cqs 0.5.1

# 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).

## [0.5.1] - 2026-02-05


### Fixed

- Algorithm correctness: glob filter applied BEFORE heap in brute-force search (was producing wrong results)
- `note_weight=0` now correctly excludes notes from unified search (was only zeroing scores)
- Windows path extraction in brute-force search uses `origin` column instead of string splitting
- GPU-to-CPU fallback no longer double-windows chunks
- Atomic note replacement (single transaction instead of delete+insert)
- Error propagation: 6 silent error swallowing sites now propagate errors
- Non-finite score validation (NaN/infinity checks in cosine similarity and search filters)
- FTS5 name query: terms now quoted to prevent syntax errors
- Empty query guard for `search_by_name`
- `split_into_windows` returns Result instead of panicking via assert
- Store Drop: `catch_unwind` around `block_on` to prevent panic in async contexts
- Stdio transport: line reads capped at 1MB
- `follow_links(false)` on filesystem walker (prevents symlink loops)
- `.cq/` directory created with 0o700 permissions
- `parse_file_calls` file size guard matching `parse_file`
- HNSW `count_vectors` size guard matching `load()`
- SQL IN clause batching for `get_embeddings_by_hashes` (chunks of 500)
- SQLite cache_size reduced from 64MB to 16MB per connection
- Path normalization gaps fixed in call_graph, graph, stats, filesystem source

### Changed

- `strip_unc_prefix` deduplicated into shared `path_utils` module
- `load_hnsw_index` deduplicated into `HnswIndex::try_load()`
- `index_notes_from_file` deduplicated — CLI now calls `cqs::index_notes()`
- MCP JSON-RPC types restricted to `pub(crate)` visibility
- Regex in `sanitize_error_message` compiled once via `LazyLock`
- `EMBEDDING_DIM` consolidated to single constant in `lib.rs`
- MCP stats uses `count_vectors()` instead of full HNSW load
- `note_stats` returns named struct instead of tuple
- Pipeline call graph upserts batched into single transaction
- HTTP server logging: `eprintln!` replaced with `tracing`
- MCP search: timing span added for observability
- GPU/CPU thread termination now logged
- Error sanitization regex covers `/mnt/` paths
- Watch mode: mtime cached per-file for efficiency
- Batch metadata checks on Store::open (single query)
- Consolidated note_stats and call_stats into fewer queries
- Dead code removed from `cli::run()`
- HNSW save uses streaming checksum (BufReader)
- Model BLAKE3 checksums populated for E5-base-v2

### Added

- 15 new search tests (HNSW-guided, brute-force, glob, language, unified, FTS)
- Test count: 379 (no GPU) up from 364

### Documentation

- `lib.rs` language list updated (C, Java)
- HNSW params corrected (M=24, ef_search=100)
- Cache size corrected (32 not 100)
- Roadmap phase updated
- Chunk cap documented as 100 lines
- Architecture tree updated with CLI/MCP submodules

## [0.5.0] - 2026-02-05


### Added

- **C and Java language support** (#222)
  - tree-sitter-c and tree-sitter-java grammars
  - 7 languages total (Rust, Python, TypeScript, JavaScript, Go, C, Java)
- **Test coverage expansion** (#224)
  - 50 new tests across 6 modules (cagra, index, MCP tools, pipeline, CLI)
  - Total: 375 tests (GPU) / 364 (no GPU)

### Changed

- **Model evaluation complete** (#221)
  - E5-base-v2 confirmed as best option: 100% Recall@5 (50/50 eval queries)
- **Parser/registry consolidation** (#223)
  - parser.rs reduced from 1469 to 1056 lines (28% reduction)
  - Parser re-exports Language, ChunkType from language module

## [0.4.6] - 2026-02-05


### Added

- **Schema migration framework** (#188, #215)
  - Migrations run automatically when opening older indexes
  - Falls back to error if no migration path exists
  - Framework ready for future schema changes
- **CLI integration tests** (#206, #213)
  - 12 end-to-end tests using `assert_cmd`
  - Tests for init, index, search, stats, completions
- **Server transport tests** (#205, #213)
  - 3 tests for stdio transport (initialize, tools/list, invalid JSON)
- **Stress tests** (#207, #213)
  - 5 ignored tests for heavy load scenarios
  - Run with `cargo test --test stress_test -- --ignored`
- **`--api-key-file` option** for secure API key loading (#202, #213)
  - Reads key from file, keeps secret out of process list
  - Uses `zeroize` crate for secure memory wiping

### Changed

- **Lazy grammar loading** (#208, #213)
  - Tree-sitter queries compile on first use, not at startup
  - Reduces startup time by 50-200ms
- **Pipeline resource sharing** (#204, #213)
  - Store shared via `Arc` across pipeline threads
  - Single Tokio runtime instead of 3 separate ones
- Note search warning now logs at WARN level when hitting 1000-note limit (#203, #213)

### Fixed

- **Atomic HNSW writes** (#186, #213)
  - Uses temp directory + rename pattern for crash safety
  - All 4 files written atomically together
- CLI test serialization to prevent HuggingFace Hub lock contention in CI

## [0.4.5] - 2026-02-05


### Added

- **20-category audit complete** - All P1-P4 items addressed (#199, #200, #201, #209)
  - ~243 findings across security, correctness, maintainability, and test coverage
  - Future improvements tracked in issues #202-208

### Changed

- FTS errors now propagate instead of silently failing (#201)
- Note scan capped at 1000 entries for memory safety (#201)
- HNSW build progress logging shows chunk/note breakdown (#201)

### Fixed

- Unicode/emoji handling in FTS5 search (#201)
- Go return type extraction for multiple returns (#201)
- CAGRA batch progress logging (#201)

## [0.4.4] - 2026-02-05


### Added

- **`note_weight` parameter** for controlling note prominence in search results (#183)
  - CLI: `--note-weight 0.5` (0.0-1.0, default 1.0)
  - MCP: `note_weight` parameter in cqs_search
  - Lower values make notes rank below code with similar semantic scores

### Changed

- CAGRA GPU index now uses streaming embeddings and includes notes (#180)
- Removed dead `search_unified()` function (#182) - only `search_unified_with_index()` was used

## [0.4.3] - 2026-02-05


### Added

- **Streaming HNSW build** for large repos (#107)
  - `Store::embedding_batches()` streams embeddings in 10k batches via LIMIT/OFFSET
  - `HnswIndex::build_batched()` builds index incrementally
  - Memory: O(batch_size) instead of O(n) - ~30MB peak instead of ~300MB for 100k chunks
- **Notes in HNSW index** for O(log n) search (#103)
  - Note IDs prefixed with `note:` in unified HNSW index
  - `Store::note_embeddings()` and `search_notes_by_ids()` for indexed note search
  - Index output now shows: `HNSW index: N vectors (X chunks, Y notes)`

### Changed

- HNSW build moved after note indexing to include notes in unified index

### Fixed

- O(n) brute-force note search eliminated - now uses HNSW candidates

## [0.4.2] - 2026-02-05


### Added

- GPU failures counter in index summary output
- `VectorIndex::name()` method for HNSW/CAGRA identification
- `active_index` field in cqs_stats showing which vector index is in use

### Changed

- `Config::merge` renamed to `override_with` for clarity
- `Language::FromStr` now returns `ParserError::UnknownLanguage` (thiserror) instead of anyhow
- `--verbose` flag now sets tracing subscriber to debug level
- Note indexing logic deduplicated into shared `cqs::index_notes()` function

### Fixed

- `check_cq_version` now logs errors at debug level instead of silently discarding
- Doc comments added for `IndexStats`, `UnifiedResult`, `CURRENT_SCHEMA_VERSION`

## [0.4.1] - 2026-02-05


### Changed

- Updated crates.io keywords for discoverability: added `mcp-server`, `vector-search`
- Added GitHub topics: `model-context-protocol`, `ai-coding`, `vector-search`, `onnx`

## [0.4.0] - 2026-02-05


### Added

- **Definition search mode** (`name_only`) for cqs_search (#165)
  - Use `name_only=true` for "where is X defined?" queries
  - Skips semantic embedding, searches function/struct names directly
  - Scoring: exact match 1.0, prefix 0.9, contains 0.7
  - Faster than glob for definition lookups
- `count_vectors()` method for fast HNSW stats without loading full index

### Changed

- CLI refactoring: extracted `watch.rs` from `mod.rs` (274 lines)
  - `cli/mod.rs` reduced from 2167 to 1893 lines

### Fixed

- P2 audit fixes (PRs #161-163):
  - HNSW checksum efficiency (hash from memory, not re-read file)
  - TOML injection prevention in note mentions
  - Memory caps for watch mode and note parsing (10k limits)
  - Platform-specific libc dependency (cfg(unix))

## [0.3.0] - 2026-02-04


### Added

- `cqs_audit_mode` MCP tool for bias-free code reviews (#101)
  - Excludes notes from search/read results during audits
  - Auto-expires after configurable duration (default 30m)
- Error path test coverage (#126, #149)
  - HNSW corruption tests: checksum mismatch, truncation, missing files
  - Schema validation tests: future/old version rejection, model mismatch
  - MCP edge cases: unicode queries, concurrent requests, nested JSON
- Unit tests for embedder.rs and cli.rs (#62, #132)
  - `pad_2d_i64` edge cases (4 tests)
  - `EmbedderError` display formatting (2 tests)
  - `apply_config_defaults` behavior (3 tests)
  - `ExitCode` values (1 test)
- Doc comments for CLI command functions (#70, #137)
- Test helper module `tests/common/mod.rs` (#137)
  - `TestStore` for automatic temp directory setup
  - `test_chunk()` and `mock_embedding()` utilities

### Changed

- Refactored `cmd_serve` to use `ServeConfig` struct (#138)
  - Removes clippy `too_many_arguments` warning
- Removed unused `ExitCode` variants (`IndexMissing`, `ModelMissing`) (#138)
- **Refactored Store module** (#125, #133): Split 1,916-line god object into focused modules
  - `src/store/mod.rs` (468 lines) - Store struct, open/init, FTS5, RRF
  - `src/store/chunks.rs` (352 lines) - Chunk CRUD operations
  - `src/store/notes.rs` (197 lines) - Note CRUD and search
  - `src/store/calls.rs` (220 lines) - Call graph storage/queries
  - `src/store/helpers.rs` (245 lines) - Types, embedding conversion
  - `src/search.rs` (531 lines) - Search algorithms, scoring
  - Largest file reduced from 1,916 to 531 lines (3.6x reduction)

### Fixed

- **CRITICAL**: MCP server concurrency issues (#128)
  - Embedder: `Option<T>` → `OnceLock<T>` for thread-safe lazy init
  - Audit mode: direct field → `Mutex<T>` for safe concurrent access
  - HTTP handler: `write()` → `read()` lock (concurrent reads safe)
- `name_match_score` now preserves camelCase boundaries (#131, #133)
  - Tokenizes before lowercasing instead of after

### Closed Issues

- #62, #70, #101, #102-#114, #121-#126, #142-#146, #148

## [0.2.1] - 2026-02-04


### Added

- Minimum Supported Rust Version (MSRV) declared: 1.88 (required by `ort` dependency)
- `homepage` and `readme` fields in Cargo.toml

### Changed

- Exclude internal files from crate package (AI context, audit docs, dev tooling)

## [0.2.0] - 2026-02-03


### Security

- **CRITICAL**: Fixed timing attack in API key validation using `subtle::ConstantTimeEq`
- Removed `rsa` vulnerability (RUSTSEC-2023-0071) by disabling unused sqlx default features

### Added

- IPv6 localhost support in origin validation (`http://[::1]`, `https://[::1]`)
- Property-based tests (9 total) for RRF fusion, embedder normalization, search bounds
- Fuzz tests (17 total) across nl.rs, note.rs, store.rs, mcp.rs for parser robustness
- MCP protocol edge case tests (malformed JSON-RPC, oversized payloads, unicode)
- FTS5 special character tests (wildcards, quotes, colons)
- Expanded SECURITY.md with threat model, trust boundaries, attack surface documentation
- Discrete sentiment scale documentation in CLAUDE.md

### Changed

- Split cli.rs into cli/ module (mod.rs + display.rs) for maintainability
- Test count: 75 → 162 (2x+ increase)
- `proptest` added to dev-dependencies

### Fixed

- RRF score bound calculation (duplicates can boost scores above naive maximum)
- `unwrap()` → `expect()` with descriptive messages (10 locations)
- CAGRA initialization returns empty vec instead of panic on failure
- Symlink logging in embedder (warns instead of silently skipping)
- clamp fix in `get_chunk_by_id` for edge cases

### Closed Issues

- #64, #66, #67, #68, #69, #74, #75, #76, #77, #78, #79, #80, #81, #82, #83, #84, #85, #86

## [0.1.18] - 2026-02-03


### Added

- `--api-key` flag and `CQS_API_KEY` env var for HTTP transport authentication
  - Required for non-localhost network exposure
  - Constant-time comparison to prevent timing attacks
- `--bind` flag to specify listen address (default: 127.0.0.1)
  - Non-localhost binding requires `--dangerously-allow-network-bind` and `--api-key`

### Changed

- Migrated from rusqlite to sqlx async SQLite (schema v10)
- Extracted validation functions for better code discoverability
  - `validate_api_key`, `validate_origin_header`, `validate_query_length`
  - `verify_hnsw_checksums` with extension allowlist
- Replaced `unwrap()` with `expect()` for better panic messages
- Added SAFETY comments to all unsafe blocks

### Fixed

- Path traversal vulnerability in HNSW checksum verification
- Integer overflow in saturating i64→u32 casts for database fields

### Security

- Updated `bytes` to 1.11.1 (RUSTSEC-2026-0007 integer overflow fix)
- HNSW checksum verification now validates extensions against allowlist

## [0.1.17] - 2026-02-01


### Added

- `--gpu` flag for `cqs serve` to enable GPU-accelerated query embedding
  - CPU (default): cold 0.52s, warm 22ms
  - GPU: cold 1.15s, warm 12ms (~45% faster warm queries)

### Changed
- Hybrid CAGRA/HNSW startup: HNSW loads instantly (~30ms), CAGRA builds in background
  - Server ready immediately, upgrades to GPU index transparently
  - Eliminates 1.2s blocking startup delay

### Fixed

- Search results now prioritize code over notes (60/40 split)
  - Notes enhance but don't dominate results
  - Reserve 60% of slots for code, notes fill the rest

## [0.1.16] - 2026-02-01


### Added

- Tracing spans for major operations (`cmd_index`, `cmd_query`, `embed_batch`, `search_filtered`)
- Version check warning when index was created by different cqs version
- `Embedding` type encapsulation with `as_slice()`, `as_vec()`, `len()` methods

### Fixed

- README: Corrected call graph documentation (cross-file works, not within-file only)
- Bug report template: Updated version placeholder

### Documentation

- Added security doc comment for MCP origin validation behavior

## [0.1.15] - 2026-02-01


### Added

- Full call graph coverage for large functions (>100 lines)
  - Separate `function_calls` table captures all calls regardless of chunk size limits
  - CLI handlers like `cmd_index` now have call graph entries
  - 1889 calls captured vs ~200 previously

### Changed
- Schema version: 4 → 5 (requires `cqs index --force` to rebuild)

## [0.1.14] - 2026-01-31


### Added

- Call graph analysis (`cqs callers`, `cqs callees`)
  - Extract function call relationships from source code
  - Find what calls a function and what a function calls
  - MCP tools: `cqs_callers`, `cqs_callees`
  - tree-sitter queries for call extraction across all 5 languages

### Changed

- Schema version: 3 → 4 (adds `calls` table)

## [0.1.13] - 2026-01-31


### Added

- NL module extraction (src/nl.rs)
  - `generate_nl_description()` for code→NL→embed pipeline
  - `tokenize_identifier()` for camelCase/snake_case splitting
  - JSDoc parsing for JavaScript (@param, @returns tags)
- Eval improvements
  - Eval suite uses NL pipeline (matches production)
  - Runs in CI on tagged releases

## [0.1.12] - 2026-01-31


### Added

- Code→NL embedding pipeline (Greptile approach)
  - Embeds natural language descriptions instead of raw code
  - Generates: "A function named X. Takes parameters Y. Returns Z."
  - Doc comments prioritized as human-written NL
  - Identifier normalization: `parseConfig` → "parse config"

### Changed

- Schema version: 2 → 3 (requires `cqs index --force` to rebuild)

### Breaking Changes

- Existing indexes must be rebuilt with `--force`

## [0.1.11] - 2026-01-31


### Added

- MCP: `semantic_only` parameter to disable RRF hybrid search when needed
- MCP: HNSW index status in `cqs_stats` output

### Changed

- tree-sitter-rust: 0.23 -> 0.24
- tree-sitter-python: 0.23 -> 0.25
- Raised brute-force warning threshold from 50k to 100k chunks

### Documentation

- Simplified CLAUDE.md and tears system
- Added docs/SCARS.md for failed approaches
- Consolidated PROJECT_CONTINUITY.md (removed dated files)

## [0.1.10] - 2026-01-31


### Added

- RRF (Reciprocal Rank Fusion) hybrid search combining semantic + FTS5 keyword search
- FTS5 virtual table for full-text keyword search
- `normalize_for_fts()` for splitting camelCase/snake_case identifiers into searchable words
- Chunk-level incremental indexing (skip re-embedding unchanged chunks via content_hash)
- `Store::get_embeddings_by_hashes()` for batch embedding lookup

### Changed

- Schema version bumped from 1 to 2 (FTS5 support)
- RRF enabled by default in CLI and MCP for improved recall

## [0.1.9] - 2026-01-31


### Added

- HNSW-guided filtered search (10-100x faster for filtered queries)
- SIMD-accelerated cosine similarity via simsimd crate
- Shell completion generation (`cqs completions bash/zsh/fish/powershell`)
- Config file support (`.cqs.toml` in project, `~/.config/cqs/config.toml` for user)
- Lock file with PID for stale lock detection
- Rustdoc documentation for public API

### Changed

- Error messages now include actionable hints
- Improved unknown language/tool error messages

## [0.1.8] - 2026-01-31


### Added

- HNSW index for O(log n) search on large codebases (>50k chunks)
- Automatic HNSW index build after indexing
- Query embedding LRU cache (32 entries)

### Fixed

- RwLock poison recovery in HTTP handler
- LRU cache poison recovery in embedder
- Query length validation (8KB max)
- Embedding byte validation with warning

## [0.1.7] - 2026-01-31


### Fixed

- Removed `Parser::default()` panic risk
- Added logging for silent search errors
- Clarified embedder unwrap with expect()
- Added parse error logging in watch mode
- Added 100KB chunk byte limit (handles minified files)
- Graceful HTTP shutdown with Ctrl+C handler
- Protocol version constant consistency

## [0.1.6] - 2026-01-31


### Added

- Connection pooling with r2d2-sqlite (4 max connections)
- Request body limit (1MB) via tower middleware
- Secure UUID generation (timestamp + random)

### Fixed

- lru crate vulnerability (0.12 -> 0.16, GHSA-rhfx-m35p-ff5j)

### Changed

- Store methods now take `&self` instead of `&mut self`

## [0.1.5] - 2026-01-31


### Added

- SSE stream support via GET /mcp
- GitHub Actions CI workflow (build, test, clippy, fmt)
- Issue templates for bug reports and feature requests
- GitHub releases with changelogs

## [0.1.4] - 2026-01-31


### Changed

- MCP 2025-11-25 compliance (Origin validation, Protocol-Version header)
- Batching removed per MCP spec update

## [0.1.3] - 2026-01-31


### Added

- Watch mode (`cqs watch`) with debounce
- HTTP transport (MCP Streamable HTTP spec)
- .gitignore support via ignore crate

### Changed

- CLI restructured (query as positional arg, flags work anywhere)
- Replaced walkdir with ignore crate

### Fixed

- Compiler warnings

## [0.1.2] - 2026-01-31


### Added

- New chunk types: Class, Struct, Enum, Trait, Interface, Constant
- Hybrid search with `--name-boost` flag
- Context display with `-C N` flag
- Doc comments included in embeddings

## [0.1.1] - 2026-01-31


### Fixed

- Path pattern filtering (relative paths)
- Invalid language error handling

## [0.1.0] - 2026-01-31


### Added

- Initial release
- Semantic code search for 5 languages (Rust, Python, TypeScript, JavaScript, Go)
- tree-sitter parsing for function/method extraction
- nomic-embed-text-v1.5 embeddings (768-dim) [later changed to E5-base-v2 in v0.1.16]
- GPU acceleration (CUDA/TensorRT) with CPU fallback
- SQLite storage with WAL mode
- MCP server (stdio transport)
- CLI commands: init, doctor, index, stats, serve
- Filter by language (`-l`) and path pattern (`-p`)

[0.5.1]: https://github.com/jamie8johnson/cqs/compare/v0.5.0...v0.5.1
[0.5.0]: https://github.com/jamie8johnson/cqs/compare/v0.4.6...v0.5.0
[0.4.6]: https://github.com/jamie8johnson/cqs/compare/v0.4.5...v0.4.6
[0.4.5]: https://github.com/jamie8johnson/cqs/compare/v0.4.4...v0.4.5
[0.4.4]: https://github.com/jamie8johnson/cqs/compare/v0.4.3...v0.4.4
[0.4.3]: https://github.com/jamie8johnson/cqs/compare/v0.4.2...v0.4.3
[0.4.2]: https://github.com/jamie8johnson/cqs/compare/v0.4.1...v0.4.2
[0.4.1]: https://github.com/jamie8johnson/cqs/compare/v0.4.0...v0.4.1
[0.4.0]: https://github.com/jamie8johnson/cqs/compare/v0.3.0...v0.4.0
[0.3.0]: https://github.com/jamie8johnson/cqs/compare/v0.2.1...v0.3.0
[0.2.1]: https://github.com/jamie8johnson/cqs/compare/v0.2.0...v0.2.1
[0.2.0]: https://github.com/jamie8johnson/cqs/compare/v0.1.18...v0.2.0
[0.1.18]: https://github.com/jamie8johnson/cqs/compare/v0.1.17...v0.1.18
[0.1.17]: https://github.com/jamie8johnson/cqs/compare/v0.1.16...v0.1.17
[0.1.16]: https://github.com/jamie8johnson/cqs/compare/v0.1.15...v0.1.16
[0.1.15]: https://github.com/jamie8johnson/cqs/compare/v0.1.14...v0.1.15
[0.1.14]: https://github.com/jamie8johnson/cqs/compare/v0.1.13...v0.1.14
[0.1.13]: https://github.com/jamie8johnson/cqs/compare/v0.1.12...v0.1.13
[0.1.12]: https://github.com/jamie8johnson/cqs/compare/v0.1.11...v0.1.12
[0.1.11]: https://github.com/jamie8johnson/cqs/compare/v0.1.10...v0.1.11
[0.1.10]: https://github.com/jamie8johnson/cqs/compare/v0.1.9...v0.1.10
[0.1.9]: https://github.com/jamie8johnson/cqs/compare/v0.1.8...v0.1.9
[0.1.8]: https://github.com/jamie8johnson/cqs/compare/v0.1.7...v0.1.8
[0.1.7]: https://github.com/jamie8johnson/cqs/compare/v0.1.6...v0.1.7
[0.1.6]: https://github.com/jamie8johnson/cqs/compare/v0.1.5...v0.1.6
[0.1.5]: https://github.com/jamie8johnson/cqs/compare/v0.1.4...v0.1.5
[0.1.4]: https://github.com/jamie8johnson/cqs/compare/v0.1.3...v0.1.4
[0.1.3]: https://github.com/jamie8johnson/cqs/compare/v0.1.2...v0.1.3
[0.1.2]: https://github.com/jamie8johnson/cqs/compare/v0.1.1...v0.1.2
[0.1.1]: https://github.com/jamie8johnson/cqs/compare/v0.1.0...v0.1.1
[0.1.0]: https://github.com/jamie8johnson/cqs/releases/tag/v0.1.0