[akushonkamen]

Summary

launch() starts every ACP agent with start_new_session=True, so the only teardown path is close()'s killpg (raven/acp_client/client.py:261, raven/acp_client/client.py:301-315). A host that dies without reaching close_pool — SIGKILL, OOM killer, crash — leaves the agent orphaned, parented to init. Observed in #928: an agent running for 22 days holding 523 MiB PSS.

Root cause: no fallback teardown when the owning host process dies abruptly.

Fix: launch() records (pgid, agent birth, owner pid, owner birth) to a per-boot temp file, and every launch() sweeps that file first. A group is killed with killpg only when its owner is dead (or its pid was recycled, detected via birth mismatch) and the agent's birth still matches — so a recycled pid is neither killed nor mistaken for a live owner. close() removes its own record. On Windows the same record is used and the tree is killed with taskkill /T /F; no real Windows machine was available, so that branch is written to the #918/#919 standard and declared untested here.

Process: timeline check confirmed no earlier claim/PR by this account on #928; a claim comment with AI disclosure was posted first (issuecomment-6135634396).

Type

• Fix

Verification

• Relevant tests pass locally

Red (bug reproduced, pre-fix): timeout 20 $V /tmp/repro_928.py + ps -o pid,ppid,stat -p <AGENT_PID> → after the stub agent's host was SIGKILLed the agent stayed alive with ppid=1 (bug evidence; first reproduction pid 40717).

Green (post-fix e2e): same script, host kill -9ed, then
$V -c "from raven.acp_client.client import sweep_orphaned_agents; print(sweep_orphaned_agents())" → KILLED [58468]; ps -p 58468 confirms the process is gone.

New tests: $V -m pytest tests/test_acp_orphan_sweep.py -v → 5 passed (owner dead → group killed; live owner spared; recycled pid not killed; dead pid with birth=None; boot token).

Regression subsets ($V = pre-existing venv python, Python 3.13.15):

• $V -m pytest tests/test_acp_methods.py tests/test_acp_protocol.py tests/test_acp_stdio.py tests/test_acp_journal.py tests/test_acp_orphan_sweep.py -q → 228 passed
• $V -m pytest tests/test_subagent_host_env.py tests/test_acp_backend_cwd.py tests/test_cli_gateway_commands.py tests/test_cli_tui_commands.py -q → 171 passed
• $V -m pytest tests/test_subagent_acp_mcp_lifecycle.py -q → 21 passed

Risk

• Security impact considered — sweep only kills process groups it recorded itself in the same boot, and only after birth verification, so pid reuse cannot cause a misdirected kill.
• Backward compatibility considered — sweep runs inside launch(), no API change; close() semantics unchanged.
• Rollback path is clear — single commit; reverting restores the prior behavior exactly.

Related Issues

Fixes #928

AI disclosure: written with AI assistance (Claude / GLM-5.3-Flash); commit carries Co-authored-by: Claude (GLM-5.3-Flash) <noreply@anthropic.com>.

[LivXue]

This PR has not been verified on a real Windows machine, and its CI has been red (coverage gates and pull request title, since 2026-10-13) with no follow-up since. Closing it; #928 is now addressed by #951 (Windows) and #953 (macOS).

You are welcome to open PRs against either branch: fix/windows_acp_process_lifetime for #951, and fix/acp_orphans_on_macos for #953. Thanks for the work on #928.