CHANGELOG.md

Changelog

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

[2.3.0] - 2026-08-06

Two new dashboard controls, Kotlin and Swift, and three Viewer fixes — one of which could freeze the whole MCP server.

Added

• Dashboard switches and buttons — the Live tab could only ever set a value (slider, number). Two control types join them: toggle for state (a two-position switch, value 0/1, with unit doubling as the "ON|OFF" captions) and button for events. A button's value is a monotonically rising press counter, not a flag, because a source polls GET /control at its own pace — with a boolean, every press landing between two polls would be lost. The source compares against the last count it saw and learns both that and how often it was pressed; a backwards jump (wrap at 1e6, /panel/clear, hub restart) means "restart, adopt the value". Presses arrive via the new POST /control/press, which takes only an id: the hub owns the counter, so two open dashboards both count instead of overwriting each other. Re-announcing a widget (a device rebooting) no longer clobbers a value the user dialed in, and the press count survives it.

• Kotlin and Swift support — .kt, .kts and .swift files are indexed via tree-sitter and classified as code across every extension list. Brings the language count to 14. A small share of files using very recent syntax falls back gracefully when the community grammar can't parse them.

Changed

• The Viewer's Debug tab is now called Live — sitting right next to Logs, the old name said nothing about what set the two apart. Logs is history that scrolls past; Live is the current state, held in fixed slots that overwrite in place. The heading inside the tab reads Live Dashboard. Nothing about the API changed: the panel endpoints, widget types and the debug tab id are all untouched, so existing senders and deep links keep working.

Fixed

• A single indexed file could freeze the entire MCP server — the per-file stats in buildTree used three LEFT JOINs with COUNT(DISTINCT) in one GROUP BY. The joins multiply per file before the DISTINCT collapses them again: occurrences × methods × types. One generated header in an indexed project (191.134 occurrences × 3.138 methods × 786 types) came to 471 billion intermediate rows for that one file. better-sqlite3 runs synchronously in native code, so the first tree request from a connecting browser killed the process for good: event loop dead, 98 % CPU for hours, 1,5 GB RSS, MCP calls hanging, dashboard empty — while the ports kept listening, which made it look "on but broken". Now three correlated subqueries per file, each walking one table by its file_id index, returning the same numbers in milliseconds. Both the code and all branches carried the pattern.

• The Viewer's file watcher opened tens of thousands of OS handles — chokidar removed glob support in v4, but the watcher still passed v3 globs ('**/node_modules/**' and friends). In v5 a string matcher is an exact comparison, so the ignore list silently never matched and every directory below the project root got watched — one non-recursive fs.watch, and therefore one OS handle, each. Measured on this repo (2.520 directories): 19.619 handles and 224 MB RSS, down to 2.375 and 84 MB after the fix; a control run without the Viewer stayed flat at 191. ignored is now a predicate that returns true for the directory itself (matching only files inside it still lets chokidar descend), and it reuses the exclusion list the indexer already had — a directory not worth indexing is not worth watching. Fixes the spurious tree rebuilds too: every node_modules event used to trigger a full rebuild and broadcast, for data that is not even in the database.

• managed_components is now excluded as well — the predicate fix above was necessary but not sufficient on ESP-IDF projects, where that directory is the embedded world's node_modules. The arithmetic on one such repo is the real finding: 14.858 handles = 1.865 watched directories + 12.713 watched files + base. On Windows chokidar holds a handle per watched file, not merely per directory, and 12.407 of those files sat in managed_components. This one entry brings the process back to a few hundred handles. It disappears from global-init scanning too, for the same reason as node_modules.

• The ▷ Demo button was missing from an empty dashboard — the Live tab rendered two different toolbars, and the empty-state one carried only the heading, dropping Demo, Pause and Clear. That hid the button exactly when it was needed: it is the way to get your first widgets, but you only saw it once widgets already existed. Both states now share one toolbar, and the empty-state hint points at the button instead of only naming the POST endpoint.

[2.2.2] - 2026-06-30

Community contribution. Adds Astro component support.

Added

• .astro file support (#15, thanks @zlegein) — Astro components are now indexed by parsing their TypeScript frontmatter (the code between the --- fences) with the existing TSX grammar. The template/markup below the frontmatter is skipped. Line numbers are preserved exactly: template lines are blanked (not removed) before parsing, so every reported method/type position matches the original file. No new dependency — reuses the bundled TypeScript grammar.

[2.2.1] - 2026-06-30

Install fix. On Node.js 24, npm install -g aidex-mcp failed for many users with node-gyp errors — they had to install a full C++ toolchain (Visual Studio Build Tools on Windows, the Xcode sysroot on macOS) just to build better-sqlite3 from source. Fixes #13.

Fixed

• better-sqlite3 build-from-source on Node 24 — better-sqlite3@^11 ships no prebuilt binary for Node 24 (ABI node-v137), so prebuild-install fell through to node-gyp rebuild, which needs a local C/C++ toolchain most users don't have. Bumped to better-sqlite3@^12, which adds Node 24 (and newer) prebuilds — the install now downloads a ready binary on every supported platform, no compiler required. No API changes; the database layer is untouched.

Changed

• Minimum Node.js is now 20 (was 18). better-sqlite3@12 dropped end-of-life Node 18 from its build matrix, so that is the new floor. engines, .nvmrc, the runtime version check (src/index.ts), and the README are aligned. Node 18 is itself past EOL.

[2.2.0] - 2026-06-30

Feature release. Turns the Debug Dashboard from a one-way display into a two-way control surface: a source can now expose interactive sliders/numbers whose values flow back to it, and both the user and the AI can tune a running program live. Also gives plots sender-controlled Y-axis scaling, separates a gauge LED's colour from its text, stops dashboard flicker, and trims the npm package.

Added

• Interactive controls + /control back-channel — two new widget types, slider and number, are editable in the viewer; when the user changes one, the new value flows back to the source. This is the first path on which data travels from AiDex to the program (everything else is source → AiDex). The mechanism is a deliberately dumb, source-agnostic { id: value } store — it knows nothing about what a value means.
  • HTTP: POST /control (set one value, mirrors it onto the card and broadcasts to every viewer) and GET /control (the whole store as a flat { id: value } map, which the source polls at its own pace to learn the current set-points). Cleared together with their widgets via POST /panel/clear — there is no separate control-clear endpoint.
  • MCP: the AI can drive controls too. aidex_log gains control_get (read all control values) and control_set (change one — the same set-point the user's dashboard slider drives). So Claude can tune a live program — e.g. a barge-in threshold, a gain, a sample rate — and watch the effect, without touching the source.
  • New step field (slider/number increment per tick, default 1). New file: src/loghub/control-store.ts. First real consumer: the GeminiPod (ESP32) barge-in tuning; the same API works unchanged from C#, Python, or shell.
• Sender-controlled plot Y-axis — three new plot fields, all decided by the sender (the renderer only renders):
  • scale — "linear" (default) or "log". Logarithmic scaling suits high-dynamic signals like audio levels, where quiet speech and a loud peak need to be visible at once. Bounds are lifted to ≥ 1.
  • autoMin — the plot's lower bound follows the data minimum (the ceiling stays fixed at max), so a noise floor sits at the bottom edge and the full plot height goes to the signal above it.
  • decimals — decimal places in the footer (cur/min/max/avg); 0 for integers. The recipe that finally made an audio-level plot readable across the whole loudness range: scale:"log" + autoMin:true + fixed max + decimals:0.
• Panel-Dashboard user guide (docs/loghub-panel-dashboard.md) — full walkthrough: stream vs. dashboard, quickstart, the complete HTTP API (display + control back-channel), all six widget types, a field reference checked against panel-types.ts, plot-scaling deep-dive, best practices from real (ESP32) use, and an end-to-end audio-dashboard example.

Changed

• Gauge LED colour separated from its text — a new state field drives a gauge's LED colour ("ok"/"warn"/"error"/…) independently of value, which stays the free display text. Previously the status word was the displayed text, so you were stuck looking at a literal "WARN"/"OK". Now the LED can be red while the card shows whatever text you want.
• npm package trimmed — .npmignore now ships only build/ plus the postinstall hook; docs/, the CHANGELOG, test scripts, and source are excluded from the published tarball. Smaller install, nothing functional removed. (scripts/verify-package.ps1, scripts/check-npm-auth.ps1 updated to match.)

Fixed

• Dashboard no longer flickers on unchanged values — updateCardValue re-rendered (and flashed) label/progress/gauge cards on every sample, even when the value was identical, causing constant flicker at high update rates. Cards now redraw only when the value actually changes.
• aidex_global_guideline list token overflow — list dumped the full text of every guideline (~1,246 lines / 68 KB for 15 guidelines), overflowing the MCP tool result so the content spilled to the swap file instead of the chat, making the overview unusable. list now renders a compact one-line-per-guideline index (key — short description (updated date)) via a new summarizeGuideline() helper; use get with a key to read the full text. 1,246 → 17 lines.

[2.1.2] - 2026-06-02

Feature + stability release. Adds the live Debug Dashboard and closes the WebSocket memory leak the v2.1.0/2.1.1 fixes had missed.

Added

• Debug Dashboard (Panel API) — a live, fixed-slot dashboard alongside the scrolling log stream. External programs POST /panel with { id, type, value, group? }; sending the same id again overwrites the value in place instead of scrolling away. Built for high-frequency / repeated values (audio levels, buffer fill, FPS, sensors).
  • Four widget types: label, progress (with warn/crit threshold colouring), gauge (radial tachometer in the MSI Afterburner / ASUS GPU Tweak style, or a pulsing status LED for string values), and plot (real-time line graph in the HWiNFO style with grid + min/max/avg).
  • Endpoints POST /panel, POST /panels (batch), POST /panel/clear. The server keeps the last state per id, so a freshly-connected or reloaded viewer gets the full dashboard snapshot immediately; cards with no update for ~3 s grey out as "stale". Clear is a full reset — a source reappears only if it re-sends widgets with their type.
  • New Viewer Debug tab (Tokyo-Night cockpit look). Plots are redrawn on a single requestAnimationFrame tick so audio-rate updates stay smooth. All panel broadcasts use the WS backpressure guard.
  • New files: src/loghub/panel-types.ts, src/loghub/panel-store.ts.
• Showcase demo (scripts/demo-dashboard.mjs) — an endless animation of all widget types (audio waveform, GPU gauges drifting through their zones, a signal generator cycling sine → sawtooth → triangle → square, latency spikes). A ▷ Demo button on the Debug tab copies the run command to the clipboard. Launcher: scripts/demo-dashboard.ps1.

Fixed

• WebSocket backpressure leak in broadcastTreeUpdate (src/viewer/server.ts): the v2.1.0 fix added a bufferedAmount guard to broadcastLogEntry and broadcastTaskUpdate, but broadcastTreeUpdate still called client.send() with no guard. On actively-changing projects (chokidar fires continuously while files churn) it broadcast full code + all trees (~0.5 MB/frame); a slow viewer client let ws's internal send-queue grow without bound → 50+ GB committed. This was the dominant leak the v2.1.0 fix missed. The same guard (drop + wsDropCounts + rate-limited stderr) is now applied to broadcastTreeUpdate and broadcastFocusTab. Proof: scripts/test-tree-backpressure.mjs — unguarded 2559 MB buffered, guarded 1 MB.
• Dashboard layout flicker: latency spikes / growing numbers made plot-stats wrap and value rows widen, changing card height and reflowing every card below. Plot-stats are now fixed-height nowrap cells, value rows are nowrap, and the grid uses a stable grid-auto-rows baseline.
• Plot-stats overlap: a cur value with its unit (e.g. -19.71 dB) ran into the next stat. Stats are now key+value cells with a fixed gap; long values clip with an ellipsis instead of colliding.

Tests

• scripts/test-panel-store.mjs (25 checks: validation, all widget types, plot ring + array frames + NaN filter + caps, snapshot, clear) and scripts/test-panel-http.mjs (8 checks: endpoints + error handling) cover the new panel layer. Backpressure proofs: scripts/test-backpressure.mjs, scripts/test-tree-backpressure.mjs.

[2.1.1] - 2026-05-31

Bugfix release. C and C++ functions were silently dropped from the index — aidex_signature and aidex_signatures returned types but zero methods for every C/C++ file.

Fixed

• C/C++ function names were never extracted (src/parser/extractor.ts): The generic method extractor only looked for a direct identifier child of a function_definition node. But the tree-sitter-c/c++ grammar never places the name there — it nests it under a declarator chain: function_definition → (pointer_declarator)* → function_declarator → identifier. As a result name stayed null and every function was discarded, so C/C++ files showed types but no methods at all.

  Fix: A new findCFunctionName() helper walks the declarator field chain (through any number of pointer_declarator / reference_declarator wrappers) down to the function_declarator and reads the real name. Applied only on the c / cpp path — other languages are untouched.

  Verified: loghub_client.c went from 0 → 11 functions extracted (including pointer-return functions like uint8_t *slot_at(...)); regression-tested against C#, TypeScript and C++ with no change to their output.

  Action required: Already-indexed C/C++ projects keep the old (empty) result in their DB because the file hash is unchanged and won't auto-reindex. Force a re-index with aidex_init (or remove + update the affected files) to pick up the now-extracted functions.

[2.1.0] - 2026-05-19

Stability release. Two independent root causes that crashed the embedding pipeline in production — one silently killing the entire MCP server process, the other silently generating impossible memory allocations — are now both permanently prevented.

Fixed

• ONNX OOM killed the entire MCP server process: Projects with large generated files (Blazor WASM bundles, large C++ translation units, minified JS) caused the ONNX runtime to request allocations of 67–93 GB of RAM. Because the ONNX model ran inside the MCP server process, the resulting std::bad_alloc killed the server — disconnecting Claude from every tool and requiring a manual /mcp reconnect. Root cause: the embedding model ran directly in the MCP server's Node.js heap with no crash isolation.

  Fix — Worker process isolation (src/commands/init.ts + src/embeddings/embed-worker.ts): aidex_init({ embeddings: true }) now spawns a dedicated child process via spawn(process.execPath, [workerPath], { stdio: ['pipe','pipe','pipe'] }) for each embedding run. The worker reads the project path from stdin as JSON, runs the full pipeline, and writes the result to stdout. When ONNX OOMs, only the worker dies — the MCP server survives and reports a Warning in the aidex_init result instead of crashing. A 10-minute timeout kills runaway workers automatically.

  Verified: projects that previously killed the server (YouTubeVoiceOver, CiscoWebExTranslator/cpp, UcHome) now complete with a Warning rather than disconnecting Claude.

• ONNX OOM on large source files: Even with worker isolation, files larger than ~50 KB reliably triggered the ONNX Non-zero status code returned while running Add node error inside the worker, causing the worker to exit with code 3228369023 (Windows STATUS_ACCESS_VIOLATION). Root cause: the jina-code model's attention layers allocate O(n²) memory relative to token count — a 63 KB file with deeply nested methods produces a sequence too long for the model's fixed-size buffers.

  Fix — 25 KB file size limit (src/embeddings/pipeline.ts): Methods, types, and doc-sections from files larger than 25 KB are silently skipped during embedding. File sizes are resolved once per file via statSync and cached for the run. The limit applies in collectCodeWrites(), collectDocsWrites(), and updateFile() — covering both full re-index and incremental update paths. Files below the limit are embedded normally; files above it are excluded without error or warning.

  Verified: TwSudoku (Blazor WASM JS bundles >1 MB + 14 C# files), YouTubeVoiceOver (63 KB TranscriptProcessor.cs), and UcHome (ESP32 SDK managed components) all complete cleanly. A 6-worker parallel stress test across all 252 registered projects completed in 33 minutes with 0 crashes and 0 errors.

Added

• src/embeddings/embed-worker.ts: Standalone embedding worker entry point. Reads { projectPath, force? } from stdin, calls createRealModule().enable() + .indexProject(), writes { ok, embedded, skipped, removed, durationMs } (or { ok: false, error }) to stdout, then exits. Has no IPC channel — crash isolation relies entirely on process separation.

Changed

• aidex_init with embeddings: true no longer runs the embedding model in-process. The spawn-based worker adds ~100 ms overhead per project but eliminates the possibility of an ONNX crash affecting the MCP server.

[2.0.1] - 2026-05-05

Bug-fix release. Two issues that surfaced when running aidex_init({ embeddings: true }) in real-world workflows with mixed-age project DBs and concurrent calls.

Fixed

• Embedder crashed on legacy project DBs with "no such column": Projects whose index.db predated the v1.19a / v1.15 / archive-summary migrations crashed aidex_init({ embeddings: true }) with SqliteError: no such column: m.body_text (or tasks.summary / note_history.summary). The embedder opens each project's index.db read-only, so the writeable migrateLegacySchema() path in db/database.ts never ran on those DBs — and the SELECTs in embeddings/store.ts referenced the missing columns directly. openProjectIndexDb() now opens the DB writeable briefly to apply the same idempotent ALTER TABLEs, then reopens read-only. Belt-and-suspenders: readMethods / readMethodsForFile / readAllTasks / readNoteHistory now check PRAGMA table_info and substitute NULL for missing columns, so even a locked / read-only DB no longer crashes — it just produces embeddings without the optional context fields. migrateLegacySchema itself was missing the note_history.summary ALTER and got it added.
• Concurrent indexProject calls loaded N copies of the embedding model: When several aidex_init (or indexProject) calls fired before the embedder was warm — a typical user pattern after global enable — every caller raced past the this.embedder == null check in RealEmbeddings.getEmbedder() and started its own createEmbedder(). Each load pulled the ~7 GB ONNX model into RAM. 14 parallel calls reproducibly produced ~98 GB RSS. Fix: the in-flight load is now cached as a Promise<Embedder>, so concurrent callers await the same load and the model is loaded exactly once. Verified with a 14-parallel stress harness: peak RSS during the run dropped from 12 GB to 92 MB; end-of-run RSS dropped from 98 GB to 12 GB.

Added

• Regression tests for the above (tests/embedder-fixes.test.js): 12 jest tests covering legacy-schema migration on each affected table, defensive read paths when columns are missing, idempotency of openProjectIndexDb, and the Promise-cache mechanism that prevents the N-way embedder load.

[2.0.0] - 2026-05-04

Major release — AiDex grew a brain. Semantic search across code, docs, and workspace items via locally-run embeddings (jina-code, 768d). Optional LLM layer for multilingual queries and reranking. New Settings tab in the Viewer. Schema migrated to v1.2 (additive — existing indexes keep working).

19 commits, ~8400 lines. Restore points: tag v1.18.0-pre-bugsweep (clean v1.18.0 working tree) and v2.0-pre-cleanup (mid-branch checkpoint).

Added

• Method body storage (v1.19a, prerequisite for embeddings): aidex_init accepts store_bodies (or embeddings) flag. When enabled, full method bodies are stored in methods.body_text for snippet display and re-embedding. Bodies >8000 chars are truncated to head + tail. Setting persists in metadata so aidex_update continues to store bodies on incremental updates. Schema migrated to v1.1 (additive, backward compatible).

• Embeddings module skeleton (v1.19b): New encapsulated module src/embeddings/ with stable public API (getEmbeddings()), lazy loading, and additive schema migration on ~/.aidex/global.db. Adds the embeddings table and embedding_model_id / embedding_dim / embedding_version / last_full_embed_at / files_changed_since columns to projects. Model registry includes jina-code (default), nomic-text, bge-small. Heavy dependencies (@xenova/transformers, sqlite-vec) are declared as optionalDependencies and only loaded when enable() is called — AiDex still installs and runs without them.

• Code embeddings (v1.19c): aidex_init({ embeddings: true }) and getEmbeddings().indexProject() produce semantic vectors for every method and type via three-tier chunking (signature + doc-comment + weighted identifier bag with string literals). Vectors live in a vec0 virtual table per dimension (vec_embeddings_<dim>); metadata in the embeddings table joins them by rowid. Hash-based skip-on-no-change avoids redundant model calls on re-runs (~500ms for a fully cached re-index). Default model jina-embeddings-v2-base-code is downloaded into ~/.aidex/models on first use. Indexing AiDex itself (402 methods + 132 types) takes ~30s on CPU, ~600ms on subsequent runs.

• Workspace embeddings (v1.20): Tasks, task logs, the active session note, and all archived note-history entries are embedded alongside code. commands/task.ts and commands/note.ts fire fire-and-forget onTaskChanged / onNoteChanged hooks that re-embed only the affected items. Deleted tasks and stale anchors are pruned automatically (pruneEmbeddingsByType). Combined index for AiDex now covers 402 methods + 132 types + 144 workspace items (35 tasks, 85 task-logs, 1 active note, 23 history entries). Semantic queries like "release process for npm publishing" return the right tasks; "visualization of progress in the browser" pinpoints the startProgress / getProgressHTML / stopProgress methods in src/viewer/progress.ts.

• Docs embeddings (v1.21): All Markdown / MDX files surfaced by project_files.type='doc' (README, CHANGELOG, CLAUDE.md, docs/**, plan/release notes, etc.) are split at heading boundaries and embedded as doc-sections. The chunker keeps code fences intact, sub-splits oversize sections with overlap, and prepends the document title for context — so a single retrieval is "Auth: token rotation", not just "token rotation". Stale sections and removed files are pruned via pruneDocSectionsExcept. AiDex's own index grew to 984 embeddings (534 code + 144 workspace + 306 doc-sections from 10 files). Hybrid retrieval works as designed: a query like "how to write logs from external programs" returns the README's ## Log Hub section first, then the log method in src/commands/log.ts.

• aidex_search MCP tool (v1.22): New tool for semantic / exact / hybrid retrieval across embedded code, docs, and workspace items. Three modes: semantic (pure vec0 KNN against the query embedding), exact (identifier match — same as aidex_query), and hybrid (default — fuses both via Reciprocal Rank Fusion with k=60). Filters: scope (current project / all enabled / linked), project_filter (glob over project paths), source_kinds (code / docs / workspace), source_types (method, type, doc-section, task, task-log, note, note-history), and k. Output is a ranked list with file:line, snippet (≤180 chars), distance, and project name when scope is global. The module auto-loads on first search call — no explicit enable() needed for read-only use.

• Viewer Search tab (v1.24): Interactive embedding search inside the project Viewer. New "Search" tab in the detail panel with a debounced text input (350 ms), mode selector (Hybrid / Semantic / Exact), source-kind checkboxes (Code / Docs / Workspace), and a k slider (5–50). Results stream in via WebSocket (searchEmbeddings → searchResults with a requestId to discard stale responses). Hits render as clickable cards with a kind/type badge, distance score, file:line, and snippet — clicking a code/docs hit opens the file in the Overview tab; workspace hits switch to the Tasks tab. State survives tab switches and the panel rebuilds itself if other tabs overwrite the shared #detail container.

• Code chunk display split: The chunker now produces two separate texts — embeddingText (still uses weighted token repetition for the model) and displayText (compact Identifiers: files (8), Filter (7), path (7), … form for the DB and viewer). Search results stop showing the noisy "files files files files files" pattern. The migrateDisplayText() helper rewrites source_text for all existing code embeddings without re-running the model — ~130 ms for AiDex's 534 vectors.

• Embeddings lifecycle (v1.23): Embeddings now track and refresh themselves automatically as code, docs, and workspace items change. aidex_update and aidex_remove fire fire-and-forget hooks that re-embed the affected file (or drop it). The pipeline uses content-hash skip-on-no-change so a no-op edit costs nothing. Anchors that disappear (renamed methods, deleted sections) are pruned. aidex_session reports an embeddings status block (fresh / drifting / stale) with a human-readable hint — "55 files changed since last full embed, consider re-indexing". Model migration: migrate({ fromModel, toModel, apply }) re-embeds every project pinned to the source model after a transactional wipe; dry-run mode reports affected counts without touching anything. The lazy-loaded module stub now auto-promotes itself to the real implementation on hooks for projects that have embeddings enabled — previously the hook silently no-op'd in fresh processes until something explicitly called enable().

• Settings tab + aidex_settings tool: A single Settings tab in the viewer is now the home for all user-facing configuration — embeddings on/off and model choice, LLM provider / endpoint / API key / model, and the llm_send_code privacy switch. Live status probes show what's already active (e.g. "OpenAI / gpt-4o-mini from env"). A "Test connection" button verifies the LLM round-trip with latency. API keys persist to ~/.aidex/llm.json (chmod 600). The new MCP tool aidex_settings({ path, open: true }) opens the viewer and jumps straight to the Settings tab — so the user can say "open my AiDex settings" and the AI just calls it. First session after install/upgrade prepends a one-time welcome banner to aidex_session ("🎉 AiDex X.Y — new features available, open Settings to enable") that stops once the user opens Settings.

• LLM layer (v1.25): Optional intelligence layer over aidex_search. When an Anthropic / OpenAI / OpenRouter API key is configured, or a local Ollama is running, AiDex can: translate non-English queries into English (so "wie speichere ich Logs lokal" finds the right code), expand vague queries into 2-4 concrete subqueries (RRF-merged), and rerank top-N retrieval candidates. New llm parameter on aidex_search (auto / off / translate / rerank / expand+rerank). Per-project privacy switch llm_send_code (default off): when off, only the user's literal query and metadata (paths, names, anchors) are sent to the LLM — never source bodies, doc sections, or notes. New aidex_init parameters: llm_endpoint, llm_model, llm_send_code. Credentials discovered in this order: project config → env vars → aidex_global_guideline keys → ~/.aidex/llm.json → local Ollama probe. If nothing is found, AiDex stays in pure-embeddings mode. Encapsulated in src/llm/ with a stable getLlm() proxy and a strict safety.ts chokepoint that defensively asserts no body content leaks when sendCode=false.

• HuggingFace backend: Sixth LLM provider option, alongside Anthropic / OpenAI / OpenRouter / Ollama / Custom. Targets the OpenAI-compatible Router endpoint (https://router.huggingface.co/v1); auto-detected from HF_TOKEN (HuggingFace standard) or HUGGINGFACE_API_KEY env vars. Suggested model dropdown includes Llama-3.1-8B-Instruct, Llama-3.3-70B-Instruct, Qwen2.5-Coder-32B-Instruct, Qwen2.5-72B-Instruct, Mistral-7B-Instruct-v0.3, and DeepSeek-V3 — but free typing works for any HF model. 503 cold-start responses surface as an explicit error with the model-loading hint. Endpoint URLs containing huggingface.co are recognised by inferBackend so per-project overrides and ~/.aidex/llm.json entries route correctly.

Fixed

• aidex_search: semantic results no longer depend on k: With small k (e.g. k=5) the KNN candidate pool was only k * 3 = 15 vectors, while a k=10 call pulled 30. sqlite-vec's KNN is not strictly stable across k variations, so the user-facing top-5 was sometimes a different set than the top-5 of a k=10 call — occasionally returning "No matches" while k=10 returned plausible hits for the same query. Pinned the underlying KNN candidate pool to a floor of 60 (SEMANTIC_FETCH_FLOOR); the JS slice still trims to the user's requested k. Result: aidex_search k=5 is now a proper subset of k=10 for the same query.
• Viewer: task description rendering: Task descriptions in the viewer's Tasks tab were piped through escapeHtml() and rendered as a single wall-of-text — Markdown headings, lists, tables, and code fences collapsed into one unreadable line. Added a small inline Markdown renderer (no dependencies) that handles H1-H6, fenced + inline code, bold/italic/strikethrough, ordered/unordered lists, blockquotes, GFM tables, horizontal rules, and links. Anyone who writes multi-paragraph task descriptions sees them properly formatted now.
• Viewer: Settings tab auto-switch via aidex_settings({ open: true }): The auto-switch never fired because readPendingFocusTab and writePendingFocusTab in src/viewer/server.ts used require('better-sqlite3') inside a pure-ESM module ("type": "module"). Node threw ReferenceError: require is not defined, the surrounding silent try/catch returned null, and the metadata round-trip via ~/.aidex/global.db was a no-op. Replaced with the existing top-level import Database from 'better-sqlite3' and a getGlobalDbPath() helper using process.env.USERPROFILE. The same anti-pattern in writeLlmConfigFile was fixed alongside.

[1.18.0] - 2026-04-25

Added

• HCL/Terraform support (#9): Indexes .tf, .tfvars, and .hcl files — now 12 supported languages
  • Blocks (resource, module, variable, output, data, locals, provider, ...) → types with dotted names (e.g. resource.aws_instance.web)
  • Function calls → methods
  • Attributes → properties
  • Block labels (incl. keyword-like names: default, root, type, data) indexed as searchable items
  • Terraform projects auto-discovered via *.tf and .terraform.lock.hcl markers in global_init
  • .terraform/ excluded from indexing
  • Uses @tree-sitter-grammars/tree-sitter-hcl grammar

Changed

• tree-sitter upgrade (#8): Bumped tree-sitter from 0.21 to 0.25 and all 10 grammar packages to latest — enables newer grammars requiring tree-sitter ^0.25.0
• Parser refactor: Centralized grammar mapping in GRAMMAR_MAP — adding a new language is now a one-line change
• File watcher: Viewer's live re-indexing watcher now uses parser.isSupported() instead of a hardcoded extension regex — automatically tracks every supported language

Notes

• Node.js 22+ recommended: tree-sitter 0.25 requires native compilation; Node 24 not yet compatible. .node-version pins to 22.
• tree-sitter-c-sharp pinned to ^0.23.1: 0.23.5+ switched to ESM-only with top-level await, breaking CJS imports.

[1.17.1] - 2026-04-20

Changed

• Node.js 18+ enforcement: Server now exits immediately with an OS-specific install hint when run on Node <18, instead of crashing later on native modules
• Prerequisites section: Added to README with install commands for macOS (brew/nvm), Linux (nvm), and Windows
• .nvmrc: Added so nvm use picks the correct version automatically

[1.17.0] - 2026-04-07

Added

• Task Scheduler: Tasks can now have due dates (due), repeat intervals (interval), actions (task_action), and auto-execute flag (auto_go)
  • Due dates: Relative ("3d", "1w") or ISO date ("2026-04-10")
  • Recurring tasks: Automatically advance due date by interval after each trigger
  • One-shot tasks: Due date cleared after trigger
  • Cross-project: Overdue tasks reported at every aidex_session call — even from other projects
  • Global mirror: scheduled_tasks table in ~/.aidex/global.db for fast cross-project lookups
  • Auto-migration: Existing databases get new columns automatically
• Test suite: First tests for AiDex — 26 tests covering scheduler logic, task CRUD, global sync

[1.16.3] - 2026-03-26

Fixed

• Monorepo support: Removed **/packages/** from DEFAULT_EXCLUDE — was incorrectly blocking indexing of JS/TS monorepo workspaces (pnpm, npm workspaces, etc.) (#4)

[1.16.2] - 2026-03-20

Changed

• Auto-setup instructions overhaul: Complete rewrite of the CLAUDE.md/GEMINI.md block installed by aidex setup
  • Added Log Hub section with full usage guide (init, query, HTTP API, Viewer integration)
  • Added missing query parameters: modified_before, file_filter, type_filter
  • Added show_progress for global_init
  • Added task summary field documentation
  • Added note summary field documentation
  • Updated tool count from 28 → 30
  • Updated Viewer description to include Logs tab
  • Added "Debug my app" to Question → Right Tool table
• README: Added aidex_log to Available Tools table, updated tool counts to 30
• Projekt-CLAUDE.md: Updated tool count to 30

[1.16.1] - 2026-03-20

Added

• Log Hub consume pattern: New consume parameter on aidex_log query — returned entries are removed from the buffer, ideal for polling without duplicates
• Viewer: Clear Logs button: "Clear" button in the Logs tab to reset the log display
• Viewer: WebSocket auto-reconnect: Viewer automatically reconnects when WebSocket connection drops (2s retry)
• LogHub Developer Guide: Added comprehensive integration guide to project CLAUDE.md with code examples for C#, Python, JavaScript, C/C++, PowerShell

Fixed

• Log entry data: null display: Entries with data: null no longer show "null" text in Viewer and MCP output

[1.16.0] - 2026-03-20

Added

• Log Hub (aidex_log): Universal logging system — any program (C#, Python, Node, etc.) sends logs via HTTP POST to AiDex, queryable by the LLM via MCP tool. Zero-cost when not used — no server, no buffer, no resources until init is called.
  • HTTP Server on port 3335 (configurable): POST /log (single), POST /logs (batch), GET /health
  • Ring Buffer: In-memory circular buffer (default 10,000 entries), oldest entries overwritten
  • Query: Filter by since, level, source, contains, limit — newest first
  • Write: LLM can inject entries (source: "claude")
  • Persistence: Optional SQLite storage with 7-day auto-cleanup
  • Viewer integration: New "Logs" tab with WebSocket live-stream, level/source/text filters, auto-scroll

[1.15.0] - 2026-03-17

Added

• Task summaries: Tasks now support a summary field — a one-sentence table-of-contents entry (~150 chars) that the AI writes on create/update. aidex_tasks shows summaries inline so you can scan the backlog without reading full details.
• Note history summaries: Archived notes now get an optional summary field. When a note is overwritten or cleared, a summary can be provided for the archive. aidex_note with history: true shows summaries (with fallback to truncated preview for older notes). Search also matches summaries.
• Auto-migration: Existing databases are automatically upgraded with ALTER TABLE ADD COLUMN summary — no manual migration needed.
• Viewer integration: Task summaries shown in italic between title and description in the browser viewer.

[1.14.0] - 2026-03-10

Added

• aidex_global_guideline: New tool (#28) — persistent key-value store in ~/.aidex/global.db for AI guidelines and coding conventions. Store named instructions like "review" → review checklist, "release-prep" → release steps. Actions: set, get, list, delete. Works without prior global_init.
• Viewer file size limit: getFileContent() now refuses files larger than 1 MB — prevents browser from freezing on large binary or generated files

Fixed

• Command injection in Linux screenshot tools: All execSync calls with shell string interpolation replaced with execFileSync using argument arrays — window titles, file paths and IDs are no longer injectable via shell
• Global query cache grows unbounded: Cache entries are now evicted on write when they exceed the 5-minute TTL — prevents memory leak in long-running sessions
• Viewer race condition on file change: pendingChanges set is now snapshotted and cleared before processing — new events arriving during async re-indexing are no longer silently dropped
• Viewer buildTree() N+1 queries: Correlated subqueries for methods and types counts replaced with LEFT JOIN — single query instead of one subquery per file
• WebSocket unknown message type: Viewer now sends an error response for unrecognized message types instead of silently ignoring them
• Viewer taskId not validated: updateTaskStatus now checks Number.isInteger(taskId) before processing
• Viewer mode not whitelisted: getTree message mode is now constrained to 'code' | 'all' — arbitrary values no longer passed through
• getProjects() SQL injection via tag/namePattern: escapeLikeTerm() now applied to both filter parameters in global-database.ts
• Silent fails in viewer and global DB: catch {} blocks now log errors via console.error
• Git status refresh on every file event: Added 5-second minimum interval between git status refreshes — reduces git subprocess spam during rapid file saves
• Global query cache not invalidated after init/update: aidex_init and aidex_update now call invalidateGlobalCache() so global searches immediately see fresh data

Refactored

• screenshot/shared.ts: New module with centralized hasTool() and runPowerShell() — both use execFileSync (no shell). Imported by platform-win32.ts, platform-linux.ts, and post-process.ts — eliminates duplicate implementations
• normalizePath(): Private duplicates in global-database.ts and git-status.ts removed — both now import from commands/shared.ts
• escapeLikeTerm(): Exported from commands/shared.ts and used consistently across all LIKE queries
• macOS sips output parsing: Replaced shell pipe (| tail -1 | awk) with regex on direct sips output

[1.13.0] - 2026-03-09

Added

• Screenshot optimization: New scale (0.1-1.0) and colors (2/4/16/256) parameters for aidex_screenshot
  • Reduces file size up to 95% (e.g., 108 KB → 5 KB with scale: 0.5, colors: 2)
  • Black & white mode (colors: 2) ideal for text-only screenshots — saves thousands of tokens
  • Cross-platform post-processing: Windows (System.Drawing), macOS (sips + ImageMagick), Linux (ImageMagick)
  • Size reporting: shows original → optimized size and percentage saved
• LLM auto-optimization strategy: Tool description guides AI assistants to start with aggressive settings, retry if unreadable, and remember working settings per app during the session

[1.12.1] - 2026-03-07

Added

• Update notifications: aidex_session now shows "What's New" when AiDex was updated since the last session
  • Compares installed version with last_seen_version stored in project DB
  • Shows highlights + changelog link, only once per version update

[1.12.0] - 2026-03-07

Added

• Auto-setup on install: npm install -g aidex-mcp now automatically registers AiDex with all detected AI clients and installs AI instructions (CLAUDE.md, GEMINI.md)
  • Opt-out via AIDEX_NO_SETUP=1 or CI environment variable
  • Graceful fallback: shows manual hint if auto-setup fails
• Comprehensive AI instructions: The CLAUDE.md block installed by aidex setup now covers all 27 tools
  • Decision tree: "Do I want to search code? → .aidex/ exists? → STOP, use AiDex"
  • Explicit ❌/✅ examples (never Grep when .aidex exists)
  • Search modes explained (exact/contains/starts_with)
  • Session notes, task backlog, global search, screenshots — all with examples
• Duplicate detection: aidex setup skips CLAUDE.md/GEMINI.md if manual AiDex instructions already exist (avoids double entries)

Changed

• Refactored commands: Extracted shared utilities into shared.ts and global-shared.ts
  • validateIndex(), noIndexError(), withDatabase(), withProjectDb() — ~200 lines of boilerplate eliminated across 10 command files
  • withGlobalDb(), EMPTY_TOTALS — 4 global commands refactored
• README: Expanded "Make your AI use it" section with full best-practice instruction block

Fixed

• DB transactions: clearFileData() and bulkInsert*() now wrapped in transactions
• N+1 query: Batch getOccurrencesByItems() replaces per-item queries
• Stats query: getStats() reduced from 7 queries to 1
• SQL injection: global-signatures.ts — t.kind = '${kind}' → parameterized query
• Session: Eliminated duplicate getMetadata calls
• Tasks: Added null-check for tableInfo result

[1.11.0] - 2026-03-07

Added

• Global Search: Search across ALL indexed projects at once — 5 new tools
  • aidex_global_init — Scan directory tree, register indexed projects in ~/.aidex/global.db, detect unindexed projects by project markers (.csproj, package.json, Cargo.toml, etc.)
  • aidex_global_status — List all registered projects with stats, sortable by name/size/recent
  • aidex_global_query — Cross-project term search (exact/contains/starts_with) with in-memory session caching (5-min TTL)
  • aidex_global_signatures — Search methods/types by name across all projects, filterable by kind
  • aidex_global_refresh — Update stats and remove stale projects
  • Uses SQLite ATTACH DATABASE for zero-copy queries — each project DB remains the single source of truth
  • exclude parameter on global_init to skip external repos (e.g., ["llama.cpp"])
  • Auto-updates global registry after aidex_init / aidex_update
• Bulk Indexing: global_init can auto-index all unindexed projects in one call
  • index_unindexed: true — Auto-index projects with ≤500 code files
  • Large projects (>500 files) are listed separately for user decision
  • File count estimation uses code-only extensions (matches what init() actually processes)
• Progress UI: Browser-based progress display for bulk indexing
  • show_progress: true — Opens http://localhost:3334 with live progress bar
  • Server-Sent Events (SSE) for real-time updates
  • Shows per-project status (indexing/done/error), progress bar, scrolling log
  • Dark theme, auto-closes after completion
• Project deduplication: Parent projects that contain sub-projects are automatically removed
  • e.g., AudioGrabber/ is skipped when AudioGrabber/AudioGrabber/ and AudioGrabber/AudioGrabber2/ exist
  • Existing duplicates in global DB are cleaned up on next global_init run
  • Reduced test index from 215 to 167 projects (48 parent-duplicates removed)
• Extended excludes: Better handling of embedded runtimes and external code
  • init.ts: Added **/site-packages/**, **/Lib/**, **/fdk-aac/** to DEFAULT_EXCLUDE
  • global-init.ts: Added Python venvs, embedded Python runtimes (Python310-313), .cargo, packages, fdk-aac to DEFAULT_EXCLUDED_DIRS

[1.10.1] - 2026-03-07

Fixed

• npm package: Exclude token files and futureWork.md from published package
• gitignore negation patterns: Filter out ! negation patterns in .gitignore to prevent excluding all files
  • Negation patterns (e.g., !.vscode/settings.json) were passed to minimatch, which interpreted ! as "NOT this pattern" — matching ALL files
  • This caused the entire index to be purged after initialization in projects with negation patterns (common in monorepos)

[1.10.0] - 2026-02-17

Added

• Note History: Archived notes are now searchable across sessions
  • Old notes are automatically archived when overwritten or cleared
  • history: true parameter to browse archived notes (newest first)
  • search: "term" parameter to search note history (case-insensitive)
  • limit parameter to control how many history entries are returned (default: 20)

[1.9.1] - 2026-02-10

Added

• Rect Screenshot Mode: New mode: "rect" for coordinate-based screen capture
  • Specify exact x, y, width, height in pixels
  • Useful with accessibility bounds (e.g., from WinfoMCP get_element_details)

Fixed

• Region screenshot flicker on Windows: Fixed visual flicker during interactive region selection

[1.9.0] - 2026-02-06

Added

• Cross-Platform Screenshots: New aidex_screenshot tool for capturing screenshots directly from AI assistants
  • 4 capture modes: fullscreen, active_window, window (by title), region (interactive selection)
  • Cross-platform: Windows (PowerShell + .NET), macOS (screencapture), Linux (maim/scrot)
  • Multi-monitor support (select monitor by index)
  • Delay parameter (wait N seconds before capture)
  • Default: Saves to temp directory with fixed filename (overwrites for quick iteration)
  • Custom filename and save path supported
  • Returns file path so AI can immediately Read the image
  • No project index required - standalone utility
• Window Listing: New aidex_windows tool to list all open windows
  • Shows title, PID, and process name
  • Optional substring filter (case-insensitive)
  • Helper for aidex_screenshot mode="window"

Technical

• New directory module: src/commands/screenshot/ with platform-specific implementations
• Windows: PowerShell scripts written to temp .ps1 files (avoids quoting issues with inline C#)
• macOS: Uses native screencapture command (interactive selection built-in)
• Linux: Uses maim (preferred) with scrot fallback; xdotool/wmctrl for window operations
• Synchronous delay via Atomics.wait (Node >= 18)

[1.8.1] - 2026-02-02

Added

• Cancelled status for tasks: backlog → active → done | cancelled
  • Cancelled tasks preserved as documentation (not deleted)
  • Viewer: collapsible ❌ Cancelled section with strikethrough styling

Fixed

• aidex_update now respects exclude patterns: Files in build/, node_modules/, .gitignore patterns are rejected
  • Previously the viewer's file watcher could re-index excluded files via aidex_update

Technical

• Auto-migration: existing tasks table CHECK constraint updated to include cancelled
• Exported DEFAULT_EXCLUDE and readGitignore from init.ts for reuse

[1.8.0] - 2026-02-02

Added

• Task Backlog: Built-in project task management persisted in AiDex database
  • aidex_task - Create, read, update, delete tasks with priority, tags, and descriptions
  • aidex_tasks - List and filter tasks by status, priority, or tag
  • Auto-logging: Status changes and task creation are automatically recorded in task history
  • Manual log entries: Add notes to any task with the log action
  • Priorities: high (🔴), medium (🟡), low (⚪)
  • Statuses: backlog → active → done
  • Sort order support for custom ordering within same priority
• Viewer Tasks Tab: Interactive task management in the browser viewer
  • Priority-colored task list grouped by status
  • Done toggle directly from the viewer
  • Tag display

Technical

• New database tables: tasks and task_log with auto-migration
• Tasks survive between sessions (persisted in SQLite)

[1.7.0] - 2026-02-01

Added

• Gemini CLI support: aidex setup now detects and registers AiDex with Gemini CLI (~/.gemini/settings.json)
• VS Code Copilot support: aidex setup now detects and registers AiDex with VS Code (mcp.json with "servers" key and "type": "stdio")

Changed

• JSON client config is now flexible: supports custom server key (serversKey) and extra fields (extraFields) per client
• Updated README with Gemini CLI and VS Code Copilot config examples

[1.6.1] - 2026-02-01

Fixed

• MCP Server version: Now reads version dynamically from package.json (was hardcoded to 1.3.0)
• aidex setup for local installs: Detects if aidex is globally available; falls back to node /full/path/index.js when not installed globally

[1.6.0] - 2026-02-01

Added

• Auto CLAUDE.md instructions: aidex setup now installs AI instructions in ~/.claude/CLAUDE.md
  • Tells Claude to auto-run aidex_init when no .aidex/ exists
  • Provides tool usage guide (prefer AiDex over Grep/Glob)
  • aidex unsetup cleanly removes the instructions block
• Idempotent setup: Re-running aidex setup updates existing config without errors

[1.5.2] - 2026-02-01

Fixed

• aidex setup for Claude Code: Uses claude mcp add --scope user instead of editing settings.json directly
• Claude Desktop, Cursor, Windsurf still use JSON config editing

[1.5.1] - 2026-01-31

Fixed

• aidex setup: Now creates config file if client directory exists but config is missing (e.g. fresh Claude Code install)

[1.5.0] - 2026-01-31

Added

• aidex setup: Auto-register AiDex as MCP server in all detected AI clients
  • Supports: Claude Code, Claude Desktop, Cursor, Windsurf
  • Cross-platform: Windows, macOS, Linux
• aidex unsetup: Remove AiDex registration from all clients
• Postinstall hint: Shows Run "aidex setup" after npm install

[1.4.2] - 2026-01-31

Added

• npm package: Published as aidex-mcp on npm (npm install -g aidex-mcp)
• Dual CLI commands: Both aidex and aidex-mcp work as command names
• npm-publish.bat: Script for easy npm publishing

Changed

• README updated with npm install instructions

[1.4.1] - 2026-01-31

Fixed

• Git Status for Subfolder Projects: Viewer now correctly shows git status for projects that are subdirectories of a git repo (e.g., a library inside a monorepo)
  • isGitRepo() now uses simpleGit().checkIsRepo() instead of checking for .git directory — traverses parent dirs
  • New toProjectRelative() helper maps git-root-relative paths to project-relative paths
  • Files outside the project subfolder are properly filtered out

[1.4.0] - 2026-01-31

Breaking Changes

• Renamed from CodeGraph to AiDex: Package name, MCP server name, and all internal references updated
  • MCP prefix changes from mcp__codegraph__ to mcp__aidex__ (requires config update)
  • Index directory changed from .codegraph/ to .aidex/
  • Batch scripts renamed: codegraph-scan.bat → aidex-scan.bat, codegraph-init-all.bat → aidex-init-all.bat
  • Old .codegraph/ directories can be safely deleted

Added

• Automatic Cleanup: aidex_init now removes files that became excluded (e.g., build outputs)
  • Reports filesRemoved count in result
  • Uses minimatch for proper glob pattern matching
• Git Status in Viewer: File tree now shows git status with cat icons
  • 🟢 Pushed (committed and up-to-date)
  • 🟡 Modified (uncommitted changes)
  • 🔵 Staged (added to index)
  • ⚪ Untracked (new files)
• aidex-init-all.bat: New batch script to recursively index all git projects in a directory tree

Changed

• Added minimatch dependency for exclude pattern handling
• Updated all documentation (README, CLAUDE.md, MCP-API-REFERENCE) with correct MCP prefix info

[1.3.0] - 2026-01-27

Added

• Interactive Viewer: New aidex_viewer tool opens a browser-based project explorer
  • Interactive file tree (click to expand directories)
  • Click files to view signatures (types, methods)
  • Tabs: Code files / All files, Overview / Source code
  • Live reload with chokidar file watcher
  • WebSocket for real-time updates
  • Syntax highlighting with highlight.js
  • Runs on http://localhost:3333
• Recent Files Filter: New modified_since parameter for aidex_files
  • Find files changed in current session: modified_since: "30m"
  • Supports relative time (2h, 1d, 1w) and ISO dates

Changed

• Viewer auto-reindexes changed files before refreshing tree

Fixed

• Server version now correctly reports 1.3.0

[1.2.0] - 2026-01-27

Added

• Session Notes: New aidex_note tool to persist reminders between sessions
  • Write, append, read, and clear notes
  • Stored in SQLite database (survives restarts)
  • Use cases: handover notes, test reminders, context for next session
• Session Tracking: New aidex_session tool for automatic session management
  • Detects new sessions (>5 min since last activity)
  • Records session start/end times
  • Detects files modified externally (outside sessions)
  • Auto-reindexes changed files on session start
  • Returns session note if one exists

Changed

• Database schema: Added metadata table for key-value storage (session times, notes)

[1.1.0] - 2026-01-27

Added

• Time-based Filtering: New modified_since and modified_before parameters for aidex_query
  • Relative time: 30m, 2h, 1d, 1w
  • ISO dates: 2026-01-27 or 2026-01-27T14:30:00
  • Track line-level changes across updates
• Project Structure: New aidex_files tool to query all project files
  • File types: code, config, doc, asset, test, other, dir
  • Glob pattern filtering
  • Statistics by file type

Changed

• aidex_init now indexes complete project structure (all files, not just code)
• aidex_update preserves modification timestamps for unchanged lines (hash-based diff)
• Path normalization to forward slashes across all commands

Technical

• New project_files table in database schema
• New line_hash and modified columns in lines table
• Hash-based change detection for accurate timestamps

[1.0.0] - 2026-01-27

Added

• 11 Language Support: C#, TypeScript, JavaScript, Rust, Python, C, C++, Java, Go, PHP, Ruby
• Core Tools:
  • aidex_init - Index a project
  • aidex_query - Search terms (exact/contains/starts_with)
  • aidex_signature - Get file signatures (methods, types)
  • aidex_signatures - Batch signatures with glob patterns
  • aidex_update - Re-index single files
  • aidex_remove - Remove files from index
  • aidex_summary - Project overview with auto-detected entry points
  • aidex_tree - File tree with statistics
  • aidex_describe - Add documentation to summary
  • aidex_status - Index statistics
• Cross-Project Support:
  • aidex_link - Link dependency projects
  • aidex_unlink - Remove linked projects
  • aidex_links - List all linked projects
• Discovery:
  • aidex_scan - Find all indexed projects in directory tree
  • CLI commands: scan, init
• Technical:
  • Tree-sitter parsing for accurate identifier extraction
  • SQLite with WAL mode for fast, reliable storage
  • Keyword filtering per language (excludes language keywords from index)
  • 1MB parser buffer for large files

Infrastructure

• MCP Server protocol implementation
• MIT License
• Comprehensive documentation