skill-stats-reporter

The thesis: a 340-skill catalog needs an observable shape — which are alive, which are dead, which fail too often, which take too long. Without it, the catalog is opaque. With it, Cooper can promote winners and retire losers in 5-minute batches.

Procedure

bash skills/skill-stats-reporter/run.sh           # full daily run
bash skills/skill-stats-reporter/run.sh --dry-run # build report, don't write

Pipeline

1. Read data/skill-telemetry/invocations.jsonl (entire log; small for a year).
2. Bucket events by skill: start count, end count, end status distribution, durations.
3. Walk the catalog directories (skills/, ~/.openclaw/skills/) for the universe of skill names.
4. Cross-reference:

• dead skills: in catalog, 0 start events in last 30 days.
• hot skills: >10 start events/day on average over last 7 days.
• failure-prone: ≥5 end status=error events in last 7 days.
• slow skills: median duration_sec > 60s over last 7 days.

1. Write data/skill-telemetry/<YYYY-MM-DD>.md (full report).
2. Auto-publish the dead-skill list to data/skill-telemetry/dead-skills-pending-review.md (replaces last week's; one canonical surface for batch review).

Output

data/skill-telemetry/<YYYY-MM-DD>.md contains six sections:

# Skill stats — <YYYY-MM-DD>

## Summary
- Total skills in catalog: N
- Skills with telemetry: M
- Coverage: M/N (%)

## Dead skills (no invocations in 30d)
...

## Hot skills (>10/day, last 7d)
...

## Failure-prone skills (≥5 errors in 7d)
...

## Slow skills (median > 60s in 7d)
...

Status codes

STATUS=ok report=<path> dead=<N> hot=<M> / STATUS=skip reason=no-telemetry-yet / STATUS=error reason=<...>

P12 — counter-monitor for the reporter

If the reporter has not run in 25 hours, the heartbeat-silence pattern picks it up via the data/skill-telemetry/<YYYY-MM-DD>.md mtime. (Cooper-side: add the file to the silence-detector's watch list when this skill graduates from cont-18.)