fix: Correctly parse Gemini reasoning for both new thought field an…
#801
Conversation
…d legacy `<think>` tags.
WalkthroughIntroduces dual-mode streaming for Gemini responses with detection of a new thought format, updates buffering and end-of-stream behavior accordingly, adds normalized response processing, expands message formatting for tool calls and multi-part content, integrates safety settings mapping, supports thinking budgets for eligible models, and maintains token usage reporting. Changes
Sequence Diagram(s)sequenceDiagram
autonumber
participant UI as UI
participant Provider as GeminiProvider
participant Gemini as Gemini API
UI->>Provider: streamRequest(messages, config)
Provider->>Provider: formatGeminiMessages(messages)
Provider->>Provider: getGenerationConfig(+thinkingBudget?, +thinkingConfig)
Provider->>Provider: getFormattedSafetySettings()
Provider->>Gemini: startStream(formattedMessages, genConfig, safety)
loop on chunk
Gemini-->>Provider: chunk(parts[])
alt new thought format detected (part.thought === true)
Provider->>Provider: isNewThoughtFormatDetected = true
Provider-->>UI: emit text(part.text) in real time
Provider->>Provider: accumulate thought segments
else old format (uses <think> buffering)
Provider->>Provider: buffer until </think>
Provider-->>UI: emit reasoning_content then content
end
end
Provider->>Provider: finalize
alt new format
Provider->>Provider: suppress leftover buffered content
else old format
Provider-->>UI: emit leftover buffered content
end
Provider->>Provider: processGeminiResponse(collect text, thoughts, usage)
Provider-->>UI: complete({content, reasoning_content, usage})
sequenceDiagram
autonumber
participant Caller as Caller
participant Provider as GeminiProvider
Caller->>Provider: formatGeminiMessages(messages)
Note over Provider: Translates tool_calls, user/model parts, inlineData
Caller->>Provider: getFormattedSafetySettings()
Note over Provider: Maps harm categories and thresholds
Caller->>Provider: getGenerationConfig(modelId)
alt model supports thinking budget
Provider-->>Caller: config + thinkingBudget + thinkingConfig.includeThoughts
else
Provider-->>Caller: config without thinkingBudget
end
Estimated code review effort🎯 4 (Complex) | ⏱️ ~60 minutes Possibly related PRs
Poem
Tip 🔌 Remote MCP (Model Context Protocol) integration is now available!Pro plan users can now connect to remote MCP servers from the Integrations page. Connect with popular remote MCPs such as Notion and Linear to add more context to your reviews and chats. ✨ Finishing Touches
🧪 Generate unit tests
Thanks for using CodeRabbit! It's free for OSS, and your support helps us grow. If you like it, consider giving us a shout-out. 🪧 TipsChatThere are 3 ways to chat with CodeRabbit:
SupportNeed help? Create a ticket on our support page for assistance with any issues or questions. CodeRabbit Commands (Invoked using PR/Issue comments)Type Other keywords and placeholders
CodeRabbit Configuration File (
|
![coderabbitai[bot]](https://avatars.githubusercontent.com/in/347564?s=60&v=4){kind=small}
There was a problem hiding this comment.
Actionable comments posted: 0
Caution
Some comments are outside the diff and can’t be posted inline due to platform limitations.
⚠️ Outside diff range comments (2)
src/main/presenter/llmProviderPresenter/providers/geminiProvider.ts (2)
361-366: Bug: comparing enum values to string literals — safety thresholds never filtered.threshold is HarmBlockThreshold (an enum), but it’s compared to strings, so the condition is always true for non-zero enums. Fix the comparisons to enum members.
Apply this diff:
- if ( - threshold && - category && - threshold !== 'BLOCK_NONE' && - threshold !== 'HARM_BLOCK_THRESHOLD_UNSPECIFIED' - ) { + if ( + category && + threshold !== undefined && + threshold !== HarmBlockThreshold.BLOCK_NONE && + threshold !== HarmBlockThreshold.HARM_BLOCK_THRESHOLD_UNSPECIFIED + ) {
929-1090: Stream error handling: add try/catch and yield standardized error events.coreStream must yield error events instead of throwing. Wrap the streaming request + loop; on failure, emit error and stop('error').
Apply this diff:
- // 发送流式请求 - const result = await this.genAI.models.generateContentStream(requestParams) - - // 状态变量 + // 发送流式请求 + try { + const result = await this.genAI.models.generateContentStream(requestParams) + + // 状态变量 let buffer = '' let isInThinkTag = false let toolUseDetected = false let usageMetadata: GenerateContentResponseUsageMetadata | undefined let isNewThoughtFormatDetected = false @@ - for await (const chunk of result) { + for await (const chunk of result) { // 处理用量统计 if (chunk.usageMetadata) { usageMetadata = chunk.usageMetadata } @@ - } - } - - if (usageMetadata) { + } + if (usageMetadata) { yield { type: 'usage', usage: { prompt_tokens: usageMetadata.promptTokenCount || 0, completion_tokens: usageMetadata.candidatesTokenCount || 0, total_tokens: usageMetadata.totalTokenCount || 0 } } - } - - // 处理剩余缓冲区内容 - if (!isNewThoughtFormatDetected && buffer) { + } + // 处理剩余缓冲区内容 + if (!isNewThoughtFormatDetected && buffer) { if (isInThinkTag) { yield { type: 'reasoning', reasoning_content: buffer } } else { yield { type: 'text', content: buffer } } - } - - // 发送停止事件 - yield { type: 'stop', stop_reason: toolUseDetected ? 'tool_use' : 'complete' } + } + // 发送停止事件 + yield { type: 'stop', stop_reason: toolUseDetected ? 'tool_use' : 'complete' } + } catch (error) { + console.error('Gemini stream error:', error) + yield { + type: 'error', + error_message: error instanceof Error ? error.message : String(error) + } + yield { type: 'stop', stop_reason: 'error' } + return + }
🧹 Nitpick comments (4)
src/main/presenter/llmProviderPresenter/providers/geminiProvider.ts (4)
946-946: Reduce noisy debug logging in hot loop.JSON.stringify of candidates on every chunk is expensive and floods logs. Gate behind a DEBUG flag or remove.
Apply this diff:
- console.log('chunk.candidates', JSON.stringify(chunk.candidates, null, 2)) + if (process.env.DEEPCHAT_DEBUG_GEMINI === '1') { + console.debug('chunk.candidates', JSON.stringify(chunk.candidates)?.slice(0, 5_000)) + }
693-733: Deduplicate response parsing.completions reimplements the same thought/text extraction as processGeminiResponse. Call processGeminiResponse(result) to reduce drift.
- // 处理响应内容，支持新格式的思考内容 - let textContent = '' - let thoughtContent = '' - ... - resultResp.content = textContent - if (thoughtContent) { - resultResp.reasoning_content = thoughtContent - } - - return resultResp + return this.processGeminiResponse(result)
262-265: Logs/comments language and structure.Guidelines require English logs/comments and structured logging with levels. Convert Chinese log lines and consider INFO/WARN/ERROR with context. Avoid leaking sensitive data.
Also applies to: 307-309, 755-768, 847-849, 1151-1157
979-980: Optional: keep reading after tool call to collect usage.Breaking immediately after tool_call may miss late usageMetadata chunks. Consider continuing the loop while suppressing further content events, then emit usage and stop(tool_use).
📜 Review details
Configuration used: CodeRabbit UI
Review profile: CHILL
Plan: Pro
💡 Knowledge Base configuration:
- MCP integration is disabled by default for public repositories
- Jira integration is disabled by default for public repositories
- Linear integration is disabled by default for public repositories
You can enable these sources in your CodeRabbit configuration.
📒 Files selected for processing (1)
src/main/presenter/llmProviderPresenter/providers/geminiProvider.ts(4 hunks)
🧰 Additional context used
📓 Path-based instructions (8)
**/*.{js,jsx,ts,tsx}
📄 CodeRabbit inference engine (.cursor/rules/development-setup.mdc)
**/*.{js,jsx,ts,tsx}: 使用 OxLint 进行代码检查
Log和注释使用英文书写
Files:
src/main/presenter/llmProviderPresenter/providers/geminiProvider.ts
src/{main,renderer}/**/*.ts
📄 CodeRabbit inference engine (.cursor/rules/electron-best-practices.mdc)
src/{main,renderer}/**/*.ts: Use context isolation for improved security
Implement proper inter-process communication (IPC) patterns
Optimize application startup time with lazy loading
Implement proper error handling and logging for debugging
Files:
src/main/presenter/llmProviderPresenter/providers/geminiProvider.ts
src/main/**/*.ts
📄 CodeRabbit inference engine (.cursor/rules/electron-best-practices.mdc)
Use Electron's built-in APIs for file system and native dialogs
From main to renderer, broadcast events via EventBus using mainWindow.webContents.send()
Files:
src/main/presenter/llmProviderPresenter/providers/geminiProvider.ts
**/*.{ts,tsx}
📄 CodeRabbit inference engine (.cursor/rules/error-logging.mdc)
**/*.{ts,tsx}: 始终使用 try-catch 处理可能的错误
提供有意义的错误信息
记录详细的错误日志
优雅降级处理
日志应包含时间戳、日志级别、错误代码、错误描述、堆栈跟踪（如适用）、相关上下文信息
日志级别应包括 ERROR、WARN、INFO、DEBUG
不要吞掉错误
提供用户友好的错误信息
实现错误重试机制
避免记录敏感信息
使用结构化日志
设置适当的日志级别Enable and adhere to strict TypeScript type checking
Files:
src/main/presenter/llmProviderPresenter/providers/geminiProvider.ts
src/main/presenter/llmProviderPresenter/providers/*.ts
📄 CodeRabbit inference engine (.cursor/rules/llm-agent-loop.mdc)
src/main/presenter/llmProviderPresenter/providers/*.ts: Each file insrc/main/presenter/llmProviderPresenter/providers/*.tsshould handle interaction with a specific LLM API, including request/response formatting, tool definition conversion, native/non-native tool call management, and standardizing output streams to a common event format.
Provider implementations must use acoreStreammethod that yields standardized stream events to decouple the main loop from provider-specific details.
ThecoreStreammethod in each Provider must perform a single streaming API request per conversation round and must not contain multi-round tool call loop logic.
Provider files should implement helper methods such asformatMessages,convertToProviderTools,parseFunctionCalls, andprepareFunctionCallPromptas needed for provider-specific logic.
All provider implementations must parse provider-specific data chunks and yield standardized events for text, reasoning, tool calls, usage, errors, stop reasons, and image data.
When a provider does not support native function calling, it must prepare messages using prompt wrapping (e.g.,prepareFunctionCallPrompt) before making the API call.
When a provider supports native function calling, MCP tools must be converted to the provider's format (e.g., usingconvertToProviderTools) and included in the API request.
Provider implementations should aggregate and yield usage events as part of the standardized stream.
Provider implementations should yield image data events in the standardized format when applicable.
Provider implementations should yield reasoning events in the standardized format when applicable.
Provider implementations should yield tool call events (tool_call_start,tool_call_chunk,tool_call_end) in the standardized format.
Provider implementations should yield stop events with appropriatestop_reasonin the standardized format.
Provider implementations should yield error events in the standardized format...
Files:
src/main/presenter/llmProviderPresenter/providers/geminiProvider.ts
src/main/**/*.{ts,js,tsx,jsx}
📄 CodeRabbit inference engine (.cursor/rules/project-structure.mdc)
主进程代码放在
src/main
Files:
src/main/presenter/llmProviderPresenter/providers/geminiProvider.ts
src/**/*.{ts,tsx,vue}
📄 CodeRabbit inference engine (CLAUDE.md)
Use English for all logs and comments
Files:
src/main/presenter/llmProviderPresenter/providers/geminiProvider.ts
src/main/presenter/**/*.ts
📄 CodeRabbit inference engine (CLAUDE.md)
Maintain one presenter per functional domain in src/main/presenter/
Files:
src/main/presenter/llmProviderPresenter/providers/geminiProvider.ts
🧠 Learnings (7)
📓 Common learnings
Learnt from: neoragex2002
PR: ThinkInAIXYZ/deepchat#550
File: src/main/presenter/mcpPresenter/inMemoryServers/meetingServer.ts:258-324
Timestamp: 2025-06-21T15:53:34.641Z
Learning: User neoragex2002 prefers concise, brief responses rather than chatty or verbose communication. Avoid lengthy explanations, excessive enthusiasm, or ASCII art in responses.
📚 Learning: 2025-07-21T01:46:52.880Z
Learnt from: CR
PR: ThinkInAIXYZ/deepchat#0
File: .cursor/rules/llm-agent-loop.mdc:0-0
Timestamp: 2025-07-21T01:46:52.880Z
Learning: Applies to src/main/presenter/llmProviderPresenter/providers/*.ts : All provider implementations must parse provider-specific data chunks and yield standardized events for text, reasoning, tool calls, usage, errors, stop reasons, and image data.
Applied to files:
src/main/presenter/llmProviderPresenter/providers/geminiProvider.ts
📚 Learning: 2025-07-21T01:46:52.880Z
Learnt from: CR
PR: ThinkInAIXYZ/deepchat#0
File: .cursor/rules/llm-agent-loop.mdc:0-0
Timestamp: 2025-07-21T01:46:52.880Z
Learning: Applies to src/main/presenter/llmProviderPresenter/providers/*.ts : Provider implementations must use a `coreStream` method that yields standardized stream events to decouple the main loop from provider-specific details.
Applied to files:
src/main/presenter/llmProviderPresenter/providers/geminiProvider.ts
📚 Learning: 2025-07-21T01:46:52.880Z
Learnt from: CR
PR: ThinkInAIXYZ/deepchat#0
File: .cursor/rules/llm-agent-loop.mdc:0-0
Timestamp: 2025-07-21T01:46:52.880Z
Learning: Applies to src/main/presenter/llmProviderPresenter/providers/*.ts : Provider implementations should yield text events in the standardized format.
Applied to files:
src/main/presenter/llmProviderPresenter/providers/geminiProvider.ts
📚 Learning: 2025-07-21T01:46:52.880Z
Learnt from: CR
PR: ThinkInAIXYZ/deepchat#0
File: .cursor/rules/llm-agent-loop.mdc:0-0
Timestamp: 2025-07-21T01:46:52.880Z
Learning: Applies to src/main/presenter/llmProviderPresenter/providers/*.ts : Provider implementations should yield reasoning events in the standardized format when applicable.
Applied to files:
src/main/presenter/llmProviderPresenter/providers/geminiProvider.ts
📚 Learning: 2025-08-26T14:13:46.601Z
Learnt from: CR
PR: ThinkInAIXYZ/deepchat#0
File: CLAUDE.md:0-0
Timestamp: 2025-08-26T14:13:46.601Z
Learning: Applies to src/main/presenter/llmProviderPresenter/providers/*.ts : Each LLM provider must implement provider-specific API interactions, convert MCP tools, and normalize streaming responses
Applied to files:
src/main/presenter/llmProviderPresenter/providers/geminiProvider.ts
📚 Learning: 2025-08-26T14:13:46.601Z
Learnt from: CR
PR: ThinkInAIXYZ/deepchat#0
File: CLAUDE.md:0-0
Timestamp: 2025-08-26T14:13:46.601Z
Learning: Applies to src/main/presenter/llmProviderPresenter/providers/*.ts : Implement a coreStream method for new providers following the standardized event interface
Applied to files:
src/main/presenter/llmProviderPresenter/providers/geminiProvider.ts
⏰ 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). (1)
- GitHub Check: build-check (x64)
🔇 Additional comments (6)
src/main/presenter/llmProviderPresenter/providers/geminiProvider.ts (6)
937-938: Fix: Gating legacy parsing behind a new-thought flag is correct.Introducing isNewThoughtFormatDetected prevents the old matcher from touching new-format streams. This directly addresses the parsing bug described in the PR.
989-996: Set the flag at detection site (part.thought) and emit reasoning immediately — good.Detecting thought in-chunk and switching mode avoids mid-stream mis-splits. No further buffering is applied once the flag flips.
1021-1058: Bypass legacy path once new-thought is detected.Directly yielding text and skipping the buffer/ logic eliminates body corruption. Logic is sound and minimal.
1073-1085: End-of-stream: suppress leftover buffer when new-thought was present.Prevents accidental emission of stale buffered chunks. Matches the intended behavior.
443-489: Harden JSON parsing ingeminiProvider.tsand verifyfunctionResponseschema• In
src/main/presenter/llmProviderPresenter/providers/geminiProvider.ts(around lines 443–489), wrap theJSON.parsecall in a try/catch so that invalid JSON intoolCall.function.argumentswon’t crash the provider.
• Confirm the exact shape the Gemini Node SDK expects for a function response—whether it’s
{ functionResponse: { name: string; response: any } }
or
{ functionResponse: { name: string; response: { result: any } } }—and adjust the code below accordingly.Suggested (optional) diff for the parsing guard; once you’ve confirmed the SDK schema, update the inner
functionResponse.responsereference if needed:--- a/src/main/presenter/llmProviderPresenter/providers/geminiProvider.ts +++ b/src/main/presenter/llmProviderPresenter/providers/geminiProvider.ts @@ -460,7 +460,18 @@ for (const toolCall of message.tool_calls || []) { role: 'model', parts: [ { - functionCall: { - name: toolCall.function.name, - args: JSON.parse(toolCall.function.arguments || '{}') - } + functionCall: { + name: toolCall.function.name, + args: (() => { + try { + return toolCall.function.arguments + ? JSON.parse(toolCall.function.arguments) + : {}; + } catch (err) { + console.warn( + '[GeminiProvider] Invalid tool_call.arguments JSON, defaulting to {}', + err + ); + return {}; + } + })() + } } ] });After updating the parsing, please verify and align the
functionResponseblock to match the SDK’s expected schema.
646-651: No change needed:systemInstructionbelongs insideconfigThe official Node.js SDK shows
systemInstructionpassed under theconfigobject, not as a top-level field:
- In the Google AI for Developers Node.js example,
generateContentis called withconfig: { systemInstruction: "…" }(not a top-level property) (ai.google.dev)Please leave the existing placement unchanged.
Likely an incorrect or invalid review comment.
老版geminiProvider在处理完新版 API 的
thought字段后，没有停止，而是继续用旧的、不健壮的<think>标签匹配逻辑去处理正文内容，导致了正文被错误地识别和分割。Summary by CodeRabbit