git-reword

Bulk edit git commit messages in your editor.

git reword                  # all commits on the current branch
git reword abc123           # a single commit
git reword HEAD~5..HEAD     # a range

Commit messages for the range are written to REWORD_EDITMSG at the worktree root, opened in $EDITOR, and the changed ones are applied by writing new commit objects (git commit-tree) and moving HEAD once. The index and the working tree are never touched, so a dirty worktree is fine. HEAD@{1} and git reset --soft get the old history back. The file is transient; the tool adds its name to .git/info/exclude on first use unless .gitignore already covers it.

The file looks like git log output. Column 0 is structure, indented lines are the message:

commit 7dcfdad1afb39b697a8632f0c450c555abe7d5b6
Author:     Ole Jørgen Brønner <ole@example.com>
AuthorDate: 2026-02-25 05:59:29 +0100

    test-env-cli: refactor the CLI interface

    The previous subcommand-based interface made it difficult to
    perform multiple operations in a single invocation.

---

---

# lines at column 0 are comments. Inside a message, # is just content. The full format is specified in prose/spec/reword-format.md.

Flags:

• --commit-link adds a forge URL comment per commit
• --author-info adds Author: and AuthorDate: lines
• --commit-info adds Commit: and CommitDate: lines, the committer
• --edit-info makes the info lines editable (implies --author-info): a run of commits made with the wrong email is then a search and replace. Every info line in the file is applied as written, so with --commit-info rewritten commits keep their committer and committer date instead of being stamped anew. The file records the mode in a # git-reword-options: edit-info line, which is how the language server and --continue know about it
• --stat adds the files each commit touched as comment lines after the message
• --no-abbrev writes full shas instead of git's abbreviations (--abbrev, the default)
• --continue reopens the file from an aborted run
• --force discards any previous aborted run

Abbreviated shas are resolved against the commits of the range when the file is read back; an ambiguous or unknown prefix is an error.

Merge commits are kept, and their messages can be reworded like any other commit's. Commits after the range up to HEAD are carried along onto the new history; commits before the first change keep their sha. The range must be in HEAD's history.

Language server

git-reword-lsp speaks LSP over stdio and finds the repository by walking up from the file, so it works on REWORD_EDITMSG and on .reword files anywhere inside a worktree. It provides:

• diagnostics: format errors with quick-fixes, subject length, unknown shas, and a hint on every commit whose message changed
• document symbols (outline): one per commit, subject plus short sha
• hover on a commit line: author, committer, dates, original message
• code actions: revert a commit to its original message, reflow the paragraph under the cursor to 72 columns, indent misindented lines, open the commit in the forge (from origin), and in Zed open it in Zed's commit view
• document links on shas, and formatting that normalises indentation
• folding ranges that keep the subject visible: fold the body under a subject, or a whole block down to its commit line with the subject as placeholder; runs of comment lines fold to their first line, so a --stat block collapses to its summary (Zed needs document_folding_ranges: "on" for the language)

Opening a commit uses window/showDocument when the editor supports it. Zed does not, so there the server runs xdg-open (or open on macOS) for forge URLs and the zed CLI for zed://git/commit/... URLs; zed must be on the PATH the server sees.

Development

uv sync
uv run pytest
uv run ruff check
uv run ty check

Disclaimer

This is a vibe-coded project in the sense that I have read little of the code.