recall-contextlisted

Use at the start of orchestrator turns and before dispatching specialists. Queries the Cognee KG+RAG MCP server for the current repo's dataset and returns structured context (decisions, runbooks, glossary terms, adjacent specs) that the orchestrator folds into the next subagent's prompt. Do NOT use inside leaf executors — the orchestrator pre-loads recall and passes it down.
syndberg/asandberg-harness · ★ 1 · AI & Automation · score 67

Install: claude install-skill syndberg/asandberg-harness

# recall-context Surface relevant prior context from the centralized Cognee store before deciding what to do next. Backend stack: - **LLM** (graph extraction during ingest + GRAPH_COMPLETION at query): configured in `~/.cognee/.env`. Default is Anthropic Haiku via `$ANTHROPIC_API_KEY`. - **Embeddings**: configured in `~/.cognee/.env`. Default is local Ollama `nomic-embed-text`; Ollama uses your GPU automatically if one is available. - **MCP server**: `cognee` (registered in `~/.claude.json`), runs `cognee-shim.sh cognee-mcp --transport stdio`. ## When to use - Beginning of a session where the user references prior work ("the auth thing", "what we decided last week"). - Right before `/spec.implement` dispatches the first specialist — pre-load decisions and glossary. - Right before `/spec.reconcile` — surface adjacent specs that the impl might have stepped on. ## When NOT to use - Inside `spec-implementer`, `test-runner`, `spec-conformance-evaluator`, or any leaf executor. They receive recall *from* the orchestrator and must not query the graph themselves. (This keeps depth = 1 honest and prevents recall floods.) - For trivial single-file lookups the user can grep faster. ## How to call Primary tool: `mcp__cognee__search`. Arguments (Cognee 1.0 surface): ```json { "query": "<the actual question, in natural language>", "query_type": "GRAPH_COMPLETION", "datasets": ["<basename of pwd repo>"] } ``` - `query_type=GRAPH_COMPLETION` returns a synthesized answer from t