ce-compound-refresh

Maintains the quality of learning and pattern docs stored under docs/solutions/ by reviewing them against the current codebase and then updating, consolidating, replacing, or deleting any that have drifted. It is intended for requests like refreshing learnings, auditing docs/solutions/, cleaning up stale learnings, or consolidating overlapping docs, and it also responds when a related skill flags an older doc as superseded. It deliberately does not trigger for general refactor, debugging, or code-review work unless the user explicitly points at docs/solutions/.

The workflow runs in either interactive or headless mode. Interactive mode asks the user for decisions on ambiguous cases one question at a time using the platform's blocking question tool, leading with a recommendation and preferring multiple choice. Headless mode, selected by passing mode:headless in the arguments, performs no user interaction: it applies all unambiguous actions, records successful writes as applied and failed writes as recommended, marks genuinely ambiguous cases as stale with stale_reason and stale_date frontmatter, uses conservative confidence, and always produces a summary report split into Applied and Recommended sections. Docs are refreshed in a fixed order, reviewing individual learning docs first as the primary evidence, then any pattern docs derived from them, because stale learnings can make a pattern look more valid than it is.

Each candidate artifact is classified into one of five outcomes: Keep (accurate and useful, no edit by default), Update (correct solution but drifted references, apply evidence-backed in-place edits), Consolidate (merge overlapping correct docs into a canonical one and delete the subsumed doc), Replace (create a trustworthy successor then delete the misleading original), or Delete (remove a doc that is no longer useful or distinct, relying on git history for recovery). Core rules emphasize evidence informing judgment rather than a mechanical scorecard, preferring a no-write Keep over leaving review breadcrumbs, and matching docs to current reality. The skill also handles CONCEPTS.md bootstrap requests by disambiguating between building the repo-wide concept map and running a refresh cycle.