The first DeepSeek Flash response works, but the request after a tool call asks for the previous thinking content. Where should I look?
This answer refers to the version in the original report. Check the controls and behavior in your own version.
Record the model ID, provider, and API protocol, then check which round fails. DeepSeek’s Chat Completions docs require preserving earlier reasoning_content in thinking-mode requests carrying tools. The report mentions content[].thinking, so verify the field for the protocol you actually use.
Try a fresh session with one read-only tool. If direct access works but a gateway fails, compare redacted responses and follow-up requests to see where the field disappears. The original report did not establish which layer caused it.
