Skip to content

v0.4.0: Contract Extractor, Route Mapper, and Project Topology

Choose a tag to compare

View all tags
@blackrain02 blackrain02 released this 30 Sep 12:39
· 19 commits to main since this release
v0.4.0
912d74a

Changelog

All notable changes to this project will be documented in this file.

The format is based on Keep a Changelog,
and this project adheres to Semantic Versioning.

[0.4.0] - 2026-09-30

Added

  • lg contract <file|component>: AST contract extractor for Vue SFCs (defineProps, defineEmits, defineModel, slots), PHP classes/interfaces (namespace, public properties, constructor promoted properties, and public method signatures), and TypeScript modules. Slashes file-reading token burn by up to 90% (<10ms execution).
  • lg route '<query>': Fast route-to-controller mapper. Scans Laravel routes/*.php and modular packages (vendor/bina/*/routes/), resolving route methods, URIs, route names, and exact controller target file line numbers in <30ms.
  • lg topo: Zero-token project topology card. Generates an ultra-dense, 150-token executive summary of framework stack (Laravel, PHP), active frontend dependencies (Vue, Inertia, Tailwind, Vite), database/queue drivers, modular monorepo packages, and entry points (<15ms).
  • MCP Server Expansion: Added localgrep_contract, localgrep_route, and localgrep_topo tools to mcp_server.py.
  • Automated test suite in tests/test_phase1.py covering contract, route, and topology extraction.

[0.3.3] - 2026-09-30

Added

  • Tier 1 Fast-Path Query Routing (<15ms): Automatically detects symbol, component, identifier, and PascalCase/camelCase queries. When high-confidence file or symbol declaration matches exist, bypasses the neural Cross-Encoder entirely, slashing query latency from >5000ms to sub-100ms.
  • Hybrid Path + Content Discovery: Decoupled path-based file discovery from content search constraints. rg --files runs concurrently with dedicated dynamic path scoring (compute_file_path_score), guaranteeing component declarations (e.g. BinaTextEditor.vue for BinaEditor) outrank incidental callers.
  • Noise Dampening for Localization & Dictionaries: Automatically excludes **/lang/** and **/locales/** from default code search scope (unless explicitly requested) and applies negative score penalties to eliminate dictionary saturation.
  • Fully Offline Tokenizer Initialization: Enforced local_files_only=True on Hugging Face tokenizer loading with automatic fallback, preventing network round-trips and cold-start latency.

Changed

  • Bounded neural Cross-Encoder candidate evaluation to the top 12 pre-ranked candidates with max sequence length 256, capping conceptual query execution time under 250ms.

[0.3.2] - 2026-09-29

Added

  • Algorithmic Morphological Stemmer (stem_word): In-memory rule-based inflection analysis for prefixes (re-, un-), participles (-ed, -ing), and plural/verb suffixes (-s, -es, -ies), with grammatical sibilant exclusions (status, wherehas, alias).
  • Software Engineering Action Taxonomy (PROGRAMMING_SYNONYMS): Zero-index lexical-semantic mapping covering 40+ common lifecycle verbs, framework relations, and domain entities (recalculate ↔ calculate/recalc/compute/update, variation ↔ variant/option/sku, bug ↔ issue/fix/wherehas/orwherehas).
  • Compound Identifier Preservation: Preserves raw camelCase and snake_case symbols (such as orWhereHas) alongside tokenized components during search term extraction.
  • Dynamic Modular Monorepo Discovery: Automatically inspects config/modules.php (paths.modules) to register custom module directories (e.g. vendor/bina/, Modules/) without hardcoding proprietary vendor names.
  • Configurable Search Overrides (LOCALGREP_DIRS): Added environment variable support to scan arbitrary non-standard monorepo directories.
  • Term Density Pre-scoring: Candidate ripgrep blocks are sorted and prioritized by query term density and source code weights prior to neural cross-encoder re-ranking.

Changed

  • Decreased default cross-encoder sequence length to 256 and batch size to 16 to reduce peak working set memory footprint.
  • Added meta-programming stop words (function, method, class, variable, trait, interface) to prevent candidate pollution.
  • Excluded internal prompt directories (!**/prompts/**) from default source code search scope.

[0.3.1] - 2026-09-29

Added

  • lg stop subcommand for graceful daemon shutdown with signal handling.
  • Process-level locking using fcntl.flock on the daemon PID file to prevent startup race conditions.
  • Automatic non-Latin query detection and user guidance message.
  • Explicit non-TTY stream detection via stat.S_ISFIFO and stat.S_ISREG for seamless shell piping (cat file | lg "q").

Fixed

  • Daemon Boot on Clean Install: Automatically downloads missing model.onnx from Hugging Face Hub during first run without crashing.
  • Linear Memory Leak: Disabled ONNX Runtime CPU memory arena (enable_cpu_mem_arena = False) and pattern memory allocator, keeping daemon RSS constant across long-running query cycles.
  • Header Snippet Extraction: Candidate path snippets now anchor on AST declaration keywords rather than generic namespace/license file headers.

Security

  • Hardened daemon socket permissions from 0o777 in /tmp to 0o600 inside ~/.local/localgrep/ (mode 0o700), restricting socket communication strictly to the owning user.
  • Added 30-second socket timeout and 10MB payload size limit on incoming connections to defend against socket starvation and buffer exhaustion.

[0.3.0] - 2026-09-28

Added

  • lg skill "<task>" CLI command for instant (<25ms) dynamic recommendation of AI agent skills.
  • Native Model Context Protocol (MCP) tool localgrep_skill for JIT skill resolution in Claude Code, Antigravity, and Cursor.
  • Recursive skill discovery covering .claude/skills, .agents/skills, .gemini/skills, and builtin/skills.

[0.2.0] - 2026-09-28

Added

  • Tree-sitter AST context pruner: lg prune <file> "<query>".
  • Grammar parsers for PHP, JavaScript, TypeScript, Python, and Vue Single File Components (SFC).
  • Model Context Protocol (MCP) server support (localgrep_search, localgrep_prune).
  • ONNX Runtime cross-encoder backend with automatic fallback to PyTorch.

[0.1.0] - 2026-09-28

Added

  • Initial release of LocalGrep (lg).
  • Zero-index local semantic search combining ripgrep candidate extraction with cross-encoder re-ranking.
  • Unix domain socket daemon architecture for sub-100ms response times.
  • Stream distilling utility for large CLI command outputs (<command> | lg "<query>").
Assets 2
Loading