Fleet self review

Fleet Self-Review (daily)

Goal: keep the fleet safe, minimise repeated errors, improve, and capture learnings as custom skills — without taking any risky action itself.

Hard safety rules (non-negotiable)

• READ-ONLY on all systems. NO git push, NO deploys, NO rm/trash of anything you did not create this run, NO edits to code/config outside the two write targets below.
• WRITE TARGETS (the only places you may create/modify files):
  1. ~/clawd/reports/fleet-self-review/ — the dated report.
  2. ~/clawd/skills/ — new or improved custom skills (additive; never delete an existing skill).
• NO sending to third parties. The ONLY outbound message is the end-of-run Telegram summary to Seyed (the runner script handles it).
• NO secret access beyond reading config you already have; NEVER print secrets into the report or the summary.
• If a proposed fix is destructive or outward-facing, DO NOT do it — write it as a recommendation for Seyed to approve.
• Token discipline: reference paths/IDs, no state dumps. Keep the report tight.

Inputs to review (gather, read-only)

1. Recent memory + observations — ~/clawd/memory/ (today + last 2 days), ~/clawd/MEMORY.md index, and claude-mem if available.
2. Errors & fabrications — scan recent session memory/logs for: fabricated output (claims of success not backed by real tool output), stale-shell/cache symptoms, retractions, repeated failures. This is the top priority (see the fabrication incidents on 2026-06-02).
3. Missed / stalled tasks — Notion "Task List" (canonical), any tasks.json, open obligations in daily notes.
4. Cron / automation health — ~/Library/LaunchAgents/com.weblyfe.*, recent ~/clawd/logs/**, which jobs are disabled or failing. Check for 0-byte cron output files (ls -la ~/clawd/logs/security-scan-*.log etc.) — a 0-byte file means a job fires but produces no output, which is a silent failure, not a clean pass. Also verify pipeline last-success timestamps: if invoice-pipeline.jsonl or sync-status.json show no entry in >48h, the pipeline is broken even if the cron job "ran".
5. Fleet reachability — tailscale status (which nodes online/offline). Read-only; do NOT ssh-mutate.
6. Repo hygiene — git status in active repos (uncommitted/secret risks); flag, don't fix.

What to produce

Write ~/clawd/reports/fleet-self-review/<YYYY-MM-DD>.md with:

• Safety — any red lines crossed or at risk (fabrication, secret exposure, public exposure, destructive risk). Most important section.
• Errors to minimise — concrete recurring errors with root cause + the cheapest durable fix.
• Improvements — prioritised, concrete, each with the file/command to apply (for Seyed or a future run to action).
• Custom skills authored/updated this run — list them with one-line purpose. Author a skill ONLY when a learning is reusable and durable (e.g. a recurring-error guard). Follow the skills format; keep skills small and specific.
• Open obligations — dated follow-ups, flags, gates.
• Keep it scannable. Cite real paths/IDs. No fabrication — if you didn't verify something, say "unverified".

Authoring custom skills (the improvement loop)

When you find a recurring, durable failure or a reusable best-practice, write a skill under ~/clawd/skills/<kebab-name>/SKILL.md with proper frontmatter (name, description). Prefer guard-rail skills that prevent the errors you found (e.g. "verify-before-claiming-success", "never-relay-uncaptured-output"). Do not duplicate an existing skill — improve it in place instead.

Self-check before finishing

"Did I change anything outside reports/ and skills/? Did I claim anything I didn't verify? Did I leak a secret?" If yes to any, undo/redact before writing the summary.