错误处理挂钩

当会话执行期间发生错误时调用 onErrorOccurred 钩子函数。 使用它可执行以下操作：

在本文中

  • 实现自定义错误日志记录
  • 跟踪错误模式
  • 提供用户友好的错误消息
  • 触发关键错误的警报

挂钩签名

代码语言 navigation

TypeScript
import type { ErrorOccurredHookInput, HookInvocation, ErrorOccurredHookOutput } from "@github/copilot-sdk";
type ErrorOccurredHandler = (
  input: ErrorOccurredHookInput,
  invocation: HookInvocation
) => Promise<ErrorOccurredHookOutput | null | undefined>;

type ErrorOccurredHandler = (
  input: ErrorOccurredHookInput,
  invocation: HookInvocation
) => Promise<ErrorOccurredHookOutput | null | undefined>;

输入

领域类型Description
timestampnumber发生错误的 Unix 时间戳
cwd字符串当前工作目录
error字符串错误消息
errorContext字符串发生错误的位置："model_call"、、"tool_execution"``"system"或"user_input"
recoverableboolean是否可以可能从错误中恢复

输出

返回 null 或 undefined 以使用默认错误处理。 否则，返回一个对象，其中包含：

领域类型Description
suppressOutputboolean如果为 true，则不向用户显示错误输出
errorHandling字符串如何处理："retry"、"skip" 或 "abort"
retryCountnumber重试次数（如果 errorHandling 为 "retry"）
userNotification字符串显示给用户的自定义消息

示例

基本错误日志记录

代码语言 navigation

TypeScript
const session = await client.createSession({
  hooks: {
    onErrorOccurred: async (input, invocation) => {
      console.error(`[${invocation.sessionId}] Error: ${input.error}`);
      console.error(`  Context: ${input.errorContext}`);
      console.error(`  Recoverable: ${input.recoverable}`);
      return null;
    },
  },
});

将错误发送到监控服务

import { captureException } from "@sentry/node"; // or your monitoring service

const session = await client.createSession({
  hooks: {
    onErrorOccurred: async (input, invocation) => {
      captureException(new Error(input.error), {
        tags: {
          sessionId: invocation.sessionId,
          errorContext: input.errorContext,
        },
        extra: {
          error: input.error,
          recoverable: input.recoverable,
          cwd: input.cwd,
        },
      });
      
      return null;
    },
  },
});

用户友好型错误消息

const ERROR_MESSAGES: Record<string, string> = {
  "model_call": "There was an issue communicating with the AI model. Please try again.",
  "tool_execution": "A tool failed to execute. Please check your inputs and try again.",
  "system": "A system error occurred. Please try again later.",
  "user_input": "There was an issue with your input. Please check and try again.",
};

const session = await client.createSession({
  hooks: {
    onErrorOccurred: async (input) => {
      const friendlyMessage = ERROR_MESSAGES[input.errorContext];
      
      if (friendlyMessage) {
        return {
          userNotification: friendlyMessage,
        };
      }
      
      return null;
    },
  },
});

抑制非关键错误

const session = await client.createSession({
  hooks: {
    onErrorOccurred: async (input) => {
      // Suppress tool execution errors that are recoverable
      if (input.errorContext === "tool_execution" && input.recoverable) {
        console.log(`Suppressed recoverable error: ${input.error}`);
        return { suppressOutput: true };
      }
      return null;
    },
  },
});

添加恢复上下文

const session = await client.createSession({
  hooks: {
    onErrorOccurred: async (input) => {
      if (input.errorContext === "tool_execution") {
        return {
          userNotification: `
The tool failed. Here are some recovery suggestions:
- Check if required dependencies are installed
- Verify file paths are correct
- Try a simpler approach
          `.trim(),
        };
      }
      
      if (input.errorContext === "model_call" && input.error.includes("rate")) {
        return {
          errorHandling: "retry",
          retryCount: 3,
          userNotification: "Rate limit hit. Retrying...",
        };
      }
      
      return null;
    },
  },
});

跟踪错误模式

interface ErrorStats {
  count: number;
  lastOccurred: number;
  contexts: string[];
}

const errorStats = new Map<string, ErrorStats>();

const session = await client.createSession({
  hooks: {
    onErrorOccurred: async (input, invocation) => {
      const key = `${input.errorContext}:${input.error.substring(0, 50)}`;
      
      const existing = errorStats.get(key) || {
        count: 0,
        lastOccurred: 0,
        contexts: [],
      };
      
      existing.count++;
      existing.lastOccurred = input.timestamp;
      existing.contexts.push(invocation.sessionId);
      
      errorStats.set(key, existing);
      
      // Alert if error is recurring
      if (existing.count >= 5) {
        console.warn(`Recurring error detected: ${key} (${existing.count} times)`);
      }
      
      return null;
    },
  },
});

针对严重错误发出警报

const CRITICAL_CONTEXTS = ["system", "model_call"];

const session = await client.createSession({
  hooks: {
    onErrorOccurred: async (input, invocation) => {
      if (CRITICAL_CONTEXTS.includes(input.errorContext) && !input.recoverable) {
        await sendAlert({
          level: "critical",
          message: `Critical error in session ${invocation.sessionId}`,
          error: input.error,
          context: input.errorContext,
          timestamp: new Date(input.timestamp).toISOString(),
        });
      }
      
      return null;
    },
  },
});

与其他上下文钩子结合使用

const sessionContext = new Map<string, { lastTool?: string; lastPrompt?: string }>();

const session = await client.createSession({
  hooks: {
    onPreToolUse: async (input, invocation) => {
      const ctx = sessionContext.get(invocation.sessionId) || {};
      ctx.lastTool = input.toolName;
      sessionContext.set(invocation.sessionId, ctx);
      return { permissionDecision: "allow" };
    },
    
    onUserPromptSubmitted: async (input, invocation) => {
      const ctx = sessionContext.get(invocation.sessionId) || {};
      ctx.lastPrompt = input.prompt.substring(0, 100);
      sessionContext.set(invocation.sessionId, ctx);
      return null;
    },
    
    onErrorOccurred: async (input, invocation) => {
      const ctx = sessionContext.get(invocation.sessionId);
      
      console.error(`Error in session ${invocation.sessionId}:`);
      console.error(`  Error: ${input.error}`);
      console.error(`  Context: ${input.errorContext}`);
      if (ctx?.lastTool) {
        console.error(`  Last tool: ${ctx.lastTool}`);
      }
      if (ctx?.lastPrompt) {
        console.error(`  Last prompt: ${ctx.lastPrompt}...`);
      }
      
      return null;
    },
  },
});

最佳做法

  1. 始终记录错误 - 即使禁止用户显示错误，也要保留日志以供调试。

  2. 对错误进行分类 - 用于 errorType 适当处理不同的错误。

  3.           **切勿屏蔽严重错误**：仅可忽略你确信不严重的错误。

  4.           **保证挂钩高效运行** - 错误处理不应减慢恢复速度。

  5. 提供有用的上下文 - 发生错误时， additionalContext 可以帮助模型恢复。

  6. 监视错误模式 - 跟踪定期错误以识别系统性问题。

另见