DeepSeek V4 tool call fails: reasoning_content must be passed back

[Anoxian]

DeepSeek V4 tool call fails with:
"The reasoning_content in the thinking mode must be passed back to the API."

AltCode v0.10.5 drops thinking from config and does not pass reasoning_content back after tool calls.
Expected: either support thinking disabled, or preserve reasoning_content in follow-up requests.

[xiaoshuo1988130]

The 400 error happens because DeepSeek V4 strictly requires the reasoning_content from previous turns to be passed back, but many frameworks strip it out when a tool_calls is present.

Until a permanent upstream fix lands, you can use deepseek-compat-kit as a quick zero-config local relay to automatically cache and re-inject the missing reasoning_content.

Just run it and point your client's baseURL to http://127.0.0.1:8787/v1 to bypass the 400 rejection instantly.

[norika1207-lab]

There are two separate knobs here and I would keep them explicit:

1. If thinking is disabled in AltCode config, the provider request should not enter a replay mode that requires reasoning_content.
2. If thinking is enabled, every assistant tool-call message that goes back to DeepSeek has to carry the provider-required reasoning field.

The bad middle state is “UI/config says thinking is off or ignored, but the transcript still contains a thinking-mode tool-call contract.” That produces the 400 and is hard for users to diagnose.

A practical fix is to add a final DeepSeek request-normalization step before dispatch:

• preserve any existing reasoning_content
• for assistant messages with tool_calls and no field, either backfill an empty field when thinking mode is truly enabled, or strip/disable thinking consistently for that request
• log how many messages were repaired and at which indices

The log matters because backfill fixes schema replay, not semantic recovery. Users should know whether the original reasoning survived or the client only repaired the provider contract.