Review your last commit in two minutes

Install the binary

Probe is a single executable. Git must be installed; nothing else is required for this guide. Download the archive for your operating system and processor, then extract it. Follow the steps for your system below.

sudo install probe /usr/local/bin/
probe version

$env:Path += ";$env:LOCALAPPDATA\Programs\probe"
probe version

The last command prints the installed version, for example probe v0.5.1. To check the archive against the published checksums, see Verify a download.

Compare your last commit with the previous one

Go to any Git repository and ask Probe what changed between HEAD~1 and HEAD:

cd path/to/your/repository
probe lint HEAD~1..HEAD

lint is the static half of Probe. It compares the two commits, collects risk signals and writes a report. It never executes repository code, never calls an AI provider and needs no .probe.json: without one, the built-in defaults for the detected language apply.

Read the report

Open .probe/CONFIDENCE_REPORT.md in your editor. Start with Suggested Human Review, a list of file and line ranges ranked by severity, each with the reason it was flagged:

## Suggested Human Review
- **high** auth/auth.go:7–7 (new): Authentication or authorization function body changed
- **high** payment/refund.go (whole file): Lines added and removed in a sensitive file; No nearby test file changed
- **high** payment/refund.go:21–21 (old): Possible input validation removed

(new) points to a line in the candidate commit and (old) to a removed line in the base. (whole file) marks a signal about the file itself, such as a sensitive path, rather than about one of its lines. Review Surface tells you how many changed lines those ranges cover. The same data, with commit IDs and signal evidence, is in confidence-report.json for scripts and CI.

Signals are reasons to look, not confirmed bugs. The signal reference explains each one.

Choose what to compare

Any two revisions work. A few common choices:

Optional: run your checks in a sandbox

review does everything lint does, then runs the repository's test, typecheck, build and coverage commands in disposable, network-less Docker containers. It needs Docker with Linux containers and a preloaded image containing your toolchain: Probe never pulls images or installs dependencies itself. For a Go project with no third-party dependencies, the default image is enough:

docker pull golang:1.26-bookworm
probe review HEAD~1..HEAD --reviewer=false

For Node.js, Python and Rust projects, stock images do not contain your dependencies: build an image that does and point sandbox.image at it in a policy (next step), or let the trusted policy build that layer with a prepare object. If Docker or the image is missing, Probe stops with exit code 4 and never falls back to running code on your machine.

Adopt it in your repository

Run probe init to generate .probe.json for your detected language. Review its commands and sandbox image, then try the policy locally:

probe review --base main --config .probe.json

Commit .probe.json to your base branch. Subsequent reviews read that trusted policy. On your feature branch, run:

probe review --base main --ci

Working with a coding agent? Adopt the plan-first process: the agent runs probe plan --intent-file task.md --ci before writing code (a flagged plan goes to a human), implements the plan, and CI runs probe review --plan .probe/PLAN.json --ci. Exit 0 means the change conforms to a low-risk plan and passed its checks, so it can merge without a human review; exit 2 lists why a human must look (plan gate).

Next steps: every flag, exit code and policy key is in the reference. To let an AI reviewer try to reproduce issues, configure a provider as described in AI reviewer. For GitHub Actions, see the CI integration guide.