Bug type
Regression (worked before, now fails)
Beta release blocker
No
Summary
When using deepseek/deepseek-v4-flash through OpenRouter in OpenClaw, follow-up turns can fail with:
code: 20015
The reasoning_content in the thinking mode must be passed back to the API.
This appears to happen because OpenClaw's DeepSeek-specific handling for reasoning_content only activates for provider = "deepseek", not for OpenRouter-routed DeepSeek models. As a result, the conversation history is not normalized/replayed in the way DeepSeek expects on subsequent turns. [docs mention special handling for DeepSeek V4 models in the native DeepSeek provider path.]
Steps to reproduce
- Configure OpenClaw to use OpenRouter
- Set model to
deepseek/deepseek-v4-flash
- Start a conversation
- Send an initial prompt
- Send a follow-up prompt, or trigger a tool call and then continue the conversation
- Observe upstream failure with error
20015
Expected behavior
OpenClaw should detect DeepSeek V4 models even when routed through OpenRouter and apply the same normalization/replay behavior as the native DeepSeek provider path, including handling of reasoning_content across follow-up and tool-call turns.
At minimum, OpenClaw should not break multi-turn conversations for OpenRouter-routed DeepSeek V4 models that require reasoning continuity.
Actual behavior
- Model:
deepseek/deepseek-v4-flash via OpenRouter
- Provider in config:
openrouter
- OpenClaw works on the first turn, but follow-up turns and especially tool-call flows may fail
- Error returned by upstream:
code: 20015The reasoning_content in the thinking mode must be passed back to the API.
In practice:
- if the model emits
reasoning_content, DeepSeek expects that content to be echoed/replayed correctly on subsequent requests
- OpenClaw seems to do the right thing only for native
provider=deepseek
- for
provider=openrouter, the DeepSeek V4 wrapper/path does not seem to run, so the request format/history becomes inconsistent for DeepSeek
OpenClaw version
2026.4.29
Operating system
debian 12.1
Install method
No response
Model
deepseek-v4-flash
Provider / routing chain
openrouter, various providers
Additional provider/model setup details
No response
Logs, screenshots, and evidence
Impact and severity
No response
Additional information
No response
Bug type
Regression (worked before, now fails)
Beta release blocker
No
Summary
When using
deepseek/deepseek-v4-flashthrough OpenRouter in OpenClaw, follow-up turns can fail with:code: 20015The reasoning_content in the thinking mode must be passed back to the API.This appears to happen because OpenClaw's DeepSeek-specific handling for
reasoning_contentonly activates forprovider = "deepseek", not for OpenRouter-routed DeepSeek models. As a result, the conversation history is not normalized/replayed in the way DeepSeek expects on subsequent turns. [docs mention special handling for DeepSeek V4 models in the native DeepSeek provider path.]Steps to reproduce
deepseek/deepseek-v4-flash20015Expected behavior
OpenClaw should detect DeepSeek V4 models even when routed through OpenRouter and apply the same normalization/replay behavior as the native DeepSeek provider path, including handling of
reasoning_contentacross follow-up and tool-call turns.At minimum, OpenClaw should not break multi-turn conversations for OpenRouter-routed DeepSeek V4 models that require reasoning continuity.
Actual behavior
deepseek/deepseek-v4-flashvia OpenRouteropenroutercode: 20015The reasoning_content in the thinking mode must be passed back to the API.In practice:
reasoning_content, DeepSeek expects that content to be echoed/replayed correctly on subsequent requestsprovider=deepseekprovider=openrouter, the DeepSeek V4 wrapper/path does not seem to run, so the request format/history becomes inconsistent for DeepSeekOpenClaw version
2026.4.29
Operating system
debian 12.1
Install method
No response
Model
deepseek-v4-flash
Provider / routing chain
openrouter, various providers
Additional provider/model setup details
No response
Logs, screenshots, and evidence
Impact and severity
No response
Additional information
No response