mercury

Apply a fresh-context reviewer's critique to the current work through a bounded, validated, human-gated loop. Use after a second-opinion review returns a critique — parse it, plan a few scoped fixes, run the build/tests and memory-lint, then summarize what was applied versus rejected and why.

Safely move a decayed fact or completed thread from continuity.md into the archive — the mechanical half of a memory review, done deterministically. Use during the REVIEW.md ritual (step 4) or whenever a specific fact must be archived, instead of hand-editing continuity.md (which has truncated the file twice). The agent decides what to archive; this performs the move safely.

Re-scan the repo's human-authored documentation (docs trees, ADRs, decision logs, design specs, roadmaps, kanban) and fold newly-durable facts into the shared memory/ layer — additively, on demand. Use when docs have grown or changed since enable, or to backfill a repo enabled before the curious harvest existed. Distills into neutral memory; never overwrites curated facts.

Deterministic integrity check for the agent-memory layer. Use after a memory review, before committing memory/ changes, or in CI to catch decay miscounts — facts archived while still referenced, an id in both continuity and the archive, tier or supersession drift. The agent judges meaning; the script does the counting.

Recompute continuity.md fact footers (last_used, uses, tier) from the session reference log — REVIEW.md steps 2–3, done deterministically. Use during a review, or whenever memory-lint reports [stale-metadata], instead of updating footers by hand. Pure arithmetic from the references; it never decides what to archive.

Generate a fresh-context review snapshot of the current task for a clean-memory AI reviewer (any vendor or a clean session). Use at a milestone that looks complete, or when blocked, uncertain, or facing a risky change, to get an independent second opinion before proceeding; shows a security advisory before exporting any state.

Regenerate the per-vendor skill adapters from the committed agent-skills/ layer. Run after authoring or editing a skill, or after a clone or pull, so Claude, Gemini, Cursor, Kiro, GitHub Copilot, and Google Antigravity each get a current native adapter.