Skip to content

web-fetch fails with DeepSeek in thinking mode — reasoning_content not passed back between tool calls #3094

@Giri-Aayush

Description

@Giri-Aayush

User Iphilo reports that triggering web-fetch consistently fails when using DeepSeek Flash (thinking mode) as the LLM provider. Logs provided with exact error.

Error (400 Bad Request from DeepSeek API)
The reasoning_content in the thinking mode must be passed back to the API.

Root cause
DeepSeek's thinking/reasoning models require the reasoning_content field from a previous assistant message to be echoed back in the next API request. OpenHuman's message-building logic does not currently preserve or replay this field when constructing subsequent requests, causing DeepSeek to reject the call with a 400. This affects the web-fetch tool and likely any multi-turn tool use with DeepSeek thinking models.

Environment
Provider: DeepSeek Flash (thinking mode via custom API)
Build path visible in logs: D:\a\openhuman\openhuman\src (Windows build)
Full logs in message.txt (3 KB) attached in Discord.


Reported by: @.iphilo
Discord ticket: https://discord.com/channels/1470856511193616498/1509448788077576243
Priority: high
Attachments (1): message (1).txt — ask support team for files

Metadata

Metadata

Labels

Type

No type
No fields configured for issues without a type.

Projects

No projects

Milestone

No milestone

Relationships

None yet

Development

No branches or pull requests

Issue actions