fix: preserve reasoning_content in multi-turn tool-call sessions (MiMo compatibility)

[Noogear]

Problem

When using models like Xiaomi MiMo that require reasoning_content (or thinking blocks) to be preserved in historical assistant messages containing tool_calls, the API returns 400 error on subsequent requests in multi-turn agent sessions.

This happens because VS Code's LanguageModelThinkingPart is not guaranteed to be preserved in conversation history between requests, causing reasoning content to be lost when reconstructing messages.

See: https://platform.xiaomimimo.com/docs/zh-CN/usage-guide/passing-back-reasoning_content

Solution

Implemented a 3-level reasoning content caching mechanism, gated by the existing include_reasoning_in_request config option (opt-in only):

1. Map cache (_reasoningContentCache): keyed by sorted tool-call IDs, persists across requests within a session
2. Last reasoning fallback (_lastReasoningContent): stores the most recent reasoning content as a cross-restart fallback
3. Placeholder fallback: "Next step." when no cached content is available

Files changed

File	Change
src/commonApi.ts	Added reasoning content accumulation (_accumulatedReasoningContent) and tool-call ID tracking (_emittedToolCallIds) during streaming
src/openai/openaiApi.ts	convertMessages() now accepts optional reasoning cache; restores reasoning_content for assistant messages with tool_calls
src/anthropic/anthropicApi.ts	convertMessages() now accepts optional reasoning cache; restores thinking content blocks for assistant messages with tool_use (was completely missing before)
src/provider.ts	Added _reasoningContentCache Map and _lastReasoningContent string; passes cache to both OpenAI and Anthropic paths; stores reasoning content after streaming completes

Behavior

• Only activates when include_reasoning_in_request: true is set in model config
• No model ID detection or auto-detection — purely opt-in
• Works for both OpenAI Chat Completions and Anthropic protocol modes

Testing

Verified with Xiaomi MiMo (mimo-v2.5-pro) using Anthropic protocol mode.
Multi-turn agent sessions with tool calls no longer produce 400 errors.

Config example:

{
  "id": "mimo-v2.5-pro",
  "apiMode": "anthropic",
  "baseUrl": "https://token-plan-sgp.xiaomimimo.com/anthropic",
  "include_reasoning_in_request": true
}

[RayWangQvQ]

Same issue:

Sorry, your request failed. Please try again.

Copilot Request id: a0cd499a-e5b7-46b0-947d-0c4f9eead4ec

Reason: OAI Compatible API error: [400] Bad Request {"error":{"code":"400","message":"Param Incorrect","param":"The reasoning_content in the thinking mode must be passed back to the API.","type":""}} URL: [https://token-plan-cn.xiaomimimo.com/v1/chat/completions](vscode-file://vscode-app/d:/Program%20Files/Microsoft%20VS%20Code/0958016b2a/resources/app/out/vs/code/electron-browser/workbench/workbench.html): Error: OAI Compatible API error: [400] Bad Request {"error":{"code":"400","message":"Param Incorrect","param":"The reasoning_content in the thinking mode must be passed back to the API.","type":""}} URL: [https://token-plan-cn.xiaomimimo.com/v1/chat/completions](vscode-file://vscode-app/d:/Program%20Files/Microsoft%20VS%20Code/0958016b2a/resources/app/out/vs/code/electron-browser/workbench/workbench.html) at c:\Users\ray.vscode\extensions\johnny-zhao.oai-compatible-copilot-0.4.2\out\provider.js:428:31 at process.processTicksAndRejections (node:internal/process/task_queues:103:5) at async executeWithRetry (c:\Users\ray.vscode\extensions\johnny-zhao.oai-compatible-copilot-0.4.2\out\utils.js:302:20) at async HuggingFaceChatModelProvider.provideLanguageModelChatResponse (c:\Users\ray.vscode\extensions\johnny-zhao.oai-compatible-copilot-0.4.2\out\provider.js:419:34)

Pls merge it.

[JohnnyZ93]

Did you set "include_reasoning_in_request": true?