wait)
                     continue
@@ -225,9 +228,9 @@ class LiteLLMProvider(LLMProvider):
                     error_type="rate_limit",
                     attempt=attempt,
                 )
-                if attempt == RATE_LIMIT_MAX_RETRIES:
+                if attempt == retries:
                     logger.error(
-                        f"[retry] GAVE UP on {model} after {RATE_LIMIT_MAX_RETRIES + 1} "
+                        f"[retry] GAVE UP on {model} after {retries + 1} "
                         f"attempts — rate limit error: {e!s}. "
                         f"~{token_count} tokens ({token_method}). "
                         f"Full request dumped to: {dump_path}"
@@ -239,7 +242,7 @@ class LiteLLMProvider(LLMProvider):
                     f"~{token_count} tokens ({token_method}). "
                     f"Full request dumped to: {dump_path}. "
                     f"Retrying in {wait}s "
-                    f"(attempt {attempt + 1}/{RATE_LIMIT_MAX_RETRIES})"
+                    f"(attempt {attempt + 1}/{retries})"
                 )
                 time.sleep(wait)
         # unreachable, but satisfies type checker
@@ -253,6 +256,7 @@ class LiteLLMProvider(LLMProvider):
         max_tokens: int = 1024,
         response_format: dict[str, Any] | None = None,
         json_mode: bool = False,
+        max_retries: int | None = None,
     ) -> LLMResponse:
         """Generate a completion using LiteLLM."""
         # Prepare messages with system prompt
@@ -293,12 +297,17 @@ class LiteLLMProvider(LLMProvider):
             kwargs["response_format"] = response_format
 
         # Make the call
-        response = self._completion_with_rate_limit_retry(**kwargs)
+        response = self._completion_with_rate_limit_retry(max_retries=max_retries, **kwargs)
 
         # Extract content
         content = response.choices[0].message.content or ""
 
-        # Get usage info
+        # Get usage info.
+        # NOTE: completion_tokens includes reasoning/thinking tokens for models
+        # that use them (o1, gpt-5-mini, etc.). LiteLLM does not reliably expose
+        # usage.completion_tokens_details.reasoning_tokens across all providers.
+        # This means output_tokens may be inflated for reasoning models.
+        # Compaction is unaffected — it uses prompt_tokens (input-side only).
         usage = response.usage
         input_tokens = usage.prompt_tokens if usage else 0
         output_tokens = usage.completion_tokens if usage else 0

core/framework/llm/mock.py

@@ -120,6 +120,7 @@ class MockLLMProvider(LLMProvider):
         max_tokens: int = 1024,
         response_format: dict[str, Any] | None = None,
         json_mode: bool = False,
+        max_retries: int | None = None,
     ) -> LLMResponse:
         """
         Generate a mock completion without calling a real LLM.

core/framework/llm/provider.py

@@ -65,6 +65,7 @@ class LLMProvider(ABC):
         max_tokens: int = 1024,
         response_format: dict[str, Any] | None = None,
         json_mode: bool = False,
+        max_retries: int | None = None,
     ) -> LLMResponse:
         """
         Generate a completion from the LLM.
@@ -79,6 +80,8 @@ class LLMProvider(ABC):
                 - {"type": "json_schema", "json_schema": {"name": "...", "schema": {...}}}
                   for strict JSON schema enforcement
             json_mode: If True, request structured JSON output from the LLM
+            max_retries: Override retry count for rate-limit/empty-response retries.
+                None uses the provider default.
 
         Returns:
             LLMResponse with content and metadata

core/framework/mcp/agent_builder_server.py

@@ -8,19 +8,22 @@ Usage:
 """
 
 import json
+import logging
 import os
 import sys
 from datetime import datetime
 from pathlib import Path
 from typing import Annotated
 
+# Project root resolution.  This file lives at core/framework/mcp/agent_builder_server.py,
+# so the project root (where exports/ lives) is four parents up.
+_PROJECT_ROOT = Path(__file__).resolve().parent.parent.parent.parent
+
 # Ensure exports/ is on sys.path so AgentRunner can import agent modules.
-_framework_dir = Path(__file__).resolve().parent.parent  # core/framework/ -> core/
-_project_root = _framework_dir.parent  # core/ -> project root
-_exports_dir = _project_root / "exports"
+_exports_dir = _PROJECT_ROOT / "exports"
 if _exports_dir.is_dir() and str(_exports_dir) not in sys.path:
     sys.path.insert(0, str(_exports_dir))
-del _framework_dir, _project_root, _exports_dir
+del _exports_dir
 
 from mcp.server import FastMCP  # noqa: E402
 from pydantic import ValidationError  # noqa: E402
@@ -176,8 +179,8 @@ def _load_active_session() -> BuildSession | None:
 
         if session_id:
             return _load_session(session_id)
-    except Exception:
-        pass
+    except Exception as e:
+        logging.warning("Failed to load active session: %s", e)
 
     return None
 
@@ -542,6 +545,9 @@ def _validate_agent_path(agent_path: str) -> tuple[Path | None, str | None]:
     """
     Validate and normalize agent_path.
 
+    Resolves relative paths against _PROJECT_ROOT since the MCP server's
+    cwd (core/) differs from the user's cwd (project root).
+
     Returns:
         (Path, None) if valid
         (None, error_json) if invalid
@@ -556,6 +562,12 @@ def _validate_agent_path(agent_path: str) -> tuple[Path | None, str | None]:
 
     path = Path(agent_path)
 
+    # Resolve relative paths against project root (not MCP server's cwd)
+    if not path.is_absolute() and not path.exists():
+        resolved = _PROJECT_ROOT / path
+        if resolved.exists():
+            path = resolved
+
     if not path.exists():
         return None, json.dumps(
             {
@@ -3019,18 +3031,15 @@ def _format_success_criteria(criteria: list[SuccessCriterion]) -> str:
 
 # Test template for Claude to use when writing tests
 CONSTRAINT_TEST_TEMPLATE = '''@pytest.mark.asyncio
-async def test_constraint_{constraint_id}_{scenario}(mock_mode):
+async def test_constraint_{constraint_id}_{scenario}(runner, auto_responder, mock_mode):
     """Test: {description}"""
-    result = await default_agent.run({{"key": "value"}}, mock_mode=mock_mode)
-
-    # IMPORTANT: result is an ExecutionResult object with these attributes:
-    # - result.success: bool - whether the agent succeeded
-    # - result.output: dict - the agent's output data (access data here!)
-    # - result.error: str or None - error message if failed
+    await auto_responder.start()
+    try:
+        result = await runner.run({{"key": "value"}})
+    finally:
+        await auto_responder.stop()
 
     assert result.success, f"Agent failed: {{result.error}}"
-
-    # Access output data via result.output
     output_data = result.output or {{}}
 
     # Add constraint-specific assertions here
@@ -3038,18 +3047,15 @@ async def test_constraint_{constraint_id}_{scenario}(mock_mode):
 '''
 
 SUCCESS_TEST_TEMPLATE = '''@pytest.mark.asyncio
-async def test_success_{criteria_id}_{scenario}(mock_mode):
+async def test_success_{criteria_id}_{scenario}(runner, auto_responder, mock_mode):
     """Test: {description}"""
-    result = await default_agent.run({{"key": "value"}}, mock_mode=mock_mode)
-
-    # IMPORTANT: result is an ExecutionResult object with these attributes:
-    # - result.success: bool - whether the agent succeeded
-    # - result.output: dict - the agent's output data (access data here!)
-    # - result.error: str or None - error message if failed
+    await auto_responder.start()
+    try:
+        result = await runner.run({{"key": "value"}})
+    finally:
+        await auto_responder.stop()
 
     assert result.success, f"Agent failed: {{result.error}}"
-
-    # Access output data via result.output
     output_data = result.output or {{}}
 
     # Add success criteria-specific assertions here
@@ -3105,7 +3111,6 @@ def generate_constraint_tests(
         test_type="Constraint",
         agent_name=agent_module,
         description=f"Tests for constraints defined in goal: {goal.name}",
-        agent_module=agent_module,
     )
 
     # Return guidelines + data for Claude to write tests directly
@@ -3121,14 +3126,22 @@ def generate_constraint_tests(
                 "max_tests": 5,
                 "naming_convention": "test_constraint_<constraint_id>_<scenario>",
                 "required_decorator": "@pytest.mark.asyncio",
-                "required_fixture": "mock_mode",
-                "agent_call_pattern": "await default_agent.run(input_dict, mock_mode=mock_mode)",
+                "required_fixtures": "runner, auto_responder, mock_mode",
+                "agent_call_pattern": "await runner.run(input_dict)",
+                "auto_responder_pattern": (
+                    "await auto_responder.start()\n"
+                    "try:\n"
+                    "    result = await runner.run(input_dict)\n"
+                    "finally:\n"
+                    "    await auto_responder.stop()"
+                ),
                 "result_type": "ExecutionResult with .success, .output (dict), .error",
                 "critical_rules": [
                     "Every test function MUST be async with @pytest.mark.asyncio",
-                    "Every test MUST accept mock_mode as a parameter",
-                    "Use await default_agent.run(input, mock_mode=mock_mode)",
-                    "default_agent is already imported - do NOT add imports",
+                    "Every test MUST accept runner, auto_responder, and mock_mode fixtures",
+                    "Use await runner.run(input) -- NOT default_agent.run()",
+                    "Start auto_responder before running, stop in finally block",
+                    "runner and auto_responder are from conftest.py -- do NOT import them",
                     "NEVER call result.get() - use result.output.get() instead",
                     "Always check result.success before accessing result.output",
                 ],
@@ -3192,7 +3205,6 @@ def generate_success_tests(
         test_type="Success criteria",
         agent_name=agent_module,
         description=f"Tests for success criteria defined in goal: {goal.name}",
-        agent_module=agent_module,
     )
 
     # Return guidelines + data for Claude to write tests directly
@@ -3214,14 +3226,22 @@ def generate_success_tests(
                 "max_tests": 12,
                 "naming_convention": "test_success_<criteria_id>_<scenario>",
                 "required_decorator": "@pytest.mark.asyncio",
-                "required_fixture": "mock_mode",
-                "agent_call_pattern": "await default_agent.run(input_dict, mock_mode=mock_mode)",
+                "required_fixtures": "runner, auto_responder, mock_mode",
+                "agent_call_pattern": "await runner.run(input_dict)",
+                "auto_responder_pattern": (
+                    "await auto_responder.start()\n"
+                    "try:\n"
+                    "    result = await runner.run(input_dict)\n"
+                    "finally:\n"
+                    "    await auto_responder.stop()"
+                ),
                 "result_type": "ExecutionResult with .success, .output (dict), .error",
                 "critical_rules": [
                     "Every test function MUST be async with @pytest.mark.asyncio",
-                    "Every test MUST accept mock_mode as a parameter",
-                    "Use await default_agent.run(input, mock_mode=mock_mode)",
-                    "default_agent is already imported - do NOT add imports",
+                    "Every test MUST accept runner, auto_responder, and mock_mode fixtures",
+                    "Use await runner.run(input) -- NOT default_agent.run()",
+                    "Start auto_responder before running, stop in finally block",
+                    "runner and auto_responder are from conftest.py -- do NOT import them",
                     "NEVER call result.get() - use result.output.get() instead",
                     "Always check result.success before accessing result.output",
                 ],
@@ -3318,11 +3338,13 @@ def run_tests(
     # Add short traceback and quiet summary
     cmd.append("--tb=short")
 
-    # Set PYTHONPATH to project root so agents can import from core.framework
+    # Set PYTHONPATH so framework and agent packages are importable
     env = os.environ.copy()
     pythonpath = env.get("PYTHONPATH", "")
     project_root = Path(__file__).parent.parent.parent.parent.resolve()
-    env["PYTHONPATH"] = f"{project_root}:{pythonpath}"
+    core_path = project_root / "core"
+    exports_path = project_root / "exports"
+    env["PYTHONPATH"] = f"{core_path}:{exports_path}:{project_root}:{pythonpath}"
 
     # Run pytest
     try:
@@ -3792,7 +3814,11 @@ def check_missing_credentials(
 
         from framework.runner import AgentRunner
 
-        runner = AgentRunner.load(agent_path)
+        path, err = _validate_agent_path(agent_path)
+        if err:
+            return err
+
+        runner = AgentRunner.load(str(path))
         runner.validate()
 
         store = _get_credential_store()
@@ -3992,7 +4018,11 @@ def verify_credentials(
     try:
         from framework.runner import AgentRunner
 
-        runner = AgentRunner.load(agent_path)
+        path, err = _validate_agent_path(agent_path)
+        if err:
+            return err
+
+        runner = AgentRunner.load(str(path))
         validation = runner.validate()
 
         return json.dumps(
@@ -4009,6 +4039,382 @@ def verify_credentials(
         return json.dumps({"error": str(e)})
 
 
+# =============================================================================
+# SESSION & CHECKPOINT TOOLS (read-only, no build session required)
+# =============================================================================
+
+_MAX_DIFF_VALUE_LEN = 500
+
+
+def _read_session_json(path: Path) -> dict | None:
+    """Read a JSON file, returning None on failure."""
+    if not path.exists():
+        return None
+    try:
+        return json.loads(path.read_text(encoding="utf-8"))
+    except (json.JSONDecodeError, OSError):
+        return None
+
+
+def _scan_agent_sessions(agent_work_dir: Path) -> list[tuple[str, Path]]:
+    """Find session directories with state.json, sorted most-recent-first."""
+    sessions: list[tuple[str, Path]] = []
+    sessions_dir = agent_work_dir / "sessions"
+    if not sessions_dir.exists():
+        return sessions
+    for session_dir in sessions_dir.iterdir():
+        if session_dir.is_dir() and session_dir.name.startswith("session_"):
+            state_path = session_dir / "state.json"
+            if state_path.exists():
+                sessions.append((session_dir.name, state_path))
+    sessions.sort(key=lambda t: t[0], reverse=True)
+    return sessions
+
+
+def _truncate_value(value: object, max_len: int = _MAX_DIFF_VALUE_LEN) -> object:
+    """Truncate a value's JSON representation if too long."""
+    s = json.dumps(value, default=str)
+    if len(s) <= max_len:
+        return value
+    return {"_truncated": True, "_preview": s[:max_len] + "...", "_length": len(s)}
+
+
+@mcp.tool()
+def list_agent_sessions(
+    agent_work_dir: Annotated[
+        str,
+        "Path to the agent's working directory (e.g., ~/.hive/agents/my_agent)",
+    ],
+    status: Annotated[
+        str,
+        "Filter by status: 'active', 'paused', 'completed', 'failed', 'cancelled'. Empty for all.",
+    ] = "",
+    limit: Annotated[int, "Maximum number of results (default 20)"] = 20,
+    offset: Annotated[int, "Number of sessions to skip for pagination"] = 0,
+) -> str:
+    """
+    List sessions for an agent with optional status filter.
+
+    Use this to discover which sessions exist, find resumable sessions,
+    or identify failed sessions for debugging. Combines well with
+    query_runtime_logs for correlating session state with log data.
+    """
+    work_dir = Path(agent_work_dir)
+    all_sessions = _scan_agent_sessions(work_dir)
+
+    if not all_sessions:
+        return json.dumps({"sessions": [], "total": 0, "offset": offset, "limit": limit})
+
+    summaries = []
+    for session_id, state_path in all_sessions:
+        data = _read_session_json(state_path)
+        if data is None:
+            continue
+
+        session_status = data.get("status", "")
+        if status and session_status != status:
+            continue
+
+        timestamps = data.get("timestamps", {})
+        progress = data.get("progress", {})
+        checkpoint_dir = state_path.parent / "checkpoints"
+
+        summaries.append(
+            {
+                "session_id": session_id,
+                "status": session_status,
+                "goal_id": data.get("goal_id", ""),
+                "started_at": timestamps.get("started_at", ""),
+                "updated_at": timestamps.get("updated_at", ""),
+                "completed_at": timestamps.get("completed_at"),
+                "is_resumable": data.get("is_resumable", False),
+                "is_resumable_from_checkpoint": data.get("is_resumable_from_checkpoint", False),
+                "current_node": progress.get("current_node"),
+                "paused_at": progress.get("paused_at"),
+                "steps_executed": progress.get("steps_executed", 0),
+                "execution_quality": progress.get("execution_quality", ""),
+                "has_checkpoints": checkpoint_dir.exists()
+                and any(checkpoint_dir.glob("cp_*.json")),
+            }
+        )
+
+    total = len(summaries)
+    page = summaries[offset : offset + limit]
+    return json.dumps(
+        {"sessions": page, "total": total, "offset": offset, "limit": limit}, indent=2
+    )
+
+
+@mcp.tool()
+def get_agent_session_state(
+    agent_work_dir: Annotated[str, "Path to the agent's working directory"],
+    session_id: Annotated[str, "The session ID (e.g., 'session_20260208_143022_abc12345')"],
+) -> str:
+    """
+    Load full session state for a specific session.
+
+    Returns complete session data including status, progress, result,
+    metrics, and checkpoint info. Memory values are excluded to prevent
+    context bloat -- use get_agent_session_memory to retrieve memory contents.
+    """
+    state_path = Path(agent_work_dir) / "sessions" / session_id / "state.json"
+    data = _read_session_json(state_path)
+    if data is None:
+        return json.dumps({"error": f"Session not found: {session_id}"})
+
+    memory = data.get("memory", {})
+    data["memory_keys"] = list(memory.keys()) if isinstance(memory, dict) else []
+    data["memory_size"] = len(memory) if isinstance(memory, dict) else 0
+    data.pop("memory", None)
+
+    return json.dumps(data, indent=2, default=str)
+
+
+@mcp.tool()
+def get_agent_session_memory(
+    agent_work_dir: Annotated[str, "Path to the agent's working directory"],
+    session_id: Annotated[str, "The session ID"],
+    key: Annotated[str, "Specific memory key to retrieve. Empty for all."] = "",
+) -> str:
+    """
+    Get memory contents from a session.
+
+    Memory stores intermediate results passed between nodes. Use this
+    to inspect what data was produced during execution.
+
+    If key is provided, returns only that memory key's value.
+    If key is empty, returns all memory keys and their values.
+    """
+    state_path = Path(agent_work_dir) / "sessions" / session_id / "state.json"
+    data = _read_session_json(state_path)
+    if data is None:
+        return json.dumps({"error": f"Session not found: {session_id}"})
+
+    memory = data.get("memory", {})
+    if not isinstance(memory, dict):
+        memory = {}
+
+    if key:
+        if key not in memory:
+            return json.dumps(
+                {
+                    "error": f"Memory key not found: '{key}'",
+                    "available_keys": list(memory.keys()),
+                }
+            )
+        value = memory[key]
+        return json.dumps(
+            {
+                "session_id": session_id,
+                "key": key,
+                "value": value,
+                "value_type": type(value).__name__,
+            },
+            indent=2,
+            default=str,
+        )
+
+    return json.dumps(
+        {"session_id": session_id, "memory": memory, "total_keys": len(memory)},
+        indent=2,
+        default=str,
+    )
+
+
+@mcp.tool()
+def list_agent_checkpoints(
+    agent_work_dir: Annotated[str, "Path to the agent's working directory"],
+    session_id: Annotated[str, "The session ID to list checkpoints for"],
+    checkpoint_type: Annotated[
+        str,
+        "Filter by type: 'node_start', 'node_complete', 'loop_iteration'. Empty for all.",
+    ] = "",
+    is_clean: Annotated[str, "Filter by clean status: 'true', 'false', or empty for all."] = "",
+) -> str:
+    """
+    List checkpoints for a specific session.
+
+    Checkpoints capture execution state at node boundaries for
+    crash recovery and resume. Use with get_agent_checkpoint for
+    detailed checkpoint inspection.
+    """
+    session_dir = Path(agent_work_dir) / "sessions" / session_id
+    checkpoint_dir = session_dir / "checkpoints"
+
+    if not session_dir.exists():
+        return json.dumps({"error": f"Session not found: {session_id}"})
+
+    if not checkpoint_dir.exists():
+        return json.dumps(
+            {
+                "session_id": session_id,
+                "checkpoints": [],
+                "total": 0,
+                "latest_checkpoint_id": None,
+            }
+        )
+
+    # Try index.json first
+    index_data = _read_session_json(checkpoint_dir / "index.json")
+    if index_data and "checkpoints" in index_data:
+        checkpoints = index_data["checkpoints"]
+    else:
+        # Fallback: scan individual checkpoint files
+        checkpoints = []
+        for cp_file in sorted(checkpoint_dir.glob("cp_*.json")):
+            cp_data = _read_session_json(cp_file)
+            if cp_data:
+                checkpoints.append(
+                    {
+                        "checkpoint_id": cp_data.get("checkpoint_id", cp_file.stem),
+                        "checkpoint_type": cp_data.get("checkpoint_type", ""),
+                        "created_at": cp_data.get("created_at", ""),
+                        "current_node": cp_data.get("current_node"),
+                        "next_node": cp_data.get("next_node"),
+                        "is_clean": cp_data.get("is_clean", True),
+                        "description": cp_data.get("description", ""),
+                    }
+                )
+
+    # Apply filters
+    if checkpoint_type:
+        checkpoints = [c for c in checkpoints if c.get("checkpoint_type") == checkpoint_type]
+    if is_clean:
+        clean_val = is_clean.lower() == "true"
+        checkpoints = [c for c in checkpoints if c.get("is_clean") == clean_val]
+
+    latest_id = None
+    if index_data:
+        latest_id = index_data.get("latest_checkpoint_id")
+    elif checkpoints:
+        latest_id = checkpoints[-1].get("checkpoint_id")
+
+    return json.dumps(
+        {
+            "session_id": session_id,
+            "checkpoints": checkpoints,
+            "total": len(checkpoints),
+            "latest_checkpoint_id": latest_id,
+        },
+        indent=2,
+    )
+
+
+@mcp.tool()
+def get_agent_checkpoint(
+    agent_work_dir: Annotated[str, "Path to the agent's working directory"],
+    session_id: Annotated[str, "The session ID"],
+    checkpoint_id: Annotated[str, "Specific checkpoint ID, or empty for latest"] = "",
+) -> str:
+    """
+    Load a specific checkpoint with full state data.
+
+    Returns the complete checkpoint including shared memory snapshot,
+    execution path, accumulated outputs, and metrics. If checkpoint_id
+    is empty, loads the latest checkpoint.
+    """
+    session_dir = Path(agent_work_dir) / "sessions" / session_id
+    checkpoint_dir = session_dir / "checkpoints"
+
+    if not checkpoint_dir.exists():
+        return json.dumps({"error": f"No checkpoints found for session: {session_id}"})
+
+    if not checkpoint_id:
+        index_data = _read_session_json(checkpoint_dir / "index.json")
+        if index_data and index_data.get("latest_checkpoint_id"):
+            checkpoint_id = index_data["latest_checkpoint_id"]
+        else:
+            cp_files = sorted(checkpoint_dir.glob("cp_*.json"))
+            if not cp_files:
+                return json.dumps({"error": f"No checkpoints found for session: {session_id}"})
+            checkpoint_id = cp_files[-1].stem
+
+    cp_path = checkpoint_dir / f"{checkpoint_id}.json"
+    data = _read_session_json(cp_path)
+    if data is None:
+        return json.dumps({"error": f"Checkpoint not found: {checkpoint_id}"})
+
+    return json.dumps(data, indent=2, default=str)
+
+
+@mcp.tool()
+def compare_agent_checkpoints(
+    agent_work_dir: Annotated[str, "Path to the agent's working directory"],
+    session_id: Annotated[str, "The session ID"],
+    checkpoint_id_before: Annotated[str, "The earlier checkpoint ID"],
+    checkpoint_id_after: Annotated[str, "The later checkpoint ID"],
+) -> str:
+    """
+    Compare memory state between two checkpoints.
+
+    Shows what memory keys were added, removed, or changed between
+    two points in execution. Useful for understanding how data flows
+    through the agent graph.
+    """
+    checkpoint_dir = Path(agent_work_dir) / "sessions" / session_id / "checkpoints"
+
+    before = _read_session_json(checkpoint_dir / f"{checkpoint_id_before}.json")
+    if before is None:
+        return json.dumps({"error": f"Checkpoint not found: {checkpoint_id_before}"})
+
+    after = _read_session_json(checkpoint_dir / f"{checkpoint_id_after}.json")
+    if after is None:
+        return json.dumps({"error": f"Checkpoint not found: {checkpoint_id_after}"})
+
+    mem_before = before.get("shared_memory", {})
+    mem_after = after.get("shared_memory", {})
+
+    keys_before = set(mem_before.keys())
+    keys_after = set(mem_after.keys())
+
+    added = {k: _truncate_value(mem_after[k]) for k in keys_after - keys_before}
+    removed = list(keys_before - keys_after)
+    unchanged = []
+    changed = {}
+
+    for k in keys_before & keys_after:
+        if mem_before[k] == mem_after[k]:
+            unchanged.append(k)
+        else:
+            changed[k] = {
+                "before": _truncate_value(mem_before[k]),
+                "after": _truncate_value(mem_after[k]),
+            }
+
+    path_before = before.get("execution_path", [])
+    path_after = after.get("execution_path", [])
+    new_nodes = path_after[len(path_before) :]
+
+    return json.dumps(
+        {
+            "session_id": session_id,
+            "before": {
+                "checkpoint_id": checkpoint_id_before,
+                "current_node": before.get("current_node"),
+                "created_at": before.get("created_at", ""),
+            },
+            "after": {
+                "checkpoint_id": checkpoint_id_after,
+                "current_node": after.get("current_node"),
+                "created_at": after.get("created_at", ""),
+            },
+            "memory_diff": {
+                "added": added,
+                "removed": removed,
+                "changed": changed,
+                "unchanged": unchanged,
+            },
+            "execution_path_diff": {
+                "new_nodes": new_nodes,
+                "path_before": path_before,
+                "path_after": path_after,
+            },
+        },
+        indent=2,
+        default=str,
+    )
+
+
 # =============================================================================
 # MAIN
 # =============================================================================

core/framework/runner/cli.py

@@ -332,10 +332,67 @@ def register_commands(subparsers: argparse._SubParsersAction) -> None:
     resume_parser.set_defaults(func=cmd_resume)
 
 
+def _load_resume_state(
+    agent_path: str, session_id: str, checkpoint_id: str | None = None
+) -> dict | None:
+    """Load session or checkpoint state for headless resume.
+
+    Args:
+        agent_path: Path to the agent folder (e.g., exports/my_agent)
+        session_id: Session ID to resume from
+        checkpoint_id: Optional checkpoint ID within the session
+
+    Returns:
+        session_state dict for executor, or None if not found
+    """
+    agent_name = Path(agent_path).name
+    agent_work_dir = Path.home() / ".hive" / "agents" / agent_name
+    session_dir = agent_work_dir / "sessions" / session_id
+
+    if not session_dir.exists():
+        return None
+
+    if checkpoint_id:
+        # Checkpoint-based resume: load checkpoint and extract state
+        cp_path = session_dir / "checkpoints" / f"{checkpoint_id}.json"
+        if not cp_path.exists():
+            return None
+        try:
+            cp_data = json.loads(cp_path.read_text())
+        except (json.JSONDecodeError, OSError):
+            return None
+        return {
+            "resume_session_id": session_id,
+            "memory": cp_data.get("shared_memory", {}),
+            "paused_at": cp_data.get("next_node") or cp_data.get("current_node"),
+            "execution_path": cp_data.get("execution_path", []),
+            "node_visit_counts": {},
+        }
+    else:
+        # Session state resume: load state.json
+        state_path = session_dir / "state.json"
+        if not state_path.exists():
+            return None
+        try:
+            state_data = json.loads(state_path.read_text())
+        except (json.JSONDecodeError, OSError):
+            return None
+        progress = state_data.get("progress", {})
+        paused_at = progress.get("paused_at") or progress.get("resume_from")
+        return {
+            "resume_session_id": session_id,
+            "memory": state_data.get("memory", {}),
+            "paused_at": paused_at,
+            "execution_path": progress.get("path", []),
+            "node_visit_counts": progress.get("node_visit_counts", {}),
+        }
+
+
 def cmd_run(args: argparse.Namespace) -> int:
     """Run an exported agent."""
     import logging
 
+    from framework.credentials.models import CredentialError
     from framework.runner import AgentRunner
 
     # Set logging level (quiet by default for cleaner output)
@@ -374,8 +431,10 @@ def cmd_run(args: argparse.Namespace) -> int:
                     runner = AgentRunner.load(
                         args.agent_path,
                         model=args.model,
-                        enable_tui=True,
                     )
+                except CredentialError as e:
+                    print(f"\n{e}", file=sys.stderr)
+                    return
                 except Exception as e:
                     print(f"Error loading agent: {e}")
                     return
@@ -415,12 +474,35 @@ def cmd_run(args: argparse.Namespace) -> int:
             runner = AgentRunner.load(
                 args.agent_path,
                 model=args.model,
-                enable_tui=False,
             )
+        except CredentialError as e:
+            print(f"\n{e}", file=sys.stderr)
+            return 1
         except FileNotFoundError as e:
             print(f"Error: {e}", file=sys.stderr)
             return 1
 
+        # Load session/checkpoint state for resume (headless mode)
+        session_state = None
+        resume_session = getattr(args, "resume_session", None)
+        checkpoint = getattr(args, "checkpoint", None)
+        if resume_session:
+            session_state = _load_resume_state(args.agent_path, resume_session, checkpoint)
+            if session_state is None:
+                print(
+                    f"Error: Could not load session state for {resume_session}",
+                    file=sys.stderr,
+                )
+                return 1
+            if not args.quiet:
+                resume_node = session_state.get("paused_at", "unknown")
+                if checkpoint:
+                    print(f"Resuming from checkpoint: {checkpoint}")
+                else:
+                    print(f"Resuming session: {resume_session}")
+                print(f"Resume point: {resume_node}")
+                print()
+
         # Auto-inject user_id if the agent expects it but it's not provided
         entry_input_keys = runner.graph.nodes[0].input_keys if runner.graph.nodes else []
         if "user_id" in entry_input_keys and context.get("user_id") is None:
@@ -440,7 +522,7 @@ def cmd_run(args: argparse.Namespace) -> int:
             print("=" * 60)
             print()
 
-        result = asyncio.run(runner.run(context))
+        result = asyncio.run(runner.run(context, session_state=session_state))
 
     # Format output
     output = {
@@ -520,10 +602,14 @@ def cmd_run(args: argparse.Namespace) -> int:
 
 def cmd_info(args: argparse.Namespace) -> int:
     """Show agent information."""
+    from framework.credentials.models import CredentialError
     from framework.runner import AgentRunner
 
     try:
         runner = AgentRunner.load(args.agent_path)
+    except CredentialError as e:
+        print(f"\n{e}", file=sys.stderr)
+        return 1
     except FileNotFoundError as e:
         print(f"Error: {e}", file=sys.stderr)
         return 1
@@ -583,10 +669,14 @@ def cmd_info(args: argparse.Namespace) -> int:
 
 def cmd_validate(args: argparse.Namespace) -> int:
     """Validate an exported agent."""
+    from framework.credentials.models import CredentialError
     from framework.runner import AgentRunner
 
     try:
         runner = AgentRunner.load(args.agent_path)
+    except CredentialError as e:
+        print(f"\n{e}", file=sys.stderr)
+        return 1
     except FileNotFoundError as e:
         print(f"Error: {e}", file=sys.stderr)
         return 1
@@ -903,6 +993,7 @@ def cmd_shell(args: argparse.Namespace) -> int:
     """Start an interactive agent session."""
     import logging
 
+    from framework.credentials.models import CredentialError
     from framework.runner import AgentRunner
 
     # Configure logging to show runtime visibility
@@ -927,6 +1018,9 @@ def cmd_shell(args: argparse.Namespace) -> int:
 
     try:
         runner = AgentRunner.load(agent_path)
+    except CredentialError as e:
+        print(f"\n{e}", file=sys.stderr)
+        return 1
     except FileNotFoundError as e:
         print(f"Error: {e}", file=sys.stderr)
         return 1
@@ -1136,6 +1230,7 @@ def cmd_tui(args: argparse.Namespace) -> int:
     """Browse agents and launch the interactive TUI dashboard."""
     import logging
 
+    from framework.credentials.models import CredentialError
     from framework.runner import AgentRunner
     from framework.tui.app import AdenTUI
 
@@ -1185,8 +1280,10 @@ def cmd_tui(args: argparse.Namespace) -> int:
             runner = AgentRunner.load(
                 agent_path,
                 model=args.model,
-                enable_tui=True,
             )
+        except CredentialError as e:
+            print(f"\n{e}", file=sys.stderr)
+            return
         except Exception as e:
             print(f"Error loading agent: {e}")
             return
@@ -1352,6 +1449,7 @@ def _select_agent(agents_dir: Path) -> str | None:
     for path in agents_dir.iterdir():
         if _is_valid_agent_dir(path):
             agents.append(path)
+    agents.sort(key=lambda p: p.name)
 
     if not agents:
         print(f"No agents found in {agents_dir}", file=sys.stderr)

core/framework/runner/runner.py

@@ -9,6 +9,10 @@ from pathlib import Path
 from typing import TYPE_CHECKING, Any
 
 from framework.config import get_hive_config, get_preferred_model
+from framework.credentials.validation import (
+    ensure_credential_key_env as _ensure_credential_key_env,
+    validate_agent_credentials,
+)
 from framework.graph import Goal
 from framework.graph.edge import (
     DEFAULT_MAX_TOKENS,
@@ -17,17 +21,13 @@ from framework.graph.edge import (
     EdgeSpec,
     GraphSpec,
 )
-from framework.graph.executor import ExecutionResult, GraphExecutor
+from framework.graph.executor import ExecutionResult
 from framework.graph.node import NodeSpec
 from framework.llm.provider import LLMProvider, Tool
 from framework.runner.tool_registry import ToolRegistry
-
-# Multi-entry-point runtime imports
 from framework.runtime.agent_runtime import AgentRuntime, create_agent_runtime
-from framework.runtime.core import Runtime
 from framework.runtime.execution_stream import EntryPointSpec
 from framework.runtime.runtime_log_store import RuntimeLogStore
-from framework.runtime.runtime_logger import RuntimeLogger
 
 if TYPE_CHECKING:
     from framework.runner.protocol import AgentMessage, CapabilityResponse
@@ -35,32 +35,6 @@ if TYPE_CHECKING:
 
 logger = logging.getLogger(__name__)
 
-
-def _ensure_credential_key_env() -> None:
-    """Load HIVE_CREDENTIAL_KEY from shell config if not already in environment.
-
-    The setup-credentials skill writes the encryption key to ~/.zshrc or ~/.bashrc.
-    If the user hasn't sourced their config in the current shell, this reads it
-    directly so the runner (and any MCP subprocesses it spawns) can unlock the
-    encrypted credential store.
-
-    Only HIVE_CREDENTIAL_KEY is loaded this way — all other secrets (API keys, etc.)
-    come from the credential store itself.
-    """
-    if os.environ.get("HIVE_CREDENTIAL_KEY"):
-        return
-
-    try:
-        from aden_tools.credentials.shell_config import check_env_var_in_shell_config
-
-        found, value = check_env_var_in_shell_config("HIVE_CREDENTIAL_KEY")
-        if found and value:
-            os.environ["HIVE_CREDENTIAL_KEY"] = value
-            logger.debug("Loaded HIVE_CREDENTIAL_KEY from shell config")
-    except ImportError:
-        pass
-
-
 CLAUDE_CREDENTIALS_FILE = Path.home() / ".claude" / ".credentials.json"
 
 
@@ -271,7 +245,7 @@ class AgentRunner:
         mock_mode: bool = False,
         storage_path: Path | None = None,
         model: str | None = None,
-        enable_tui: bool = False,
+        intro_message: str = "",
     ):
         """
         Initialize the runner (use AgentRunner.load() instead).
@@ -283,14 +257,14 @@ class AgentRunner:
             mock_mode: If True, use mock LLM responses
             storage_path: Path for runtime storage (defaults to temp)
             model: Model to use (reads from agent config or ~/.hive/configuration.json if None)
-            enable_tui: If True, forces use of AgentRuntime with EventBus
+            intro_message: Optional greeting shown to user on TUI load
         """
         self.agent_path = agent_path
         self.graph = graph
         self.goal = goal
         self.mock_mode = mock_mode
         self.model = model or self._resolve_default_model()
-        self.enable_tui = enable_tui
+        self.intro_message = intro_message
 
         # Set up storage
         if storage_path:
@@ -310,15 +284,17 @@ class AgentRunner:
 
         # Initialize components
         self._tool_registry = ToolRegistry()
-        self._runtime: Runtime | None = None
         self._llm: LLMProvider | None = None
-        self._executor: GraphExecutor | None = None
         self._approval_callback: Callable | None = None
 
-        # Multi-entry-point support (AgentRuntime)
+        # AgentRuntime — unified execution path for all agents
         self._agent_runtime: AgentRuntime | None = None
         self._uses_async_entry_points = self.graph.has_async_entry_points()
 
+        # Validate credentials before spawning MCP servers.
+        # Fails fast with actionable guidance — no MCP noise on screen.
+        self._validate_credentials()
+
         # Auto-discover tools from tools.py
         tools_path = agent_path / "tools.py"
         if tools_path.exists():
@@ -329,6 +305,13 @@ class AgentRunner:
         if mcp_config_path.exists():
             self._load_mcp_servers_from_config(mcp_config_path)
 
+    def _validate_credentials(self) -> None:
+        """Check that required credentials are available before spawning MCP servers.
+
+        Raises CredentialError with actionable guidance if any are missing.
+        """
+        validate_agent_credentials(self.graph.nodes)
+
     @staticmethod
     def _import_agent_module(agent_path: Path):
         """Import an agent package from its directory path.
@@ -372,7 +355,6 @@ class AgentRunner:
         mock_mode: bool = False,
         storage_path: Path | None = None,
         model: str | None = None,
-        enable_tui: bool = False,
     ) -> "AgentRunner":
         """
         Load an agent from an export folder.
@@ -386,7 +368,6 @@ class AgentRunner:
             mock_mode: If True, use mock LLM responses
             storage_path: Path for runtime storage (defaults to ~/.hive/agents/{name})
             model: LLM model to use (reads from agent's default_config if None)
-            enable_tui: If True, forces use of AgentRuntime with EventBus
 
         Returns:
             AgentRunner instance ready to run
@@ -420,6 +401,12 @@ class AgentRunner:
                 hive_config = get_hive_config()
                 max_tokens = hive_config.get("llm", {}).get("max_tokens", DEFAULT_MAX_TOKENS)
 
+            # Read intro_message from agent metadata (shown on TUI load)
+            agent_metadata = getattr(agent_module, "metadata", None)
+            intro_message = ""
+            if agent_metadata and hasattr(agent_metadata, "intro_message"):
+                intro_message = agent_metadata.intro_message
+
             # Build GraphSpec from module-level variables
             graph = GraphSpec(
                 id=f"{agent_path.name}-graph",
@@ -432,6 +419,7 @@ class AgentRunner:
                 nodes=nodes,
                 edges=edges,
                 max_tokens=max_tokens,
+                loop_config=getattr(agent_module, "loop_config", {}),
             )
 
             return cls(
@@ -441,7 +429,7 @@ class AgentRunner:
                 mock_mode=mock_mode,
                 storage_path=storage_path,
                 model=model,
-                enable_tui=enable_tui,
+                intro_message=intro_message,
             )
 
         # Fallback: load from agent.json (legacy JSON-based agents)
@@ -459,7 +447,6 @@ class AgentRunner:
             mock_mode=mock_mode,
             storage_path=storage_path,
             model=model,
-            enable_tui=enable_tui,
         )
 
     def register_tool(
@@ -549,9 +536,6 @@ class AgentRunner:
             callback: Function to call for approval (receives node info, returns bool)
         """
         self._approval_callback = callback
-        # If executor already exists, update it
-        if self._executor is not None:
-            self._executor.approval_callback = callback
 
     def _setup(self) -> None:
         """Set up runtime, LLM, and executor."""
@@ -600,7 +584,10 @@ class AgentRunner:
                 self._llm = LiteLLMProvider(model=self.model, api_key=api_key)
             else:
                 # Fall back to environment variable
-                api_key_env = self._get_api_key_env_var(self.model)
+                # First check api_key_env_var from config (set by quickstart)
+                api_key_env = llm_config.get("api_key_env_var") or self._get_api_key_env_var(
+                    self.model
+                )
                 if api_key_env and os.environ.get(api_key_env):
                     self._llm = LiteLLMProvider(model=self.model)
                 else:
@@ -616,16 +603,11 @@ class AgentRunner:
                         print(f"Warning: {api_key_env} not set. LLM calls will fail.")
                         print(f"Set it with: export {api_key_env}=your-api-key")
 
-        # Get tools for executor/runtime
+        # Get tools for runtime
         tools = list(self._tool_registry.get_tools().values())
         tool_executor = self._tool_registry.get_executor()
 
-        if self._uses_async_entry_points or self.enable_tui:
-            # Multi-entry-point mode or TUI mode: use AgentRuntime
-            self._setup_agent_runtime(tools, tool_executor)
-        else:
-            # Single-entry-point mode: use legacy GraphExecutor
-            self._setup_legacy_executor(tools, tool_executor)
+        self._setup_agent_runtime(tools, tool_executor)
 
     def _get_api_key_env_var(self, model: str) -> str | None:
         """Get the environment variable name for the API key based on model name."""
@@ -640,7 +622,7 @@ class AgentRunner:
         elif model_lower.startswith("anthropic/") or model_lower.startswith("claude"):
             return "ANTHROPIC_API_KEY"
         elif model_lower.startswith("gemini/") or model_lower.startswith("google/"):
-            return "GOOGLE_API_KEY"
+            return "GEMINI_API_KEY"
         elif model_lower.startswith("mistral/"):
             return "MISTRAL_API_KEY"
         elif model_lower.startswith("groq/"):
@@ -686,26 +668,6 @@ class AgentRunner:
         except Exception:
             return None
 
-    def _setup_legacy_executor(self, tools: list, tool_executor: Callable | None) -> None:
-        """Set up legacy single-entry-point execution using GraphExecutor."""
-        # Create runtime
-        self._runtime = Runtime(storage_path=self._storage_path)
-
-        # Create runtime logger
-        log_store = RuntimeLogStore(base_path=self._storage_path / "runtime_logs")
-        runtime_logger = RuntimeLogger(store=log_store, agent_id=self.graph.id)
-
-        # Create executor
-        self._executor = GraphExecutor(
-            runtime=self._runtime,
-            llm=self._llm,
-            tools=tools,
-            tool_executor=tool_executor,
-            approval_callback=self._approval_callback,
-            runtime_logger=runtime_logger,
-            loop_config=self.graph.loop_config,
-        )
-
     def _setup_agent_runtime(self, tools: list, tool_executor: Callable | None) -> None:
         """Set up multi-entry-point execution using AgentRuntime."""
         # Convert AsyncEntryPointSpec to EntryPointSpec for AgentRuntime
@@ -723,9 +685,9 @@ class AgentRunner:
             )
             entry_points.append(ep)
 
-        # If TUI enabled but no entry points (single-entry agent), create default
-        if not entry_points and self.enable_tui and self.graph.entry_node:
-            logger.info("Creating default entry point for TUI")
+        # Single-entry agent with no async entry points: create a default entry point
+        if not entry_points and self.graph.entry_node:
+            logger.info("Creating default entry point for single-entry agent")
             entry_points.append(
                 EntryPointSpec(
                     id="default",
@@ -762,6 +724,9 @@ class AgentRunner:
             checkpoint_config=checkpoint_config,
         )
 
+        # Pass intro_message through for TUI display
+        self._agent_runtime.intro_message = self.intro_message
+
     async def run(
         self,
         input_data: dict | None = None,
@@ -801,32 +766,9 @@ class AgentRunner:
                 error=error_msg,
             )
 
-        if self._uses_async_entry_points or self.enable_tui:
-            # Multi-entry-point mode: use AgentRuntime
-            return await self._run_with_agent_runtime(
-                input_data=input_data or {},
-                entry_point_id=entry_point_id,
-            )
-        else:
-            # Legacy single-entry-point mode
-            return await self._run_with_executor(
-                input_data=input_data or {},
-                session_state=session_state,
-            )
-
-    async def _run_with_executor(
-        self,
-        input_data: dict,
-        session_state: dict | None = None,
-    ) -> ExecutionResult:
-        """Run using legacy GraphExecutor (single entry point)."""
-        if self._executor is None:
-            self._setup()
-
-        return await self._executor.execute(
-            graph=self.graph,
-            goal=self.goal,
-            input_data=input_data,
+        return await self._run_with_agent_runtime(
+            input_data=input_data or {},
+            entry_point_id=entry_point_id,
             session_state=session_state,
         )
 
@@ -834,8 +776,11 @@ class AgentRunner:
         self,
         input_data: dict,
         entry_point_id: str | None = None,
+        session_state: dict | None = None,
     ) -> ExecutionResult:
-        """Run using AgentRuntime (multi-entry-point)."""
+        """Run using AgentRuntime."""
+        import sys
+
         if self._agent_runtime is None:
             self._setup()
 
@@ -843,6 +788,52 @@ class AgentRunner:
         if not self._agent_runtime.is_running:
             await self._agent_runtime.start()
 
+        # Set up stdin-based I/O for client-facing nodes in headless mode.
+        # When a client_facing EventLoopNode calls ask_user(), it emits
+        # CLIENT_INPUT_REQUESTED on the event bus and blocks.  We subscribe
+        # a handler that prints the prompt and reads from stdin, then injects
+        # the user's response back into the node to unblock it.
+        has_client_facing = any(n.client_facing for n in self.graph.nodes)
+        sub_ids: list[str] = []
+
+        if has_client_facing and sys.stdin.isatty():
+            from framework.runtime.event_bus import EventType
+
+            runtime = self._agent_runtime
+
+            async def _handle_client_output(event):
+                """Print agent output to stdout as it streams."""
+                content = event.data.get("content", "")
+                if content:
+                    print(content, end="", flush=True)
+
+            async def _handle_input_requested(event):
+                """Read user input from stdin and inject it into the node."""
+                import asyncio
+
+                node_id = event.node_id
+                try:
+                    loop = asyncio.get_event_loop()
+                    user_input = await loop.run_in_executor(None, input, "\n>>> ")
+                except EOFError:
+                    user_input = ""
+
+                # Inject into the waiting EventLoopNode via runtime
+                await runtime.inject_input(node_id, user_input)
+
+            sub_ids.append(
+                runtime.subscribe_to_events(
+                    event_types=[EventType.CLIENT_OUTPUT_DELTA],
+                    handler=_handle_client_output,
+                )
+            )
+            sub_ids.append(
+                runtime.subscribe_to_events(
+                    event_types=[EventType.CLIENT_INPUT_REQUESTED],
+                    handler=_handle_input_requested,
+                )
+            )
+
         # Determine entry point
         if entry_point_id is None:
             # Use first entry point or "default" if no entry points defined
@@ -852,44 +843,38 @@ class AgentRunner:
             else:
                 entry_point_id = "default"
 
-        # Trigger and wait for result
-        result = await self._agent_runtime.trigger_and_wait(
-            entry_point_id=entry_point_id,
-            input_data=input_data,
-        )
-
-        # Return result or create error result
-        if result is not None:
-            return result
-        else:
-            return ExecutionResult(
-                success=False,
-                error="Execution timed out or failed to complete",
+        try:
+            # Trigger and wait for result
+            result = await self._agent_runtime.trigger_and_wait(
+                entry_point_id=entry_point_id,
+                input_data=input_data,
+                session_state=session_state,
             )
 
-    # === Multi-Entry-Point API (for agents with async_entry_points) ===
+            # Return result or create error result
+            if result is not None:
+                return result
+            else:
+                return ExecutionResult(
+                    success=False,
+                    error="Execution timed out or failed to complete",
+                )
+        finally:
+            # Clean up subscriptions
+            for sub_id in sub_ids:
+                self._agent_runtime.unsubscribe_from_events(sub_id)
+
+    # === Runtime API ===
 
     async def start(self) -> None:
-        """
-        Start the agent runtime (for multi-entry-point agents).
-
-        This starts all registered entry points and allows concurrent execution.
-        For single-entry-point agents, this is a no-op.
-        """
-        if not self._uses_async_entry_points:
-            return
-
+        """Start the agent runtime."""
         if self._agent_runtime is None:
             self._setup()
 
         await self._agent_runtime.start()
 
     async def stop(self) -> None:
-        """
-        Stop the agent runtime (for multi-entry-point agents).
-
-        For single-entry-point agents, this is a no-op.
-        """
+        """Stop the agent runtime."""
         if self._agent_runtime is not None:
             await self._agent_runtime.stop()
 
@@ -902,7 +887,7 @@ class AgentRunner:
         """
         Trigger execution at a specific entry point (non-blocking).
 
-        For multi-entry-point agents only. Returns execution ID for tracking.
+        Returns execution ID for tracking.
 
         Args:
             entry_point_id: Which entry point to trigger
@@ -911,16 +896,7 @@ class AgentRunner:
 
         Returns:
             Execution ID for tracking
-
-        Raises:
-            RuntimeError: If agent doesn't use async entry points
         """
-        if not self._uses_async_entry_points:
-            raise RuntimeError(
-                "trigger() is only available for multi-entry-point agents. "
-                "Use run() for single-entry-point agents."
-            )
-
         if self._agent_runtime is None:
             self._setup()
 
@@ -937,19 +913,9 @@ class AgentRunner:
         """
         Get goal progress across all execution streams.
 
-        For multi-entry-point agents only.
-
         Returns:
             Dict with overall_progress, criteria_status, constraint_violations, etc.
-
-        Raises:
-            RuntimeError: If agent doesn't use async entry points
         """
-        if not self._uses_async_entry_points:
-            raise RuntimeError(
-                "get_goal_progress() is only available for multi-entry-point agents."
-            )
-
         if self._agent_runtime is None:
             self._setup()
 
@@ -957,14 +923,11 @@ class AgentRunner:
 
     def get_entry_points(self) -> list[EntryPointSpec]:
         """
-        Get all registered entry points (for multi-entry-point agents).
+        Get all registered entry points.
 
         Returns:
             List of EntryPointSpec objects
         """
-        if not self._uses_async_entry_points:
-            return []
-
         if self._agent_runtime is None:
             self._setup()
 
@@ -1388,7 +1351,7 @@ Respond with JSON only:
             self._temp_dir = None
 
     async def cleanup_async(self) -> None:
-        """Clean up resources (asynchronous - for multi-entry-point agents)."""
+        """Clean up resources (asynchronous)."""
         # Stop agent runtime if running
         if self._agent_runtime is not None and self._agent_runtime.is_running:
             await self._agent_runtime.stop()
@@ -1399,8 +1362,7 @@ Respond with JSON only:
     async def __aenter__(self) -> "AgentRunner":
         """Context manager entry."""
         self._setup()
-        # Start runtime for multi-entry-point agents
-        if self._uses_async_entry_points and self._agent_runtime is not None:
+        if self._agent_runtime is not None:
             await self._agent_runtime.start()
         return self
 

core/framework/runtime/README.md

@@ -0,0 +1,172 @@
+# Agent Runtime
+
+Unified execution system for all Hive agents. Every agent — single-entry or multi-entry, headless or TUI — runs through the same runtime stack.
+
+## Topology
+
+```
+                     AgentRunner.load(agent_path)
+                              |
+                         AgentRunner
+                     (factory + public API)
+                              |
+                       _setup_agent_runtime()
+                              |
+                        AgentRuntime
+                   (lifecycle + orchestration)
+                      /       |       \
+               Stream A   Stream B   Stream C    ← one per entry point
+                  |           |          |
+            GraphExecutor  GraphExecutor  GraphExecutor
+                  |           |          |
+              Node → Node → Node  (graph traversal)
+```
+
+Single-entry agents get a `"default"` entry point automatically. There is no separate code path.
+
+## Components
+
+| Component | File | Role |
+|---|---|---|
+| `AgentRunner` | `runner/runner.py` | Load agents, configure tools/LLM, expose high-level API |
+| `AgentRuntime` | `runtime/agent_runtime.py` | Lifecycle management, entry point routing, event bus |
+| `ExecutionStream` | `runtime/execution_stream.py` | Per-entry-point execution queue, session persistence |
+| `GraphExecutor` | `graph/executor.py` | Node traversal, tool dispatch, checkpointing |
+| `EventBus` | `runtime/event_bus.py` | Pub/sub for execution events (streaming, I/O) |
+| `SharedStateManager` | `runtime/shared_state.py` | Cross-stream state with isolation levels |
+| `OutcomeAggregator` | `runtime/outcome_aggregator.py` | Goal progress tracking across streams |
+| `SessionStore` | `storage/session_store.py` | Session state persistence (`sessions/{id}/state.json`) |
+
+## Programming Interface
+
+### AgentRunner (high-level)
+
+```python
+from framework.runner import AgentRunner
+
+# Load and run
+runner = AgentRunner.load("exports/my_agent", model="anthropic/claude-sonnet-4-20250514")
+result = await runner.run({"query": "hello"})
+
+# Resume from paused session
+result = await runner.run({"query": "continue"}, session_state=saved_state)
+
+# Lifecycle
+await runner.start()                           # Start the runtime
+await runner.stop()                            # Stop the runtime
+exec_id = await runner.trigger("default", {})  # Non-blocking trigger
+progress = await runner.get_goal_progress()    # Goal evaluation
+entry_points = runner.get_entry_points()       # List entry points
+
+# Context manager
+async with AgentRunner.load("exports/my_agent") as runner:
+    result = await runner.run({"query": "hello"})
+
+# Cleanup
+runner.cleanup()          # Synchronous
+await runner.cleanup_async()  # Asynchronous
+```
+
+### AgentRuntime (lower-level)
+
+```python
+from framework.runtime.agent_runtime import AgentRuntime, create_agent_runtime
+from framework.runtime.execution_stream import EntryPointSpec
+
+# Create runtime with entry points
+runtime = create_agent_runtime(
+    graph=graph,
+    goal=goal,
+    storage_path=Path("~/.hive/agents/my_agent"),
+    entry_points=[
+        EntryPointSpec(id="default", name="Default", entry_node="start", trigger_type="manual"),
+    ],
+    llm=llm,
+    tools=tools,
+    tool_executor=tool_executor,
+    checkpoint_config=checkpoint_config,
+)
+
+# Lifecycle
+await runtime.start()
+await runtime.stop()
+
+# Execution
+exec_id = await runtime.trigger("default", {"query": "hello"})              # Non-blocking
+result = await runtime.trigger_and_wait("default", {"query": "hello"})      # Blocking
+result = await runtime.trigger_and_wait("default", {}, session_state=state) # Resume
+
+# Client-facing node I/O
+await runtime.inject_input(node_id="chat", content="user response")
+
+# Events
+sub_id = runtime.subscribe_to_events(
+    event_types=[EventType.CLIENT_OUTPUT_DELTA],
+    handler=my_handler,
+)
+runtime.unsubscribe_from_events(sub_id)
+
+# Inspection
+runtime.is_running           # bool
+runtime.event_bus            # EventBus
+runtime.state_manager        # SharedStateManager
+runtime.get_stats()          # Runtime statistics
+```
+
+## Execution Flow
+
+1. `AgentRunner.run()` calls `AgentRuntime.trigger_and_wait()`
+2. `AgentRuntime` routes to the `ExecutionStream` for the entry point
+3. `ExecutionStream` creates a `GraphExecutor` and calls `execute()`
+4. `GraphExecutor` traverses nodes, dispatches tools, manages checkpoints
+5. `ExecutionResult` flows back up through the stack
+6. `ExecutionStream` writes session state to disk
+
+## Session Resume
+
+All execution paths support session resume:
+
+```python
+# First run (agent pauses at a client-facing node)
+result = await runner.run({"query": "start task"})
+# result.paused_at = "review-node"
+# result.session_state = {"memory": {...}, "paused_at": "review-node", ...}
+
+# Resume
+result = await runner.run({"input": "approved"}, session_state=result.session_state)
+```
+
+Session state flows: `AgentRunner.run()` → `AgentRuntime.trigger_and_wait()` → `ExecutionStream.execute()` → `GraphExecutor.execute()`.
+
+Checkpoints are saved at node boundaries (`sessions/{id}/checkpoints/`) for crash recovery.
+
+## Event Bus
+
+The `EventBus` provides real-time execution visibility:
+
+| Event | When |
+|---|---|
+| `NODE_STARTED` | Node begins execution |
+| `NODE_COMPLETED` | Node finishes |
+| `TOOL_CALL_STARTED` | Tool invocation begins |
+| `TOOL_CALL_COMPLETED` | Tool invocation finishes |
+| `CLIENT_OUTPUT_DELTA` | Agent streams text to user |
+| `CLIENT_INPUT_REQUESTED` | Agent needs user input |
+| `EXECUTION_COMPLETED` | Full execution finishes |
+
+In headless mode, `AgentRunner` subscribes to `CLIENT_OUTPUT_DELTA` and `CLIENT_INPUT_REQUESTED` to print output and read stdin. In TUI mode, `AdenTUI` subscribes to route events to UI widgets.
+
+## Storage Layout
+
+```
+~/.hive/agents/{agent_name}/
+  sessions/
+    session_YYYYMMDD_HHMMSS_{uuid}/
+      state.json              # Session state (status, memory, progress)
+      checkpoints/            # Node-boundary snapshots
+      logs/
+        summary.json          # Execution summary
+        details.jsonl         # Detailed event log
+        tool_logs.jsonl       # Tool call log
+  runtime_logs/               # Cross-session runtime logs
+```

core/framework/runtime/RESUMABLE_SESSIONS_DESIGN.md

@@ -478,7 +478,7 @@ async def resume_session(
 
 ```bash
 # List resumable sessions
-hive sessions list --agent twitter_outreach --status failed
+hive sessions list --agent deep_research_agent --status failed
 
 # Show checkpoints for a session
 hive sessions checkpoints session_20260208_143022_abc12345

core/framework/runtime/RUNTIME_LOGGING.md

@@ -224,7 +224,7 @@ Three MCP tools provide access to the logging system:
 
 ```python
 query_runtime_logs(
-    agent_work_dir: str,        # e.g., "~/.hive/agents/twitter_outreach"
+    agent_work_dir: str,        # e.g., "~/.hive/agents/deep_research_agent"
     status: str = "",           # "needs_attention", "success", "failure", "degraded"
     limit: int = 20
 ) -> dict  # {"runs": [...], "total": int}
@@ -371,14 +371,14 @@ query_runtime_log_raw(agent_work_dir, run_id)
 ```python
 # 1. Find problematic runs (L1)
 result = query_runtime_logs(
-    agent_work_dir="~/.hive/agents/twitter_outreach",
+    agent_work_dir="~/.hive/agents/deep_research_agent",
     status="needs_attention"
 )
 run_id = result["runs"][0]["run_id"]
 
 # 2. Identify failing nodes (L2)
 details = query_runtime_log_details(
-    agent_work_dir="~/.hive/agents/twitter_outreach",
+    agent_work_dir="~/.hive/agents/deep_research_agent",
     run_id=run_id,
     needs_attention_only=True
 )
@@ -386,7 +386,7 @@ problem_node = details["nodes"][0]["node_id"]
 
 # 3. Analyze root cause (L3)
 raw = query_runtime_log_raw(
-    agent_work_dir="~/.hive/agents/twitter_outreach",
+    agent_work_dir="~/.hive/agents/deep_research_agent",
     run_id=run_id,
     node_id=problem_node
 )
@@ -496,7 +496,7 @@ logger.start_run(goal_id, session_id=execution_id)
 ```json
 {
   "run_id": "session_20260206_115718_e22339c5",
-  "goal_id": "twitter-outreach-multi-loop",
+  "goal_id": "deep-research",
   "status": "degraded",
   "started_at": "2026-02-06T11:57:18.593081",
   "ended_at": "2026-02-06T11:58:45.123456",

core/framework/runtime/agent_runtime.py

@@ -8,7 +8,7 @@ while preserving the goal-driven approach.
 import asyncio
 import logging
 from collections.abc import Callable
-from dataclasses import dataclass
+from dataclasses import dataclass, field
 from pathlib import Path
 from typing import TYPE_CHECKING, Any
 
@@ -39,6 +39,11 @@ class AgentRuntimeConfig:
     max_history: int = 1000
     execution_result_max: int = 1000
     execution_result_ttl_seconds: float | None = None
+    # Webhook server config (only starts if webhook_routes is non-empty)
+    webhook_host: str = "127.0.0.1"
+    webhook_port: int = 8080
+    webhook_routes: list[dict] = field(default_factory=list)
+    # Each dict: {"source_id": str, "path": str, "methods": ["POST"], "secret": str|None}
 
 
 class AgentRuntime:
@@ -150,10 +155,18 @@ class AgentRuntime:
         self._entry_points: dict[str, EntryPointSpec] = {}
         self._streams: dict[str, ExecutionStream] = {}
 
+        # Webhook server (created on start if webhook_routes configured)
+        self._webhook_server: Any = None
+        # Event-driven entry point subscriptions
+        self._event_subscriptions: list[str] = []
+
         # State
         self._running = False
         self._lock = asyncio.Lock()
 
+        # Optional greeting shown to user on TUI load (set by AgentRunner)
+        self.intro_message: str = ""
+
     def register_entry_point(self, spec: EntryPointSpec) -> None:
         """
         Register a named entry point for the agent.
@@ -231,6 +244,63 @@ class AgentRuntime:
                 await stream.start()
                 self._streams[ep_id] = stream
 
+            # Start webhook server if routes are configured
+            if self._config.webhook_routes:
+                from framework.runtime.webhook_server import (
+                    WebhookRoute,
+                    WebhookServer,
+                    WebhookServerConfig,
+                )
+
+                wh_config = WebhookServerConfig(
+                    host=self._config.webhook_host,
+                    port=self._config.webhook_port,
+                )
+                self._webhook_server = WebhookServer(self._event_bus, wh_config)
+
+                for rc in self._config.webhook_routes:
+                    route = WebhookRoute(
+                        source_id=rc["source_id"],
+                        path=rc["path"],
+                        methods=rc.get("methods", ["POST"]),
+                        secret=rc.get("secret"),
+                    )
+                    self._webhook_server.add_route(route)
+
+                await self._webhook_server.start()
+
+            # Subscribe event-driven entry points to EventBus
+            from framework.runtime.event_bus import EventType as _ET
+
+            for ep_id, spec in self._entry_points.items():
+                if spec.trigger_type != "event":
+                    continue
+
+                tc = spec.trigger_config
+                event_types = [_ET(et) for et in tc.get("event_types", [])]
+                if not event_types:
+                    logger.warning(
+                        f"Entry point '{ep_id}' has trigger_type='event' "
+                        "but no event_types in trigger_config"
+                    )
+                    continue
+
+                # Capture ep_id in closure
+                def _make_handler(entry_point_id: str):
+                    async def _on_event(event):
+                        if self._running and entry_point_id in self._streams:
+                            await self.trigger(entry_point_id, {"event": event.to_dict()})
+
+                    return _on_event
+
+                sub_id = self._event_bus.subscribe(
+                    event_types=event_types,
+                    handler=_make_handler(ep_id),
+                    filter_stream=tc.get("filter_stream"),
+                    filter_node=tc.get("filter_node"),
+                )
+                self._event_subscriptions.append(sub_id)
+
             self._running = True
             logger.info(f"AgentRuntime started with {len(self._streams)} streams")
 
@@ -240,6 +310,16 @@ class AgentRuntime:
             return
 
         async with self._lock:
+            # Unsubscribe event-driven entry points
+            for sub_id in self._event_subscriptions:
+                self._event_bus.unsubscribe(sub_id)
+            self._event_subscriptions.clear()
+
+            # Stop webhook server
+            if self._webhook_server:
+                await self._webhook_server.stop()
+                self._webhook_server = None
+
             # Stop all streams
             for stream in self._streams.values():
                 await stream.stop()
@@ -445,6 +525,11 @@ class AgentRuntime:
         """Access the outcome aggregator."""
         return self._outcome_aggregator
 
+    @property
+    def webhook_server(self) -> Any:
+        """Access the webhook server (None if no webhook entry points)."""
+        return self._webhook_server
+
     @property
     def is_running(self) -> bool:
         """Check if runtime is running."""

core/framework/runtime/core.py

@@ -56,6 +56,12 @@ class Runtime:
     """
 
     def __init__(self, storage_path: str | Path):
+        # Validate and create storage path if needed
+        path = Path(storage_path) if isinstance(storage_path, str) else storage_path
+        if not path.exists():
+            logger.warning(f"Storage path does not exist, creating: {path}")
+            path.mkdir(parents=True, exist_ok=True)
+
         self.storage = FileStorage(storage_path)
         self._current_run: Run | None = None
         self._current_node: str = "unknown"

core/framework/runtime/event_bus.py

@@ -63,6 +63,22 @@ class EventType(StrEnum):
     NODE_INPUT_BLOCKED = "node_input_blocked"
     NODE_STALLED = "node_stalled"
 
+    # Judge decisions
+    JUDGE_VERDICT = "judge_verdict"
+
+    # Output tracking
+    OUTPUT_KEY_SET = "output_key_set"
+
+    # Retry / edge tracking
+    NODE_RETRY = "node_retry"
+    EDGE_TRAVERSED = "edge_traversed"
+
+    # Context management
+    CONTEXT_COMPACTED = "context_compacted"
+
+    # External triggers
+    WEBHOOK_RECEIVED = "webhook_received"
+
     # Custom events
     CUSTOM = "custom"
 
@@ -633,6 +649,158 @@ class EventBus:
             )
         )
 
+    # === JUDGE / OUTPUT / RETRY / EDGE PUBLISHERS ===
+
+    async def emit_judge_verdict(
+        self,
+        stream_id: str,
+        node_id: str,
+        action: str,
+        feedback: str = "",
+        judge_type: str = "implicit",
+        iteration: int = 0,
+        execution_id: str | None = None,
+    ) -> None:
+        """Emit judge verdict event."""
+        await self.publish(
+            AgentEvent(
+                type=EventType.JUDGE_VERDICT,
+                stream_id=stream_id,
+                node_id=node_id,
+                execution_id=execution_id,
+                data={
+                    "action": action,
+                    "feedback": feedback,
+                    "judge_type": judge_type,
+                    "iteration": iteration,
+                },
+            )
+        )
+
+    async def emit_output_key_set(
+        self,
+        stream_id: str,
+        node_id: str,
+        key: str,
+        execution_id: str | None = None,
+    ) -> None:
+        """Emit output key set event."""
+        await self.publish(
+            AgentEvent(
+                type=EventType.OUTPUT_KEY_SET,
+                stream_id=stream_id,
+                node_id=node_id,
+                execution_id=execution_id,
+                data={"key": key},
+            )
+        )
+
+    async def emit_node_retry(
+        self,
+        stream_id: str,
+        node_id: str,
+        retry_count: int,
+        max_retries: int,
+        error: str = "",
+        execution_id: str | None = None,
+    ) -> None:
+        """Emit node retry event."""
+        await self.publish(
+            AgentEvent(
+                type=EventType.NODE_RETRY,
+                stream_id=stream_id,
+                node_id=node_id,
+                execution_id=execution_id,
+                data={
+                    "retry_count": retry_count,
+                    "max_retries": max_retries,
+                    "error": error,
+                },
+            )
+        )
+
+    async def emit_edge_traversed(
+        self,
+        stream_id: str,
+        source_node: str,
+        target_node: str,
+        edge_condition: str = "",
+        execution_id: str | None = None,
+    ) -> None:
+        """Emit edge traversed event."""
+        await self.publish(
+            AgentEvent(
+                type=EventType.EDGE_TRAVERSED,
+                stream_id=stream_id,
+                node_id=source_node,
+                execution_id=execution_id,
+                data={
+                    "source_node": source_node,
+                    "target_node": target_node,
+                    "edge_condition": edge_condition,
+                },
+            )
+        )
+
+    async def emit_execution_paused(
+        self,
+        stream_id: str,
+        node_id: str,
+        reason: str = "",
+        execution_id: str | None = None,
+    ) -> None:
+        """Emit execution paused event."""
+        await self.publish(
+            AgentEvent(
+                type=EventType.EXECUTION_PAUSED,
+                stream_id=stream_id,
+                node_id=node_id,
+                execution_id=execution_id,
+                data={"reason": reason},
+            )
+        )
+
+    async def emit_execution_resumed(
+        self,
+        stream_id: str,
+        node_id: str,
+        execution_id: str | None = None,
+    ) -> None:
+        """Emit execution resumed event."""
+        await self.publish(
+            AgentEvent(
+                type=EventType.EXECUTION_RESUMED,
+                stream_id=stream_id,
+                node_id=node_id,
+                execution_id=execution_id,
+                data={},
+            )
+        )
+
+    async def emit_webhook_received(
+        self,
+        source_id: str,
+        path: str,
+        method: str,
+        headers: dict[str, str],
+        payload: dict[str, Any],
+        query_params: dict[str, str] | None = None,
+    ) -> None:
+        """Emit webhook received event."""
+        await self.publish(
+            AgentEvent(
+                type=EventType.WEBHOOK_RECEIVED,
+                stream_id=source_id,
+                data={
+                    "path": path,
+                    "method": method,
+                    "headers": headers,
+                    "payload": payload,
+                    "query_params": query_params or {},
+                },
+            )
+        )
+
     # === QUERY OPERATIONS ===
 
     def get_history(

core/framework/runtime/execution_stream.py

@@ -293,8 +293,13 @@ class ExecutionStream:
         if not self._running:
             raise RuntimeError(f"ExecutionStream '{self.stream_id}' is not running")
 
-        # Generate execution ID using unified session format
-        if self._session_store:
+        # When resuming, reuse the original session ID so the execution
+        # continues in the same session directory instead of creating a new one.
+        resume_session_id = session_state.get("resume_session_id") if session_state else None
+
+        if resume_session_id:
+            execution_id = resume_session_id
+        elif self._session_store:
             execution_id = self._session_store.generate_session_id()
         else:
             # Fallback to old format if SessionStore not available (shouldn't happen)
@@ -597,10 +602,22 @@ class ExecutionStream:
                     entry_point=self.entry_spec.id,
                 )
             else:
-                # Create initial state
-                from framework.schemas.session_state import SessionTimestamps
+                # Create initial state — when resuming, preserve the previous
+                # execution's progress so crashes don't lose track of state.
+                from framework.schemas.session_state import (
+                    SessionProgress,
+                    SessionTimestamps,
+                )
 
                 now = datetime.now().isoformat()
+                ss = ctx.session_state or {}
+                progress = SessionProgress(
+                    current_node=ss.get("paused_at") or ss.get("resume_from"),
+                    paused_at=ss.get("paused_at"),
+                    resume_from=ss.get("paused_at") or ss.get("resume_from"),
+                    path=ss.get("execution_path", []),
+                    node_visit_counts=ss.get("node_visit_counts", {}),
+                )
                 state = SessionState(
                     session_id=execution_id,
                     stream_id=self.stream_id,
@@ -613,6 +630,8 @@ class ExecutionStream:
                         started_at=ctx.started_at.isoformat(),
                         updated_at=now,
                     ),
+                    progress=progress,
+                    memory=ss.get("memory", {}),
                     input_data=ctx.input_data,
                 )
 
@@ -651,6 +670,7 @@ class ExecutionStream:
             max_tokens=self.graph.max_tokens,
             max_steps=self.graph.max_steps,
             cleanup_llm_model=self.graph.cleanup_llm_model,
+            loop_config=self.graph.loop_config,
         )
 
     async def wait_for_completion(

core/framework/runtime/tests/test_webhook_server.py

@@ -0,0 +1,717 @@
+"""
+Tests for WebhookServer and event-driven entry points.
+"""
+
+import asyncio
+import hashlib
+import hmac as hmac_mod
+import json
+import tempfile
+from pathlib import Path
+from unittest.mock import patch
+
+import aiohttp
+import pytest
+
+from framework.runtime.agent_runtime import AgentRuntime, AgentRuntimeConfig
+from framework.runtime.event_bus import AgentEvent, EventBus, EventType
+from framework.runtime.execution_stream import EntryPointSpec
+from framework.runtime.webhook_server import (
+    WebhookRoute,
+    WebhookServer,
+    WebhookServerConfig,
+)
+
+
+def _make_server(event_bus: EventBus, routes: list[WebhookRoute] | None = None):
+    """Helper to create a WebhookServer with port=0 for OS-assigned port."""
+    config = WebhookServerConfig(host="127.0.0.1", port=0)
+    server = WebhookServer(event_bus, config)
+    for route in routes or []:
+        server.add_route(route)
+    return server
+
+
+def _base_url(server: WebhookServer) -> str:
+    """Get the base URL for a running server."""
+    return f"http://127.0.0.1:{server.port}"
+
+
+class TestWebhookServerLifecycle:
+    """Tests for server start/stop."""
+
+    @pytest.mark.asyncio
+    async def test_start_stop(self):
+        bus = EventBus()
+        server = _make_server(
+            bus,
+            [
+                WebhookRoute(source_id="test", path="/webhooks/test", methods=["POST"]),
+            ],
+        )
+
+        await server.start()
+        assert server.is_running
+        assert server.port is not None
+
+        await server.stop()
+        assert not server.is_running
+        assert server.port is None
+
+    @pytest.mark.asyncio
+    async def test_no_routes_skips_start(self):
+        bus = EventBus()
+        server = _make_server(bus)  # no routes
+
+        await server.start()
+        assert not server.is_running
+
+    @pytest.mark.asyncio
+    async def test_stop_when_not_started(self):
+        bus = EventBus()
+        server = _make_server(bus)
+
+        # Should be a no-op, not raise
+        await server.stop()
+        assert not server.is_running
+
+
+class TestWebhookEventPublishing:
+    """Tests for HTTP request -> EventBus event publishing."""
+
+    @pytest.mark.asyncio
+    async def test_post_publishes_webhook_received(self):
+        bus = EventBus()
+        received = []
+
+        async def handler(event):
+            received.append(event)
+
+        bus.subscribe([EventType.WEBHOOK_RECEIVED], handler)
+
+        server = _make_server(
+            bus,
+            [
+                WebhookRoute(source_id="gh", path="/webhooks/github", methods=["POST"]),
+            ],
+        )
+        await server.start()
+
+        try:
+            async with aiohttp.ClientSession() as session:
+                async with session.post(
+                    f"{_base_url(server)}/webhooks/github",
+                    json={"action": "opened", "number": 42},
+                ) as resp:
+                    assert resp.status == 202
+                    body = await resp.json()
+                    assert body["status"] == "accepted"
+
+            # Give event bus time to dispatch
+            await asyncio.sleep(0.05)
+
+            assert len(received) == 1
+            event = received[0]
+            assert event.type == EventType.WEBHOOK_RECEIVED
+            assert event.stream_id == "gh"
+            assert event.data["path"] == "/webhooks/github"
+            assert event.data["method"] == "POST"
+            assert event.data["payload"] == {"action": "opened", "number": 42}
+            assert isinstance(event.data["headers"], dict)
+            assert event.data["query_params"] == {}
+        finally:
+            await server.stop()
+
+    @pytest.mark.asyncio
+    async def test_query_params_included(self):
+        bus = EventBus()
+        received = []
+
+        async def handler(event):
+            received.append(event)
+
+        bus.subscribe([EventType.WEBHOOK_RECEIVED], handler)
+
+        server = _make_server(
+            bus,
+            [
+                WebhookRoute(source_id="hook", path="/webhooks/hook", methods=["POST"]),
+            ],
+        )
+        await server.start()
+
+        try:
+            async with aiohttp.ClientSession() as session:
+                async with session.post(
+                    f"{_base_url(server)}/webhooks/hook?source=test&v=2",
+                    json={"data": "hello"},
+                ) as resp:
+                    assert resp.status == 202
+
+            await asyncio.sleep(0.05)
+
+            assert len(received) == 1
+            assert received[0].data["query_params"] == {"source": "test", "v": "2"}
+        finally:
+            await server.stop()
+
+    @pytest.mark.asyncio
+    async def test_non_json_body(self):
+        bus = EventBus()
+        received = []
+
+        async def handler(event):
+            received.append(event)
+
+        bus.subscribe([EventType.WEBHOOK_RECEIVED], handler)
+
+        server = _make_server(
+            bus,
+            [
+                WebhookRoute(source_id="raw", path="/webhooks/raw", methods=["POST"]),
+            ],
+        )
+        await server.start()
+
+        try:
+            async with aiohttp.ClientSession() as session:
+                async with session.post(
+                    f"{_base_url(server)}/webhooks/raw",
+                    data=b"plain text body",
+                    headers={"Content-Type": "text/plain"},
+                ) as resp:
+                    assert resp.status == 202
+
+            await asyncio.sleep(0.05)
+
+            assert len(received) == 1
+            assert received[0].data["payload"] == {"raw_body": "plain text body"}
+        finally:
+            await server.stop()
+
+    @pytest.mark.asyncio
+    async def test_empty_body(self):
+        bus = EventBus()
+        received = []
+
+        async def handler(event):
+            received.append(event)
+
+        bus.subscribe([EventType.WEBHOOK_RECEIVED], handler)
+
+        server = _make_server(
+            bus,
+            [
+                WebhookRoute(source_id="empty", path="/webhooks/empty", methods=["POST"]),
+            ],
+        )
+        await server.start()
+
+        try:
+            async with aiohttp.ClientSession() as session:
+                async with session.post(f"{_base_url(server)}/webhooks/empty") as resp:
+                    assert resp.status == 202
+
+            await asyncio.sleep(0.05)
+
+            assert len(received) == 1
+            assert received[0].data["payload"] == {}
+        finally:
+            await server.stop()
+
+    @pytest.mark.asyncio
+    async def test_multiple_routes(self):
+        bus = EventBus()
+        received = []
+
+        async def handler(event):
+            received.append(event)
+
+        bus.subscribe([EventType.WEBHOOK_RECEIVED], handler)
+
+        server = _make_server(
+            bus,
+            [
+                WebhookRoute(source_id="a", path="/webhooks/a", methods=["POST"]),
+                WebhookRoute(source_id="b", path="/webhooks/b", methods=["POST"]),
+            ],
+        )
+        await server.start()
+
+        try:
+            async with aiohttp.ClientSession() as session:
+                async with session.post(
+                    f"{_base_url(server)}/webhooks/a", json={"from": "a"}
+                ) as resp:
+                    assert resp.status == 202
+
+                async with session.post(
+                    f"{_base_url(server)}/webhooks/b", json={"from": "b"}
+                ) as resp:
+                    assert resp.status == 202
+
+            await asyncio.sleep(0.05)
+
+            assert len(received) == 2
+            stream_ids = {e.stream_id for e in received}
+            assert stream_ids == {"a", "b"}
+        finally:
+            await server.stop()
+
+    @pytest.mark.asyncio
+    async def test_filter_stream_subscription(self):
+        """Subscribers can filter by stream_id (source_id)."""
+        bus = EventBus()
+        a_events = []
+        b_events = []
+
+        async def handle_a(event):
+            a_events.append(event)
+
+        async def handle_b(event):
+            b_events.append(event)
+
+        bus.subscribe([EventType.WEBHOOK_RECEIVED], handle_a, filter_stream="a")
+        bus.subscribe([EventType.WEBHOOK_RECEIVED], handle_b, filter_stream="b")
+
+        server = _make_server(
+            bus,
+            [
+                WebhookRoute(source_id="a", path="/webhooks/a", methods=["POST"]),
+                WebhookRoute(source_id="b", path="/webhooks/b", methods=["POST"]),
+            ],
+        )
+        await server.start()
+
+        try:
+            async with aiohttp.ClientSession() as session:
+                await session.post(f"{_base_url(server)}/webhooks/a", json={"x": 1})
+                await session.post(f"{_base_url(server)}/webhooks/b", json={"x": 2})
+
+            await asyncio.sleep(0.05)
+
+            assert len(a_events) == 1
+            assert a_events[0].data["payload"] == {"x": 1}
+            assert len(b_events) == 1
+            assert b_events[0].data["payload"] == {"x": 2}
+        finally:
+            await server.stop()
+
+
+class TestHMACVerification:
+    """Tests for HMAC-SHA256 signature verification."""
+
+    @pytest.mark.asyncio
+    async def test_valid_signature_accepted(self):
+        bus = EventBus()
+        received = []
+
+        async def handler(event):
+            received.append(event)
+
+        bus.subscribe([EventType.WEBHOOK_RECEIVED], handler)
+
+        secret = "test-secret-key"
+        server = _make_server(
+            bus,
+            [
+                WebhookRoute(
+                    source_id="secure",
+                    path="/webhooks/secure",
+                    methods=["POST"],
+                    secret=secret,
+                ),
+            ],
+        )
+        await server.start()
+
+        try:
+            body = json.dumps({"event": "push"}).encode()
+            sig = hmac_mod.new(secret.encode(), body, hashlib.sha256).hexdigest()
+
+            async with aiohttp.ClientSession() as session:
+                async with session.post(
+                    f"{_base_url(server)}/webhooks/secure",
+                    data=body,
+                    headers={
+                        "Content-Type": "application/json",
+                        "X-Hub-Signature-256": f"sha256={sig}",
+                    },
+                ) as resp:
+                    assert resp.status == 202
+
+            await asyncio.sleep(0.05)
+            assert len(received) == 1
+        finally:
+            await server.stop()
+
+    @pytest.mark.asyncio
+    async def test_invalid_signature_rejected(self):
+        bus = EventBus()
+        received = []
+
+        async def handler(event):
+            received.append(event)
+
+        bus.subscribe([EventType.WEBHOOK_RECEIVED], handler)
+
+        server = _make_server(
+            bus,
+            [
+                WebhookRoute(
+                    source_id="secure",
+                    path="/webhooks/secure",
+                    methods=["POST"],
+                    secret="real-secret",
+                ),
+            ],
+        )
+        await server.start()
+
+        try:
+            async with aiohttp.ClientSession() as session:
+                async with session.post(
+                    f"{_base_url(server)}/webhooks/secure",
+                    json={"event": "push"},
+                    headers={"X-Hub-Signature-256": "sha256=invalidsignature"},
+                ) as resp:
+                    assert resp.status == 401
+
+            await asyncio.sleep(0.05)
+            assert len(received) == 0  # No event published
+        finally:
+            await server.stop()
+
+    @pytest.mark.asyncio
+    async def test_missing_signature_rejected(self):
+        bus = EventBus()
+        received = []
+
+        async def handler(event):
+            received.append(event)
+
+        bus.subscribe([EventType.WEBHOOK_RECEIVED], handler)
+
+        server = _make_server(
+            bus,
+            [
+                WebhookRoute(
+                    source_id="secure",
+                    path="/webhooks/secure",
+                    methods=["POST"],
+                    secret="my-secret",
+                ),
+            ],
+        )
+        await server.start()
+
+        try:
+            async with aiohttp.ClientSession() as session:
+                # No X-Hub-Signature-256 header
+                async with session.post(
+                    f"{_base_url(server)}/webhooks/secure",
+                    json={"event": "push"},
+                ) as resp:
+                    assert resp.status == 401
+
+            await asyncio.sleep(0.05)
+            assert len(received) == 0
+        finally:
+            await server.stop()
+
+    @pytest.mark.asyncio
+    async def test_no_secret_skips_verification(self):
+        """Routes without a secret accept any request."""
+        bus = EventBus()
+        received = []
+
+        async def handler(event):
+            received.append(event)
+
+        bus.subscribe([EventType.WEBHOOK_RECEIVED], handler)
+
+        server = _make_server(
+            bus,
+            [
+                WebhookRoute(
+                    source_id="open",
+                    path="/webhooks/open",
+                    methods=["POST"],
+                    secret=None,
+                ),
+            ],
+        )
+        await server.start()
+
+        try:
+            async with aiohttp.ClientSession() as session:
+                async with session.post(
+                    f"{_base_url(server)}/webhooks/open",
+                    json={"data": "test"},
+                ) as resp:
+                    assert resp.status == 202
+
+            await asyncio.sleep(0.05)
+            assert len(received) == 1
+        finally:
+            await server.stop()
+
+
+class TestEventDrivenEntryPoints:
+    """Tests for event-driven entry points wired through AgentRuntime."""
+
+    def _make_graph_and_goal(self):
+        """Minimal graph + goal for testing entry point triggering."""
+        from framework.graph import Goal
+        from framework.graph.edge import GraphSpec
+        from framework.graph.goal import SuccessCriterion
+        from framework.graph.node import NodeSpec
+
+        nodes = [
+            NodeSpec(
+                id="process-event",
+                name="Process Event",
+                description="Process incoming event",
+                node_type="llm_generate",
+                input_keys=["event"],
+                output_keys=["result"],
+            ),
+        ]
+        graph = GraphSpec(
+            id="test-graph",
+            goal_id="test-goal",
+            version="1.0.0",
+            entry_node="process-event",
+            entry_points={"start": "process-event"},
+            async_entry_points=[],
+            terminal_nodes=[],
+            pause_nodes=[],
+            nodes=nodes,
+            edges=[],
+        )
+        goal = Goal(
+            id="test-goal",
+            name="Test Goal",
+            description="Test",
+            success_criteria=[
+                SuccessCriterion(
+                    id="sc-1",
+                    description="Done",
+                    metric="done",
+                    target="yes",
+                    weight=1.0,
+                ),
+            ],
+        )
+        return graph, goal
+
+    @pytest.mark.asyncio
+    async def test_event_entry_point_subscribes_to_bus(self):
+        """Entry point with trigger_type='event' subscribes and triggers on matching events."""
+        graph, goal = self._make_graph_and_goal()
+
+        config = AgentRuntimeConfig(
+            webhook_host="127.0.0.1",
+            webhook_port=0,
+            webhook_routes=[
+                {"source_id": "gh", "path": "/webhooks/github"},
+            ],
+        )
+
+        with tempfile.TemporaryDirectory() as tmpdir:
+            runtime = AgentRuntime(
+                graph=graph,
+                goal=goal,
+                storage_path=Path(tmpdir),
+                config=config,
+            )
+
+            runtime.register_entry_point(
+                EntryPointSpec(
+                    id="gh-handler",
+                    name="GitHub Handler",
+                    entry_node="process-event",
+                    trigger_type="event",
+                    trigger_config={
+                        "event_types": ["webhook_received"],
+                        "filter_stream": "gh",
+                    },
+                )
+            )
+
+            trigger_calls = []
+
+            async def mock_trigger(ep_id, data, **kwargs):
+                trigger_calls.append((ep_id, data))
+
+            with patch.object(runtime, "trigger", side_effect=mock_trigger):
+                await runtime.start()
+
+                try:
+                    assert runtime.webhook_server is not None
+                    assert runtime.webhook_server.is_running
+
+                    port = runtime.webhook_server.port
+                    async with aiohttp.ClientSession() as session:
+                        async with session.post(
+                            f"http://127.0.0.1:{port}/webhooks/github",
+                            json={"action": "push", "ref": "main"},
+                        ) as resp:
+                            assert resp.status == 202
+
+                    await asyncio.sleep(0.1)
+
+                    assert len(trigger_calls) == 1
+                    ep_id, data = trigger_calls[0]
+                    assert ep_id == "gh-handler"
+                    assert "event" in data
+                    assert data["event"]["type"] == "webhook_received"
+                    assert data["event"]["stream_id"] == "gh"
+                    assert data["event"]["data"]["payload"] == {
+                        "action": "push",
+                        "ref": "main",
+                    }
+                finally:
+                    await runtime.stop()
+
+            assert runtime.webhook_server is None
+
+    @pytest.mark.asyncio
+    async def test_event_entry_point_filter_stream(self):
+        """Entry point only triggers for matching stream_id (source_id)."""
+        graph, goal = self._make_graph_and_goal()
+
+        config = AgentRuntimeConfig(
+            webhook_routes=[
+                {"source_id": "github", "path": "/webhooks/github"},
+                {"source_id": "stripe", "path": "/webhooks/stripe"},
+            ],
+            webhook_port=0,
+        )
+
+        with tempfile.TemporaryDirectory() as tmpdir:
+            runtime = AgentRuntime(
+                graph=graph,
+                goal=goal,
+                storage_path=Path(tmpdir),
+                config=config,
+            )
+
+            runtime.register_entry_point(
+                EntryPointSpec(
+                    id="gh-only",
+                    name="GitHub Only",
+                    entry_node="process-event",
+                    trigger_type="event",
+                    trigger_config={
+                        "event_types": ["webhook_received"],
+                        "filter_stream": "github",
+                    },
+                )
+            )
+
+            trigger_calls = []
+
+            async def mock_trigger(ep_id, data, **kwargs):
+                trigger_calls.append((ep_id, data))
+
+            with patch.object(runtime, "trigger", side_effect=mock_trigger):
+                await runtime.start()
+
+                try:
+                    port = runtime.webhook_server.port
+                    async with aiohttp.ClientSession() as session:
+                        # POST to stripe — should NOT trigger
+                        await session.post(
+                            f"http://127.0.0.1:{port}/webhooks/stripe",
+                            json={"type": "payment"},
+                        )
+                        # POST to github — should trigger
+                        await session.post(
+                            f"http://127.0.0.1:{port}/webhooks/github",
+                            json={"action": "opened"},
+                        )
+
+                    await asyncio.sleep(0.1)
+
+                    assert len(trigger_calls) == 1
+                    assert trigger_calls[0][0] == "gh-only"
+                finally:
+                    await runtime.stop()
+
+    @pytest.mark.asyncio
+    async def test_no_webhook_routes_skips_server(self):
+        """Runtime without webhook_routes does not start a webhook server."""
+        graph, goal = self._make_graph_and_goal()
+
+        with tempfile.TemporaryDirectory() as tmpdir:
+            runtime = AgentRuntime(
+                graph=graph,
+                goal=goal,
+                storage_path=Path(tmpdir),
+            )
+
+            runtime.register_entry_point(
+                EntryPointSpec(
+                    id="manual",
+                    name="Manual",
+                    entry_node="process-event",
+                    trigger_type="manual",
+                )
+            )
+
+            await runtime.start()
+            try:
+                assert runtime.webhook_server is None
+            finally:
+                await runtime.stop()
+
+    @pytest.mark.asyncio
+    async def test_event_entry_point_custom_event(self):
+        """Entry point can subscribe to CUSTOM events, not just webhooks."""
+        graph, goal = self._make_graph_and_goal()
+
+        with tempfile.TemporaryDirectory() as tmpdir:
+            runtime = AgentRuntime(
+                graph=graph,
+                goal=goal,
+                storage_path=Path(tmpdir),
+            )
+
+            runtime.register_entry_point(
+                EntryPointSpec(
+                    id="custom-handler",
+                    name="Custom Handler",
+                    entry_node="process-event",
+                    trigger_type="event",
+                    trigger_config={
+                        "event_types": ["custom"],
+                    },
+                )
+            )
+
+            trigger_calls = []
+
+            async def mock_trigger(ep_id, data, **kwargs):
+                trigger_calls.append((ep_id, data))
+
+            with patch.object(runtime, "trigger", side_effect=mock_trigger):
+                await runtime.start()
+
+                try:
+                    await runtime.event_bus.publish(
+                        AgentEvent(
+                            type=EventType.CUSTOM,
+                            stream_id="some-source",
+                            data={"key": "value"},
+                        )
+                    )
+
+                    await asyncio.sleep(0.1)
+
+                    assert len(trigger_calls) == 1
+                    assert trigger_calls[0][0] == "custom-handler"
+                    assert trigger_calls[0][1]["event"]["type"] == "custom"
+                    assert trigger_calls[0][1]["event"]["data"]["key"] == "value"
+                finally:
+                    await runtime.stop()

core/framework/runtime/webhook_server.py

@@ -0,0 +1,175 @@
+"""
+Webhook HTTP Server - Receives HTTP requests and publishes them as EventBus events.
+
+Only starts if webhook-type entry points are registered. Uses aiohttp for
+a lightweight embedded HTTP server that runs within the existing asyncio loop.
+"""
+
+import hashlib
+import hmac
+import json
+import logging
+from dataclasses import dataclass
+
+from aiohttp import web
+
+from framework.runtime.event_bus import EventBus
+
+logger = logging.getLogger(__name__)
+
+
+@dataclass
+class WebhookRoute:
+    """A registered webhook route derived from an EntryPointSpec."""
+
+    source_id: str
+    path: str
+    methods: list[str]
+    secret: str | None = None  # For HMAC-SHA256 signature verification
+
+
+@dataclass
+class WebhookServerConfig:
+    """Configuration for the webhook HTTP server."""
+
+    host: str = "127.0.0.1"
+    port: int = 8080
+
+
+class WebhookServer:
+    """
+    Embedded HTTP server that receives webhook requests and publishes
+    them as WEBHOOK_RECEIVED events on the EventBus.
+
+    The server's only job is: receive HTTP -> publish AgentEvent.
+    Subscribers decide what to do with the event.
+
+    Lifecycle:
+        server = WebhookServer(event_bus, config)
+        server.add_route(WebhookRoute(...))
+        await server.start()
+        # ... server running ...
+        await server.stop()
+    """
+
+    def __init__(
+        self,
+        event_bus: EventBus,
+        config: WebhookServerConfig | None = None,
+    ):
+        self._event_bus = event_bus
+        self._config = config or WebhookServerConfig()
+        self._routes: dict[str, WebhookRoute] = {}  # path -> route
+        self._app: web.Application | None = None
+        self._runner: web.AppRunner | None = None
+        self._site: web.TCPSite | None = None
+
+    def add_route(self, route: WebhookRoute) -> None:
+        """Register a webhook route."""
+        self._routes[route.path] = route
+
+    async def start(self) -> None:
+        """Start the HTTP server. No-op if no routes registered."""
+        if not self._routes:
+            logger.debug("No webhook routes registered, skipping server start")
+            return
+
+        self._app = web.Application()
+
+        for path, route in self._routes.items():
+            for method in route.methods:
+                self._app.router.add_route(method, path, self._handle_request)
+
+        self._runner = web.AppRunner(self._app)
+        await self._runner.setup()
+        self._site = web.TCPSite(
+            self._runner,
+            self._config.host,
+            self._config.port,
+        )
+        await self._site.start()
+        logger.info(
+            f"Webhook server started on {self._config.host}:{self._config.port} "
+            f"with {len(self._routes)} route(s)"
+        )
+
+    async def stop(self) -> None:
+        """Stop the HTTP server gracefully."""
+        if self._runner:
+            await self._runner.cleanup()
+            self._runner = None
+            self._app = None
+            self._site = None
+            logger.info("Webhook server stopped")
+
+    async def _handle_request(self, request: web.Request) -> web.Response:
+        """Handle an incoming webhook request."""
+        path = request.path
+        route = self._routes.get(path)
+
+        if route is None:
+            return web.json_response({"error": "Not found"}, status=404)
+
+        # Read body
+        try:
+            body = await request.read()
+        except Exception:
+            return web.json_response(
+                {"error": "Failed to read request body"},
+                status=400,
+            )
+
+        # Verify HMAC signature if secret is configured
+        if route.secret:
+            if not self._verify_signature(request, body, route.secret):
+                return web.json_response({"error": "Invalid signature"}, status=401)
+
+        # Parse body as JSON (fall back to raw text for non-JSON)
+        try:
+            payload = json.loads(body) if body else {}
+        except (json.JSONDecodeError, ValueError):
+            payload = {"raw_body": body.decode("utf-8", errors="replace")}
+
+        # Publish event to bus
+        await self._event_bus.emit_webhook_received(
+            source_id=route.source_id,
+            path=path,
+            method=request.method,
+            headers=dict(request.headers),
+            payload=payload,
+            query_params=dict(request.query),
+        )
+
+        return web.json_response({"status": "accepted"}, status=202)
+
+    def _verify_signature(
+        self,
+        request: web.Request,
+        body: bytes,
+        secret: str,
+    ) -> bool:
+        """Verify HMAC-SHA256 signature from X-Hub-Signature-256 header."""
+        signature_header = request.headers.get("X-Hub-Signature-256", "")
+        if not signature_header.startswith("sha256="):
+            return False
+
+        expected_sig = signature_header[7:]  # strip "sha256="
+        computed_sig = hmac.new(
+            secret.encode("utf-8"),
+            body,
+            hashlib.sha256,
+        ).hexdigest()
+
+        return hmac.compare_digest(expected_sig, computed_sig)
+
+    @property
+    def is_running(self) -> bool:
+        """Check if the server is running."""
+        return self._site is not None
+
+    @property
+    def port(self) -> int | None:
+        """Return the actual listening port (useful when configured with port=0)."""
+        if self._site and self._site._server and self._site._server.sockets:
+            return self._site._server.sockets[0].getsockname()[1]
+        return None

core/framework/schemas/session_state.py

@@ -156,8 +156,13 @@ class SessionState(BaseModel):
     @computed_field
     @property
     def is_resumable(self) -> bool:
-        """Can this session be resumed?"""
-        return self.status == SessionStatus.PAUSED and self.progress.resume_from is not None
+        """Can this session be resumed?
+
+        Every non-completed session is resumable. If resume_from/paused_at
+        aren't set, the executor falls back to the graph entry point —
+        so we don't gate on those. Even catastrophic failures are resumable.
+        """
+        return self.status != SessionStatus.COMPLETED
 
     @computed_field
     @property
@@ -279,9 +284,16 @@ class SessionState(BaseModel):
 
     def to_session_state_dict(self) -> dict[str, Any]:
         """Convert to session_state format for GraphExecutor.execute()."""
+        # Derive resume target: explicit > last node in path > entry point
+        resume_from = (
+            self.progress.resume_from
+            or self.progress.paused_at
+            or (self.progress.path[-1] if self.progress.path else None)
+        )
         return {
-            "paused_at": self.progress.paused_at,
-            "resume_from": self.progress.resume_from,
+            "paused_at": resume_from,
+            "resume_from": resume_from,
             "memory": self.memory,
-            "next_node": None,
+            "execution_path": self.progress.path,
+            "node_visit_counts": self.progress.node_visit_counts,
         }

core/framework/storage/session_store.py

@@ -37,7 +37,7 @@ class SessionStore:
         Initialize session store.
 
         Args:
-            base_path: Base path for storage (e.g., ~/.hive/agents/twitter_outreach)
+            base_path: Base path for storage (e.g., ~/.hive/agents/deep_research_agent)
         """
         self.base_path = Path(base_path)
         self.sessions_dir = self.base_path / "sessions"

core/framework/testing/prompts.py

@@ -3,6 +3,10 @@ Pytest templates for test file generation.
 
 These templates provide headers and fixtures for pytest-compatible async tests.
 Tests are written to exports/{agent}/tests/ as Python files and run with pytest.
+
+Tests use AgentRunner.load() — the canonical runtime path — which creates
+AgentRuntime, ExecutionStream, and proper session/log storage. For agents
+with client-facing nodes, an auto_responder fixture handles input injection.
 """
 
 # Template for the test file header (imports and fixtures)
@@ -11,17 +15,19 @@ PYTEST_TEST_FILE_HEADER = '''"""
 
 {description}
 
-REQUIRES: API_KEY (OpenAI or Anthropic) for real testing.
+REQUIRES: API_KEY for execution tests. Structure tests run without keys.
 """
 
 import os
 import pytest
-from {agent_module} import default_agent
+from pathlib import Path
+
+# Agent path resolved from this test file's location
+AGENT_PATH = Path(__file__).resolve().parents[1]
 
 
 def _get_api_key():
     """Get API key from CredentialStoreAdapter or environment."""
-    # 1. Try CredentialStoreAdapter for Anthropic
     try:
         from aden_tools.credentials import CredentialStoreAdapter
         creds = CredentialStoreAdapter.default()
@@ -29,28 +35,43 @@ def _get_api_key():
             return creds.get("anthropic")
     except (ImportError, KeyError):
         pass
-
-    # 2. Fallback to standard environment variables for OpenAI and others
     return (
         os.environ.get("OPENAI_API_KEY") or
         os.environ.get("ANTHROPIC_API_KEY") or
         os.environ.get("CEREBRAS_API_KEY") or
-        os.environ.get("GROQ_API_KEY")
+        os.environ.get("GROQ_API_KEY") or
+        os.environ.get("GEMINI_API_KEY")
     )
 
 
 # Skip all tests if no API key and not in mock mode
 pytestmark = pytest.mark.skipif(
     not _get_api_key() and not os.environ.get("MOCK_MODE"),
-    reason="API key required. Please set OPENAI_API_KEY, ANTHROPIC_API_KEY, or use MOCK_MODE=1."
+    reason="API key required. Set ANTHROPIC_API_KEY or use MOCK_MODE=1 for structure tests."
 )
 '''
 
 # Template for conftest.py with shared fixtures
 PYTEST_CONFTEST_TEMPLATE = '''"""Shared test fixtures for {agent_name} tests."""
 
+import json
 import os
+import re
+import sys
+from pathlib import Path
+
+# Add exports/ and core/ to sys.path so the agent package and framework are importable
+_repo_root = Path(__file__).resolve().parents[3]
+for _p in ["exports", "core"]:
+    _path = str(_repo_root / _p)
+    if _path not in sys.path:
+        sys.path.insert(0, _path)
+
 import pytest
+from framework.runner.runner import AgentRunner
+from framework.runtime.event_bus import EventType
+
+AGENT_PATH = Path(__file__).resolve().parents[1]
 
 
 def _get_api_key():
@@ -62,19 +83,80 @@ def _get_api_key():
             return creds.get("anthropic")
     except (ImportError, KeyError):
         pass
-
     return (
         os.environ.get("OPENAI_API_KEY") or
         os.environ.get("ANTHROPIC_API_KEY") or
         os.environ.get("CEREBRAS_API_KEY") or
-        os.environ.get("GROQ_API_KEY")
+        os.environ.get("GROQ_API_KEY") or
+        os.environ.get("GEMINI_API_KEY")
     )
 
 
-@pytest.fixture
+@pytest.fixture(scope="session")
 def mock_mode():
-    """Check if running in mock mode."""
-    return bool(os.environ.get("MOCK_MODE"))
+    """Return True if running in mock mode (no API key or MOCK_MODE=1)."""
+    if os.environ.get("MOCK_MODE"):
+        return True
+    return not bool(_get_api_key())
+
+
+@pytest.fixture(scope="session")
+async def runner(tmp_path_factory, mock_mode):
+    """Create an AgentRunner using the canonical runtime path.
+
+    Uses tmp_path_factory for storage so tests don't pollute ~/.hive/agents/.
+    Goes through AgentRunner.load() -> _setup() -> AgentRuntime, the same
+    path as ``hive run``.
+    """
+    storage = tmp_path_factory.mktemp("agent_storage")
+    r = AgentRunner.load(
+        AGENT_PATH,
+        mock_mode=mock_mode,
+        storage_path=storage,
+    )
+    r._setup()
+    yield r
+    await r.cleanup_async()
+
+
+@pytest.fixture
+def auto_responder(runner):
+    """Auto-respond to client-facing node input requests.
+
+    Subscribes to CLIENT_INPUT_REQUESTED events and injects a response
+    to unblock the node. Customize the response before calling start():
+
+        auto_responder.response = "approve the report"
+        await auto_responder.start()
+    """
+    class AutoResponder:
+        def __init__(self, runner_instance):
+            self._runner = runner_instance
+            self.response = "yes, proceed"
+            self.interactions = []
+            self._sub_id = None
+
+        async def start(self):
+            runtime = self._runner._agent_runtime
+            if runtime is None:
+                return
+
+            async def _handle(event):
+                self.interactions.append(event.node_id)
+                await runtime.inject_input(event.node_id, self.response)
+
+            self._sub_id = runtime.subscribe_to_events(
+                event_types=[EventType.CLIENT_INPUT_REQUESTED],
+                handler=_handle,
+            )
+
+        async def stop(self):
+            runtime = self._runner._agent_runtime
+            if self._sub_id and runtime:
+                runtime.unsubscribe_from_events(self._sub_id)
+                self._sub_id = None
+
+    return AutoResponder(runner)
 
 
 @pytest.fixture(scope="session", autouse=True)
@@ -82,19 +164,51 @@ def check_api_key():
     """Ensure API key is set for real testing."""
     if not _get_api_key():
         if os.environ.get("MOCK_MODE"):
-            print("\\n⚠️  Running in MOCK MODE - structure validation only")
-            print("   This does NOT test LLM behavior or agent quality")
-            print("   Set OPENAI_API_KEY or ANTHROPIC_API_KEY for real testing\\n")
+            print("\\n  Running in MOCK MODE - structure validation only")
+            print("  Set ANTHROPIC_API_KEY for real testing\\n")
         else:
             pytest.fail(
-                "\\n❌ No API key found!\\n\\n"
-                "Real testing requires an API key. Choose one:\\n"
-                "1. Set OpenAI key:\\n"
-                "   export OPENAI_API_KEY='your-key-here'\\n"
-                "2. Set Anthropic key:\\n"
-                "   export ANTHROPIC_API_KEY='your-key-here'\\n"
-                "3. Run structure validation only:\\n"
-                "   MOCK_MODE=1 pytest exports/{agent_name}/tests/\\n\\n"
-                "Note: Mock mode does NOT validate agent behavior or quality."
+                "\\nNo API key found!\\n"
+                "Set ANTHROPIC_API_KEY or use MOCK_MODE=1 for structure tests.\\n"
             )
+
+
+def parse_json_from_output(result, key):
+    """Parse JSON from agent output (framework may store full LLM response as string)."""
+    val = result.output.get(key, "")
+    if isinstance(val, (dict, list)):
+        return val
+    if isinstance(val, str):
+        json_text = re.sub(r"```json\\s*|\\s*```", "", val).strip()
+        try:
+            return json.loads(json_text)
+        except (json.JSONDecodeError, TypeError):
+            return val
+    return val
+
+
+def safe_get_nested(result, key_path, default=None):
+    """Safely get nested value from result.output."""
+    output = result.output or {{}}
+    current = output
+    for key in key_path:
+        if isinstance(current, dict):
+            current = current.get(key)
+        elif isinstance(current, str):
+            try:
+                json_text = re.sub(r"```json\\s*|\\s*```", "", current).strip()
+                parsed = json.loads(json_text)
+                if isinstance(parsed, dict):
+                    current = parsed.get(key)
+                else:
+                    return default
+            except json.JSONDecodeError:
+                return default
+        else:
+            return default
+    return current if current is not None else default
+
+
+pytest.parse_json_from_output = parse_json_from_output
+pytest.safe_get_nested = safe_get_nested
 '''

core/framework/tui/app.py

@@ -5,14 +5,13 @@ import time
 
 from textual.app import App, ComposeResult
 from textual.binding import Binding
-from textual.containers import Container, Horizontal, Vertical
-from textual.widgets import Footer, Input, Label
+from textual.containers import Container, Horizontal
+from textual.widgets import Footer, Label
 
 from framework.runtime.agent_runtime import AgentRuntime
 from framework.runtime.event_bus import AgentEvent, EventType
 from framework.tui.widgets.chat_repl import ChatRepl
 from framework.tui.widgets.graph_view import GraphOverview
-from framework.tui.widgets.log_pane import LogPane
 from framework.tui.widgets.selectable_rich_log import SelectableRichLog
 
 
@@ -136,28 +135,15 @@ class AdenTUI(App):
         background: $surface;
     }
 
-    #left-pane {
-        width: 60%;
-        height: 100%;
-        layout: vertical;
-        background: $surface;
-    }
-
     GraphOverview {
-        height: 40%;
+        width: 40%;
+        height: 100%;
         background: $panel;
         padding: 0;
     }
 
-    LogPane {
-        height: 60%;
-        background: $surface;
-        padding: 0;
-        margin-bottom: 1;
-    }
-
     ChatRepl {
-        width: 40%;
+        width: 60%;
         height: 100%;
         background: $panel;
         border-left: tall $primary;
@@ -180,13 +166,13 @@ class AdenTUI(App):
         scrollbar-color: $primary;
     }
 
-    Input {
+    ChatTextArea {
         background: $surface;
         border: tall $primary;
         margin-top: 1;
     }
 
-    Input:focus {
+    ChatTextArea:focus {
         border: tall $accent;
     }
 
@@ -208,6 +194,7 @@ class AdenTUI(App):
         Binding("ctrl+c", "ctrl_c", "Interrupt", show=False, priority=True),
         Binding("super+c", "ctrl_c", "Copy", show=False, priority=True),
         Binding("ctrl+s", "screenshot", "Screenshot (SVG)", show=True, priority=True),
+        Binding("ctrl+l", "toggle_logs", "Toggle Logs", show=True, priority=True),
         Binding("ctrl+z", "pause_execution", "Pause", show=True, priority=True),
         Binding("ctrl+r", "show_sessions", "Sessions", show=True, priority=True),
         Binding("tab", "focus_next", "Next Panel", show=True),
@@ -223,7 +210,6 @@ class AdenTUI(App):
         super().__init__()
 
         self.runtime = runtime
-        self.log_pane = LogPane()
         self.graph_view = GraphOverview(runtime)
         self.chat_repl = ChatRepl(runtime, resume_session, resume_checkpoint)
         self.status_bar = StatusBar(graph_id=runtime.graph.id)
@@ -253,11 +239,7 @@ class AdenTUI(App):
         yield self.status_bar
 
         yield Horizontal(
-            Vertical(
-                self.log_pane,
-                self.graph_view,
-                id="left-pane",
-            ),
+            self.graph_view,
             self.chat_repl,
         )
 
@@ -328,7 +310,7 @@ class AdenTUI(App):
                 if record.name.startswith(("textual", "LiteLLM", "litellm")):
                     continue
 
-                self.log_pane.write_python_log(record)
+                self.chat_repl.write_python_log(record)
         except Exception:
             pass
 
@@ -350,6 +332,14 @@ class AdenTUI(App):
         EventType.CONSTRAINT_VIOLATION,
         EventType.STATE_CHANGED,
         EventType.NODE_INPUT_BLOCKED,
+        EventType.CONTEXT_COMPACTED,
+        EventType.NODE_INTERNAL_OUTPUT,
+        EventType.JUDGE_VERDICT,
+        EventType.OUTPUT_KEY_SET,
+        EventType.NODE_RETRY,
+        EventType.EDGE_TRAVERSED,
+        EventType.EXECUTION_PAUSED,
+        EventType.EXECUTION_RESUMED,
     ]
 
     _LOG_PANE_EVENTS = frozenset(_EVENT_TYPES) - {
@@ -371,12 +361,22 @@ class AdenTUI(App):
         """Called from the agent thread — bridge to Textual's main thread."""
         try:
             self.call_from_thread(self._route_event, event)
-        except Exception:
-            pass
+        except Exception as e:
+            logging.getLogger("tui.events").error(
+                "call_from_thread failed for %s (node=%s): %s",
+                event.type.value,
+                event.node_id or "?",
+                e,
+            )
 
     def _route_event(self, event: AgentEvent) -> None:
         """Route incoming events to widgets. Runs on Textual's main thread."""
         if not self.is_ready:
+            logging.getLogger("tui.events").warning(
+                "Event dropped (not ready): %s node=%s",
+                event.type.value,
+                event.node_id or "?",
+            )
             return
 
         try:
@@ -407,6 +407,35 @@ class AdenTUI(App):
                 self.chat_repl.handle_input_requested(
                     event.node_id or event.data.get("node_id", ""),
                 )
+            elif et == EventType.NODE_LOOP_STARTED:
+                self.chat_repl.handle_node_started(event.node_id or "")
+            elif et == EventType.NODE_LOOP_ITERATION:
+                self.chat_repl.handle_loop_iteration(event.data.get("iteration", 0))
+
+            # Track active node in chat_repl for mid-execution input
+            if et == EventType.NODE_LOOP_STARTED:
+                self.chat_repl.handle_node_started(event.node_id or "")
+            elif et == EventType.NODE_LOOP_COMPLETED:
+                self.chat_repl.handle_node_completed(event.node_id or "")
+
+            # Non-client-facing node output → chat repl
+            if et == EventType.NODE_INTERNAL_OUTPUT:
+                content = event.data.get("content", "")
+                if content.strip():
+                    self.chat_repl.handle_internal_output(event.node_id or "", content)
+
+            # Execution paused/resumed → chat repl
+            if et == EventType.EXECUTION_PAUSED:
+                reason = event.data.get("reason", "")
+                self.chat_repl.handle_execution_paused(event.node_id or "", reason)
+            elif et == EventType.EXECUTION_RESUMED:
+                self.chat_repl.handle_execution_resumed(event.node_id or "")
+
+            # Goal achieved / constraint violation → chat repl
+            if et == EventType.GOAL_ACHIEVED:
+                self.chat_repl.handle_goal_achieved(event.data)
+            elif et == EventType.CONSTRAINT_VIOLATION:
+                self.chat_repl.handle_constraint_violation(event.data)
 
             # --- Graph view events ---
             if et in (
@@ -444,6 +473,13 @@ class AdenTUI(App):
                     started=False,
                 )
 
+            # Edge traversal → graph view
+            if et == EventType.EDGE_TRAVERSED:
+                self.graph_view.handle_edge_traversed(
+                    event.data.get("source_node", ""),
+                    event.data.get("target_node", ""),
+                )
+
             # --- Status bar events ---
             if et == EventType.EXECUTION_STARTED:
                 entry_node = event.data.get("entry_node") or (
@@ -464,12 +500,36 @@ class AdenTUI(App):
                 self.status_bar.set_node_detail("thinking...")
             elif et == EventType.NODE_STALLED:
                 self.status_bar.set_node_detail(f"stalled: {event.data.get('reason', '')}")
+            elif et == EventType.CONTEXT_COMPACTED:
+                before = event.data.get("usage_before", "?")
+                after = event.data.get("usage_after", "?")
+                self.status_bar.set_node_detail(f"compacted: {before}% \u2192 {after}%")
+            elif et == EventType.JUDGE_VERDICT:
+                action = event.data.get("action", "?")
+                self.status_bar.set_node_detail(f"judge: {action}")
+            elif et == EventType.OUTPUT_KEY_SET:
+                key = event.data.get("key", "?")
+                self.status_bar.set_node_detail(f"set: {key}")
+            elif et == EventType.NODE_RETRY:
+                retry = event.data.get("retry_count", "?")
+                max_r = event.data.get("max_retries", "?")
+                self.status_bar.set_node_detail(f"retry {retry}/{max_r}")
+            elif et == EventType.EXECUTION_PAUSED:
+                self.status_bar.set_node_detail("paused")
+            elif et == EventType.EXECUTION_RESUMED:
+                self.status_bar.set_node_detail("resumed")
 
-            # --- Log pane events ---
+            # --- Log events (inline in chat) ---
             if et in self._LOG_PANE_EVENTS:
-                self.log_pane.write_event(event)
-        except Exception:
-            pass
+                self.chat_repl.write_log_event(event)
+        except Exception as e:
+            logging.getLogger("tui.events").error(
+                "Route failed for %s (node=%s): %s",
+                event.type.value,
+                event.node_id or "?",
+                e,
+                exc_info=True,
+            )
 
     def save_screenshot(self, filename: str | None = None) -> str:
         """Save a screenshot of the current screen as SVG (viewable in browsers).
@@ -504,8 +564,8 @@ class AdenTUI(App):
         original_chat_border = chat_widget.styles.border_left
         chat_widget.styles.border_left = ("none", "transparent")
 
-        # Hide all Input widget borders
-        input_widgets = self.query("Input")
+        # Hide all TextArea widget borders
+        input_widgets = self.query("ChatTextArea")
         original_input_borders = []
         for input_widget in input_widgets:
             original_input_borders.append(input_widget.styles.border)
@@ -535,6 +595,12 @@ class AdenTUI(App):
         except Exception as e:
             self.notify(f"Screenshot failed: {e}", severity="error", timeout=5)
 
+    def action_toggle_logs(self) -> None:
+        """Toggle inline log display in chat (bound to Ctrl+L)."""
+        self.chat_repl.toggle_logs()
+        mode = "ON" if self.chat_repl._show_logs else "OFF"
+        self.notify(f"Logs {mode}", severity="information", timeout=2)
+
     def action_pause_execution(self) -> None:
         """Immediately pause execution by cancelling task (bound to Ctrl+Z)."""
         try:
@@ -575,19 +641,12 @@ class AdenTUI(App):
                 timeout=5,
             )
 
-    def action_show_sessions(self) -> None:
+    async def action_show_sessions(self) -> None:
         """Show sessions list (bound to Ctrl+R)."""
         # Send /sessions command to chat input
         try:
             chat_repl = self.query_one(ChatRepl)
-            chat_input = chat_repl.query_one("#chat-input", Input)
-            chat_input.value = "/sessions"
-            # Trigger submission
-            self.notify(
-                "💡 Type /sessions in the chat to see all sessions",
-                severity="information",
-                timeout=3,
-            )
+            await chat_repl._submit_input("/sessions")
         except Exception:
             self.notify(
                 "Use /sessions command to see all sessions",

core/framework/tui/widgets/chat_repl.py

@@ -15,6 +15,7 @@ Client-facing input:
 """
 
 import asyncio
+import logging
 import re
 import threading
 from pathlib import Path
@@ -22,12 +23,40 @@ from typing import Any
 
 from textual.app import ComposeResult
 from textual.containers import Vertical
-from textual.widgets import Input, Label
+from textual.message import Message
+from textual.widgets import Label, TextArea
 
 from framework.runtime.agent_runtime import AgentRuntime
+from framework.runtime.event_bus import AgentEvent
+from framework.tui.widgets.log_pane import format_event, format_python_log
 from framework.tui.widgets.selectable_rich_log import SelectableRichLog as RichLog
 
 
+class ChatTextArea(TextArea):
+    """TextArea that submits on Enter and inserts newlines on Shift+Enter."""
+
+    class Submitted(Message):
+        """Posted when the user presses Enter."""
+
+        def __init__(self, text: str) -> None:
+            super().__init__()
+            self.text = text
+
+    async def _on_key(self, event) -> None:
+        if event.key == "enter":
+            text = self.text.strip()
+            self.clear()
+            if text:
+                self.post_message(self.Submitted(text))
+            event.stop()
+            event.prevent_default()
+        elif event.key == "shift+enter":
+            event.key = "enter"
+            await super()._on_key(event)
+        else:
+            await super()._on_key(event)
+
+
 class ChatRepl(Vertical):
     """Widget for interactive chat/REPL."""
 
@@ -56,16 +85,17 @@ class ChatRepl(Vertical):
         display: none;
     }
 
-    ChatRepl > Input {
+    ChatRepl > ChatTextArea {
         width: 100%;
         height: auto;
+        max-height: 7;
         dock: bottom;
         background: $surface;
         border: tall $primary;
         margin-top: 1;
     }
 
-    ChatRepl > Input:focus {
+    ChatRepl > ChatTextArea:focus {
         border: tall $accent;
     }
     """
@@ -82,8 +112,13 @@ class ChatRepl(Vertical):
         self._streaming_snapshot: str = ""
         self._waiting_for_input: bool = False
         self._input_node_id: str | None = None
+        self._pending_ask_question: str = ""
+        self._active_node_id: str | None = None  # Currently executing node
         self._resume_session = resume_session
         self._resume_checkpoint = resume_checkpoint
+        self._session_index: list[str] = []  # IDs from last listing
+        self._show_logs: bool = False  # Clean mode by default
+        self._log_buffer: list[str] = []  # Buffered log lines for backfill on toggle ON
 
         # Dedicated event loop for agent execution.
         # Keeps blocking runtime code (LLM calls, MCP tools) off
@@ -106,7 +141,7 @@ class ChatRepl(Vertical):
             min_width=0,
         )
         yield Label("Agent is processing...", id="processing-indicator")
-        yield Input(placeholder="Enter input for agent...", id="chat-input")
+        yield ChatTextArea(id="chat-input", placeholder="Enter input for agent...")
 
     # Regex for file:// URIs that are NOT already inside Rich [link=...] markup
     _FILE_URI_RE = re.compile(r"(?<!\[link=)(file://[^\s)\]>*]+)")
@@ -129,6 +164,31 @@ class ChatRepl(Vertical):
         if was_at_bottom:
             history.scroll_end(animate=False)
 
+    def toggle_logs(self) -> None:
+        """Toggle inline log display on/off. Backfills buffered logs on toggle ON."""
+        self._show_logs = not self._show_logs
+        if self._show_logs and self._log_buffer:
+            self._write_history("[dim]--- Backfilling logs ---[/dim]")
+            for line in self._log_buffer:
+                self._write_history(line)
+            self._write_history("[dim]--- Live logs ---[/dim]")
+        mode = "ON (dirty)" if self._show_logs else "OFF (clean)"
+        self._write_history(f"[dim]Logs {mode}[/dim]")
+
+    def write_log_event(self, event: AgentEvent) -> None:
+        """Buffer a formatted agent event. Display inline if logs are ON."""
+        formatted = format_event(event)
+        self._log_buffer.append(formatted)
+        if self._show_logs:
+            self._write_history(formatted)
+
+    def write_python_log(self, record: logging.LogRecord) -> None:
+        """Buffer a formatted Python log record. Display inline if logs are ON."""
+        formatted = format_python_log(record)
+        self._log_buffer.append(formatted)
+        if self._show_logs:
+            self._write_history(formatted)
+
     async def _handle_command(self, command: str) -> None:
         """Handle slash commands for session and checkpoint operations."""
         parts = command.split(maxsplit=2)
@@ -138,8 +198,9 @@ class ChatRepl(Vertical):
             self._write_history("""[bold cyan]Available Commands:[/bold cyan]
   [bold]/sessions[/bold]                    - List all sessions for this agent
   [bold]/sessions[/bold] <session_id>       - Show session details and checkpoints
-  [bold]/resume[/bold]                     - Resume latest paused/failed session
-  [bold]/resume[/bold] <session_id>         - Resume session from where it stopped
+  [bold]/resume[/bold]                      - List sessions and pick one to resume
+  [bold]/resume[/bold] <number>             - Resume session by list number
+  [bold]/resume[/bold] <session_id>         - Resume session by ID
   [bold]/recover[/bold] <session_id> <cp_id> - Recover from specific checkpoint
   [bold]/pause[/bold]                      - Pause current execution (Ctrl+Z)
   [bold]/help[/bold]                       - Show this help message
@@ -147,8 +208,9 @@ class ChatRepl(Vertical):
 [dim]Examples:[/dim]
   /sessions                              [dim]# List all sessions[/dim]
   /sessions session_20260208_143022      [dim]# Show session details[/dim]
-  /resume                                [dim]# Resume latest session (from state)[/dim]
-  /resume session_20260208_143022        [dim]# Resume specific session (from state)[/dim]
+  /resume                                [dim]# Show numbered session list[/dim]
+  /resume 1                              [dim]# Resume first listed session[/dim]
+  /resume session_20260208_143022        [dim]# Resume by full session ID[/dim]
   /recover session_20260208_143022 cp_xxx [dim]# Recover from specific checkpoint[/dim]
   /pause                                 [dim]# Pause (or Ctrl+Z)[/dim]
 """)
@@ -156,15 +218,25 @@ class ChatRepl(Vertical):
             session_id = parts[1].strip() if len(parts) > 1 else None
             await self._cmd_sessions(session_id)
         elif cmd == "/resume":
-            # Resume from session state (not checkpoint-based)
             if len(parts) < 2:
-                session_id = await self._find_latest_resumable_session()
-                if not session_id:
-                    self._write_history("[bold red]No resumable sessions found[/bold red]")
-                    self._write_history("  Tip: Use [bold]/sessions[/bold] to see all sessions")
+                # No arg → show session list so user can pick one
+                await self._cmd_sessions(None)
+                return
+
+            arg = parts[1].strip()
+
+            # Numeric index → resolve from last listing
+            if arg.isdigit():
+                idx = int(arg) - 1  # 1-based to 0-based
+                if 0 <= idx < len(self._session_index):
+                    session_id = self._session_index[idx]
+                else:
+                    self._write_history(f"[bold red]Error:[/bold red] No session at index {arg}")
+                    self._write_history("  Use [bold]/resume[/bold] to see available sessions")
                     return
             else:
-                session_id = parts[1].strip()
+                session_id = arg
+
             await self._cmd_resume(session_id)
         elif cmd == "/recover":
             # Recover from specific checkpoint
@@ -241,6 +313,15 @@ class ChatRepl(Vertical):
         except Exception:
             return None
 
+    def _get_session_label(self, state: dict) -> str:
+        """Extract the first user message from input_data as a human-readable label."""
+        input_data = state.get("input_data", {})
+        for value in input_data.values():
+            if isinstance(value, str) and value.strip():
+                label = value.strip()
+                return label[:60] + "..." if len(label) > 60 else label
+        return "(no input)"
+
     async def _list_sessions(self, storage_path: Path) -> None:
         """List all sessions for the agent."""
         self._write_history("[bold cyan]Available Sessions:[/bold cyan]")
@@ -264,6 +345,11 @@ class ChatRepl(Vertical):
 
         self._write_history(f"[dim]Found {len(session_dirs)} session(s)[/dim]\n")
 
+        # Reset the session index for numeric lookups
+        self._session_index = []
+
+        import json
+
         for session_dir in session_dirs[:10]:  # Show last 10 sessions
             session_id = session_dir.name
             state_file = session_dir / "state.json"
@@ -273,12 +359,15 @@ class ChatRepl(Vertical):
 
             # Read session state
             try:
-                import json
-
                 with open(state_file) as f:
                     state = json.load(f)
 
+                # Track this session for /resume <number> lookup
+                self._session_index.append(session_id)
+                index = len(self._session_index)
+
                 status = state.get("status", "unknown").upper()
+                label = self._get_session_label(state)
 
                 # Status with color
                 if status == "COMPLETED":
@@ -292,25 +381,17 @@ class ChatRepl(Vertical):
                 else:
                     status_colored = f"[dim]{status}[/dim]"
 
-                # Check for checkpoints
-                checkpoint_dir = session_dir / "checkpoints"
-                checkpoint_count = 0
-                if checkpoint_dir.exists():
-                    checkpoint_files = list(checkpoint_dir.glob("cp_*.json"))
-                    checkpoint_count = len(checkpoint_files)
-
-                # Session line
-                self._write_history(f"📋 [bold]{session_id}[/bold]")
-                self._write_history(f"   Status: {status_colored}  Checkpoints: {checkpoint_count}")
-
-                if checkpoint_count > 0:
-                    self._write_history(f"   [dim]Resume: /resume {session_id}[/dim]")
-
+                # Session line with index and label
+                self._write_history(f"  [bold]{index}.[/bold] {label}  {status_colored}")
+                self._write_history(f"     [dim]{session_id}[/dim]")
                 self._write_history("")  # Blank line
 
             except Exception as e:
                 self._write_history(f"   [dim red]Error reading: {e}[/dim red]")
 
+        if self._session_index:
+            self._write_history("[dim]Use [bold]/resume <number>[/bold] to resume a session[/dim]")
+
     async def _show_session_details(self, storage_path: Path, session_id: str) -> None:
         """Show detailed information about a specific session."""
         self._write_history(f"[bold cyan]Session Details:[/bold cyan] {session_id}\n")
@@ -428,6 +509,7 @@ class ChatRepl(Vertical):
             if paused_at:
                 # Has paused_at - resume from there
                 resume_session_state = {
+                    "resume_session_id": session_id,
                     "paused_at": paused_at,
                     "memory": state.get("memory", {}),
                     "execution_path": progress.get("path", []),
@@ -435,8 +517,13 @@ class ChatRepl(Vertical):
                 }
                 resume_info = f"From node: [cyan]{paused_at}[/cyan]"
             else:
-                # No paused_at - just retry with same input
-                resume_session_state = {}
+                # No paused_at - retry with same input but reuse session directory
+                resume_session_state = {
+                    "resume_session_id": session_id,
+                    "memory": state.get("memory", {}),
+                    "execution_path": progress.get("path", []),
+                    "node_visit_counts": progress.get("node_visit_counts", {}),
+                }
                 resume_info = "Retrying with same input"
 
             # Display resume info
@@ -462,7 +549,7 @@ class ChatRepl(Vertical):
             indicator.display = True
 
             # Update placeholder
-            chat_input = self.query_one("#chat-input", Input)
+            chat_input = self.query_one("#chat-input", ChatTextArea)
             chat_input.placeholder = "Commands: /pause, /sessions (agent resuming...)"
 
             # Trigger execution with resume state
@@ -540,6 +627,7 @@ class ChatRepl(Vertical):
 
             # Create session_state for checkpoint recovery
             recover_session_state = {
+                "resume_session_id": session_id,
                 "resume_from_checkpoint": checkpoint_id,
             }
 
@@ -549,7 +637,7 @@ class ChatRepl(Vertical):
             indicator.display = True
 
             # Update placeholder
-            chat_input = self.query_one("#chat-input", Input)
+            chat_input = self.query_one("#chat-input", ChatTextArea)
             chat_input.placeholder = "Commands: /pause, /sessions (agent recovering...)"
 
             # Trigger execution with checkpoint recovery
@@ -638,10 +726,14 @@ class ChatRepl(Vertical):
         # Check for resumable sessions
         self._check_and_show_resumable_sessions()
 
-        history.write(
-            "[dim]Quick start: /sessions to see previous sessions, "
-            "/pause to pause execution[/dim]\n"
-        )
+        # Show agent intro message if available
+        if self.runtime.intro_message:
+            history.write(f"[bold blue]Agent:[/bold blue] {self.runtime.intro_message}\n")
+        else:
+            history.write(
+                "[dim]Quick start: /sessions to see previous sessions, "
+                "/pause to pause execution[/dim]\n"
+            )
 
     def _check_and_show_resumable_sessions(self) -> None:
         """Check for non-terminated sessions and prompt user."""
@@ -678,16 +770,20 @@ class ChatRepl(Vertical):
                             {
                                 "session_id": session_dir.name,
                                 "status": status.upper(),
+                                "label": self._get_session_label(state),
                             }
                         )
                 except Exception:
                     continue
 
             if resumable:
-                self._write_history("\n[bold yellow]⚠ Non-terminated sessions found:[/bold yellow]")
+                # Populate session index so /resume <number> works immediately
+                self._session_index = [s["session_id"] for s in resumable[:3]]
+
+                self._write_history("\n[bold yellow]Non-terminated sessions found:[/bold yellow]")
                 for i, session in enumerate(resumable[:3], 1):  # Show top 3
                     status = session["status"]
-                    session_id = session["session_id"]
+                    label = session["label"]
 
                     # Color code status
                     if status == "PAUSED":
@@ -699,23 +795,21 @@ class ChatRepl(Vertical):
                     else:
                         status_colored = f"[dim]{status}[/dim]"
 
-                    self._write_history(f"  {i}. {session_id[:32]}... [{status_colored}]")
+                    self._write_history(f"  [bold]{i}.[/bold] {label}  {status_colored}")
 
-                self._write_history("\n[bold cyan]What would you like to do?[/bold cyan]")
-                self._write_history("  • Type [bold]/resume[/bold] to continue the latest session")
-                self._write_history(
-                    f"  • Type [bold]/resume {resumable[0]['session_id']}[/bold] "
-                    "for specific session"
-                )
-                self._write_history("  • Or just type your input to start a new session\n")
+                self._write_history("\n  Type [bold]/resume <number>[/bold] to continue a session")
+                self._write_history("  Or just type your input to start a new session\n")
 
         except Exception:
             # Silently fail - don't block TUI startup
             pass
 
-    async def on_input_submitted(self, message: Input.Submitted) -> None:
-        """Handle input submission — either start new execution or inject input."""
-        user_input = message.value.strip()
+    async def on_chat_text_area_submitted(self, message: ChatTextArea.Submitted) -> None:
+        """Handle chat input submission."""
+        await self._submit_input(message.text)
+
+    async def _submit_input(self, user_input: str) -> None:
+        """Handle submitted text — either start new execution or inject input."""
         if not user_input:
             return
 
@@ -723,16 +817,14 @@ class ChatRepl(Vertical):
         # Commands work during execution, during client-facing input, anytime
         if user_input.startswith("/"):
             await self._handle_command(user_input)
-            message.input.value = ""
             return
 
         # Client-facing input: route to the waiting node
         if self._waiting_for_input and self._input_node_id:
             self._write_history(f"[bold green]You:[/bold green] {user_input}")
-            message.input.value = ""
 
             # Keep input enabled for commands (but change placeholder)
-            chat_input = self.query_one("#chat-input", Input)
+            chat_input = self.query_one("#chat-input", ChatTextArea)
             chat_input.placeholder = "Commands: /pause, /sessions (agent processing...)"
             self._waiting_for_input = False
 
@@ -752,16 +844,29 @@ class ChatRepl(Vertical):
                 self._write_history(f"[bold red]Error delivering input:[/bold red] {e}")
             return
 
-        # Double-submit guard: reject input while an execution is in-flight
+        # Mid-execution input: inject into the active node's conversation
+        if self._current_exec_id is not None and self._active_node_id:
+            self._write_history(f"[bold green]You:[/bold green] {user_input}")
+            node_id = self._active_node_id
+            try:
+                future = asyncio.run_coroutine_threadsafe(
+                    self.runtime.inject_input(node_id, user_input),
+                    self._agent_loop,
+                )
+                await asyncio.wrap_future(future)
+            except Exception as e:
+                self._write_history(f"[bold red]Error delivering input:[/bold red] {e}")
+            return
+
+        # Double-submit guard: no active node to inject into
         if self._current_exec_id is not None:
             self._write_history("[dim]Agent is still running — please wait.[/dim]")
             return
 
         indicator = self.query_one("#processing-indicator", Label)
 
-        # Append user message and clear input
+        # Append user message
         self._write_history(f"[bold green]You:[/bold green] {user_input}")
-        message.input.value = ""
 
         try:
             # Get entry point
@@ -787,7 +892,7 @@ class ChatRepl(Vertical):
             indicator.display = True
 
             # Keep input enabled for commands during execution
-            chat_input = self.query_one("#chat-input", Input)
+            chat_input = self.query_one("#chat-input", ChatTextArea)
             chat_input.placeholder = "Commands available: /pause, /sessions, /help"
 
             # Submit execution to the dedicated agent loop so blocking
@@ -808,12 +913,32 @@ class ChatRepl(Vertical):
             indicator.display = False
             self._current_exec_id = None
             # Re-enable input on error
-            chat_input = self.query_one("#chat-input", Input)
+            chat_input = self.query_one("#chat-input", ChatTextArea)
             chat_input.disabled = False
             self._write_history(f"[bold red]Error:[/bold red] {e}")
 
     # -- Event handlers called by app.py _handle_event --
 
+    def handle_node_started(self, node_id: str) -> None:
+        """Reset streaming state and track active node when a new node begins.
+
+        Flushes any stale ``_streaming_snapshot`` left over from the
+        previous node and resets the processing indicator so the user
+        sees a clean transition between graph nodes.
+        """
+        self._active_node_id = node_id
+        if self._streaming_snapshot:
+            self._write_history(f"[bold blue]Agent:[/bold blue] {self._streaming_snapshot}")
+            self._streaming_snapshot = ""
+        indicator = self.query_one("#processing-indicator", Label)
+        indicator.update("Thinking...")
+
+    def handle_loop_iteration(self, iteration: int) -> None:
+        """Flush accumulated streaming text when a new loop iteration starts."""
+        if self._streaming_snapshot:
+            self._write_history(f"[bold blue]Agent:[/bold blue] {self._streaming_snapshot}")
+            self._streaming_snapshot = ""
+
     def handle_text_delta(self, content: str, snapshot: str) -> None:
         """Handle a streaming text token from the LLM."""
         self._streaming_snapshot = snapshot
@@ -829,23 +954,42 @@ class ChatRepl(Vertical):
 
     def handle_tool_started(self, tool_name: str, tool_input: dict[str, Any]) -> None:
         """Handle a tool call starting."""
-        # Update indicator to show tool activity
         indicator = self.query_one("#processing-indicator", Label)
+
+        if tool_name == "ask_user":
+            # Stash the question for handle_input_requested() to display.
+            # Suppress the generic "Tool: ask_user" line.
+            self._pending_ask_question = tool_input.get("question", "")
+            indicator.update("Preparing question...")
+            return
+
+        # Update indicator to show tool activity
         indicator.update(f"Using tool: {tool_name}...")
 
-        # Write a discrete status line to history
-        self._write_history(f"[dim]Tool: {tool_name}[/dim]")
+        # Buffer and conditionally display tool status line
+        line = f"[dim]Tool: {tool_name}[/dim]"
+        self._log_buffer.append(line)
+        if self._show_logs:
+            self._write_history(line)
 
     def handle_tool_completed(self, tool_name: str, result: str, is_error: bool) -> None:
         """Handle a tool call completing."""
+        if tool_name == "ask_user":
+            # Suppress the synthetic "Waiting for user input..." result.
+            # The actual question is displayed by handle_input_requested().
+            return
+
         result_str = str(result)
         preview = result_str[:200] + "..." if len(result_str) > 200 else result_str
         preview = preview.replace("\n", " ")
 
         if is_error:
-            self._write_history(f"[dim red]Tool {tool_name} error: {preview}[/dim red]")
+            line = f"[dim red]Tool {tool_name} error: {preview}[/dim red]"
         else:
-            self._write_history(f"[dim]Tool {tool_name} result: {preview}[/dim]")
+            line = f"[dim]Tool {tool_name} result: {preview}[/dim]"
+        self._log_buffer.append(line)
+        if self._show_logs:
+            self._write_history(line)
 
         # Restore thinking indicator
         indicator = self.query_one("#processing-indicator", Label)
@@ -868,9 +1012,12 @@ class ChatRepl(Vertical):
         self._streaming_snapshot = ""
         self._waiting_for_input = False
         self._input_node_id = None
+        self._active_node_id = None
+        self._pending_ask_question = ""
+        self._log_buffer.clear()
 
         # Re-enable input
-        chat_input = self.query_one("#chat-input", Input)
+        chat_input = self.query_one("#chat-input", ChatTextArea)
         chat_input.disabled = False
         chat_input.placeholder = "Enter input for agent..."
         chat_input.focus()
@@ -886,10 +1033,13 @@ class ChatRepl(Vertical):
         self._current_exec_id = None
         self._streaming_snapshot = ""
         self._waiting_for_input = False
+        self._pending_ask_question = ""
         self._input_node_id = None
+        self._active_node_id = None
+        self._log_buffer.clear()
 
         # Re-enable input
-        chat_input = self.query_one("#chat-input", Input)
+        chat_input = self.query_one("#chat-input", ChatTextArea)
         chat_input.disabled = False
         chat_input.placeholder = "Enter input for agent..."
         chat_input.focus()
@@ -902,17 +1052,54 @@ class ChatRepl(Vertical):
         and sets a flag so the next submission routes to inject_input().
         """
         # Flush accumulated streaming text as agent output
-        if self._streaming_snapshot:
-            self._write_history(f"[bold blue]Agent:[/bold blue] {self._streaming_snapshot}")
+        flushed_snapshot = self._streaming_snapshot
+        if flushed_snapshot:
+            self._write_history(f"[bold blue]Agent:[/bold blue] {flushed_snapshot}")
             self._streaming_snapshot = ""
 
+        # Display the ask_user question if stashed and not already
+        # present in the streaming snapshot (avoids double-display).
+        question = self._pending_ask_question
+        self._pending_ask_question = ""
+        if question and question not in flushed_snapshot:
+            self._write_history(f"[bold blue]Agent:[/bold blue] {question}")
+
         self._waiting_for_input = True
         self._input_node_id = node_id or None
 
         indicator = self.query_one("#processing-indicator", Label)
         indicator.update("Waiting for your input...")
 
-        chat_input = self.query_one("#chat-input", Input)
+        chat_input = self.query_one("#chat-input", ChatTextArea)
         chat_input.disabled = False
         chat_input.placeholder = "Type your response..."
         chat_input.focus()
+
+    def handle_node_completed(self, node_id: str) -> None:
+        """Clear active node when it finishes."""
+        if self._active_node_id == node_id:
+            self._active_node_id = None
+
+    def handle_internal_output(self, node_id: str, content: str) -> None:
+        """Show output from non-client-facing nodes."""
+        self._write_history(f"[dim cyan]⟨{node_id}⟩[/dim cyan] {content}")
+
+    def handle_execution_paused(self, node_id: str, reason: str) -> None:
+        """Show that execution has been paused."""
+        msg = f"[bold yellow]⏸ Paused[/bold yellow] at [cyan]{node_id}[/cyan]"
+        if reason:
+            msg += f" [dim]({reason})[/dim]"
+        self._write_history(msg)
+
+    def handle_execution_resumed(self, node_id: str) -> None:
+        """Show that execution has been resumed."""
+        self._write_history(f"[bold green]▶ Resumed[/bold green] from [cyan]{node_id}[/cyan]")
+
+    def handle_goal_achieved(self, data: dict[str, Any]) -> None:
+        """Show goal achievement prominently."""
+        self._write_history("[bold green]★ Goal achieved![/bold green]")
+
+    def handle_constraint_violation(self, data: dict[str, Any]) -> None:
+        """Show constraint violation as a warning."""
+        desc = data.get("description", "Unknown constraint")
+        self._write_history(f"[bold red]⚠ Constraint violation:[/bold red] {desc}")

core/framework/tui/widgets/graph_view.py

@@ -1,7 +1,14 @@
 """
 Graph/Tree Overview Widget - Displays real agent graph structure.
+
+Supports rendering loops (back-edges) via right-side return channels:
+arrows drawn on the right margin that visually point back up to earlier nodes.
 """
 
+from __future__ import annotations
+
+import re
+
 from textual.app import ComposeResult
 from textual.containers import Vertical
 
@@ -9,6 +16,17 @@ from framework.runtime.agent_runtime import AgentRuntime
 from framework.runtime.event_bus import EventType
 from framework.tui.widgets.selectable_rich_log import SelectableRichLog as RichLog
 
+# Width of each return-channel column (padding + │ + gap)
+_CHANNEL_WIDTH = 5
+
+# Regex to strip Rich markup tags for measuring visible width
+_MARKUP_RE = re.compile(r"\[/?[^\]]*\]")
+
+
+def _plain_len(s: str) -> int:
+    """Return the visible character length of a Rich-markup string."""
+    return len(_MARKUP_RE.sub("", s))
+
 
 class GraphOverview(Vertical):
     """Widget to display Agent execution graph/tree with real data."""
@@ -47,6 +65,10 @@ class GraphOverview(Vertical):
         """Display initial graph structure."""
         self._display_graph()
 
+    # ------------------------------------------------------------------
+    # Graph analysis helpers
+    # ------------------------------------------------------------------
+
     def _topo_order(self) -> list[str]:
         """BFS from entry_node following edges."""
         graph = self.runtime.graph
@@ -68,6 +90,39 @@ class GraphOverview(Vertical):
                 visited.append(node.id)
         return visited
 
+    def _detect_back_edges(self, ordered: list[str]) -> list[dict]:
+        """Find edges where target appears before (or equal to) source in topo order.
+
+        Returns a list of dicts with keys: edge, source, target, source_idx, target_idx.
+        """
+        order_idx = {nid: i for i, nid in enumerate(ordered)}
+        back_edges: list[dict] = []
+        for node_id in ordered:
+            for edge in self.runtime.graph.get_outgoing_edges(node_id):
+                target_idx = order_idx.get(edge.target, -1)
+                source_idx = order_idx.get(node_id, -1)
+                if target_idx != -1 and target_idx <= source_idx:
+                    back_edges.append(
+                        {
+                            "edge": edge,
+                            "source": node_id,
+                            "target": edge.target,
+                            "source_idx": source_idx,
+                            "target_idx": target_idx,
+                        }
+                    )
+        return back_edges
+
+    def _is_back_edge(self, source: str, target: str, order_idx: dict[str, int]) -> bool:
+        """Check whether an edge from *source* to *target* is a back-edge."""
+        si = order_idx.get(source, -1)
+        ti = order_idx.get(target, -1)
+        return ti != -1 and ti <= si
+
+    # ------------------------------------------------------------------
+    # Line rendering (Pass 1)
+    # ------------------------------------------------------------------
+
     def _render_node_line(self, node_id: str) -> str:
         """Render a single node with status symbol and optional status text."""
         graph = self.runtime.graph
@@ -95,43 +150,290 @@ class GraphOverview(Vertical):
         suffix = f"  [italic]{status}[/italic]" if status else ""
         return f"  {sym} {name}{suffix}"
 
-    def _render_edges(self, node_id: str) -> list[str]:
-        """Render edge connectors from this node to its targets."""
-        edges = self.runtime.graph.get_outgoing_edges(node_id)
-        if not edges:
+    def _render_edges(self, node_id: str, order_idx: dict[str, int]) -> list[str]:
+        """Render forward-edge connectors from *node_id*.
+
+        Back-edges are excluded here — they are drawn by the return-channel
+        overlay in Pass 2.
+        """
+        all_edges = self.runtime.graph.get_outgoing_edges(node_id)
+        if not all_edges:
             return []
-        if len(edges) == 1:
+
+        # Split into forward and back
+        forward = [e for e in all_edges if not self._is_back_edge(node_id, e.target, order_idx)]
+
+        if not forward:
+            # All edges are back-edges — nothing to render here
+            return []
+
+        if len(forward) == 1:
             return ["  │", "  ▼"]
+
         # Fan-out: show branches
         lines: list[str] = []
-        for i, edge in enumerate(edges):
-            connector = "└" if i == len(edges) - 1 else "├"
+        for i, edge in enumerate(forward):
+            connector = "└" if i == len(forward) - 1 else "├"
             cond = ""
             if edge.condition.value not in ("always", "on_success"):
                 cond = f" [dim]({edge.condition.value})[/dim]"
             lines.append(f"  {connector}──▶ {edge.target}{cond}")
         return lines
 
+    # ------------------------------------------------------------------
+    # Return-channel overlay (Pass 2)
+    # ------------------------------------------------------------------
+
+    def _overlay_return_channels(
+        self,
+        lines: list[str],
+        node_line_map: dict[str, int],
+        back_edges: list[dict],
+        available_width: int,
+    ) -> list[str]:
+        """Overlay right-side return channels onto the line buffer.
+
+        Each back-edge gets a vertical channel on the right margin.  Channels
+        are allocated left-to-right by increasing span length so that shorter
+        (inner) loops are closer to the graph body and longer (outer) loops are
+        further right.
+
+        If the terminal is too narrow to fit even one channel, we fall back to
+        simple inline ``↺`` annotations instead.
+        """
+        if not back_edges:
+            return lines
+
+        num_channels = len(back_edges)
+
+        # Sort by span length ascending → inner loops get nearest channel
+        sorted_be = sorted(back_edges, key=lambda b: b["source_idx"] - b["target_idx"])
+
+        # --- Insert dedicated connector lines for back-edge sources ---
+        # Each back-edge source gets a blank line inserted after its node
+        # section (after any forward-edge lines).  We process insertions in
+        # reverse order so that earlier indices remain valid.
+        all_node_lines_set = set(node_line_map.values())
+
+        insertions: list[tuple[int, int]] = []  # (insert_after_line, be_index)
+        for be_idx, be in enumerate(sorted_be):
+            source_node_line = node_line_map.get(be["source"])
+            if source_node_line is None:
+                continue
+            # Walk forward to find the last line in this node's section
+            last_section_line = source_node_line
+            for li in range(source_node_line + 1, len(lines)):
+                if li in all_node_lines_set:
+                    break
+                last_section_line = li
+            insertions.append((last_section_line, be_idx))
+
+        source_line_for_be: dict[int, int] = {}
+        for insert_after, be_idx in sorted(insertions, reverse=True):
+            insert_at = insert_after + 1
+            lines.insert(insert_at, "")  # placeholder for connector
+            source_line_for_be[be_idx] = insert_at
+            # Shift node_line_map entries that come after the insertion point
+            for nid in node_line_map:
+                if node_line_map[nid] > insert_after:
+                    node_line_map[nid] += 1
+            # Also shift already-assigned source lines
+            for prev_idx in source_line_for_be:
+                if prev_idx != be_idx and source_line_for_be[prev_idx] > insert_after:
+                    source_line_for_be[prev_idx] += 1
+
+        # Recompute max content width after insertions
+        max_content_w = max(_plain_len(ln) for ln in lines) if lines else 0
+
+        # Check if we have room for channels
+        channels_total_w = num_channels * _CHANNEL_WIDTH
+        if max_content_w + channels_total_w + 2 > available_width:
+            return self._inline_back_edge_fallback(lines, node_line_map, back_edges)
+
+        content_pad = max_content_w + 3  # gap between content and first channel
+
+        # Build channel info with final line positions
+        channel_info: list[dict] = []
+        for ch_idx, be in enumerate(sorted_be):
+            target_line = node_line_map.get(be["target"])
+            source_line = source_line_for_be.get(ch_idx)
+            if target_line is None or source_line is None:
+                continue
+            col = content_pad + ch_idx * _CHANNEL_WIDTH
+            channel_info.append(
+                {
+                    "target_line": target_line,
+                    "source_line": source_line,
+                    "col": col,
+                }
+            )
+
+        if not channel_info:
+            return lines
+
+        # Build overlay grid — one row per line, columns for channel area
+        total_width = content_pad + num_channels * _CHANNEL_WIDTH + 1
+        overlay_width = total_width - max_content_w
+        overlays: list[list[str]] = [[" "] * overlay_width for _ in range(len(lines))]
+
+        for ci in channel_info:
+            tl = ci["target_line"]
+            sl = ci["source_line"]
+            col_offset = ci["col"] - max_content_w
+
+            if col_offset < 0 or col_offset >= overlay_width:
+                continue
+
+            # Target line: ◄──...──┐
+            if 0 <= tl < len(overlays):
+                for c in range(col_offset):
+                    if overlays[tl][c] == " ":
+                        overlays[tl][c] = "─"
+                overlays[tl][col_offset] = "┐"
+
+            # Source line: ──...──┘
+            if 0 <= sl < len(overlays):
+                for c in range(col_offset):
+                    if overlays[sl][c] == " ":
+                        overlays[sl][c] = "─"
+                overlays[sl][col_offset] = "┘"
+
+            # Vertical lines between target+1 and source-1
+            for li in range(tl + 1, sl):
+                if 0 <= li < len(overlays) and overlays[li][col_offset] == " ":
+                    overlays[li][col_offset] = "│"
+
+        # Merge overlays into the line strings
+        result: list[str] = []
+        for i, line in enumerate(lines):
+            pw = _plain_len(line)
+            pad = max_content_w - pw
+            overlay_chars = overlays[i] if i < len(overlays) else []
+            overlay_str = "".join(overlay_chars)
+            overlay_trimmed = overlay_str.rstrip()
+            if overlay_trimmed:
+                is_target_line = any(ci["target_line"] == i for ci in channel_info)
+                if is_target_line:
+                    overlay_trimmed = "◄" + overlay_trimmed[1:]
+
+                is_source_line = any(ci["source_line"] == i for ci in channel_info)
+                if is_source_line and not line.strip():
+                    # Inserted blank line → build └───┘ connector.
+                    # "  └" = 3 chars of content prefix, so remaining pad = max_content_w - 3
+                    remaining_pad = max_content_w - 3
+                    full = list(" " * remaining_pad + overlay_trimmed)
+                    # Find the ┘ corner for this source connector
+                    corner_pos = -1
+                    for ci_s in channel_info:
+                        if ci_s["source_line"] == i:
+                            corner_pos = remaining_pad + (ci_s["col"] - max_content_w)
+                            break
+                    # Fill everything up to the corner with ─
+                    if corner_pos >= 0:
+                        for c in range(corner_pos):
+                            if full[c] not in ("│", "┘", "┐"):
+                                full[c] = "─"
+                    connector = "  └" + "".join(full).rstrip()
+                    result.append(f"[dim]{connector}[/dim]")
+                    continue
+
+                colored_overlay = f"[dim]{' ' * pad}{overlay_trimmed}[/dim]"
+                result.append(f"{line}{colored_overlay}")
+            else:
+                result.append(line)
+
+        return result
+
+    def _inline_back_edge_fallback(
+        self,
+        lines: list[str],
+        node_line_map: dict[str, int],
+        back_edges: list[dict],
+    ) -> list[str]:
+        """Fallback: add inline ↺ annotations when terminal is too narrow for channels."""
+        # Group back-edges by source node
+        source_to_be: dict[str, list[dict]] = {}
+        for be in back_edges:
+            source_to_be.setdefault(be["source"], []).append(be)
+
+        result = list(lines)
+        # Insert annotation lines after each source node's section
+        offset = 0
+        all_node_lines = sorted(node_line_map.values())
+        for source, bes in source_to_be.items():
+            source_line = node_line_map.get(source)
+            if source_line is None:
+                continue
+            # Find end of source node section
+            end_line = source_line
+            for nl in all_node_lines:
+                if nl > source_line:
+                    end_line = nl - 1
+                    break
+            else:
+                end_line = len(lines) - 1
+            # Insert after last content line of this node's section
+            insert_at = end_line + offset + 1
+            for be in bes:
+                cond = ""
+                edge = be["edge"]
+                if edge.condition.value not in ("always", "on_success"):
+                    cond = f" [dim]({edge.condition.value})[/dim]"
+                annotation = f"  [yellow]↺[/yellow] {be['target']}{cond}"
+                result.insert(insert_at, annotation)
+                insert_at += 1
+                offset += 1
+
+        return result
+
+    # ------------------------------------------------------------------
+    # Main display
+    # ------------------------------------------------------------------
+
     def _display_graph(self) -> None:
-        """Display the graph as an ASCII DAG with edge connectors."""
+        """Display the graph as an ASCII DAG with edge connectors and loop channels."""
         display = self.query_one("#graph-display", RichLog)
         display.clear()
 
         graph = self.runtime.graph
         display.write(f"[bold cyan]Agent Graph:[/bold cyan] {graph.id}\n")
 
-        # Render each node in topological order with edges
         ordered = self._topo_order()
+        order_idx = {nid: i for i, nid in enumerate(ordered)}
+
+        # --- Pass 1: Build line buffer ---
+        lines: list[str] = []
+        node_line_map: dict[str, int] = {}
+
         for node_id in ordered:
-            display.write(self._render_node_line(node_id))
-            for edge_line in self._render_edges(node_id):
-                display.write(edge_line)
+            node_line_map[node_id] = len(lines)
+            lines.append(self._render_node_line(node_id))
+            for edge_line in self._render_edges(node_id, order_idx):
+                lines.append(edge_line)
+
+        # --- Pass 2: Overlay return channels for back-edges ---
+        back_edges = self._detect_back_edges(ordered)
+        if back_edges:
+            # Try to get actual widget width; default to a reasonable value
+            try:
+                available_width = self.size.width or 60
+            except Exception:
+                available_width = 60
+            lines = self._overlay_return_channels(lines, node_line_map, back_edges, available_width)
+
+        # Write all lines
+        for line in lines:
+            display.write(line)
 
         # Execution path footer
         if self.execution_path:
             display.write("")
             display.write(f"[dim]Path:[/dim] {' → '.join(self.execution_path[-5:])}")
 
+    # ------------------------------------------------------------------
+    # Public API (called by app.py)
+    # ------------------------------------------------------------------
+
     def update_active_node(self, node_id: str) -> None:
         """Update the currently active node."""
         self.active_node = node_id
@@ -177,6 +479,8 @@ class GraphOverview(Vertical):
     def handle_node_loop_completed(self, node_id: str) -> None:
         """A node's event loop completed."""
         self._node_status.pop(node_id, None)
+        if self.active_node == node_id:
+            self.active_node = None
         self._display_graph()
 
     def handle_tool_call(self, node_id: str, tool_name: str, *, started: bool) -> None:
@@ -192,3 +496,8 @@ class GraphOverview(Vertical):
         """Highlight a stalled node."""
         self._node_status[node_id] = f"[red]stalled: {reason}[/red]"
         self._display_graph()
+
+    def handle_edge_traversed(self, source_node: str, target_node: str) -> None:
+        """Highlight an edge being traversed."""
+        self._node_status[source_node] = f"[dim]→ {target_node}[/dim]"
+        self._display_graph()

core/framework/tui/widgets/log_pane.py

@@ -1,5 +1,8 @@
 """
-Log Pane Widget - Uses RichLog for reliable rendering.
+Log formatting utilities and LogPane widget.
+
+The module-level functions (format_event, extract_event_text, format_python_log)
+can be used by any widget that needs to render log lines without instantiating LogPane.
 """
 
 import logging
@@ -11,36 +14,108 @@ from textual.containers import Container
 from framework.runtime.event_bus import AgentEvent, EventType
 from framework.tui.widgets.selectable_rich_log import SelectableRichLog as RichLog
 
+# --- Module-level formatting constants ---
+
+EVENT_FORMAT: dict[EventType, tuple[str, str]] = {
+    EventType.EXECUTION_STARTED: (">>", "bold cyan"),
+    EventType.EXECUTION_COMPLETED: ("<<", "bold green"),
+    EventType.EXECUTION_FAILED: ("!!", "bold red"),
+    EventType.TOOL_CALL_STARTED: ("->", "yellow"),
+    EventType.TOOL_CALL_COMPLETED: ("<-", "green"),
+    EventType.NODE_LOOP_STARTED: ("@@", "cyan"),
+    EventType.NODE_LOOP_ITERATION: ("..", "dim"),
+    EventType.NODE_LOOP_COMPLETED: ("@@", "dim"),
+    EventType.NODE_STALLED: ("!!", "bold yellow"),
+    EventType.NODE_INPUT_BLOCKED: ("!!", "yellow"),
+    EventType.GOAL_PROGRESS: ("%%", "blue"),
+    EventType.GOAL_ACHIEVED: ("**", "bold green"),
+    EventType.CONSTRAINT_VIOLATION: ("!!", "bold red"),
+    EventType.STATE_CHANGED: ("~~", "dim"),
+    EventType.CLIENT_INPUT_REQUESTED: ("??", "magenta"),
+}
+
+LOG_LEVEL_COLORS: dict[int, str] = {
+    logging.DEBUG: "dim",
+    logging.INFO: "",
+    logging.WARNING: "yellow",
+    logging.ERROR: "red",
+    logging.CRITICAL: "bold red",
+}
+
+
+# --- Module-level formatting functions ---
+
+
+def extract_event_text(event: AgentEvent) -> str:
+    """Extract human-readable text from an event's data dict."""
+    et = event.type
+    data = event.data
+
+    if et == EventType.EXECUTION_STARTED:
+        return "Execution started"
+    elif et == EventType.EXECUTION_COMPLETED:
+        return "Execution completed"
+    elif et == EventType.EXECUTION_FAILED:
+        return f"Execution FAILED: {data.get('error', 'unknown')}"
+    elif et == EventType.TOOL_CALL_STARTED:
+        return f"Tool call: {data.get('tool_name', 'unknown')}"
+    elif et == EventType.TOOL_CALL_COMPLETED:
+        name = data.get("tool_name", "unknown")
+        if data.get("is_error"):
+            preview = str(data.get("result", ""))[:80]
+            return f"Tool error: {name} - {preview}"
+        return f"Tool done: {name}"
+    elif et == EventType.NODE_LOOP_STARTED:
+        return f"Node started: {event.node_id or 'unknown'}"
+    elif et == EventType.NODE_LOOP_ITERATION:
+        return f"{event.node_id or 'unknown'} iteration {data.get('iteration', '?')}"
+    elif et == EventType.NODE_LOOP_COMPLETED:
+        return f"Node done: {event.node_id or 'unknown'}"
+    elif et == EventType.NODE_STALLED:
+        reason = data.get("reason", "")
+        node = event.node_id or "unknown"
+        return f"Node stalled: {node} - {reason}" if reason else f"Node stalled: {node}"
+    elif et == EventType.NODE_INPUT_BLOCKED:
+        return f"Node input blocked: {event.node_id or 'unknown'}"
+    elif et == EventType.GOAL_PROGRESS:
+        return f"Goal progress: {data.get('progress', '?')}"
+    elif et == EventType.GOAL_ACHIEVED:
+        return "Goal achieved"
+    elif et == EventType.CONSTRAINT_VIOLATION:
+        return f"Constraint violated: {data.get('description', 'unknown')}"
+    elif et == EventType.STATE_CHANGED:
+        return f"State changed: {data.get('key', 'unknown')}"
+    elif et == EventType.CLIENT_INPUT_REQUESTED:
+        return "Waiting for user input"
+    else:
+        return f"{et.value}: {data}"
+
+
+def format_event(event: AgentEvent) -> str:
+    """Format an AgentEvent as a Rich markup string with timestamp + symbol."""
+    ts = event.timestamp.strftime("%H:%M:%S")
+    symbol, color = EVENT_FORMAT.get(event.type, ("--", "dim"))
+    text = extract_event_text(event)
+    return f"[dim]{ts}[/dim] [{color}]{symbol} {text}[/{color}]"
+
+
+def format_python_log(record: logging.LogRecord) -> str:
+    """Format a Python log record as a Rich markup string with timestamp and severity color."""
+    ts = datetime.fromtimestamp(record.created).strftime("%H:%M:%S")
+    color = LOG_LEVEL_COLORS.get(record.levelno, "")
+    msg = record.getMessage()
+    if color:
+        return f"[dim]{ts}[/dim] [{color}]{record.levelname}[/{color}] {msg}"
+    else:
+        return f"[dim]{ts}[/dim] {record.levelname} {msg}"
+
+
+# --- LogPane widget (kept for backward compatibility) ---
+
 
 class LogPane(Container):
     """Widget to display logs with reliable rendering."""
 
-    _EVENT_FORMAT: dict[EventType, tuple[str, str]] = {
-        EventType.EXECUTION_STARTED: (">>", "bold cyan"),
-        EventType.EXECUTION_COMPLETED: ("<<", "bold green"),
-        EventType.EXECUTION_FAILED: ("!!", "bold red"),
-        EventType.TOOL_CALL_STARTED: ("->", "yellow"),
-        EventType.TOOL_CALL_COMPLETED: ("<-", "green"),
-        EventType.NODE_LOOP_STARTED: ("@@", "cyan"),
-        EventType.NODE_LOOP_ITERATION: ("..", "dim"),
-        EventType.NODE_LOOP_COMPLETED: ("@@", "dim"),
-        EventType.NODE_STALLED: ("!!", "bold yellow"),
-        EventType.NODE_INPUT_BLOCKED: ("!!", "yellow"),
-        EventType.GOAL_PROGRESS: ("%%", "blue"),
-        EventType.GOAL_ACHIEVED: ("**", "bold green"),
-        EventType.CONSTRAINT_VIOLATION: ("!!", "bold red"),
-        EventType.STATE_CHANGED: ("~~", "dim"),
-        EventType.CLIENT_INPUT_REQUESTED: ("??", "magenta"),
-    }
-
-    _LOG_LEVEL_COLORS = {
-        logging.DEBUG: "dim",
-        logging.INFO: "",
-        logging.WARNING: "yellow",
-        logging.ERROR: "red",
-        logging.CRITICAL: "bold red",
-    }
-
     DEFAULT_CSS = """
     LogPane {
         width: 100%;
@@ -58,84 +133,27 @@ class LogPane(Container):
     """
 
     def compose(self) -> ComposeResult:
-        # RichLog is designed for log display and doesn't have TextArea's rendering issues
         yield RichLog(id="main-log", highlight=True, markup=True, auto_scroll=False)
 
     def write_event(self, event: AgentEvent) -> None:
         """Format an AgentEvent with timestamp + symbol and write to the log."""
-        ts = event.timestamp.strftime("%H:%M:%S")
-        symbol, color = self._EVENT_FORMAT.get(event.type, ("--", "dim"))
-        text = self._extract_event_text(event)
-        self.write_log(f"[dim]{ts}[/dim] [{color}]{symbol} {text}[/{color}]")
-
-    def _extract_event_text(self, event: AgentEvent) -> str:
-        """Extract human-readable text from an event's data dict."""
-        et = event.type
-        data = event.data
-
-        if et == EventType.EXECUTION_STARTED:
-            return "Execution started"
-        elif et == EventType.EXECUTION_COMPLETED:
-            return "Execution completed"
-        elif et == EventType.EXECUTION_FAILED:
-            return f"Execution FAILED: {data.get('error', 'unknown')}"
-        elif et == EventType.TOOL_CALL_STARTED:
-            return f"Tool call: {data.get('tool_name', 'unknown')}"
-        elif et == EventType.TOOL_CALL_COMPLETED:
-            name = data.get("tool_name", "unknown")
-            if data.get("is_error"):
-                preview = str(data.get("result", ""))[:80]
-                return f"Tool error: {name} - {preview}"
-            return f"Tool done: {name}"
-        elif et == EventType.NODE_LOOP_STARTED:
-            return f"Node started: {event.node_id or 'unknown'}"
-        elif et == EventType.NODE_LOOP_ITERATION:
-            return f"{event.node_id or 'unknown'} iteration {data.get('iteration', '?')}"
-        elif et == EventType.NODE_LOOP_COMPLETED:
-            return f"Node done: {event.node_id or 'unknown'}"
-        elif et == EventType.NODE_STALLED:
-            reason = data.get("reason", "")
-            node = event.node_id or "unknown"
-            return f"Node stalled: {node} - {reason}" if reason else f"Node stalled: {node}"
-        elif et == EventType.NODE_INPUT_BLOCKED:
-            return f"Node input blocked: {event.node_id or 'unknown'}"
-        elif et == EventType.GOAL_PROGRESS:
-            return f"Goal progress: {data.get('progress', '?')}"
-        elif et == EventType.GOAL_ACHIEVED:
-            return "Goal achieved"
-        elif et == EventType.CONSTRAINT_VIOLATION:
-            return f"Constraint violated: {data.get('description', 'unknown')}"
-        elif et == EventType.STATE_CHANGED:
-            return f"State changed: {data.get('key', 'unknown')}"
-        elif et == EventType.CLIENT_INPUT_REQUESTED:
-            return "Waiting for user input"
-        else:
-            return f"{et.value}: {data}"
+        self.write_log(format_event(event))
 
     def write_python_log(self, record: logging.LogRecord) -> None:
         """Format a Python log record with timestamp and severity color."""
-        ts = datetime.fromtimestamp(record.created).strftime("%H:%M:%S")
-        color = self._LOG_LEVEL_COLORS.get(record.levelno, "")
-        msg = record.getMessage()
-        if color:
-            self.write_log(f"[dim]{ts}[/dim] [{color}]{record.levelname}[/{color}] {msg}")
-        else:
-            self.write_log(f"[dim]{ts}[/dim] {record.levelname} {msg}")
+        self.write_log(format_python_log(record))
 
     def write_log(self, message: str) -> None:
         """Write a log message to the log pane."""
         try:
-            # Check if widget is mounted
             if not self.is_mounted:
                 return
 
             log = self.query_one("#main-log", RichLog)
 
-            # Check if log is mounted
             if not log.is_mounted:
                 return
 
-            # Only auto-scroll if user is already at the bottom
             was_at_bottom = log.is_vertical_scroll_end
 
             log.write(message)

core/framework/tui/widgets/selectable_rich_log.py

@@ -195,12 +195,27 @@ def _copy_to_clipboard(text: str) -> None:
     try:
         if sys.platform == "darwin":
             subprocess.run(["pbcopy"], input=text.encode(), check=True, timeout=5)
-        elif sys.platform.startswith("linux"):
+        elif sys.platform == "win32":
             subprocess.run(
-                ["xclip", "-selection", "clipboard"],
-                input=text.encode(),
+                ["clip.exe"],
+                input=text.encode("utf-16le"),
                 check=True,
                 timeout=5,
             )
+        elif sys.platform.startswith("linux"):
+            try:
+                subprocess.run(
+                    ["xclip", "-selection", "clipboard"],
+                    input=text.encode(),
+                    check=True,
+                    timeout=5,
+                )
+            except (subprocess.SubprocessError, FileNotFoundError):
+                subprocess.run(
+                    ["xsel", "--clipboard", "--input"],
+                    input=text.encode(),
+                    check=True,
+                    timeout=5,
+                )
     except (subprocess.SubprocessError, FileNotFoundError):
         pass

core/pyproject.toml

@@ -20,6 +20,7 @@ dependencies = [
 
 [project.optional-dependencies]
 tui = ["textual>=0.75.0"]
+webhook = ["aiohttp>=3.9.0"]
 
 [project.scripts]
 hive = "framework.cli:main"

core/tests/test_builder.py

@@ -1,342 +0,0 @@
-"""Tests for the BuilderQuery interface - how Builder analyzes agent runs.
-
-DEPRECATED: These tests rely on the deprecated FileStorage backend.
-BuilderQuery and Runtime both use FileStorage which is deprecated.
-New code should use unified session storage instead.
-"""
-
-from pathlib import Path
-
-import pytest
-
-from framework import BuilderQuery, Runtime
-from framework.schemas.run import RunStatus
-
-# Mark all tests in this module as skipped - they rely on deprecated FileStorage
-pytestmark = pytest.mark.skip(reason="Tests rely on deprecated FileStorage backend")
-
-
-def create_successful_run(runtime: Runtime, goal_id: str = "test_goal") -> str:
-    """Helper to create a successful run with decisions."""
-    run_id = runtime.start_run(goal_id, f"Test goal: {goal_id}")
-
-    runtime.set_node("search-node")
-    d1 = runtime.decide(
-        intent="Search for data",
-        options=[
-            {"id": "web", "description": "Web search", "pros": ["Fresh data"]},
-            {"id": "cache", "description": "Use cache", "pros": ["Fast"]},
-        ],
-        chosen="web",
-        reasoning="Need fresh data",
-    )
-    runtime.record_outcome(d1, success=True, result={"items": 3}, tokens_used=50)
-
-    runtime.set_node("process-node")
-    d2 = runtime.decide(
-        intent="Process results",
-        options=[{"id": "filter", "description": "Filter and transform"}],
-        chosen="filter",
-        reasoning="Standard processing",
-    )
-    runtime.record_outcome(d2, success=True, result={"processed": 3}, tokens_used=30)
-
-    runtime.end_run(success=True, narrative="Successfully processed data")
-    return run_id
-
-
-def create_failed_run(runtime: Runtime, goal_id: str = "test_goal") -> str:
-    """Helper to create a failed run."""
-    run_id = runtime.start_run(goal_id, f"Test goal: {goal_id}")
-
-    runtime.set_node("search-node")
-    d1 = runtime.decide(
-        intent="Search for data",
-        options=[{"id": "web", "description": "Web search"}],
-        chosen="web",
-        reasoning="Need data",
-    )
-    runtime.record_outcome(d1, success=True, result={"items": 0})
-
-    runtime.set_node("process-node")
-    d2 = runtime.decide(
-        intent="Process results",
-        options=[{"id": "process", "description": "Process data"}],
-        chosen="process",
-        reasoning="Continue pipeline",
-    )
-    runtime.record_outcome(d2, success=False, error="No data to process")
-
-    runtime.report_problem(
-        severity="critical",
-        description="Processing failed due to empty input",
-        decision_id=d2,
-        suggested_fix="Add empty input handling",
-    )
-
-    runtime.end_run(success=False, narrative="Failed to process - no data")
-    return run_id
-
-
-class TestBuilderQueryBasics:
-    """Test basic query operations."""
-
-    def test_get_run_summary(self, tmp_path: Path):
-        """Test getting a run summary."""
-        runtime = Runtime(tmp_path)
-        run_id = create_successful_run(runtime)
-
-        query = BuilderQuery(tmp_path)
-        summary = query.get_run_summary(run_id)
-
-        assert summary is not None
-        assert summary.run_id == run_id
-        assert summary.status == RunStatus.COMPLETED
-        assert summary.decision_count == 2
-        assert summary.success_rate == 1.0
-
-    def test_get_full_run(self, tmp_path: Path):
-        """Test getting the full run details."""
-        runtime = Runtime(tmp_path)
-        run_id = create_successful_run(runtime)
-
-        query = BuilderQuery(tmp_path)
-        run = query.get_full_run(run_id)
-
-        assert run is not None
-        assert run.id == run_id
-        assert len(run.decisions) == 2
-        assert run.decisions[0].node_id == "search-node"
-        assert run.decisions[1].node_id == "process-node"
-
-    def test_list_runs_for_goal(self, tmp_path: Path):
-        """Test listing all runs for a goal."""
-        runtime = Runtime(tmp_path)
-        create_successful_run(runtime, "goal_a")
-        create_successful_run(runtime, "goal_a")
-        create_successful_run(runtime, "goal_b")
-
-        query = BuilderQuery(tmp_path)
-        summaries = query.list_runs_for_goal("goal_a")
-
-        assert len(summaries) == 2
-        for s in summaries:
-            assert s.goal_id == "goal_a"
-
-    def test_get_recent_failures(self, tmp_path: Path):
-        """Test getting recent failed runs."""
-        runtime = Runtime(tmp_path)
-        create_successful_run(runtime)
-        create_failed_run(runtime)
-        create_failed_run(runtime)
-
-        query = BuilderQuery(tmp_path)
-        failures = query.get_recent_failures()
-
-        assert len(failures) == 2
-        for f in failures:
-            assert f.status == RunStatus.FAILED
-
-
-class TestFailureAnalysis:
-    """Test failure analysis capabilities."""
-
-    def test_analyze_failure(self, tmp_path: Path):
-        """Test analyzing why a run failed."""
-        runtime = Runtime(tmp_path)
-        run_id = create_failed_run(runtime)
-
-        query = BuilderQuery(tmp_path)
-        analysis = query.analyze_failure(run_id)
-
-        assert analysis is not None
-        assert analysis.run_id == run_id
-        assert "No data to process" in analysis.root_cause
-        assert len(analysis.decision_chain) >= 2
-        assert len(analysis.problems) == 1
-        assert "critical" in analysis.problems[0].lower()
-
-    def test_analyze_failure_returns_none_for_success(self, tmp_path: Path):
-        """analyze_failure returns None for successful runs."""
-        runtime = Runtime(tmp_path)
-        run_id = create_successful_run(runtime)
-
-        query = BuilderQuery(tmp_path)
-        analysis = query.analyze_failure(run_id)
-
-        assert analysis is None
-
-    def test_failure_analysis_has_suggestions(self, tmp_path: Path):
-        """Failure analysis should include suggestions."""
-        runtime = Runtime(tmp_path)
-        run_id = create_failed_run(runtime)
-
-        query = BuilderQuery(tmp_path)
-        analysis = query.analyze_failure(run_id)
-
-        assert len(analysis.suggestions) > 0
-        # Should include the suggested fix from the problem
-        assert any("empty input" in s.lower() for s in analysis.suggestions)
-
-    def test_get_decision_trace(self, tmp_path: Path):
-        """Test getting a readable decision trace."""
-        runtime = Runtime(tmp_path)
-        run_id = create_successful_run(runtime)
-
-        query = BuilderQuery(tmp_path)
-        trace = query.get_decision_trace(run_id)
-
-        assert len(trace) == 2
-        assert "search-node" in trace[0]
-        assert "process-node" in trace[1]
-
-
-class TestPatternAnalysis:
-    """Test pattern detection across runs."""
-
-    def test_find_patterns_basic(self, tmp_path: Path):
-        """Test basic pattern finding."""
-        runtime = Runtime(tmp_path)
-        create_successful_run(runtime, "goal_x")
-        create_successful_run(runtime, "goal_x")
-        create_failed_run(runtime, "goal_x")
-
-        query = BuilderQuery(tmp_path)
-        patterns = query.find_patterns("goal_x")
-
-        assert patterns is not None
-        assert patterns.goal_id == "goal_x"
-        assert patterns.run_count == 3
-        assert 0 < patterns.success_rate < 1  # 2/3 success
-
-    def test_find_patterns_common_failures(self, tmp_path: Path):
-        """Test finding common failures."""
-        runtime = Runtime(tmp_path)
-        # Create multiple runs with the same failure
-        for _ in range(3):
-            create_failed_run(runtime, "failing_goal")
-
-        query = BuilderQuery(tmp_path)
-        patterns = query.find_patterns("failing_goal")
-
-        assert len(patterns.common_failures) > 0
-        # "No data to process" should be a common failure
-        failure_messages = [f[0] for f in patterns.common_failures]
-        assert any("No data to process" in msg for msg in failure_messages)
-
-    def test_find_patterns_problematic_nodes(self, tmp_path: Path):
-        """Test finding problematic nodes."""
-        runtime = Runtime(tmp_path)
-        # Create runs where process-node always fails
-        for _ in range(3):
-            create_failed_run(runtime, "node_test")
-
-        query = BuilderQuery(tmp_path)
-        patterns = query.find_patterns("node_test")
-
-        # process-node should be flagged as problematic
-        problematic_node_ids = [n[0] for n in patterns.problematic_nodes]
-        assert "process-node" in problematic_node_ids
-
-    def test_compare_runs(self, tmp_path: Path):
-        """Test comparing two runs."""
-        runtime = Runtime(tmp_path)
-        run1 = create_successful_run(runtime)
-        run2 = create_failed_run(runtime)
-
-        query = BuilderQuery(tmp_path)
-        comparison = query.compare_runs(run1, run2)
-
-        assert comparison["run_1"]["status"] == "completed"
-        assert comparison["run_2"]["status"] == "failed"
-        assert len(comparison["differences"]) > 0
-
-
-class TestImprovementSuggestions:
-    """Test improvement suggestion generation."""
-
-    def test_suggest_improvements(self, tmp_path: Path):
-        """Test generating improvement suggestions."""
-        runtime = Runtime(tmp_path)
-        # Create runs with failures to trigger suggestions
-        for _ in range(3):
-            create_failed_run(runtime, "improve_goal")
-
-        query = BuilderQuery(tmp_path)
-        suggestions = query.suggest_improvements("improve_goal")
-
-        assert len(suggestions) > 0
-        # Should suggest improving the problematic node
-        node_suggestions = [s for s in suggestions if s["type"] == "node_improvement"]
-        assert len(node_suggestions) > 0
-
-    def test_suggest_improvements_for_low_success_rate(self, tmp_path: Path):
-        """Should suggest architecture review for low success rate."""
-        runtime = Runtime(tmp_path)
-        # 4 failures, 1 success = 20% success rate
-        for _ in range(4):
-            create_failed_run(runtime, "low_success")
-        create_successful_run(runtime, "low_success")
-
-        query = BuilderQuery(tmp_path)
-        suggestions = query.suggest_improvements("low_success")
-
-        arch_suggestions = [s for s in suggestions if s["type"] == "architecture"]
-        assert len(arch_suggestions) > 0
-        assert arch_suggestions[0]["priority"] == "high"
-
-    def test_get_node_performance(self, tmp_path: Path):
-        """Test getting performance metrics for a node."""
-        runtime = Runtime(tmp_path)
-        create_successful_run(runtime)
-        create_successful_run(runtime)
-
-        query = BuilderQuery(tmp_path)
-        perf = query.get_node_performance("search-node")
-
-        assert perf["node_id"] == "search-node"
-        assert perf["total_decisions"] == 2
-        assert perf["success_rate"] == 1.0
-        assert perf["total_tokens"] == 100  # 50 tokens per run
-
-
-class TestBuilderWorkflow:
-    """Test complete Builder workflows."""
-
-    def test_builder_investigation_workflow(self, tmp_path: Path):
-        """Test a complete investigation workflow as Builder would use it."""
-        runtime = Runtime(tmp_path)
-
-        # Set up scenario: some successes, some failures
-        for _ in range(2):
-            create_successful_run(runtime, "customer_goal")
-        for _ in range(2):
-            create_failed_run(runtime, "customer_goal")
-
-        query = BuilderQuery(tmp_path)
-
-        # Step 1: Get overview of the goal
-        summaries = query.list_runs_for_goal("customer_goal")
-        assert len(summaries) == 4
-
-        # Step 2: Find patterns
-        patterns = query.find_patterns("customer_goal")
-        assert patterns.success_rate == 0.5  # 2/4
-
-        # Step 3: Get recent failures
-        failures = query.get_recent_failures()
-        assert len(failures) == 2
-
-        # Step 4: Analyze a specific failure
-        failure_id = failures[0].run_id
-        analysis = query.analyze_failure(failure_id)
-        assert analysis is not None
-        assert len(analysis.suggestions) > 0
-
-        # Step 5: Generate improvement suggestions
-        suggestions = query.suggest_improvements("customer_goal")
-        assert len(suggestions) > 0
-
-        # Step 6: Check node performance
-        perf = query.get_node_performance("process-node")
-        assert perf["success_rate"] < 1.0  # process-node fails in failed runs

core/tests/test_concurrent_storage.py

@@ -1,185 +0,0 @@
-"""Tests for ConcurrentStorage race condition and cache invalidation fixes."""
-
-import asyncio
-from pathlib import Path
-
-import pytest
-
-from framework.schemas.run import Run, RunMetrics, RunStatus
-from framework.storage.concurrent import ConcurrentStorage
-
-
-def create_test_run(
-    run_id: str, goal_id: str = "test-goal", status: RunStatus = RunStatus.RUNNING
-) -> Run:
-    """Create a minimal test Run object."""
-    return Run(
-        id=run_id,
-        goal_id=goal_id,
-        status=status,
-        narrative="Test run",
-        metrics=RunMetrics(
-            nodes_executed=[],
-        ),
-        decisions=[],
-        problems=[],
-    )
-
-
-@pytest.mark.skip(
-    reason="FileStorage.save_run() is deprecated and now a no-op. "
-    "ConcurrentStorage wraps FileStorage, so these tests no longer work. "
-    "New sessions use unified storage at sessions/{session_id}/state.json"
-)
-@pytest.mark.asyncio
-async def test_cache_invalidation_on_save(tmp_path: Path):
-    """Test that summary cache is invalidated when a run is saved.
-
-    This tests the fix for the cache invalidation bug where load_summary()
-    would return stale data after a run was updated.
-    """
-    storage = ConcurrentStorage(tmp_path)
-    await storage.start()
-
-    try:
-        run_id = "test-run-1"
-
-        # Create and save initial run
-        run = create_test_run(run_id, status=RunStatus.RUNNING)
-        await storage.save_run(run, immediate=True)
-
-        # Load summary to populate the cache
-        summary = await storage.load_summary(run_id)
-        assert summary is not None
-        assert summary.status == RunStatus.RUNNING
-
-        # Update run with new status
-        run.status = RunStatus.COMPLETED
-        await storage.save_run(run, immediate=True)
-
-        # Load summary again - should get fresh data, not cached stale data
-        summary = await storage.load_summary(run_id)
-        assert summary is not None
-        assert summary.status == RunStatus.COMPLETED, (
-            "Summary cache should be invalidated on save - got stale data"
-        )
-    finally:
-        await storage.stop()
-
-
-@pytest.mark.skip(
-    reason="FileStorage.save_run() is deprecated and now a no-op. "
-    "ConcurrentStorage wraps FileStorage, so these tests no longer work. "
-    "New sessions use unified storage at sessions/{session_id}/state.json"
-)
-@pytest.mark.asyncio
-async def test_batched_write_cache_consistency(tmp_path: Path):
-    """Test that cache is only updated after successful batched write.
-
-    This tests the fix for the race condition where cache was updated
-    before the batched write completed.
-    """
-    storage = ConcurrentStorage(tmp_path, batch_interval=0.05)
-    await storage.start()
-
-    try:
-        run_id = "test-run-2"
-
-        # Save via batching (immediate=False)
-        run = create_test_run(run_id, status=RunStatus.RUNNING)
-        await storage.save_run(run, immediate=False)
-
-        # Before batch flush, cache should NOT contain the run
-        # (This is the fix - previously cache was updated immediately)
-        cache_key = f"run:{run_id}"
-        assert cache_key not in storage._cache, (
-            "Cache should not be updated before batch is flushed"
-        )
-
-        # Wait for batch to flush (poll instead of fixed sleep for CI reliability)
-        for _ in range(500):  # 500 * 0.01s = 5s max
-            if cache_key in storage._cache:
-                break
-            await asyncio.sleep(0.01)
-
-        # After batch flush, cache should contain the run
-        assert cache_key in storage._cache, "Cache should be updated after batch flush"
-
-        # Verify data on disk matches cache
-        loaded_run = await storage.load_run(run_id, use_cache=False)
-        assert loaded_run is not None
-        assert loaded_run.id == run_id
-        assert loaded_run.status == RunStatus.RUNNING
-    finally:
-        await storage.stop()
-
-
-@pytest.mark.skip(
-    reason="FileStorage.save_run() is deprecated and now a no-op. "
-    "ConcurrentStorage wraps FileStorage, so these tests no longer work. "
-    "New sessions use unified storage at sessions/{session_id}/state.json"
-)
-@pytest.mark.asyncio
-async def test_immediate_write_updates_cache(tmp_path: Path):
-    """Test that immediate writes still update cache correctly."""
-    storage = ConcurrentStorage(tmp_path)
-    await storage.start()
-
-    try:
-        run_id = "test-run-3"
-
-        # Save with immediate=True
-        run = create_test_run(run_id, status=RunStatus.COMPLETED)
-        await storage.save_run(run, immediate=True)
-
-        # Cache should be updated immediately for immediate writes
-        cache_key = f"run:{run_id}"
-        assert cache_key in storage._cache, "Cache should be updated after immediate write"
-
-        # Verify cached value is correct
-        cached_run = storage._cache[cache_key].value
-        assert cached_run.id == run_id
-        assert cached_run.status == RunStatus.COMPLETED
-    finally:
-        await storage.stop()
-
-
-@pytest.mark.skip(
-    reason="FileStorage.save_run() is deprecated and now a no-op. "
-    "ConcurrentStorage wraps FileStorage, so these tests no longer work. "
-    "New sessions use unified storage at sessions/{session_id}/state.json"
-)
-@pytest.mark.asyncio
-async def test_summary_cache_invalidated_on_multiple_saves(tmp_path: Path):
-    """Test that summary cache is invalidated on each save, not just the first."""
-    storage = ConcurrentStorage(tmp_path)
-    await storage.start()
-
-    try:
-        run_id = "test-run-4"
-
-        # First save
-        run = create_test_run(run_id, status=RunStatus.RUNNING)
-        await storage.save_run(run, immediate=True)
-
-        # Load summary to cache it
-        summary1 = await storage.load_summary(run_id)
-        assert summary1.status == RunStatus.RUNNING
-
-        # Second save with new status
-        run.status = RunStatus.RUNNING
-        await storage.save_run(run, immediate=True)
-
-        # Load summary - should be fresh
-        summary2 = await storage.load_summary(run_id)
-        assert summary2.status == RunStatus.RUNNING
-
-        # Third save with final status
-        run.status = RunStatus.COMPLETED
-        await storage.save_run(run, immediate=True)
-
-        # Load summary - should be fresh again
-        summary3 = await storage.load_summary(run_id)
-        assert summary3.status == RunStatus.COMPLETED
-    finally:
-        await storage.stop()

core/tests/test_continuous_conversation.py

@@ -0,0 +1,538 @@
+"""Tests for the Continuous Agent architecture (conversation threading + cumulative tools).
+
+Validates:
+  - conversation_mode="isolated" preserves existing behavior
+  - conversation_mode="continuous" threads one conversation across nodes
+  - Transition markers are inserted at phase boundaries
+  - System prompt updates at each transition (layered prompt composition)
+  - Tools accumulate across nodes in continuous mode
+  - prompt_composer functions work correctly
+"""
+
+from __future__ import annotations
+
+from collections.abc import AsyncIterator
+from typing import Any
+from unittest.mock import MagicMock
+
+import pytest
+
+from framework.graph.conversation import NodeConversation
+from framework.graph.edge import EdgeCondition, EdgeSpec, GraphSpec
+from framework.graph.executor import GraphExecutor
+from framework.graph.goal import Goal
+from framework.graph.node import NodeResult, NodeSpec, SharedMemory
+from framework.graph.prompt_composer import (
+    build_narrative,
+    build_transition_marker,
+    compose_system_prompt,
+)
+from framework.llm.provider import LLMProvider, LLMResponse, Tool
+from framework.llm.stream_events import FinishEvent, TextDeltaEvent, ToolCallEvent
+from framework.runtime.core import Runtime
+
+# ---------------------------------------------------------------------------
+# Mock LLM
+# ---------------------------------------------------------------------------
+
+
+class MockStreamingLLM(LLMProvider):
+    """Mock LLM that yields pre-programmed StreamEvent sequences."""
+
+    def __init__(self, scenarios: list[list] | None = None):
+        self.scenarios = scenarios or []
+        self._call_index = 0
+        self.stream_calls: list[dict] = []
+
+    async def stream(
+        self,
+        messages: list[dict[str, Any]],
+        system: str = "",
+        tools: list[Tool] | None = None,
+        max_tokens: int = 4096,
+    ) -> AsyncIterator:
+        self.stream_calls.append({"messages": messages, "system": system, "tools": tools})
+        if not self.scenarios:
+            return
+        events = self.scenarios[self._call_index % len(self.scenarios)]
+        self._call_index += 1
+        for event in events:
+            yield event
+
+    def complete(self, messages, system="", **kwargs) -> LLMResponse:
+        return LLMResponse(content="Summary.", model="mock", stop_reason="stop")
+
+    def complete_with_tools(self, messages, system, tools, tool_executor, **kwargs) -> LLMResponse:
+        return LLMResponse(content="", model="mock", stop_reason="stop")
+
+
+# ---------------------------------------------------------------------------
+# Helpers
+# ---------------------------------------------------------------------------
+
+
+def _set_output_scenario(key: str, value: str) -> list:
+    """LLM calls set_output then finishes."""
+    return [
+        ToolCallEvent(
+            tool_use_id=f"call_{key}",
+            tool_name="set_output",
+            tool_input={"key": key, "value": value},
+        ),
+        FinishEvent(stop_reason="tool_calls", input_tokens=10, output_tokens=5, model="mock"),
+    ]
+
+
+def _text_then_set_output(text: str, key: str, value: str) -> list:
+    """LLM produces text, then calls set_output, then finishes (2 turns needed)."""
+    return [
+        TextDeltaEvent(content=text, snapshot=text),
+        ToolCallEvent(
+            tool_use_id=f"call_{key}",
+            tool_name="set_output",
+            tool_input={"key": key, "value": value},
+        ),
+        FinishEvent(stop_reason="tool_calls", input_tokens=10, output_tokens=5, model="mock"),
+    ]
+
+
+def _text_finish(text: str) -> list:
+    """LLM produces text and stops (triggers judge)."""
+    return [
+        TextDeltaEvent(content=text, snapshot=text),
+        FinishEvent(stop_reason="stop", input_tokens=10, output_tokens=5, model="mock"),
+    ]
+
+
+def _make_runtime():
+    rt = MagicMock(spec=Runtime)
+    rt.start_run = MagicMock(return_value="run_1")
+    rt.end_run = MagicMock()
+    rt.report_problem = MagicMock()
+    rt.decide = MagicMock(return_value="dec_1")
+    rt.record_outcome = MagicMock()
+    rt.set_node = MagicMock()
+    return rt
+
+
+def _make_goal():
+    return Goal(id="g1", name="test", description="test goal")
+
+
+def _make_tool(name: str) -> Tool:
+    return Tool(
+        name=name,
+        description=f"Tool {name}",
+        parameters={"type": "object", "properties": {}},
+    )
+
+
+# ===========================================================================
+# prompt_composer unit tests
+# ===========================================================================
+
+
+class TestComposeSystemPrompt:
+    def test_all_layers(self):
+        result = compose_system_prompt(
+            identity_prompt="I am a research agent.",
+            focus_prompt="Focus on writing the report.",
+            narrative="We found 5 sources on topic X.",
+        )
+        assert "I am a research agent." in result
+        assert "Focus on writing the report." in result
+        assert "We found 5 sources on topic X." in result
+        # Identity comes first
+        assert result.index("I am a research agent.") < result.index("Focus on writing")
+
+    def test_identity_only(self):
+        result = compose_system_prompt(identity_prompt="I am an agent.", focus_prompt=None)
+        assert result == "I am an agent."
+
+    def test_focus_only(self):
+        result = compose_system_prompt(identity_prompt=None, focus_prompt="Do the thing.")
+        assert "Current Focus" in result
+        assert "Do the thing." in result
+
+    def test_empty(self):
+        result = compose_system_prompt(identity_prompt=None, focus_prompt=None)
+        assert result == ""
+
+
+class TestBuildNarrative:
+    def test_with_execution_path(self):
+        memory = SharedMemory()
+        memory.write("findings", "some findings")
+
+        node_a = NodeSpec(
+            id="a", name="Research", description="Research the topic", node_type="event_loop"
+        )
+        node_b = NodeSpec(id="b", name="Report", description="Write report", node_type="event_loop")
+        graph = GraphSpec(
+            id="g1",
+            goal_id="g1",
+            entry_node="a",
+            nodes=[node_a, node_b],
+            edges=[],
+        )
+
+        result = build_narrative(memory, ["a"], graph)
+        assert "Research" in result
+        assert "findings" in result
+
+    def test_empty_state(self):
+        memory = SharedMemory()
+        graph = GraphSpec(id="g1", goal_id="g1", entry_node="a", nodes=[], edges=[])
+        result = build_narrative(memory, [], graph)
+        assert result == ""
+
+
+class TestBuildTransitionMarker:
+    def test_basic_marker(self):
+        prev = NodeSpec(
+            id="research", name="Research", description="Find sources", node_type="event_loop"
+        )
+        next_n = NodeSpec(
+            id="report", name="Report", description="Write report", node_type="event_loop"
+        )
+        memory = SharedMemory()
+        memory.write("findings", "important stuff")
+
+        marker = build_transition_marker(
+            previous_node=prev,
+            next_node=next_n,
+            memory=memory,
+            cumulative_tool_names=["web_search", "save_data"],
+        )
+
+        assert "PHASE TRANSITION" in marker
+        assert "Research" in marker
+        assert "Report" in marker
+        assert "findings" in marker
+        assert "web_search" in marker
+        assert "reflect" in marker.lower()
+
+
+# ===========================================================================
+# NodeConversation.update_system_prompt
+# ===========================================================================
+
+
+class TestUpdateSystemPrompt:
+    def test_update(self):
+        conv = NodeConversation(system_prompt="original")
+        assert conv.system_prompt == "original"
+        conv.update_system_prompt("updated")
+        assert conv.system_prompt == "updated"
+
+
+# ===========================================================================
+# Conversation threading through executor
+# ===========================================================================
+
+
+class TestContinuousConversation:
+    """Test that conversation_mode='continuous' threads a single conversation."""
+
+    @pytest.mark.asyncio
+    async def test_isolated_mode_no_conversation_in_result(self):
+        """In isolated mode, NodeResult.conversation should be None."""
+        runtime = _make_runtime()
+        llm = MockStreamingLLM(
+            scenarios=[
+                _set_output_scenario("result", "done"),
+                _text_finish("accepted"),
+            ]
+        )
+
+        spec = NodeSpec(
+            id="n1",
+            name="Node1",
+            description="test",
+            node_type="event_loop",
+            output_keys=["result"],
+        )
+        graph = GraphSpec(
+            id="g1",
+            goal_id="g1",
+            entry_node="n1",
+            nodes=[spec],
+            edges=[],
+            conversation_mode="isolated",
+        )
+
+        executor = GraphExecutor(runtime=runtime, llm=llm)
+        result = await executor.execute(graph=graph, goal=_make_goal())
+        assert result.success
+
+    @pytest.mark.asyncio
+    async def test_continuous_threads_conversation(self):
+        """In continuous mode, second node sees messages from first node."""
+        runtime = _make_runtime()
+
+        # Node A: set_output("brief", "the brief"), then finish (accept)
+        # Node B: set_output("report", "the report"), then finish (accept)
+        llm = MockStreamingLLM(
+            scenarios=[
+                _text_then_set_output("I'll research this.", "brief", "the brief"),
+                _text_finish(""),  # triggers accept for node A (all keys set)
+                _text_then_set_output("Here's the report.", "report", "the report"),
+                _text_finish(""),  # triggers accept for node B
+            ]
+        )
+
+        node_a = NodeSpec(
+            id="a",
+            name="Intake",
+            description="Gather requirements",
+            node_type="event_loop",
+            output_keys=["brief"],
+        )
+        node_b = NodeSpec(
+            id="b",
+            name="Report",
+            description="Write report",
+            node_type="event_loop",
+            input_keys=["brief"],
+            output_keys=["report"],
+        )
+
+        graph = GraphSpec(
+            id="g1",
+            goal_id="g1",
+            entry_node="a",
+            nodes=[node_a, node_b],
+            edges=[EdgeSpec(id="e1", source="a", target="b", condition=EdgeCondition.ON_SUCCESS)],
+            terminal_nodes=["b"],
+            conversation_mode="continuous",
+            identity_prompt="You are a thorough research agent.",
+        )
+
+        executor = GraphExecutor(runtime=runtime, llm=llm)
+        result = await executor.execute(graph=graph, goal=_make_goal())
+
+        assert result.success
+        assert result.path == ["a", "b"]
+
+        # Verify the LLM saw the identity prompt in system messages
+        # The second node's system prompt should contain the identity
+        if len(llm.stream_calls) >= 3:
+            system_at_node_b = llm.stream_calls[2]["system"]
+            assert "thorough research agent" in system_at_node_b
+
+    @pytest.mark.asyncio
+    async def test_continuous_transition_marker_present(self):
+        """Transition marker should appear in messages when switching nodes."""
+        runtime = _make_runtime()
+
+        llm = MockStreamingLLM(
+            scenarios=[
+                _text_then_set_output("Research done.", "brief", "the brief"),
+                _text_finish(""),
+                _text_then_set_output("Report done.", "report", "the report"),
+                _text_finish(""),
+            ]
+        )
+
+        node_a = NodeSpec(
+            id="a",
+            name="Research",
+            description="Do research",
+            node_type="event_loop",
+            output_keys=["brief"],
+        )
+        node_b = NodeSpec(
+            id="b",
+            name="Report",
+            description="Write report",
+            node_type="event_loop",
+            input_keys=["brief"],
+            output_keys=["report"],
+        )
+
+        graph = GraphSpec(
+            id="g1",
+            goal_id="g1",
+            entry_node="a",
+            nodes=[node_a, node_b],
+            edges=[EdgeSpec(id="e1", source="a", target="b", condition=EdgeCondition.ON_SUCCESS)],
+            terminal_nodes=["b"],
+            conversation_mode="continuous",
+        )
+
+        executor = GraphExecutor(runtime=runtime, llm=llm)
+        result = await executor.execute(graph=graph, goal=_make_goal())
+        assert result.success
+
+        # When node B's first LLM call happens, its messages should contain
+        # the transition marker from the executor
+        if len(llm.stream_calls) >= 3:
+            node_b_messages = llm.stream_calls[2]["messages"]
+            all_content = " ".join(
+                m.get("content", "") for m in node_b_messages if isinstance(m.get("content"), str)
+            )
+            assert "PHASE TRANSITION" in all_content
+
+
+# ===========================================================================
+# Cumulative tools
+# ===========================================================================
+
+
+class TestCumulativeTools:
+    """Test that tools accumulate in continuous mode."""
+
+    @pytest.mark.asyncio
+    async def test_isolated_mode_tools_scoped(self):
+        """In isolated mode, each node only gets its own declared tools."""
+        runtime = _make_runtime()
+        tool_a = _make_tool("web_search")
+        tool_b = _make_tool("save_data")
+
+        llm = MockStreamingLLM(
+            scenarios=[
+                _text_then_set_output("Done.", "brief", "brief"),
+                _text_finish(""),
+                _text_then_set_output("Done.", "report", "report"),
+                _text_finish(""),
+            ]
+        )
+
+        node_a = NodeSpec(
+            id="a",
+            name="Research",
+            description="Research",
+            node_type="event_loop",
+            output_keys=["brief"],
+            tools=["web_search"],
+        )
+        node_b = NodeSpec(
+            id="b",
+            name="Report",
+            description="Report",
+            node_type="event_loop",
+            input_keys=["brief"],
+            output_keys=["report"],
+            tools=["save_data"],
+        )
+
+        graph = GraphSpec(
+            id="g1",
+            goal_id="g1",
+            entry_node="a",
+            nodes=[node_a, node_b],
+            edges=[EdgeSpec(id="e1", source="a", target="b", condition=EdgeCondition.ON_SUCCESS)],
+            terminal_nodes=["b"],
+            conversation_mode="isolated",
+        )
+
+        executor = GraphExecutor(
+            runtime=runtime,
+            llm=llm,
+            tools=[tool_a, tool_b],
+        )
+        result = await executor.execute(graph=graph, goal=_make_goal())
+        assert result.success
+
+        # In isolated mode, node B should NOT have web_search
+        if len(llm.stream_calls) >= 3:
+            node_b_tools = llm.stream_calls[2].get("tools") or []
+            tool_names = [t.name for t in node_b_tools]
+            assert "save_data" in tool_names or "set_output" in tool_names
+            # web_search should NOT be present (only set_output + save_data)
+            real_tools = [n for n in tool_names if n != "set_output"]
+            assert "web_search" not in real_tools
+
+    @pytest.mark.asyncio
+    async def test_continuous_mode_tools_accumulate(self):
+        """In continuous mode, node B should have both web_search and save_data."""
+        runtime = _make_runtime()
+        tool_a = _make_tool("web_search")
+        tool_b = _make_tool("save_data")
+
+        llm = MockStreamingLLM(
+            scenarios=[
+                _text_then_set_output("Done.", "brief", "brief"),
+                _text_finish(""),
+                _text_then_set_output("Done.", "report", "report"),
+                _text_finish(""),
+            ]
+        )
+
+        node_a = NodeSpec(
+            id="a",
+            name="Research",
+            description="Research",
+            node_type="event_loop",
+            output_keys=["brief"],
+            tools=["web_search"],
+        )
+        node_b = NodeSpec(
+            id="b",
+            name="Report",
+            description="Report",
+            node_type="event_loop",
+            input_keys=["brief"],
+            output_keys=["report"],
+            tools=["save_data"],
+        )
+
+        graph = GraphSpec(
+            id="g1",
+            goal_id="g1",
+            entry_node="a",
+            nodes=[node_a, node_b],
+            edges=[EdgeSpec(id="e1", source="a", target="b", condition=EdgeCondition.ON_SUCCESS)],
+            terminal_nodes=["b"],
+            conversation_mode="continuous",
+        )
+
+        executor = GraphExecutor(
+            runtime=runtime,
+            llm=llm,
+            tools=[tool_a, tool_b],
+        )
+        result = await executor.execute(graph=graph, goal=_make_goal())
+        assert result.success
+
+        # In continuous mode, node B should have BOTH tools
+        if len(llm.stream_calls) >= 3:
+            node_b_tools = llm.stream_calls[2].get("tools") or []
+            tool_names = [t.name for t in node_b_tools]
+            real_tools = [n for n in tool_names if n != "set_output"]
+            assert "web_search" in real_tools
+            assert "save_data" in real_tools
+
+
+# ===========================================================================
+# Schema field defaults
+# ===========================================================================
+
+
+class TestSchemaDefaults:
+    def test_graphspec_defaults(self):
+        """New fields should have safe defaults."""
+        graph = GraphSpec(
+            id="g1",
+            goal_id="g1",
+            entry_node="n1",
+            nodes=[],
+            edges=[],
+        )
+        assert graph.conversation_mode == "continuous"
+        assert graph.identity_prompt is None
+
+    def test_nodespec_defaults(self):
+        """NodeSpec.success_criteria should default to None."""
+        spec = NodeSpec(
+            id="n1",
+            name="test",
+            description="test",
+            node_type="event_loop",
+        )
+        assert spec.success_criteria is None
+
+    def test_noderesult_defaults(self):
+        """NodeResult.conversation should default to None."""
+        result = NodeResult(success=True)
+        assert result.conversation is None

core/tests/test_conversation_judge.py

@@ -0,0 +1,380 @@
+"""Tests for Level 2 conversation-aware judge.
+
+Validates:
+  - No success_criteria → Level 0 only (existing behavior)
+  - success_criteria set, good conversation → Level 2 ACCEPT
+  - success_criteria set, poor conversation → Level 2 RETRY with feedback
+  - Custom explicit judge takes priority over Level 2
+  - Level 2 fires only when Level 0 passes (all keys set)
+  - _parse_verdict correctly parses LLM responses
+"""
+
+from __future__ import annotations
+
+from collections.abc import AsyncIterator
+from typing import Any
+from unittest.mock import MagicMock
+
+import pytest
+
+from framework.graph.conversation import NodeConversation
+from framework.graph.conversation_judge import (
+    _parse_verdict,
+    evaluate_phase_completion,
+)
+from framework.graph.edge import GraphSpec
+from framework.graph.executor import GraphExecutor
+from framework.graph.goal import Goal
+from framework.graph.node import NodeSpec
+from framework.llm.provider import LLMProvider, LLMResponse, Tool
+from framework.llm.stream_events import FinishEvent, TextDeltaEvent, ToolCallEvent
+from framework.runtime.core import Runtime
+
+# ---------------------------------------------------------------------------
+# Mock LLM
+# ---------------------------------------------------------------------------
+
+
+class MockStreamingLLM(LLMProvider):
+    """Mock LLM that yields pre-programmed StreamEvent sequences."""
+
+    def __init__(self, scenarios: list[list] | None = None, complete_response: str = ""):
+        self.scenarios = scenarios or []
+        self._call_index = 0
+        self.stream_calls: list[dict] = []
+        self.complete_response = complete_response
+        self.complete_calls: list[dict] = []
+
+    async def stream(
+        self,
+        messages: list[dict[str, Any]],
+        system: str = "",
+        tools: list[Tool] | None = None,
+        max_tokens: int = 4096,
+    ) -> AsyncIterator:
+        self.stream_calls.append({"messages": messages, "system": system, "tools": tools})
+        if not self.scenarios:
+            return
+        events = self.scenarios[self._call_index % len(self.scenarios)]
+        self._call_index += 1
+        for event in events:
+            yield event
+
+    def complete(self, messages, system="", **kwargs) -> LLMResponse:
+        self.complete_calls.append({"messages": messages, "system": system})
+        return LLMResponse(content=self.complete_response, model="mock", stop_reason="stop")
+
+    def complete_with_tools(self, messages, system, tools, tool_executor, **kwargs) -> LLMResponse:
+        return LLMResponse(content="", model="mock", stop_reason="stop")
+
+
+# ---------------------------------------------------------------------------
+# Helpers
+# ---------------------------------------------------------------------------
+
+
+def _set_output_scenario(key: str, value: str) -> list:
+    return [
+        ToolCallEvent(
+            tool_use_id=f"call_{key}",
+            tool_name="set_output",
+            tool_input={"key": key, "value": value},
+        ),
+        FinishEvent(stop_reason="tool_calls", input_tokens=10, output_tokens=5, model="mock"),
+    ]
+
+
+def _text_then_set_output(text: str, key: str, value: str) -> list:
+    return [
+        TextDeltaEvent(content=text, snapshot=text),
+        ToolCallEvent(
+            tool_use_id=f"call_{key}",
+            tool_name="set_output",
+            tool_input={"key": key, "value": value},
+        ),
+        FinishEvent(stop_reason="tool_calls", input_tokens=10, output_tokens=5, model="mock"),
+    ]
+
+
+def _text_finish(text: str) -> list:
+    return [
+        TextDeltaEvent(content=text, snapshot=text),
+        FinishEvent(stop_reason="stop", input_tokens=10, output_tokens=5, model="mock"),
+    ]
+
+
+def _make_runtime():
+    rt = MagicMock(spec=Runtime)
+    rt.start_run = MagicMock(return_value="run_1")
+    rt.end_run = MagicMock()
+    rt.report_problem = MagicMock()
+    rt.decide = MagicMock(return_value="dec_1")
+    rt.record_outcome = MagicMock()
+    rt.set_node = MagicMock()
+    return rt
+
+
+def _make_goal():
+    return Goal(id="g1", name="test", description="test goal")
+
+
+# ===========================================================================
+# Unit tests for _parse_verdict
+# ===========================================================================
+
+
+class TestParseVerdict:
+    def test_accept(self):
+        v = _parse_verdict("ACTION: ACCEPT\nCONFIDENCE: 0.9\nFEEDBACK:")
+        assert v.action == "ACCEPT"
+        assert v.confidence == 0.9
+        assert v.feedback == ""
+
+    def test_retry_with_feedback(self):
+        v = _parse_verdict("ACTION: RETRY\nCONFIDENCE: 0.6\nFEEDBACK: Research is too shallow.")
+        assert v.action == "RETRY"
+        assert v.confidence == 0.6
+        assert "shallow" in v.feedback
+
+    def test_defaults_on_garbage(self):
+        v = _parse_verdict("some random text\nno structured output")
+        assert v.action == "ACCEPT"  # default
+        assert v.confidence == 0.8  # default
+
+    def test_invalid_action_defaults_to_accept(self):
+        v = _parse_verdict("ACTION: ESCALATE\nCONFIDENCE: 0.5")
+        assert v.action == "ACCEPT"  # ESCALATE not valid for Level 2
+
+
+# ===========================================================================
+# Unit tests for evaluate_phase_completion
+# ===========================================================================
+
+
+class TestEvaluatePhaseCompletion:
+    @pytest.mark.asyncio
+    async def test_accept_on_good_response(self):
+        """LLM says ACCEPT → verdict is ACCEPT."""
+        llm = MockStreamingLLM(complete_response="ACTION: ACCEPT\nCONFIDENCE: 0.95\nFEEDBACK:")
+        conv = NodeConversation(system_prompt="test")
+        await conv.add_user_message("Do research on topic X")
+        await conv.add_assistant_message("I found 5 high-quality sources on X.")
+
+        verdict = await evaluate_phase_completion(
+            llm=llm,
+            conversation=conv,
+            phase_name="Research",
+            phase_description="Research the topic",
+            success_criteria="Find at least 3 credible sources",
+            accumulator_state={"findings": "5 sources found"},
+        )
+        assert verdict.action == "ACCEPT"
+        assert verdict.confidence == 0.95
+
+    @pytest.mark.asyncio
+    async def test_retry_on_poor_response(self):
+        """LLM says RETRY → verdict is RETRY with feedback."""
+        llm = MockStreamingLLM(
+            complete_response=(
+                "ACTION: RETRY\nCONFIDENCE: 0.4\nFEEDBACK: Only found 1 source, need 3."
+            )
+        )
+        conv = NodeConversation(system_prompt="test")
+        await conv.add_user_message("Do research")
+        await conv.add_assistant_message("I found 1 source.")
+
+        verdict = await evaluate_phase_completion(
+            llm=llm,
+            conversation=conv,
+            phase_name="Research",
+            phase_description="Research the topic",
+            success_criteria="Find at least 3 credible sources",
+            accumulator_state={"findings": "1 source"},
+        )
+        assert verdict.action == "RETRY"
+        assert "1 source" in verdict.feedback
+
+    @pytest.mark.asyncio
+    async def test_llm_failure_defaults_to_accept(self):
+        """When LLM fails, Level 2 should not block (Level 0 already passed)."""
+        llm = MockStreamingLLM()
+        # Make complete() raise an exception
+        llm.complete = MagicMock(side_effect=RuntimeError("LLM unavailable"))
+
+        conv = NodeConversation(system_prompt="test")
+        await conv.add_assistant_message("Done.")
+
+        verdict = await evaluate_phase_completion(
+            llm=llm,
+            conversation=conv,
+            phase_name="Test",
+            phase_description="Test phase",
+            success_criteria="Do the thing",
+            accumulator_state={"result": "done"},
+        )
+        assert verdict.action == "ACCEPT"
+        assert verdict.confidence == 0.5
+
+
+# ===========================================================================
+# Integration: Level 2 in EventLoopNode implicit judge
+# ===========================================================================
+
+
+class TestLevel2InImplicitJudge:
+    @pytest.mark.asyncio
+    async def test_no_success_criteria_level0_only(self):
+        """Without success_criteria, Level 0 accepts normally (existing behavior)."""
+        runtime = _make_runtime()
+        llm = MockStreamingLLM(
+            scenarios=[
+                _set_output_scenario("result", "done"),
+                _text_finish("accepted"),
+            ]
+        )
+
+        spec = NodeSpec(
+            id="n1",
+            name="Node1",
+            description="test",
+            node_type="event_loop",
+            output_keys=["result"],
+            # No success_criteria!
+        )
+        graph = GraphSpec(
+            id="g1",
+            goal_id="g1",
+            entry_node="n1",
+            nodes=[spec],
+            edges=[],
+        )
+
+        executor = GraphExecutor(runtime=runtime, llm=llm)
+        result = await executor.execute(graph=graph, goal=_make_goal())
+        assert result.success
+        # LLM.complete should NOT have been called for Level 2
+        assert len(llm.complete_calls) == 0
+
+    @pytest.mark.asyncio
+    async def test_success_criteria_accept(self):
+        """With success_criteria and good work, Level 2 accepts."""
+        runtime = _make_runtime()
+        llm = MockStreamingLLM(
+            scenarios=[
+                _text_then_set_output("I did thorough research.", "result", "done"),
+                _text_finish(""),  # triggers judge
+            ],
+            complete_response="ACTION: ACCEPT\nCONFIDENCE: 0.9\nFEEDBACK:",
+        )
+
+        spec = NodeSpec(
+            id="n1",
+            name="Research",
+            description="Do research",
+            node_type="event_loop",
+            output_keys=["result"],
+            success_criteria="Provide thorough research with multiple sources.",
+        )
+        graph = GraphSpec(
+            id="g1",
+            goal_id="g1",
+            entry_node="n1",
+            nodes=[spec],
+            edges=[],
+        )
+
+        executor = GraphExecutor(runtime=runtime, llm=llm)
+        result = await executor.execute(graph=graph, goal=_make_goal())
+        assert result.success
+        # LLM.complete should have been called for Level 2
+        assert len(llm.complete_calls) >= 1
+
+    @pytest.mark.asyncio
+    async def test_success_criteria_retry_then_accept(self):
+        """Level 2 rejects first attempt, LLM tries again, Level 2 accepts."""
+        runtime = _make_runtime()
+
+        # Track complete calls to alternate responses
+        complete_responses = [
+            "ACTION: RETRY\nCONFIDENCE: 0.4\nFEEDBACK: Need more detail.",
+            "ACTION: ACCEPT\nCONFIDENCE: 0.9\nFEEDBACK:",
+        ]
+        call_count = [0]
+
+        class SequentialLLM(MockStreamingLLM):
+            def complete(self, messages, system="", **kwargs):
+                idx = call_count[0]
+                call_count[0] += 1
+                resp = complete_responses[idx % len(complete_responses)]
+                return LLMResponse(content=resp, model="mock", stop_reason="stop")
+
+        llm = SequentialLLM(
+            scenarios=[
+                # Turn 1: set output, then stop → Level 2 RETRY
+                _text_then_set_output("Brief research.", "result", "brief"),
+                _text_finish(""),  # triggers judge → Level 2 RETRY
+                # Turn 2: after retry feedback, set output again, stop → Level 2 ACCEPT
+                _text_then_set_output("Much more detailed research.", "result", "detailed"),
+                _text_finish(""),  # triggers judge → Level 2 ACCEPT
+            ]
+        )
+
+        spec = NodeSpec(
+            id="n1",
+            name="Research",
+            description="Do research",
+            node_type="event_loop",
+            output_keys=["result"],
+            success_criteria="Provide thorough research with multiple sources.",
+        )
+        graph = GraphSpec(
+            id="g1",
+            goal_id="g1",
+            entry_node="n1",
+            nodes=[spec],
+            edges=[],
+        )
+
+        executor = GraphExecutor(runtime=runtime, llm=llm)
+        result = await executor.execute(graph=graph, goal=_make_goal())
+        assert result.success
+        # Should have had 2 complete calls (first RETRY, second ACCEPT)
+        assert call_count[0] >= 2
+
+    @pytest.mark.asyncio
+    async def test_level2_only_fires_when_level0_passes(self):
+        """Level 2 should NOT fire when output keys are missing."""
+        runtime = _make_runtime()
+
+        llm = MockStreamingLLM(
+            scenarios=[
+                # Turn 1: just text, no set_output → Level 0 RETRY (missing keys)
+                _text_finish("I did some thinking."),
+                # Turn 2: set output → Level 0 ACCEPT, Level 2 check
+                _text_then_set_output("Now I have output.", "result", "done"),
+                _text_finish(""),  # triggers judge
+            ],
+            complete_response="ACTION: ACCEPT\nCONFIDENCE: 0.9\nFEEDBACK:",
+        )
+
+        spec = NodeSpec(
+            id="n1",
+            name="Research",
+            description="Do research",
+            node_type="event_loop",
+            output_keys=["result"],
+            success_criteria="Provide results.",
+        )
+        graph = GraphSpec(
+            id="g1",
+            goal_id="g1",
+            entry_node="n1",
+            nodes=[spec],
+            edges=[],
+        )
+
+        executor = GraphExecutor(runtime=runtime, llm=llm)
+        result = await executor.execute(graph=graph, goal=_make_goal())
+        assert result.success
+        # Level 2 should only fire once (when Level 0 passes)
+        assert len(llm.complete_calls) == 1

core/tests/test_event_loop_integration.py

@@ -929,6 +929,7 @@ async def test_context_handoff_between_nodes(runtime):
 
 
 @pytest.mark.asyncio
+@pytest.mark.skip(reason="Hangs in non-interactive shells (client-facing blocks on stdin)")
 async def test_client_facing_node_streams_output():
     """Client-facing node emits CLIENT_OUTPUT_DELTA events."""
     recorded: list[AgentEvent] = []
@@ -951,7 +952,7 @@ async def test_client_facing_node_streams_output():
         config=LoopConfig(max_iterations=5),
     )
 
-    # Text-only on client_facing no longer blocks (no ask_user called),
+    # Text-only on client_facing does not block (no ask_user called),
     # so the node completes without needing a shutdown workaround.
     result = await node.execute(ctx)
 

core/tests/test_event_loop_node.py

@@ -425,6 +425,7 @@ class TestEventBusLifecycle:
         assert EventType.NODE_LOOP_COMPLETED in received_events
 
     @pytest.mark.asyncio
+    @pytest.mark.skip(reason="Hangs in non-interactive shells (client-facing blocks on stdin)")
     async def test_client_facing_uses_client_output_delta(self, runtime, memory):
         """client_facing=True should emit CLIENT_OUTPUT_DELTA instead of LLM_TEXT_DELTA."""
         spec = NodeSpec(
@@ -475,6 +476,7 @@ class TestClientFacingBlocking:
         )
 
     @pytest.mark.asyncio
+    @pytest.mark.skip(reason="Hangs in non-interactive shells (client-facing blocks on stdin)")
     async def test_text_only_no_blocking(self, runtime, memory, client_spec):
         """client_facing + text-only (no ask_user) should NOT block."""
         llm = MockStreamingLLM(
@@ -630,6 +632,7 @@ class TestClientFacingBlocking:
         assert received[0].type == EventType.CLIENT_INPUT_REQUESTED
 
     @pytest.mark.asyncio
+    @pytest.mark.skip(reason="Hangs in non-interactive shells (client-facing blocks on stdin)")
     async def test_ask_user_with_real_tools(self, runtime, memory):
         """ask_user alongside real tool calls still triggers blocking."""
         spec = NodeSpec(

core/tests/test_graph_executor.py

@@ -143,6 +143,18 @@ class FakeEventBus:
     async def emit_node_loop_completed(self, **kwargs):
         self.events.append(("completed", kwargs))
 
+    async def emit_edge_traversed(self, **kwargs):
+        self.events.append(("edge_traversed", kwargs))
+
+    async def emit_execution_paused(self, **kwargs):
+        self.events.append(("execution_paused", kwargs))
+
+    async def emit_execution_resumed(self, **kwargs):
+        self.events.append(("execution_resumed", kwargs))
+
+    async def emit_node_retry(self, **kwargs):
+        self.events.append(("node_retry", kwargs))
+
 
 @pytest.mark.asyncio
 async def test_executor_emits_node_events():
@@ -201,15 +213,19 @@ async def test_executor_emits_node_events():
     assert result.success is True
     assert result.path == ["n1", "n2"]
 
-    # Should have 4 events: started/completed for n1, then started/completed for n2
-    assert len(event_bus.events) == 4
+    # Should have 5 events: started/completed for n1, edge_traversed, then started/completed for n2
+    assert len(event_bus.events) == 5
     assert event_bus.events[0] == ("started", {"stream_id": "test-stream", "node_id": "n1"})
     assert event_bus.events[1] == (
         "completed",
         {"stream_id": "test-stream", "node_id": "n1", "iterations": 1},
     )
-    assert event_bus.events[2] == ("started", {"stream_id": "test-stream", "node_id": "n2"})
-    assert event_bus.events[3] == (
+    assert event_bus.events[2] == (
+        "edge_traversed",
+        {"stream_id": "test-stream", "source_node": "n1", "target_node": "n2"},
+    )
+    assert event_bus.events[3] == ("started", {"stream_id": "test-stream", "node_id": "n2"})
+    assert event_bus.events[4] == (
         "completed",
         {"stream_id": "test-stream", "node_id": "n2", "iterations": 1},
     )

core/tests/test_litellm_streaming.py

@@ -143,6 +143,7 @@ def _has_api_key(env_var: str) -> bool:
 # ---------------------------------------------------------------------------
 # Real API tests — text streaming
 # ---------------------------------------------------------------------------
+@pytest.mark.skip(reason="Requires valid live API keys — run manually")
 class TestRealAPITextStreaming:
     """Stream a simple text response from each provider and dump events."""
 
@@ -204,6 +205,7 @@ class TestRealAPITextStreaming:
 # ---------------------------------------------------------------------------
 # Real API tests — tool call streaming
 # ---------------------------------------------------------------------------
+@pytest.mark.skip(reason="Requires valid live API keys — run manually")
 class TestRealAPIToolCallStreaming:
     """Stream a tool call response from each provider and dump events."""
 

core/tests/test_phase_compaction.py

@@ -0,0 +1,280 @@
+"""Tests for phase-aware compaction in continuous conversation mode.
+
+Validates:
+  - Phase tags persist through storage roundtrip
+  - Transition markers survive compaction
+  - Current phase messages protected during compaction
+  - Older phase tool results pruned first
+  - Phase metadata fields have safe defaults
+"""
+
+from __future__ import annotations
+
+import pytest
+
+from framework.graph.conversation import Message, NodeConversation
+
+
+class TestPhaseMetadata:
+    """Phase metadata on Message dataclass."""
+
+    def test_defaults(self):
+        msg = Message(seq=0, role="user", content="hello")
+        assert msg.phase_id is None
+        assert msg.is_transition_marker is False
+
+    def test_set_phase(self):
+        msg = Message(seq=0, role="user", content="hello", phase_id="research")
+        assert msg.phase_id == "research"
+
+    def test_transition_marker(self):
+        msg = Message(
+            seq=0,
+            role="user",
+            content="PHASE TRANSITION",
+            is_transition_marker=True,
+            phase_id="report",
+        )
+        assert msg.is_transition_marker is True
+        assert msg.phase_id == "report"
+
+    def test_storage_roundtrip(self):
+        """Phase metadata should survive to_storage_dict → from_storage_dict."""
+        msg = Message(
+            seq=5,
+            role="user",
+            content="transition",
+            phase_id="review",
+            is_transition_marker=True,
+        )
+        d = msg.to_storage_dict()
+        assert d["phase_id"] == "review"
+        assert d["is_transition_marker"] is True
+
+        restored = Message.from_storage_dict(d)
+        assert restored.phase_id == "review"
+        assert restored.is_transition_marker is True
+
+    def test_storage_roundtrip_no_phase(self):
+        """Messages without phase metadata should roundtrip cleanly."""
+        msg = Message(seq=0, role="assistant", content="hello")
+        d = msg.to_storage_dict()
+        assert "phase_id" not in d
+        assert "is_transition_marker" not in d
+
+        restored = Message.from_storage_dict(d)
+        assert restored.phase_id is None
+        assert restored.is_transition_marker is False
+
+    def test_to_llm_dict_no_metadata(self):
+        """Phase metadata should NOT appear in LLM-facing dicts."""
+        msg = Message(
+            seq=0,
+            role="user",
+            content="hello",
+            phase_id="research",
+            is_transition_marker=True,
+        )
+        d = msg.to_llm_dict()
+        assert "phase_id" not in d
+        assert "is_transition_marker" not in d
+        assert d == {"role": "user", "content": "hello"}
+
+
+class TestPhaseStamping:
+    """Messages are stamped with current phase."""
+
+    @pytest.mark.asyncio
+    async def test_messages_stamped_with_phase(self):
+        conv = NodeConversation(system_prompt="test")
+        conv.set_current_phase("research")
+
+        msg1 = await conv.add_user_message("search for X")
+        msg2 = await conv.add_assistant_message("Found it.")
+
+        assert msg1.phase_id == "research"
+        assert msg2.phase_id == "research"
+
+    @pytest.mark.asyncio
+    async def test_phase_changes_stamp(self):
+        conv = NodeConversation(system_prompt="test")
+        conv.set_current_phase("research")
+
+        msg1 = await conv.add_user_message("research msg")
+
+        conv.set_current_phase("report")
+        msg2 = await conv.add_user_message("report msg")
+
+        assert msg1.phase_id == "research"
+        assert msg2.phase_id == "report"
+
+    @pytest.mark.asyncio
+    async def test_no_phase_no_stamp(self):
+        conv = NodeConversation(system_prompt="test")
+        msg = await conv.add_user_message("no phase")
+        assert msg.phase_id is None
+
+    @pytest.mark.asyncio
+    async def test_transition_marker_flag(self):
+        conv = NodeConversation(system_prompt="test")
+        conv.set_current_phase("report")
+
+        msg = await conv.add_user_message(
+            "PHASE TRANSITION: Research → Report",
+            is_transition_marker=True,
+        )
+        assert msg.is_transition_marker is True
+        assert msg.phase_id == "report"
+
+    @pytest.mark.asyncio
+    async def test_tool_result_stamped(self):
+        conv = NodeConversation(system_prompt="test")
+        conv.set_current_phase("research")
+
+        msg = await conv.add_tool_result("call_1", "tool output here")
+        assert msg.phase_id == "research"
+
+
+class TestPhaseAwareCompaction:
+    """prune_old_tool_results protects current phase and transition markers."""
+
+    @pytest.mark.asyncio
+    async def test_transition_marker_survives_compaction(self):
+        """Transition markers should never be pruned."""
+        conv = NodeConversation(system_prompt="test")
+
+        # Old phase with a big tool result
+        conv.set_current_phase("research")
+        await conv.add_assistant_message(
+            "calling tool",
+            tool_calls=[
+                {
+                    "id": "call_1",
+                    "type": "function",
+                    "function": {"name": "search", "arguments": "{}"},
+                }
+            ],
+        )
+        await conv.add_tool_result("call_1", "x" * 20000)  # big tool result
+
+        # Transition marker
+        await conv.add_user_message(
+            "PHASE TRANSITION: Research → Report",
+            is_transition_marker=True,
+        )
+
+        # New phase
+        conv.set_current_phase("report")
+        await conv.add_assistant_message(
+            "calling another tool",
+            tool_calls=[
+                {
+                    "id": "call_2",
+                    "type": "function",
+                    "function": {"name": "save", "arguments": "{}"},
+                }
+            ],
+        )
+        await conv.add_tool_result("call_2", "y" * 200)
+
+        pruned = await conv.prune_old_tool_results(protect_tokens=0, min_prune_tokens=100)
+        assert pruned >= 1
+
+        # Transition marker should still be intact
+        marker_msgs = [m for m in conv.messages if m.is_transition_marker]
+        assert len(marker_msgs) == 1
+        assert "PHASE TRANSITION" in marker_msgs[0].content
+
+    @pytest.mark.asyncio
+    async def test_current_phase_protected(self):
+        """Tool results in the current phase should not be pruned."""
+        conv = NodeConversation(system_prompt="test")
+
+        # Old phase
+        conv.set_current_phase("research")
+        await conv.add_assistant_message(
+            "tool call",
+            tool_calls=[
+                {"id": "c1", "type": "function", "function": {"name": "s", "arguments": "{}"}}
+            ],
+        )
+        await conv.add_tool_result("c1", "old_data " * 5000)
+
+        # Current phase
+        conv.set_current_phase("report")
+        await conv.add_assistant_message(
+            "tool call",
+            tool_calls=[
+                {"id": "c2", "type": "function", "function": {"name": "s", "arguments": "{}"}}
+            ],
+        )
+        await conv.add_tool_result("c2", "current_data " * 5000)
+
+        await conv.prune_old_tool_results(protect_tokens=0, min_prune_tokens=100)
+
+        # Old phase's tool result should be pruned
+        msgs = conv.messages
+        old_tool = [m for m in msgs if m.role == "tool" and m.phase_id == "research"]
+        assert len(old_tool) == 1
+        assert old_tool[0].content.startswith("[Pruned tool result")
+
+        # Current phase's tool result should be intact
+        current_tool = [m for m in msgs if m.role == "tool" and m.phase_id == "report"]
+        assert len(current_tool) == 1
+        assert "current_data" in current_tool[0].content
+
+    @pytest.mark.asyncio
+    async def test_no_phase_metadata_works_normally(self):
+        """Without phase metadata, compaction works as before (no regression)."""
+        conv = NodeConversation(system_prompt="test")
+
+        # No phase set — messages have phase_id=None
+        await conv.add_assistant_message(
+            "tool call",
+            tool_calls=[
+                {"id": "c1", "type": "function", "function": {"name": "s", "arguments": "{}"}}
+            ],
+        )
+        await conv.add_tool_result("c1", "data " * 5000)  # ~6250 tokens
+
+        await conv.add_assistant_message(
+            "another tool call",
+            tool_calls=[
+                {"id": "c2", "type": "function", "function": {"name": "s", "arguments": "{}"}}
+            ],
+        )
+        await conv.add_tool_result("c2", "more " * 100)  # ~125 tokens
+
+        # protect_tokens=100: c2 (~125 tokens) fills the budget,
+        # c1 (~6250 tokens) becomes pruneable
+        pruned = await conv.prune_old_tool_results(protect_tokens=100, min_prune_tokens=100)
+        assert pruned >= 1
+
+    @pytest.mark.asyncio
+    async def test_pruned_message_preserves_phase_metadata(self):
+        """Pruned messages should keep their phase_id."""
+        conv = NodeConversation(system_prompt="test")
+        conv.set_current_phase("research")
+
+        await conv.add_assistant_message(
+            "tool call",
+            tool_calls=[
+                {"id": "c1", "type": "function", "function": {"name": "s", "arguments": "{}"}}
+            ],
+        )
+        await conv.add_tool_result("c1", "data " * 5000)
+
+        # Switch to new phase so research messages become pruneable
+        conv.set_current_phase("report")
+        await conv.add_assistant_message(
+            "recent",
+            tool_calls=[
+                {"id": "c2", "type": "function", "function": {"name": "s", "arguments": "{}"}}
+            ],
+        )
+        await conv.add_tool_result("c2", "x" * 200)
+
+        await conv.prune_old_tool_results(protect_tokens=0, min_prune_tokens=100)
+
+        pruned_msg = [m for m in conv.messages if m.content.startswith("[Pruned")][0]
+        assert pruned_msg.phase_id == "research"

docs/antigravity-setup.md

@@ -0,0 +1,214 @@
+# Antigravity IDE Setup
+
+Use the Hive agent framework (MCP servers and skills) inside [Antigravity IDE](https://antigravity.google/) (Google’s AI IDE).
+
+---
+
+## Quick start (3 steps)
+
+**Repo root** = the folder that contains `core/`, `tools/`, and `.agent/` (where you cloned the project).
+
+1. **Open a terminal** and go to the hive repo root (e.g. `cd ~/hive`).
+2. **Run the setup script** (use `./` so the script runs from this repo; don't use `/scripts/...`):
+   ```bash
+   ./scripts/setup-antigravity-mcp.sh
+   ```
+3. **Restart Antigravity IDE.** You should see **agent-builder** and **tools** as available MCP servers.
+
+> **Important:** Always restart/refresh Antigravity IDE after running the setup script or making any changes to MCP configuration. The IDE only loads MCP servers on startup.
+
+Done. For details, prerequisites, and troubleshooting, read on.
+
+---
+
+## What you get after setup
+
+- **agent-builder** – Create and manage agents (goals, nodes, edges).
+- **tools** – File operations, web search, and other agent tools.
+- **Skills** – Guided docs for building and testing agents (in `.agent/skills/` or `.claude/skills/`).
+
+---
+
+## Prerequisites
+
+- [Antigravity IDE](https://antigravity.google/) installed.
+- **Python 3.11+** and project dependencies. If you haven’t set up the repo yet, from repo root run:
+  ```bash
+  ./quickstart.sh
+  ```
+- **MCP server dependencies** (one-time). From repo root:
+  ```bash
+  cd core && ./setup_mcp.sh
+  ```
+
+---
+
+## Full setup (step by step)
+
+### Step 1: Install MCP dependencies (one-time)
+
+From the **repo root**:
+
+```bash
+cd core
+./setup_mcp.sh
+```
+
+This installs the framework and MCP packages and checks that the server can start.
+
+### Step 2: Register MCP servers with Antigravity
+
+Antigravity reads MCP config from your **user config file** (`~/.gemini/antigravity/mcp_config.json`), not from the project. The easiest way is to run the setup script from the **hive repo folder**:
+
+```bash
+./scripts/setup-antigravity-mcp.sh
+```
+
+The script finds the repo root, writes `~/.gemini/antigravity/mcp_config.json` with the right paths, and you don't edit any paths by hand.
+
+> **Important:** Always restart/refresh Antigravity IDE after running the setup script. MCP servers are only loaded on IDE startup.
+
+The **agent-builder** and **tools** servers should show up after restart.
+
+**Using Claude Code instead?** Run:
+
+```bash
+./scripts/setup-antigravity-mcp.sh --claude
+```
+
+That writes `~/.claude/mcp.json` as well.
+
+**Prefer to do it manually?** See [Manual MCP config](#manual-mcp-config-template) below. You’ll create `~/.gemini/mcp.json` (or `~/.claude/mcp.json`) with absolute paths to your repo’s `core` and `tools` folders.
+
+### Step 3: Use skills
+
+Skills are guides (workflow, building, testing) in `.agent/skills/` (they point to `.claude/skills/`). If Antigravity doesn’t show a “skills” UI, open those folders in the project and use the files as reference while you use the MCP tools.
+
+| Skill | What it's for |
+|-------|----------------|
+| **hive** | End-to-end workflow for building and testing agents |
+| **hive-concepts** | Core ideas for goal-driven agents |
+| **hive-create** | Step-by-step agent construction |
+| **hive-patterns** | Patterns and best practices |
+| **hive-test** | Goal-based evaluation and testing |
+| **hive-credentials** | Set up and manage agent credentials |
+
+---
+
+## What’s in the repo (`.agent/`)
+
+```
+.agent/
+├── mcp_config.json   # Template for MCP servers (agent-builder, tools)
+└── skills/           # Symlinks to .claude/skills/
+```
+
+The **setup script** writes your **user** config (`~/.gemini/antigravity/mcp_config.json`) using paths from **this repo**. The file in `.agent/` is the template; Antigravity itself uses the file in your home directory.
+
+---
+
+## Troubleshooting
+
+**MCP servers don’t connect**
+
+- Run the setup script again from the hive repo root: `./scripts/setup-antigravity-mcp.sh`, then restart Antigravity.
+- Make sure Python and deps are installed: from repo root run `./quickstart.sh`.
+- Check that the servers can start: from repo root run
+  `cd core && uv run -m framework.mcp.agent_builder_server` (Ctrl+C to stop), and in another terminal
+  `cd tools && uv run mcp_server.py --stdio` (Ctrl+C to stop).
+  If those fail, fix the errors first (e.g. install deps with `uv sync`).
+
+**"Module not found" or import errors**
+
+- Open the **repo root** as the project in the IDE (the folder that has `core/` and `tools/`).
+- If you edited `~/.gemini/antigravity/mcp_config.json` by hand, make sure `--directory` paths are **absolute** (e.g. `/Users/you/hive/core` and `/Users/you/hive/tools`).
+
+**Skills don’t show up in the UI**
+
+- Antigravity may not have a dedicated “skills” panel. Use the files in `.claude/skills/` or `.agent/skills/` as docs; the MCP tools (agent-builder, tools) still work.
+
+---
+
+## Verification prompt (optional)
+
+Paste this into Antigravity to check that MCP and skills are set up. It doesn’t use your machine’s paths; anyone can use it.
+
+```
+Check the Hive + Antigravity integration:
+
+1. MCP: List available MCP servers/tools. Confirm that "agent-builder" and "tools" (or equivalent) are connected. If not, tell the user to run ./scripts/setup-antigravity-mcp.sh from the hive repo root, then restart Antigravity (see docs/antigravity-setup.md).
+
+2. Skills: Confirm that the project has .agent/skills/ (or .claude/skills/) with: hive, hive-concepts, hive-create, hive-patterns, hive-test, hive-credentials.
+
+3. Result: Reply with PASS (MCP + skills OK), PARTIAL (only skills or only MCP), or FAIL (neither), and one line on what to fix if not PASS.
+```
+
+If you get **PARTIAL** (e.g. MCP not connected), run `./scripts/setup-antigravity-mcp.sh` from the repo root and restart Antigravity.
+
+---
+
+## Manual MCP config template
+
+Use this only if you don’t want to run the setup script. Replace `/path/to/hive` with your actual repo root (e.g. the output of `pwd` when you’re in the hive folder).
+
+Save as `~/.gemini/antigravity/mcp_config.json` (Antigravity) or `~/.claude/mcp.json` (Claude Code), then **restart the IDE** to load the new configuration.
+
+```json
+{
+  "mcpServers": {
+    "agent-builder": {
+      "command": "uv",
+      "args": ["run", "--directory", "/path/to/hive/core", "-m", "framework.mcp.agent_builder_server"],
+      "disabled": false
+    },
+    "tools": {
+      "command": "uv",
+      "args": ["run", "--directory", "/path/to/hive/tools", "mcp_server.py", "--stdio"],
+      "disabled": false
+    }
+  }
+}
+```
+
+Make sure `uv` is installed and available in your PATH. Note: Use `--directory` in args instead of `cwd` for Antigravity compatibility.
+
+---
+
+## Verify from the command line (optional)
+
+From the **repo root**:
+
+**Check that config and skills exist**
+
+```bash
+test -f .agent/mcp_config.json && echo "OK: mcp_config.json" || echo "MISSING"
+for s in hive hive-concepts hive-create hive-patterns hive-test hive-credentials; do
+  test -L .agent/skills/$s && test -d .agent/skills/$s && echo "OK: $s" || echo "BROKEN: $s"
+done
+```
+
+**Check that the config is valid JSON**
+
+```bash
+python3 -c "import json; json.load(open('.agent/mcp_config.json')); print('OK: valid JSON')"
+```
+
+**Test that MCP servers start** (two terminals)
+
+```bash
+# Terminal 1
+cd core && uv run -m framework.mcp.agent_builder_server
+
+# Terminal 2
+cd tools && uv run mcp_server.py --stdio
+```
+
+If both start without errors, the config is fine.
+
+---
+
+## See also
+
+- [Cursor IDE support](../README.md#cursor-ide-support) – Same MCP servers and skills for Cursor
+- [MCP Integration Guide](../core/MCP_INTEGRATION_GUIDE.md) – How the framework MCP works
+- [Environment setup](../ENVIRONMENT_SETUP.md) – Repo and Python setup

docs/contributing-lint-setup.md

@@ -102,6 +102,14 @@ The repository includes a `.claude/settings.json` hook that automatically runs `
 
 The `.cursorrules` file at the repo root tells Cursor's AI the project's style rules (line length, import order, quote style, etc.) so generated code follows convention.
 
+### Antigravity IDE
+
+Antigravity IDE (Google's AI-powered IDE) is supported via `.antigravity/mcp_config.json` and `.antigravity/skills/` (symlinks to `.claude/skills/`). See [antigravity-setup.md](antigravity-setup.md) for setup and troubleshooting.
+
+### Codex CLI
+
+Codex CLI (OpenAI, v0.101.0+) is supported via `.codex/config.toml` (MCP server config) and `.agents/skills/` (symlinks to `.claude/skills/`). These files are tracked in git. Run `codex` in the repo root and type `use hive` to start. See the [Codex CLI section in the README](../README.md#codex-cli) for details.
+
 ---
 
 ## CI Pipeline

docs/developer-guide.md

@@ -25,7 +25,8 @@ Aden Agent Framework is a Python-based system for building goal-driven, self-imp
 | **framework** | `/core`    | Core runtime, graph executor, protocols   | Python 3.11+ |
 | **tools**     | `/tools`   | MCP tools for agent capabilities          | Python 3.11+ |
 | **exports**   | `/exports` | Agent packages (user-created, gitignored) | Python 3.11+ |
-| **skills**    | `.claude`  | Claude Code skills for building/testing   | Markdown     |
+| **skills**    | `.claude`, `.agents`, `.agent` | Shared skills for Claude/Codex/other coding agents | Markdown     |
+| **codex**     | `.codex`   | Codex CLI project configuration (MCP servers) | TOML         |
 
 ### Key Principles
 
@@ -46,7 +47,8 @@ Ensure you have installed:
 - **Python 3.11+** - [Download](https://www.python.org/downloads/) (3.12 or 3.13 recommended)
 - **uv** - Python package manager ([Install](https://docs.astral.sh/uv/getting-started/installation/))
 - **git** - Version control
-- **Claude Code** - [Install](https://docs.anthropic.com/claude/docs/claude-code) (optional, for using building skills)
+- **Claude Code** - [Install](https://docs.anthropic.com/claude/docs/claude-code) (optional)
+- **Codex CLI** - [Install](https://github.com/openai/codex) (optional)
 
 Verify installation:
 
@@ -116,6 +118,33 @@ Skills are also available in Cursor. To enable:
 3. Restart Cursor to load the MCP servers from `.cursor/mcp.json`
 4. Type `/` in Agent chat and search for skills (e.g., `/hive-create`)
 
+### Codex CLI Support
+
+Hive supports [OpenAI Codex CLI](https://github.com/openai/codex) (v0.101.0+).
+
+Configuration files are tracked in git:
+- `.codex/config.toml` — MCP server config (`agent-builder`)
+- `.agents/skills/` — Symlinks to Hive skills
+
+To use Codex with Hive:
+1. Run `codex` in the repo root
+2. Type `use hive` to start the agent workflow
+
+Example:
+```
+codex> use hive
+```
+
+
+### Opencode Support
+To enable Opencode integration:
+
+1. Create/Ensure `.opencode/` directory exists
+2. Configure MCP servers in `.opencode/mcp.json`
+3. Restart Opencode to load the MCP servers
+4. Switch to the Hive agent
+* **Tools:** Accesses `agent-builder` and standard `tools` via standard MCP protocols over stdio.
+
 ### Verify Setup
 
 ```bash
@@ -154,6 +183,10 @@ hive/                                    # Repository root
 │       ├── hive-concepts/               # Fundamental concepts
 │       ├── hive-patterns/               # Best practices
 │       └── hive-test/                   # Test and validate agents
+├── .codex/                              # Codex CLI project config
+│   └── config.toml                      # Codex MCP server definitions
+├── .agents/                             # Shared skill mountpoint
+│   └── skills/                          # Symlinks to Hive skills
 │
 ├── core/                                # CORE FRAMEWORK PACKAGE
 │   ├── framework/                       # Main package code
@@ -205,10 +238,10 @@ hive/                                    # Repository root
 ├── scripts/                             # Utility scripts
 │   └── auto-close-duplicates.ts         # GitHub duplicate issue closer
 │
+├── .agent/                        # Antigravity IDE: mcp_config.json + skills (symlinks)
 ├── quickstart.sh                        # Interactive setup wizard
 ├── README.md                            # Project overview
 ├── CONTRIBUTING.md                      # Contribution guidelines
-├── CHANGELOG.md                         # Version history
 ├── LICENSE                              # Apache 2.0 License
 ├── docs/CODE_OF_CONDUCT.md              # Community guidelines
 └── SECURITY.md                          # Security policy
@@ -646,7 +679,6 @@ lsof -i :4000
 # Kill process
 kill -9 <PID>
 
-# Or change ports in config.yaml and regenerate
 ```
 
 ### Environment Variables Not Loading

docs/environment-setup.md

@@ -65,28 +65,26 @@ source .venv/bin/activate
 
 If you prefer to set up manually or the script fails:
 
-### 1. Install Core Framework
+### 1. Sync Workspace Dependencies
 
 ```bash
-cd core
-uv pip install -e .
+# From repository root - this creates a single .venv at the root
+uv sync
 ```
 
-### 2. Install Tools Package
+> **Note:** The `uv sync` command uses the workspace configuration in `pyproject.toml` to install both `core` (framework) and `tools` (aden_tools) packages together. This is the recommended approach over individual `pip install -e` commands which may fail due to circular dependencies.
+
+### 2. Activate the Virtual Environment
 
 ```bash
-cd tools
-uv pip install -e .
+# Linux/macOS
+source .venv/bin/activate
+
+# Windows (PowerShell)
+.venv\Scripts\Activate.ps1
 ```
 
-### 3. Upgrade OpenAI Package
-
-```bash
-# litellm requires openai >= 1.0.0
-uv pip install --upgrade "openai>=1.0.0"
-```
-
-### 4. Verify Installation
+### 3. Verify Installation
 
 ```bash
 uv run python -c "import framework; print('✓ framework OK')"
@@ -133,15 +131,15 @@ hive run exports/my_agent --tui
 
 ### CLI Command Reference
 
-| Command | Description |
-|---------|-------------|
-| `hive tui` | Browse agents and launch TUI dashboard |
-| `hive run <path>` | Execute an agent (`--tui`, `--model`, `--mock`, `--quiet`, `--verbose`) |
-| `hive shell [path]` | Interactive REPL (`--multi`, `--no-approve`) |
-| `hive info <path>` | Show agent details |
-| `hive validate <path>` | Validate agent structure |
-| `hive list [dir]` | List available agents |
-| `hive dispatch [dir]` | Multi-agent orchestration |
+| Command                | Description                                                             |
+| ---------------------- | ----------------------------------------------------------------------- |
+| `hive tui`             | Browse agents and launch TUI dashboard                                  |
+| `hive run <path>`      | Execute an agent (`--tui`, `--model`, `--mock`, `--quiet`, `--verbose`) |
+| `hive shell [path]`    | Interactive REPL (`--multi`, `--no-approve`)                            |
+| `hive info <path>`     | Show agent details                                                      |
+| `hive validate <path>` | Validate agent structure                                                |
+| `hive list [dir]`      | List available agents                                                   |
+| `hive dispatch [dir]`  | Multi-agent orchestration                                               |
 
 ### Using Python directly (alternative)
 
@@ -186,10 +184,16 @@ Skills are also available in Cursor. To enable:
 
 ### 2. Build an Agent
 
+**Claude Code:**
 ```
 claude> /hive
 ```
 
+**Codex CLI:**
+```
+codex> use hive
+```
+
 Follow the prompts to:
 
 1. Define your agent's goal
@@ -281,18 +285,20 @@ Set-ExecutionPolicy -Scope Process -ExecutionPolicy Bypass
 
 ### "ModuleNotFoundError: No module named 'framework'"
 
-**Solution:** Install the core package:
+**Solution:** Sync the workspace dependencies:
 
 ```bash
-cd core && uv pip install -e .
+# From repository root
+uv sync
 ```
 
 ### "ModuleNotFoundError: No module named 'aden_tools'"
 
-**Solution:** Install the tools package:
+**Solution:** Sync the workspace dependencies:
 
 ```bash
-cd tools && uv pip install -e .
+# From repository root
+uv sync
 ```
 
 Or run the setup script:
@@ -350,15 +356,14 @@ The Hive framework consists of three Python packages:
 
 ```
 hive/
+├── .venv/                   # Single workspace venv (created by uv sync)
 ├── core/                    # Core framework (runtime, graph executor, LLM providers)
 │   ├── framework/
-│   ├── .venv/              # Created by quickstart.sh
 │   └── pyproject.toml
 │
 ├── tools/                   # Tools and MCP servers
 │   ├── src/
 │   │   └── aden_tools/     # Actual package location
-│   ├── .venv/              # Created by quickstart.sh
 │   └── pyproject.toml
 │
 ├── exports/                 # Agent packages (user-created, gitignored)
@@ -368,28 +373,29 @@ hive/
     └── templates/           # Pre-built template agents
 ```
 
-## Separate Virtual Environments
+## Virtual Environment Setup
 
-Hive primarily uses **uv** to create and manage separate virtual environments for `core` and `tools`.
+Hive uses **uv workspaces** to manage dependencies. When you run `uv sync` from the repository root, a **single `.venv`** is created at the root containing both packages.
 
-The project uses separate virtual environments to:
+### Benefits of Workspace Mode
 
-- Isolate dependencies and avoid conflicts
-- Allow independent development and testing of each package
-- Enable MCP servers to run with their specific dependencies
+- **Single environment** - No need to switch between multiple venvs
+- **Unified dependencies** - Consistent package versions across core and tools
+- **Simpler development** - One activation, access to everything
 
 ### How It Works
 
-When you run `./quickstart.sh`, `uv` sets up:
+When you run `./quickstart.sh` or `uv sync`:
 
-1. **core/.venv/** - Contains the `framework` package and its dependencies (anthropic, litellm, mcp, etc.)
-2. **tools/.venv/** - Contains the `aden_tools` package and its dependencies (beautifulsoup4, pandas, etc.)
+1. **/.venv/** - Single root virtual environment is created
+2. Both `framework` (from core/) and `aden_tools` (from tools/) are installed
+3. All dependencies (anthropic, litellm, beautifulsoup4, pandas, etc.) are resolved together
 
-If you need to refresh environments manually, use `uv`:
+If you need to refresh the environment:
 
 ```bash
-cd core && uv sync
-cd ../tools && uv sync
+# From repository root
+uv sync
 ```
 
 ### Cross-Package Imports
@@ -522,11 +528,42 @@ export ADEN_CREDENTIALS_PATH="/custom/path"
 export AGENT_STORAGE_PATH="/custom/storage"
 ```
 
+## Opencode Setup
+
+[Opencode](https://github.com/opencode-ai/opencode) is fully supported as a coding agent.
+
+### Automatic Setup
+
+Run the quickstart script in the root directory:
+
+```bash
+./quickstart.sh
+```
+
+## Codex Setup
+
+[OpenAI Codex CLI](https://github.com/openai/codex) (v0.101.0+) is supported with project-level config:
+
+- `.codex/config.toml` — MCP server configuration (`agent-builder`)
+- `.agents/skills/` — Symlinks to Hive skills
+
+These files are tracked in git and available on clone. To use Codex with Hive:
+
+1. Run `codex` in the repo root
+2. Type `use hive` to start the agent workflow
+
+Quick verification:
+
+```bash
+test -f .codex/config.toml && echo "OK: Codex config" || echo "MISSING: .codex/config.toml"
+test -d .agents/skills/hive && echo "OK: Skills" || echo "MISSING: .agents/skills/"
+```
+
 ## Additional Resources
 
 - **Framework Documentation:** [core/README.md](../core/README.md)
 - **Tools Documentation:** [tools/README.md](../tools/README.md)
-- **Example Agents:** [exports/](../exports/)
+- **Example Agents:** [examples/](../examples/)
 - **Agent Building Guide:** [.claude/skills/hive-create/SKILL.md](../.claude/skills/hive-create/SKILL.md)
 - **Testing Guide:** [.claude/skills/hive-test/SKILL.md](../.claude/skills/hive-test/SKILL.md)
 

docs/key_concepts/goals_outcome.md

@@ -32,26 +32,26 @@ Each goal has weighted success criteria that define what "done" looks like. Thes
 
 ```python
 Goal(
-    id="twitter-outreach",
-    name="Personalized Twitter Outreach",
+    id="deep-research",
+    name="Deep Research Report",
     success_criteria=[
         SuccessCriterion(
-            id="personalized",
-            description="Messages reference specific details from the prospect's profile",
+            id="comprehensive",
+            description="Report covers all major aspects of the research topic",
             metric="llm_judge",
             weight=0.4
         ),
         SuccessCriterion(
-            id="compliant",
-            description="Messages follow brand voice guidelines",
+            id="cited",
+            description="All claims are backed by cited sources",
             metric="llm_judge",
             weight=0.3
         ),
         SuccessCriterion(
-            id="actionable",
-            description="Each message includes a clear call to action",
+            id="structured",
+            description="Report has clear sections with headings and a summary",
             metric="output_contains",
-            target="CTA",
+            target="## Summary",
             weight=0.3
         ),
     ],

docs/quizzes/README.md

@@ -40,7 +40,7 @@ Welcome to the Aden Engineering Challenges! These quizzes are designed for stude
 After completing challenges, submit your work by:
 
 1. Creating a GitHub Gist with your answers
-2. Emailing the link to `careers@adenhq.com` with subject: `[Engineering Challenge] Your Name - Track Name`
+2. Emailing the link to `contact@adenhq.com` with subject: `[Engineering Challenge] Your Name - Track Name`
 3. Include your GitHub username in the email
 
 ## Getting Help

docs/roadmap-developer-success.md

@@ -236,6 +236,17 @@ By default, Hive supports deployment through Docker.
 
 If you want zero-ops deployment, easier integration and credential management, and built-in logging, Aden Cloud is ideal. You get secure defaults, scaling, and observability out of the box—at the cost of less low-level control and some vendor lock-in.
 
+## Autonomous Agent Deployment
+
+Hive is designed to support 
+
+- Memory sustainalibility (what are the memory to keep and what to discard)
+- Event source management
+- Recoverablility
+- Repeatability
+- Volume - Multiple approach to support batch operation
+
+
 ## Deployment Strategy
 
 Autonomous and interactive modes look different, but the core remains the same, and your deployment strategy should be consistent across both.

docs/roadmap.md

@@ -68,6 +68,7 @@ subgraph Expansion
         j2["Cursor"]
         j3["Opencode"]
         j4["Antigravity"]
+        j5["Codex CLI"]
     end
     subgraph plat["Platform"]
         k1["JavaScript/TypeScript SDK"]
@@ -254,6 +255,7 @@ classDef done fill:#9e9e9e,color:#fff,stroke:#757575
 - [ ] Cursor
 - [ ] Opencode
 - [ ] Antigravity
+- [ ] Codex CLI (in progress)
 
 ### File System Enhancement
 - [ ] Semantic Search integration

docs/why-conditional-edge-priority.md

@@ -0,0 +1,42 @@
+# Why Conditional Edges Need Priority (Function Nodes)
+
+## The problem
+
+Function nodes return everything they computed. They don't pick one output key — they return all of them.
+
+```python
+def score_lead(inputs):
+    score = compute_score(inputs["profile"])
+    return {
+        "score": score,
+        "is_high_value": score > 80,
+        "needs_enrichment": score > 50 and not inputs["profile"].get("company"),
+    }
+```
+
+Lead comes in: score 92, no company on file. Output: `{"score": 92, "is_high_value": True, "needs_enrichment": True}`.
+
+Two conditional edges leaving this node:
+
+```
+Edge A: needs_enrichment == True  → enrichment node
+Edge B: is_high_value == True     → outreach node
+```
+
+Both are true. Without priority, the graph either fans out to both (wrong — you'd email someone while still enriching their data) or picks one randomly (wrong — non-deterministic).
+
+## Priority fixes it
+
+```
+Edge A: needs_enrichment == True   priority=2  (higher = checked first)
+Edge B: is_high_value == True      priority=1
+Edge C: is_high_value == False     priority=0
+```
+
+Executor keeps only the highest-priority matching group. A wins. Lead gets enriched first, loops back, gets re-scored — now `needs_enrichment` is false, B wins, outreach happens.
+
+## Why event loop nodes don't need this
+
+The LLM understands "if/else." You tell it in the prompt: "if needs enrichment, set `needs_enrichment`. Otherwise if high value, set `approved`." It picks one. Only one conditional edge matches.
+
+A function just returns a dict. It doesn't do "otherwise." Priority is the "otherwise" for function nodes.

examples/templates/README.md

@@ -44,4 +44,3 @@ uv run python -m exports.my_research_agent --input '{"topic": "..."}'
 |----------|-------------|
 | [deep_research_agent](deep_research_agent/) | Interactive research agent that searches diverse sources, evaluates findings with user checkpoints, and produces a cited HTML report |
 | [tech_news_reporter](tech_news_reporter/) | Researches the latest technology and AI news from the web and produces a well-organized report |
-| [twitter_outreach](twitter_outreach/) | Researches a Twitter/X profile, crafts a personalized outreach email, and sends it after user approval |

examples/templates/deep_research_agent/agent.json

@@ -17,6 +17,8 @@
     "terminal_nodes": [
       "report"
     ],
+    "conversation_mode": "continuous",
+    "identity_prompt": "You are a rigorous research agent. You search for information from diverse, authoritative sources, analyze findings critically, and produce well-cited reports. You never fabricate information \u2014 every claim must trace back to a source you actually retrieved.",
     "nodes": [
       {
         "id": "intake",
@@ -32,7 +34,8 @@
         "nullable_output_keys": [],
         "input_schema": {},
         "output_schema": {},
-        "system_prompt": "You are a research intake specialist. The user wants to research a topic.\nHave a brief conversation to clarify what they need.\n\n**STEP 1 \u2014 Read and respond (text only, NO tool calls):**\n1. Read the topic provided\n2. If it's vague, ask 1-2 clarifying questions (scope, angle, depth)\n3. If it's already clear, confirm your understanding and ask the user to confirm\n\nKeep it short. Don't over-ask.\n\nAfter your message, call ask_user() to wait for the user's response.\n\n**STEP 2 \u2014 After the user confirms, call set_output:**\n- set_output(\"research_brief\", \"A clear paragraph describing exactly what to research, what questions to answer, what scope to cover, and how deep to go.\")",
+        "success_criteria": "The research brief is specific and actionable: it states the topic, the key questions to answer, the desired scope, and depth.",
+        "system_prompt": "You are a research intake specialist. The user wants to research a topic.\nHave a brief conversation to clarify what they need.\n\n**STEP 1 \u2014 Read and respond (text only, NO tool calls):**\n1. Read the topic provided\n2. If it's vague, ask 1-2 clarifying questions (scope, angle, depth)\n3. If it's already clear, confirm your understanding and ask the user to confirm\n\nKeep it short. Don't over-ask.\n\n**STEP 2 \u2014 After the user confirms, call set_output:**\n- set_output(\"research_brief\", \"A clear paragraph describing exactly what to research, what questions to answer, what scope to cover, and how deep to go.\")",
         "tools": [],
         "model": null,
         "function": null,
@@ -63,7 +66,8 @@
         ],
         "input_schema": {},
         "output_schema": {},
-        "system_prompt": "You are a research agent. Given a research brief, find and analyze sources.\n\nIf feedback is provided, this is a follow-up round \u2014 focus on the gaps identified.\n\nWork in phases:\n1. **Search**: Use web_search with 3-5 diverse queries covering different angles.\n   Prioritize authoritative sources (.edu, .gov, established publications).\n2. **Fetch**: Use web_scrape on the most promising URLs (aim for 5-8 sources).\n   Skip URLs that fail. Extract the substantive content.\n3. **Analyze**: Review what you've collected. Identify key findings, themes,\n   and any contradictions between sources.\n\nImportant:\n- Work in batches of 3-4 tool calls at a time to manage context\n- After each batch, assess whether you have enough material\n- Prefer quality over quantity \u2014 5 good sources beat 15 thin ones\n- Track which URL each finding comes from (you'll need citations later)\n\nWhen done, use set_output:\n- set_output(\"findings\", \"Structured summary: key findings with source URLs for each claim. Include themes, contradictions, and confidence levels.\")\n- set_output(\"sources\", [{\"url\": \"...\", \"title\": \"...\", \"summary\": \"...\"}])\n- set_output(\"gaps\", \"What aspects of the research brief are NOT well-covered yet, if any.\")",
+        "success_criteria": "Findings reference at least 3 distinct sources with URLs. Key claims are substantiated by fetched content, not generated.",
+        "system_prompt": "You are a research agent. Given a research brief, find and analyze sources.\n\nIf feedback is provided, this is a follow-up round \u2014 focus on the gaps identified.\n\nWork in phases:\n1. **Search**: Use web_search with 3-5 diverse queries covering different angles.\n   Prioritize authoritative sources (.edu, .gov, established publications).\n2. **Fetch**: Use web_scrape on the most promising URLs (aim for 5-8 sources).\n   Skip URLs that fail. Extract the substantive content.\n3. **Analyze**: Review what you've collected. Identify key findings, themes,\n   and any contradictions between sources.\n\nImportant:\n- Work in batches of 3-4 tool calls at a time \u2014 never more than 10 per turn\n- After each batch, assess whether you have enough material\n- Prefer quality over quantity \u2014 5 good sources beat 15 thin ones\n- Track which URL each finding comes from (you'll need citations later)\n- Call set_output for each key in a SEPARATE turn (not in the same turn as other tool calls)\n\nWhen done, use set_output (one key at a time, separate turns):\n- set_output(\"findings\", \"Structured summary: key findings with source URLs for each claim. Include themes, contradictions, and confidence levels.\")\n- set_output(\"sources\", [{\"url\": \"...\", \"title\": \"...\", \"summary\": \"...\"}])\n- set_output(\"gaps\", \"What aspects of the research brief are NOT well-covered yet, if any.\")",
         "tools": [
           "web_search",
           "web_scrape",
@@ -99,7 +103,8 @@
         "nullable_output_keys": [],
         "input_schema": {},
         "output_schema": {},
-        "system_prompt": "Present the research findings to the user clearly and concisely.\n\n**STEP 1 \u2014 Present (your first message, text only, NO tool calls):**\n1. **Summary** (2-3 sentences of what was found)\n2. **Key Findings** (bulleted, with confidence levels)\n3. **Sources Used** (count and quality assessment)\n4. **Gaps** (what's still unclear or under-covered)\n\nEnd by asking: Are they satisfied, or do they want deeper research? Should we proceed to writing the final report?\n\nAfter your presentation, call ask_user() to wait for the user's response.\n\n**STEP 2 \u2014 After the user responds, call set_output:**\n- set_output(\"needs_more_research\", \"true\")  \u2014 if they want more\n- set_output(\"needs_more_research\", \"false\") \u2014 if they're satisfied\n- set_output(\"feedback\", \"What the user wants explored further, or empty string\")",
+        "success_criteria": "The user has been presented with findings and has explicitly indicated whether they want more research or are ready for the report.",
+        "system_prompt": "Present the research findings to the user clearly and concisely.\n\n**STEP 1 \u2014 Present (your first message, text only, NO tool calls):**\n1. **Summary** (2-3 sentences of what was found)\n2. **Key Findings** (bulleted, with confidence levels)\n3. **Sources Used** (count and quality assessment)\n4. **Gaps** (what's still unclear or under-covered)\n\nEnd by asking: Are they satisfied, or do they want deeper research? Should we proceed to writing the final report?\n\n**STEP 2 \u2014 After the user responds, call set_output:**\n- set_output(\"needs_more_research\", \"true\")  \u2014 if they want more\n- set_output(\"needs_more_research\", \"false\") \u2014 if they're satisfied\n- set_output(\"feedback\", \"What the user wants explored further, or empty string\")",
         "tools": [],
         "model": null,
         "function": null,
@@ -127,7 +132,8 @@
         "nullable_output_keys": [],
         "input_schema": {},
         "output_schema": {},
-        "system_prompt": "Write a comprehensive research report as an HTML file and present it to the user.\n\n**STEP 1 \u2014 Write the HTML report (tool calls, NO text to user yet):**\n\n1. Compose a complete, self-contained HTML document with embedded CSS styling.\n   Use a clean, readable design: max-width container, pleasant typography,\n   numbered citation links, a table of contents, and a references section.\n\n   Report structure inside the HTML:\n   - Title & date\n   - Executive Summary (2-3 paragraphs)\n   - Table of Contents\n   - Findings (organized by theme, with [n] citation links)\n   - Analysis (synthesis, implications, areas of debate)\n   - Conclusion (key takeaways, confidence assessment)\n   - References (numbered list with clickable URLs)\n\n   Requirements:\n   - Every factual claim must cite its source with [n] notation\n   - Be objective \u2014 present multiple viewpoints where sources disagree\n   - Distinguish well-supported conclusions from speculation\n   - Answer the original research questions from the brief\n\n2. Save the HTML file:\n   save_data(filename=\"report.html\", data=<your_html>)\n\n3. Get the clickable link:\n   serve_file_to_user(filename=\"report.html\", label=\"Research Report\")\n\n**STEP 2 \u2014 Present the link to the user (text only, NO tool calls):**\n\nTell the user the report is ready and include the file:// URI from\nserve_file_to_user so they can click it to open. Give a brief summary\nof what the report covers. Ask if they have questions.\n\nAfter presenting the link, call ask_user() to wait for the user's response.\n\n**STEP 3 \u2014 After the user responds:**\n- Answer follow-up questions from the research material\n- Call ask_user() again if they might have more questions\n- When the user is satisfied: set_output(\"delivery_status\", \"completed\")",
+        "success_criteria": "An HTML report has been saved, the file link has been presented to the user, and the user has acknowledged receipt.",
+        "system_prompt": "Write a research report as an HTML file and present it to the user.\n\nIMPORTANT: save_data requires TWO separate arguments: filename and data.\nCall it like: save_data(filename=\"report.html\", data=\"<html>...</html>\")\nDo NOT use _raw, do NOT nest arguments inside a JSON string.\n\n**STEP 1 \u2014 Write and save the HTML report (tool calls, NO text to user yet):**\n\nBuild a clean HTML document. Keep the HTML concise \u2014 aim for clarity over length.\nUse minimal embedded CSS (a few lines of style, not a full framework).\n\nReport structure:\n- Title & date\n- Executive Summary (2-3 paragraphs)\n- Key Findings (organized by theme, with [n] citation links)\n- Analysis (synthesis, implications)\n- Conclusion (key takeaways)\n- References (numbered list with clickable URLs)\n\nRequirements:\n- Every factual claim must cite its source with [n] notation\n- Be objective \u2014 present multiple viewpoints where sources disagree\n- Answer the original research questions from the brief\n\nSave the HTML:\n  save_data(filename=\"report.html\", data=\"<html>...</html>\")\n\nThen get the clickable link:\n  serve_file_to_user(filename=\"report.html\", label=\"Research Report\")\n\nIf save_data fails, simplify and shorten the HTML, then retry.\n\n**STEP 2 \u2014 Present the link to the user (text only, NO tool calls):**\n\nTell the user the report is ready and include the file:// URI from\nserve_file_to_user so they can click it to open. Give a brief summary\nof what the report covers. Ask if they have questions.\n\n**STEP 3 \u2014 After the user responds:**\n- Answer follow-up questions from the research material\n- When the user is satisfied: set_output(\"delivery_status\", \"completed\")",
         "tools": [
           "save_data",
           "serve_file_to_user",

examples/templates/deep_research_agent/agent.py

@@ -1,12 +1,15 @@
 """Agent graph construction for Deep Research Agent."""
 
+from pathlib import Path
+
 from framework.graph import EdgeSpec, EdgeCondition, Goal, SuccessCriterion, Constraint
 from framework.graph.edge import GraphSpec
-from framework.graph.executor import ExecutionResult, GraphExecutor
-from framework.runtime.event_bus import EventBus
-from framework.runtime.core import Runtime
+from framework.graph.executor import ExecutionResult
+from framework.graph.checkpoint_config import CheckpointConfig
 from framework.llm import LiteLLMProvider
 from framework.runner.tool_registry import ToolRegistry
+from framework.runtime.agent_runtime import AgentRuntime, create_agent_runtime
+from framework.runtime.execution_stream import EntryPointSpec
 
 from .config import default_config, metadata
 from .nodes import (
@@ -102,22 +105,40 @@ edges = [
         condition=EdgeCondition.ON_SUCCESS,
         priority=1,
     ),
-    # review -> research (feedback loop, checked first)
+    # review -> research (feedback loop)
     EdgeSpec(
         id="review-to-research-feedback",
         source="review",
         target="research",
         condition=EdgeCondition.CONDITIONAL,
-        condition_expr="str(needs_more_research).lower() == 'true'",
-        priority=2,
+        condition_expr="needs_more_research == True",
+        priority=1,
     ),
-    # review -> report (complementary condition — proceed to report when no more research needed)
+    # review -> report (user satisfied)
     EdgeSpec(
         id="review-to-report",
         source="review",
         target="report",
         condition=EdgeCondition.CONDITIONAL,
-        condition_expr="str(needs_more_research).lower() != 'true'",
+        condition_expr="needs_more_research == False",
+        priority=2,
+    ),
+    # report -> research (user wants deeper research on current topic)
+    EdgeSpec(
+        id="report-to-research",
+        source="report",
+        target="research",
+        condition=EdgeCondition.CONDITIONAL,
+        condition_expr="str(next_action).lower() == 'more_research'",
+        priority=2,
+    ),
+    # report -> intake (user wants a new topic — default when not more_research)
+    EdgeSpec(
+        id="report-to-intake",
+        source="report",
+        target="intake",
+        condition=EdgeCondition.CONDITIONAL,
+        condition_expr="str(next_action).lower() != 'more_research'",
         priority=1,
     ),
 ]
@@ -126,7 +147,7 @@ edges = [
 entry_node = "intake"
 entry_points = {"start": "intake"}
 pause_nodes = []
-terminal_nodes = ["report"]
+terminal_nodes = []
 
 
 class DeepResearchAgent:
@@ -136,6 +157,12 @@ class DeepResearchAgent:
     Flow: intake -> research -> review -> report
                       ^           |
                       +-- feedback loop (if user wants more)
+
+    Uses AgentRuntime for proper session management:
+    - Session-scoped storage (sessions/{session_id}/)
+    - Checkpointing for resume capability
+    - Runtime logging
+    - Data folder for save_data/load_data
     """
 
     def __init__(self, config=None):
@@ -147,10 +174,10 @@ class DeepResearchAgent:
         self.entry_points = entry_points
         self.pause_nodes = pause_nodes
         self.terminal_nodes = terminal_nodes
-        self._executor: GraphExecutor | None = None
         self._graph: GraphSpec | None = None
-        self._event_bus: EventBus | None = None
+        self._agent_runtime: AgentRuntime | None = None
         self._tool_registry: ToolRegistry | None = None
+        self._storage_path: Path | None = None
 
     def _build_graph(self) -> GraphSpec:
         """Build the GraphSpec."""
@@ -173,82 +200,99 @@ class DeepResearchAgent:
             },
         )
 
-    def _setup(self) -> GraphExecutor:
+    def _setup(self, mock_mode: bool = False) -> None:
         """Set up the executor with all components."""
         from pathlib import Path
 
         storage_path = Path.home() / ".hive" / "agents" / "deep_research_agent"
         storage_path.mkdir(parents=True, exist_ok=True)
 
-        self._event_bus = EventBus()
         self._tool_registry = ToolRegistry()
 
         mcp_config_path = Path(__file__).parent / "mcp_servers.json"
         if mcp_config_path.exists():
             self._tool_registry.load_mcp_config(mcp_config_path)
 
-        llm = LiteLLMProvider(
-            model=self.config.model,
-            api_key=self.config.api_key,
-            api_base=self.config.api_base,
-        )
+        llm = None
+        if not mock_mode:
+            llm = LiteLLMProvider(
+                model=self.config.model,
+                api_key=self.config.api_key,
+                api_base=self.config.api_base,
+            )
 
         tool_executor = self._tool_registry.get_executor()
         tools = list(self._tool_registry.get_tools().values())
 
         self._graph = self._build_graph()
-        runtime = Runtime(storage_path)
 
-        self._executor = GraphExecutor(
-            runtime=runtime,
+        checkpoint_config = CheckpointConfig(
+            enabled=True,
+            checkpoint_on_node_start=False,
+            checkpoint_on_node_complete=True,
+            checkpoint_max_age_days=7,
+            async_checkpoint=True,
+        )
+
+        entry_point_specs = [
+            EntryPointSpec(
+                id="default",
+                name="Default",
+                entry_node=self.entry_node,
+                trigger_type="manual",
+                isolation_level="shared",
+            )
+        ]
+
+        self._agent_runtime = create_agent_runtime(
+            graph=self._graph,
+            goal=self.goal,
+            storage_path=self._storage_path,
+            entry_points=entry_point_specs,
             llm=llm,
             tools=tools,
             tool_executor=tool_executor,
-            event_bus=self._event_bus,
-            storage_path=storage_path,
-            loop_config=self._graph.loop_config,
+            checkpoint_config=checkpoint_config,
         )
 
-        return self._executor
-
-    async def start(self) -> None:
-        """Set up the agent (initialize executor and tools)."""
-        if self._executor is None:
-            self._setup()
+    async def start(self, mock_mode=False) -> None:
+        """Set up and start the agent runtime."""
+        if self._agent_runtime is None:
+            self._setup(mock_mode=mock_mode)
+        if not self._agent_runtime.is_running:
+            await self._agent_runtime.start()
 
     async def stop(self) -> None:
-        """Clean up resources."""
-        self._executor = None
-        self._event_bus = None
+        """Stop the agent runtime and clean up."""
+        if self._agent_runtime and self._agent_runtime.is_running:
+            await self._agent_runtime.stop()
+        self._agent_runtime = None
 
     async def trigger_and_wait(
         self,
-        entry_point: str,
-        input_data: dict,
+        entry_point: str = "default",
+        input_data: dict | None = None,
         timeout: float | None = None,
         session_state: dict | None = None,
     ) -> ExecutionResult | None:
         """Execute the graph and wait for completion."""
-        if self._executor is None:
+        if self._agent_runtime is None:
             raise RuntimeError("Agent not started. Call start() first.")
-        if self._graph is None:
-            raise RuntimeError("Graph not built. Call start() first.")
 
-        return await self._executor.execute(
-            graph=self._graph,
-            goal=self.goal,
-            input_data=input_data,
+        return await self._agent_runtime.trigger_and_wait(
+            entry_point_id=entry_point,
+            input_data=input_data or {},
             session_state=session_state,
         )
 
     async def run(
-        self, context: dict, session_state=None
+        self, context: dict, mock_mode=False, session_state=None
     ) -> ExecutionResult:
         """Run the agent (convenience method for single execution)."""
-        await self.start()
+        await self.start(mock_mode=mock_mode)
         try:
             result = await self.trigger_and_wait(
-                "start", context, session_state=session_state
+                "default", context, session_state=session_state
             )
             return result or ExecutionResult(success=False, error="Execution timeout")
         finally:

examples/templates/deep_research_agent/config.py

@@ -16,6 +16,11 @@ class AgentMetadata:
         "multi-source search, quality evaluation, and synthesis - with TUI conversation "
         "at key checkpoints for user guidance and feedback."
     )
+    intro_message: str = (
+        "Hi! I'm your deep research assistant. Tell me a topic and I'll investigate it "
+        "thoroughly — searching multiple sources, evaluating quality, and synthesizing "
+        "a comprehensive report. What would you like me to research?"
+    )
 
 
 metadata = AgentMetadata()