[](https://crates.io/crates/git-agent-verdict)
[](https://github.com/fredrikolis/git-agent-verdict/actions/workflows/ci.yml)
[](LICENSE)
`git-agent-verdict` is a CLI tool that blocks agents from committing code that does not comply with
your organization's standards. It uses git's built-in `commit-msg` hook to enforce that every commit
carries a passing verdict. Built for repos where an agent is the sole maintainer and no human is
expected to see the diff before it lands.
**Review is mandatory**, and happens before the commit exists.
**No self-review.** The tool invokes the reviewer, not the agent, and the one line the agent supplies
is auto-rejected if it argues for the change instead of stating what it does.
**The verdict rides in the commit**, with the neutral intent as the subject. `git log` is the audit
trail: what changed, who reviewed it, what they found.
## Expected agent usage
The agent finds the tool by hitting it. Nothing has to be read first.
```
Bash(git commit -am "add retry to the uploader")
⎿ git-agent-verdict: error: standards: no reviewable trailer
use `git agent-verdict attest --intent "<one line>"` to commit
```
The agent states in one neutral line what the diff does, spawns the reviewer, and waits for it.
```
Bash(git agent-verdict attest --intent "retry a failed upload three times")
⎿ git-agent-verdict: spawned attestation process (pid 48213)
use `git agent-verdict await` to wait for it
Bash(git agent-verdict await)
⎿ git-agent-verdict: BLOCKED
1-standards.log # BLOCKED
address the reported findings, then: git agent-verdict attest
```
The agent reads the log, fixes the MAJOR findings, and attests again. A gate that has passed is not
re-reviewed. Committing is its own call, so findings under a passing verdict are read before
anything lands.
```
Bash(git agent-verdict commit)
⎿ [main 3f9a1c2] retry a failed upload three times
Reviewed-standards: reviewer=claude-opus-5 major=0 moderate=2 minor=1 token=1f0c...
Reviewed-docs: reviewer=claude-opus-5 major=0 moderate=0 minor=0 token=8ba7...
```
## Install
```bash
cargo install git-agent-verdict
git config --global agent-verdict.runner claude
```
Installs as `git agent-verdict`.
## Configure a repo
Declare one gate per line in the `commit-msg` hook. Line order is review order, and the gate name
becomes the trailer key.
```bash
#!/usr/bin/env bash
set -euo pipefail
git agent-verdict --require-version 2.0
git agent-verdict "$1" standards --model opus \
--standard programming --standard testing --path .
git agent-verdict "$1" docs --simple --doc docs/house-style.md --path "*.md"
```
`.git/hooks/commit-msg` is the standard location and needs no configuration, but git does not track
it, so each clone and each agent would install its own. To make the gates a property of the repo,
commit the hook and point git at it:
```bash
chmod +x .githooks/commit-msg
git config core.hooksPath .githooks
```
`git agent-verdict --repo-setup-guide` prints the full reference: every flag a declaration takes,
what `--simple` and `--read-only` change, and how to feed a gate a rubric that lives outside the
repo. `--help` is the flag grammar.
## Contributing
The reviewer runs on Claude Code today. **Please help us add support for other agents.** The seam is
`src/runner.rs` and `src/agent.rs`: an agent is an argv, a session, and an answer this tool parses
verdicts out of. Nothing above that layer knows which one answered.