{
  "skill_name": "cmux-orchestrator",
  "evals": [
    {
      "id": 1,
      "prompt": "Set up a workshop for bread-builder",
      "expected_output": "Creates three-pane workshop layout (terminal left, browser right, test watcher bottom full-width), detects project commands from lockfile/scripts, verifies commands started, sets sidebar dashboard",
      "expectations": [
        "Recognizes 'workshop' as a named convention and builds the specific layout",
        "Sets up sidebar dashboard with meaningful status entries and renames workspace",
        "Splits down FIRST for full-width bottom pane, then splits top pane right",
        "Verifies commands worked via read-screen after sending them"
      ],
      "files": []
    },
    {
      "id": 2,
      "prompt": "Set up an ops deck to run tests and lint on planventura",
      "expected_output": "Spawns separate agent workspaces with agent-inbox coordination, sidebar progress tracking, and dual-channel monitoring",
      "expectations": [
        "Recognizes 'ops deck' as a named convention and spawns multiple agent workspaces",
        "Sets up agent-inbox directories for orchestrator, test-runner, and linter",
        "Bakes inbox protocol instructions into each spawned agent's prompt",
        "Uses sidebar dashboard with per-agent status tracking and progress bar",
        "Describes the dual-channel monitoring pattern (read-screen for real-time, inbox for results)"
      ],
      "files": []
    },
    {
      "id": 3,
      "prompt": "Open the Cloudflare Workers docs in a browser and find the D1 bindings section",
      "expected_output": "Opens browser pane, uses snapshot discovery to navigate, sets sidebar status with findings",
      "expectations": [
        "Opens browser pane with Cloudflare Workers docs URL",
        "Uses snapshot --interactive or --compact to discover navigation elements rather than guessing selectors",
        "Navigates to the D1 bindings section using discovered elements",
        "Sets sidebar status summarizing what was found"
      ],
      "files": []
    },
    {
      "id": 4,
      "prompt": "I have two Claude agents running in separate cmux workspaces — one doing a code review in workspace:3 and one running tests in workspace:4. Set up agent-inbox so they can coordinate, and have the test runner send its results to the reviewer when done.",
      "expected_output": "Creates inbox directories, constructs proper messages with YAML frontmatter, uses atomic write pattern, shows bidirectional message flow",
      "expectations": [
        "Creates inbox directories for both agents with proper new/tmp/archive structure",
        "Constructs agent-inbox messages with correct YAML frontmatter (from, to, reply_to, timestamp, thread)",
        "Uses atomic write pattern: writes to tmp/ then mv to new/",
        "Shows the message flow direction correctly (test-runner sends results TO reviewer)"
      ],
      "files": []
    },
    {
      "id": 5,
      "prompt": "Set up a respond-to-pr-review workshop for PR #42 on the workspaces repo",
      "expected_output": "Fetches PR context and review comments, reads unresolved comments before acting, reflects in context, summarizes do/defer/decline decisions, creates/reuses worktree on PR branch, builds workshop layout with browser on PR Files Changed tab, launches agent with review context and correct permissions, sidebar shows PR number and comment progress",
      "expectations": [
        "Recognizes 'respond-to-pr-review workshop' as a named convention and follows the Respond to PR Review Workshop flow",
        "Fetches PR metadata and review comments using gh pr view and gh api before creating workspace",
        "Requires the agent to read all unresolved comments before execution and classify each as do, defer, or decline",
        "Creates or reuses worktree idempotently (doesn't error if worktree already exists)",
        "Opens browser pane to the PR's Files Changed URL",
        "Includes the permission profile for autonomous verification (swift test, uv run, python)",
        "Describes the re-review closing flow: commit, push, comment summary, request re-review via script",
        "Sets sidebar with PR number in title and comment-addressed progress tracking"
      ],
      "files": []
    }
  ]
}