semantic-code-mcp

# Changelog All notable changes to this project will be documented in this file. The format is based on [Keep a Changelog](https://keepachangelog.com/en/1.1.0/), and this project adheres to [Semantic Versioning](https://semver.org/spec/v2.0.0.html). ## [0.4.0] - 2026-02-01 ### Added - Markdown support — indexes `.md` files by heading-based sections (`ChunkType.section`) - Setext heading support (`Title\n=====` syntax) in addition to ATX headings (`# Title`) - `ChunkType.section` enum value for document sections (backward-compatible with existing indexes) - YAML frontmatter correctly excluded from chunks - Decision doc: `docs/decisions/005-markdown-chunking.md` ### Changed - `tree-sitter-markdown` added as dependency - `CompositeChunker` now routes `.md` files to `MarkdownChunker` ## [0.3.0] - 2026-01-31 ### Added - Multi-language chunker architecture (`BaseTreeSitterChunker` + `CompositeChunker` dispatcher) - Rust support — functions, structs, enums, traits, impl blocks, `//!` module doc comments - Published to PyPI — installable via `uvx semantic-code-mcp` - GitHub Actions workflow for automated publishing on tag push (trusted publishers OIDC) - Platform-specific install docs (macOS/Windows vs Linux CPU-only torch) - "Adding a New Language" guide in README ### Changed - Flattened package — `chunkers/` and `embedder.py` at top level, `indexer/` collapsed to single module - `IndexService` owns scanning, change detection, chunking, status; `Indexer` handles embedding and storage only - `CompositeChunker` with extension collision detection (renamed from `MultiLanguageChunker`) - `IndexService` takes `CompositeChunker` directly — no separate `supported_extensions` parameter - All enums standardized to `StrEnum` + `auto()` — `ChunkType`, Python `NodeType`, Rust `NodeType` - `__init__.py` files are docstring-only (no re-exports); all imports use full module paths - Container uses `cached_property` for model/embedder singletons, per-project store caching ### Fixed - Tree-sitter `Parser` thread-safety — create fresh parser per `chunk_string` call (was shared, mutates on `parse()`) - `mock_embedder.embed_batch` returns correct embedding count via `side_effect` (was hardcoded to 1) - `Indexer` is now pure data pipeline — no `Settings`, no `FileChangeCache`, no `cache_dir`; all cache bookkeeping owned by `IndexService` - Clean shutdown on Ctrl+C (SIGINT handler instead of traceback) - Removed redundant `.value` calls on `StrEnum` members ## [0.1.0] - 2026-01-31 ### Added - `IndexService` orchestrating full index pipeline (scan → detect → chunk → embed) with progress callbacks - `SearchService` with auto-indexing via `IndexService` (replaces manual orchestration in server.py) - `services/` package replacing `search/` directory - `duration_seconds` on `IndexResult` for end-to-end timing - Strict ruff rules: C901, DTZ, ASYNC, SLF, PIE, T20, PERF, FURB, PLC0415 - ty type-checking rules in pre-commit and pyproject.toml - Profiling support with pyinstrument for dev (enable with `SEMANTIC_CODE_MCP_PROFILE=1`) - MCP server with three tools: `search_code`, `index_codebase`, `index_status` - Semantic code search using sentence-transformers embeddings (all-MiniLM-L6-v2) - LanceDB vector storage for embeddings - Module-level docstring chunking — conceptual queries now match files by their self-description (decision 003) - Tree-sitter based AST chunking for Python (functions, classes, methods, module docstrings) - Incremental indexing with mtime-based change detection - Debug timing info in search results (status_check_ms, embedding_ms, search_ms) - Hybrid search: keyword boost (up to 20%) and recency boost (up to 5% for files < 1 week old) - Score threshold filtering (< 0.3 filtered as noise) - Result truncation (> 50 lines shows "... truncated") - Results grouped by file for cleaner output - Pre-load embedding model at startup (avoids 2s cold start) - Parallel file chunking with asyncio.gather - Project documentation structure (CLAUDE.md, README.md, TODO.md, CHANGELOG.md) - Claude Code rules in `.claude/rules/` - Architecture decision records in `docs/decisions/` - Pre-commit hooks (ruff, bandit, conventional commits) ### Fixed - SQL injection in `delete_by_file` — escape single quotes in file paths - Force reindex now clears vector store to prevent duplicate results - Atomic cache writes (tempfile + rename) to prevent corruption on crash - FTS index failures logged at WARNING instead of DEBUG - Timezone-aware datetimes throughout (DTZ compliance) ### Changed - CPU-only PyTorch via `[tool.uv.sources]` — venv reduced from 7.8GB to 1.7GB (no CUDA/nvidia/triton) - Lazy `sentence-transformers` import — startup no longer loads torch (~4s saved) - `server.py` is now a thin tool layer delegating to `IndexService`/`SearchService` - `SearchOutcome.index_result` always present (default zeros) — eliminates None guards - `container.create_search_service()` reuses indexer instead of creating duplicates - Removed unused `status_cache_ttl` config option - Chunker complexity reduced by extracting `_extract_decorated` and `_extract_class_with_methods` - File scanning uses `git ls-files` for 100x speedup (falls back to os.walk for non-git repos) - Removed sync `index()` method - only async version remains (no code duplication) - Removed unused Searcher class - Skip FTS index rebuild if already exists (~80ms saved per search)