Amabie

A verification gate for coding agents. The agent implements a change; Amabie runs the project's checks and compares their result with the starting code. A claim of “done” is never proof.

npm install --global amabie-cli
amabie setup --yes

Run setup once, from your own terminal, inside a Git repository with declared checks and Claude Code or Codex CLI. Requires Node.js 22+, npm and Git.

Version: this source is 2.0.2. It fixes false regression evidence when a baseline cannot find installed dependencies, and preserves open-session tracking outside the repository. Upgrade from 2.0.1 before relying on dependency-backed baseline evidence. The session guards and CLI registration fallback described below were introduced in 2.0.1. These fixes are absent from 2.0.0. Check the published version with npm view amabie-cli version. To test this checkout, run npm ci, npm run build:force, node tooling/pack.mjs, then install ./pack/npm/amabie-cli-2.0.2.tgz.

Reproduce this case: original upstream tests, a green replacement command, a blocked completion, and restoration. This is a simulated agent completion; no model or reliability rate is measured.

Start with a failing test

1. Write a test for the required behaviour and confirm it fails. Commit that reviewed test and run amabie setup --yes before starting the agent task.
2. Ask the agent to implement the behaviour without changing the existing test.
3. At completion, Amabie runs the sealed checks. A check that fails without the implementation and passes with it provides evidence about the change.

The workflow guide also explains new agent-written tests, existing-test edits and refactoring.

Result	What it means
READY	The checks and applicable proof obligations passed. Read any evidence-origin caveat.
NOT PROVEN	The available evidence does not establish the change or a requirement. This is not a finding that the code is wrong.
NEEDS ATTENTION	A check failed or verification could not run safely. The output points to the cause and full log.

Claude Code registration

Approve the project server once in Claude's /mcp menu. The agent registers each request with amabie_task. If approval is pending, the request hook gives it a CLI alternative:

amabie task "implement the requested behaviour" --requirement "exact criterion" --session <id-from-hook>

Registration records the request; it does not approve or prove it. The agent cannot use setup, bind, accept, uninstall, or direct protected-file edits to approve its own work through the Claude hook. Administrative changes during an open session require review and confirmation in a human terminal.

Scope and evidence

Claude Code hooks can block completion. Codex hooks require the user's trust; other integrations can be advisory. The current capability is LOCAL: worker and verifier share an OS user. These guards are not an OS sandbox against a determined same-user process.

No token-saving or comparative superiority claim is established. Whether an agent is helped by the gate in everyday use is also not established: the saved benchmarks measure what the gate catches and blocks, not better agent outcomes. Historical benchmarks and their withdrawn claims are kept in the dated reference.

• Runnable quickstart
• Documentation index and how the gate works
• Compatibility
• Security contract
• Contributing and module map

Apache-2.0. Historical canary command names are compatibility aliases.