Weavatrix Refactor

The explicit source-write layer of the Weavatrix ecosystem; Core remains read-only.

Refactor across files without letting the agent guess.

Weavatrix Refactor is a native MCP server for coding agents. It finds a symbol through the repository graph, previews byte-exact edits, reports every same-named occurrence it could not prove, and writes only after an explicit, hash-bound confirmation.

The useful difference is not “AI can rename a word.” It is that the agent gets evidence before the write and a recoverable transaction after it.

Install

Run the native binary through npm:

npx -y weavatrix-refactor mcp /absolute/path/to/repository

Or install it with Cargo:

cargo install weavatrix-refactor --locked
weavatrix-refactor mcp /absolute/path/to/repository

An MCP client can start it directly:

{
  "mcpServers": {
    "weavatrix-refactor": {
      "command": "weavatrix-refactor",
      "args": ["mcp", "/absolute/path/to/repository"],
      "env": {"WEAVATRIX_ALLOW_SOURCE_EDITS": "1"}
    }
  }
}

Leave WEAVATRIX_ALLOW_SOURCE_EDITS unset for preview-only sessions. Preview is a read; only a confirmed apply or rollback requires the gate.

Install as a plugin

The repository also ships one plugin bundle for Cursor, Codex, Claude Code, and Grok Build. The plugin starts the extension-only profile, so it contributes the 11 Refactor methods without repeating Weavatrix Core's 43 read-only method definitions. Install the separate Weavatrix plugin when both surfaces are needed.

The bundled skill is optional and activates only for explicitly requested refactoring work. Its entry point stays short; each method has an individual reference card that is loaded only when needed.

• Cursor: search for Weavatrix Refactor after marketplace approval.
• Codex: add the Weavatrix/weavatrix-refactor marketplace from .agents/plugins, then add weavatrix-refactor@weavatrix-refactor.
• Claude Code: add the same repository marketplace from .claude-plugin, then install weavatrix-refactor@weavatrix-refactor.
• Grok Build: run grok plugin marketplace add Weavatrix/weavatrix-refactor, open /marketplace, and install it.

The plugin opens WEAVATRIX_ALLOW_SOURCE_EDITS=1, but every write still needs the exact preview's short-lived, plan-bound single-use token.

For a rename task, expose only the two tools the agent needs:

weavatrix-refactor mcp /absolute/path/to/repository --profile=rename

The rename profile keeps rename_symbol and rollback_last_apply; every unrelated graph and refactor tool is absent and cannot be called. Omit the option (or use --profile=refactor) for all 11 Refactor methods. The narrower catalog reduces repeated agent context; it does not weaken preview hashes, the write gate, the confirmation token, or rollback checks.

When Weavatrix Core is already installed as a separate MCP, expose only the 11 methods that Refactor adds:

weavatrix-refactor mcp /absolute/path/to/repository --profile=refactor

The refactor profile is the default for MCP and list-tools, including the marketplace plugin. The former full compatibility profile has been removed and is rejected by both commands: Refactor no longer republishes Core methods. Install Weavatrix Core as its own MCP when read-only repository intelligence is also needed.

A safe cross-file rename in three calls

Suppose two files declare resolveTarget. A bare name is ambiguous, so the server refuses to pick one and returns exact ids:

// rename_symbol
{"symbol":"resolveTarget","new_name":"locateTarget","output_format":"json"}

{
  "status": "NOT_FOUND",
  "reason": "the name matches more than one symbol; pass one of the candidate ids",
  "candidates": [
    "symbol:src/core.ts#function:resolveTarget@2:1",
    "symbol:src/shadow.ts#function:resolveTarget@3:1"
  ]
}

Retry with the intended id. The result is still read-only: it contains the edit plan, content hashes, the proven edit count, and the references the backend refused to guess at.

// rename_symbol
{
  "symbol": "symbol:src/core.ts#function:resolveTarget@2:1",
  "new_name": "locateTarget",
  "output_format": "json"
}

{
  "status": "PLANNED",
  "completeness": "PARTIAL",
  "renamedEdits": 7,
  "uncertainReferences": ["... three named locations ..."],
  "confirmToken": "single-use-token",
  "plan": {"schemaVersion":"weavatrix.edit-plan.v1","files":["..."]}
}

After reviewing the plan, repeat the same operation with the same inputs and its token:

// rename_symbol
{
  "symbol": "symbol:src/core.ts#function:resolveTarget@2:1",
  "new_name": "locateTarget",
  "mode": "apply",
  "confirm_token": "single-use-token",
  "output_format": "json"
}

{"status":"APPLIED"}

This exact flow is exercised against the real stdio MCP server. The repository fixture requires seven edits across three files, preserves four traps (a longer identifier, string, comment, and unrelated shadow), and must still compile; it scores 12/12.

What is automatic today

The eleven tools do not all claim the same level of automation:

Refactor uses the Rust Weavatrix graph internally for evidence but never exports Core methods. Install the separate Weavatrix plugin for read-only repository intelligence. The Refactor MCP owns exactly its 11 extension methods; --profile=rename is an optional narrower two-method workflow.

Safety model and exact limits

• Ambiguous names fail with candidate ids instead of selecting the first match.
• PARTIAL means the graph proved the edits in the plan but did not prove that no other semantic references exist. uncertainReferences names the places requiring review.
• Every file in a plan carries a content hash. A changed working tree makes the preview stale instead of applying the old coordinates.
• Apply requires WEAVATRIX_ALLOW_SOURCE_EDITS=1 plus a short-lived, single-use token bound to that repository and exact recomputed plan.
• rollback_last_apply refuses to overwrite files that drifted after the transaction.
• Multi-file writes are journaled and crash-recoverable. They are not observationally atomic to unrelated processes: another process may briefly observe an intermediate filesystem state.

Only edits with proven provenance are applied. A lexical or graph backend is deliberately more conservative than a language server and must not be read as a claim of full compiler semantics.

Measured benchmark, with unlike layers kept separate

The checked-in August 2026 benchmark uses equivalent adversarial TypeScript, Rust, and Python fixtures. Each has seven required edits, four traps, and a build/test gate. Protocol runs count exact o200k_base initialization, catalog, request, and response tokens. Agent runs use the actual cumulative usage reported by Codex CLI with pinned gpt-5.6-sol / medium reasoning. No byte proxy or estimated tokens-per-second conversion is used.

The new rename-only profile kept all nine protocol runs at 12/12 while reducing fixed MCP context from 9,331 to 485 tokens. Its guided TypeScript agent A/B is:

The full agent matrix is 12/12 in all 27 TypeScript/Rust/Python runs; every run also completed operationally and passed the sanitized edit-channel audit. The deterministic protocol matrix also exposed a language-specific Serena defect: its suggested find_symbol → rename_symbol flow scored 7/12 on TypeScript in all three runs, while Rust and Python passed. It also found and drove fixes for Rust macro calls and Python f-string/module-level call evidence.

On this toy TypeScript repository naked Codex is still faster and cheaper than the narrowed profile. Weavatrix's demonstrated value is preview-before-write, named uncertainty, stale-plan refusal, explicit authorization, and rollback—not a universal small-repository token win.

See the full methodology, all language medians, raw-result links, and limitations.

Contract

The 11 Refactor tool names, schemas, and result states are frozen in contract/refactor-tools.v1.json. All refactor operations are native Rust; the host uses the read-only weavatrix-rust engine internally for graph evidence without exporting its methods on the default surface.

Versions 0.1.x used the JavaScript engine. That line continues as weavatrix-refactor-js and keeps a separate state directory, token store, and rollback journal.

License

MIT.