新增错误规则：messages.4.content.0: unexpected `tool_use_id` found in `tool_result` blocks

[ding113]

错误响应示例：

Provider cubence returned 500: Provider returned 500: 500 | Upstream: {"error":{"type":"request_fail","upstream_error":{"error":{"message":"messages.4.content.0: unexpected `tool_use_id` found in `tool_result` blocks: toolu_01GeLJZMgXCwn6LRTHbGBbCa. Each `tool_result` block must have a corresponding `tool_use` block in the previous message.","type":"invalid_request_error"},"request_id":"req_011CWXTTNqQoG2JZJqgjf9pE","type":"error"}},"request_id":"req_011CWXTTNqQoG2JZJqgjf9pE","type":"error"}

[ding113]

@claude fix it

[ding113]

I've created a fix for this issue.

Pull Request: #473

Changes made:

• Added a new default error rule to handle Claude API validation errors where tool_use_id field is incorrectly placed in tool_result blocks
• The regex pattern matches variations including different message indices and quote styles
• Classified as validation_error (non-retryable client error)
• Priority: 89 (same as other validation errors)
• Provides user-friendly Chinese error message

Please review and let me know if you need any adjustments.

---

Response from Claude AI

[github-actions]

Thank you for this feature request.

Analysis:
After investigating the codebase, I found that this error pattern is already covered by an existing default error rule.

Existing error rule:

• Location: src/repository/error-rules.ts (lines 476-491)
• Pattern: unexpected.*tool_use_id.*tool_result|tool_result.*must have.*corresponding.*tool_use
• Match Type: regex
• Category: validation_error
• Priority: 88
• Description: "tool_result block missing corresponding tool_use (client error)"
• Override Response:

  {
    "type": "error",
    "error": {
      "type": "validation_error",
      "message": "tool_result 缺少对应的 tool_use，请检查工具调用链"
    }
  }

Verification:
I verified that the existing regex pattern matches your error message:

• Your error: messages.4.content.0: unexpected \tool_use_id` found in `tool_result` blocks`
• Pattern test result: ✅ Matches successfully

How it works:

1. When the proxy receives a 500 error from the provider with this message
2. ErrorRuleDetector (src/lib/error-rule-detector.ts) detects the pattern match
3. The system categorizes it as validation_error (non-retryable client error)
4. Returns a user-friendly Chinese error message

Notes:

• This rule is enabled by default (isEnabled: true)
• It is marked as a default rule (isDefault: true)
• It will be automatically synced to the database on startup via syncDefaultErrorRules() in src/repository/error-rules.ts:671

If you're still seeing the raw error message instead of the friendly one, please verify:

1. The error rules have been synced to the database (check the error_rules table)
2. The ErrorRuleDetector cache has been loaded (check startup logs for [ErrorRuleDetector] Loaded X error rules)

---

Automated response from Claude AI

[github-actions]

感谢反馈。

从当前仓库代码来看，这个报错场景已经有对应的默认错误规则：

• src/repository/error-rules.ts (lines 475-490) 默认规则包含 regex：unexpected.*tool_use_id.*tool_result|tool_result.*must have.*corresponding.*tool_use，并归类为 validation_error，同时提供了更友好的 overrideResponse 提示。

该规则在运行时会影响两处行为：

1. 判定为不可重试的客户端错误：错误分类会先做规则匹配，再决定是否按 ProxyError 走供应商故障路径

• src/app/v1/_lib/proxy/errors.ts (lines 778-797) categorizeErrorAsync() 在判断 ProxyError 前先调用 isNonRetryableClientErrorAsync()
• src/app/v1/_lib/proxy/forwarder.ts (lines 448-507) 命中 NON_RETRYABLE_CLIENT_ERROR 后会立即停止重试/停止切换供应商并抛出该错误（代码注释也说明这类白名单错误不会调用 recordFailure()）

2. 返回给客户端时可覆写响应：如果规则配置了 overrideResponse/overrideStatusCode，会在错误响应阶段应用

• src/app/v1/_lib/proxy/error-handler.ts (lines 97-216) 通过 getErrorOverrideAsync() 应用覆写并返回 JSON

另外，你贴的 Provider ... | Upstream: ... 这种拼接格式是“日志用”的详细错误信息：

• src/app/v1/_lib/proxy/errors.ts (lines 417-437) getDetailedErrorMessage() 负责拼接 provider + upstream body
• src/app/v1/_lib/proxy/error-handler.ts (lines 66-72) 会把它作为 logErrorMessage 记录，但对客户端使用 getClientSafeMessage()（src/app/v1/_lib/proxy/errors.ts lines 447-449）

如果你这边仍然希望“新增/补齐”规则，想先确认下实际情况：

• 你看到的示例是服务端日志/数据库记录，还是客户端收到的 HTTP 响应体？
• 你的部署里 DEFAULT_ERROR_RULES 是否已同步到数据库（启动初始化会尝试同步并 reload：src/instrumentation.ts lines 27-38, 118-127）？

---

Automated response from Codex AI