🩹 clawpatch

Automated code review that lands fixes.

clawpatch maps a repo into semantic feature slices, reviews each slice with a provider, persists findings, and can run an explicit fix loop for one finding at a time.

Current status: early CLI. Review/report/state are implemented; patching exists behind clawpatch fix --finding <id> and still requires manual review of the resulting worktree changes.

Install

pnpm add -g clawpatch

From source:

pnpm install
pnpm build
pnpm link --global

Workflow

clawpatch init
clawpatch map
clawpatch review --limit 3 --jobs 3
clawpatch report
clawpatch next
clawpatch show --finding <id>
clawpatch triage --finding <id> --status false-positive --note "covered by tests"
clawpatch fix --finding <id>
clawpatch revalidate --finding <id>
clawpatch revalidate --all --status open

fix does not commit, push, open PRs, or land changes. It runs configured validation commands and records a patch attempt under .clawpatch/.

What It Maps Today

• npm package bins
• selected package scripts: start, build, test, lint, typecheck, format
• Next.js app/ and pages/ routes
• Go package slices from go list ./..., including command packages
• Go package tests and same-repo imports as review context
• Rust src/main.rs, src/bin/*.rs, src/lib.rs, crates/*, and tests/*.rs
• SwiftPM Sources/* targets and Tests/* suites
• common project config files

Deeper framework mappers and agent-assisted enrichment are next steps.

Provider

The default provider is the local Codex CLI.

codex --version
clawpatch doctor

Provider calls use codex exec with strict JSON schemas. Review and revalidate run read-only; fix planning runs with workspace-write because Codex may edit the working tree during the explicit fix command.

Supported provider names today:

• codex: local Codex CLI
• mock: deterministic test provider
• mock-fail: failure test provider

Direct OpenAI, Claude, Gemini, and provider panels are not implemented yet.

Commands

• clawpatch init: create .clawpatch/, detect project basics, write config
• clawpatch map: write feature records
• clawpatch status: show project, dirty state, feature/finding counts
• clawpatch review: review pending or selected features
• clawpatch report: print or write a Markdown findings report
• clawpatch next: print the next actionable finding
• clawpatch show --finding <id>: inspect one finding with evidence and suggested validation
• clawpatch triage --finding <id> --status <status>: mark a finding with optional history note
• clawpatch fix --finding <id>: run the explicit patch loop for one finding
• clawpatch revalidate --finding <id>: re-check one finding
• clawpatch revalidate --all: re-check open findings with report-style filters
• clawpatch doctor: check provider availability
• clawpatch clean-locks: clear feature locks

Useful flags:

• --root <path>
• --state-dir <path>
• --config <path>
• --json
• --plain
• --limit <n>
• --jobs <n>
• --feature <id>
• --finding <id>
• --status <status>
• --severity <severity>
• --provider <name>
• --model <name>
• --output <path> / -o <path>
• --dry-run
• --force

Unknown flags fail fast.

State

State is project-local by default:

.clawpatch/
  config.json
  project.json
  features/*.json
  findings/*.json
  patches/*.json
  reports/*.md
  runs/*.json

Feature records are the durable work units. Findings and patch attempts link back to features so runs can resume and be audited.

Safety

• Review does not edit files.
• Fix is explicit and selected by finding ID.
• Fix refuses a dirty source worktree by default.
• Clawpatch never commits, pushes, opens PRs, or lands changes today.
• Provider output is parsed through strict schemas.
• Symlinked directories and generated build output are skipped during mapping.

See docs/spec.md for the longer product and implementation spec.