metrics

/debt-ops: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)

# Locate this repo's plugin cache. CLAUDE_PLUGIN_DATA is set in hook
# subprocesses but NOT in the skill's bash env, so glob the standard
# Claude Code plugin-data dirs and fall back to the legacy cache path.
CACHE_DIR=""
for D in \
    ${CLAUDE_PLUGIN_DATA:+"$CLAUDE_PLUGIN_DATA/cache/$REPO_HASH"} \
    "$HOME/.claude/plugins/data"/debt-ops*/cache/"$REPO_HASH" \
    "$HOME/.cache/debt-ops/cache/$REPO_HASH"; do
  [ -d "$D" ] && { CACHE_DIR="$D"; break; }
done

LOG="$CACHE_DIR/metrics.jsonl"
if [ -n "$CACHE_DIR" ] && [ -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 Claude 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.