| name | durable-bug-hunt-harness |
| description | Drive a forked bug-hunt PR (scripts/bug-hunt-daily.sh) to durable green by chunking hermes -z workers under one global deadline, retrying failed chunks on a proven surviving model, repairing truncated JSON, and verifying with real-agent dry runs before requesting review. |
durable-bug-hunt-harness
When the daily bug hunt cron in jleechanorg/jleechanclaw fires, the underlying scripts/bug-hunt-daily.sh script dispatches hermes -z workers (one per agent lane) over the merged-PRs list. The original version ran each lane sequentially, ate 30 minutes per lane, and silently dropped any chunk that timed out, ran out of tokens, or emitted truncated JSON — producing false-clean reports.
This recipe drives the PR that fixes that harness end-to-end on the fix/durable-bug-hunt-harness branch.
Trigger
- Cron fires (
ai.hermes.schedule.bug-hunt-9am), every chunk JSON is 0 bytes, and the report says "All bug hunt chunks failed — 0 bugs recorded".
- LLM rewrites PR body or labels mention "durable bug hunt harness", "chunk", "retry on proven", or "truncated JSON repair".
- AO session named
bughunt-dedupe (or any AO worker) starts a fix branch for scripts/bug-hunt-daily.sh.
Workflow
-
Confirm the failure mode
ls /tmp/bughunt-* — inspect the most recent run dir; look for 0-byte JSON outputs and oversized stderr.- Read the launchd plist
~/Library/LaunchAgents/ai.hermes.schedule.bug-hunt-9am.plist and the script header for the active lane config.
- Capture the symptom (timeout? truncated JSON?
|| true swallowed exit code?) before designing a fix.
-
Branch from
