Skip to content
github-actions[bot] edited this page Nov 22, 2026 · 4 revisions

clippa audit tells you how the week went, and once a month a deep audit reads the saved weeks to find what keeps going wrong. Where Checkup asks "what is wrong right now", this one looks for patterns.

The weekly digest

clippa audit         # the digest, at most three things to do
clippa audit --all   # every thing to do, and where to look

It reads files and makes no model call, so it costs nothing. It covers:

Section What it says
To do At most three things, what is breaking now first
Ran Streams done, partial and failed this week
Out Uploads by kind, and any that missed their slot
Spend The week's model and Modal spend, biggest stages first
Review What you approved and rejected
Disk Free space on the streams share
Repo Open PRs and branches besides main
Best The most-watched video published this week

A section that can't be read becomes a thing to do, and the others still run. A stream that failed and then finished counts as finished, because run.json keeps only the last result.

The claude container runs it after the morning checkup on AUDIT_WEEKDAY (deploy/claude/env.local; 1 is Monday, 7 is Sunday, the default) and DMs it to you. Each week is kept as streams/.audit/weekly-<date>.json, which is what the deep audit reads. A week it misses (the container was down) is simply missing, and nothing catches up.

The settings are [audit] in the config.

Backups

clippa backup (Backups) writes to [backup] folder, and serve makes one every [backup] every_days. With the folder set to "", the digest says every week that nothing is backing up your files. Once it is set, the digest warns when the newest file there is older than [backup] max_age_days. It only looks at the date of the newest file, so restore from a backup now and then to know it works.

Reminders

For a decision you want to come back to, add a dated line to reminders in clipper.toml:

[audit]
reminders = ["2026-11-16 compare Kino v8 with v7"]

It shows from a week before the date until you take the line out.

The monthly deep audit

On AUDIT_MONTHDAY (default the 1st) the container gives clippa audit --brief, the last deep_weeks weeks side by side with repeat problems first, to a headless claude -p run (prompt: deploy/claude/audit-deep.md). It follows the repeat problems into the stage logs and the code, looks for drift in spend, failures, disk and uploads, reads why picks were rejected, and checks what came of last month's findings.

You get one DM with the top three findings. The full report, with the evidence, is streams/.audit/deep-<YYYY-MM>.md, and next month's run reads it. The deep audit needs at least one saved week.

It runs on the Claude subscription, not the API key, and it is read-only in the same way as the nightly report: the read tools, a fixed list of commands, and one DM. It can't change anything. When a fix needs a change, the report says what to do. If Claude can't run (signed out, usage limit), you get the repeat problems alone.

To try either one by hand:

ssh nas 'cd clippa-claude && docker compose exec claude clippa-audit --weekly --dry-run'
ssh nas 'cd clippa-claude && docker compose exec claude clippa-audit --deep --dry-run'

--dry-run prints the DM instead of sending it and saves nothing. The timer's output is in /tmp/audit.log in the container.

Dependency updates

.github/dependabot.yml opens up to two grouped PRs a week, one for Python packages and one for GitHub Actions. CI runs on them and nothing merges by itself. They get no code review comment, since a Dependabot PR has no secrets, so merge on green CI and your own word.

Clone this wiki locally