sag-knowledgelisted

# sag-knowledge — reading the knowledge base via MCP `sag` SAG is a **read-model of Git** — an approved knowledge base (ADRs, PRDs, design docs, API contracts, runbooks, postmortems, research), not session memory. Source priority when there's a conflict: **current Code/Git > SAG > Honcho/session memory**. ## Retrieval funnel (escalate step by step, don't skip) ```text list_sources # confirm which sources exist, source_id → list_documents(source_id?) # view documents in scope → outline(document_id) # view heading structure before reading deeply → search(query) | grep(pattern) # search=semantic, grep=exact match (identifiers, function names) → get_chunk(chunk_id) | read(...) # fetch verbatim text for ONLY the section needed, don't read the whole file → get_entity(name) # when you need to clarify a specific entity ``` **Default budget per lookup pass**: at most 1 `search` + 2 `outline` + 3 `get_chunk` before concluding "insufficient evidence" and reporting back to the user instead of guessing. **Never `read` the full text of a document** unless it's under ~5KB — use `outline` + `get_chunk` to fetch exactly the part needed. Use `grep`, not `search`, when you need: function/variable names, ticket IDs, error codes, API endpoints, or any technical identifier string that needs an EXACT match — semantic search can miss these due to rephrasing. ## Citation — use the path, not document_id/chunk_id SAG has no