commit-delta

git bisect for a dirty working tree.

You have a known-good HEAD, a pile of uncommitted changes, and a test that fails on the full working tree. commit-delta finds a minimal failure-inducing change set — the smallest 1-minimal subset of those hunks that still fails the test.

It will not tell you that one line is “the cause”. Several hunks may need each other. Another 1-minimal subset may also fail. The output is a reduced patch you can read, apply, and hand to a reviewer or an agent.

HEAD:          PASS
Working tree:  14 files, 93 hunks, 567 lines, FAIL

$ commit-delta -- ./reproduce.sh

Minimal failure-inducing change set:

calc.py
  hunk 2
    -        acc = acc + item
    +        acc = acc - item

Same failure. 93 hunks → 1 hunk.

Download the demo as MP4 — live xfce4-terminal capture, 15s.

On a real multi-file Python tree (four feature commits left uncommitted): 28 files / 2230 lines → 2 files / 69 lines in about 8 seconds. That run returned a 1-minimal fail, not the only one. Two other independent fails lived in the same tree. That is the tool working as designed.

Install

Requires Git 2.5+ (git worktree) and Python 3.10+.

pip install commit-delta

From a clone, for development:

git clone https://github.com/robbyczgw-cla/commit-delta.git
cd commit-delta
python3 -m venv .venv
source .venv/bin/activate   # Windows: .venv\Scripts\activate
pip install -e ".[dev]"

Use

Run the same command that is already failing:

commit-delta -- ./reproduce.sh
commit-delta -- pytest -q
commit-delta --output reduced.patch --json reduced.json -- npm test

Everything after -- is your reproduction command. It runs inside an isolated worktree. Your dirty tree is only read, never rewritten.

What your command should return

Same contract as git bisect run:

If “cannot compile” and “assertion failed” both exit 1, wrap the command and map unbuildable subsets to 125. For Python, SyntaxError / ImportError and (when no AssertionError appears) TypeError / AttributeError are already treated as UNRESOLVED.

The command must work from a clean checkout plus the candidate hunks. The probe tree does not include your node_modules or local virtualenv.

How it works

1. Verify HEAD is GOOD and the full dirty tree is BAD.
2. Snapshot git diff HEAD -U0 and untracked text files.
3. Reduce at file granularity, then hunk granularity (ddmin).
4. Apply each candidate in a temporary worktree. Cache results.
5. Print a report; optionally write reduced.patch and reduced.json.

For coding agents

See AGENTS.md. After tests go red, run once:

commit-delta --output reduced.patch --json reduced.json -- <the-same-test>

Prefer reduced.json over scraping the text report. Work only on that subset. Do not loop. The field unique_not_guaranteed is always true.

Compared to nearby tools

Limitations (v0.1)

• Atoms are hunks, not statements. A 40-line rewrite is one piece.
• The result is a 1-minimal FAIL set. Uniqueness is not guaranteed.
• Text files only. No binaries, submodules, or ignored build trees.
• The isolated worktree has no local node_modules / venv unless your command creates them.
• Renames and unusual encodings are best-effort.
• npm test, tsc, and cargo need an exit-125 wrapper if they use 1 for both compile errors and test failures.

Details: docs/limitations.md.

Develop

pip install -e ".[dev]"
make test    # unit + fixture e2e
make bench   # candidate counts / reduction ratios
make demo    # 567-line dirty tree → one hunk

Docs

• AGENTS.md — when to run, what to do with the output
• docs/algorithm.md — ddmin, heuristics, apply
• docs/limitations.md — what v0.1 will not do
• docs/benchmark.md — trial counts and the demo
• docs/competition.md — why this is not git bisect

License

Apache-2.0. See LICENSE.