sdlc-recall

Run knowledge_search with several differently phrased queries, then read the full nodes with document_read before citing them. Vary wording the way the original author might have written it, and search without a type filter when you mean everything the team knows.

knowledge_search fuses full-text relevance, tag overlap, and initiative evidence into a score, returning matched excerpts with provenance. Filters include query, type, status, initiative, flow, tags, include_superseded, and limit, but type spans two corpora so filtering on one silently hides the other.

The index uses PostgreSQL's english configuration, which splits CJK on whitespace, so unspaced Chinese becomes one token that queries cannot match. Search Chinese topics in English too, prefer one Chinese word per query, and report empty Chinese results as inconclusive retrieval rather than no prior learning.

document_read needs the initiative and path joined, such as _knowledge/nodes/0010-….md, not the path alone. For nodes on the platform shelf, you must also pass scope: "platform" or the read fails against the wrong store.

No, a superseded decision is often the most valuable finding because it shows the team already tried that direction and moved on. Report it clearly marked with its superseded_by reference, excluding it only if the caller explicitly asked for current state only.

No, it is strictly read-only and never calls knowledge_add or knowledge_supersede. Recording new learnings is the handover stage's job, which runs after the initiative closes.