MCP knowledge tools

Module: apps/mcp/src/tools/knowledge/ Scopes: knowledge:read for search_knowledge, knowledge:write for store_knowledge and supersede_knowledge.

`store_knowledge`

Argument	Type	Notes
`title`	`string`, 1–200
`body`	`string`, 1–32000
`tags`	`string[≤16]?`	lowercase kebab-case enforced; mixed-case input is normalised
`confidence`	`0..100?`	default 80
`expiresAt`	`ISO-8601?`	must be strictly in the future
`taskRef`	`{ taskId, kind: 'references' \	'produced' }?`	wires a `task_knowledge` edge
`clientToken`	`uuid?`	60s idempotence key in Redis (`idem:k.store:{org}:{agent}:{token}`)

Body is inserted with source='agent' and created_by_agent_id from the auth context. The embedding generation job runs out-of-band (BullMQ embeddings queue); the response carries embeddingStatus: 'pending'.

Quota: knowledge_stored is checked before insert and incremented after commit.

`search_knowledge`

Argument	Type	Notes
`query`	`string`, 1–2000
`limit`	`1..50?`	default 10
`tags`	`string[≤16]?`	lowercased before query
`includeSuperseded`	`boolean?`	default `false`
`includeExpired`	`boolean?`	default `false`
`minScore`	`0..1?`	filter hits with fused score below
`efSearch`	`1..200?`	HNSW probe width (default 64)

Hybrid search = pgvector cosine (embedding <=> queryVec) plus tsvector @@ plainto_tsquery('simple', query) plus title ILIKE %query%, fused by RRF (k=60). Returns a snippet (240 chars centred on the first match) — fetch the full body via the task-29 resource read once it lands.

`supersede_knowledge`

Argument	Type	Notes
`oldId`	`uuid`	will be marked superseded
`newId`	`uuid`	replacement

Authorisation: the calling agent must own at least one of the two entries. Cross-author supersession requires a future knowledge:write_others scope (not in today's catalogue). Wraps supersedeKnowledge from @drobek/knowledge (task 21) — it handles cycle detection and updates the transitive superseded_by_id cache so search short-circuits stay correct.