L’interface est localisée ; les présentations et documents des Skills sont traduits automatiquement par IA dans votre langue. Le code, les commandes, chemins et termes techniques restent en version originale, et le document de l’auteur fait foi.
Explorer les Skills
Productivité · SKILL.md

metrics

metrics

Print a debt-ops health summary from the metrics log, covering registration rate, hook feedback action rate, ADR creation, and AI-authored share. Use when the user asks for "debt-ops metrics", "debt health", "registry stats", or invokes $metrics. Read-only, never writes the log.

Independent GitHub snapshot · Automated classification · Original documentation governs

Documentation de la Skill

GitHub

name: metrics description: Print a debt-ops health summary from the metrics log, covering registration rate, hook feedback action rate, ADR creation, and AI-authored share. Use when the user asks for "debt-ops metrics", "debt health", "registry stats", or invokes $metrics. Read-only, never writes the log.

metrics

Read the hidden metrics log and tell the user whether v1's tripwires are tripping.

1. Find the log

TOPLEVEL=$(git rev-parse --show-toplevel)
REPO_HASH=$(printf '%s' "$TOPLEVEL" | shasum | cut -c1-12)

# Hooks and skills share one deterministic cache base (ADR 0011). Override the
# base with DEBT_OPS_CACHE; default is ~/.cache/debt-ops.
CACHE_DIR="${DEBT_OPS_CACHE:-$HOME/.cache/debt-ops}/cache/$REPO_HASH"

LOG="$CACHE_DIR/metrics.jsonl"
if [ -f "$LOG" ]; then
  tail -n 500 "$LOG"
else
  echo "MISSING: no metrics.jsonl found for repo hash $REPO_HASH"
fi

If the file is missing or empty, tell the user the hooks haven't fired yet in this repo and stop.

2. The log format

One JSON object per line, three event shapes:

  • {"event":"edit","file":"...","registry_count":N,"ts":"..."} — every agent edit
  • {"event":"feedback","file":"...","result":"pass|fail","ts":"..."} — every quality-check fire
  • {"event":"session","registry_count":N,"adr_count":M,"ai_authored_count":K,"ts":"..."} — start of each session

Timestamps are ISO-8601 UTC.

3. Compute the tripwires

Filter to the last 7 days. Then compute:

  • Edits / sessions — counts of event:edit and event:session.
  • Registry growth — last registry_count minus first (across either edit or session events). >0 means Discipline 1 is firing.
  • ADR growth — last adr_count minus first (session events only).
  • AI-authored share trend — first vs. last session percentage (ai_authored_count / registry_count, when registry_count>0).
  • Feedback pass rate — count(result:pass) / count(event:feedback).
  • FAIL → PASS rate — for each feedback event with result:fail, look at the next feedback event for the same file. Count those that flipped to pass. Divide by total fails. Below 50% means the agent isn't reliably acting on hook output — the architectural alarm bell.

If there are fewer than 5 sessions in the window, say "need more data" and skip the verdict.

4. Report

One screen. No padding. Use → and ↑/↓ for trends. Example shape:

debt-ops metrics — last 7 days
─────────────────────────────────
edits           : 142  (8 sessions, ~18 edits/session)
registry        : +3
adrs            : +1
ai-authored     : 50% → 60% ↑

feedback ran    : 89 times
pass rate       : 88%
fail → pass rate: 80% (8/10)

verdict: ok

End with one judgment line:

  • ok — registry growth >0 AND fail→pass rate ≥50%.
  • investigate: — name the specific tripwire that tripped.

Don't

  • Don't write to the log.
  • Don't compute metrics not listed above.
  • Don't guess at health when data is thin — say "need more data" instead.
Skills enregistrées