Repository Janitor

A conservative Linux repository auditor with a deterministic Python executor, private finding history, and an optional OpenClaw scheduler watchdog.

Workflow

1. Define the inventory. Read applicable project instructions and references/configuration.md. Create a private configuration based on references/repositories.example.json. List only repositories the operator has authorized. Keep runtime checkout and administrative Git checkout separate. Done: each target has a verified identity and scope; examples are not treated as installed configuration.
2. Run the audit. From this skill directory, run:

   python3 scripts/audit.py --config /absolute/path/to/private-config.json --fetch --github --persist --report
   

   The executor owns its lock. Do not acquire it again or run a second fetch around this command. Omit --fetch, --github, or --persist when those actions are not requested; omitted network checks remain explicitly unknown. Done: a report or an explicit execution failure is available.
3. Interpret the result. Read references/findings.md. Separate successful execution from repository cleanliness. Preserve open findings when coverage is incomplete. A clean index is not proof that PRs, commits, or worktrees need no attention. Done: coverage, evidence, open findings, and unknown checks are stated separately.
4. Handle any requested organization manually. The Python executor implements only AUDIT. For ORGANIZE_SAFE or PREPARE_CLEANUP, follow references/supervised-actions.md under explicit authorization for that run. Done: approved actions have preconditions, verified backup, and postconditions; other targets remain untouched.
5. Configure monitoring only when requested. Read references/scheduling.md. Test execution and delivery independently. Do not overwrite an existing job whose owner or authority cannot be verified. Done: schedule, destination, owner, and watchdog behavior have been checked, or the activation blocker is reported.
6. Validate changes. Run python3 -B scripts/test_audit.py. Use temporary repositories for tests, not production changes. Confirm that only the current run's lock was released. Done: focused tests pass and a partial audit is never described as a successful complete audit.

Requirements and effects

• Linux, Python 3, Git, and readable /proc for process observations.
• GitHub CLI (gh) with separately configured authentication for --github.
• Optional systemd/SQLite runtime checks require the explicit runtime fields documented in the configuration reference.
• OpenClaw CLI is required only for the optional scheduler watchdog.
• The audit reads Git metadata, worktree status, filesystem metadata, and optional runtime/database signals. It does not read changed-file contents.
• --fetch performs git fetch --prune against the configured remote: remote-tracking refs can change, but working files are not updated.
• The executor creates an exclusive run marker at the configured lock path, then removes only its own marker. An existing marker is preserved, even if it looks stale.
• --persist writes private JSON state atomically. A corrupt baseline blocks replacement. Keep state outside the skill, source repositories, indexed memory, and public artifacts.
• Optional process environment inspection selects only the configured database directory key and never emits the environment contents.
• A read-only SQLite connection runs PRAGMA quick_check with an interrupt deadline; no migrations, vacuum, or business-row queries are executed.

Modes

AUDIT — default and only automated mode

Inspect repositories and report risks. No cleanup, automatic commits, fast-forward, deployment, or pruning of worktree registrations. A successful process exit indicates that a report was produced: inspect its errors and findings before claiming complete coverage.

ORGANIZE_SAFE — explicit authorization for one supervised run

Allows only verified pruning of administrative records for genuinely missing worktree directories, or a verified fast-forward of a clean non-runtime checkout. This is a manual workflow, not a CLI flag. Existing dirty, locked, or active worktrees remain untouched.

PREPARE_CLEANUP — explicit authorization

Identify exact targets and prepare verified backups. Do not delete or overwrite targets. Actual material cleanup requires a separate supervised authorization for those exact targets.

Permanent automated limits

The scheduled auditor never deletes code, untracked files, branches, tags, worktrees, caches, builds, backups, releases, runtime dependencies, databases, uploads, credentials, or logs. It never performs git clean, automatic stash, reset, rebase, force push, commit, push, PR mutation, operation continuation/abort, process termination, migration, restart, or deployment.

Age, size, inactivity, absent approval, and a clean Git status never authorize deletion. A free persistent deployment lock is normal and is not removed.

Coverage boundaries

• Worktrees are inspected individually; registered repositories sharing a common Git directory are deduplicated.
• Process observations use readable process working directories, not a complete attribution system. Partial visibility is reported in observations; absence of a PID does not prove inactivity.
• WIP age is based on available dirty-path mtimes, not the moment the worktree became dirty. Pure deletions can have unknown age. Automation ownership and handoffs require external evidence.
• PR listing is bounded to 100 results; reaching the limit is incomplete coverage. Review threads are not queried by this version. Draft status and missing approval are not defects by themselves.
• Discovery and artifact measurements are bounded metadata inventories, not licenses to clean.
• A watchdog on the same host/Gateway is not an external server availability monitor.

The public package contains examples only: no publisher deployment paths, scheduler IDs, account credentials, repository history, or operational findings.