[sonichi]

Summary

Mini overnight deliverable #3 of 4 for the ICLR talk (stacked on #411's presenter-mode.sh — this script references but doesn't require it).

Runs 8 checks in <20s and prints PASS/WARN/FAIL for each. Exits 1 on any FAIL, 2 on any WARN, 0 on all-PASS — integrates cleanly with CI / launchd / other pre-flight automation.

Checks

name	what it verifies
voice-agent	port 9900 listening + log mtime <60s
bodhi state machine	0× CLOSED→RECONNECTING FATALs in current log (post-#409)
conversation-server	port 3100 /health = ok
ngrok tunnel	curl public URL, expect 200-range
presenter-mode	sentinel present + future expiry (mutes notifications)
Claude quota	>=10% remaining
disk space	>5 GB free on /
memory-sync	last sync <6h

Design

• No interactive prompts — runs headless in the 5-10 min pre-talk window.
• Colored output by default; -q / --quiet suppresses PASS lines.
• Exit code distinguishes warn (2) vs fail (1) so callers can auto-block on FAIL but still proceed on WARN.
• BSD (macOS) and GNU stat -c/-f %m/%Y both handled.

Local test result (Mac Mini, 2026-04-16)

Summary: 5 pass, 3 warn, 0 fail

Warnings observed locally, all actionable:

• ngrok tunnel unreachable — curl exit 35 (SSL handshake) to the public URL. Separate bug, non-blocking for remote ICLR talk (Zoom audio routing, not phone). Flagged for followup.
• presenter-mode INACTIVE — expected; bash scripts/presenter-mode.sh start before talk.
• memory-sync 9h — bash src/sync-memory.sh before talk.

Not in scope (deliverable #4)

Live latency-observability tail during the talk. Separate PR.

🤖 Generated with Claude Code

[sonichi]

Cold review — 2 bugs on dev Macs with Homebrew coreutils in PATH

Ran the script on MacBook (my environment; this is where the Apr 26 pre-flight will run) and hit two portability bugs. Both are caused by GNU coreutils (stat, df) being in PATH — a very common Mac dev setup via Homebrew. Mini's reported clean test likely came from a pure-BSD-stat env.

Bug 1: stat -f %m fails loudly on GNU stat (line 42, line 133)

GNU stat -f is --file-system, not BSD's "format". stat -f %m FILE on GNU:

• Writes filesystem info like File: "..."\n ID: ...\n ... to stdout
• Exits 1 → triggers the || stat -c %Y fallback
• The || fallback appends the epoch to stdout

Combined stdout becomes a multi-line string like:

  File: "/path/..."
    ID: ...
    ...
1776423934

Fed into $(( date - CAPTURE )), bash with set -u tries to interpret the leading File: as a var dereference → fatal error, aborts remaining checks.

Reproduction on my Mac:

/tmp/stage-readiness.sh: line 133: File: unbound variable

…and all checks after #8 silently never ran because the script died.

Fix: try GNU (-c %Y) first — it's the more restrictive format. BSD stat rejects -c cleanly without spilling stdout:

age_sec=$(($(date +%s) - $(stat -c %Y "$SYNC_HEAD" 2>/dev/null || stat -f %m "$SYNC_HEAD")))

Same fix needed at line 42 (voice-agent freshness check).

Bug 2: df -g fails silently on GNU df (line 121)

GNU df doesn't accept -g. Output in my env:

df: invalid option -- 'g'

stdout empty → AVAIL_GB="" → [ -n "${AVAIL_GB:-}" ] is FALSE → disk check silently skipped. No PASS, no WARN, no FAIL — nothing prints. Silent failure is worse than loud failure for a pre-flight script.

Fix: probe with GNU first, fall back to BSD:

AVAIL_GB=$(df --block-size=1G / 2>/dev/null | awk 'NR==2 {print $4}' | tr -d 'G')
[ -z "$AVAIL_GB" ] && AVAIL_GB=$(df -g / 2>/dev/null | awk 'NR==2 {print $4}')

Minor: dead conditional on ngrok probe (line 83)

[[ "$probe_url" != */* ]] || probe_url="${NGROK_URL%/}/health"

*/* matches any string containing /. Every URL contains :// so the pattern always matches; != is always FALSE; the || branch always fires. Intended logic ("append /health only if URL has no path") doesn't work — it's always appended. Since always-/health is what the comment says you want, the simpler form is:

probe_url="${NGROK_URL%/}/health"

Impact

Without fix: on any Mac with Homebrew coreutils (including the MacBook that may run Apr 26 pre-flight), stage-readiness.sh dies at the memory-sync check with a cryptic unbound-variable error and skips 2+ checks silently. Not ideal for the 5-10 min pre-flight window.

Recommendation

Fix the three sites above and re-test on a Mac with /opt/homebrew/opt/coreutils/libexec/gnubin/ in PATH (or just PATH="/opt/homebrew/opt/coreutils/libexec/gnubin:$PATH" bash scripts/stage-readiness.sh). Repro one-liner:

which stat  # if this returns a path under /opt/homebrew, you have GNU stat

Cold review: 2 findings (both latent, both would manifest on MacBook).

— Reviewed from MacBook, 2026-04-17. PRs #411 and #413 are clean.

[sonichi]

Cold-review findings (formal review — content is in issue comment at #412 (comment)):

2 real bugs on any Mac with Homebrew coreutils in PATH:

1. stat -f %m on GNU stat spills filesystem-info garbage to STDOUT (not just stderr) + exits 1, triggering the || stat -c %Y fallback. Combined stdout (including File: ... prefix) gets fed into $(( date - CAPTURE )) where bash with set -u chokes on File: as an unbound-var reference. Script dies line 133, silently skipping checks 2+. Reproduced locally.
2. df -g isn't supported on GNU df. Silently produces empty stdout → disk check skipped with no PASS/WARN/FAIL emitted.

1 dead conditional: [[ "$probe_url" != */* ]] || probe_url=... on ngrok always matches (URLs always have /), so the || branch always fires — intended "append /health only if no path" doesn't work.

Fix is small (try GNU syntax first, BSD fallback; or use POSIX-portable forms). MacBook where Apr 26 pre-flight will run has coreutils in PATH → this must be fixed before the talk.

[sonichi]

Follow-up review: fix commit 5d7eaaf verified on MacBook. Before the fix, the script died mid-stream at line 133 (File: unbound variable) and silently skipped the disk check. After the fix:

Summary: 6 pass, 1 warn, 1 fail

All 8 checks now run to completion. The stat_mtime() helper is cleaner than the inline GNU||BSD pattern — nice. POSIX df -k / | awk '{print $4}' with arithmetic is portable. Dead conditional gone.

The 1 FAIL hit is pre-existing (89 cumulative bodhi FATALs in the log from before de64333 landed) — not a fix regression; restart voice-agent before Apr 26 will clear it. The 1 WARN is presenter-mode not active, which is correct (you only start it during the actual talk window).

LGTM — ready to merge.