Skill Sweep — Every Skill, Until A Pass Finds Nothing

The runnable form of the skill pass over the whole tree. What one skill's pass does is the skill-authoring skill's (references/skill-pass.md), and every rule a pass checks is that skill's too; this page is the order the whole run takes and what it reports. Rerun it whenever the rules change, a newer model arrives, or the ledger shows open rows — a run that finds nothing is the result it exists to reach.

Settled — do not re-propose

• Fanning the skills out to subagents. A reading pass is priced by files read, and a finding one skill teaches applies to the next; the run stays in the main session, one skill at a time (the model-delegation skill, references/reading-passes.md).
• Reading by grep or by skimming. The findings that matter — a claim the code no longer bears out, a copy between pages — live in prose a grep does not reach; every SKILL.md and every references/ page is read whole.
• An installed third-party skill in the run. A skill skills-lock.json installs is upstream's, outside the docs/skills ledger, and edited by nobody here.
• Fixing the same kind of finding by hand twice. The second one is an enforcer (the sweeps skill, references/handing-to-an-enforcer.md).

The run

1. Setup. git pull --rebase over a clean tree, then pnpm ai:sweep:ledger-coverage to date the rows from the trailers (the sweeps skill, references/ledger-files.md).
2. Mechanical checks first, and fix what they report before reading prose: pnpm ai:sweep:skill-docs, pnpm ai:sweep:duplicate-prose, and from scripts/ pnpm test src/workspace --run — citations, ledger units, package scripts. A timeout on a cold cache reruns the one file alone before it is read as a failure.
3. Per skill, in the main session, each through the skill-authoring skill's pass (references/skill-pass.md):
   • read SKILL.md and every references/ page whole;
   • check it against the audit below;
   • verify every code claim against the tree — identifiers, paths, signatures and example code as they exist now, written in the newest APIs and the error-handling skill's result chains;
   • fix what the audit finds, code included when the claim is right and the code is wrong, in one commit per skill carrying Ledger: docs/skills | ``; a clean skill's trailer rides the next commit, never an empty one.
4. Finish. pnpm ai:sweep:ledger-coverage again and commit the ledger dates; run the finishing skill; then the checks in the background per the running-checks skill, with repairs committed behind the units; commit by pathspec, git pull --rebase over a clean tree and a plain push (the review-queue skill).
5. Report a table of every skill — clean or fixed, with one line on what changed — and whether the run reached zero findings. A run that did not ends with the prompt for the next: this skill, rerun.

The audit

A skill may owe	Owner
One topic per page, and a mixed page split	skill-authoring (references/splitting-a-skill.md)
SKILL.md as one-line rules plus an index whose lines name the trigger	skill-authoring (references/splitting-a-skill.md)
A page opening Read when …	skill-authoring (references/splitting-a-skill.md)
A description opening Apply when …, naming the domain and never indexing the body	skill-authoring (references/frontmatter.md)
Citations as backticked repo paths, another skill by name plus page	skill-authoring (references/citations.md)
No positional pointer that does not name its target	skill-authoring (references/skill-pass.md)
Magnitudes, never counts; no roster or hand copy of a derived fact	skill-authoring (references/derived-surfaces.md)
No restatement of a rule an enforcer owns	skill-authoring (references/enforced-rules.md)
No one-off and no history	skill-authoring (references/what-belongs.md)
A Settled list first	skill-authoring (references/settled-lists.md)
An exception naming the forcing agent outside our control	skill-authoring (references/exceptions.md)
One owner per rule across skills	skill-authoring (references/one-owner-per-topic.md)

Every row is answered per skill, and "nothing owed" is an answer.

After a clean run

The ledger resweep is next: pnpm ai:sweep:ledger-coverage lists each ledger's open rows and the entries it could not match, and the unmatched entries are cleaned up first (the sweeps skill, references/ledger-files.md).