Changelog

Tool `smart_context` now supports 3 new modes:

`hnindex init --mcp` now generates all 16 env vars by default (INDEX_WORKERS, SEARCH_STREAM_ENABLED, SEARCH_FUZZY_ENABLED, SEARCH_CACHE_SIZE, and more)

`vibe-hnindex` v0.10.0, `hnindex-cli` v0.10.0.

`hnindex init-skill` no longer reads SKILL.md from disk. Content is inlined to avoid ENOENT errors when installed via npm.

`vibe-hnindex` v0.9.8, `hnindex-cli` v0.9.8.

New `use-vibe-hnindex` skill for AI agents (Claude, Antigravity, Cursor, Codex). Auto-loaded to understand all 20 tools, search modes, streaming, fuzzy matching, and best practices.

New `hnindex init-skill --target <editor>` command creates the skill file in the correct directory for your editor.

New `benchmark_search` tool runs a suite of test queries with and without streaming, reports avg/min/max timing and speedup ratios.

`vibe-hnindex` v0.9.5, `hnindex-cli` v0.9.5.

New `benchmark_search` tool runs a suite of test queries (keyword, hybrid, regex, fuzzy) with and without streaming.

`vibe-hnindex` v0.9.4, `hnindex-cli` v0.9.4.

Dependency, export, and symbol parsing is now done during the initial file scan, eliminating the second full directory scan entirely.

File change detection now uses SHA-1 instead of SHA-256 (~2× faster, still reliable for code deduplication).

`vibe-hnindex` v0.9.1, `hnindex-cli` v0.9.1.

New `stream: true` flag (or env `SEARCH_STREAM_ENABLED=true`) enables parallel search execution: keyword FTS5 + semantic Qdrant run simultaneously instead of sequentially.

Extracted `parallelSearch()` orchestrator, `sendProgress()` notifications, `sendSearchPreview()`, and `applyFuzzyBoost()` into `services/streaming-search.ts` for cleaner architecture.

`vibe-hnindex` v0.9.0, `hnindex-cli` v0.9.0.

New `fuzzy: true` flag uses Levenshtein distance to detect typos and re-rank search results.

`tools-reference.md`: added sections for Regex Search, Symbol Filters, Search Cache, and Fuzzy Search with examples and best practices.

`index_codebase` now uses a worker threads pool (auto = CPU count − 1, min 1 worker) to chunk and embed files in parallel.

LRU cache with 100 entries and 5-minute TTL. Cache key = `project|query|mode|limit|filters`.

New `mode: 'regex'` using JavaScript RegExp to match patterns in chunk content.

Filter search results by `symbol_kind`: `function`, `class`, `method`, `interface`, `type`, `variable`, `enum`, `export`.

`vibe-hnindex` v0.8.0, `hnindex-cli` v0.8.0.

Added timeouts to prevent the MCP server from hanging when Ollama or Qdrant are unresponsive.

`search` wrapped with `withTimeout()`. On timeout returns a clear error message instead of hanging indefinitely.

Added a dedicated Timeouts section listing all timeout environment variables.

Separated `create-release` job so GitHub Releases are created independently of npm publish (no more blocked releases).

Fixed `realpath` crash when traversing deeply nested folders.

Deleted the hard-coded `SUPPORTED_EXTENSIONS` set.

`hnindex-cli` synced to v0.7.1.

One-call health check: Ollama, Qdrant, config summary, optional embedding probe.

Short markdown for CLAUDE.md / AGENTS.md with project path, last index time, top language + stats, and `package.json` script hints.

SQLite `projects.indexed_git_head` stores `git rev-parse HEAD` after indexing.

MCP `server.json`, `.claude-plugin` metadata, npm packages, and startup log all updated to v0.7.0.

Documented in MCP `server.json`.

MCP server stderr now shows v0.6.1.

New global CLI: `hnindex init --mcp <claude|claude-desktop|antigravity|cursor|cursor-project|windsurf|vscode>`, `hnindex update`, `hnindex version`.

Heuristic symbol extraction (TS/JS/TSX/JSX, Python) on index/reindex/watch. SQLite `symbols` table with name-based lookup.

Resolve symbols by name with optional kind and file pattern filters.

Find chunks via symbol name (explicit mode; `auto` does not route to symbol).

`RERANK_URL` for POST `{ query, documents }` → `{ scores }`. Falls back to Qdrant semantic score reorder.

Richer detection: Prisma, Tailwind, shadcn/ui (`components.json`), Turborepo/Nx, lockfile → package manager, monorepo hints.

`getProjectWithRetry` eliminates rare "project not found" errors right after indexing.

Summary ends with `Ready: yes|no`. When Qdrant is up, shows `qdrant_vectors: <count>` after collection verify.

Heuristic routing between keyword and hybrid. Set `SEARCH_AUTO_ROUTE=true` to treat omitted `mode` as `auto`.

If keyword returns no hits and Ollama+Qdrant are OK, one semantic pass runs automatically.

Optional per-result score breakdown: path multiplier, RRF hints, semantic raw score.

`search` exposes `content_mode`, `max_content_chars`, `deprioritize_generated_paths`, `mode` (including `auto`), and `explain`.

Rule-based briefing from README, CLAUDE.md, `package.json`, and index statistics.

Single markdown block: briefing + stats + optional recent git activity. Truncated by `max_chars`.

New `watch` parameter (default `false`). Enables file watcher after indexing, same as `watch_project`.

Shared `startWatchingProject` logic used by both `watch_project` and `index_codebase` + `watch`.

Optional for local Docker; required for Qdrant Cloud.

If `QDRANT_URL` looks like Cloud (HTTPS + `qdrant` host) but no API key is set, a clear warning is logged on first Qdrant use.

Getting Started and Configuration pages now spell out Cloud vs self-hosted env vars. Default `OLLAMA_URL` is `http://localhost:11434`.

Default `compact` truncates chunk bodies to save tokens. Use `full` for entire chunks.

If strict FTS AND returns nothing with multi-token queries, automatically retries with OR and shows a warning.

Down-ranks results from `dist/`, `.next/`, `build/`, `coverage/`, `*.min.js`, `node_modules`, etc.

Reduces rare "project not found" errors right after indexing by retrying the stats query.

Tokenizes punctuation-heavy queries for better FTS5 matching.

One chunk per file by default (`dedupe_by_file`). Cleaner, less noisy results.

Exclude paths from indexing using minimatch patterns (gitignore-style syntax).