Skip to content

CHANGELOG.md

Latest commit

 

History

History
284 lines (218 loc) · 17.3 KB

CHANGELOG.md

File metadata and controls

284 lines (218 loc) · 17.3 KB

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.

Unreleased

0.7.0 - 2026-04-14

Added

  • Knowledge base support: Added add_knowledge_base, list_knowledge_bases, and remove_knowledge_base tools to manage external document folders indexed alongside the project
  • Reranking with SiliconFlow: Added BAAI/bge-reranker-v2-m3 reranking support via SiliconFlow API for improved search result quality
  • Routing hints for local discovery: Added dynamic routing hints so local search can steer retrieval toward more relevant code paths before semantic reranking
  • TXT/HTML file support: Added *.txt, *.html, *.htm to default include patterns for document indexing
  • Config merging: Global and project configs are now merged, allowing shared provider settings at global level and knowledge base paths at project level
  • Hidden file exclusion: Files and folders starting with . are now excluded from indexing and file watching
  • Build folder exclusion: Folders containing "build" in their name (e.g., build, mingwBuildDebug) are now excluded from indexing and file watching
  • additionalInclude config: Added new config option to extend default file patterns without replacing them
  • Eval diversity quality gates: Added raw and distinct top-k diversity metrics, budgets, and regression coverage for eval runs and reranker benchmarking

Changed

  • Default verbose=false: Changed /index command default to verbose=false to reduce token consumption
  • Dependency hardening: Added targeted npm overrides and refreshed lockfile resolution to keep vulnerable transitive packages patched in release builds

Fixed

  • Knowledge base refresh behavior: Adding or removing knowledge bases now rebuilds the shared in-memory indexer immediately instead of requiring a restart
  • Watcher-triggered reindexing: Restored automatic reindexing on file changes so watched projects and attached knowledge bases stay current during a live session
  • Parser and call-graph stability: Fixed recursion-limit and segmentation-fault regressions, removed unsupported parent traversal paths, and improved PHP method-call extraction reliability
  • Plugin/runtime packaging: Kept @opencode-ai/plugin available at runtime by shipping it as a dependency instead of relying on dev-only installation
  • Eval workflow rate limiting: Throttled GitHub Models quality runs to avoid rate-limit failures in the release verification pipeline

0.6.1 - 2026-03-29

Added

  • Custom provider batch caps: Added customProvider.maxBatchSize / max_batch_size support so OpenAI-compatible embedding servers can cap inputs per /embeddings request
  • Environment placeholders in config: Added {env:VAR_NAME} placeholder support for string config values so secrets and endpoints can be supplied from the environment instead of committed files

Changed

  • Release documentation alignment: Updated release metadata to publish the post-v0.6.0 config improvements as v0.6.1

0.6.0 - 2026-03-28

Added

  • Evaluation harness: First-class eval CLI, golden datasets, budgets, compare mode, run artifacts, and smoke/quality workflows for measuring retrieval quality over time
  • Implementation lookup workflow: Added a dedicated definition/implementation retrieval path across the CLI, plugin tools, MCP server, indexer, and tests for faster code lookup by intent
  • Cross-repo benchmarking: Added a benchmark runner with ripgrep and ast-grep baselines plus reproducible benchmarking documentation and golden datasets for external repos
  • Release automation guardrails: Added Release Drafter automation and CI enforcement for release-category and semver labels on pull requests
  • Contributor language-support guide: Added an agent-ready guide for extending semantic parsing and call-graph support to new languages
  • PHP language support: Added semantic parsing, chunking, and call-graph extraction for PHP, including fixtures and tests for constructors, imports, method calls, and simple calls

Changed

  • Evaluation CI strategy: Split the default GitHub Models quality gate from explicit external-provider budget checks and documented the active CI budget paths
  • Documentation refresh: Reorganized contributor and maintenance docs, expanded evaluation and benchmarking guidance, and updated README benchmark snapshots and workflow references

Fixed

  • Release Drafter permissions: Restored draft-release updates so release automation can keep draft notes current
  • Eval/CI correctness: Closed CI gating gaps, normalized baseline paths, and pinned the Rust toolchain action input used by CI
  • Benchmark auditability: Fixed scoped ast-grep metric accounting and dataset/result mutability issues in the benchmark runner and reporting flow
  • Supply-chain hardening: Tightened dependency and repository security posture, including stronger git/worktree handling coverage in tests
  • Native test reliability: test:run and test:coverage now rebuild the native module first so newly added parser/call-graph language support is exercised against a current binary during release verification

0.5.2 - 2026-03-21

Added

  • Call graph extraction and query: Tree-sitter query-based extraction of function calls, method calls, constructors, and imports across 5 languages (TypeScript/JavaScript, Python, Go, Rust)
  • call_graph tool: Query callers or callees of any function/method with branch-aware filtering
  • DB schema v2: symbols, call_edges, and branch_symbols tables with full CRUD, GC, and batch operations
  • Same-file call resolution: Automatically resolves call edges to symbols defined in the same file during indexing
  • /call-graph slash command: Added command support for call graph workflows

Changed

  • Documentation updates: Expanded README, CHANGELOG, and skill guide to document call graph usage and behavior

Fixed

  • Missing call_graph export: The call_graph tool was not exported from the plugin entry point — now available to OpenCode users
  • JavaScript call extraction routing: JavaScript now uses a dedicated query file instead of TypeScript query routing
  • Caller output context: Caller results now include caller symbol/file context for clearer navigation
  • Call graph consistency/integrity: Improved branch filtering and database integrity handling for call graph data

0.5.1 - 2026-03-01

Added

  • Custom embedding provider: Support for any OpenAI-compatible embedding endpoint (custom provider with baseUrl, model, dimensions config). Works with llama.cpp, vLLM, text-embeddings-inference, LiteLLM, etc.

Fixed

  • Critical: infinite recursion on stale lock file: When a stale indexing.lock existed from a crashed session, initialize() entered infinite recursion via recoverFromInterruptedIndexing() → healthCheck() → ensureInitialized() → initialize(), causing 70GB+ memory usage and OOM. Recovery now runs after store/database initialization.
  • Relative path storage: Index now stores relative paths for project portability. Detects and warns about legacy absolute-path indexes.
  • MCP status prompt: Removed empty args schema from status prompt that caused validation errors

Changed

  • Changelog and README: Fixed bullet formatting, added platform support table

0.5.0 - 2026-02-23

Added

  • MCP server: Standalone MCP server (opencode-codebase-index-mcp CLI) exposing all 8 tools and 4 prompts over stdio transport, enabling integration with Cursor, Claude Code, and Windsurf
  • Crash-safe indexing: Lock file and atomic writes prevent index corruption from interrupted indexing sessions, with automatic recovery on next run
  • Git worktree support: Branch detection now works correctly in git worktrees by resolving .git file pointers to the actual git directory
  • Index metadata contract: Stores embedding provider, model, and dimensions in the database; blocks searches against incompatible indexes with clear error messages and force=true rebuild instructions
  • Google gemini-embedding-001 model: Support for Google's latest embedding model with Matryoshka truncation (3072D → 1536D) and task-specific embeddings (CODE_RETRIEVAL_QUERY / RETRIEVAL_DOCUMENT)
  • Google batch embedding: Batch requests up to 20 texts per API call via batchEmbedContents endpoint
  • Compatibility warnings: Provider mismatch (same model + dimensions) now logs a warning instead of forcing a rebuild
  • Windows support: Native binaries now build on Windows MSVC across all 5 platform targets (macOS x86/ARM, Linux x86/ARM, Windows x86)

Changed

  • Embedding API split: embed() replaced by embedQuery() and embedDocument() to support task-specific embeddings (Google)
  • Type-safe embedding models: EMBEDDING_MODELS constant as single source of truth; EmbeddingProvider, EmbeddingModelName, and related types derived at compile time
  • Google default model: Updated from deprecated text-embedding-004 to text-embedding-005
  • Tool formatting: Extracted all formatting functions from src/tools/index.ts to src/tools/utils.ts
  • Exhaustive provider check: createEmbeddingProvider uses never exhaustive check instead of default branch
  • ESM compatibility: Build config adds createRequire shim for ESM entry points

Fixed

  • SQLite bind parameter limit: get_missing_embeddings and get_embeddings_batch now batch IN (...) queries to stay under SQLITE_MAX_VARIABLE_NUMBER (999) — fixes crash on large codebases (thanks @zb1749)
  • Google embedding API endpoints: Corrected single and batch request URLs
  • Index compatibility on force rebuild: clearIndex() now deletes stale index metadata so provider changes take effect
  • Search/findSimilar initialization: Both now call ensureInitialized() before compatibility check
  • Windows MSVC build: Disabled usearch simsimd feature on Windows — MSVC lacks _mm512_reduce_add_ph intrinsic. Pinned usearch to 2.23.0 to avoid 2.24.0 MAP_FAILED regression. Committed Cargo.lock for reproducible CI builds.

0.4.1 - 2025-01-19

Added

  • requireProjectMarker config option: Prevents plugin from hanging when opened in non-project directories like home. When true (default), requires a project marker (.git, package.json, Cargo.toml, etc.) to enable file watching and auto-indexing.

Fixed

  • Plugin no longer hangs when OpenCode is opened in home directory or other large non-project directories

0.4.0 - 2025-01-18

Added

  • find_similar tool: Find code similar to a given snippet for duplicate detection, pattern discovery, and refactoring prep. Paste code and find semantically similar implementations elsewhere in the codebase.
  • codebase_peek tool: Token-efficient semantic search returning metadata only (file, line, name, type) without code content. Saves ~90% tokens compared to codebase_search for discovery workflows.

0.3.2 - 2025-01-18

Fixed

  • Rust code formatting (cargo fmt)
  • CI publish workflow: use Node 24 + npm OIDC trusted publishing (no token required)

0.3.1 - 2025-01-18

Added

  • Query embedding cache: LRU cache (100 entries, 5min TTL) avoids redundant API calls for repeated searches
  • Query similarity matching: Reuses cached embeddings for similar queries (Jaccard similarity ≥0.85)
  • Batch metadata lookup: VectorStore.getMetadata() and getMetadataBatch() for efficient chunk retrieval
  • Parse timing metrics: Tracks parseMs for tree-sitter parsing duration
  • Query cache stats: Separate tracking for exact hits, similar hits, and misses

Changed

  • BM25 keyword search now uses getMetadataBatch() - O(n) instead of O(total) for result metadata lookup

Fixed

  • Remove console output from Logger (was leaking to stdout)
  • Record embedding API metrics for search queries (previously only tracked during indexing)
  • Record embedding API metrics during batch retries

0.3.0 - 2025-01-16

Added

  • Language support: Java, C#, Ruby, Bash, C, and C++ parsing via tree-sitter
  • CI improvements: Rust caching, cargo fmt --check, cargo clippy, and cargo test in workflows
  • /status command: Check index health and provider info
  • Batch operations: High-performance bulk inserts for embeddings and chunks (~10-18x speedup)
  • Auto garbage collection: Configurable automatic cleanup of orphaned embeddings/chunks
  • Documentation: ARCHITECTURE.md, TROUBLESHOOTING.md, comprehensive AGENTS.md

Changed

  • Upgraded tree-sitter from 0.20 to 0.24 (new LANGUAGE constant API)
  • Optimized embedBatch for Google and Ollama providers with Promise.all
  • Enhanced skill documentation with filter examples

Fixed

  • Node version consistency in publish workflow (Node 24 → Node 22)
  • Clippy warnings in Rust code

0.2.1 - 2025-01-10

Fixed

  • Rate limit handling and error messages
  • TypeScript errors in delta.ts

0.2.0 - 2025-01-09

Added

  • Branch-aware indexing: Embeddings stored by content hash, branch catalog tracks membership
  • SQLite storage: Persistent storage for embeddings, chunks, and branch catalog
  • Slash commands: /search, /find, /index, /status registered via config hook
  • Global config support: ~/.config/opencode/codebase-index.json
  • Provider-specific rate limiting: Ollama has no limits, GitHub Copilot has strict limits

Changed

  • Migrated from JSON file storage to SQLite database
  • Improved rate limit handling for GitHub Models API (15 req/min)

0.1.11 - 2025-01-07

Added

  • Community standards: LICENSE, Code of Conduct, Contributing guide, Security policy, Issue templates

Fixed

  • Clippy warnings and TypeScript type errors

0.1.10 - 2025-01-06

Added

  • F16 quantization: 50% memory reduction for vector storage
  • Dead-letter queue: Failed embedding batches are tracked for retry
  • JSDoc/docstring extraction: Comments included with semantic nodes
  • Overlapping chunks: Improved context continuity across chunk boundaries
  • maxChunksPerFile config: Control token costs for large files
  • semanticOnly config: Only index functions/classes, skip generic blocks

Changed

  • Moved inverted index from TypeScript to Rust native module (performance improvement)

Fixed

  • GitHub Models API for embeddings instead of Copilot API

0.1.9 - 2025-01-05

Fixed

  • Use GitHub Models API for embeddings instead of Copilot API

0.1.8 - 2025-01-04

Fixed

  • Only export default plugin to prevent OpenCode loader crash
  • Downgrade to zod v3 to match OpenCode SDK version

0.1.3 - 2025-01-02

Changed

  • Use Node.js 24 for npm 11+ trusted publishing support
  • Externalize @opencode-ai/plugin to prevent runtime conflicts

Fixed

  • ESM output as main entry for Bun/OpenCode compatibility
  • Native binding loading in CJS context

0.1.1 - 2025-01-01

Added

  • CI/CD workflows for testing and publishing
  • Comprehensive README with badges, diagrams, and examples

Fixed

  • NAPI configuration for OIDC trusted publishing

0.1.0 - 2024-12-30

Added

  • Initial release
  • Semantic codebase indexing with tree-sitter parsing
  • Vector similarity search with usearch (HNSW algorithm)
  • Hybrid search combining semantic + BM25 keyword matching
  • Support for TypeScript, JavaScript, Python, Rust, Go, JSON
  • Multiple embedding providers: GitHub Copilot, OpenAI, Google, Ollama
  • Incremental indexing with file hash caching
  • File watcher for automatic re-indexing
  • OpenCode tools: codebase_search, index_codebase, index_status, index_health_check
⚡