non-interactive-git-rebase（非交互式 git rebase）

Non-Interactive Git Rebase

Overview

git rebase -i normally opens an editor for the todo list. By setting GIT_SEQUENCE_EDITOR to a command that replaces the todo file with a pre-prepared one, you can perform any interactive rebase operation — reorder, drop, squash, edit, amend — entirely from scripts.

CRITICAL: Always Create a Backup Branch

Before any rebase, create a backup branch:

git branch backup-before-rebase HEAD

After the rebase succeeds and verification passes, do NOT delete the backup branch automatically. Always ask the user before deleting it, unless the user explicitly requested automatic cleanup. The backup is cheap to keep and the user may discover problems later.

Core Technique

How GIT_SEQUENCE_EDITOR Works

The editor receives the todo file path as $1. Any command that overwrites that file works:

# Replace the todo with our prepared version
GIT_SEQUENCE_EDITOR="cp /tmp/prepared-todo" git rebase -i <base>

Git generates the default todo (all pick lines), then invokes cp /tmp/prepared-todo <todo-file>, which replaces it with your version. Git then executes the rewritten todo.

Preparing the Todo File

Extract the default todo without actually rebasing — copy it out, then fail the editor so git aborts:

GIT_SEQUENCE_EDITOR='cp $1 /tmp/default-todo && false' git rebase -i <base>

The false exits non-zero, so git aborts the rebase. You have the todo file, no rebase happened.

Or build it directly from git log:

git rev-list --reverse <base>..HEAD | while read hash; do
  echo "pick $hash $(git log --oneline -1 $hash | cut -d' ' -f2-)"
done > /tmp/prepared-todo

Then edit the file: reorder lines, change pick to drop/squash/edit/etc.

Using --exec for Bulk Amendments

The --exec <cmd> flag inserts an exec <cmd> line after every pick in the todo. Combined with GIT_SEQUENCE_EDITOR=: (a no-op that succeeds, leaving the default todo unchanged), this runs a command after each commit is replayed — no editor needed.

Change author email on all commits in a range:

GIT_SEQUENCE_EDITOR=: git rebase -i <base> --exec \
  'if [ "$(git log -1 --format=%ae)" != "correct@email.com" ]; then git commit --amend --no-edit --author="Name <correct@email.com>"; fi'

Change commit message pattern:

GIT_SEQUENCE_EDITOR=: git rebase -i <base> --exec \
  'git commit --amend -m "$(git log -1 --format=%B | sed s/old/new/)"'

Operations

Reorder Commits

Rearrange lines in the todo file. Example — move commit abc123 to the front:

# Build todo with abc123 first, then everything else in original order
{
  echo "pick abc123 The commit to move first"
  git rev-list --reverse <base>..HEAD | while read hash; do
    full=$(git rev-parse abc123)
    [ "$hash" != "$full" ] && echo "pick $hash $(git log --oneline -1 $hash | cut -d' ' -f2-)"
  done
} > /tmp/prepared-todo

GIT_SEQUENCE_EDITOR="cp /tmp/prepared-todo" git rebase -i <base>

Drop Commits

Remove lines from the todo, or change pick to drop.

Reword Commit Messages

NEVER use reword in the todo. The reword command opens an interactive editor for the new message, which cannot be scripted via GIT_SEQUENCE_EDITOR. Instead, use edit and amend with -m:

# Step 1: Prepare todo with 'edit' on commits to reword

# Step 2: At each stop, amend the message non-interactively:
git commit --amend -m "New subject line

New body text."

# Step 3: Continue to the next stop
git rebase --continue

Squash / Fixup

Change pick to squash or fixup. For squash, you also need GIT_SEQUENCE_EDITOR for the commit message editor that opens after squashing:

# Squash second commit into first, keep first message
# In the todo: pick aaa, fixup bbb
GIT_SEQUENCE_EDITOR="cp /tmp/prepared-todo" git rebase -i <base>

Split a Commit

This requires edit in the todo. When rebase stops, you run commands to split, then continue.

# Step 1: Prepare todo with 'edit' on the target commit
sed 's/^pick TARGET_HASH/edit TARGET_HASH/' /tmp/default-todo > /tmp/prepared-todo
GIT_SEQUENCE_EDITOR="cp /tmp/prepared-todo" git rebase -i <base>

# Step 2: Rebase stops at the target commit. Split it:
git reset HEAD~1

# Step 3: Selectively stage and commit pieces.
# If the commit has N hunks in a file and you want hunk 3:
printf 'n\nn\ny\n' | git add -p path/to/file
git commit -m "First piece: just hunk 3"

# Stage the rest
git add .
git commit -m "Second piece: remaining changes"

# Step 4: Continue the rebase
git rebase --continue

Automating hunk selection with git add -p: Each y/n in the printf corresponds to a hunk in order. This works reliably when hunks are well-separated (different regions of the file). If hunks are adjacent, git may present them as one hunk, and you may need s (split) first:

# Split first hunk, then accept/reject sub-hunks
printf 's\ny\nn\ny\n' | git add -p path/to/file

Alternative — patch-based splitting (more reliable for complex cases):

git reset HEAD~1
git diff path/to/file > /tmp/full.patch
git checkout -- path/to/file

# Apply only the lines you want for commit 1
# (manually prepare /tmp/part1.patch from full.patch)
git apply /tmp/part1.patch
git add path/to/file
git commit -m "First piece"

git apply /tmp/full.patch  # won't re-apply already-applied hunks
# Or: git checkout <original-hash> -- path/to/file
git add .
git commit -m "Second piece"

Verification

After any rebase, verify the tree content is identical to before (unless you intentionally dropped commits):

git diff backup-before-rebase HEAD --stat
# Empty output = content preserved correctly

Common Mistakes

Quick Reference