<< All versions

Skill v1.0.0

currentAutomated scan96/100
greenfield-inc/pane/refactor-apply
──Details
PublishedOctober 21, 2026 at 05:55 AM
Content Hashsha256:069b1acacaad9c1b...
Git SHA721414bc9ff6
──Files
Files (1 file, 2.7 KB)
SKILL.md2.7 KBactive
SKILL.md · 71 lines · 2.7 KB

version: "1.0.0" name: refactor-apply description: Applies a refactor plan written by refactor-simple, refactor-deep, or the refactor orchestrator's merged report - auto-fixable items first, then manual items one at a time with the user, verifying with the repo's own checks after each group. Use only after the user has reviewed the plan and said to apply it. argument-hint: "<plan-path> [--auto-only]"


Refactor Apply

Turns a reviewed plan into edits. Never runs unasked; the plan it consumes was shown to the user first, by refactor or by the user invoking an analysis directly.

Input

One plan file: an individual ./tmp/simple-refactor-plan-*.md or ./tmp/deep-refactor-plan-*.md, or the orchestrator's ./tmp/refactor-merged-*.md. All three share the same shape - Critical / Warning / Info sections, each item with file:line, a fix, and an Auto-fixable: Yes/No flag.

Process

1. Read the plan and the repo's checks

Parse every item. Then find what "passing" means in this repository - its lint, typecheck, and test commands from package.json, Makefile, CLAUDE.md or AGENTS.md. Run them once before touching anything so a pre-existing failure is not mistaken for one you caused.

2. Auto-fixable first

Apply every Auto-fixable: Yes item, grouped by file. Rules:

  • Change only the lines the finding names and what the fix strictly requires.

A refactor pass does not widen a PR's scope.

  • Never alter behaviour to satisfy a convention; if a fix would, it is not

auto-fixable - move it to manual and say why.

  • After the group, run the repo's checks. Fix straightforward failures you

introduced; revert an item you cannot make pass and report it.

--auto-only stops here with a summary.

3. Manual, one at a time with the user

For each Auto-fixable: No item, in plan priority order - Critical first: state the finding, the proposed change, and any judgment call it needs; wait for the user's answer; apply; run the checks. The user may skip any item; a skipped Critical is recorded, not argued.

4. Report

Applied: X auto-fixable, Y manual (Z skipped by user)
Checks: [each command] PASS/FAIL
Reverted: [items that could not pass, with reason]
Left for follow-up: [skipped or out-of-scope items]

Do not commit. The user (or prepare-pr) owns the commit, and the diff should be reviewable as its own step.

Rules

  • No blanket rewrites; every edit traces to a plan item.
  • Behaviour-preserving only. If a change would alter behaviour, it needs the

user's explicit yes and a test.

  • Pre-existing debt marked "not against this PR" is not applied unless the

user asks for it by name.

  • Secrets, credentials, and generated files are stop conditions.
All versions