review-shift

review-shift reviews your local git branches overnight and leaves a report and an apply-able git patch on your desk by morning — instead of the review you keep postponing until after the merge.

Русская версия

Status: v0.1.0 released. pipx install review-shift installs it from PyPI.

What it does

At night, with nobody present, review-shift:

1. finds local branches that moved recently (refs/heads/, by committer date);
2. builds each branch's diff against its merge base with the base branch;
3. masks secret values before anything is sent to the model;
4. runs a read-only code review at the configured depth;
5. writes a markdown report, a findings.json, and two patch files — and validates every patch with git apply --check before writing it to disk.

In the morning you read one file and apply one patch. review-shift never applies anything itself, never commits, never pushes, and never touches your working tree.

Requirements

• macOS for the scheduled path (launchd), on mains power and awake — see Limitations. The CLI itself runs on Linux and macOS, x86 and ARM.
• Claude Code CLI 2.1.0 or newer, already authenticated.
• Python 3.11+.

Install

pipx install review-shift

Golden path

# set the repo up: writes .review-shift/config.yml, keeps run artifacts out of git
cd ~/proj/myrepo && review-shift init

# check the environment before trusting it with a night
review-shift doctor

# schedule the nightly run (renders the launchd job, registers the wake-up)
review-shift init launchd

Then, in the morning:

$EDITOR .review-shift/runs/latest/report.md

Demo

asciinema play demo/review-shift.cast

Findings in the recording are scripted for a fast, free, reproducible playback — init and the patch/report artifacts you see are real review-shift output, produced by the same patch.resolve / patch.generate_and_verify / report.render code a real overnight run uses (the recording skips the live model call review-shift run makes, and doesn't run doctor either, since its own auth check is a live call too).

In-session review (/review-shift)

review-shift also works from inside a Claude Code session, as a skill that shells out to the same CLI (ADR-006) — same prompts, same lock, same report. It is capped to one branch and depth <= medium. Two ways to get it, not alternatives to pick between:

# zero marketplace dependency, exact `/review-shift` command, no auto-update
review-shift init skill

# in a Claude Code session: self-hosted marketplace, auto-updatable, namespaced command
/plugin marketplace add https://github.com/yushman/review-shift
/plugin install review-shift@review-shift
# invoked as /review-shift:review-shift (Claude Code always namespaces plugin skills)

init skill writes .claude/skills/review-shift/SKILL.md in the current repository; re-run it after upgrading review-shift to pick up a changed skill. Both channels can be installed at once — Claude Code keeps the original /skill-name and the plugin copy side by side.

Applying a patch

Deliberately manual, and deliberately three steps. The patch is bound to the branch head the review ran against.

# 1. the sha the patch was built against is in the patch header and in run.json
git -C . rev-parse feature/payments-v2

# 2. check applicability
git switch feature/payments-v2
git apply --check .review-shift/runs/latest/patches/auto_fixed.patch

# 3. apply
git apply .review-shift/runs/latest/patches/auto_fixed.patch

If the branch has moved past that sha, the CLI says so and prints an explanation instead of the recipe, rather than pretending the patch still applies.

Output

.review-shift/
├── config.yml                  # meant to be committed
└── runs/
    ├── 2026-08-22T03-30-00Z-feature-payments-v2/
    │   ├── report.md           # what you read in the morning
    │   ├── findings.json
    │   ├── run.json            # branch, shas, depth, cost, timings, counters
    │   ├── events.jsonl
    │   └── patches/
    │       ├── auto_fixed.patch  # severity >= patch.auto_fix_min_severity, default high
    │       └── all.patch
    ├── index.json
    └── latest -> …

Exit codes

The scheduler templates pass --exit-zero-on-findings, so a night that honestly finds problems does not look like a broken job.

Limitations — read these

• A closed laptop on battery will not wake up. launchd does not wake the machine; the run would happen at your next wake, which is far too late to be useful. The tool registers a pmset wake-up and wraps the run in caffeinate, but neither helps a machine on battery with the lid shut.
• Your code is sent to the model provider. Check your employer's policy and your plan's terms before pointing this at work code.
• Secret masking reduces exposure, it does not guarantee it. Regex heuristics miss custom token formats, and the agent has its own filesystem access. Not for code under regulatory constraints.
• No quality numbers are published yet. Recall and precision are only claimed once the benchmark bench exists (v0.2). What v0.1 measures is patch applicability.
• Run artifacts contain code fragments and live in your working tree.
• v0.1 scope: depths low and medium, local branches only, one repository per run. Diffs above ~2 000 changed lines are skipped with an explicit reason rather than truncated. high, chunking, retention and incremental review are v0.2; remote branches and PR integration are v0.3.

License

MIT — see LICENSE.