🤖 RepoNova 🔭

Turn your codebase into a knowledge graph. Query it with AI.

Knowledge graph builder & MCP server for AI code assistants.
Extracts symbols, relationships, and semantics from your code — then exposes the entire structure
as 11 graph tools that any MCP-compatible agent can use.

⚠️ Alpha — Active Development APIs, config format, and CLI may change between releases. Already usable in production workflows. Open an issue if something doesn't work.

Why RepoNova?

AI agents read files one at a time. They don't understand how your codebase fits together — which functions call what, which modules depend on which, where the architectural bottlenecks are.

RepoNova fixes that. It builds a persistent knowledge graph of your entire codebase (or multiple repos) and gives your AI agent 11 specialized tools to query it: search, impact analysis, shortest path, semantic similarity, community detection, and more.

One build. Persistent graph. Instant queries across sessions. No re-reading files. No burning tokens on context. The graph remembers everything.

What makes it different

• Zero external dependencies — no Python, no Docker, no database servers. Pure Node.js
• Multi-repo support — build one graph spanning multiple repositories
• Smart incremental builds — SHA256 file hashing, config change detection, semantic graph diffing, selective subsystem regeneration
• Local LLM-enhanced — optional local LLM for richer community summaries and node descriptions (runs on CPU)
• 11 MCP tools — from text search to weighted Dijkstra, semantic similarity to structural queries
• Works with any MCP client — OpenCode, Cursor, Claude Code, VS Code Copilot

How it works

  Your Codebase                      reponova build                    AI Agent
  ─────────────                      ──────────────                    ────────

  Python / TS / JS                   1. tree-sitter AST parsing        graph_search
  Markdown / Docs    ──────────►     2. Symbol + edge extraction   ──► graph_impact
  Diagrams / SVG                     3. Louvain communities            graph_path
  Multi-repo                         4. TF-IDF / ONNX embeddings       graph_similar
                                     5. Community summaries
                                     6. HTML visualizations            ... (11 tools)

Install

npm install -g reponova

Or run directly without installing:

npx reponova

Requires Node.js >= 18.

Quick Start

1. Install into your editor

reponova install --target opencode

This registers the MCP server, installs hooks/skills, and writes the default reponova.yml config.

Supported editors: opencode, cursor, claude, vscode

2. Build the knowledge graph

reponova build

3. Use it

The MCP server starts automatically with your editor. Your AI agent now has access to all 11 graph tools.

You: "What would be the impact of refactoring the authenticate function?"
Agent: [calls graph_impact] → shows upstream/downstream blast radius across repos

MCP Tools

11 specialized tools exposed over MCP (stdio). Each tool is designed for a specific query pattern.

Agentic Workflows

RepoNova is designed to be the structural memory layer for AI coding agents. Here's how to use it effectively in agentic workflows.

Recommended agent patterns

Before any refactoring:

1. graph_impact "TargetFunction" → understand blast radius
2. graph_path "ModuleA" "ModuleB" → see dependency chain
3. graph_community 5 → understand the module cluster
4. Make changes with full structural awareness

When exploring unfamiliar code:

1. graph_status → understand graph size and repos
2. graph_hotspots → identify architectural pillars
3. graph_search "authentication" → find entry points
4. graph_explain "Function:authenticate" → deep dive

When answering "where is X used?":

1. graph_search "X" → find the node
2. graph_impact "X" direction=downstream → who depends on it
3. graph_similar "X" → find semantically related code

Integration with editor skills

The reponova install command installs a skill file and a hook/rule that teaches your AI agent when and how to use each tool. The agent automatically reaches for graph tools when it needs structural information.

Keeping the graph fresh

# Incremental rebuild — only processes changed files
reponova build

# Full clean rebuild
reponova build --force

Tip: Add reponova build to your CI pipeline or as a post-commit hook to keep the graph always up-to-date.

How incremental builds work

RepoNova's incremental build goes beyond simple file-change detection. It minimizes redundant work at every stage of the pipeline:

The build config fingerprint is stored in graph.json metadata, so it persists across builds without extra files.

CLI Reference

reponova install

Set up editor integration. Creates MCP config, hook, skill, and reponova.yml.

reponova install --target <editor> [--graph <path>]

reponova build

Build (or rebuild) the knowledge graph.

reponova build [--config <path>] [--force]

Build pipeline:

1. Detect source files, documentation, and diagrams (with centralized glob matching via picomatch)
2. Diff files against previous build — detect changed, added, and removed files
3. Compare build config fingerprint — detect config-only changes (e.g. switching embeddings.method)
4. Parse changed files with tree-sitter WASM — extract symbols, calls, imports, inheritance (cached extractions reused for unchanged files)
5. Build directed graph with cross-file / cross-repo edges
6. Compute semantic graph hash — skip downstream steps when graph structure is unchanged
7. Clean up stale artifacts when config changes invalidate them (e.g. old TF-IDF files after switching to ONNX)
8. Detect communities (Louvain algorithm)
9. Generate embeddings incrementally — only re-embed nodes whose text content changed (TF-IDF or ONNX MiniLM)
10. Selectively regenerate subsystems affected by config changes (embeddings, summaries, descriptions, outlines, HTML)
11. Generate community summaries + node descriptions (algorithmic or LLM-enhanced)
12. Generate graph.html and graph_communities.html interactive visualizations
13. Generate SQLite search index (graph_search.db)
14. Generate code outlines (SHA256 per-file hashing — skip unchanged) and report.md

reponova index

Regenerate the SQLite search index (graph_search.db) from an existing graph.json. Useful when the index is corrupted or deleted without needing a full rebuild.

reponova index [--graph <path>]

reponova outline

Pre-compute tree-sitter code outlines for files matching outlines.patterns. Runs the same outline logic as reponova build, but standalone — useful to regenerate outlines without re-running extraction or graph construction.

reponova outline [--config <path>] [--graph <path>] [--force]

reponova mcp

Start the MCP server over stdio. Normally launched automatically by the editor.

reponova mcp [--graph <path>]

reponova models

Manage local AI models (ONNX embeddings, LLM). See Models for details.

reponova models status              # Show configured and cached models
reponova models download            # Pre-download all models needed by config
reponova models remove <name>       # Remove a specific cached model
reponova models clear               # Remove all cached models

reponova check

Verify graph installation, build integrity, and report stats.

reponova check [--graph <path>]

Checks performed:

• Graph file exists and is loadable
• Node/edge counts and repo list
• Build metadata presence (build_config fingerprint, reponova version)
• Embedding artifacts consistency (TF-IDF IDF file, ONNX vectors)
• Warns if embedding method in config doesn't match the built artifacts

Supported Languages

Extraction (AST parsing + graph building)

Outline (tree-sitter code outline)

Adding a new language: Create src/extract/languages/<lang>.ts implementing LanguageExtractor, register it in registry.ts, add the .wasm grammar to grammars/. See Contributing > Adding Language Support for the full interface reference.

Note: Extraction and outline are separate systems with different registries and interfaces. Registering an extractor gives you graph building (symbols, edges, imports). For code outlines (graph_outline), you also need a LanguageSupport implementation in src/outline/languages/ — see Contributing > Adding Outline Support.

Edge Types

Every edge in the graph has a type that describes the relationship:

Configuration

Config Resolution

The config file is auto-detected from these locations (first match wins):

1. Explicit --config argument
2. reponova.yml in the project root
3. .opencode/reponova.yml
4. .cursor/reponova.yml
5. .claude/reponova.yml
6. .vscode/reponova.yml

All paths inside the config are relative to the config file's location. When placed inside an editor directory (e.g. .opencode/), use ../ to reference the project root.

Pattern Resolution

All glob patterns (build.patterns, build.exclude, docs.patterns, outlines.patterns, etc.) are matched against workspace-relative paths. How those paths look depends on the number of repos.

Single-repo

With one repo, file paths are relative to the repo root — no prefix:

src/core.py
src/utils/helpers.py
tests/test_core.py

Patterns work as you'd expect:

repos:
  - name: my-project
    path: .
build:
  patterns: ["src/**/*.py"]          # matches src/core.py ✓
  exclude: ["tests/**"]              # excludes tests/test_core.py ✓

Multi-repo

With multiple repos, each file path is prefixed with the repo name from the config:

api/src/routes.py          # ← "api" comes from repos[].name
api/src/handlers.py
core/src/models.py         # ← "core" comes from repos[].name
core/src/db.py

Patterns are tested against both forms — the full prefixed path and the repo-relative path — so the same pattern works in single and multi-repo:

repos:
  - name: api
    path: ../services/api
  - name: core
    path: ../services/core
build:
  patterns: ["src/**/*.py"]          # matches api/src/routes.py, api/src/handlers.py, core/src/models.py, core/src/db.py ✓ (via repo-relative)
  exclude: ["**/test_*.py"]          # works across all repos

Filtering a specific repo

Use the repo name as a path prefix to target one repo only:

build:
  exclude:
    - "api/src/generated/**"         # excludes only in the api repo
    - "**/migrations/**"             # excludes in all repos

outlines:
  patterns:
    - "core/src/**/*.py"             # outlines only for the core repo

This works because the full workspace path is always <repo-name>/<path>. The repo name is the name field from your repos config — not the directory name on disk.

Full Config Reference

Every field, every valid value, every default.

# ──────────────────────────────────────────────────────────────────────────────
# reponova.yml — Full Configuration Reference
# ──────────────────────────────────────────────────────────────────────────────

# Where to write build output (graph.json, graph.html, graph_search.db, etc.)
# Type: string
# Default: "reponova-out"
output: ../reponova-out

# ── Repositories ──────────────────────────────────────────────────────────────
# List of repositories to include in the build.
# Each repo needs a unique name and a path (relative to this config file).
repos:
  - name: api-service           # string — unique identifier for this repo
    path: ../services/api       # string — path to repo root (relative to this file)
  - name: core-lib
    path: ../services/core

# ── Centralized Model Management ─────────────────────────────────────────────
# Shared settings for all models (LLM, ONNX embeddings).
# Individual features (community_summaries, node_descriptions) can specify
# their own model via a `model` field. These settings apply to all of them.
models:
  # Directory to cache downloaded models (ONNX embeddings + LLM weights)
  # Type: string
  # Default: "~/.cache/reponova/models"
  cache_dir: ~/.cache/reponova/models

  # GPU acceleration backend for LLM inference
  # Values: "auto" | "cpu" | "cuda" | "metal" | "vulkan"
  #   - auto:    auto-detect best available backend
  #   - cpu:     force CPU inference (slower but always works)
  #   - cuda:    NVIDIA GPU (requires CUDA drivers)
  #   - metal:   Apple Silicon GPU (macOS only)
  #   - vulkan:  Cross-platform GPU (AMD, Intel, NVIDIA)
  # Default: "auto"
  gpu: auto

  # Number of CPU threads for LLM inference
  # Type: number
  # Default: 0 (auto-detect based on available cores)
  threads: 0

  # Automatically download models on first use
  # Type: boolean
  # Default: true
  download_on_first_use: true

# ── Build Options ─────────────────────────────────────────────────────────────
build:

  # Glob patterns for source code files to include
  # Type: string[]
  # Default: [] (empty = auto-detect by file extension using registered extractors)
  # Example: ["src/**/*.py", "lib/**/*.ts"]
  patterns: []

  # Glob patterns to exclude from source code detection
  # Type: string[]
  # Default: []
  # Example: ["**/generated/**", "**/*.test.ts", "**/vendor/**"]
  exclude: []

  # Exclude common non-source directories from all file detection
  # (source code, documentation, diagrams, and outlines).
  # When true, the following directories are skipped at any depth:
  #   node_modules, __pycache__, .git, .svn, .hg, venv, .venv, env, .env, .tox,
  #   site-packages, dist, build, .eggs, .mypy_cache, .pytest_cache, .ruff_cache,
  #   target, bin, obj
  # Set to false if you need to index files inside these directories
  # (e.g. vendored code in node_modules). You can still exclude specific
  # directories via the `exclude` patterns above.
  # Type: boolean
  # Default: true
  exclude_common: true

  # Incremental builds: only re-process files whose SHA256 hash changed
  # Type: boolean
  # Default: true
  incremental: true

  # Generate interactive HTML visualizations (graph.html + graph_communities.html)
  # Type: boolean
  # Default: true
  html: true

  # Minimum node degree to include in HTML visualization
  # Useful for large graphs — filters out leaf nodes to reduce clutter
  # Type: integer (>= 1)
  # Default: not set (include all nodes)
  # html_min_degree: 3

  # ── Documentation Extraction ──────────────────────────────────────────────
  docs:
    # Enable/disable documentation extraction
    # Type: boolean
    # Default: true
    enabled: true

    # Glob patterns for documentation files (relative to repo root)
    # Type: string[]
    # Default: ["**/*.md", "**/*.txt", "**/*.rst"]
    patterns:
      - "**/*.md"
      - "**/*.txt"
      - "**/*.rst"

    # Glob patterns to exclude from documentation extraction
    # Type: string[]
    # Default: ["**/CHANGELOG.md", "**/node_modules/**"]
    # Note: if your output dir is inside the workspace (e.g. output: ./reponova-out),
    # add it here to prevent generated files from being re-ingested on rebuild.
    exclude:
      - "**/CHANGELOG.md"
      - "**/node_modules/**"
      - "reponova-out/**"

    # Maximum file size in KB — files larger than this are skipped
    # Type: number
    # Default: 500
    max_file_size_kb: 500

  # ── Diagram / Image Extraction ────────────────────────────────────────────
  images:
    # Enable/disable diagram extraction
    # Type: boolean
    # Default: true
    enabled: true

    # Glob patterns for diagram files (relative to repo root)
    # Type: string[]
    # Default: ["**/*.puml", "**/*.plantuml", "**/*.svg"]
    patterns:
      - "**/*.puml"
      - "**/*.plantuml"
      - "**/*.svg"

    # Glob patterns to exclude
    # Type: string[]
    # Default: ["**/node_modules/**"]
    exclude:
      - "**/node_modules/**"

    # Parse PlantUML files to extract components and relationships
    # Type: boolean
    # Default: true
    parse_puml: true

    # Extract text content from SVG files
    # Type: boolean
    # Default: true
    parse_svg_text: true

  # ── Embeddings ────────────────────────────────────────────────────────────
  # Vector representations for semantic search (graph_similar, graph_context)
  embeddings:
    # Enable/disable embedding generation
    # Type: boolean
    # Default: true
    enabled: true

    # Embedding method
    # Values: "tfidf" | "onnx"
    #   - tfidf:  Feature-hashed TF-IDF (384-dim). Fast (milliseconds). No model download.
    #   - onnx:   MiniLM-L6-v2 via ONNX Runtime (384-dim). More accurate. ~86MB model download.
    # Default: "tfidf"
    method: tfidf

    # ONNX model name (only used when method: "onnx")
    # Must be a sentence-transformers/ model on HuggingFace with ONNX export
    # and BERT-compatible tokenizer. Dimensions must match 'dimensions' below.
    # See the "Models" section for compatible models and details.
    # Type: string
    # Default: "all-MiniLM-L6-v2"
    model: all-MiniLM-L6-v2

    # Embedding vector dimensions
    # Type: number
    # Default: 384
    dimensions: 384

    # Batch size for ONNX inference
    # Type: number
    # Default: 128
    batch_size: 128

  # ── Community Summaries ───────────────────────────────────────────────────
  # Natural-language summaries for each detected community (cluster of related symbols).
  # Independent from node descriptions — can enable one without the other.
  community_summaries:
    # Enable/disable community summary generation
    # Type: boolean
    # Default: true
    enabled: true

    # Maximum number of communities to summarize
    # Type: integer (>= 0)
    # Default: 0 (no limit — summarize all communities)
    # Communities are sorted by size (largest first). When max_number > 0,
    # only the top N largest communities are summarized.
    # Communities with fewer than 3 nodes are always excluded.
    max_number: 0

    # LLM model for richer summaries (optional)
    # Uses hf: URI notation — see the "Models" section for details.
    # Type: string | null
    # Default: null (algorithmic summaries — still useful, just less prose)
    # model: "hf:Qwen/Qwen2.5-0.5B-Instruct-GGUF:Q4_K_M"

    # Context window size for LLM inference (only used when model is set)
    # Type: number
    # Default: 512
    context_size: 512

  # ── Node Descriptions ─────────────────────────────────────────────────────
  # Natural-language descriptions for high-degree (important) nodes.
  # Independent from community summaries — can enable one without the other.
  node_descriptions:
    # Enable/disable node description generation
    # Type: boolean
    # Default: true
    enabled: true

    # Degree threshold for node description generation
    # Type: number (0.0 – 1.0)
    # Default: 0.8
    # Meaning: top (1 - threshold)% of nodes by degree get descriptions.
    #   - 0.8 = top 20% of nodes
    #   - 0.5 = top 50% of nodes
    #   - 0.0 = all nodes (expensive!)
    #   - 1.0 = no nodes
    threshold: 0.8

    # LLM model for richer descriptions (optional)
    # Uses hf: URI notation — see the "Models" section for details.
    # Type: string | null
    # Default: null (algorithmic descriptions)
    # model: "hf:Qwen/Qwen2.5-0.5B-Instruct-GGUF:Q4_K_M"

    # Context window size for LLM inference (only used when model is set)
    # Type: number
    # Default: 512
    context_size: 512

# ── Outlines ──────────────────────────────────────────────────────────────────
# Tree-sitter code outlines: functions, classes, imports with signatures.
# Language is auto-detected from file extension (no need to specify it).
outlines:
  # Enable/disable outline generation
  # Type: boolean
  # Default: true
  enabled: true

  # Glob patterns for files to outline (relative to repo root)
  # Type: string[]
  # Default: ["src/**/*.ts", "src/**/*.py", "src/**/*.js"]
  patterns:
    - "src/**/*.py"
    - "src/**/*.ts"
    - "src/**/*.js"

  # Glob patterns to exclude from outline generation
  # Type: string[]
  # Default: ["**/node_modules/**", "**/.git/**", "**/dist/**"]
  exclude:
    - "**/__pycache__/**"
    - "**/node_modules/**"
    - "**/.git/**"
    - "**/dist/**"

# ── Server ────────────────────────────────────────────────────────────────────
# MCP server options (reserved for future use)
# Type: object
# Default: {}
server: {}

Minimal Config

Most fields have sensible defaults. A minimal config for a single repo:

output: ../reponova-out
repos:
  - name: my-project
    path: ..

Multi-repo Config

output: ../reponova-out
repos:
  - name: api
    path: ../services/api
  - name: core
    path: ../services/core
  - name: shared
    path: ../libs/shared

LLM-enhanced Config

For richer, natural-language community summaries and node descriptions:

output: ../reponova-out
repos:
  - name: my-project
    path: ..
models:
  gpu: auto                 # auto-detect GPU, falls back to CPU
  download_on_first_use: true
build:
  community_summaries:
    enabled: true
    model: "hf:Qwen/Qwen2.5-0.5B-Instruct-GGUF:Q4_K_M"   # ~350MB download
  node_descriptions:
    enabled: true
    threshold: 0.5          # describe top 50% nodes by degree
    model: "hf:Qwen/Qwen2.5-0.5B-Instruct-GGUF:Q4_K_M"   # same model, auto-shared

When community_summaries.model and node_descriptions.model resolve to the same file, RepoNova shares a single engine instance — no double memory usage.

File Filtering Config

Control which source files are included in the graph:

output: ../reponova-out
repos:
  - name: my-project
    path: ..
build:
  patterns:                    # only include files matching these globs
    - "src/**/*.py"
    - "lib/**/*.ts"
  exclude:                     # exclude files matching these globs
    - "**/test/**"
    - "**/tests/**"
    - "**/migrations/**"
    - "**/*.generated.ts"

When patterns is empty (default), RepoNova auto-detects source files by extension using all registered extractors. When exclude_common: true (default), the following directories are skipped at any depth: node_modules, __pycache__, .git, .svn, .hg, venv, .venv, env, .env, .tox, site-packages, dist, build, .eggs, .mypy_cache, .pytest_cache, .ruff_cache, target, bin, obj. Set exclude_common: false to disable this behavior and use explicit exclude patterns instead.

Models

RepoNova uses two types of AI models, both downloaded automatically on first use and cached locally. No API keys, no cloud services.

ONNX Embeddings

Sentence-transformer models for semantic similarity search (graph_similar, graph_context).

Compatible models (all 384-dim, must match embeddings.dimensions):

Any model under the sentence-transformers/ org on HuggingFace that provides an ONNX export with BERT-compatible tokenizer (WordPiece) should work. The dimensions config field must match the model's output dimension.

LLM (GGUF)

Local language models for richer community summaries and node descriptions, powered by node-llama-cpp.

When both community_summaries.model and node_descriptions.model resolve to the same file, RepoNova shares a single engine instance — no double memory usage.

Why different notations? ONNX embeddings use direct HTTP fetch from a fixed HuggingFace org (sentence-transformers/), downloading specific files (model.onnx, vocab.txt). LLM models delegate entirely to node-llama-cpp's resolveModelFile(), which handles the hf: URI protocol, download, and caching. The two systems are technically incompatible — the notation reflects this.

Model Management CLI

reponova models status              # Show configured and cached models
reponova models download            # Pre-download all models needed by config
reponova models remove <name>       # Remove a specific cached model
reponova models clear               # Remove all cached models

Models are also downloaded automatically during reponova build when models.download_on_first_use: true (default). The CLI commands let you manage the cache independently of the build.

Build Output

After reponova build, the output directory contains:

reponova-out/
├── graph.json                          # Full graph: nodes, edges, community assignments, metadata
│                                       #   metadata.build_config: config fingerprint for change detection
│                                       #   nodes include: docstring, signature, bases (when available)
├── graph.html                          # Interactive visualization (vis.js) — click, search, filter
├── graph_communities.html              # Community-focused visualization with summary labels
├── graph_search.db                     # SQLite search index (sql.js WASM) — structural queries
├── report.md                           # Build report: stats, hotspots, community breakdown
├── community_summaries.json            # Community summaries (algorithmic or LLM-enhanced)
├── node_descriptions.json              # Descriptions for high-degree nodes
├── tfidf_idf.json                      # TF-IDF vocabulary weights (for query-time embedding)
├── vectors/                            # LanceDB vector store — semantic similarity search
│   └── (LanceDB internal files)        #   fallback: vectors.json when @lancedb/lancedb unavailable
├── outlines/                           # Pre-computed code outlines per file
│   └── <repo>/<path>.outline.json
└── .cache/                             # Incremental build cache (SHA256 content hashing)
    ├── hashes.json                     #   file path → SHA256 hex map
    ├── semantic-graph-hash.txt         #   SHA256 of graph structure (nodes + edges)
    ├── outline-hashes.json             #   file path → SHA256 map for outline generation
    ├── node-texts.json                 #   node id → text hash map for incremental embeddings
    └── extractions/                    #   cached FileExtraction per file
        └── <hash>.json

Two storage engines serve different purposes:

• SQLite (graph_search.db) — structural index for exact lookups, graph traversal, FTS. Used by graph_search, graph_impact, graph_path, graph_explain, and more.
• LanceDB (vectors/) — vector index for semantic similarity. Used by graph_similar and graph_context. Falls back to brute-force cosine similarity (JSON) when @lancedb/lancedb is not installed.

Programmatic API

Use RepoNova as a library in your own Node.js tools.

Build API

Run the full build pipeline programmatically — useful for CI integrations, custom tooling, or workflows that register custom extractors/languages before building.

import { build } from "reponova";

const result = await build("./reponova.yml");
console.log(`Built: ${result.nodeCount} nodes, ${result.edgeCount} edges`);
console.log(`Output: ${result.outputDir}`);

// Force rebuild (deletes output and rebuilds from scratch)
const result = await build("./reponova.yml", { force: true });

build() returns a BuildResult:

If configPath is omitted, config is auto-detected from standard locations (see Config Resolution).

Runtime Registration + Build

Register custom extractors or outline languages before calling build():

import {
  build,
  registerExtractor,
  registerOutlineLanguage,
} from "reponova";
import type { LanguageExtractor, LanguageSupport } from "reponova";

// 1. Register a custom extractor (graph building)
const myExtractor: LanguageExtractor = { /* ... */ };
registerExtractor(myExtractor);

// 2. Register outline support (graph_outline)
const myOutline: LanguageSupport = { /* ... */ };
registerOutlineLanguage("rust", ["rs"], myOutline);

// 3. Build — all registrations are picked up automatically
const result = await build("./reponova.yml");

Query API

After building, load and query the graph:

import {
  openDatabase,
  initializeSchema,
  populateDatabase,
  loadGraphData,
  searchNodes,
  analyzeImpact,
  findShortestPath,
  getNodeDetail,
} from "reponova";

// Load and index the graph
const graphData = loadGraphData("./reponova-out/graph.json");
const db = await openDatabase(":memory:");
initializeSchema(db);
populateDatabase(db, graphData);

// Search
const results = searchNodes(db, "authentication", { top_k: 5, type: "function" });

// Impact analysis
const impact = analyzeImpact(db, "Function:authenticate_user", { max_depth: 3 });

// Shortest path
const path = findShortestPath(db, graphData, "ModuleA", "ModuleB");

// Node detail
const detail = getNodeDetail(db, graphData, "Function:process_payment");

Advanced API

import {
  ContextBuilder,
  loadConfig,
} from "reponova";

// Smart context assembly (search + vectors + graph expansion)
const { config } = loadConfig("./reponova.yml");
const builder = new ContextBuilder(db, graphData, "./reponova-out");
await builder.initialize(config.build.embeddings);
const context = await builder.buildContext({
  query: "authentication flow",
  maxTokens: 4000,
});

FAQ

Do I need an API key?

No. Everything runs locally. The optional LLM is a local model (Qwen 0.5B) — no cloud, no API keys, no data leaves your machine.

How big are the models?

How long does a build take?

Depends on codebase size. Rough benchmarks:

• Small project (500 files): ~5-10 seconds
• Medium project (5,000 files): ~30-60 seconds
• Large monorepo (20,000+ files): 2-5 minutes
• LLM summaries add ~2-3 seconds per community

Can I use it without an editor?

Yes. Use the CLI (reponova build, reponova check) and the programmatic API. The MCP server is just one way to query the graph.

What about TypeScript / JavaScript extraction?

Tree-sitter grammars are ready. The extractor implementation is on the roadmap — contributions welcome.

Contributing

Contributions are welcome.

Adding Language Support (Extraction)

Add new programming language extractors via tree-sitter. An extractor teaches RepoNova how to parse a language's AST and extract symbols, imports, and references for graph building.

Steps

1. Create src/extract/languages/<lang>.ts implementing the LanguageExtractor interface
2. Register it in src/extract/languages/registry.ts (or at runtime via registerExtractor())
3. Add the tree-sitter WASM grammar to grammars/ (e.g., tree-sitter-javascript.wasm)

LanguageExtractor Interface

interface LanguageExtractor {
  /** Language identifier — must match tree-sitter grammar name (e.g., "javascript") */
  readonly languageId: string;

  /** File extensions this extractor handles (e.g., [".js", ".mjs", ".cjs"]) */
  readonly extensions: string[];

  /**
   * WASM grammar filename (e.g., "tree-sitter-javascript.wasm").
   * If provided: pipeline parses with tree-sitter and passes the SyntaxTree.
   * If omitted: extract() receives a null tree and must work from sourceCode directly.
   * (Markdown and diagram extractors use this — no WASM needed.)
   */
  readonly wasmFile?: string;

  /**
   * Extract symbols, imports, and references from a single source file.
   * @param tree - Parsed tree-sitter AST (null if wasmFile not set)
   * @param sourceCode - Raw file content
   * @param filePath - Relative path (normalized, forward slashes)
   */
  extract(tree: SyntaxTree | null, sourceCode: string, filePath: string): FileExtraction;

  /**
   * Resolve an import module path to candidate file paths.
   * Example: "config.loader" → ["config/loader.py", "config/loader/__init__.py"]
   * Return empty array for external/third-party modules.
   */
  resolveImportPath(importModule: string, currentFilePath: string): string[];
}

FileExtraction Return Type

interface FileExtraction {
  filePath: string;           // Relative path (forward slashes)
  language: string;           // Must match languageId
  symbols: SymbolNode[];      // Functions, classes, methods, variables
  imports: ImportDeclaration[];  // Import/export statements
  references: SymbolReference[];  // Calls, type annotations, inheritance refs
}

Key types your extractor produces:

See src/extract/types.ts for full type definitions and JSDoc.

How tree-sitter Parsing Works

1. If wasmFile is set, the pipeline loads grammars/<wasmFile>, parses the source, and passes a SyntaxTree to extract()
2. If wasmFile is omitted, extract() receives null as the tree and must work from sourceCode directly
3. WASM grammars are loaded from the grammars/ directory relative to the package root
4. SyntaxTree / SyntaxNode types match the web-tree-sitter WASM interface

Runtime Registration

You can also register extractors at runtime via the public API (must be called before build):

import { registerExtractor } from "reponova";
import type { LanguageExtractor } from "reponova";

const myExtractor: LanguageExtractor = { /* ... */ };
registerExtractor(myExtractor);

Note: duplicate languageId or extensions silently overwrite the previous extractor.

Reference Implementation

See src/extract/languages/python.ts for a full tree-sitter-based extractor, or src/extract/languages/markdown.ts for a non-tree-sitter (regex) extractor.

Adding Outline Support

Outlines (graph_outline) use a separate system from extraction. They have their own registry, interface, and implementations.

Steps

1. Create src/outline/languages/<lang>.ts implementing the LanguageSupport interface
2. Register it in src/outline/languages/registry.ts via registerOutlineLanguage()
3. The same WASM grammar from grammars/ is shared with the extraction system

LanguageSupport Interface

interface LanguageSupport {
  /** WASM grammar filename (e.g., "tree-sitter-python.wasm") */
  readonly wasmFile: string;

  /** Extract outline from tree-sitter AST (primary method) */
  treeSitterExtract(rootNode: SyntaxNode, filePath: string, lineCount: number): FileOutline;

  /** Extract outline from raw source via regex (fallback when WASM unavailable) */
  regexExtract(filePath: string, source: string, lineCount: number): FileOutline;
}

Runtime Registration

You can also register outline languages at runtime via the public API (must be called before build):

import { registerOutlineLanguage } from "reponova";
import type { LanguageSupport } from "reponova";

const myOutline: LanguageSupport = { /* ... */ };
registerOutlineLanguage("rust", ["rs"], myOutline);

Note: duplicate language names or extensions silently overwrite the previous registration.

See src/outline/languages/python.ts for the reference implementation.

License

MIT — CristianoCiuti/reponova