Recomposing Commits

Recomposing Commits

Overview

Analyze messy commits → isolate in a worktree → let user review → apply back.

Core principle: The original branch is NEVER modified until the user explicitly approves the new history. Worktree isolation and the review gate are non-negotiable — even if the user says "just do it" or "I trust you, skip review."

Announce at start: "I'm using the recomposing-commits skill to reorganize your branch history."

Safety Check — Do This First

git rev-parse --abbrev-ref HEAD

Refuse immediately if the current branch is any of: main, master, dev, develop, development, staging, production, release, or matches release/* or hotfix/*.

⛔ Recomposing commits on '<branch>' is not allowed.
This skill only works on feature branches.
Please switch to your feature branch first.

Also refuse if HEAD is detached (the command above prints HEAD instead of a branch name):

⛔ You're in a detached HEAD state, not on a branch.
This skill needs a real branch to reset back onto at the end. Check out a branch first.

Also check for a dirty working tree before doing anything else:

git status --porcelain

If this prints anything, stop:

⛔ You have uncommitted changes. These won't be included in the recompose,
and they'll be at risk when the original branch is reset at the end.
Please commit or stash them first.

Do NOT create a workaround branch. Do NOT proceed on any of the above. Refuse and stop.

When to Use

• Commits are WIP/messy and a PR is coming
• Multiple unrelated changes were mixed into the same commits
• Commit messages are unclear ("fix", "WIP", "stuff")
• Commits need to be split, merged, or reordered

Do NOT use when:

• History is already clean
• You are on a protected/shared branch (refuse instead)
• There is only one commit to restructure
• The range contains a merge commit (see Step 3 — not supported, linear history only)

Workflow (Steps 1–8)

Step 1: Show current history

git log --oneline

Display the output to the user.

Step 2: Ask for the start SHA

Which commit should be the start of the recompose?
Provide the SHA of the first commit you want to reorganize.
(All commits from that SHA through HEAD will be recomposed.)

Enter SHA:

Validate before proceeding:

# Must actually be an ancestor of HEAD
git merge-base --is-ancestor <user-sha> HEAD || echo "NOT AN ANCESTOR"

If it prints NOT AN ANCESTOR, tell the user the SHA isn't in the current branch's history and ask again.

Root-commit edge case: if <user-sha> is the repository's very first commit, it has no parent:

git rev-parse <user-sha>^ 2>&1   # will error: "unknown revision" if it's the root commit

If so, stop and tell the user this case isn't automated: recomposing from the repo's root commit needs manual handling that this skill doesn't cover safely. Ask them to either pick a start point after the root commit, or handle the root commit's history by hand.

Otherwise compute the base normally:

BASE=$(git rev-parse <user-sha>^)

Step 3: Analyze the commits to recompose

git log --oneline $BASE..HEAD
git diff $BASE HEAD

Read all diffs. Identify logical groups. Flag any files that appear in multiple logical groups (overlapping files — see Step 6).

Refuse if the range contains a merge commit:

git log --merges --oneline $BASE..HEAD

If this prints anything:

⛔ This range contains a merge commit. This skill only handles linear history —
recomposing across a merge isn't supported. Pick a start point after the merge,
or restructure the branches involved separately.

Check for multiple authors (recomposing resets authorship and commit dates to whoever runs this):

git log --format='%ae' $BASE..HEAD | sort -u

If more than one address appears, tell the user before continuing:

⚠️  Commits in this range have multiple authors. Recomposing will reset every
new commit's author and date to yours. Continue anyway?

Step 4: Create isolated worktree

Always use an absolute path. Relative paths like .git/recompose/feat/HISBA-mastra-ai silently break when branch names contain slashes AND when shell state doesn't persist between tool calls.

BRANCH_NAME=$(git rev-parse --abbrev-ref HEAD)
REPO_ROOT=$(git rev-parse --show-toplevel)
WORKTREE_PATH="$REPO_ROOT/.git/recompose/$BRANCH_NAME"
WORKTREE_BRANCH="recompose/$BRANCH_NAME"

Clean up leftovers from a previous or interrupted run first — otherwise worktree add/-b fails on a path or branch that already exists:

if git worktree list | grep -q "$WORKTREE_PATH"; then
  git worktree remove "$WORKTREE_PATH" --force
fi
if git rev-parse --verify "$WORKTREE_BRANCH" >/dev/null 2>&1; then
  git branch -D "$WORKTREE_BRANCH"
fi

git worktree add "$WORKTREE_PATH" -b "$WORKTREE_BRANCH"

All subsequent git operations happen inside $WORKTREE_PATH. The original branch is untouched.

Step 5: Collapse commits to unstaged

Use git -C with the absolute path — do NOT rely on cd. Each Bash tool invocation starts with a fresh shell; cd from a previous call does not carry over.

git -C "$WORKTREE_PATH" reset --mixed $BASE

All changes are now unstaged in the worktree, ready to be re-staged logically.

Step 6: Stage and commit each logical group

Use git -C "$WORKTREE_PATH" for every command. Never assume the shell is cd'd into the worktree.

For files that belong cleanly to ONE logical commit:

git -C "$WORKTREE_PATH" add <file> <file2>
git -C "$WORKTREE_PATH" commit --no-verify -m "type: description"

For overlapping files (same file spans multiple logical commits):

# git add -p requires a TTY — run it inside a single shell block
cd "$WORKTREE_PATH" && git add -p <file>   # stage hunk by hunk: y/n/s/e
git -C "$WORKTREE_PATH" commit --no-verify -m "type: first concern"
cd "$WORKTREE_PATH" && git add -p <file>
git -C "$WORKTREE_PATH" commit --no-verify -m "type: second concern"

For a NEW/untracked file that spans multiple commits, git add -p alone won't offer hunk selection — it only works on tracked files. Mark it intent-to-add first:

git -C "$WORKTREE_PATH" add -N <new-file>   # tracks it for diff purposes, stays unstaged
cd "$WORKTREE_PATH" && git add -p <new-file>

NEVER use git add -i or git rebase -i — these require interactive TTY input and will hang or fail in this context.

Step 7: Review gate — MANDATORY, non-negotiable

After all commits are created in the worktree, always present this to the user:

Recomposition complete in isolated worktree.

To review:
  cd .git/recompose/<branch-name>
  git log --oneline

Original branch is UNCHANGED. Nothing will be applied until you approve.

Note: commits were created with --no-verify (hooks skipped) — run your
linter/formatter after applying. Authorship and commit dates on the new
commits reflect whoever ran this, not the original commits.

Does the new history look good?
  [yes]       → Apply to original branch
  [no/edit]   → Describe what to change, I'll update the worktree
  [cancel]    → Discard worktree, original branch stays as-is

Wait for explicit user response. Do not proceed to Step 8 without it.

If the user previously said "no review, just do it": You must still present this review gate. Explain: "The review step is part of this skill's safety guarantee. I've isolated the changes — reviewing takes 30 seconds and means the original branch is safe. Here's what the new history looks like:"

Step 8: Handle user response

[yes] — Apply back:

Before touching the original branch, verify the diff is identical (see Verification section below) — do not proceed past a "DIFFERENCES FOUND" result.

Check for remote upstream first:

git -C "$REPO_ROOT" remote -v
git -C "$REPO_ROOT" branch -vv

If branch tracks a remote:

⚠️  This branch has a remote upstream. Applying recomposed commits will
   require a force-push. Do you want to proceed?

Wait for confirmation, then:

# Safety net: tag where the original branch was, in case anything needs to be undone
git -C "$REPO_ROOT" tag "backup/$BRANCH_NAME-$(date +%s)" "$BRANCH_NAME"

# Confirm we're actually resetting the branch we think we are
CURRENT=$(git -C "$REPO_ROOT" rev-parse --abbrev-ref HEAD)
if [ "$CURRENT" != "$BRANCH_NAME" ]; then
  echo "⛔ Expected to be on $BRANCH_NAME but found $CURRENT — aborting."
  exit 1
fi

# DO NOT run `git checkout $BRANCH_NAME` — you are already on it in the main worktree.
# git checkout will fail with "already used by worktree". Skip it entirely.
git -C "$REPO_ROOT" reset --hard "recompose/$BRANCH_NAME"

# Verify it worked
git -C "$REPO_ROOT" log --oneline -5

# Clean up — worktree FIRST, then branch
git -C "$REPO_ROOT" worktree remove "$WORKTREE_PATH"
git -C "$REPO_ROOT" branch -D "recompose/$BRANCH_NAME"

If a remote upstream was confirmed above, push it:

git -C "$REPO_ROOT" push --force-with-lease

Prefer --force-with-lease over --force — it fails safely if someone else pushed to the remote branch in the meantime, instead of silently overwriting their work.

Mention the backup tag to the user (backup/$BRANCH_NAME-<timestamp>) so they know how to recover the pre-recompose state if needed, and that it's safe to delete once they're happy with the result.

[no/edit] — Iterate:

To adjust a commit already made in the worktree without touching rebase -i, uncollapse back to just before it and redo Step 6 from there:

# Find the SHA of the commit right before the one you need to change
git -C "$WORKTREE_PATH" log --oneline

# Uncollapse everything from that point forward back to unstaged
git -C "$WORKTREE_PATH" reset --mixed <sha-before-the-commit-to-fix>

Then restage and recommit following Step 6's approach, and re-present the review gate (Step 7).

[cancel] — Discard:

git worktree remove "$WORKTREE_PATH"
git branch -D "recompose/$BRANCH_NAME"

Original branch is untouched.

Overlapping Files Strategy

Verification (Before Declaring Done)

# 1. Confirm new clean history
git -C "$WORKTREE_PATH" log --oneline $BASE..HEAD

# 2. Full content diff — stats alone are NOT enough
# Stats can match while content differs. Always do a byte-for-byte comparison.
diff \
  <(git -C "$REPO_ROOT" diff $BASE HEAD) \
  <(git -C "$WORKTREE_PATH" diff $BASE HEAD) \
&& echo "IDENTICAL — diffs match exactly" || echo "DIFFERENCES FOUND — investigate before proceeding"

Stats are not sufficient. 66 files changed, 8393 insertions matching does not prove content matches — only a full diff comparison does. Always use the diff <(...) <(...) form and confirm "IDENTICAL" before presenting the review gate.

If it says "DIFFERENCES FOUND": stop. Do not present the review gate and do not apply back. Go back to Step 5/6, figure out what was missed or misassigned (a hunk staged to the wrong commit, a file left out), and re-run this check until it says "IDENTICAL".

Quick Reference

Common Mistakes / Rationalization Table

Red Flags — STOP

If you find yourself thinking any of these, stop and follow the skill:

• "I'll use git rebase -i"
• "User said skip review, so I'll skip it"
• "I'll work directly on the original branch since it's faster"
• "I'll create a backup branch instead of a worktree"
• "The protected branch thing doesn't matter here"
• "I can do git add <file> for the overlapping file since it's cleaner"
• "I'll cd into the worktree, it's simpler than git -C"
• "The stats match so the diff is fine, I'll skip the content check"
• "I need to checkout the branch before resetting to it"
• "The working tree is a little dirty, it's probably fine to continue"
• "There's a merge commit in range, I'll just reset through it anyway"
• "I don't need a backup tag, nothing will go wrong"
• "I mentioned the force-push is needed, that's the same as doing it"