fix(acp): cancelled turn with final_response=None permanently bricks the session (every prompt queued forever)

[israellot]

Problem

Cancelling an ACP turn (session/cancel, e.g. hitting Stop in Zed / a VS Code ACP client) while a tool call is in flight can permanently brick the session: every subsequent user prompt is answered only with Queued for the next turn. (N queued) and never runs. The only recovery is restarting the hermes acp process.

Root cause

When a cancel lands mid tool call, run_conversation returns {"final_response": None, "interrupted": True, ...} — the key is present with value None, so in acp_adapter/server.py::prompt:

final_response = result.get("final_response", "")   # keeps None — default doesn't apply
...
suppress_interrupt_response = interrupted and final_response.startswith(  # AttributeError
    INTERRUPT_WAITING_FOR_MODEL_PREFIX
)

The AttributeError escapes prompt() before the state.is_running = False reset at the end of the post-turn tail. The session is then stuck marked running forever:

• the client sees prompt failed ...: Internal error (acp.exceptions.RequestError via acp/task/supervisor.py);
• every later prompt hits the if state.is_running: guard and is appended to state.queued_prompts → Queued for the next turn. (N queued);
• nothing ever drains the queue, because draining only happens at the end of a successful prompt() call that can never start.

Observed live (agent log excerpt):

acp_adapter.server: Cancelled session 145febe7-...
agent.tool_executor: Tool terminal returned error (89.92s): {"output": "[Command interrupted]", "exit_code": 130}
agent.conversation_loop: Turn ended: reason=interrupted_by_user ... response_len=0
prompt failed for 145febe7-...: Internal error
acp.exceptions.RequestError: Internal error

Trigger condition: the cancel must land while the turn has produced no partial response text yet (e.g. during a long-running terminal tool call), so final_response comes back None rather than "" or partial text.

Fix (two layers)

1. Coerce the trigger: final_response = result.get("final_response") or "" — None can no longer reach .startswith().
2. Fix the class, not just the instance: wrap the whole post-turn tail (history persist, provenance update, auto-title, final-response delivery) in try/finally so the state.is_running / state.current_prompt_text reset is guaranteed. Any future exception in that tail can no longer leave the session stuck busy — it propagates to the client as a failed request, but the next prompt runs normally.

The diff is larger than the logic change because the tail block is re-indented under the try:; git diff -w shows the real delta.

Tests

Two regression tests in tests/acp/test_server.py:

• test_prompt_cancelled_turn_with_none_response_does_not_brick_session — cancelled turn returning final_response=None completes with stop_reason="cancelled", leaves is_running=False and an empty queue, and a follow-up prompt runs as a normal turn instead of being queued.
• test_prompt_tail_exception_still_releases_session — a forced exception inside the guarded tail (save_session raising) still resets is_running/current_prompt_text.

python -m pytest tests/acp/ -q
308 passed

[alt-glitch]

This was generated by AI during triage.

Related: competing fix for the same None-final_response ACP crash as #61927 and #62776 (the latter combines #61927 + #50461). Flagging the cluster so a maintainer can pick the canonical fix.

[tonydwb]

Code Review Summary

Verdict: Approved

Fix for ACP cancelled turns with final_response=None causing permanent blocking. Substantial fix with 170 additions, 78 deletions. No security concerns.

---

Reviewed by Hermes Agent

[teknium1]

Thanks for the targeted ACP regression fix. The current-main premise is verified: acp_adapter/server.py:1599 preserves an explicit None from result["final_response"], and :1606 calls .startswith() before the normal session release at :1646-1648. Because the busy guard at :1374-1383 queues subsequent prompts while the state remains running, the reported permanent queueing behavior follows directly.

The PR normalizes the value before that prefix check and places the post-turn release in finally; its two added tests cover both the None cancellation path and an exception during history persistence. Current origin/main has no commits after this PR's base touching acp_adapter/server.py or tests/acp/test_server.py, so this should be a clean salvage.

This is an automated hermes-sweeper review.

[hopyrez366]

Confirmed on my end. Running Hermes ACP with local Ollama model (qwen3.6-27B) via Zed agent panel.

Reproduction:

1. Start a conversation in Zed agent panel
2. Agent starts tool calls
3. Hit Stop/cancel mid-tool-call
4. Session bricks permanently — every subsequent prompt returns Queued for the next turn. (N queued)
5. Only recovery is restarting hermes acp process

This has been blocking my workflow consistently. The fix in this PR looks correct — wrapping post-turn tail in try/finally and normalizing None before .startswith() should prevent the permanent brick.

Environment: Hermes latest main, Zed stable, Linux, local Ollama at http://192.168.1.15:9090/v1