@op1/code-intel

Semantic code graph engine for OpenCode. Builds an AST-aware symbol index with vector embeddings and graph relationships, then exposes hybrid retrieval tools to agents.

Install

bun add @op1/code-intel

Add to your opencode.json:

{
	"plugin": ["@op1/code-intel"]
}

How it works

The plugin uses lazy initialization — no index is built until the first tool call. On first use it:

1. Parses the workspace with tree-sitter (TypeScript and Python by default)
2. Extracts symbols, edges (calls, imports, inheritance), and chunks
3. Generates 768-dim embeddings via UniXcoder (@huggingface/transformers)
4. Stores everything in a local SQLite database with sqlite-vec for vector search
5. Computes a PageRank-based repo map using graphology

Subsequent sessions reuse the index and refresh incrementally.

Tools

| Tool | Description | |------|-------------| | smart_query | Natural language code search — hybrid vector + BM25 retrieval with RRF fusion and graph expansion. Token-budget aware. | | symbol_impact | Change impact analysis. "What breaks if I modify X?" with risk levels and transitive dependents. | | call_graph | Caller/callee visualization with depth-limited traversal. | | symbol_search | BM25 keyword search for symbols by name or pattern. Filterable by type. | | repo_map | File importance rankings based on PageRank over the import/call graph. | | code_intel_status | Index statistics — file counts, symbol/edge totals, embedding model, schema version. | | code_intel_rebuild | Force a full reindex from scratch. | | code_intel_refresh | Incremental update — reindexes only changed files. |

Storage

Default paths (relative to workspace root):

| File | Purpose | |------|---------| | .opencode/code-intel/index.db | SQLite database with symbols, edges, chunks, and vector embeddings | | .opencode/code-intel/cache.json | Merkle hash cache for fast change detection |

Add .opencode/code-intel/ to .gitignore.

Configuration defaults

Defaults from source (types.ts):

| Setting | Default | |---------|---------| | Languages | typescript, python | | Embedding model | microsoft/unixcoder-base (768-dim) | | Max query tokens | 8000 | | Graph depth | 2 (max 3) | | Max fan-out | 10 | | Rerank mode | hybrid | | Index external deps | true | | Ignored patterns | node_modules, .git, dist, build, *.min.js, *.bundle.js |

Voyage AI reranking (optional)

The smart_query tool supports a rerankMode parameter:

• none — no reranking
• heuristic — BM25-based reranking
• llm — Voyage AI reranking (requires API key)
• hybrid — Voyage AI with BM25 fallback (default)

To enable Voyage AI reranking, set:

export VOYAGE_AI_API_KEY=your-key

Without the key, the plugin falls back to heuristic reranking automatically.

Branch awareness

Indices are scoped per git branch. Switching branches uses a separate index partition so results stay accurate without a rebuild.

License

MIT