Built-in Skills¶

Seven skills are installed into every enabled repo, marked provenance: agent-memory-builtin and tool-managed (overwritten on upgrade — fork or upstream to customize). They sort along the judgment-vs-arithmetic boundary: the deterministic moves are mechanized; the judgment stays with the agent.

Skill	Side of the boundary
`memory-lint`	arithmetic — read-only
`refresh-metadata`	arithmetic
`archive-fact`	arithmetic
`sync-adapters`	arithmetic
`harvest-knowledge`	judgment-assisted
`second-opinion`	judgment-assisted
`apply-critique`	judgment-assisted

A hello-world demo skill also ships in the tool's own repo but is not installed into targets.

memory-lint¶

The deterministic integrity verifier the review ritual relies on. Ships in both Python and Node at output parity, so it runs on a node-only box too. Nine checks:

duplicate facts ([both])
over-archived ([over-archived])
overdue archival ([overdue])
dangling supersession links ([dangling])
date-only session filenames ([date-only-session])
empty/malformed version manifest ([version-manifest])
leftover VCS merge-conflict markers ([conflict-marker])
review cadence + size ([review-overdue] / [continuity-bloat])
stale metadata — stored tier ≠ recomputed tier ([stale-metadata])

python agent-skills/memory-lint/scripts/memory-lint.py      # or
node   agent-skills/memory-lint/scripts/memory-lint.mjs

Read-only — it verifies, it never mutates. Errors fail CI; advisories warn.

refresh-metadata¶

Executes the review's deterministic re-tier (REVIEW.md steps 2–3): recomputes every fact's last_used / uses / tier from the ## Memory References across memory/sessions/ and writes the footers back (read into memory, write once). Pure arithmetic — core / superseded and never-referenced facts are untouched; it clamps at archive-candidate (it never archives — that's archive-fact plus the agent's judgment); pinned threads keep their tier. Python or Node, --dry-run, idempotent.

archive-fact¶

Performs the archive move safely, given fact id(s) the agent has decided to retire. It extracts each block (bullet through its footer), appends it to the quarter archive + INDEX.md, and rewrites continuity.md without it — reading the whole file into memory and writing once, so truncation is structurally impossible. Guards (refuse, exit 1): id with no footer, id already archived, or a move that would empty continuity. All-or-nothing; --dry-run to preview. It never decides what to archive (never-pick-a-winner).

sync-adapters¶

Regenerates the six vendor adapters from each neutral SKILL.md, pruning orphaned generated adapters (signature-guarded — never deletes a hand-authored file). Ships in bash, Node, and Python at byte-identical output parity; the bash runtime needs no install and is preferred.

bash agent-skills/sync-adapters/scripts/sync-adapters.sh --dry-run

harvest-knowledge¶

Re-scans the repo's human-authored documentation (recursive docs/ trees + a root sweep for ADRs, decision logs, design specs, roadmaps) and distils durable facts into the neutral memory/ layer additively: map-don't-mirror, check-existing-first (a re-run never duplicates), conflicts → a Contradiction thread, genuine replacements → supersession. Scoped incrementally by a last_harvest marker. Not a vendor /init — the output goes to memory/, never a vendor steering file.

second-opinion¶

Distils a compact snapshot from continuity.md + recent session logs (never a parallel state file) and, behind a security advisory you must acknowledge, hands it to a reviewer with clean memory — a fresh session or a different vendor that didn't live the work. The reviewer is a hypothesis generator, not an authority.

apply-critique¶

Feeds a returned critique through a bounded, validated, human-gated loop: a few scoped fixes → build/tests + memory-lint → an applied-vs-rejected summary. Deterministic checks and a human gate the result — the lesson baked in after a clean-context reviewer once over-archived still-referenced facts.

Zero overhead by default

Installed ≠ run. The built-ins sit in agent-skills/ until an agent's task matches one's description (or you invoke it explicitly).