@softerist/heuristic-mcp 3.2.3 → 3.2.4

package/README.md

@@ -1,376 +1,387 @@
-# Heuristic MCP Server
-
-An enhanced MCP server for your codebase. It provides intelligent semantic search, find-similar-code, recency-aware ranking, call-graph proximity boosts, and smart chunking. Optimized for Antigravity, Cursor, Claude Desktop, and VS Code.
-
----
-
-## Key Features
-
-- Zero-touch setup: postinstall auto-registers the MCP server with supported IDEs when possible.
-- Smart indexing: detects project type and applies smart ignore patterns on top of your excludes.
-- Semantic search: find code by meaning, not just keywords.
-- Find similar code: locate near-duplicate or related patterns from a snippet.
-- Package version lookup: check latest versions from npm, PyPI, crates.io, Maven, and more.
-- Workspace switching: change workspace at runtime without restarting the server.
-- Recency ranking and call-graph boosting: surfaces fresh and related code.
-- Optional ANN index: faster candidate retrieval for large codebases.
-- Optional binary vector store: mmap-friendly cache format for large repos.
-- Flexible embedding dimensions: MRL-compatible dimension reduction (64-768d) for speed/quality tradeoffs.
-
----
-
-## Installation
-
-Install globally (recommended):
-
-```bash
-npm install -g @softerist/heuristic-mcp
-```
-
-What happens during install:
-
-- Registration runs automatically (`scripts/postinstall.js`).
-- Model pre-download is attempted (`scripts/download-model.js`). If offline, it will be skipped and downloaded on first run.
-
-If auto-registration did not update your IDE config, run:
-
-```bash
-heuristic-mcp --start
-```
-
----
-
-## CLI Commands
-
-The `heuristic-mcp` binary manages the server lifecycle.
-
-### Status
-
-```bash
-heuristic-mcp --status
-```
-
-Shows server PID(s) and cache stats.
-
-### Logs
-
-```bash
-heuristic-mcp --logs
-```
-
-Tails the server log for the current workspace (defaults to last 200 lines and follows).
-
-Optional flags:
-
-```bash
-heuristic-mcp --logs --tail 100
-heuristic-mcp --logs --no-follow
-```
-
-### Version
-
-```bash
-heuristic-mcp --version
-```
-
-### Start/Stop
-
-```bash
-heuristic-mcp --start
-heuristic-mcp --start antigravity
-heuristic-mcp --start codex
-heuristic-mcp --start cursor
-heuristic-mcp --start vscode
-heuristic-mcp --start windsurf
-heuristic-mcp --start warp
-heuristic-mcp --start "Claude Desktop"
-heuristic-mcp --stop
-```
-
-`--start` registers (if needed) and enables the MCP server entry. `--stop` disables it so the IDE won't immediately respawn it. Restart/reload the IDE after `--start` to launch.
-
-Warp note: this package now targets `~/.warp/mcp_settings.json` (and `%APPDATA%\\Warp\\mcp_settings.json` on Windows when present). If no local Warp MCP config is writable yet, use Warp MCP settings/UI once to initialize it, then re-run `--start warp`.
-
-### Clear Cache
-
-```bash
-heuristic-mcp --clear-cache
-```
-
-Clears the cache for the current working directory (or `--workspace` if provided) and removes stale cache directories without metadata.
-
----
-
-## Configuration (`config.jsonc`)
-
-Configuration is loaded from your workspace root when the server runs with `--workspace`. If not provided by the IDE, the server auto-detects workspace via environment variables and current working directory. In server mode, it falls back to the package `config.jsonc` (or `config.json`) and then your current working directory.
-
-Example `config.jsonc`:
-
-```json
-{
-  "excludePatterns": ["**/legacy-code/**", "**/*.test.ts"],
-  "fileNames": ["Dockerfile", ".env.example", "Makefile"],
-  "indexing": {
-    "smartIndexing": true
-  },
-  "worker": {
-    "workerThreads": 0
-  },
-  "embedding": {
-    "embeddingModel": "jinaai/jina-embeddings-v2-base-code",
-    "embeddingBatchSize": null,
-    "embeddingProcessNumThreads": 8
-  },
-  "search": {
-    "recencyBoost": 0.1,
-    "recencyDecayDays": 30
-  },
-  "callGraph": {
-    "callGraphEnabled": true,
-    "callGraphBoost": 0.15
-  },
-  "ann": {
-    "annEnabled": true
-  },
-  "vectorStore": {
-    "vectorStoreFormat": "binary",
-    "vectorStoreContentMode": "external",
-    "vectorStoreLoadMode": "disk",
-    "contentCacheEntries": 256,
-    "vectorCacheEntries": 64
-  },
-  "memoryCleanup": {
-    "clearCacheAfterIndex": true
-  }
-}
-```
-
-Preferred style is namespaced keys (shown above). Legacy top-level keys are still supported for backward compatibility.
-
-### Embedding Model & Dimension Options
-
-**Default model:** `jinaai/jina-embeddings-v2-base-code` (768 dimensions)
-
-> **Important:** The default Jina model was **not** trained with Matryoshka Representation Learning (MRL). Dimension reduction (`embeddingDimension`) will significantly degrade search quality with this model. Only use dimension reduction with MRL-trained models.
-
-For faster search with smaller embeddings, switch to an MRL-compatible model:
-
-```json
-{
-  "embedding": {
-    "embeddingModel": "nomic-ai/nomic-embed-text-v1.5",
-    "embeddingDimension": 128
-  }
-}
-```
-
-**MRL-compatible models:**
-- `nomic-ai/nomic-embed-text-v1.5` — recommended for 128d/256d
-- Other models explicitly trained with Matryoshka loss
-
-**embeddingDimension values:** `64 | 128 | 256 | 512 | 768 | null` (null = full dimensions)
-
-Cache location:
-
-- By default, the cache is stored in a global OS cache directory under `heuristic-mcp/<hash>`.
-- You can override with `cacheDirectory` in your config file.
-
-### Environment Variables
-
-Selected overrides (prefix `SMART_CODING_`):
-
-Environment overrides target runtime keys and are synced back into namespaces by `lib/config.js`.
-
-- `SMART_CODING_VERBOSE=true|false` — enable detailed logging.
-- `SMART_CODING_WORKER_THREADS=auto|N` — worker thread count.
-- `SMART_CODING_BATCH_SIZE=100` — files per indexing batch.
-- `SMART_CODING_CHUNK_SIZE=25` — lines per chunk.
-- `SMART_CODING_MAX_RESULTS=5` — max search results.
-- `SMART_CODING_EMBEDDING_BATCH_SIZE=64` — embedding batch size (1–256, overrides auto).
-- `SMART_CODING_EMBEDDING_THREADS=8` — ONNX threads for the embedding child process.
-- `SMART_CODING_RECENCY_BOOST=0.1` — boost for recently edited files.
-- `SMART_CODING_RECENCY_DECAY_DAYS=30` — days until recency boost decays to 0.
-- `SMART_CODING_ANN_ENABLED=true|false` — enable ANN index.
-- `SMART_CODING_ANN_EF_SEARCH=64` — ANN search quality/speed tradeoff.
-- `SMART_CODING_VECTOR_STORE_FORMAT=json|binary|sqlite` — on-disk vector store format.
-- `SMART_CODING_VECTOR_STORE_CONTENT_MODE=external|inline` — where content is stored for binary format.
-- `SMART_CODING_VECTOR_STORE_LOAD_MODE=memory|disk` — vector loading strategy.
-- `SMART_CODING_CONTENT_CACHE_ENTRIES=256` — LRU entries for decoded content.
-- `SMART_CODING_VECTOR_CACHE_ENTRIES=64` — LRU entries for vectors (disk mode).
-- `SMART_CODING_CLEAR_CACHE_AFTER_INDEX=true|false` — drop in-memory vectors after indexing.
-- `SMART_CODING_UNLOAD_MODEL_AFTER_INDEX=true|false` — unload embedding model after indexing to free RAM (~500MB-1GB).
-- `SMART_CODING_EXPLICIT_GC=true|false` — opt-in to explicit GC (requires `--expose-gc`).
-- `SMART_CODING_INCREMENTAL_GC_THRESHOLD_MB=2048` — RSS threshold for running incremental GC after watcher updates (requires explicit GC).
-- `SMART_CODING_EMBEDDING_DIMENSION=64|128|256|512|768` — MRL dimension reduction (only for MRL-trained models).
-
-See `lib/config.js` for the full list.
-
-### Binary Vector Store
-
-Set `vectorStore.vectorStoreFormat` to `binary` to use the on-disk binary cache. This keeps vectors and content out of JS heap
-and reads on demand. Recommended for large repos.
-
-- `vectorStore.vectorStoreContentMode=external` keeps content in the binary file and only loads for top-N results.
-- `vectorStore.contentCacheEntries` controls the small in-memory LRU for decoded content strings.
-- `vectorStore.vectorStoreLoadMode=disk` streams vectors from disk to reduce memory usage.
-- `vectorStore.vectorCacheEntries` controls the small in-memory LRU for vectors when using disk mode.
-- `memoryCleanup.clearCacheAfterIndex=true` drops in-memory vectors after indexing and reloads lazily on next query.
-- `memoryCleanup.unloadModelAfterIndex=true` (default) unloads the embedding model after indexing to free ~500MB-1GB of RAM; the model will reload on the next search query.
-- Note: `ann.annEnabled=true` with `vectorStore.vectorStoreLoadMode=disk` can increase disk reads during ANN rebuilds on large indexes.
-
-### SQLite Vector Store
-
-Set `vectorStore.vectorStoreFormat` to `sqlite` to use SQLite for persistence. This provides:
-
-- ACID transactions for reliable writes
-- Simpler concurrent access
-- Standard database format for inspection
-
-```json
-{
-  "vectorStore": {
-    "vectorStoreFormat": "sqlite"
-  }
-}
-```
-
-The vectors and content are stored in `vectors.sqlite` in your cache directory. You can inspect it with any SQLite browser.
-`vectorStore.vectorStoreContentMode` and `vectorStore.vectorStoreLoadMode` are respected for SQLite (use `vectorStore.vectorStoreLoadMode=disk` to avoid loading vectors into memory).
-
-**Tradeoffs vs Binary:**
-- Slightly higher read overhead (SQL queries vs direct memory access)
-- Better write reliability (transactions)
-- Easier debugging (standard SQLite file)
-
-### Benchmarking Search
-
-Use the built-in script to compare memory vs latency tradeoffs:
-
-```bash
-node tools/scripts/benchmark-search.js --query "database connection" --runs 10
-```
-
-Compare modes quickly:
-
-```bash
-SMART_CODING_VECTOR_STORE_LOAD_MODE=memory node tools/scripts/benchmark-search.js --runs 10
-SMART_CODING_VECTOR_STORE_LOAD_MODE=disk node tools/scripts/benchmark-search.js --runs 10
-SMART_CODING_VECTOR_STORE_FORMAT=binary SMART_CODING_VECTOR_STORE_LOAD_MODE=disk node tools/scripts/benchmark-search.js --runs 10
-```
-
-Note: On small repos, disk mode may be slightly slower and show noisy RSS deltas; benefits are clearer on large indexes with a small `vectorStore.vectorCacheEntries`.
-
----
-
-## MCP Tools Reference
-
-### `a_semantic_search`
-Find code by meaning. Ideal for natural language queries like "authentication logic" or "database queries".
-
-### `b_index_codebase`
-Manually trigger a full reindex. Useful after large code changes.
-
-### `c_clear_cache`
-Clear the embeddings cache and force reindex.
-
-### `d_ann_config`
-Configure the ANN (Approximate Nearest Neighbor) index. Actions: `stats`, `set_ef_search`, `rebuild`.
-
-### `d_find_similar_code`
-Find similar code patterns given a snippet. Useful for finding duplicates or refactoring opportunities.
-
-### `e_check_package_version`
-Fetch the latest version of a package from its official registry.
-
-**Supported registries:**
-- **npm** (default): `lodash`, `@types/node`
-- **PyPI**: `pip:requests`, `pypi:django`
-- **crates.io**: `cargo:serde`, `rust:tokio`
-- **Maven**: `maven:org.springframework:spring-core`
-- **Go**: `go:github.com/gin-gonic/gin`
-- **RubyGems**: `gem:rails`
-- **NuGet**: `nuget:Newtonsoft.Json`
-- **Packagist**: `composer:laravel/framework`
-- **Hex**: `hex:phoenix`
-- **pub.dev**: `pub:flutter`
-- **Homebrew**: `brew:node`
-- **Conda**: `conda:numpy`
-
-### `f_set_workspace`
-Change the workspace directory at runtime. Updates search directory, cache location, and optionally triggers reindex.
-
-The server also attempts this automatically before each tool call when it detects a new workspace path from environment variables (for example `CODEX_WORKSPACE`, `CODEX_PROJECT_ROOT`, `WORKSPACE_FOLDER`).
-
-**Parameters:**
-- `workspacePath` (required): Absolute path to the new workspace
-- `reindex` (optional, default: `true`): Whether to trigger a full reindex
-
----
-
-## Troubleshooting
-
-**Server isn't starting**
-
-1. Run `heuristic-mcp --status` to check config and cache status.
-2. Run `heuristic-mcp --logs` to see startup errors.
-
-**Native ONNX backend unavailable (falls back to WASM)**
-
-If you see log lines like:
-
-```
-Native ONNX backend unavailable: The operating system cannot run %1.
-...onnxruntime_binding.node. Falling back to WASM.
-```
-
-The server will automatically disable workers and force `embedding.embeddingProcessPerBatch` to reduce memory spikes, but you
-should fix the native binding to restore stable memory usage:
-
-- Ensure you are running **64-bit Node.js** (`node -p "process.arch"` should be `x64`).
-- Install **Microsoft Visual C++ 2015–2022 Redistributable (x64)**.
-- Reinstall dependencies (clears locked native binaries):
-
-```bash
-Remove-Item -Recurse -Force node_modules\\onnxruntime-node, node_modules\\.onnxruntime-node-* -ErrorAction SilentlyContinue
-npm install
-```
-
-If you see a warning about **version mismatch** (e.g. "onnxruntime-node 1.23.x incompatible with transformers.js
-expectation 1.14.x"), install the matching version:
-
-```bash
-npm install onnxruntime-node@1.14.0
-```
-
-**Search returns no results**
-
-- Check `heuristic-mcp --status` for indexing progress.
-- If indexing shows zero files, review `excludePatterns` and `fileExtensions`.
-
-**Model download fails**
-
-- The install step tries to pre-download the model, but it can be skipped offline.
-- The server will download on first run; ensure network access at least once.
-
-**Clear cache**
-
-- Use the MCP tool `c_clear_cache`, run `heuristic-mcp --clear-cache`, or delete the cache directory. For local dev, run `npm run clean`.
-
-**Inspect cache**
-
-```bash
-node tools/scripts/cache-stats.js --workspace <path>
-```
-
-**Stop doesn't stick**
-
-- The IDE will auto-restart the server if it's still enabled in its config. `--stop` now disables the server entry for Antigravity, Cursor (including `~/.cursor/mcp.json`), Windsurf (`~/.codeium/windsurf/mcp_config.json`), Warp (`~/.warp/mcp_settings.json` and `%APPDATA%\\Warp\\mcp_settings.json` when present), Claude Desktop, and VS Code (when using common MCP settings keys). Restart the IDE after `--start` to re-enable.
-
----
-
-## Contributing
-
-See `CONTRIBUTING.md` for guidelines.
-
-License: MIT
+# Heuristic MCP Server
+
+An enhanced MCP server for your codebase. It provides intelligent semantic search, find-similar-code, recency-aware ranking, call-graph proximity boosts, and smart chunking. Optimized for Antigravity, Cursor, Claude Desktop, and VS Code.
+
+---
+
+## Key Features
+
+- Zero-touch setup: postinstall auto-registers the MCP server with supported IDEs when possible.
+- Smart indexing: detects project type and applies smart ignore patterns on top of your excludes.
+- Semantic search: find code by meaning, not just keywords.
+- Find similar code: locate near-duplicate or related patterns from a snippet.
+- Package version lookup: check latest versions from npm, PyPI, crates.io, Maven, and more.
+- Workspace switching: change workspace at runtime without restarting the server.
+- Recency ranking and call-graph boosting: surfaces fresh and related code.
+- Optional ANN index: faster candidate retrieval for large codebases.
+- Optional binary vector store: mmap-friendly cache format for large repos.
+- Flexible embedding dimensions: MRL-compatible dimension reduction (64-768d) for speed/quality tradeoffs.
+
+---
+
+## Installation
+
+Install globally (recommended):
+
+```bash
+npm install -g @softerist/heuristic-mcp
+```
+
+What happens during install:
+
+- Registration runs automatically (`scripts/postinstall.js`).
+- Model pre-download is attempted (`scripts/download-model.js`). If offline, it will be skipped and downloaded on first run.
+
+If auto-registration did not update your IDE config, run:
+
+```bash
+heuristic-mcp --start
+```
+
+---
+
+## CLI Commands
+
+The `heuristic-mcp` binary manages the server lifecycle.
+
+### Status
+
+```bash
+heuristic-mcp --status
+```
+
+Shows server PID(s) and cache stats.
+
+### Logs
+
+```bash
+heuristic-mcp --logs
+```
+
+Tails the server log for the current workspace (defaults to last 200 lines and follows).
+
+Optional flags:
+
+```bash
+heuristic-mcp --logs --tail 100
+heuristic-mcp --logs --no-follow
+```
+
+### Version
+
+```bash
+heuristic-mcp --version
+```
+
+### Start/Stop
+
+```bash
+heuristic-mcp --start
+heuristic-mcp --start antigravity
+heuristic-mcp --start codex
+heuristic-mcp --start cursor
+heuristic-mcp --start vscode
+heuristic-mcp --start windsurf
+heuristic-mcp --start warp
+heuristic-mcp --start "Claude Desktop"
+heuristic-mcp --stop
+```
+
+`--start` registers (if needed) and enables the MCP server entry. `--stop` disables it so the IDE won't immediately respawn it. Restart/reload the IDE after `--start` to launch.
+
+Warp note: this package now targets `~/.warp/mcp_settings.json` (and `%APPDATA%\\Warp\\mcp_settings.json` on Windows when present). If no local Warp MCP config is writable yet, use Warp MCP settings/UI once to initialize it, then re-run `--start warp`.
+
+### Clear Cache
+
+```bash
+heuristic-mcp --clear-cache
+```
+
+Clears the cache for the current working directory (or `--workspace` if provided) and removes stale cache directories without metadata.
+
+---
+
+## Configuration (`config.jsonc`)
+
+Configuration is loaded from your workspace root when the server runs with `--workspace`. If not provided by the IDE, the server auto-detects workspace via environment variables and current working directory. In server mode, it falls back to the package `config.jsonc` (or `config.json`) and then your current working directory.
+
+Example `config.jsonc`:
+
+```json
+{
+  "excludePatterns": ["**/legacy-code/**", "**/*.test.ts"],
+  "fileNames": ["Dockerfile", ".env.example", "Makefile"],
+  "indexing": {
+    "smartIndexing": true
+  },
+  "worker": {
+    "workerThreads": 0
+  },
+  "embedding": {
+    "embeddingModel": "jinaai/jina-embeddings-v2-base-code",
+    "embeddingBatchSize": null,
+    "embeddingProcessNumThreads": 8
+  },
+  "search": {
+    "recencyBoost": 0.1,
+    "recencyDecayDays": 30
+  },
+  "callGraph": {
+    "callGraphEnabled": true,
+    "callGraphBoost": 0.15
+  },
+  "ann": {
+    "annEnabled": true
+  },
+  "vectorStore": {
+    "vectorStoreFormat": "binary",
+    "vectorStoreContentMode": "external",
+    "vectorStoreLoadMode": "disk",
+    "contentCacheEntries": 256,
+    "vectorCacheEntries": 64
+  },
+  "memoryCleanup": {
+    "clearCacheAfterIndex": true
+  }
+}
+```
+
+Preferred style is namespaced keys (shown above). Legacy top-level keys are still supported for backward compatibility.
+
+### Embedding Model & Dimension Options
+
+**Default model:** `jinaai/jina-embeddings-v2-base-code` (768 dimensions)
+
+> **Important:** The default Jina model was **not** trained with Matryoshka Representation Learning (MRL). Dimension reduction (`embeddingDimension`) will significantly degrade search quality with this model. Only use dimension reduction with MRL-trained models.
+
+For faster search with smaller embeddings, switch to an MRL-compatible model:
+
+```json
+{
+  "embedding": {
+    "embeddingModel": "nomic-ai/nomic-embed-text-v1.5",
+    "embeddingDimension": 128
+  }
+}
+```
+
+**MRL-compatible models:**
+
+- `nomic-ai/nomic-embed-text-v1.5` — recommended for 128d/256d
+- Other models explicitly trained with Matryoshka loss
+
+**embeddingDimension values:** `64 | 128 | 256 | 512 | 768 | null` (null = full dimensions)
+
+Cache location:
+
+- By default, the cache is stored in a global OS cache directory under `heuristic-mcp/<hash>`.
+- You can override with `cacheDirectory` in your config file.
+
+### Environment Variables
+
+Selected overrides (prefix `SMART_CODING_`):
+
+Environment overrides target runtime keys and are synced back into namespaces by `lib/config.js`.
+
+- `SMART_CODING_VERBOSE=true|false` — enable detailed logging.
+- `SMART_CODING_WORKER_THREADS=auto|N` — worker thread count.
+- `SMART_CODING_BATCH_SIZE=100` — files per indexing batch.
+- `SMART_CODING_CHUNK_SIZE=25` — lines per chunk.
+- `SMART_CODING_MAX_RESULTS=5` — max search results.
+- `SMART_CODING_EMBEDDING_BATCH_SIZE=64` — embedding batch size (1–256, overrides auto).
+- `SMART_CODING_EMBEDDING_THREADS=8` — ONNX threads for the embedding child process.
+- `SMART_CODING_RECENCY_BOOST=0.1` — boost for recently edited files.
+- `SMART_CODING_RECENCY_DECAY_DAYS=30` — days until recency boost decays to 0.
+- `SMART_CODING_ANN_ENABLED=true|false` — enable ANN index.
+- `SMART_CODING_ANN_EF_SEARCH=64` — ANN search quality/speed tradeoff.
+- `SMART_CODING_VECTOR_STORE_FORMAT=json|binary|sqlite` — on-disk vector store format.
+- `SMART_CODING_VECTOR_STORE_CONTENT_MODE=external|inline` — where content is stored for binary format.
+- `SMART_CODING_VECTOR_STORE_LOAD_MODE=memory|disk` — vector loading strategy.
+- `SMART_CODING_CONTENT_CACHE_ENTRIES=256` — LRU entries for decoded content.
+- `SMART_CODING_VECTOR_CACHE_ENTRIES=64` — LRU entries for vectors (disk mode).
+- `SMART_CODING_CLEAR_CACHE_AFTER_INDEX=true|false` — drop in-memory vectors after indexing.
+- `SMART_CODING_UNLOAD_MODEL_AFTER_INDEX=true|false` — unload embedding model after indexing to free RAM (~500MB-1GB).
+- `SMART_CODING_EXPLICIT_GC=true|false` — opt-in to explicit GC (requires `--expose-gc`).
+- `SMART_CODING_INCREMENTAL_GC_THRESHOLD_MB=2048` — RSS threshold for running incremental GC after watcher updates (requires explicit GC).
+- `SMART_CODING_EMBEDDING_DIMENSION=64|128|256|512|768` — MRL dimension reduction (only for MRL-trained models).
+
+See `lib/config.js` for the full list.
+
+### Binary Vector Store
+
+Set `vectorStore.vectorStoreFormat` to `binary` to use the on-disk binary cache. This keeps vectors and content out of JS heap
+and reads on demand. Recommended for large repos.
+
+- `vectorStore.vectorStoreContentMode=external` keeps content in the binary file and only loads for top-N results.
+- `vectorStore.contentCacheEntries` controls the small in-memory LRU for decoded content strings.
+- `vectorStore.vectorStoreLoadMode=disk` streams vectors from disk to reduce memory usage.
+- `vectorStore.vectorCacheEntries` controls the small in-memory LRU for vectors when using disk mode.
+- `memoryCleanup.clearCacheAfterIndex=true` drops in-memory vectors after indexing and reloads lazily on next query.
+- `memoryCleanup.unloadModelAfterIndex=true` (default) unloads the embedding model after indexing to free ~500MB-1GB of RAM; the model will reload on the next search query.
+- Note: `ann.annEnabled=true` with `vectorStore.vectorStoreLoadMode=disk` can increase disk reads during ANN rebuilds on large indexes.
+
+### SQLite Vector Store
+
+Set `vectorStore.vectorStoreFormat` to `sqlite` to use SQLite for persistence. This provides:
+
+- ACID transactions for reliable writes
+- Simpler concurrent access
+- Standard database format for inspection
+
+```json
+{
+  "vectorStore": {
+    "vectorStoreFormat": "sqlite"
+  }
+}
+```
+
+The vectors and content are stored in `vectors.sqlite` in your cache directory. You can inspect it with any SQLite browser.
+`vectorStore.vectorStoreContentMode` and `vectorStore.vectorStoreLoadMode` are respected for SQLite (use `vectorStore.vectorStoreLoadMode=disk` to avoid loading vectors into memory).
+
+**Tradeoffs vs Binary:**
+
+- Slightly higher read overhead (SQL queries vs direct memory access)
+- Better write reliability (transactions)
+- Easier debugging (standard SQLite file)
+
+### Benchmarking Search
+
+Use the built-in script to compare memory vs latency tradeoffs:
+
+```bash
+node tools/scripts/benchmark-search.js --query "database connection" --runs 10
+```
+
+Compare modes quickly:
+
+```bash
+SMART_CODING_VECTOR_STORE_LOAD_MODE=memory node tools/scripts/benchmark-search.js --runs 10
+SMART_CODING_VECTOR_STORE_LOAD_MODE=disk node tools/scripts/benchmark-search.js --runs 10
+SMART_CODING_VECTOR_STORE_FORMAT=binary SMART_CODING_VECTOR_STORE_LOAD_MODE=disk node tools/scripts/benchmark-search.js --runs 10
+```
+
+Note: On small repos, disk mode may be slightly slower and show noisy RSS deltas; benefits are clearer on large indexes with a small `vectorStore.vectorCacheEntries`.
+
+---
+
+## MCP Tools Reference
+
+### `a_semantic_search`
+
+Find code by meaning. Ideal for natural language queries like "authentication logic" or "database queries".
+
+### `b_index_codebase`
+
+Manually trigger a full reindex. Useful after large code changes.
+
+### `c_clear_cache`
+
+Clear the embeddings cache and force reindex.
+
+### `d_ann_config`
+
+Configure the ANN (Approximate Nearest Neighbor) index. Actions: `stats`, `set_ef_search`, `rebuild`.
+
+### `d_find_similar_code`
+
+Find similar code patterns given a snippet. Useful for finding duplicates or refactoring opportunities.
+
+### `e_check_package_version`
+
+Fetch the latest version of a package from its official registry.
+
+**Supported registries:**
+
+- **npm** (default): `lodash`, `@types/node`
+- **PyPI**: `pip:requests`, `pypi:django`
+- **crates.io**: `cargo:serde`, `rust:tokio`
+- **Maven**: `maven:org.springframework:spring-core`
+- **Go**: `go:github.com/gin-gonic/gin`
+- **RubyGems**: `gem:rails`
+- **NuGet**: `nuget:Newtonsoft.Json`
+- **Packagist**: `composer:laravel/framework`
+- **Hex**: `hex:phoenix`
+- **pub.dev**: `pub:flutter`
+- **Homebrew**: `brew:node`
+- **Conda**: `conda:numpy`
+
+### `f_set_workspace`
+
+Change the workspace directory at runtime. Updates search directory, cache location, and optionally triggers reindex.
+
+The server also attempts this automatically before each tool call when it detects a new workspace path from environment variables (for example `CODEX_WORKSPACE`, `CODEX_PROJECT_ROOT`, `WORKSPACE_FOLDER`).
+
+**Parameters:**
+
+- `workspacePath` (required): Absolute path to the new workspace
+- `reindex` (optional, default: `true`): Whether to trigger a full reindex
+
+---
+
+## Troubleshooting
+
+**Server isn't starting**
+
+1. Run `heuristic-mcp --status` to check config and cache status.
+2. Run `heuristic-mcp --logs` to see startup errors.
+
+**Native ONNX backend unavailable (falls back to WASM)**
+
+If you see log lines like:
+
+```
+Native ONNX backend unavailable: The operating system cannot run %1.
+...onnxruntime_binding.node. Falling back to WASM.
+```
+
+The server will automatically disable workers and force `embedding.embeddingProcessPerBatch` to reduce memory spikes, but you
+should fix the native binding to restore stable memory usage:
+
+- Ensure you are running **64-bit Node.js** (`node -p "process.arch"` should be `x64`).
+- Install **Microsoft Visual C++ 2015–2022 Redistributable (x64)**.
+- Reinstall dependencies (clears locked native binaries):
+
+```bash
+Remove-Item -Recurse -Force node_modules\\onnxruntime-node, node_modules\\.onnxruntime-node-* -ErrorAction SilentlyContinue
+npm install
+```
+
+If you see a warning about **version mismatch** (e.g. "onnxruntime-node 1.23.x incompatible with transformers.js
+expectation 1.14.x"), install the matching version:
+
+```bash
+npm install onnxruntime-node@1.14.0
+```
+
+**Search returns no results**
+
+- Check `heuristic-mcp --status` for indexing progress.
+- If indexing shows zero files, review `excludePatterns` and `fileExtensions`.
+
+**Model download fails**
+
+- The install step tries to pre-download the model, but it can be skipped offline.
+- The server will download on first run; ensure network access at least once.
+
+**Clear cache**
+
+- Use the MCP tool `c_clear_cache`, run `heuristic-mcp --clear-cache`, or delete the cache directory. For local dev, run `npm run clean`.
+
+**Inspect cache**
+
+```bash
+node tools/scripts/cache-stats.js --workspace <path>
+```
+
+**Stop doesn't stick**
+
+- The IDE will auto-restart the server if it's still enabled in its config. `--stop` now disables the server entry for Antigravity, Cursor (including `~/.cursor/mcp.json`), Windsurf (`~/.codeium/windsurf/mcp_config.json`), Warp (`~/.warp/mcp_settings.json` and `%APPDATA%\\Warp\\mcp_settings.json` when present), Claude Desktop, and VS Code (when using common MCP settings keys). Restart the IDE after `--start` to re-enable.
+
+---
+
+## Contributing
+
+See `CONTRIBUTING.md` for guidelines.
+
+License: MIT