fix(agent): sanitize unanswered tool_calls on interrupt to prevent transcript corruption

[nicoechaniz]

Summary

Fixes a bug where interrupting the agent during pending tool calls leaves the conversation transcript in an invalid state, causing the next API call to fail with a "missing tool response" error.

Closes #11351

Changes

• Adds _sanitize_unanswered_tool_calls() helper to AIAgent in run_agent.py.
• Calls the helper from _persist_session() so every exit path guarantees an API-valid transcript.
• Removes the duplicated inline backfill logic from the outer-loop error handler.

Testing

• Local pytest passes
• Rebased on latest origin/main

Backwards compatibility

No behavior change for users — this only prevents a crash condition.

[nicoechaniz]

Hi @teknium1 — this fixes transcript corruption when interrupts leave unanswered tool_calls in flight. Rebased on latest main and ready for review. Thanks!

[teknium1]

Thanks for the careful writeup and the rebase work, @nicoechaniz! This hermes-sweeper automated review found that current main already covers the reported failure mode.

The bug is already fixed on main by two existing mechanisms:

• _sanitize_api_messages() (pre-call sanitizer) — Introduced in commit 85993fbb5 (feat: pre-call sanitization and post-call tool guardrails #1732, merged 2026-03-17). It runs unconditionally before every LLM API call and injects synthetic role=tool stub results for any tool_call_id that was never answered (run_agent.py:4597 — step 2 of the method). Called at run_agent.py:9644.
• Inline cancel messages on interrupt — Both _execute_tool_calls_sequential (run_agent.py:8515) and the pre-flight check in _execute_tool_calls_concurrent (run_agent.py:8216) already append {role: "tool", content: "[Tool execution cancelled…]", tool_call_id: …} for every skipped call immediately at interrupt time.

The proposed _sanitize_unanswered_tool_calls() helper + _persist_session() hook would duplicate what _sanitize_api_messages() already does. Closing as already implemented. No action needed on your end — the fix you described shipped roughly a month ago!