Fix Review Warnings Sweep

Type: types/instruction.md

Steps

1. Build the work queue

commonplace-review status --fresh --outcome warn fail --json

Each row in pairs is one current warning: a (note, criterion) pair whose fresh result has outcome warn or fail, with criterion-id and result, the path of its result file. Group the rows by note and order the notes by row count, descending. If the user did not name a scope and the queue spans more than one collection, report the counts and ask which notes or collections to take first; to filter, pass note paths or directories before the options.

If pairs is empty, stop — there are no current warn findings to fix. Rows for stale results are excluded by --fresh; report them as needing re-review if the user asked about them, and do not fix from their text.

2. Execute or delegate

Run a single-note queue locally unless several notes have disjoint note and report paths and a fresh per-note context or parallel capacity gives a specific benefit. When that condition holds, launch one single-use worker per note with this complete packet:

• run fix-review-warnings.md (give the worker its absolute path, resolved from this file's location) on the exact {note-path}, skipping its re-review step;
• own only that note and kb/reports/state/fixes/{note-stem}.fix-report.md;
• read the inputs authorized by that instruction and no unrelated notes;
• do not use another orchestration skill, and do not run commonplace-review run;
• validate the note and return its diff summary, report path, and validation;
• defer and return the exact substantive choice when the instruction's edit boundary would be crossed.

The parent owns queue selection and collision checks. After each worker returns, verify its note diff, report, and validation, then close, terminate, or release it before dispatching more work. Stop on missing or partial output; workers are single-use and receive no follow-up task.

3. Re-review the edited notes

Run one gate sweep over every note that changed, asking for all gates:

commonplace-review run gates {edited-notes...} --all-gates --grouping bundle

The edits stale every pair whose watched parts changed, including pairs that passed before, and the sweep asks exactly those, skipping fresh results. Read its report as run review batches describes.

4. Report

After the re-review, report: - Fixed by strategy: count of fixes per taxonomy strategy name - Rejected: findings judged spurious or inapplicable, with evidence - Deferred: exact claim/evidence/source choices, affected passages, acceptable responses, and why the worker could not select one - Re-reviewed: each re-asked pair's new outcome, and any pair the sweep left without a current result - New patterns: any new-pattern classifications

5. Evolve taxonomy

If new patterns recur (3+ instances of the same pattern), propose adding them to the fix strategy taxonomy. Present the proposed entry to the user before adding.