Skip to content

model does not support thinking - Claude Code

Quick fix

Turn the Proxy Rectifier off in cc-switch advanced settings, or upgrade to a fixed version. Qwen tool call loops and empty thinking blocks remain unresolved.

Error output
thinking: model does not support thinking, stripping config | provider=iflow model=deepseek-v3.2-reasoner

cc-switch has defects across several versions in how it handles the `thinking` and `reasoning_content` fields when forwarding requests for third-party models such as DeepSeek, GLM, and Qwen. In v3.10.0 the default Proxy Rectifier wrongly strips the thinking config or triggers a 406.

Because the two-way format conversion is incomplete, other variants follow: dropping `reasoning_content` causing an infinite tool-call loop, thinking blocks repeating endlessly, an empty thinking block appended after text scrambling TUI rendering, and token counting failing to parse.

  1. In the cc-switch UI, open the provider's advanced settings and turn the 整流器 (Proxy Rectifier) off.

  2. Upgrade cc-switch to v3.11.0 or newer for the improved Thinking Budget Rectifier. On the openai_chat adapter you can also enable Thinking Mode in advanced settings where available.

ToolClaude Code / Codex
Version3.10.0 - 3.16.5
PlatformsWindowsLinuxmacOS
Why does thinking work in the cc-switch terminal while Claude Code opened another way reports 406?
The other route triggers the rectifier intercepting wrongly. Turning the rectifier off in advanced settings resolves it.
Thinking blocks loop endlessly, or TUI rendering shows only a few characters per line. What now?
Those are known bugs in how cc-switch maps `reasoning_content` and orders content_block. Upgrade to the latest version, or adjust Thinking Mode in advanced settings.

This is an unofficial community wiki with no affiliation to the cc-switch authors or the project itself. Its content is compiled from the project's public GitHub issues. This site distributes no software.