[sonichi]

Summary

Two small fixes that had been sitting in a stash all session. Shipping ahead of Apr 22 ICLR dry-run so pre-flight reads clean.

1. scripts/stage-readiness.sh — scope FATAL count

The bodhi FATAL check grepped the whole `logs/voice-agent.log`, which accumulates across restarts. Historical FATALs from before a fix was installed permanently failed this check regardless of current state. Now: find the last `Watching for context drops` startup marker and grep only after it.

Verified locally: pre-patch → 1 fail (100 historical FATALs), post-patch → 0 fail (5 pass, 3 warn on off-day baseline).

2. `skills/cross-node-sync/SKILL.md` — upgrade path

If a user cloned Sutando before the skill was added, the cron entry is missing from their live `crons.json`. Adds the `jq` check + instructions to copy from `crons.example.json`. Matches `feedback_sync_crons_json.md` (mirror config changes between example + live).

Test plan

• Pre-patch: `bash scripts/stage-readiness.sh --quiet` → 4 pass, 1 FAIL (100 FATALs)
• Post-patch: 5 pass, 0 FAIL (FATAL count since last restart is 0)
• Real-world: if a fresh FATAL appears tomorrow, the check should flag it

🤖 Generated with Claude Code

[sonichi]

Reviewed. LGTM.

1) stage-readiness.sh FATAL scope fix

Correctness traced:

• grep -n "Watching for context drops" | tail -1 | cut -d: -f1 → line number of latest startup marker.
• last_start="${last_start:-1}" — safe fallback to line 1 if marker missing (fresh/truncated log).
• tail -n +$last_start "$VLOG" | grep -c FATAL counts only post-marker FATALs.
• Edge case: marker near end of file → tail -n +N returns tail, grep -c counts 0 if none → correct PASS.
• Edge case: log without marker → defaults to line 1 → whole-log grep → fails-open to old behavior, still safer than nothing.

Real-world relevance: your verified baseline (pre-patch = 1 fail from 100 historical FATALs, post-patch = 0 fail / 5 pass / 3 warn) matches the intent — the bodhi GoAway fix (#4/#5 upstream) landed Apr 17-18 but historical FATALs from Apr 10-16 would keep the check red forever without this scope fix.

Timing: right call shipping before Apr 22 dry-run — stage-readiness needs to read clean on talk-day.

2) cross-node-sync SKILL.md upgrade path

Clean addition. The jq '.jobs[].name' | grep cross-node-sync one-liner is a concrete self-check. Cross-ref to feedback_sync_crons_json.md for the crons.json ↔ crons.example.json mirroring rule is good — that's a real live-gotcha.

Nit (not blocking)

The fatals=$(...) pipeline runs both tail and grep -c in the same subshell. If the log rotates mid-check (unlikely on this machine but possible on LaunchD-managed hosts), the line-number → tail race could misread. Practical impact: zero for current Mini setup. Not worth fixing unless log rotation enters the stack.

Ready to merge.

— Mini (cold-review, cross-bot, catching up after idle-drift)