Bug Data Generator

根据 BUG 描述生成对应的工具名和用户指令列表，供用户审核确认后再调用 data-generator 生成完整 JSONL 训练数据。触发场景：(1) 输入 BUG 描述，获取应调用的工具和触发指令列表；(2) 确认指令后自动调用 data-generator；(3) BUG 数据生成的中间步骤。

Security Scan

The skill's code, instructions, and resource needs are consistent with its stated purpose of turning bug descriptions into tool names and user-instruction lists for later JSONL generation.

Purpose & Capability

Name/description match the implementation: scripts generate instruction lists from BUG descriptions and either produce JSONL or a prompt for a separate data-generator. No unrelated credentials, binaries, or external services are requested.

Instruction Scope

SKILL.md and the scripts stick to generating instruction lists and assembling JSONL. One implementation detail: gen_bug_data.py attempts to import build_prompt from a specific path (/app/openclaw/skills/data-generator/scripts) to integrate with a separate data-generator skill; if unavailable it falls back to writing an instructions text file. This is expected behavior but means the skill assumes presence/read access to another skill's code path.

Install Mechanism

Instruction-only skill with included Python scripts and no install spec. Nothing is downloaded or installed by the skill itself.

Credentials

The skill requires no environment variables or credentials. It does write output to paths like /workspace/... and /tmp, and it inserts a hard-coded sys.path to access another skill's scripts; these file-path assumptions are legitimate for its purpose but worth confirming in your environment (permissions, shared workspace).

Persistence & Privilege

No always:true, no special persistent privileges requested, and no modifications of other skills' configurations. Autonomous invocation remains enabled by default (normal for skills).

Assessment

This skill appears to do what it claims: generate candidate user instructions from a bug description and hand them off to a separate data-generator. Before installing, consider: (1) It tries to import build_prompt from /app/openclaw/skills/data-generator/scripts — ensure the data-generator skill is present and trusted, or the script will fall back to writing text files. (2) It writes files under /workspace and /tmp; confirm those paths are acceptable and that file permissions/privacy meet your requirements. (3) There are no network calls or credential usage in the included code, but if you plan to integrate with a remote data-generator, review that component separately. If you have low trust in the environment or in other installed skills, inspect the build_prompt implementation in the data-generator skill before use.

将 BUG 描述转换为可用的训练指令序列。

工作流

用户: BUG描述 + 错误原因
        ↓
Agent: 分析 BUG → 确定工具 + 生成用户指令列表
        ↓
用户: 审核指令列表（可修改/补充/删除）
        ↓
用户: 确认 "没问题，生成数据"
        ↓
调用 data-generator → 输出 JSONL（新格式）

第一步：分析 BUG

输入以下信息，自动推断：

字段	说明	示例
BUG 描述	用户的实际指令 + 系统的错误行为	"用户说X分钟后开空调，系统却立即执行"
错误原因	为什么错 + 正确工具	"调用了 dev_control 而应该调用 scene_generator"

第二步：生成指令列表

输出内容：

工具: scene_generator

用户指令列表:
  1. 5分钟后打开空调
  2. 3分钟后关灯
  3. 10分钟后开启空气净化器
  ...（共 N 条）

发送给用户确认，等待回复。

第三步：用户修改指令

用户可直接回复修改意见：

• "把第3条改成 20分钟后打开加湿器"
• "删除第5条"
• "补充5条延时类指令"
• "指令没问题，生成数据"

第四步：调用 data-generator

收到确认后，使用最终版指令列表，调用 data-generator v2.0.0（新格式）：

输入参数：

tool_name: scene_generator
user_instructions: [确认后的完整列表]

输出 JSONL（新格式）：

{
  "conversations": [
    {"from": "human", "value": "<当前用户指令>5分钟后打开空调</当前用户指令>\n<本地设备>舒享家(空调)</本地设备>\n<当前时间>2026-03-21 20:00:00</当前时间>\n<用户场景列表>[...]</用户场景列表>\n<用户设备列表>{...}</用户设备列表>"},
    {"from": "assistant", "value": "<tool_call>{\"tool_name\":\"scene_generator\",\"query\":\"...\"}</tool_call>"},
    {"from": "observation", "value": "<tool_response>场景创建成功。</tool_response>"},
    {"from": "assistant", "value": "好的，5分钟后准时执行~"}
  ],
  "system": "",
  "history": []
}

格式规则（新格式）

1. conversations[0].value = 完整上下文（<当前用户指令> + <本地设备> + <当前时间> + <用户场景列表> + <用户设备列表>）
2. conversations[1].value = tool_call，无垫音前缀
3. conversations[2].value = <tool_response>...</tool_response>
4. conversations[3].value = 终接回复，无垫音前缀
5. system = ""，history = []

指令模板参考

生成时可参考以下模板类型：

类型	模板示例
延时-分钟	{N}分钟后打开{D}
延时-秒	{N}秒后关闭{D}
延时-小时	{H}小时后开{D}
定时-今天	今天{H}点打开{D}
定时-明天	明天{H}点关{D}
循环-每天	每天{H}点打开{D}
设备开关	打开{D}、把{D}关闭

输出格式（第一步）

{
  "tool_name": "scene_generator",
  "instruction_count": 20,
  "instructions": [
    {"id": 1, "text": "5分钟后打开空调", "type": "延时-分钟"},
    {"id": 2, "text": "3分钟后关灯", "type": "延时-分钟"}
  ],
  "note": "以上为由 AI 根据 BUG 分析生成的指令列表，请审核或修改后确认，确认后调用 data-generator 生成 JSONL。"
}

注意

• 不依赖 data-generator 内部实现：仅输出中间产物（工具名 + 指令列表）
• 可迭代：用户可多轮修改指令，直到满意再生成
• data-generator 升级时，bug-data-generator 无需更新