fix(acp): ignore idle cancel before next prompt

[MauroFigueira]

Summary

• Treat ACP cancel requests received while a session is already idle as no-ops.
• Clear any stale cancel/interrupt state in that idle path so the next prompt is not immediately reported as interrupted.
• Add regression coverage for both idle-cancel and true running-cancel behavior.

Root cause

Some ACP clients, observed with Xcode, can send a best-effort cancel immediately before submitting the next prompt. Hermes previously always set the session cancel event and called agent.interrupt() whenever a cancel request arrived, even if state.is_running was already false.

That stale interrupt then poisoned the next session/prompt: run_conversation() saw the pending interrupt before calling the model and returned interrupted_by_user. In Xcode this surfaced as:

Error: JSON-RPC Internal Error (code -32603)
The server encountered an internal error processing the request.

Fix

Only propagate cancel to the agent when the ACP session is actively running. If the session is idle, clear the cancel event and any stale agent interrupt instead.

This preserves real cancellation behavior for in-flight prompts while avoiding a stale idle cancel affecting the next user prompt.

Test plan

uv run --extra acp --extra dev python -m pytest tests/acp_adapter/test_acp_commands.py -q -o 'addopts='
# 8 passed

[teknium1]

Thanks for isolating the stale ACP interrupt path. The current-main premise is confirmed: acp_adapter/server.py:1221-1224 sets the cancel event and calls agent.interrupt() even while idle, while agent/conversation_loop.py:648-653 exits a new turn when that flag remains set.

Problems

• tests/acp/test_server.py:342-347 currently creates an idle session, invokes cancel, and asserts that cancel_event is set. This PR intentionally clears that event for idle sessions, so the existing test will fail after applying the change.

Suggested changes

• Update test_cancel_sets_event to model a running session (state.is_running = True) before asserting the real-cancel behavior. The new idle-cancel regression test can then cover the opposite contract.

Automated hermes-sweeper review.

[teknium1]

This intended idle no-op makes tests/acp/test_server.py:342-347 fail because that existing test cancels an idle session and asserts cancel_event.is_set(). Update that test to set state.is_running = True before asserting real cancellation behavior.

[MauroFigueira]

Addressed in 517b579. The existing test now sets state.is_running = True (and is renamed to test_running_cancel_sets_event) before asserting the real-cancel behavior. The idle-cancel regression remains covered separately. Verified with both relevant test files: 88 passed.

[MauroFigueira]

Addressed the sweeper feedback in 517b579: test_cancel_sets_event now explicitly models an active turn with state.is_running = True (and was renamed accordingly), while the idle-cancel regression test continues to cover the opposite contract.

Verification: 88 passed across tests/acp/test_server.py and tests/acp_adapter/test_acp_commands.py on the PR branch.