RepoWeaver

Code context fabric for AI coding agents — deterministic call-graph indexing, hybrid retrieval, and a single-tool MCP interface.

RepoWeaver builds a local, zero-cost code intelligence layer that lets AI coding agents (Codex, Claude Code, and other MCP clients) answer structural questions without reading entire files.

"Who calls this method, and what would break if I change its signature?"
→ explore(query, task="impact", repo=".", max_tokens=4000)
→ [verbatim source slices + call paths + blast radius + known blind spots]

Why

AI agents that rely on grep or full-file reads waste tokens and miss cross-file call chains. Existing tools either require cloud APIs, embed LLMs in the index layer, or carry restrictive licenses. RepoWeaver takes only the verified-consensus patterns from the landscape and assembles them locally.

Design

Six primitives, all independently validated across multiple open-source tools:

#	Primitive	Validated by
1	tree-sitter deterministic parsing, local $0	CodeGraph · GitNexus · Graft · Aider
2	Call / inherit / import edges + blast-radius	CodeGraph · GitNexus · Graft · CodeQL
3	BM25/FTS + PageRank graph diffusion	Graft · CodeGraph · Aider
4	Content-hash incremental freshness	Graft · Codebase-Memory · CodeGraph
5	MCP single strong tool explore()	CodeGraph (empirically validated)
6	AGENTS.md protocol injection + edit hook	GitNexus · CodeGraph · Graft · Potpie

Not included: LLM in the index layer (rejected by every serious tool), code sent to external APIs, opinionated semantic summaries.

Roadmap

Milestone	Tag	Status
M1 — Fabric MVP (Java parser + edges + FTS5 + PageRank + MCP)	v0.1.0	✅ shipped
M2 — Freshness & confidence (auto-sync + disambiguation + confidence edges)	v0.2.0	✅ shipped
M3 — Type precision overlay (SCIP/jdtls)	v0.3.0	planned
M4 — Runtime overlay (OTel/Jaeger trace → edge weights)	v0.4.0	planned

Each milestone ships a fabric verify --level mN gate that runs in CI.

Verification approach

Benchmarks use public repos only — no proprietary code enters this repository.

Benchmark repo	Purpose
spring-projects/spring-petclinic	Spring DI-heavy, tests injection-aware edges
google/gson	Comparable to CodeGraph's published 93.3% Java coverage
square/okhttp	Call-chain depth
mybatis/mybatis-3	Generated-code noise filtering

SOTA alignment — measured, not claimed

RepoWeaver ships a reproducible benchmark harness and refuses to count low-confidence or ambiguous edges as resolved coverage.

Pinned commit dae37cf…; apples-to-apples scope gson/src/main/; both source and target files must be inside the scope.

Metric	v0.1 whole-repo baseline	RepoWeaver v0.2 core	CodeGraph 1.5 core	Gate
Resolved cross-file dependent coverage	35.3%	90.12%	92.59%	>=90%
Ambiguous edge rate	89.3%	6.36%	not exposed	<=10%
Fixture edge precision / recall	1.0 / 1.0	1.0 / 1.0	not measured	>=.95 / >=.90

RepoWeaver passes its release gates and is in the same measured coverage band, while remaining 2.47 percentage points behind CodeGraph on this benchmark. We therefore claim alignment, not universal superiority. Coverage cannot be improved by adding ambiguous edges: the gate pairs it with ambiguity and fixture precision.

fabric verify --level benchmark
fabric benchmark run --repo /path/to/gson --name gson \
  --scope-prefix gson/src/main/ --output gson.json
fabric benchmark compare --candidate gson.json --target benchmarks/sota-targets.yaml

See benchmark methodology, the checked-in RepoWeaver v0.2 baseline, and the CodeGraph 1.5 comparison.

Install

git clone <this-repo> && cd repoweaver
uv sync --extra dev

uv run fabric build /path/to/your/java/repo     # index a repo, $0, no LLM
uv run fabric check /path/to/your/java/repo      # OK | STALE (content-hash freshness)
uv run fabric init /path/to/your/java/repo       # inject AGENTS.md protocol block
uv run fabric watch /path/to/your/java/repo      # OS-event auto-sync, 2s debounce
uv run fabric verify --level m2                  # watcher + incremental consistency gate
uv run fabric serve                              # start the MCP server (explore() tool)

v0.2.0 supports Java only, via tree-sitter. M2 adds conservative overload/type resolution, REFERENCES, annotation symbols, unresolved-candidate storage, framework entry-point metadata and auto-sync. No LLM or external network call is made at runtime.

MCP tool

explore(
  query: str,
  task: "understand" | "impact" | "locate" | "debug",
  repo: str = ".",
  max_tokens: int = 4000
) → {
  query, task, repo,
  slices: [{node_id, file, span_start, span_end, source, qualified_name, confidence, provenance}],
  stats: {nodes_visited, edges_traversed, tokens_estimated, freshness},
  blind_spots: "<frozen incompleteness contract string>",
  # task == "impact": + blast_radius: [{depth, node_id, qualified_name, file, edge_type, confidence, risk}]
  # task == "debug":  + call_path: [{step, node_id, qualified_name, file, edge_type, confidence}]
}

See docs/explore-contract.md (frozen v1.1) for the full contract.

Acknowledgements

Design patterns drawn from (not forked from): CodeGraph · Graft · GitNexus · Serena · SCIP · Aider

License

MIT