Opus finding
history array de-dup is missing — same-day re-runs grow the 12-entry budget
lowbughigh
- skills/fleet-state/SKILL.md:282-290
Two runs in the same UTC day (e.g. a dry-run rehearsal followed by the scheduled Monday run, or a manual workflow_dispatch retry) append two entries with identical `run_date`. After 6 same-day pairs the 12-entry trend table is fully consumed by 6 weeks instead of 12. The trend surface silently shrinks. The fix is one-line: `unique_by(.run_date) | sort_by(.run_date) | .[-12:]` with the *latest* entry winning (the current append puts the new one after, so unique_by would keep the *first* by default — needs `group_by` + `last` or `reverse | unique_by | reverse`).
Recommendation
Replace the tail of the jq filter with: `| group_by(.run_date) | map(.[-1]) | sort_by(.run_date) | .[-12:]` to coalesce same-day re-runs.