agent-debugger
Compare original and translation side by side
🇺🇸
Original
English🇨🇳
Translation
ChineseAgent Debugger
Agent Debugger
Skill for diagnosing and resolving failures in AI agents, tool calls and pipelines.
用于诊断和解决AI Agent、工具调用及流程故障的技能。
When to use this skill
何时使用此技能
- The agent enters a loop and never finishes
- A tool call fails silently or returns an unexpected result
- The agent does something different from what was requested
- There is a parsing error in the model response
- The context fills up and the agent loses information
- Two subagents conflict or overwrite each other
- Agent进入循环且永远无法完成任务
- 工具调用静默失败或返回意外结果
- Agent的行为与请求不符
- 模型响应存在解析错误
- 上下文已满导致Agent丢失信息
- 两个子Agent发生冲突或互相覆盖
Patterns and solutions
常见模式与解决方案
Infinite loop — add an explicit stop condition, limit iterations
Malformed tool call — validate schema before executing, add examples in the prompt
Context overflow — summarize history every N steps, truncate tool outputs to the minimum needed
Parsing error — use structured outputs if supported, add validation with retry logic
Prompt drift — re-inject the original objective every N steps as a system reminder
Silent subagent conflict — assign exclusive resources to each subagent, implement locks
无限循环 — 添加明确的终止条件，限制迭代次数
格式错误的工具调用 — 执行前验证schema，在提示词中添加示例
上下文溢出 — 每N步总结历史记录，将工具输出截断至所需的最小内容
解析错误 — 若支持则使用结构化输出，添加带重试逻辑的验证
提示词偏移 — 每N步重新注入原始目标作为系统提醒
子Agent静默冲突 — 为每个子Agent分配专属资源，实现锁机制
Verification checklist
验证清单
- Agent completes the task without loops
- All tool calls match the expected schema
- Context stays below 80% of the limit
- Model responses parse correctly
- Behavior is consistent across runs
- Subagents do not overwrite each other
- Agent无循环地完成任务
- 所有工具调用符合预期schema
- 上下文占用率保持在限制的80%以下
- 模型响应可正确解析
- 多次运行的行为保持一致
- 子Agent不会互相覆盖
Compatibility
兼容性
Claude Code, Cursor, Windsurf, Cline, Roo and any agent compatible with the Agent Skills standard.
Claude Code、Cursor、Windsurf、Cline、Roo以及任何兼容Agent Skills标准的Agent。