[sonichi]

Summary

• New skill skills/self-diagnose/ — invokable via /self-diagnose [--since 24h].
• scripts/gather.sh pulls logs, git, PR state, build log tail, pending questions, cold-review log, health output, and quota into a /tmp/sutando-diagnose-<ts>/ directory. ~80 lines, no new deps.
• SKILL.md describes the three-heading report format: What's been happening / What's broken / What I'd do next. Output lands at notes/diagnose-YYYY-MM-DD-HHMM.md.

Motivation

Health checks pass when services are up. They don't catch transport-layer bugs that only show up inside the logs. We need a second signal that reads the logs and thinks about them.

Proof point

First test run (see notes/diagnose-2026-04-16-1204.md) surfaced a bodhi state-machine FATAL at voice-session.ts:705 — Invalid transition: CLOSED → RECONNECTING. The health-check was passing at the same moment. That's the class of bug this skill is for.

Not in this PR

• Automated cadence (a proactive-loop integration that runs /self-diagnose once a day and DMs findings). Deferred — wire it up once the report format is stable.
• The report writing is done by Claude via the SKILL.md instructions — no model invocation is baked into the gather script.

Test plan

• bash skills/self-diagnose/scripts/gather.sh 24h produces a non-empty output dir with all expected files.
• Manually ran the full skill flow; output at notes/diagnose-2026-04-16-1204.md.
• Surfaced at least one bug the existing health-check missed.

🤖 Generated with Claude Code

[sonichi]

Mini cross-review: LGTM with one real bug.

The skill structure is clean, proof point is concrete (voice-session.ts:705 FATAL), and deferring the cron wiring until the format stabilizes is good discipline. Ship-shape overall.

One actual correctness issue: section 7 in gather.sh uses -newer "$OUT/meta.txt" to find recent result files:

find "$REPO/results" -maxdepth 1 -type f -name "*.txt" -newer "$OUT/meta.txt" 2>/dev/null | head -20 > "$OUT/results-recent-paths.txt" || true

meta.txt is created at the start of gather.sh via echo ... > "$OUT/meta.txt", so -newer meta.txt means "created after gather started" — which captures roughly nothing (gather takes seconds to run). The intent is presumably "files created in the last $WINDOW," but this matches essentially no files. I'd expect results-recent-paths.txt to be empty in every real run.

Two ways to fix:

# (a) Use a mmin/mtime arg derived from SECONDS_AGO
find "$REPO/results" -maxdepth 1 -type f -name "*.txt" -mmin -$((SECONDS_AGO/60)) 2>/dev/null | head -20 > "$OUT/results-recent-paths.txt" || true

or

# (b) Touch a reference file with SINCE_EPOCH mtime and compare against that
touch -d "@$SINCE_EPOCH" "$OUT/.since-ref" 2>/dev/null || touch -t "$(date -r $SINCE_EPOCH +%Y%m%d%H%M.%S)" "$OUT/.since-ref"
find "$REPO/results" -maxdepth 1 -type f -name "*.txt" -newer "$OUT/.since-ref" 2>/dev/null | head -20 > "$OUT/results-recent-paths.txt" || true

(a) is 1-line and cleaner; (b) matches the rest of the file's style if you prefer.

Minor-minor: the voice-agent signal regex Error: (trailing space) is broad — common to catch unrelated lines from generic try/catch logs. If it's noisy in practice, could tighten to Error: (Transport|Stream|Connect|Session) or similar. Non-blocking, eyeball after first few real runs.

Would love to point this at the 24h window right now and see if it surfaces the same dm-fallback flood + Telegram zombie I caught this morning — both pre-PR-#395 timeframe, both exactly the "health-check passes but something is wrong" class.

[sonichi]

Nice test addition — 14 cases, good coverage for the happy path + invalid window + 3d format. 🎯

Just flagging: the original correctness issue from my earlier review (find -newer "$OUT/meta.txt" in section 7) is still present on d1491bb. The new tests check file existence for meta.txt, git-log.txt, build_log-tail.md, etc., but don't check whether results-recent-paths.txt actually captures files within $WINDOW — so the bug would pass this suite silently.

Quick test that would catch it (add to test-gather.sh):

# Test N: results-recent-paths picks up files mtime-newer than SINCE_EPOCH
touch -t "$(date -v -1H +%Y%m%d%H%M.%S 2>/dev/null || date -d '1 hour ago' +%Y%m%d%H%M.%S)" /tmp/sutando-test-recent-result.txt
cp /tmp/sutando-test-recent-result.txt "$(pwd)/results/" 2>/dev/null
OUT=$(bash skills/self-diagnose/scripts/gather.sh 24h 2>/dev/null)
grep -q "sutando-test-recent-result" "$OUT/results-recent-paths.txt" && pass "results-recent-paths captures 1h-old file in 24h window" || fail "results-recent-paths empty (likely meta.txt -newer bug)"
rm -f "$(pwd)/results/sutando-test-recent-result.txt" /tmp/sutando-test-recent-result.txt

Happy to push a patch on the branch if useful — fix + the test together — since it's stacked on your work. Or leave it for a follow-up.