/workflow:checkpoint-refactor

Multi-phase refactoring with persistent state that survives context limits and session boundaries.

When to Use This Skill

Context

• Repo root: !git rev-parse --show-toplevel
• Plan file exists: !find . -maxdepth 1 -name REFACTOR_PLAN.md
• Git status: !git status --porcelain
• Recent commits: !git log --oneline --max-count=5

Parameters

• --init: Create a new refactoring plan interactively
• --continue: Resume from the last completed phase
• --status: Show current plan progress
• --phase=N: Execute a specific phase

Plan File Format

The plan file (REFACTOR_PLAN.md) serves as persistent state. It is the loop's compact state packet (.claude/rules/loop-integrity.md): a fresh session, or a sub-agent with no memory of prior phases, must be able to re-enter from this file alone. Every phase therefore carries not just what to do but what was verified, what changed, and what must come first — and the plan names the phase the last run chose next. Without those, resuming across a context limit silently redoes or undoes work, or re-derives the order from the transcript.

# Refactor Plan: {description}

Created: {date}
Last updated: {date}
Base commit: {hash}
Exit condition: {the literal criterion that ends the whole refactor — e.g. "all phases done, full suite + tsc green on base"}
Next target: {the phase the last run chose to do next, and why — set before a run ends}

## Overview
{What is being refactored and why}

## Phase 1: {phase name}
- **Status**: done | in-progress | pending | needs-review
- **Files**: file1.ts, file2.ts, file3.ts
- **Description**: {what this phase does}
- **Ordering / preconditions**: {what must hold before this phase starts — e.g. "none (first phase)"; keep it to a line or two}
- **Acceptance criteria**: {how to verify success — the phase's exit condition}
- **Verifier result**: {what the independent check returned — PASS/FAIL + the criterion it judged; filled in at the phase boundary}
- **Changed since last run**: {what this phase actually touched, so a successor doesn't redo or undo it}
- **Result**: {summary of changes made, filled in after completion}

## Phase 2: {phase name}
- **Status**: pending
- **Files**: file4.ts, file5.ts
- **Description**: {what this phase does}
- **Ordering / preconditions**: {e.g. "after Phase 1 — uses its shared types"}
- **Acceptance criteria**: {how to verify success}
- **Verifier result**: {empty until verified}
- **Changed since last run**: {empty until completed}
- **Result**: {empty until completed}

...

Execution

Execute this multi-phase refactoring workflow:

Step 1: Initialize refactoring plan (--init mode)

If --init flag provided:

1. Analyze scope: Read files to be refactored, understand dependencies
2. Define phases where each:
   • Touches 3-7 files (bounded scope)
   • Has clear acceptance criteria (tests, type check)
   • Can be committed independently
   • Builds on previous phases
3. Write plan file: Create REFACTOR_PLAN.md at repo root
4. Record base commit: git log --format='%H' -1

Phase ordering: Shared utilities/types first, leaf components last, tests alongside implementation.

Step 2: Resume existing refactor (--continue mode)

If --continue flag provided:

1. Read REFACTOR_PLAN.md
2. Take the plan's Next target; if it is empty, fall back to the first pending phase (status pending or needs-review)
3. Verify the phase's Ordering / preconditions hold (prior phases it names are done); if one does not, work the unmet precondition instead
4. Execute phase (go to Step 4)

Step 3: Check progress (--status mode)

If --status flag provided:

1. Read REFACTOR_PLAN.md
2. Parse plan and display status table with phases, descriptions, statuses, file counts
3. Exit

Step 4: Execute target phase (--phase=N or selected via Step 2)

For each phase:

1. Read context from plan file (current phase's details)
2. Read only the files listed for this phase
3. Implement changes according to phase description
4. Validate with appropriate tool (tsc, ty check, cargo check, or npm/pytest test)
5. If validation fails:
   • Fix errors if straightforward
   • If complex, mark phase as needs-review with error details
   • Commit partial work with WIP: prefix
6. If validation passes, gate done on an independent verifier when the acceptance criteria are a judgement (e.g. "the class is now single-purpose"), not a purely mechanical check (a green suite / clean tsc is already independent — a failing test does not care how hard you worked):
   • Dispatch a fresh Task sub-agent that reads only this phase's acceptance criteria and the resulting diff — not the worker's reasoning — and judges whether the criteria are met. The worker is biased toward declaring completion; the loop's stop condition must come from outside it (.claude/rules/loop-integrity.md, Pillar 1). When the criteria are about behaviour ("endpoint returns X", "the bug no longer reproduces"), delegate this to agent-patterns-plugin:execution-grounded-review, which runs the suite first and grounds each criterion in execution evidence rather than appearance.
   • Record the verdict in the phase's Verifier result field (PASS/FAIL + the criterion judged).
   • If the verifier returns FAIL, leave status in-progress/needs-review and address the gap before marking done.
7. Once verified, update the plan file: set status to done, fill Verifier result, Changed since last run, and Result, set the plan's Next target, then commit: git add -u && git commit -m "refactor phase N: {description}"
8. If more phases remain, proceed to next phase or suggest --continue

Step 5: Sub-agent delegation (for large phases)

For phases with 7+ files, delegate to Task sub-agent with:

• File list to modify
• Phase description and acceptance criteria
• Instructions: run validation, update plan file, stage/commit changes
• If validation fails: mark phase as needs-review with error details

Recovery Patterns

Agentic Optimizations

Quick Reference

Related Skills

• code-review-checklist - Review refactored code
• refactoring-patterns - Refactoring techniques
• adversarial-review - The isolated verifier a judgement-based phase gate delegates to
• execution-grounded-review - The execution-grounded verifier for behaviour-based phase acceptance criteria
• .claude/rules/loop-integrity.md - Why the plan file is a state packet and why done is judged independently