fix: thinking enabled + tool_use first block#577
Conversation
- Extend thinking signature rectifier to detect tool_use-first error and drop top-level thinking on retry - Add default error rule guidance for thinking/tool_use mismatch - Add unit tests for rectifier, forwarder retry path, and default rule sync
📝 WalkthroughWalkthrough新增并扩展了 thinking signature rectifier 的检测与修复路径，新增 error-rules 条目以识别“必须以 thinking 块开头”类错误，更新相关类型与运行时导出，并添加/调整多处单元测试覆盖这些行为。 Changes
Estimated code review effort🎯 3 (Moderate) | ⏱️ ~25 分钟 Possibly related PRs
🚥 Pre-merge checks | ✅ 2 | ❌ 1❌ Failed checks (1 warning)
✅ Passed checks (2 passed)
✏️ Tip: You can configure your own custom pre-merge checks in the settings. ✨ Finishing touches
🧪 Generate unit tests (beta)
Comment |
Summary of ChangesHello @ding113, I'm Gemini Code Assist1! I'm currently reviewing this pull request and will post my feedback shortly. In the meantime, here's a summary to help you and other reviewers quickly get up to speed! 此拉取请求旨在增强系统对 Claude API 特定错误模式的鲁棒性。它通过引入智能的请求整流逻辑和用户友好的错误提示，解决了在启用 thinking 功能时，如果 assistant 消息的工具使用块未正确前置 thinking 块而导致的 400 错误。这提高了用户体验，并减少了因 API 行为差异而导致的请求失败。 Highlights
Using Gemini Code AssistThe full guide for Gemini Code Assist can be found on our documentation page, here are some quick tips. Invoking Gemini You can request assistance from Gemini at any point by creating a comment using either
Customization To customize Gemini Code Assist for GitHub experience, repository maintainers can create a configuration file and/or provide a custom code review style guide (such as PEP-8 for Python) by creating and adding files to a Limitations & Feedback Gemini Code Assist may make mistakes. Please leave feedback on any instances where its feedback is incorrect or counter productive. You can react with 👍 and 👎 on @gemini-code-assist comments. If you're interested in giving your feedback about your experience with Gemini Code Assist for Github and other Google products, sign up here. You can also get AI-powered code generation, chat, as well as code reviews directly in the IDE at no cost with the Gemini Code Assist IDE Extension. Footnotes
|
🧪 测试结果
总体结果: ✅ 所有测试通过 |
![gemini-code-assist[bot]](https://avatars.githubusercontent.com/in/956858?s=60&v=4){kind=small}
There was a problem hiding this comment.
Code Review
这个 PR 修复了一个在启用 thinking 和 tool_use 时与 Claude 兼容渠道交互的边缘 case。修复方案很全面，包括了两个方面：
- 在代理层通过
thinking-signature-rectifier增加了对错误的识别和自动修复逻辑，即在重试前移除顶层的thinking字段，这能有效避免连续的 400 错误。 - 在
error-rules中增加了一条新的默认规则，当上游返回特定错误时，能给用户一个清晰、可操作的错误提示。
代码变更逻辑清晰，并且附带了非常详尽的单元测试和集成测试，覆盖了新增长的逻辑和相关的边缘场景。特别是 proxy-forwarder-thinking-signature-rectifier.test.ts 中对组合场景的测试，确保了整流器在不同情况下的行为符合预期。
整体来看，这是一个高质量的修复，不仅解决了问题，还提升了代码的可维护性和健壮性。
| if (lastAssistantContent && lastAssistantContent.length > 0) { | ||
| const hasToolUse = lastAssistantContent.some((block) => { | ||
| if (!block || typeof block !== "object") return false; | ||
| return (block as Record<string, unknown>).type === "tool_use"; | ||
| }); | ||
|
|
||
| const firstBlock = lastAssistantContent[0]; | ||
| const firstBlockType = | ||
| firstBlock && typeof firstBlock === "object" | ||
| ? (firstBlock as Record<string, unknown>).type | ||
| : null; | ||
|
|
||
| const missingThinkingPrefix = | ||
| firstBlockType !== "thinking" && firstBlockType !== "redacted_thinking"; | ||
|
|
||
| if (hasToolUse && missingThinkingPrefix && "thinking" in message) { | ||
| delete (message as any).thinking; | ||
| applied = true; | ||
| } | ||
| } |
There was a problem hiding this comment.
为了提高代码的可读性和效率，可以稍微重构这部分逻辑。
当前的实现会无条件地遍历 lastAssistantContent 来检查 hasToolUse，即使 missingThinkingPrefix 为 false。我们可以调整一下顺序，仅在确实缺少 thinking 前缀时，才去检查是否存在 tool_use。
此外，if 条件中的 "thinking" in message 检查是多余的，因为代码块已经处于 if (thinkingEnabled) 的保护之下。
建议修改如下，使逻辑更清晰，并避免不必要的操作。
if (lastAssistantContent && lastAssistantContent.length > 0) {
const firstBlock = lastAssistantContent[0];
const firstBlockType =
firstBlock && typeof firstBlock === "object"
? (firstBlock as Record<string, unknown>).type
: null;
const missingThinkingPrefix =
firstBlockType !== "thinking" && firstBlockType !== "redacted_thinking";
// 仅在缺少 thinking 前缀时才需要进一步检查
if (missingThinkingPrefix) {
const hasToolUse = lastAssistantContent.some((block) => {
if (!block || typeof block !== "object") return false;
return (block as Record<string, unknown>).type === "tool_use";
});
if (hasToolUse) {
delete (message as any).thinking;
applied = true;
}
}
}![greptile-apps[bot]](https://avatars.githubusercontent.com/in/867647?s=60&v=4){kind=small}
![coderabbitai[bot]](https://avatars.githubusercontent.com/in/347564?s=60&v=4){kind=small}
There was a problem hiding this comment.
Actionable comments posted: 2
🤖 Fix all issues with AI agents
In @src/repository/error-rules.ts:
- Around line 660-680: The rule object with pattern "must start with a thinking
block" embeds a hard-coded Chinese string in overrideResponse.message; replace
that literal with an i18n key (or a translations object) and supply localized
strings for zh-CN, en, ja, ko, de, then update any consumer (e.g.,
edit-rule-dialog.tsx) to render the translated text by looking up that key/value
instead of using the hard-coded message; specifically modify the rule entry's
overrideResponse to reference a messageKey (or messages map) and add the
five-language translations so the UI can display the proper localized message.
🧹 Nitpick comments (3)
tests/unit/lib/emit-event.test.ts (1)
38-64: 测试覆盖了 Node.js runtime 下的预期行为。测试验证了：
emitErrorRulesUpdated触发本地事件并广播缓存失效emitSensitiveWordsUpdated仅触发本地事件（无缓存失效）emitRequestFiltersUpdated触发本地事件并广播缓存失效建议在每个测试的动态导入前添加
vi.resetModules()以确保模块状态隔离，避免跨测试污染。♻️ 可选优化：确保模块隔离
test("emitErrorRulesUpdated：Node.js runtime 下应触发本地事件并广播缓存失效", async () => { + vi.resetModules(); const { emitErrorRulesUpdated } = await import("@/lib/emit-event"); await emitErrorRulesUpdated();tests/unit/proxy/proxy-forwarder-thinking-signature-rectifier.test.ts (1)
208-260: 测试用例覆盖了“tool_use 置顶 + thinking 启用”重试路径，建议降低对上游报错文案的耦合度。Line 227 的错误字符串过长且过度依赖完整文案；这里仅需包含能命中检测逻辑的关键片段（例如 “must start with a thinking block” 或 “Expected
thinking... foundtool_use”）即可，避免未来上游文案微调导致用例脆弱、也提升可读性。建议改法（缩短并聚焦触发关键词）
- throw new ProxyError( - "messages.69.content.0.type: Expected `thinking` or `redacted_thinking`, but found `tool_use`. When `thinking` is enabled, a final `assistant` message must start with a thinking block (preceeding the lastmost set of `tool_use` and `tool_result` blocks). To avoid this requirement, disable `thinking`.", + throw new ProxyError( + "Expected `thinking` or `redacted_thinking`, but found `tool_use`; must start with a thinking block", 400, { body: "", providerId: 1, providerName: "anthropic-1", } );src/app/v1/_lib/proxy/thinking-signature-rectifier.ts (1)
18-41: 检测规则建议再“抗变体”一点（但当前实现已可用）。目前同时覆盖关键短语与 “Expected thinking/redacted_thinking ... found tool_use” 正则，方向正确；若线上渠道存在更多变体（例如无 backtick、或
tool_use前后有额外字段路径），建议后续补充 1-2 个最小化样例到单测中以锁定行为。
📜 Review details
Configuration used: Organization UI
Review profile: CHILL
Plan: Pro
Cache: Disabled due to Reviews > Disable Cache setting
📒 Files selected for processing (9)
src/app/v1/_lib/proxy/thinking-signature-rectifier.test.tssrc/app/v1/_lib/proxy/thinking-signature-rectifier.tssrc/instrumentation.tssrc/repository/error-rules.tssrc/types/special-settings.tstests/unit/lib/emit-event.test.tstests/unit/lib/session-manager-helpers.test.tstests/unit/proxy/proxy-forwarder-thinking-signature-rectifier.test.tstests/unit/repository/error-rules-default-thinking-tooluse.test.ts
🧰 Additional context used
📓 Path-based instructions (3)
**/*.{js,ts,tsx,jsx}
📄 CodeRabbit inference engine (CLAUDE.md)
No emoji characters in any code, comments, or string literals
Files:
src/types/special-settings.tssrc/app/v1/_lib/proxy/thinking-signature-rectifier.test.tssrc/app/v1/_lib/proxy/thinking-signature-rectifier.tstests/unit/proxy/proxy-forwarder-thinking-signature-rectifier.test.tssrc/repository/error-rules.tssrc/instrumentation.tstests/unit/lib/emit-event.test.tstests/unit/lib/session-manager-helpers.test.tstests/unit/repository/error-rules-default-thinking-tooluse.test.ts
**/*.{ts,tsx,jsx,js}
📄 CodeRabbit inference engine (CLAUDE.md)
**/*.{ts,tsx,jsx,js}: All user-facing strings must use i18n (5 languages supported: zh-CN, en, ja, ko, de). Never hardcode display text
Use path alias @/ to map to ./src/
Use Biome for code formatting with configuration: double quotes, trailing commas, 2-space indent, 100 character line width
Prefer named exports over default exports
Use next-intl for internationalization
Use Next.js 16 App Router with Hono for API routes
Files:
src/types/special-settings.tssrc/app/v1/_lib/proxy/thinking-signature-rectifier.test.tssrc/app/v1/_lib/proxy/thinking-signature-rectifier.tstests/unit/proxy/proxy-forwarder-thinking-signature-rectifier.test.tssrc/repository/error-rules.tssrc/instrumentation.tstests/unit/lib/emit-event.test.tstests/unit/lib/session-manager-helpers.test.tstests/unit/repository/error-rules-default-thinking-tooluse.test.ts
**/*.test.{ts,tsx,js,jsx}
📄 CodeRabbit inference engine (CLAUDE.md)
Place unit tests in tests/unit/, integration tests in tests/integration/, and source-adjacent tests in src/**/*.test.ts
Files:
src/app/v1/_lib/proxy/thinking-signature-rectifier.test.tstests/unit/proxy/proxy-forwarder-thinking-signature-rectifier.test.tstests/unit/lib/emit-event.test.tstests/unit/lib/session-manager-helpers.test.tstests/unit/repository/error-rules-default-thinking-tooluse.test.ts
🧠 Learnings (3)
📚 Learning: 2026-01-05T03:01:39.354Z
Learnt from: NieiR
Repo: ding113/claude-code-hub PR: 539
File: src/types/user.ts:158-170
Timestamp: 2026-01-05T03:01:39.354Z
Learning: In TypeScript interfaces, explicitly document and enforce distinct meanings for null and undefined. Example: for numeric limits like limitTotalUsd, use 'number | null | undefined' when null signifies explicitly unlimited (e.g., matches DB schema or special UI logic) and undefined signifies 'inherit default'. This pattern should be consistently reflected in type definitions across related fields to preserve semantic clarity between database constraints and UI behavior.
Applied to files:
src/types/special-settings.tssrc/app/v1/_lib/proxy/thinking-signature-rectifier.test.tssrc/app/v1/_lib/proxy/thinking-signature-rectifier.tssrc/repository/error-rules.tssrc/instrumentation.ts
📚 Learning: 2026-01-10T06:19:58.167Z
Learnt from: NieiR
Repo: ding113/claude-code-hub PR: 573
File: src/actions/model-prices.ts:275-335
Timestamp: 2026-01-10T06:19:58.167Z
Learning: In the `ding113/claude-code-hub` repository, Server Actions (files under `src/actions/*.ts`) currently return hardcoded Chinese error messages directly. This is a codebase-wide architectural decision that applies to all action files (e.g., model-prices.ts, users.ts, system-config.ts). Changing this pattern requires a coordinated i18n refactor across all Server Actions, which should be handled as a separate unified task rather than piecemeal changes in individual PRs.
Applied to files:
src/repository/error-rules.ts
📚 Learning: 2026-01-07T17:05:36.362Z
Learnt from: CR
Repo: ding113/claude-code-hub PR: 0
File: CLAUDE.md:0-0
Timestamp: 2026-01-07T17:05:36.362Z
Learning: Applies to **/*.test.{ts,tsx,js,jsx} : Place unit tests in tests/unit/, integration tests in tests/integration/, and source-adjacent tests in src/**/*.test.ts
Applied to files:
tests/unit/lib/emit-event.test.tstests/unit/repository/error-rules-default-thinking-tooluse.test.ts
🧬 Code graph analysis (4)
src/app/v1/_lib/proxy/thinking-signature-rectifier.test.ts (1)
src/app/v1/_lib/proxy/thinking-signature-rectifier.ts (2)
detectThinkingSignatureRectifierTrigger(18-59)rectifyAnthropicRequestMessage(70-185)
src/instrumentation.ts (10)
src/app/api/actions/[...route]/route.ts (1)
runtime(45-45)src/app/v1/[...route]/route.ts (1)
runtime(16-16)src/app/api/leaderboard/route.ts (1)
runtime(31-31)src/app/api/version/route.ts (1)
runtime(7-7)src/app/api/proxy-status/route.ts (1)
runtime(7-7)src/app/api/admin/database/export/route.ts (1)
runtime(7-7)src/app/api/admin/database/import/route.ts (1)
runtime(13-13)src/app/api/auth/login/route.ts (1)
runtime(6-6)src/app/api/admin/log-level/route.ts (1)
runtime(5-5)src/app/api/admin/log-cleanup/manual/route.ts (1)
runtime(8-8)
tests/unit/lib/session-manager-helpers.test.ts (1)
src/lib/session-manager.ts (1)
parseHeaderRecord(2010-2010)
tests/unit/repository/error-rules-default-thinking-tooluse.test.ts (1)
src/repository/error-rules.ts (1)
syncDefaultErrorRules(816-896)
⏰ Context from checks skipped due to timeout of 90000ms. You can increase the timeout in your CodeRabbit configuration to a maximum of 15 minutes (900000ms). (6)
- GitHub Check: 🌐 API Tests
- GitHub Check: pr-review
- GitHub Check: pr-description
- GitHub Check: Docker Build Test
- GitHub Check: check-codex-status
- GitHub Check: dev-build-deploy
🔇 Additional comments (12)
src/instrumentation.ts (2)
6-8: LGTM!
runtime = "nodejs"导出与代码库中其他路由文件保持一致（如src/app/v1/[...route]/route.ts、src/app/api/actions/[...route]/route.ts等）。这确保了 instrumentation hook 在正确的运行时环境中执行。
82-89: LGTM!将
closeRedis改为懒加载动态导入是合理的优化。这与文件中其他 shutdown handler 内的动态导入模式一致（如stopMessageRequestWriteBuffer），避免了在 shutdown 阶段可能出现的模块解析问题。tests/unit/lib/session-manager-helpers.test.ts (2)
4-8: LGTM! 测试重构提升了可维护性。引入
PARSE_HEADER_RECORD_WARN_MESSAGE常量和getParseHeaderRecordWarnCalls()辅助函数是良好的实践：
- 避免魔法字符串，确保一致性
- 过滤器模式精确匹配目标警告，减少其他 warn 调用的干扰
- 提高测试的可读性和可维护性
67-72: 断言逻辑清晰。使用
calls[0] ?? []进行安全解构是良好的防御性编程，避免空数组时的 undefined 错误。tests/unit/repository/error-rules-default-thinking-tooluse.test.ts (2)
68-90: LGTM! 测试覆盖了新规则的关键属性。测试设计合理：
- 验证 pattern、matchType、category 等核心字段
- 检查
overrideResponse包含可操作的用户提示（tool_result、signature、关闭）- Mock 设置模拟了空数据库场景，正确测试插入逻辑
测试文件位置符合编码规范（
tests/unit/）。
21-53: Mock 设置完整且符合syncDefaultErrorRules的事务流程。通过
capturedInsertedRules.push(rule)捕获插入的规则，便于后续断言验证。returning: vi.fn(async () => [{ id: 1 }])正确模拟了成功插入的响应。tests/unit/lib/emit-event.test.ts (1)
66-80: Edge runtime 测试验证了静默跳过行为。测试确保在 Edge runtime 下所有 emit 函数都不触发任何事件或广播，符合预期的运行时条件逻辑。
src/app/v1/_lib/proxy/thinking-signature-rectifier.test.ts (2)
24-29: LGTM! 新增的触发器检测测试。测试用例使用了真实的 Claude API 错误消息格式，验证了
detectThinkingSignatureRectifierTrigger能正确识别 "assistant_message_must_start_with_thinking" 场景。
95-124: LGTM! 关键场景测试覆盖。测试验证了 PR 目标中描述的核心场景：
thinking已启用- 最后一条 assistant 消息以
tool_use开头（缺少 thinking/redacted_thinking 前缀）- 整流器应删除顶层
thinking字段这确保了在重试前正确处理请求，避免继续触发 400 错误。
src/types/special-settings.ts (1)
103-106: LGTM! 类型扩展正确。新增的
"assistant_message_must_start_with_thinking"触发器类型与以下实现保持一致：
detectThinkingSignatureRectifierTrigger函数的返回值- 相关测试用例的断言
- 新增的默认错误规则
类型安全得到保证，命名风格与现有触发器一致。
tests/unit/proxy/proxy-forwarder-thinking-signature-rectifier.test.ts (1)
262-316: “先移除 thinking block，再因 tool_use 置顶而关闭顶层 thinking” 的链式兜底验证到位。重试时同时断言 blocks 清理结果（Line 293-295）和
bodyMsg.thinking被删除（Line 295）很关键；整体用例设计符合该 PR 的修复目标。src/app/v1/_lib/proxy/thinking-signature-rectifier.ts (1)
1-5: 新增 trigger 类型扩展合理。
ThinkingSignatureRectifierTrigger增加"assistant_message_must_start_with_thinking"能更精确区分该类 400 场景。
![github-actions[bot]](https://avatars.githubusercontent.com/in/15368?s=60&v=4){kind=small}
There was a problem hiding this comment.
Code Review Summary
No significant issues identified in this PR. The implementation correctly addresses the specific error scenario where Claude-compatible channels return a 400 error when thinking is enabled but the last assistant message's content starts with tool_use instead of thinking/redacted_thinking.
PR Size: M
- Lines changed: 431
- Files changed: 9
Review Coverage
- Logic and correctness - Clean
- Security (OWASP Top 10) - Clean
- Error handling - Clean
- Type safety - Clean
- Documentation accuracy - Clean
- Test coverage - Adequate
- Code clarity - Good
Notable Observations
Core Changes Validated:
thinking-signature-rectifier.ts- New trigger typeassistant_message_must_start_with_thinkingproperly detects the error pattern and the rectification logic correctly removes the top-levelthinkingfield when tool_use is the first blockinstrumentation.ts- Runtime declaration and dynamic import forcloseRedisproperly addresses Edge runtime compatibilityerror-rules.ts- New default error rule provides actionable guidance with appropriate priorityspecial-settings.ts- Type union correctly extended
Test Coverage:
- Unit tests cover trigger detection with actual error messages
- Integration tests validate retry behavior and cascading fix scenarios
- New emit-event tests properly validate runtime-specific behavior
Automated review by Claude AI
|
|
||
| afterEach(() => { | ||
| process.env.NEXT_RUNTIME = prevRuntime; | ||
| }); |
There was a problem hiding this comment.
[High] [TEST-BRITTLE] process.env.NEXT_RUNTIME restore can leak 'undefined' across the suite
Why this is a problem: Node coerces process.env assignments to strings. If prevRuntime is undefined, process.env.NEXT_RUNTIME = prevRuntime sets the literal string "undefined", leaving the var truthy and not "nodejs". That can change runtime-guarded behavior (e.g. process.env.NEXT_RUNTIME && process.env.NEXT_RUNTIME !== "nodejs") and cause cross-test contamination/flakes.
Suggested fix:
afterEach(() => {
if (prevRuntime === undefined) {
delete process.env.NEXT_RUNTIME;
} else {
process.env.NEXT_RUNTIME = prevRuntime;
}
});
There was a problem hiding this comment.
Code Review Summary
PR adds a new thinking-signature-rectifier trigger for Anthropic 400s when tool_use precedes thinking, plus a default error rule and related tests. One high-priority test isolation issue can leak NEXT_RUNTIME across the test suite and cause flaky behavior.
PR Size: M
- Lines changed: 431
- Files changed: 9
Issues Found
| Category | Critical | High | Medium | Low |
|---|---|---|---|---|
| Logic/Bugs | 0 | 0 | 0 | 0 |
| Security | 0 | 0 | 0 | 0 |
| Error Handling | 0 | 0 | 0 | 0 |
| Types | 0 | 0 | 0 | 0 |
| Comments/Docs | 0 | 0 | 0 | 0 |
| Tests | 0 | 1 | 0 | 0 |
| Simplification | 0 | 0 | 0 | 0 |
Critical Issues (Must Fix)
None.
High Priority Issues (Should Fix)
- [TEST-BRITTLE] process.env.NEXT_RUNTIME restore can leak 'undefined' across the suite (tests/unit/lib/emit-event.test.ts:35, confidence 90)
Review Coverage
- Logic and correctness
- Security (OWASP Top 10)
- Error handling
- Type safety
- Documentation accuracy
- Test coverage
- Code clarity
Automated review by Codex AI
There was a problem hiding this comment.
- Identified PR
#577and applied labelsize/M. - Posted 1 inline review comment:
tests/unit/lib/emit-event.test.ts:35—[High] [TEST-BRITTLE]restoringprocess.env.NEXT_RUNTIMEcan leave the literal string"undefined"and leak env state across tests; suggested usingdelete process.env.NEXT_RUNTIMEwhen the previous value isundefined. (Discussion:https://github.com/ding113/claude-code-hub/pull/577#discussion_r2678499549) - Submitted the required summary review (Review:
https://github.com/ding113/claude-code-hub/pull/577#pullrequestreview-3646663258).
🧪 测试结果
总体结果: ✅ 所有测试通过 |
There was a problem hiding this comment.
Actionable comments posted: 1
🧹 Nitpick comments (1)
src/app/v1/_lib/proxy/thinking-signature-rectifier.ts (1)
135-180: 逻辑正确，建议考虑在结果中追踪 thinking 字段的移除兜底逻辑处理了各种边界情况（空 content、无 assistant 消息、多条 assistant 消息），实现符合预期。
不过当前
ThinkingSignatureRectifierResult只追踪了removedThinkingBlocks、removedRedactedThinkingBlocks、removedSignatureFields，未记录顶层thinking字段被移除的情况。如果审计或调试时需要区分「移除了 thinking block」和「禁用了顶层 thinking」，可考虑在返回结构中新增一个字段（如disabledTopLevelThinking: boolean）。♻️ 可选：扩展返回结构
export type ThinkingSignatureRectifierResult = { applied: boolean; removedThinkingBlocks: number; removedRedactedThinkingBlocks: number; removedSignatureFields: number; + disabledTopLevelThinking: boolean; };并在函数中追踪：
+ let disabledTopLevelThinking = false; // ... if (hasToolUse) { delete (message as any).thinking; applied = true; + disabledTopLevelThinking = true; } // ... return { applied, removedThinkingBlocks, removedRedactedThinkingBlocks, removedSignatureFields, + disabledTopLevelThinking, };
📜 Review details
Configuration used: Organization UI
Review profile: CHILL
Plan: Pro
Cache: Disabled due to Reviews > Disable Cache setting
📒 Files selected for processing (3)
src/app/v1/_lib/proxy/thinking-signature-rectifier.tssrc/repository/error-rules.tstests/unit/repository/error-rules-default-thinking-tooluse.test.ts
🚧 Files skipped from review as they are similar to previous changes (1)
- src/repository/error-rules.ts
🧰 Additional context used
📓 Path-based instructions (3)
**/*.{js,ts,tsx,jsx}
📄 CodeRabbit inference engine (CLAUDE.md)
No emoji characters in any code, comments, or string literals
Files:
tests/unit/repository/error-rules-default-thinking-tooluse.test.tssrc/app/v1/_lib/proxy/thinking-signature-rectifier.ts
**/*.{ts,tsx,jsx,js}
📄 CodeRabbit inference engine (CLAUDE.md)
**/*.{ts,tsx,jsx,js}: All user-facing strings must use i18n (5 languages supported: zh-CN, en, ja, ko, de). Never hardcode display text
Use path alias @/ to map to ./src/
Use Biome for code formatting with configuration: double quotes, trailing commas, 2-space indent, 100 character line width
Prefer named exports over default exports
Use next-intl for internationalization
Use Next.js 16 App Router with Hono for API routes
Files:
tests/unit/repository/error-rules-default-thinking-tooluse.test.tssrc/app/v1/_lib/proxy/thinking-signature-rectifier.ts
**/*.test.{ts,tsx,js,jsx}
📄 CodeRabbit inference engine (CLAUDE.md)
Place unit tests in tests/unit/, integration tests in tests/integration/, and source-adjacent tests in src/**/*.test.ts
Files:
tests/unit/repository/error-rules-default-thinking-tooluse.test.ts
🧠 Learnings (3)
📓 Common learnings
Learnt from: NieiR
Repo: ding113/claude-code-hub PR: 573
File: src/actions/model-prices.ts:275-335
Timestamp: 2026-01-10T06:20:04.478Z
Learning: In the `ding113/claude-code-hub` repository, Server Actions (files under `src/actions/*.ts`) currently return hardcoded Chinese error messages directly. This is a codebase-wide architectural decision that applies to all action files (e.g., model-prices.ts, users.ts, system-config.ts). Changing this pattern requires a coordinated i18n refactor across all Server Actions, which should be handled as a separate unified task rather than piecemeal changes in individual PRs.
📚 Learning: 2026-01-07T17:05:37.171Z
Learnt from: CR
Repo: ding113/claude-code-hub PR: 0
File: CLAUDE.md:0-0
Timestamp: 2026-01-07T17:05:37.171Z
Learning: Applies to **/*.test.{ts,tsx,js,jsx} : Place unit tests in tests/unit/, integration tests in tests/integration/, and source-adjacent tests in src/**/*.test.ts
Applied to files:
tests/unit/repository/error-rules-default-thinking-tooluse.test.ts
📚 Learning: 2026-01-05T03:01:39.354Z
Learnt from: NieiR
Repo: ding113/claude-code-hub PR: 539
File: src/types/user.ts:158-170
Timestamp: 2026-01-05T03:01:39.354Z
Learning: In TypeScript interfaces, explicitly document and enforce distinct meanings for null and undefined. Example: for numeric limits like limitTotalUsd, use 'number | null | undefined' when null signifies explicitly unlimited (e.g., matches DB schema or special UI logic) and undefined signifies 'inherit default'. This pattern should be consistently reflected in type definitions across related fields to preserve semantic clarity between database constraints and UI behavior.
Applied to files:
src/app/v1/_lib/proxy/thinking-signature-rectifier.ts
🧬 Code graph analysis (1)
tests/unit/repository/error-rules-default-thinking-tooluse.test.ts (1)
src/repository/error-rules.ts (1)
syncDefaultErrorRules(837-917)
⏰ Context from checks skipped due to timeout of 90000ms. You can increase the timeout in your CodeRabbit configuration to a maximum of 15 minutes (900000ms). (3)
- GitHub Check: 🌐 API Tests
- GitHub Check: Docker Build Test
- GitHub Check: dev-build-deploy
🔇 Additional comments (3)
src/app/v1/_lib/proxy/thinking-signature-rectifier.ts (2)
1-4: LGTM!新增的触发类型
assistant_message_must_start_with_thinking命名清晰，与现有类型风格一致。
25-40: LGTM!检测逻辑设计合理：
- 正则兼容带/不带反引号的多种错误消息变体
- 检测顺序正确（更具体的条件优先匹配）
- 注释清晰解释了触发场景和修复策略
tests/unit/repository/error-rules-default-thinking-tooluse.test.ts (1)
1-75: 测试设置结构合理。Mock 策略完善，通过
capturedInsertedRules捕获插入的规则对象，环境变量隔离避免了与全局测试设置的冲突，所有 mock 返回值结构正确。
| // 覆写响应需为 Claude 错误格式，且包含清晰的自助修复建议 | ||
| expect(rule.overrideResponse?.type).toBe("error"); | ||
| expect(rule.overrideResponse?.error?.type).toBe("thinking_error"); | ||
| expect(rule.overrideResponse?.error?.message).toContain("tool_result"); | ||
| expect(rule.overrideResponse?.error?.message).toContain("signature"); | ||
| expect(rule.overrideResponse?.error?.message).toContain("关闭"); |
There was a problem hiding this comment.
🧩 Analysis chain
🏁 Script executed:
# First, let's check if the test file exists and examine the context around lines 93-98
wc -l tests/unit/repository/error-rules-default-thinking-tooluse.test.tsRepository: ding113/claude-code-hub
Length of output: 137
🏁 Script executed:
# Let's look at the specific test case and surrounding context
sed -n '70,105p' tests/unit/repository/error-rules-default-thinking-tooluse.test.tsRepository: ding113/claude-code-hub
Length of output: 1350
🏁 Script executed:
# Now let's check the source error-rules.ts file for the pattern "must start with a thinking block"
rg -A 15 -B 2 'must start with a thinking block' src/repository/Repository: ding113/claude-code-hub
Length of output: 1247
错误消息应使用国际化，作为整体 i18n 重构计划的一部分。
错误规则的 overrideResponse 消息包含硬编码的中文（如"关闭"），违反了编码规范要求"所有面向用户的字符串必须使用 i18n"。虽然当前代码库中存在硬编码中文错误消息的模式，但这应作为统一的国际化重构任务处理，而不是分散在各个 PR 中单独修改。建议：
- 将此问题标记为技术债务，跟踪于统一的 i18n 重构计划中
- 该重构应覆盖 src/repository/ 中的所有用户面向错误消息，包含所有 5 种语言支持（zh-CN, en, ja, ko, de）
github-project-automation
bot
moved this from Backlog
to Done
in Claude Code Hub Roadmap
- PR ding113#580: TOML cloud price table + billing fail-open - PR ding113#578: make drizzle migrations idempotent - PR ding113#577: fix thinking enabled + tool_use first block - PR ding113#573: add manual model price management Conflict resolutions: - drizzle migrations: use upstream idempotent version - i18n messages: accept upstream additions - price-sync.ts: removed (replaced by cloud-price-table) - model-prices.ts: use upstream refactored version
1 participant
问题：当 /v1/messages 请求顶层 thinking 启用，但最后一条 assistant message 的 content 以 tool_use 开头（缺少 thinking/redacted_thinking 前缀）时，部分 Claude 兼容渠道会返回 400：Expected thinking/redacted_thinking but found tool_use。
修复：
验证：
Greptile Overview
Greptile Summary
Overview
This PR fixes a specific error scenario where Claude-compatible channels return a 400 error when thinking is enabled at the top level but the last assistant message's content starts with
tool_useinstead of athinking/redacted_thinkingblock.Key Changes
Core Fix (thinking-signature-rectifier.ts)
assistant_message_must_start_with_thinkingto detect the specific error patternthinkingfield when:Error Rule (error-rules.ts)
Runtime Fix (instrumentation.ts)
runtime = "nodejs"to ensure Node.js capabilities are availablecloseRedisimport to dynamic import in shutdown handler to prevent Edge runtime conflictsCascading Fix Behavior
The rectification is designed to handle multiple related issues in one pass:
This cascading behavior is intentional and well-tested (see test "移除 thinking block 后若 tool_use 置顶且 thinking 仍启用，应同时关闭 thinking 再重试")
Test Coverage
Architecture
The fix maintains the existing retry-on-same-provider pattern without polluting the circuit breaker or triggering provider switches. The rectification is applied in-place to the request message before retry, and audit trails are properly recorded in specialSettings.
Confidence Score: 5/5
Important Files Changed
File Analysis
Sequence Diagram
sequenceDiagram participant Client participant ProxyForwarder participant ThinkingRectifier participant AnthropicProvider participant ErrorRules Client->>ProxyForwarder: POST /v1/messages<br/>{thinking: enabled, messages: [assistant with tool_use]} ProxyForwarder->>AnthropicProvider: Forward request AnthropicProvider-->>ProxyForwarder: 400: Expected thinking/redacted_thinking,<br/>but found tool_use ProxyForwarder->>ThinkingRectifier: detectThinkingSignatureRectifierTrigger(errorMsg) ThinkingRectifier-->>ProxyForwarder: "assistant_message_must_start_with_thinking" ProxyForwarder->>ProxyForwarder: Check enableThinkingSignatureRectifier ProxyForwarder->>ThinkingRectifier: rectifyAnthropicRequestMessage(message) Note over ThinkingRectifier: Step 1: Remove thinking/redacted_thinking blocks<br/>and signature fields from message content Note over ThinkingRectifier: Step 2: Check if last assistant has tool_use<br/>and first block is not thinking ThinkingRectifier->>ThinkingRectifier: Delete top-level thinking field ThinkingRectifier-->>ProxyForwarder: {applied: true, removedThinkingBlocks: X} ProxyForwarder->>ProxyForwarder: Add specialSettings audit trail ProxyForwarder->>ProxyForwarder: Update message_request_details ProxyForwarder->>AnthropicProvider: Retry with rectified request<br/>(no top-level thinking) AnthropicProvider-->>ProxyForwarder: 200 OK {content: [...]} ProxyForwarder-->>Client: Success response alt If retry also fails AnthropicProvider-->>ProxyForwarder: Still 400 ProxyForwarder->>ErrorRules: Match error pattern ErrorRules-->>Client: Override with friendly error message end