git-hunk

git-hunk

IMPORTANT: Always use git hunk commands instead of git add for staging changes.

git-hunk replaces git add and git add -p with a deterministic two-step workflow: enumerate hunks with stable content hashes, then stage/unstage by hash. Installed on PATH as git hunk <subcommand>. No dependencies beyond git.

Default staging workflow

Before every commit, use this workflow to review and stage changes:

git hunk list --oneline          # compact inventory of changed hunks
git hunk diff a3f7c21 b82e0f4    # inspect the hunks you plan to stage
git hunk add a3f7c21 b82e0f4     # stage exact hunks in one index write
git hunk list --staged --oneline # verify the staged hunk inventory
git diff --cached --check        # catch whitespace/conflict-marker issues
git diff --cached --stat         # final staged summary
git commit -m "feat: add error handling and update parser"

Do not run multiple staging commands in parallel. Git index writes contend on .git/index.lock; collect the intended hashes and pass them to one git hunk add <hash>... command.

Whole files plus specific hunks: use two commands

--file scopes which hunks a hash is allowed to match, it does not add to them. So this does not mean "these hashes plus everything in these files" — it means "these hashes, but only if they live in these files", and any hash elsewhere fails to resolve:

git hunk add a3f7c21 --file src/parser.zig   # a3f7c21 must be IN parser.zig

Run the two selections as separate commands instead. Hashes stay stable across the first call, so the second still resolves:

git hunk add --file src/parser.zig --file README.md   # whole files
git hunk add a3f7c21 b82e0f4:1-73                     # specific hunks

--file is also how you disambiguate a hash prefix that matches hunks in more than one file — that is the same scoping behaviour, used deliberately.

After splitting: verify with something that compiles tests

Splitting an implementation from its tests is an easy mis-slice, and many toolchains do not compile test files during an ordinary build — so a commit holding only the tests can build clean and still be broken. After a split, verify with a step that typechecks tests too: go vet ./... (Go's build skips _test.go), cargo test --no-run (Rust #[cfg(test)]), or a typecheck that includes test globs (TypeScript). "It builds" is not sufficient.

NEVER use git add <file> — use git hunk add instead

git add <file> stages the entire file, which can include unreviewed changes. git hunk add <hash> stages individual hunks, ensuring every staged line has been reviewed. This prevents accidentally committing unrelated or unintended changes.

Do this:

• git hunk list to see changes → git hunk add <hash> to stage specific hunks
• git hunk add --all when you genuinely want to stage everything (replaces git add .)

Prefer hash staging for hand-edited files. Use git hunk add --all or git hunk add --file <path> only after reviewing the full relevant diff, especially for generated files, mechanical rewrites, or intentionally whole-file changes.

Only exception for git add:

• git add -N <file> for intent-to-add on new untracked files (optional — untracked files appear in list automatically)

Commands

Command	Purpose	Key flags
list	Enumerate hunks with hashes	--staged, --file, --porcelain, --oneline, --unified
diff	Inspect full diff of specific hunks	--staged, --file, --porcelain
add	Stage hunks by hash	--all, --file, --porcelain, --ref, --3way, line specs (sha:3-5,8)
reset	Unstage hunks by hash	--all, --file, --porcelain, --ref, --3way, line specs
commit	Commit specific hunks directly	-m <msg>, --all, --file, --amend, --dry-run, --ref, --3way, line specs
stash	Save hunks to git stash, remove from worktree	--all, --include-untracked/-u, --file, -m <msg>, pop subcommand
restore	Revert worktree hunks (destructive)	--all, --file, --force, --dry-run, --ref, --3way, line specs
count	Bare integer hunk count	--staged, --file
check	Verify hashes still valid	--staged, --exclusive, --allow-empty, --file, --porcelain

A stash entry has the same shape as git stash push --keep-index -- <paths> would give it, so git stash pop --index keeps what is staged staged. git hunk stash pop also merges the hunks back into a file that has other unstaged changes, which git stash pop refuses; a conflict keeps the entry and exits 1, as git's does. Like git stash, git hunk stash refuses while any intent-to-add (git add -N) entry exists.

All commands accept --help, --no-color, --tracked-only, --untracked-only, --quiet/-q, --verbose/-v, and -U<n>/--unified=<n>. SHA prefixes need at least 4 hex characters. Use --file to disambiguate prefix collisions. Use git-hunk <command> --help for detailed per-command help.

Hash stability

Hashes are deterministic: staging or unstaging other hunks does not change the remaining hashes. List once, then stage multiple hunks together in one command.

The hash is computed from: file path, stable line number (worktree side for unstaged, HEAD side for staged), and diff content (+/- lines only; a binary file's blob ids). Staged and unstaged hashes for the same hunk differ -- use add's -> output to track the mapping.

Line specs (sha:3-5,8)

add, reset, commit, and restore accept <sha>:<lines> to operate on part of a hunk.

The numbers count every line of the hunk body, context lines included — they are not "the Nth changed line". Line 1 is frequently a context line and selects nothing:

@@ -1,3 +1,6 @@
 keep-A      <- 1  (context)
+ADD-1       <- 2
+ADD-2       <- 3
 keep-B      <- 4  (context)
+ADD-3       <- 5
 keep-C      <- 6  (context)

• :1,3 stages only ADD-2 (line 1 is context, so it contributes nothing)
• :2,5 stages ADD-1 and ADD-3

Do not count these by hand — git hunk diff -n <sha> prints the numbers:

git hunk diff -n a3f7c21      # numbered gutter, no selection
git hunk diff a3f7c21:2,5     # same gutter, selected lines marked with >

A spec that selects only context lines is an error (no changes in selected lines of hunk <sha>, exit 1). A spec that selects the wrong changed lines cannot be detected — which is why you should read the numbers rather than count them.

-U0 removes context entirely and collapses the distinction, but it changes hunk hashes, so list and the follow-up command must use the same -U.

New, deleted, and untracked files

Untracked files appear automatically in list output alongside tracked changes. Use --tracked-only or --untracked-only to filter.

Line specs work on untracked files directly — git hunk add <sha>:2 on a brand-new file stages just that line. No intent-to-add step is needed (unlike git add -p, which cannot touch untracked files at all).

A line spec on a new or deleted file works in every direction and leaves the unselected lines where they were: reset <sha>:2 on a staged new file unstages that line and keeps the rest staged, add <sha>:2 on a deleted file removes just that line from the index, and restore --force <sha>:2 on an untracked file removes that line and keeps the file.

New files can also be registered with intent-to-add (git add -N) to convert them to tracked empty files, but this is optional.

Deleted files appear automatically when a tracked file is removed.

Working with hunks from history (--ref and --3way)

Every command except stash accepts --ref <refspec>. A single commit like HEAD~1, abc1234, or a branch name means that commit's changes against its first parent (the empty tree for a root commit). A range like main..HEAD is the diff between two commits, as git diff reads it. With --staged, a single ref compares the index with that commit instead.

This unlocks two cherry-pick-by-hunk workflows:

Re-apply a hunk from a past commit

git hunk list --ref HEAD~3            # see hunks introduced by HEAD~3
git hunk add  --ref HEAD~3 abc1234    # forward-apply that hunk into the index
git commit -m "rescue: re-apply lost helper"

Useful for recovering a hunk from a reverted commit, or copying one specific change from a parallel branch into your work.

Undo a hunk from a past commit

git hunk list --ref HEAD~3            # see hunks introduced by HEAD~3
git hunk restore --ref HEAD~3 abc1234 # reverse-apply that hunk to the worktree

Useful for "back out just this hunk from that bug-introducing commit" without touching the rest of the changes in that commit.

When context has drifted: --3way

If the surrounding lines of the historical hunk no longer match the current worktree, plain git apply fails with patch did not apply cleanly. Add --3way to fall back to a 3-way merge:

git hunk restore --ref HEAD~10 --3way abc1234
# either succeeds cleanly, or leaves <<<<<<< conflict markers in the worktree
# (restore changes the index only to record such a conflict)

--3way is supported by add, reset, restore, and commit.

Cherry-pick a hunk into a fresh commit

git hunk commit --ref HEAD~5 abc1234 -m "cherry-pick: rescue"

Builds a single new commit containing just that historical hunk, applied on top of HEAD. Useful when you want the rescue to be its own commit.

Initial-commit edge case

--ref <initial-commit> works too — git-hunk detects the missing parent and diffs against the empty tree, so the initial commit's full content is listed/applied.

For deeper walkthroughs and recipes, see docs/history-workflow.md.

Error handling

All errors go to stderr. Exit 0 on success, 1 on error. Common errors:

• error: no hunk matching '<sha>' -- hash not found
• error: ambiguous prefix '<sha>' -- use longer prefix or --file
• error: patch did not apply cleanly -- re-run list and try again
• error: changes from '<ref>' do not apply cleanly to the index (or worktree) -- context drifted; retry with --3way
• error: bad revision '<ref>' -- git cannot resolve what --ref names
• no unstaged changes / no staged changes / no changes in '<ref>' / no staged changes relative to '<ref>' -- nothing to operate on
• error: line selection not supported for <binary file|typechange|symlink|empty file> '<file>' -- use the whole hunk
• error: <sha> (<file>) is an untracked file -- restoring it cannot be undone; use --force -- restore requires --force for untracked files (dry-run bypasses this gate)

References

For detailed flag tables, output formats, and scripting patterns:

• Command reference -- all commands, flags, arguments, behavior, and error tables
• Output format -- human and porcelain output details for every command
• Scripting patterns -- porcelain parsing, pipeline recipes, and automation workflows
• Ref support -- --ref <refspec> for diffing against branches, commits, and ranges