ORT 1.24.x downloads CUDA provider .so from NuGet at postinstall,
but only for linux/x64. The process.arch guard correctly returns
false on arm64 (safe CPU fallback), but the prior comment implied
arm64 CUDA was supported. Clarify the actual state.
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
Address PR #300 review findings:
[CRITICAL] hasOrtCudaProvider() was checking the top-level
onnxruntime-node@1.24.3 but @huggingface/transformers loads its own
nested onnxruntime-node@1.21.0 at runtime. The guard inspected the
wrong binary, so the native crash was not prevented.
Fix: resolve onnxruntime-node from transformers' own module scope
(createRequire from transformers' package.json) so the guard always
checks the same binary that will be dlopen'd at runtime.
Also:
- Add npm overrides to force @huggingface/transformers to use our
onnxruntime-node@^1.24.0 (works for global installs where gitnexus
is the root package; npx installs get safety from the resolve fix)
- Replace hardcoded 'x64' with process.arch for arm64 support
- Remove dead napi-v3 path check (ORT 1.21.0 never shipped CUDA .so)
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
- Enhance C++ tree-sitter queries to support inline class method declarations and return types.
- Introduce `importedRawReturnTypes` in `BuildTypeEnvOptions` for cross-file raw return type handling.
- Add `FileTypeEnvBindings` interface to capture file-scope type bindings for exported symbols.
- Implement logic in `parse-worker.ts` to extract and serialize file-scope type bindings for cross-file type resolution.
- Create test fixtures for C++, Go, Ruby, and Rust to validate cross-file binding propagation.
- Update integration tests to verify correct resolution of method calls across files for C++, Go, Ruby, and Rust.
- Document Phase 14: Cross-File Binding Propagation in the type resolution roadmap and system documentation.
Fix server/bridge mode leaving the web UI with 0 nodes and broken
Query/Processes/embeddings by hydrating the worker-side LadybugDB
and BM25 indexes after loading graph data from the backend.
Also fix LadybugDB QueryResult API mismatch where result.getAll()
does not exist in some @ladybugdb/wasm-core versions — falls back
to getAllObjects() or getAllRows().
- Add .env.example with all HTTP embedding env vars documented
- Early return in httpEmbed() for empty text arrays
- Warn once if API returns vectors with different dimensions than
GITNEXUS_EMBEDDING_DIMS — helps catch misconfiguration early
- Extract shared HTTP client (http-client.ts) used by both core and MCP embedders
- Remove module-level httpConfig cache — read env vars fresh on every call
so config set after module load (e.g. via dotenv) takes effect
- Add NaN/non-positive guard on GITNEXUS_EMBEDDING_DIMS in schema.ts
- Include scrubbed URL and batch index in error messages (no API key)
- Wrap fetch rejections (DNS/timeout/connection) with same scrubbed context
- MCP embedder delegates to shared httpEmbedQuery() instead of inline logic
- apiKey confined to http-client.ts internals — not exported in any type or accessor
- Remove HttpEmbeddingConfig from types.ts (replaced by internal HttpConfig)
- All 16 HTTP embedder tests pass, tsc clean
* feat: add markdown file indexing (headings + cross-links)
Parse .md/.mdx files using regex (no tree-sitter dependency) to extract:
- Section nodes from headings (h1-h6) with hierarchy via CONTAINS edges
- Cross-file IMPORTS edges from markdown links to other repo files
Ported from #286 to resolve conflicts with kuzu→lbug rename.
Co-Authored-By: Dennis Palatov <dp-web4@users.noreply.github.com>
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
* fix: add Section to NODE_TABLES and NODE_SCHEMA_QUERIES
The Section schema was defined but not registered in NODE_TABLES or
NODE_SCHEMA_QUERIES, so the table was never created in the database.
Also adds missing FROM File TO Section relation entry.
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
* fix: update schema test counts for Section node type
NODE_TABLES: 27→28, NODE_SCHEMA_QUERIES: 27→28, SCHEMA_QUERIES: 29→30
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
* test: add diagnostic output to skills-e2e idempotency test
Show stdout/stderr in assertion message so CI failures reveal
why the second analyze --skills run exits with code 1.
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
* fix: add Section COPY query with level column in lbug-adapter
Section table has 8 columns (includes level) but getCopyQuery fell
through to the default 7-column multi-language path. Adds explicit
Section cases to getCopyQuery and insertNodeToLbug/upsertNodeToLbug.
Error was: COPY failed for Section: Number of columns mismatch. Expected 7 but got 8.
---------
Co-authored-by: Dennis Palatov <dp-web4@users.noreply.github.com>
Add MiniMax as a new LLM provider using the Anthropic-compatible API.
Changes:
- Add MiniMax to LLMProvider type and MiniMaxConfig interface
- Add MiniMax chat model creation via ChatAnthropic with custom base URL
- Add MiniMax settings persistence and model list (MiniMax-M2.5, MiniMax-M2.5-highspeed)
- Add MiniMax provider UI in SettingsPanel with API key and model selection
1. Replace fragile regex in buildExportedTypeMapFromGraph with
extractReturnTypeName() for consistent generic unwrapping.
2. Fix gap detection precision: check upstream.has(binding.exportedName)
instead of exportedTypeMap.has(binding.sourcePath) to avoid
over-counting files that import only unrelated symbols.
All 4 test blocks now share one DB lifecycle to avoid cross-block
"Database is closed" errors caused by LadybugDB's shared global DB
in a single vitest fork. Staleness detection (which triggers closeLbug)
runs last to avoid invalidating connections for other blocks.
11/11 tests pass on macOS, Ubuntu, and Windows.
Critical fixes:
- Re-resolution pass now actually re-resolves CALLS edges by calling
processCalls with importedBindingsMap (was building typeEnv but
discarding it without producing edges)
- Worker path populates ExportedTypeMap via buildExportedTypeMapFromGraph
using graph node isExported + SymbolTable returnType/declaredType
(was dead parameter in processCallsFromExtracted)
Important fixes:
- Skip threshold denominator uses totalFiles (was exportedTypeMap.size +
filesWithGaps which made threshold nearly useless)
- processCalls accepts importedBindingsMap parameter to thread cross-file
bindings into buildTypeEnv during re-resolution
All 3454 tests pass.
Add ExportedTypeMap infrastructure to propagate resolved type bindings
across file boundaries. When file A exports `const user = getUser()`
(resolved to `User`), file B importing `user` now gets seeded with
`user → User`, enabling `user.save()` to produce CALLS edges.
Key components:
- `importedBindings` option on BuildTypeEnvOptions with scopeEnv seeding
AFTER walk() to respect first-writer-wins (local declarations win)
- `collectExportedBindings()` in call-processor using graph node
isExported flag (no SymbolDefinition changes needed)
- Inline Kahn's algorithm topological sort with level grouping for
parallel-safe file ordering and cycle detection
- Re-resolution pass in pipeline.ts: topological order, 3% skip
threshold, path validation, per-file export caps (500)
- 32 new tests: 11 topological sort, 6 seeding, 15 integration
(simple cross-file, re-export chain, circular imports)
All 3454 tests pass (32 net new, 0 regressions).
Address code review findings from PR #392 senior compiler review:
- Fix Java "Yes" → "No" in optional-param-arity matrix (Java has no defaults)
- Simplify Kotlin hasDefaultValue while-as-if to direct const/if check
- Update OPTIONAL_PARAM_TYPES comment to include Ruby
- Replace per-declaration Set allocation with size-based Map iteration skip
- Add 11 unit tests for multi-declarator type association and constructorTypeMap
Add requiredParameterCount to SymbolDefinition and MethodSignature,
enabling range-based arity filtering in filterCallableCandidates.
Calls with omitted optional/default arguments now resolve correctly.
Supported: TS, Python, Kotlin, C#, C++, PHP, Ruby (7 languages).
Detection via OPTIONAL_PARAM_TYPES set + hasDefaultValue helper.
9 integration tests added across all 7 languages.
- Watchdog timer now exempts in-flight queries via activeQueryCount,
preventing premature stdout restoration during long queries (>1s)
- Stale detection uses reinitPromises Map to prevent TOCTOU race where
concurrent callers double-close the connection pool
- Throttle meta.json staleness checks to once per 5s per repo
- Add null guard for i.id in IN-clause construction
- Enrichment queries run in parallel on non-arm64 platforms to preserve
performance; sequential only on arm64 macOS where SIGSEGV occurs
- Updated AGENTS.md and CLAUDE.md to reflect new indexing metrics.
- Enhanced call-processor.ts to support cross-file inheritance tracking and improved virtual dispatch resolution.
- Added support for TypeScript overload signatures in tree-sitter queries.
- Improved type extraction for C++, C#, and Kotlin to handle smart pointers and constructor types.
- Introduced inferLiteralType for overload disambiguation across multiple languages.
- Added tests for C++ smart pointer dispatch and Kotlin virtual dispatch scenarios.
- Updated type-resolution-roadmap.md to reflect completion of phases P.1 to P.3 and outline future work on covariant return types.
- Add AbortSignal.timeout(30s) on all fetch calls
- Add retry with backoff for 429/5xx (core: 2 retries, MCP: 1 retry)
- Guard initEmbedder() and getEmbedder() to throw in HTTP mode
- Discard cached embeddings on dimension mismatch during incremental re-index
- Add MCP embedQuery retry for transient failures
- Add 16 unit tests covering both core and MCP HTTP paths
- Fix README: concise, accurate env var docs
Fixes three related issues that cause SIGSEGV crashes and stale data:
1. Impact enrichment queries (Promise.all → sequential await)
The impact() method ran 3 enrichment queries concurrently via
Promise.all against the same LadybugDB connection pool. On arm64
macOS, concurrent native DB access triggers SIGSEGV. Changed to
sequential await. Also caps IN-clause to 100 IDs to prevent
oversized queries. (#285, #290, #292)
2. Silence stdout during query execution
silenceStdout()/restoreStdout() only wrapped createConnection() and
initLbug(). Now also wraps executeQuery() and executeParameterized()
to prevent native stdout writes from corrupting the MCP stdio
stream during all DB operations. (#285)
3. Stale data after re-index
ensureInitialized() checked pool existence but never verified whether
the underlying index was rebuilt. Now reads meta.json's indexedAt
timestamp on each call and closes/re-opens the pool when the index
has changed. (#297)
- schema.ts: FLOAT[${EMBEDDING_DIMS}] reads from GITNEXUS_EMBEDDING_DIMS env
- embedding-pipeline.ts: vector search CAST uses actual query vector length
- mcp/core/embedder.ts: HTTP embedding support for MCP query-time search
Without this, using a 1024d model (e.g. bge-large) fails with
'Expected: 384, Actual: 1024' on LadybugDB vector insert.
Adds support for OpenAI-compatible embedding endpoints as an alternative
to the local transformers.js pipeline. Enables using self-hosted servers
(Infinity, vLLM, TEI, llama.cpp) over Tailscale/VPN, or any cloud
endpoint — with higher-quality models like bge-large-en-v1.5 (1024d).
Configuration via environment variables:
GITNEXUS_EMBEDDING_URL=http://your-server:8080/v1
GITNEXUS_EMBEDDING_MODEL=BAAI/bge-large-en-v1.5
GITNEXUS_EMBEDDING_API_KEY=your-key (default: 'unused')
GITNEXUS_EMBEDDING_DIMS=1024 (auto-detected if omitted)
When env vars are set:
- initEmbedder() skips local model download entirely
- embedText() and embedBatch() call the HTTP endpoint
- Dimensions auto-detected from first response
- Batches in groups of 64
When env vars are NOT set:
- Existing local transformers.js behavior is completely unchanged
Build: tsc clean
Tests: 1776 passed, 0 failed
