openclaw - 💡(How to fix) Fix [Bug] DeepSeek V4 thinking mode fails with 400: reasoning_content must be passed back on multi-turn tool calls [1 comments, 2 participants]

Bug Report

Description

When using DeepSeek V4 models (deepseek-v4-flash / deepseek-v4-pro) with thinking mode enabled, multi-turn conversations that involve tool calls fail with a 400 error:

The `content[].thinking` in the thinking mode must be passed back to the API.

This happens on every follow-up request after the first assistant response that includes both thinking and tool calls. The DeepSeek API requires reasoning_content from prior assistant turns to be replayed in subsequent requests, but OpenClaw 2026.4.23 does not properly replay the thinking/reasoning content when constructing multi-turn tool-call history.

Environment

• OpenClaw version: 2026.4.23 (a979721)
• Provider: deepseek (direct API, base_url: https://api.deepseek.com, api: openai-completions)
• Models affected: deepseek/deepseek-v4-flash, deepseek/deepseek-v4-pro
• DeepSeek API key: valid, works for single-turn requests
• OS: macOS 26.3 (arm64)

Steps to Reproduce

1. Configure DeepSeek V4 models in openclaw.json with the official OpenAI-compatible endpoint
2. Enable thinking mode (do NOT set params.thinking.type: "disabled")
3. Start a conversation that triggers a tool call (e.g., web_search, exec)
4. On the next user turn, the request fails with 400

Expected Behavior

OpenClaw should replay the prior assistant reasoning_content in the follow-up request, as documented in DeepSeek Thinking Mode and the OpenClaw DeepSeek docs. The docs state:

V4 models support DeepSeek's thinking control. OpenClaw also replays DeepSeek reasoning_content on follow-up turns so thinking sessions with tool calls can continue.

Actual Behavior

The 400 error occurs, and OpenClaw falls back to the next model in the fallback chain. Gateway logs show:

{
  "error": "LLM request failed: provider rejected the request schema or tool payload.",
  "failoverReason": "format",
  "model": "deepseek-v4-pro",
  "provider": "deepseek",
  "rawErrorPreview": "400 {\"error\":{\"message\":\"The `content[].thinking` in the thinking mode must be passed back to the API.\",\"type\":\"invalid_request_error\"}}",
  "providerRuntimeFailureKind": "schema"
}

Workaround

Setting params.thinking.type: "disabled" in the model allowlist entry prevents the error by disabling thinking mode, but this sacrifices V4's reasoning capability:

{
  "agents": {
    "defaults": {
      "models": {
        "deepseek/deepseek-v4-flash": { "alias": "ds4", "params": { "thinking": { "type": "disabled" } } },
        "deepseek/deepseek-v4-pro": { "alias": "ds4p", "params": { "thinking": { "type": "disabled" } } }
      }
    }
  }
}

Additional Context

• Same issue reported in other agent frameworks: OpenCode #24104, OpenCode #24097, QwenPaw #3782
• The bundled DeepSeek plugin catalog in 2026.4.23 still only exposes deepseek-chat / deepseek-reasoner (V3.2 surfaces), not the V4 models. V4 models require manual provider configuration.
• DeepSeek has announced that deepseek-chat and deepseek-reasoner will be retired on 2026-07-24, making V4 thinking support increasingly important.