sherlock-second-brain AGENTS.md

AGENTS.md

Sherlock's second brain MCP server: validated knowledge lives in fiches/ and skills/; everything unvalidated lives in cases/ (JSON investigations — investigation files for debugging and analysis) or memories/ (standalone MD notes with YAML frontmatter). A resolved case is promoted into a fiche or skill via the MCP; a memory can also be promoted into a fiche. Repo language is English (code, docs, generated output) — the README.md is in English because it is parsed by Glama (https://glama.ai). Some tests keep French content (tests/test_vector.py, test_slugify) to exercise multilingual support.

Commands

uv run ruff check src/ tests/                    # lint (Astral, E,F,I,W,UP,B,ANN,PYI + preview)
uv run ty check                                  # type checking (Astral)
uv run pytest tests/ -v                          # all tests (no services needed, tmp_path fixtures)
uv run pytest tests/test_storage.py::test_slugify -v   # single test
uv sync                                           # installs deps (chromadb+fastembed inclus)

Required verification order: ruff check → ty check → pytest.

Run the server with uv run python -m sherlock_second_brain.server — this is a stdio MCP server, it needs an MCP client to be useful (wired as ~/.config/opencode/opencode.json → SHERLOCK_BRAIN_DATA_DIR=/opt/infra/kb).

Architecture (hexagonal)

• src/sherlock_second_brain/:
  • server.py — composition root: instantiates adapters+services, @mcp.tool() defs
  • domain/ — rich aggregates enforcing their invariants: models/case.py (Case factory + mutations + promote) and models/memory.py (Memory factory + setters + promote), one file per class (models/{case,hypothesis,step,evidence,memory,promotion}.py), plus errors.py (StorageError hierarchy), rules.py (VALID_STATUS, VALID_PROMOTION_TARGETS) and text.py (slugify, now_iso, CASE_ID_PATTERN, MEMORY_ID_PATTERN)
  • application/ — thin orchestration only (no business rules), depending on ports.py Protocols: case_service.py (CaseService), memory_service.py (MemoryService), promotion_service.py (PromotionService)
  • adapters/ — persistence/IO only, no domain policy: filesystem.py (Storage persists by object: save_*/get/list/delete/next_*_id/ write_evidence), chroma.py (VectorIndex + FastembedEmbeddingFunction multilingue), lexical.py (LexicalIndex), hybrid.py (HybridIndex RRF), dto/case_update.py (CaseUpdateFields for the case_update fields), frontmatter.py (memory MD + YAML frontmatter render/parse), templates/ (Jinja2: fiche.md.j2, skill.md.j2, memory_fiche.md.j2 rendered by PromotionService)
• Data layout under SHERLOCK_BRAIN_DATA_DIR (default ~/sherlock-second-brain-data):
  • cases/<case-id>/case.json + cases/<case-id>/evidence/* — case-id = case-YYYY-MM-DD-NNN (per-day counter)
  • memories/<id>.md — id = mem-YYYY-MM-DD-NNN (per-day counter), YAML frontmatter + body
  • fiches/<slug>.md, skills/<slug>/SKILL.md
  • vector/ — derived ChromaDB index, always rebuildable via the index_rebuild tool
• src/sherlock_second_brain/schema/case.schema.json is the source of truth: Storage._save_case validates every case against it (SCHEMA_PATH prefers the schema bundled inside the package, falling back to walking up to the repo root). If you touch the schema, keep writes compatible or test_written_case_matches_json_schema fails.
• Search is hybrid: server.py composes HybridIndex(VectorIndex, LexicalIndex) and fuses rankings by RRF. ChromaDB is a mandatory dependency (no fallback). The embedding model is multilingual (sentence-transformers/paraphrase-multilingual-MiniLM-L12-v2, downloaded at first index_rebuild). Doc IDs are fiche:<slug>, case:<id>, skill:<slug>, memory:<id>. Document enumeration is shared in adapters/documents.py.
• CaseService maintains the index on every mutation (update/evidence/status/delete); MemoryService maintains it on every mutation including create (a memory must be immediately searchable).