Request body format invalid error - Claude Code

Quick fix

Upgrade cc-switch and enable the Rectifier, or add MAX_THINKING_TOKENS=0, for compatibility with third-party API formats.

Symptom

Error outputCopy

API Error: 400 {"error":{"code":"invalid_parameter_error","message":"event:error\ndata:{"code":"InvalidParameter","message":"Request body format invalid","request_id":"eedf1443-5885-9123-b239-ea0ad34d499a"}\n\n","param":null,"type":"invalid_request_error"},"request_id":"f7cd2ca2-718e-9a2b-93a5-56bb212e493a"}

Cause

This 400 comes from the request body Claude Code or Codex sends being incompatible with the API specification of a third-party provider such as Alibaba Bailian or DeepSeek. Three main variants: 1. Thinking mode incompatibility: Claude sends thinking: adaptive by default while some third parties only support enabled/disabled. 2. System prompt format change: Claude Code >= 2.1.154 puts the system prompt into the messages array as role: system, while third parties expect it at the top level as a string. 3. Oversized parameter conversion anomalies: very large or deeply nested tool_call parameters in Codex produce an invalid payload during protocol conversion.

Fix

1. 1

   Upgrade cc-switch to v3.11.0 or newer (v3.12.3+ preferred) for the Thinking Budget Rectifier and Normalize system messages features.

2. 2

   Enable the rectifier in cc-switch settings: go to Settings > Proxy > Rectifier and make sure Normalize system messages is on.

3. 3

   If the problem persists, force adaptive thinking mode off in the config.

   // .claude/settings.json
   {
     "env": {
       "MAX_THINKING_TOKENS": "0"
     },
     "alwaysThinkingEnabled": false
   }

4. 4

   (Alternative) On Alibaba Bailian, try switching to the OpenAI-compatible protocol: set the Base URL to https://coding.dashscope.aliyuncs.com/v1, pick openai as the API format, and enable the proxy.

Affected Versions

ToolClaude Code / Codex

VersionClaude Code >= 2.1.154 或 cc-switch < 3.11.0

PlatformsmacOSWindowsLinux

Source Issues

This page is distilled from 4 real issues

• #1095开启本地代理时，qwen的模型请求会报错
• #1128阿里百炼 coding plan 使用glm-5/glm4.7报错 API 请求体错误 thinking模式不能是adaptive
• #3303Claude Code >= 2.1.154 兼容性问题：role:system 消息导致第三方 Provider 请求失败
• #5116[Codex] Responses→chat/completions 转换在复杂/超大 tool_call arguments 时产出上游不合法 payload，返回 400 InvalidParameter

FAQ

Why does it still error after adding the MAX_THINKING_TOKENS=0 environment variable?

After changing an environment variable or config, you must save and restart the Claude Code terminal for it to apply. Also, if the cause is the system prompt format change in Claude Code >= 2.1.154, turning thinking off is not enough — you have to upgrade cc-switch and enable Normalize system messages.

After switching to the OpenAI-compatible protocol I get Cannot read properties of null (reading 'output_tokens'). What now?

That usually means the model name is not mapped correctly, or the upstream response format is non-standard. Drop the OpenAI-compatible protocol, go back to the Anthropic protocol, and rely on the cc-switch Rectifier to correct the request body automatically.

Related problems

api failed deserialize json body intoDepending on the error: roll Claude Code back, upgrade cc-switch, turn the local proxy off, or stop sending unsupported fields.7 issues mergedConsole Go: Upstream request failedThis 400 comes from the OpenCode Go gateway rejecting an incompatible payload. Restart the proxy to restore session mapping, or avoid WebSearch and images.3 issues mergedDeepSeek prefix cache hit rate drop - CC Switch fixUpgrade to cc-switch v3.16.3 to fix cache hit rate calculation errors caused by format conversion.2 issues merged