[Bug]: DeepSeek V4-flash/pro fails with 400 Bad Request (missing reasoning_content) while deepseek-reasoner works

[xjx22]

Describe the bug
When using the latest DeepSeek V4 models (deepseek-v4-flash or deepseek-v4-pro) via the OpenAI-compatible provider, claw-code fails on the second turn of a conversation or after a tool call.
The API returns a 400 Bad Request error:
api returned 400 Bad Request (invalid_request_error): The reasoning_content in the thinking mode must be passed back to the API.
Interestingly, using the legacy model name deepseek-reasoner does not trigger this error, even though it also involves a thinking process. This suggests that claw-code is dropping the reasoning_content field from the message history when constructing the next request, which is strictly required by the newer DeepSeek V4 API contract.
To Reproduce
Set OPENAI_BASE_URL to https://api.deepseek.com.
Run: claw --model openai/deepseek-v4-flash prompt "Analyze the code in this directory"
The model performs a read_file or ls call.
On the next turn, the API returns the 400 error.
Expected behavior
claw-code should preserve and pass back the reasoning_content field in the message history, as required by the DeepSeek API for models in thinking/reasoning mode.
Actual behavior
The reasoning_content is lost in the message context, causing the DeepSeek API to reject the request for breaking the "Thinking Chain" continuity.
Environment
OS: Windows (PowerShell)
Model: openai/deepseek-v4-flash / openai/deepseek-v4-pro
Claw-code version: 0.1.0

[Jengro777]

Set OPENAI_BASE_URL to https://api.deepseek.com/.

How did you change the basic URL。
Doesn't claw code not support deepseek

————————

export ANTHROPIC_API_KEY="https://api.deepseek.com/anthropic"
export ANTHROPIC_AUTH_TOKEN="******************"
Is this the configuration？

[xjx22]

No, I am not using the Anthropic-style configuration you mentioned.
Instead, I configured it using the standard OpenAI format because DeepSeek is OpenAI-compatible. Specifically, I set the following environment variables:
OPENAI_BASE_URL: https://api.deepseek.com
OPENAI_API_KEY: xxxxx (My DeepSeek API key)
Interestingly, it works perfectly when I use:
claw --model openai/deepseek-reasoner
However, when I try using openai/deepseek-v4-flash or pro, it fails and shows the error message I mentioned earlier.

[marco-peach-melba]

I encountered this issue using the OpenAI format as well，
OPENAI_BASE_URL: https://api.deepseek.com/
OPENAI_API_KEY: xxxxx (My DeepSeek API key)

PS J:\claw-code\rust\target\debug> .\claw.exe
██████╗██╗ █████╗ ██╗ ██╗
██╔════╝██║ ██╔══██╗██║ ██║
██║ ██║ ███████║██║ █╗ ██║
██║ ██║ ██╔══██║██║███╗██║
╚██████╗███████╗██║ ██║╚███╔███╔╝
╚═════╝╚══════╝╚═╝ ╚═╝ ╚══╝╚══╝ Code 🦞

Model deepseek-v4-pro
Permissions danger-full-access
Branch main
Workspace dirty · 1 files · 1 untracked
Directory J:\claw-code\rust\target\debug
Session session-1777366704068-0
Auto-save \?\J:\claw-code\rust\target\debug.claw\sessions\2bd7152c96c9d635\session-1777366704068-0.jsonl

Type /help for commands · /status for live context · /resume latest jumps back to the newest session · /diff then /commit to ship · Tab for workflow completions · Shift+Enter for newline
Connected: deepseek-v4-pro via openai

then

✘ ❌ Request failed
[error-kind: api_http_error]
error: api returned 400 Bad Request (invalid_request_error): The reasoning_content in the thinking mode must be passed back to the API.

Run claw --help for usage.
PS J:\claw-code\rust\target\debug>

that's all,It seems that we need to adapt to the new changes in deepseek

[xjx22]

I have some new findings. It seems to be caused by the DeepSeek V4 protocol requirements:
When the model's response contains tool_calls, the reasoning_content field must be completely passed back in all subsequent API requests. If this field is missing or empty, the API will return a 400 error. However, the current Clawcode does not support passing back reasoning_content when handling multi-turn calls, which ultimately leads to the API 400 validation failure

[smllb]

same here

✘ ❌ Request failed
[error-kind: api_http_error]
error: api returned 400 Bad Request (invalid_request_error): The content[].thinking in the thinking mode must be passed back to the API.

[code-yeongyu]

Fixed in PR #3001 (75c08bc9).