Ethan's last three joyos-pr-dashboard commits are all the same class of fix — hardcoded absolute paths in tests breaking on swarm/mirror hosts. This watches commit messages for repeated fix-type patterns within a project and, when the same root cause recurs 3+ times in 14 days, writes a one-paragraph 'root cause you keep re-fixing' card to the tablet instead of another alert email. It turns repetitive firefighting into a visible, fixable systemic issue.
Create ~/.local/bin/repeat-fix-radar: for each repo in ~/.claude/active-projects, run `git log --since=14.days --pretty=%h%x09%s`, filter to `^fix(` subjects, strip conventional-commit scope, then cluster by shared significant tokens (drop stopwords, require >=2 shared tokens, e.g. 'relative path' + 'test'). Any cluster with >=3 commits emits a card. Reuse pondercard-generate for rendering and drop the output into ~/Tablet/cards/ so it rides the existing tablet-sync pipeline; dedupe via a state file at ~/.local/state/repeat-fix-radar/seen.json keyed by sorted cluster tokens so the same cluster only surfaces once per 14 days. Wire as a systemd timer (daily 07:15) alongside the existing eval cycles, with a matching repeat-fix-radar-freshness-check following the house pattern. The card text asks one Socratic question ('what invariant would make this class of fix impossible?') rather than proposing a fix.
---
**Repeat-fix radar — SHIPPED, LIVE, TIMERS ARMED**
**What it does**
- Scans `fix(...)` commits per repo over a 14-day window
- Clusters by shared significant tokens (stopwords dropped, ≥2 shared tokens)
- Any cluster of **3+ commits** = a systemic issue, not three bugs
- Writes one card to `~/Tablet/inbox/` asking *"What invariant would make this class of fix impossible?"* — no fix proposed, no alert email
- Deduped via `seen.json`, so a cluster surfaces once per window
**Files created**
- `~/bin/repeat-fix-radar` — the scanner (110 lines python3)
- `~/bin/repeat-fix-radar-freshness-check` — staleness alarm on the heartbeat
- Both `.test.sh` suites, executable
- 4 systemd user units: radar (daily 07:15) + freshness check (every 6h, offset)
**Verification**
- Radar tests: **8/8 green** — clustering, non-fix exclusion, singleton exclusion, dedupe, heartbeat
- Freshness tests: **5/5 green**, including **sabotage in all three failure directions** (missing / malformed / stale all fire; fresh stays silent)
- Real E2E: service ran against real repos, `Result=success`, heartbeat written, freshness check read it back healthy
- Positive control at a loosened threshold produced real cards from linkedin-engine and legacy-map-builder — proving zero-at-production-threshold is real
**Two things you should know**
- `joyos-pr-dashboard` **does not exist on this host** — the rep