batch-health

name	batch-health
description	Daily post-batch audit — checks whether all enabled scheduled skills fired in their expected morning window, alerts on silent misses, files issues on batch-level outages
var
tags	["meta","reliability"]

Today is ${today}. Cross-reference what was scheduled to run this morning against what actually ran.

Goal

run-frequency-guard catches skills running too often. Nothing catches when they go silent. ISS-021 (2026-05-06 — 8 even-day skills missed the 06:30–07:25 UTC window, no alert fired) is the canonical example. This skill closes that gap.

Runs at 08:00 UTC daily — after the 06:00–07:30 UTC batch window completes. Skills scheduled at 07:31+ are excluded from the expected list (transient cron drift could leave them in flight when we audit). Output: OK if everything ran, WARN for 1-2 isolated misses, OUTAGE for 3+ missing skills.

Steps

1. Build the expected-run list for today

Read aeon.yml. For each skill with enabled: true, check its schedule field against today's context.

Parse today's date ${today} (YYYY-MM-DD):

• Extract day-of-month (DOM).
• Odd DOM = 1, 3, 5, 7, 9, 11, 13, 15, 17, 19, 21, 23, 25, 27, 29, 31
• Even DOM = 2, 4, 6, 8, 10, 12, 14, 16, 18, 20, 22, 24, 26, 28, 30

Include a skill in the expected list if ALL of:

1. enabled: true
2. Schedule fires in the 06:00–07:30 UTC audit window. Parse the minute and hour fields:
   • hour 6 (any minute 0–59) → in window
   • hour 7 AND minute ≤ 30 → in window
   • hour 7 AND minute > 30 → exclude (too close to the 08:00 audit; transient cron drift could leave it in flight when we check)
   • multi-hour like 7,19 → counts as hour 7 for the AM window (apply the same minute rule)
3. Day-of-month matches today:
   • * * * or any-day pattern → always included
   • 1/2 * * in DOM field → odd days only
   • 2/2 * * in DOM field → even days only
   • */N * * step in DOM field → include if (DOM - 1) % N == 0. Vixie-cron expands */N to 1-31/N → DOMs 1, 1+N, 1+2N, … For */3: DOMs 1, 4, 7, 10, 13, 16, 19, 22, 25, 28, 31. Do not blanket-include — that false-WARNs ~2/3 of days.
   • Day-of-week patterns (* * N) → include only if today's weekday matches

Exclude: disabled skills, skills firing outside the 06:00–07:30 UTC window (e.g., schedule: "30 9 * * *" at 09:30, schedule: "40 7 * * *" at 07:40 — both out of scope for this audit).

2. Fetch actual runs

Try the skill-runs script:

./scripts/skill-runs --json --hours 26

From the JSON output, extract the .skills[] array — each entry has a skill name and last_conclusion. A skill that appears here (any conclusion) means the GHA cron fired.

Fallback if skill-runs fails (sandbox network block or gh api error): Parse memory/logs/${today}.md:

cat memory/logs/${today}.md 2>/dev/null

Look for ## SkillName headers or *_OK / *_SKIP / *_FAIL markers. Any skill that logged means it ran. Note: fallback gives partial coverage only — missing from the log doesn't mean it didn't run.

3. Cross-reference

For each skill in the expected list:

• OK: appears in actual-run list (any conclusion) — scheduler fired correctly
• FAILED: in actual-run list with failure conclusion — scheduler fired, skill errored (tracked by skill-health, not our concern here)
• MISSING: not in actual-run list at all — GHA cron never triggered it

Tally: N_expected, N_ok, N_failed, N_missing.

4. Classify severity

Missing count	Severity
0	OK — no notification needed
1–2	WARN — isolated hiccup, possible transient cron delay
3+	OUTAGE — batch-level failure (ISS-021 pattern)

5. File issue on OUTAGE

If 3+ missing, check whether today's outage is already tracked:

1. Read memory/issues/INDEX.md — scan open issues for one matching today's date and "batch" keyword. If found, skip filing.
2. Determine next ISS number: scan existing memory/issues/ISS-*.md files, take highest N + 1.
3. Create memory/issues/ISS-{NNN}.md:

---
id: ISS-{NNN}
title: Batch outage — {N} skills missed morning window {today}
status: open
severity: medium
category: unknown
detected_by: batch-health
detected_at: {today}T08:00:00Z
resolved_at: null
affected_skills: [{comma-separated skill names}]
root_cause: null
fix_pr: null
---

# ISS-{NNN}: Batch Outage {today}

{N} enabled skills never triggered in the 06:00–07:30 UTC window on {today}.

## Missing Skills

{list each missing skill and its aeon.yml schedule}

## Context

Detected by `batch-health` at 08:00 UTC. The GHA scheduler fired for daily-promoted skills but the batch window was silent.

## Next Steps

- Check GitHub Actions for {today} around 06:00–07:30 UTC for cron delivery failures
- If one-off: likely GHA infrastructure delay (documented pattern)
- If recurring: investigate aeon.yml schedule changes or GHA Actions billing limits

4. Append to memory/issues/INDEX.md open table:

| ISS-{NNN} | medium | Batch outage — {N} skills missed morning window {today} | {today} | batch-health |

5. Update memory/MEMORY.md issue tracker summary if the issue changes the open count.

6. Notify (WARN or OUTAGE only)

Write to .pending-notify-temp/batch-health-${today}.md (create dir if needed), then send.

OUTAGE format:

batch outage — ${today}

${N} skills never triggered (06:00–07:30 UTC):
${list each missing skill with its schedule}

ISS-${NNN} filed. check GHA cron status.

WARN format:

batch gap — ${today}

${N} scheduled skills didn't run:
${list each}

isolated miss — transient cron delay likely. monitoring.

Send: ./notify -f .pending-notify-temp/batch-health-${today}.md

7. Log

Append to memory/logs/${today}.md:

## Batch Health
- **Expected:** ${N} skills in 06:00–07:30 UTC window
- **Expected list:** ${comma-separated skill names}
- **OK:** ${N_ok}
- **Failed:** ${N_failed} (${list or "none"})
- **Missing:** ${N_missing} (${list or "none"})
- **Status:** OK / WARN / OUTAGE
- **Issue filed:** ISS-${NNN} / n/a
- **Notification:** sent / skipped (OK)
- BATCH_HEALTH_OK

Sandbox Note

./scripts/skill-runs --json calls gh api which is authenticated via GITHUB_TOKEN in GitHub Actions — no extra setup needed. If the script errors, fall back to parsing memory/logs/${today}.md for log markers. Local fallback is less precise: treat any skill with a log header or *_OK marker as "ran", but note the data source in the log entry.

Required Env Vars

None. gh api uses GITHUB_TOKEN provided automatically by GitHub Actions.