weavatrix-refactor

Evidence-backed, transactional refactoring for coding agents.

weavatrix-refactor is the write-capable member of the Weavatrix family. It combines the complete read-only weavatrix-js code-intelligence MCP with 11 refactoring tools that can prove a change, preview it against the current working tree, apply it atomically, refresh the graph, and roll it back.

It is substantially more than a rename wrapper:

• semantic JS/TS rename through the bundled language server;
• SQL table and field rename with schema-aware evidence;
• strict graph-plus-lexical rename for Python, Rust, Go, Java, C#, and Solidity;
• several related JS/TS renames merged into one atomic transaction;
• signature, symbol-body, import, bulk-replace, move, and delete-safety workflows;
• byte-exact file hashes, before text, provenance, uncertainty, and graph revision in every applyable plan;
• stale-tree detection, repository locking, rollback bundles, and automatic rollback after a mid-write failure;
• architecture and cycle projection before structural moves;
• post-change blast-radius and verification tools inherited from the core.

Why this is a separate package

The MIT weavatrix-js core is physically read-only: its published artifact has no repository source-write path. This package is the explicit write boundary. Installing it and selecting its refactor profile makes the edit capability visible; without this package, the server cannot modify source.

The split is a safety property, not packaging cosmetics:

weavatrix-js core       weavatrix-refactor                 repository
read-only evidence  ->  plan + preview + confirmation  ->  atomic write
graph / LSP / audit     hashes / provenance / rollback     refreshed graph

Architecture

The implementation is a ports-and-adapters system with one-way boundaries:

edit-plan model  <-  filesystem / lock / token adapters
       ^                         ^
       |                         |
plan engines     ->  preview / apply / rollback workflows  ->  MCP adapter

• Model validates the frozen weavatrix.edit-plan.v1 envelope and applies byte-exact edits as pure string transformations.
• Platform adapters own real-path containment, atomic replacement, repository locks, single-use tokens, and durable rollback bundles.
• Plan engines compute rename, signature, symbol, import, move, delete, and bulk-replace evidence without owning the write workflow.
• Application workflows bind a preview to the current repository, issue and consume confirmation tokens, apply under a lock, and restore on failure.
• MCP adapter composes the 11 refactor tools with the read-only weavatrix-js catalog and exposes one stdio server.

The checked-in strict architecture contract enforces zero runtime cycles, files no longer than 300 lines, and functions no longer than 100 lines. It has no exceptions or ratchet baseline.

What makes the refactor workflow different

An ordinary editor rename answers: "Which text edits should I make now?" Weavatrix Refactor also answers:

Question	Evidence returned
Is this the exact symbol?	Stable graph symbol id plus parser/LSP selection range
Which references are proven?	Per-edit provenance: EXACT_LSP, RESOLVED, EXTRACTED, or LEXICAL_EXACT
What was not proven?	Explicit uncertainReferences, notModified, warnings, and PARTIAL completeness
Did the tree change after preview?	File sha256 plus exact before text rechecked under the write lock
Can several renames partially succeed?	No. Related renames are conflict-checked and applied as one transaction
What happens after a disk/write failure?	Already-written files are restored; a durable rollback bundle remains
Will a move worsen architecture?	Projected runtime cycles, boundary violations, improvements, and blast radius
Did the refactor preserve behavior-shaped structure?	Refreshed graph plus verified_change caller/import/reference conservation

The system fails closed when proof is insufficient. It never upgrades an INFERRED edge into an applyable edit and never hides an ambiguous reference.

The complete rename workflow

rename_symbol and rename_related_symbols are complete operations, not PLANNED-only helpers. Each method owns both phases.

1. Preview

Call the rename method normally:

{
  "symbol": "src/users.ts#getUser@12",
  "new_name": "getCustomer"
}

The method computes the rename, validates every plan file against the working tree, and returns PREVIEW_OK with a short-lived confirmToken. Preview never writes source and does not require the environment write gate.

2. Apply through the same method

Repeat the same operation inputs and add the confirmation:

{
  "symbol": "src/users.ts#getUser@12",
  "new_name": "getCustomer",
  "mode": "apply",
  "confirm_token": "<token from preview>"
}

The tool recomputes the deterministic plan, verifies that the token belongs to that plan and repository, takes the repository lock, rechecks hashes and before text, writes a rollback bundle, and applies every edit bottom-up.

The same contract applies to a coordinated set:

{
  "renames": [
    {"symbol": "src/api.ts#getUser@8", "new_name": "getCustomer"},
    {"symbol": "src/api.ts#getOrder@20", "new_name": "getPurchase"}
  ]
}

rename_related_symbols detects overlapping edits, chains, swaps, shadowing risk, and per-sub-rename failure before it issues a token. Apply is one atomic multi-file operation.

Refactoring tools

Complete write workflows

Tool	What it actually does
rename_symbol	Cross-language preview/confirm/apply rename. Dispatches to exact JS/TS LSP, SQL schema, or strict graph+lexical backends; returns honest backend completeness and every uncovered reference.
rename_related_symbols	Coordinates up to 50 JS/TS symbol renames in one shared language-server session and one atomic edit plan. Detects conflicts, chains, swaps, snapshot drift, and any failed sub-rename before writing.
apply_edit_plan	Generic two-phase executor for weavatrix.edit-plan.v1 envelopes from the other tools or weavatrix-online. Preview issues a plan-bound token; apply writes atomically with rollback.
rollback_last_apply	Restores the latest pre-apply bundle. Refuses if post-apply files drifted; retries converge after an incomplete restore.

Proven plan producers

Tool	What it actually does
change_signature	Adds or removes a JS/TS function or method parameter. Performs byte-exact declaration and call-argument surgery; spread calls and value-requiring additions remain explicit uncertainty.
edit_symbol	Uses the indexed parser range for replace_symbol_body, insert_before_symbol, or insert_after_symbol. JS/TS output is parse-gated; line endings and UTF-16 coordinates are preserved.
bulk_replace	Two-stage, occurrence-selective replacement over indexed files. First returns stable occurrence ids; the second call accepts chosen ids or an exact expected count and emits a hash-bound plan. Literal mode is the default; regex replacements use real capture expansion.
organize_imports	Removes only provably unused named JS/TS imports. Default and namespace imports stay uncertain; side-effect imports are untouched; sorting is deliberately left to the formatter.

These plans are applied with apply_edit_plan, using the same preview, token, atomic-write, and rollback protocol as rename.

Structural review and safety tools

Tool	What it actually does
move_file	Builds a JS/TS relocate review: rewrites importer specifiers and the moved file's own relative imports, then projects architecture effects. File renaming itself remains an explicit editor/agent action, so this is intentionally not an apply envelope.
move_symbol	Projects a declaration move without inventing byte edits. Reports introduced/removed runtime cycles, target-file dependencies, architecture violations or improvements, and blast radius.
delete_readiness	Returns safe: true, false, or UNPROVEN with known references, dynamic/reflection risks, confidence, and the declaration span. Exported symbols are capped at UNPROVEN; deletion is never automated.

Language and proof matrix

Surface	Backend	Applyable provenance	Completeness contract
JavaScript / TypeScript rename	Bundled TypeScript language server	EXACT_LSP	COMPLETE only when the language-server result and repository boundary are complete
SQL table rename	Schema-aware SQL scanner across SQL and host files	EXTRACTED / LEXICAL_EXACT	Reports every skipped or ambiguous reference
SQL field rename	Definition-safe SQL backend	Proven definition edits only	Usages remain UNPROVEN rather than guessed
Python / Rust / Go / Java / C# / Solidity rename	Indexed graph references plus exact lexical location on the recorded line	EXTRACTED / LEXICAL_EXACT	Always PARTIAL; ambiguous lines are never edited
JS/TS signature and imports	Parser plus graph call/reference evidence	EXTRACTED / RESOLVED	Explicitly partial where graph reach cannot prove absence
Symbol-anchored edit	Indexed parser ranges for every indexed language	EXTRACTED	JS/TS parse gate; other languages retain the parser-range evidence boundary

Edit-plan proof envelope

Every applyable plan uses weavatrix.edit-plan.v1. Its load-bearing fields are:

• operation and graph revision;
• repository-relative target paths only;
• sha256 of every target file;
• exact 1-based line and UTF-16 character ranges;
• exact before and after text;
• per-edit provenance;
• uncertainReferences, notModified, warnings, and completeness.

The applier additionally protects against:

• absolute paths, traversal, .git casing/trailing-dot tricks, NTFS streams, and escaping symlinks/junctions;
• non-UTF-8 or oversized files;
• overlapping edits, stale ranges, lone surrogates, and edits that split surrogate pairs;
• two writers interleaving in the same repository;
• token reuse, expiry, repository mismatch, or plan mismatch;
• partial writes and incomplete rollback.

createdAt is provenance metadata and is the only field excluded from the confirmation fingerprint. This allows a rename method to recompute the same plan on its apply call; every executable field remains token-bound.

Result states agents can act on

State	Meaning
PREVIEW_OK	Every hash and before text matches; a single-use token was issued.
PREVIEW_BLOCKED	The generated plan does not match the current tree; nothing can be applied.
WRITE_GATE_CLOSED	The server was not deliberately started with source edits enabled.
APPLIED	Every planned edit was written and the rollback bundle is available.
STALE	The working tree changed between preview and the locked apply check; nothing was written.
TOKEN_UNKNOWN / TOKEN_EXPIRED / TOKEN_*_MISMATCH	Confirmation is absent, consumed, expired, or belongs to another plan/repository.
REPO_BUSY	Another apply or rollback currently owns the repository lock.
ROLLED_BACK	A failed apply or explicit rollback restored the original files.
ROLLBACK_INCOMPLETE	Restoration was blocked for named files; the durable bundle remains retryable.
INVALID_PLAN	Schema, path, range, encoding, overlap, or provenance validation failed before writing.

Planner-specific states such as NOT_FOUND, NO_CHANGE, CONFLICT, BLOCKED, UNPROVEN, and NOT_SUPPORTED remain visible instead of being collapsed into a generic failure.

The three write gates

Repository source changes require all three:

1. weavatrix-refactor is installed and the refactor profile selects edit;
2. the server starts with WEAVATRIX_ALLOW_SOURCE_EDITS=1;
3. the apply call presents a valid, unexpired, single-use token bound to the exact plan and repository.

Preview and every read-only analysis remain available while the environment gate is closed.

End-to-end change proof

The package includes all 34 read-only core tools in the same MCP server. A strong refactor session can therefore stay in one evidence chain:

1. inspect_symbol, context_bundle, or get_dependents identifies the exact target;
2. rename_symbol, change_signature, move_symbol, or another refactor tool previews the change;
3. the write workflow applies atomically;
4. the next graph call auto-refreshes changed files and reverse importers;
5. verified_change compares callers, imports, and references against the merge base;
6. change_impact, verify_architecture, coverage_map, run_audit, and find_duplicates inspect the consequences.

Useful inherited surfaces include:

• architecture maps and navigation: module_map, query_graph, shortest_path, context_bundle;
• impact and proof: change_impact, get_dependents, prepare_change, verified_change;
• health: run_audit, find_dead_code, find_duplicates, coverage_map, hot_path_review;
• contracts: list_endpoints, trace_endpoint, trace_api_contract;
• target architecture: get_architecture_contract, verify_architecture, explain_architecture_violation;
• repository control: open_repo, rebuild_graph, graph_diff, list_known_repos.

See the weavatrix-js README for the complete JavaScript host catalog.

Run it

Start the merged read-only-plus-refactor MCP server for one repository:

npx -y weavatrix-refactor <repoRoot>

For an MCP client, the minimal configuration is:

{
  "mcpServers": {
    "weavatrix": {
      "command": "npx",
      "args": ["-y", "weavatrix-refactor", "/absolute/path/to/repository"]
    }
  }
}

On Windows, use npx.cmd when the client does not resolve command shims. With no environment override, every analysis and preview tool works but source writes fail closed. Add "env": {"WEAVATRIX_ALLOW_SOURCE_EDITS": "1"} only for a session in which apply and rollback are deliberately authorized.

Applications that already host weavatrix-js can compose the same extension:

import {startMcpServer} from 'weavatrix-js/mcp-runtime'
import {refactorExtension} from 'weavatrix-refactor/extension'

await startMcpServer({
  defaultCapabilities: 'refactor',
  loadExtensions: async () => [refactorExtension()],
})

The exported extension registers tools and the refactor capability profile; it does not silently open the write gate.

Scope and honest limits

• Related multi-symbol rename is currently JS/TS-only.
• move_file cannot rename the file through apply_edit_plan; it is a review plan because file relocation has different filesystem semantics.
• move_symbol is a topology/architecture dry-run, not byte-edit synthesis.
• Graph+lexical language backends cannot prove reference completeness and stay PARTIAL even when every known reference was located.
• delete_readiness never auto-deletes, and public/exported APIs cannot receive an automatic clean verdict.
• Tests, typechecking, runtime checks, and human review remain the release authority. Weavatrix supplies bounded evidence; it does not fabricate proof.

Package boundary

Package	License	Responsibility
weavatrix-js	MIT	Read-only JavaScript graph, analysis, evidence, architecture, and verification
weavatrix-refactor	MIT	Proven refactor plans, transactional writes, and rollback
weavatrix-online	MIT	Explicit public network connector and remote plan/evidence workflows

The refactor package extends the legacy JavaScript core only through weavatrix-js/extension-api and weavatrix-js/analysis-kit; it does not copy or relicense that core. The canonical weavatrix package is the native Rust engine and is not this JavaScript extension host.

License

MIT.