fix: interrupt cleanup doesn't propagate to LangGraph checkpoint

[avoidwork]

Summary

Session state cleanup on interrupt only affects in-memory state, not the LangGraph checkpoint, causing orphaned tool calls to persist and corrupt subsequent turns.

Environment

• OS: Linux 7.0.2-7-pve
• Node.js: v25.8.1
• madz version: 1.22.0
• LLM provider: Unknown — user to confirm

Reproduction

1. Start a conversation in the TUI
2. Trigger a tool call that requires user input (e.g., file read, web search)
3. Interrupt the tool execution (e.g., via command or cancel)
4. Send a new message to resume the conversation
5. Observe orphaned tool calls in the checkpoint corrupting the resumed turn

Expected Behavior

Interrupt cleanup should propagate to the LangGraph checkpoint, removing orphaned tool calls and ensuring the checkpoint state matches the in-memory conversation state.

Actual Behavior

Only in-memory state is cleaned up. The LangGraph checkpoint retains partial AIMessages with tool_calls, causing duplicate tool calls, dangling tool references in LLM API requests, and corrupted conversation history on resume.

Additional Context

• Affected files: ./src/tui/app.js, ./src/agent/react.js, ./src/session/stateManager.js
• handleChat() (app.js:922-924) calls both removeLastAssistantToolCallMessage() and popExchange(), but handleCommand() (app.js:524-526) only calls popExchange() — tool call messages are not cleaned up in the command path
• removeLastAssistantToolCallMessage() (stateManager.js:80-88) only removes from sessionState.#state.conversation. The LangGraph checkpoint — written at superstep boundaries by createReactAgentGraph — retains the partial AIMessage with tool_calls
• No resumption reconciliation: When the user sends a new message after interrupt, a fresh dispatchProvider call uses the cleaned in-memory conversation, but the LangGraph checkpoint still contains orphaned messages. The isNewThread flag only controls system prompt injection, not checkpoint state
• Orphaned tool calls corrupt resume: Partial AIMessages with tool_calls persist in the checkpoint. On resume, streamEvents replays from the checkpoint, potentially sending dangling tool references to the LLM API

Audit Checklist

The following areas were audited during investigation:

• 1. Verify abort signal propagation through the entire streaming pipeline (app.js → react.js → LangGraph streamEvents)
• 2. Verify session state cleanup on interrupt — ensure all partial messages are removed from both in-memory state and LangGraph checkpoint
• 3. Verify checkpoint state consistency after interrupt — no orphaned AIMessages with incomplete tool_calls persist in checkpoint
• 4. Verify resumption behavior — confirm the graph can resume from checkpoint after interrupt without duplicating or losing messages
• 5. Verify that interrupted tool calls are properly cleaned up in both session state and checkpoint before resume
• 6. Verify that the abort signal doesn't leave partial state in the LangGraph checkpoint that would corrupt subsequent turns

Audit Log

1. Abort signal propagation → ✅ PASS (Risk: Low)

Signal is properly propagated and checked on each event. react.js:225-228 passes signal to streamOptions, react.js:293-306 checks signal.aborted per event, clears turn hashes, emits pending tool_end, and returns cleanly.

2. Session state cleanup on interrupt → ⚠️ INCONSISTENT (Risk: Medium)

Inconsistent cleanup between code paths. handleChat() (app.js:922-924) calls both removeLastAssistantToolCallMessage() and popExchange(), but handleCommand() (app.js:524-526) only calls popExchange() — tool call messages are not cleaned up in the command path.

3. Checkpoint state consistency after interrupt → ⚠️ FAIL (Risk: High)

removeLastAssistantToolCallMessage() only modifies sessionState.#state.conversation (in-memory array). It does NOT update the LangGraph checkpoint. Orphaned AIMessages with tool_calls persist in the checkpoint and will be replayed on resume.

4. Resumption behavior → ❌ FAIL (Risk: High)

No explicit resumption mechanism exists. When the user sends a new message after interrupt, a new dispatchProvider call is made with the cleaned in-memory conversation. However, the LangGraph checkpoint retains the old state with orphaned messages. The isNewThread flag (react.js:164) only controls system prompt injection, not checkpoint state reconciliation.

5. Interrupted tool call cleanup in checkpoint → ⚠️ FAIL (Risk: High)

Tool call cleanup is in-memory only. stateManager.js:80-88 removes the last assistant message with tool_calls from the conversation array, but the LangGraph checkpoint (written by createReactAgentGraph at superstep boundaries) still contains the partial AIMessage. On resume, the checkpoint will replay the orphaned tool call.

6. Partial state in checkpoint corrupting subsequent turns → ⚠️ FAIL (Risk: High)

Partial AIMessages from interrupted streams persist in the checkpoint. When the graph resumes via streamEvents, it reads from the checkpoint which may include orphaned AIMessages with incomplete tool_calls. This can cause the LLM API to receive corrupted conversation history with dangling tool references.

Recommendations

1. Propagate checkpoint cleanup on interrupt: After cleaning in-memory state, explicitly update or reset the LangGraph checkpoint for the current thread_id to remove orphaned messages. This may require calling checkpointer.update() or checkpointer.put() with the cleaned state.

2. Add cleanup to handleCommand(): Mirror the handleChat() interrupt cleanup by calling sessionState.removeLastAssistantToolCallMessage() in the command path (app.js:524-526).

3. Implement checkpoint reconciliation on resume: Before the next dispatchProvider call after an interrupt, verify that the checkpoint state matches the in-memory conversation. If they diverge, reconcile by writing the cleaned state to the checkpoint.

4. Add integration test: Create a test that simulates interrupt during tool execution, then resumes with a new message, and verifies the checkpoint contains no orphaned tool calls.

References

• https://reference.langchain.com/python/langgraph.checkpoint.md
• https://reference.langchain.com/python/langgraph.store.md