# Weavatrix Refactor
**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:
```bash
npx -y weavatrix-refactor mcp /absolute/path/to/repository
```
Or install it with Cargo:
```bash
cargo install weavatrix-refactor --locked
weavatrix-refactor mcp /absolute/path/to/repository
```
An MCP client can start it directly:
```json
{
"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 `sergii-ziborov/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
sergii-ziborov/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:
```bash
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:
```bash
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:
```json
// 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.
```json
// 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:
```json
// 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:
| Complete preview/apply workflow | `rename_symbol`, `rename_related_symbols` | A generated plan and token, then a confirmed write by repeating the same operation |
| Plan-producing | `change_signature`, `edit_symbol`, `bulk_replace`, `organize_imports`, `move_file` | Byte-exact edits or a review plan; use `apply_edit_plan` where an edit-plan envelope is returned |
| Advisory | `move_symbol`, `delete_readiness` | Blast radius, cycle/architecture risks, or a deletion verdict; no automatic source write |
| Safety infrastructure | `apply_edit_plan`, `rollback_last_apply` | Verification/application of an edit-plan envelope and drift-safe restoration |
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:
| Naked Codex | **12/12 x3** | **42,602 ms** | **72,169** |
| Weavatrix 1.0.6 `--profile=rename` | **12/12 x3** | 48,389 ms | 108,155 |
| Weavatrix 1.0.5 full surface | **12/12 x3** | 58,719 ms | 275,784 |
| Serena | **12/12 x3** | 68,728 ms | 388,271 |
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](docs/benchmarks/refactor-vs-competitors-2026-08.md).
## Contract
The 11 Refactor tool names, schemas, and result states are frozen in
[`contract/refactor-tools.v1.json`](contract/refactor-tools.v1.json). All refactor operations are
native Rust; the host uses the read-only
[`weavatrix-rust`](https://github.com/sergii-ziborov/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`](https://github.com/sergii-ziborov/weavatrix-refactor-js) and keeps a
separate state directory, token store, and rollback journal.
## License
MIT.