Behavior tests for AI agent workflows.

import {
  artifact, command, defineDyno, sequence, skill,
} from '@dynobox/sdk';

export default defineDyno({
  name: '[commit-skill] no writes',
  target: 'commit-skill',
  scenarios: [{
    name: 'no push',
    harnesses: ['claude-code', 'codex', 'opencode'],
    prompt: 'Use the commit skill to commit the README.md '
      + 'change. Do not push or amend.',
    setup: [
      'git init',
      'git config user.email tester@dynobox.com',
      'git config user.name "Dynobox Test"',
      'git add . && git commit -m "chore: initial commit"',
      'printf "\nCommit skill smoke change.\n" >> README.md',
    ],
    assertions: [
      sequence.inOrder([
        command.called('git', {args: ['status']}),
        command.called('git', {args: ['diff']}),
        command.called('git', {args: ['commit']}),
      ]),
      skill.referenced('commit'),
      artifact.exists('.agents/skills/commit/SKILL.md'),
      command.notCalled('git', {args: ['push']}),
      command.notCalled('git', {args: ['commit', '--amend']}),
    ],
  }],
});

name: "[commit-skill] no writes"
target: commit-skill
scenarios:
  - name: no push
    harnesses: [claude-code, codex, opencode]
    prompt: >-
      Use the commit skill to commit the README.md change.
      Do not push or amend.
    setup:
      - git init
      - git config user.email tester@dynobox.com
      - git config user.name "Dynobox Test"
      - 'git add . && git commit -m "chore: initial commit"'
      - 'printf "\nCommit skill smoke change.\n" >> README.md'
    assertions:
      - type: sequence.inOrder
        steps:
          - type: command.called
            executable: git
            command: {args: [status]}
          - type: command.called
            executable: git
            command: {args: [diff]}
          - type: command.called
            executable: git
            command: {args: [commit]}
      - type: skill.referenced
        skill: commit
      - type: artifact.exists
        path: .agents/skills/commit/SKILL.md
      - type: command.notCalled
        executable: git
        command: {args: [push]}
      - type: command.notCalled
        executable: git
        command: {args: [commit, --amend]}

Assertions are evaluated against evidence recorded during the run. A model does not grade the result.