openclaw - 💡(How to fix) Fix Bug: DeepSeek V4 reasoning_content silently dropped, causing 400 on multi-turn conversations [1 comments, 2 participants]

Bug Description

When using DeepSeek V4 models (deepseek-v4-pro, deepseek-v4-flash) with multi-turn conversations, the assistant's reasoning_content from previous turns is silently dropped when constructing follow-up API requests. This causes DeepSeek's API to return a 400 error:

"The reasoning_content in the thinking mode must be passed back to the API"

This affects all AI tools built on @mariozechner/pi-ai / OpenClaw (OpenCode, Qwen Code, Roo Code, etc.), making DeepSeek V4 models unusable for multi-turn conversations.

Root Cause

The issue is in convertMessagesToInputItems (in pi-embedded-C7LhuXsi.js):

1. Receiving: DeepSeek returns reasoning_content in streaming deltas. OpenClaw correctly identifies it and stores it as a thinking block with thinkingSignature: "reasoning_content" (a plain string - the field name).

   const reasoningField = [
     "reasoning_content",
     "reasoning",
     "reasoning_text"
   ].find((field) => {
     const value = choice.delta[field];
     return typeof value === "string" && value.length > 0;
   });

2. Replaying: When building subsequent API requests, convertMessagesToInputItems encounters thinking blocks and calls parseThinkingSignature():

   if (block.type === "thinking") {
     const reasoningItem = parseThinkingSignature(block.thinkingSignature);
     if (reasoningItem) items.push(reasoningItem);
     continue;
   }

3. Failure: parseThinkingSignature expects a JSON-parseable string with an rs_ prefixed ID (OpenAI Responses API format):

   function parseThinkingSignature(value) {
     if (typeof value !== "string" || value.trim().length === 0) return null;
     try {
       const signature = toReasoningSignature(JSON.parse(value));
       return signature ? parseReasoningItem(signature) : null;
     } catch {
       return null;  // <-- "reasoning_content" is not JSON, so this fires
     }
   }

   Since thinkingSignature is the plain string "reasoning_content", JSON.parse("reasoning_content") throws, the catch returns null, and the entire thinking block is silently discarded.

4. API Error: The subsequent request to DeepSeek is missing reasoning_content in the assistant message history, which DeepSeek requires. Result: 400 error.

Expected Behavior

When thinkingSignature is a plain string like "reasoning_content" (DeepSeek's native field name, as opposed to an OpenAI Responses API JSON signature), the system should serialize the thinking block back as a reasoning_content field on the assistant message, per the DeepSeek API specification:

{
  "role": "assistant",
  "content": "actual response",
  "reasoning_content": "thinking process..."
}

Environment

• OpenClaw version: 2026.4.5 (3e72c03)
• Provider: DeepSeek (deepseek-native endpoint class)
• Models affected: deepseek-v4-pro, deepseek-v4-flash (all V4 models with default thinking: enabled)
• API adapter: openai-completions

Workaround

1. Use deepseek-chat (non-thinking mode) instead of V4 models, if still available
2. Use a proxy pool (e.g. qclaw/pool-deepseek-v4-pro) which may handle the conversion at the proxy layer
3. This is a cross-ecosystem bug also affecting OpenCode, Qwen Code, Roo Code, and other tools using @mariozechner/pi-ai

Suggested Fix

In convertMessagesToInputItems, add handling for DeepSeek's native reasoning_content signature format:

if (block.type === "thinking") {
  pushAssistantText(currentTextPhase);
  
  // Existing: OpenAI Responses API format (rs_ prefix JSON)
  const reasoningItem = parseThinkingSignature(block.thinkingSignature);
  if (reasoningItem) {
    items.push(reasoningItem);
    continue;
  }
  
  // NEW: DeepSeek native format (plain string field name like "reasoning_content")
  if (["reasoning_content", "reasoning", "reasoning_text"].includes(block.thinkingSignature)) {
    items.push({
      type: "message",
      role: "assistant",
      content: block.thinking,
      reasoning_content: block.thinking
    });
    continue;
  }
}