recipe-maintainers/cc-ci-orchestrator
1
0
Fork 0
Code Issues Pull Requests Actions Packages Projects Releases Wiki Activity

watchdog: fix _completed() false-positive that abandoned the run

Browse Source 
The 2026-07-03 finish run wedged because _completed() returned True while the run
was still mid-work — so the watchdog exited early and nothing recovered the wedge.
Cause: it scanned part.get('text') across ALL message parts, so DONE_MARKER inside
a TOOL part (a subagent `task` prompt / bash command that referenced 'print
UPGRADE RUN COMPLETE') matched. Now: require the marker in the LAST assistant
TEXT (prose) message — the genuine sign-off — ignoring tool-call args and any
mid-run echo of the instruction (work after the echo disqualifies it).

Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01WxbpH3DquKzoSTSwGvGuET
This commit is contained in:
autonomic-bot
2026-07-04 10:08:33 +00:00
parent 69dd17833a
commit ddedea80cd
1 changed files with 13 additions and 8 deletions
Download Patch File Download Diff File Expand all files Collapse all files

21
cc-ci-plan/launch-upgrader.py
View File

@ -325,21 +325,26 @@ def _run_pids(sid=None):
return out return out
def _completed(): def _completed():
# The run is done only when the MODEL prints DONE_MARKER — i.e. it appears in an ASSISTANT # Done only when the MODEL signs off with DONE_MARKER as its FINAL word: the marker in the LAST
# message's output. NOT a log grep: the kickoff/resume PROMPT (a user message) also contains # assistant TEXT (prose) message. This guards THREE false-positives that each abandoned a run:
# the marker (it instructs the model to print it), which would false-positive. # (1) the kickoff/resume PROMPT contains the marker — but that's a USER message (skipped).
# (2) the marker inside a TOOL part — a subagent `task` prompt or a bash command that echoes
# "print <marker>" — so we look ONLY at type=='text' prose, never tool-call args.
# (3) the model ECHOING the instruction mid-run ("…then I'll print <marker>") — only the FINAL
# assistant prose counts, so any further work after the echo means it's NOT done.
sid = _session_id() sid = _session_id()
msgs = _server_get(f"/session/{sid}/message") if sid else None msgs = _server_get(f"/session/{sid}/message") if sid else None
if msgs is not None: if msgs is not None:
msgs = msgs if isinstance(msgs, list) else msgs.get("data", []) msgs = msgs if isinstance(msgs, list) else msgs.get("data", [])
last_prose = None
for m in msgs: for m in msgs:
if ((m.get("info") or {}).get("role")) != "assistant": if ((m.get("info") or {}).get("role")) != "assistant":
continue continue
for part in (m.get("parts") or []): prose = "".join(p.get("text", "") for p in (m.get("parts") or [])
t = part.get("text") if p.get("type") == "text" and isinstance(p.get("text"), str))
if isinstance(t, str) and DONE_MARKER in t: if prose.strip():
return True last_prose = prose
return False return bool(last_prose and DONE_MARKER in last_prose)
# Server unreachable → conservative log fallback that excludes the prompt's own mention. # Server unreachable → conservative log fallback that excludes the prompt's own mention.
try: try:
with open(LOG_FILE, errors="ignore") as f: with open(LOG_FILE, errors="ignore") as f: