Task¶

{
  "version": 1,
  "name": "sales-summary",
  "goal": [
    {
      "type": "text",
      "text": "打开 /work/sales.xlsx，按下面的样式制作各地区销售额汇总表。"
    },
    {
      "type": "image",
      "path": "goal/1.png"
    },
    {
      "type": "text",
      "text": "保留原始数据，将结果保存为 /work/summary.xlsx。"
    }
  ],
  "metadata": {
    "source": "https://example.com/forum/posts/123",
    "software": "LibreOffice",
    "tags": [
      "spreadsheet",
      "sales"
    ]
  },
  "task_runtime_ref": {
    "backend": "docker",
    "image": "sha256:<保存后的镜像 ID>",
    "config": {
      "command": [
        "sleep",
        "infinity"
      ],
      "env": {
        "LANG": "C.UTF-8"
      },
      "user": "1000:1000",
      "memory": "4g",
      "cpus": 2,
      "network": "bridge",
      "shm_size": "1g",
      "gpus": "all",
      "args": [
        "--hostname",
        "desktop"
      ],
      "python": "python3"
    },
    "connection": {}
  },
  "context": [
    {
      "source": "context/0/sales.xlsx",
      "target": "/work/sales.xlsx"
    },
    {
      "source": "context/1/setup.py",
      "target": "/work/setup.py"
    }
  ],
  "cwd": "/work",
  "setup": {
    "command": [
      "python3",
      "/work/setup.py"
    ],
    "timeout": 120
  },
  "verifier_runtime_ref": {
    "backend": "docker",
    "image": "registry.example.com/verifiers/spreadsheet:1.0",
    "config": {
      "command": [
        "sleep",
        "infinity"
      ],
      "env": {
        "LANG": "C.UTF-8"
      },
      "memory": "1g",
      "cpus": 1,
      "python": "python3"
    },
    "connection": {}
  },
  "verifier": {
    "entry": "verifier/run.py",
    "timeout": 300,
    "required_env": [
      "LLM_BASE_URL",
      "LLM_API_KEY"
    ],
    "requires_task_instance": true
  }
}

{
  "version": 1,
  "name": "windows-sales-summary",
  "goal": [
    {
      "type": "text",
      "text": "打开 C:\\Tasks\\sales.xlsx，按下面的样式制作各地区销售额汇总表。"
    },
    {
      "type": "image",
      "path": "goal/1.png"
    },
    {
      "type": "text",
      "text": "保留原始数据，将结果保存为 C:\\Tasks\\summary.xlsx。"
    }
  ],
  "metadata": {
    "source": "https://example.com/forum/posts/123",
    "software": "LibreOffice",
    "tags": [
      "spreadsheet",
      "sales",
      "windows"
    ]
  },
  "task_runtime_ref": {
    "backend": "qemu",
    "image": "/data/envweaver/disks/runtimes/<环境 ID>/disk.qcow2",
    "config": {
      "memory": 8192,
      "cpus": 4,
      "disk_interface": "virtio",
      "network_device": "virtio-net-pci",
      "format": "qcow2",
      "machine": "q35",
      "accel": "kvm",
      "firmware": "/data/envweaver/disks/runtimes/<环境 ID>/firmware.fd",
      "nvram": "/data/envweaver/disks/runtimes/<环境 ID>/nvram.fd",
      "args": [
        "-vga",
        "std"
      ]
    },
    "connection": {
      "ssh": {
        "user": "worker",
        "python": "python"
      }
    }
  },
  "context": [
    {
      "source": "context/0/sales.xlsx",
      "target": "C:\\Tasks\\sales.xlsx"
    },
    {
      "source": "context/1/setup.py",
      "target": "C:\\Tasks\\setup.py"
    }
  ],
  "cwd": "C:\\Tasks",
  "setup": {
    "command": [
      "python",
      "C:\\Tasks\\setup.py"
    ],
    "timeout": 120
  },
  "verifier_runtime_ref": {
    "backend": "docker",
    "image": "registry.example.com/verifiers/spreadsheet:1.0",
    "config": {
      "command": [
        "sleep",
        "infinity"
      ],
      "env": {
        "LANG": "C.UTF-8"
      },
      "memory": "1g",
      "cpus": 1,
      "python": "python3"
    },
    "connection": {}
  },
  "verifier": {
    "entry": "verifier/run.py",
    "timeout": 300,
    "required_env": [
      "LLM_BASE_URL",
      "LLM_API_KEY"
    ],
    "requires_task_instance": true
  }
}

task = pool.load(task_ref=task_ref)

Docker：启动任务沙箱

task_instance = runtime.start(
    ref=task.task_runtime_ref,
    context=task.context,
    setup=task.setup,
)

QEMU：指定磁盘目录并启动任务沙箱

ref = runtime.create(
    backend="qemu",
    image=task.task_runtime_ref.image,
    config={**task.task_runtime_ref.config, "disk_dir": "/data/task-runs"},
    connection=task.task_runtime_ref.connection,
)
task_instance = runtime.start(
    ref=ref,
    context=task.context,
    setup=task.setup,
)

服务器地址填在 Runtime(host=...)；QEMU 磁盘目录填在 config.disk_dir；虚拟机 SSH 私钥和 Docker VNC 地址通过 start(connection=...) 传入。

启动验证沙箱

Docker 直接使用任务中的验证镜像引用：

verifier_instance = runtime.start(
    ref=task.verifier_runtime_ref,
)

QEMU 需要先补充磁盘目录：

verifier_ref = runtime.create(
    backend="qemu",
    image=task.verifier_runtime_ref.image,
    config={**task.verifier_runtime_ref.config, "disk_dir": "/data/verify-runs"},
    connection=task.verifier_runtime_ref.connection,
)

verifier_instance = runtime.start(ref=verifier_ref)

执行 Verify

result = task.verifier.run(
    verifier_instance=verifier_instance,
    task_instance=task_instance,
    env={"LLM_BASE_URL": base_url, "LLM_API_KEY": api_key},
)