Ollama silently drops conversation history when using previous_response_id parameter
A developer discovered that Ollama's Responses API accepts the previous_response_id parameter without error but completely ignores it, causing the model to treat every request as a fresh conversation. The flaw exists because Ollama's internal request struct has no field for previous_response_id, so the JSON decoder silently discards it, and the response struct explicitly marks it as unsupported. Requests sent with previous_response_id consume the same number of input tokens as requests with no history at all, confirming the model never receives prior context. This silent failure is especially problematic in tool-calling workflows, where the model receives a function result with no surrounding context and produces plausible but incorrect responses. The issue was identified on Ollama 0.34.0 running on Debian 13 with the qwen2.5:1.5b model, and also affects the proxy path used by Codex Desktop.
This is an AI-generated summary. ShortSingh links to the original source for the complete article.


Discussion (0)
Log in to join the discussion and vote.
Log in