Housekeeping

Runs on: local-exec - runs local scripts and edits OS files; on a read-only surface it reports what it would do.

The OS accumulates maintenance debt between sessions. The SessionStart brief and the audit detect pieces of it; nothing remediates it. This skill encodes the split, the ordering (later fixes read what earlier ones write), and the verify step.

It invents no new machinery. Every detector composes a tool that already ships: the cadence date checks, /dream, scripts/log-archive.py, /wiki-build, the lint skill, scripts/memory-diff.py, scripts/brain-snapshot.py, scripts/skill_health.py.

The maturity ladder

1. Detect (default) - read-only. One report. Writes nothing. Live here first.
2. Fix (housekeeping fix) - operator-triggered, supervised. Runs the AUTO fixes in dependency order, narrating one line per step, then prints the JUDGMENT punch-list and a verify table.
3. There is no unattended rung. The OS does not run while you sleep (that is the design, not a gap). If you sync your OS to a remote and run cloud routines, a weekly remote "housekeeping fix" is a natural candidate - your call, set up via the routines skill's remote upgrade path.

Pre-flight

If core/identity.md does not exist, stop with: Founder OS not set up here. Run /founder-os:setup first. Do not fabricate a report.

If a detector's input is missing or a script errors, do not silently drop the line - mark it UNKNOWN - <detector> unavailable (<reason>). An honest partial sweep beats a confident one built on broken inputs.

The detector set

Route AUTO = reversible, on the approval-gate Auto list, fix mode may run it. Route JUDGMENT = needs the operator's call; the skill lists it and never decides it.

Detect mode (default)

Read-only, deterministic, cheap: file reads plus three scripts (memory-diff.py, skill_health.py --json, and the date math). Render exactly this and nothing else:

HOUSEKEEPING - DETECT - <YYYY-MM-DD HH:MM>
---
AUTO (safe to fix - say "housekeeping fix" or run the command per line):
  [<severity>] <debt class> - <one-line state> -> <fix>
  (none)
JUDGMENT (your call - command given, decision is yours):
  [<severity>] <debt class> - <one-line state> -> <fix>
  (none)
UNKNOWN (detector unavailable):        [omit when all detectors ran]
  <debt class> - <reason>
---
<N> AUTO, <M> JUDGMENT. Say "housekeeping fix" to clear the AUTO half.

Severity: HIGH (cadence stale, decay-due, quarantine ACTIVE), MED (rants aging, memory gaps, broken links), LOW (log bloat, graph stale, snapshot stale, skill health). One line per item - except skill health, which can return many findings at once: render dead pointers one line each (they route AUTO and are few) and the description-bloat findings as ONE summary line with the count and the command that reveals them. A stated summary, not a silent cap.

Fix mode ("housekeeping fix")

Operator-triggered, supervised. Run the AUTO fixes in this dependency order, narrating each step in one line. Do not re-prompt per item (every AUTO item is on the approval-gate Auto list); do not touch JUDGMENT items.

1. Daily anchor bump - cheap, and everything below reads dates.
2. Rants - /dream if any are aging.
3. Log bloat - python scripts/log-archive.py.
4. Dead skill pointers - apply a repoint only when the correct target is unambiguous (the name exists at a clear alternate path, e.g. a command instead of a skill). Show each one-line edit. No obvious target: move it to the punch-list, never blind-remove.
5. Wiki graph - /wiki-build, after any link edits so the graph reflects them.
6. Snapshot last - python scripts/brain-snapshot.py --write, because it captures everything above. Without --write the script only prints; the cache on disk stays stale.

Then print:

HOUSEKEEPING - FIX - <YYYY-MM-DD HH:MM>
---
RAN (AUTO):
  <debt class> -> <command> -> <result>
PUNCH-LIST (your call):
  [<severity>] <debt class> - <state> -> <fix>
VERIFY (did it actually happen):
  | Fix | Check | Result |
  | <debt class> | <how confirmed> | <PASS / UNVERIFIED> |
---
<N> AUTO cleared, <M> on your punch-list.

Fill the VERIFY table by re-reading the side-effect, not by trusting an exit code: the anchor header shows today's date, the rant frontmatter flipped to processed: true, relations.yaml carries a fresh timestamp, .snapshot.md is dated today. Mark any row you did not confirm as UNVERIFIED and never dress it up.

Rules

• Detect mode writes nothing, ever.
• Never run the weekly review inside fix mode - it forces judgment calls that belong to the operator.
• Never blind-fix a broken wikilink, and never create a hollow stub to silence a checker.
• If a referenced file does not exist on this install (brain/flags.md, system/quarantine.md), report the detector as UNKNOWN with the reason - do not report zero debt you did not check.
• No em dashes or en dashes. Hyphens only.