hunky

git add -p for agents and scripts.

Oh, mama. You're an AI agent — or a script, or a CI job — and you just want to stage some of a file's changes, like a civilized creature. So you reach for git add -p and... it's an interactive TUI. You can't drive it. No hands, no keyboard, no party. The one git operation a non-interactive caller can't do is the exact one you need. Sure is annoying.

If only some good-lookin' tool could stage just the hunks you point at, no TUI required. Hey there, pretty mama. That's hunky: it parses the diff, rebuilds a clean patch from the hunks — or even the individual lines — you select, and pipes it to git apply --cached. Man, those are some pretty hunks.

Single-file Go, zero dependencies.

Install

go install github.com/whiskeytuesday/hunky@latest
# or, from a checkout:
go build -o hunky .

Usage

hunky --list <file>               list hunks with numbers and @@ headers
hunky --count <file>              print the number of hunks
hunky <file> <hunk>...            stage hunks by number (1-based)
hunky --invert <file> <hunk>...   stage every hunk EXCEPT the listed ones
hunky --lines START-END <file>    stage whole hunks overlapping a line range
hunky --pick-lines A-B <file>     stage only the +/- lines within A-B (splits a hunk)
hunky --dry-run <file> <hunk>...  print the patch instead of staging
hunky --stdin <file> <hunk>...    read the diff from stdin instead of `git diff`

By default it reads git diff HEAD -- <file> (HEAD vs working tree). With --stdin it reads a diff you provide — and if that diff spans multiple files, the <file> arg selects which one to operate on.

The reliable loop

hunky --list f.go          # 1. see numbered hunks
hunky --dry-run f.go 4     # 2. PREVIEW the exact patch (don't skip this)
hunky f.go 4               # 3. stage just those hunks
git diff --cached --stat   # 4. confirm what landed
git commit -m "..."        # 5. the leftover hunks are the next commit

Step 2 matters: a hunk header (@@ -1017,7 …) tells you where but not what — a one-line fix can hide between feature hunks. The dry-run shows the actual +/- lines.

Splitting a single hunk: --pick-lines

When two unrelated changes land in one hunk with no unchanged line between them, whole-hunk selection can't separate them. --pick-lines A-B reconstructs the hunk keeping only the +/- lines whose worktree line numbers fall in A-B (dropping other additions, demoting other deletions to context):

# stage one changelog entry out of an adjacent pair of new entries
hunky --pick-lines 25-37 CHANGELOG.md

Staging everything but a few hunks: --invert

When most of a file is one concern and only a hunk or two are separate, name the few and invert — cheaper than listing all the rest:

hunky --count f.go            # how many hunks are there?
hunky --invert f.go 4         # stage all of them except hunk 4

How it works

It parses git diff, reconstructs a valid unified patch from the selected hunks/lines, and pipes it to git apply --cached. It never does string-replace surgery on patch bodies, so literal text that looks like sed backrefs or special characters passes through untouched.

The one gotcha: hunk numbers are stable until you commit

The tool diffs HEAD vs the working tree, so the index already reflects what you've staged. Therefore:

• Staging more hunks does not renumber. --list once, then stage 1 3 5 and later 2 4 against the same numbering.
• Committing does renumber — committed changes leave the HEAD-vs-worktree diff, so the next --list starts at 1 with only what's left. (invalid hunk number 5 (have 4 hunks) means you reused stale numbers after a commit.)

Rule of thumb: one --list per commit.

Limitations

• One file per invocation. Multi-file diffs parse correctly and the <file> arg selects one, but each run operates on a single file.
• Base is always git diff HEAD. There's no --base <ref> mode — and on purpose: the staged patch is applied to the index with git apply --cached, so a base other than HEAD would fail to apply wherever the index and that base already diverge.
• Quoted paths containing spaces (rare) aren't matched by the <file> selector.

License

MIT — see LICENSE.