reviewing-memory

name

reviewing-memory

description

Invoke this skill as a deliberate review pass whenever stability matters: (1) before a release/migration/large refactor; (2) at the start of work in an unfamiliar subsystem; (3) when you see repeated agent confusion; (4) when thread-type working sets accumulate; (5) when you suspect stale guidance is causing defects. This is not ad-hoc maintenance; it is a structured audit to prevent systemic drift.

Review procedure:

1. Pull the highest-impact set first:

   • List by severity/weight with minimal filters: spec-driver list memories --limit N
   • Then run contextual reviews for the active area: spec-driver list memories -p <subsystem paths>... -c "<common commands>" --match-tag <subsystem tags>... --limit N

2. Explicitly include drafts only when reviewing for promotion:

   • Use --include-draft to evaluate whether drafts should become active or be removed/merged.

3. For each surfaced record, apply a strict checklist:

   • Provenance: does it cite authoritative sources (provenance.sources)? If not, add sources or reduce confidence.
   • Freshness: is verified recent enough for its type? If not, verify against code/docs now or schedule by setting review_by soon.
   • Metadata efficiency: remove low-value defaults and generated bloat (audience: [human, agent], unnecessary created, bulk links.out when not actively consumed).
   • Scope: does it surface under the correct --path/--command/--match-tag queries? If not, fix scope.*.
   • Visibility discipline: keep visibility: [pre] only for memories with strong scope.* and clear pre-hook value.
   • Actionability: does it contain executable steps/pointers (especially for pattern/signpost/system)? If not, rewrite to be procedural or convert to a signpost.
   • Duplication/conflict: does it overlap or contradict another memory? If yes, merge/supersede immediately.
   • Links: run spec-driver resolve links --dry-run to check for broken [[...]] references. Missing targets may indicate renamed/deleted artifacts; only persist resolved links when needed.

4. Thread hygiene (mandatory):

   • Threads are context-dependent and should not linger. Any thread that is not verified recently (or no longer relevant to current work) must be moved to archived, superseded, or converted into a durable fact/pattern/system record with proper scope and provenance.

5. Produce an outcome, not notes:

   • Every reviewed item must end in one of: verified (date updated), corrected (content/scope/provenance updated), superseded (replacement created/linked), archived/obsolete (removed from surfacing), or promoted from draft.

Stop conditions:

• Do not expand memories into mini-specs. If review uncovers missing design/decision authority, create or reference ADR/SPEC/design docs and convert the memory into a pointer/signpost to the canonical artifact.

Success metric:

• After review, the top results for common --path/--command/--match-tag queries should be accurate, source-backed, and recently verified, so future agents default to correct guidance without extra investigation.

---