Fix background task liveness probe after session resume#215
Conversation
![chatgpt-codex-connector[bot]](https://avatars.githubusercontent.com/in/1144995?s=60&v=4){kind=small}
There was a problem hiding this comment.
💡 Codex Review
Here are some automated review suggestions for this pull request.
Reviewed commit: d795398521
ℹ️ About Codex in GitHub
Codex has been enabled to automatically review pull requests in this repo. Reviews are triggered when you
- Open a pull request for review
- Mark a draft as ready
- Comment "@codex review".
If Codex has suggestions, it will comment; otherwise it will react with 👍.
When you sign up for Codex through ChatGPT, Codex can also answer questions or update the PR, like "@codex address that feedback".
| # literal "Output is being written to". The path ends at the first | ||
| # whitespace after ".output". | ||
| match=$(grep -F "$task_id" "$transcript_path" 2>/dev/null \ | ||
| | grep -oE 'Output is being written to: [^[:space:]]+\.output' \ |
There was a problem hiding this comment.
If the recorded Claude output path contains whitespace, this pattern fails to match anything because it requires the path to be a single non-whitespace token up to .output. In that case is_bg_task_alive falls back to the derived path for the resumed session; when that file is absent it fail-opens as alive, so a dead/orphaned background task remains parked forever for projects whose generated task-output path includes a space.
Useful? React with 👍 / 👎.
SwordFaith
force-pushed
the
fix/bg-task-output-path-from-transcript
branch
from
d795398 to
8929fb0
Compare
Fix A - TaskStop completion recognition: list_pending_background_task_ids now treats a TaskStop tool_result (.toolUseResult.message "Successfully stopped task: <id>") as a terminal completion event, because many Claude Code builds do not also emit a task_notification system record for a TaskStop. The id is read from .task_id, falling back to parsing it out of the message text when that field is absent. The message field is coerced with tostring before contains() so unrelated tool results carrying a structured (object/array/number) .message do not error and wipe the whole completed set under pipefail. Fix B - cross-session liveness glob: is_bg_task_alive now searches sibling session dirs under the same project slug when neither the transcript-recorded output path nor the session-derived path exists. This only expands where we look, so a genuinely alive task is never wrongly pruned; a file missing everywhere still fails open as before. Builds on the extract_bg_task_output_path_from_transcript helper landed in PolyArch#215. Tests: 51/51. Renames existing AC-NN cases to descriptive names and adds regression coverage for TaskStop completion, the cross-session glob, and non-string .message poisoning.
1 participant
Problem
When a Claude Code session is resumed or continued, the current transcript may have a different session id from the session that originally launched a
background task. The task's
.outputfile physically lives under the old session directory, but the stop hook's liveness probe derived the output pathfrom the current transcript's session id. As a result, dead/orphaned tasks were never pruned and the loop was blocked from reaching Codex review.
Fix
Teach
is_bg_task_aliveto look up the real output file path recorded in the transcript's launch message:extract_bg_task_output_path_from_transcript()to parse the "Output is being written to: " line.tasks_dirpath when the transcript is unreadable or the message is missing.transcript_pathargument throughprune_dead_bg_task_idsandlist_pending_background_task_ids.Test
Add regression test AC-25: simulate a session resume with a new-session transcript that records an old-session output path, then assert the dead task is
pruned using the real path rather than the derived new-session path.
42/42 tests pass.