Reimplement Branch

Create a new branch with a clean, narrative-quality commit history from an existing branch's changes.

Gather Context

Run these commands to understand the current state:

git branch --show-current          # Source branch
git status --short                 # Uncommitted changes
git log main..HEAD --oneline       # Commits since main
git diff main...HEAD --stat        # Full diff summary

Workflow

Use the autonomous workflow when the user wants the agent to complete the entire reimplementation, verification, and PR flow end to end.

1. Validate source branch

• Ensure no uncommitted changes or merge conflicts
• Confirm branch is up to date with main

2. Analyze the diff

Study all changes between source branch and main. Form a clear understanding of the final intended state.

3. Create clean branch

git checkout main
git checkout -b <new-branch-name>

Use the user-provided branch name, or {source-branch}-clean if none provided.

4. Plan commit storyline

Break the implementation into self-contained logical steps. Each step should reflect a stage of development—as if writing a tutorial.

5. Reimplement the work

Recreate changes in the clean branch, committing step by step. Each commit must:

• Introduce a single coherent idea
• Include a clear commit message and description

Use git commit --no-verify for intermediate commits. Pre-commit hooks check tests, types, and imports that may not pass until full implementation is complete.

6. Verify correctness

• Confirm final state exactly matches source branch: git diff <source-branch>
• Run final commit without --no-verify to ensure all checks pass

7. Open PR

Use the /pr skill to create a pull request. Include a link to the original branch in the PR description.

Interactive Stacked PR Workflow

Use this workflow when the user wants to collaborate on the reimplementation timeline, create one clean change at a time, or build a stack of PRs where each branch depends on the previous branch.

1. Treat the source branch as a snapshot

Record the source branch name and inspect its full diff against main, but do not assume the source branch will be merged.

git branch --show-current
git status --short
git log main..HEAD --oneline
git diff main...HEAD --stat

2. Discuss the commit and PR timeline

Break the snapshot into reviewable milestones with the user. Prefer milestones that can stand alone as stacked PRs:

• Foundation changes first, such as file moves, schema, or shared types
• Server-side behavior next
• API boundaries after the backing behavior exists
• UI integration last
• Documentation or agent instruction changes separately unless they are required

Do not start committing until the user confirms the initial timeline or the next branch to carve out.

Standalone-first carve-out pattern

It is okay, and often preferable, to break the snapshot into several small PRs instead of trying to make every branch correspond to the original feature's internal implementation order. Start by separating out changes that are independently useful and easy to review:

• Repository instructions, skills, generated metadata, or other tooling changes
• Runtime/client swaps that do not depend on the feature
• Pure UI refactors that preserve behavior and create a cleaner place for later feature work
• File moves or namespace changes that can compile without the rest of the feature

For each standalone change, create a branch from the current stack base, stage only that change, commit it, push it, and open a PR. The rest of the snapshot can remain as unstaged working-tree changes on top of the branch so the next PR can be carved out from the same source material.

When a standalone PR merges, temporarily stash the remaining unstaged snapshot changes, update the new stack base, create the next branch, and restore the snapshot changes:

git stash push --include-untracked -m "snapshot remainder before <next-branch>"
git switch main
git pull --ff-only
git switch -c codex/<next-change>
git stash pop
git restore --staged .

If restoring the snapshot conflicts with new main changes, resolve or reset only the conflicted files needed for the current PR. Keep unrelated snapshot work unstaged. This lets the agent and user repeatedly extract focused PRs while preserving the rest of the source branch as local working-tree material.

3. Move the snapshot diff onto the stack base

Switch to the desired base, usually main, and apply the source branch diff into the working tree without staging it.

git switch main
git diff main...<source-branch> | git apply
git status --short

Keep unrelated pre-existing worktree changes separate. If needed, unstage files without discarding content before starting the stack.

4. Create the next branch from the current stack tip

For the first PR, branch from main. For later PRs, branch from the previous PR branch so the stack is explicit.

git switch -c codex/<change-name>

5. Commit only the next milestone

Stage only the files or hunks that belong to the current milestone. Leave the rest of the snapshot diff unstaged for later branches.

git add -p
git commit --no-verify

Use --no-verify for intermediate stacked commits when later milestones are needed for the full app to pass checks.

6. Verify and open a PR for that branch

Run the relevant focused checks for the current branch. Open a PR whose base is the previous stack branch, or main for the first branch. Mention the source snapshot branch in the PR description.

7. Continue from the new stack tip

After each PR branch is created, continue from that branch, carve out the next milestone from the remaining unstaged snapshot diff, and repeat until the final stacked branch matches the source snapshot.

At the end, verify that the stack tip matches the snapshot:

git diff <source-branch>

Rules

• Never add yourself as author or contributor
• Never include "Generated with Claude Code" or "Co-Authored-By" lines
• End state of clean branch must be identical to source branch