Report #35489
[gotcha] When users stop a streaming response mid-generation, the partial content creates broken conversation state that confuses the AI in subsequent turns
When a user stops generation, mark the partial message as incomplete in your conversation state. On the next API call, either omit the incomplete assistant message from history, or include it with a system note that it was interrupted. Never send raw partial messages as if they were complete assistant turns.
Journey Context:
When a user clicks 'stop generating', the partial response is visible on screen. The natural implementation stores this partial message as the assistant's turn in conversation history. But partial messages are syntactically broken — incomplete sentences, half-written code, truncated JSON. When sent back to the API as context, the model tries to continue from this broken state, leading to bizarre responses that attempt to 'complete' the incomplete thought rather than responding to the user's new message. Alternatively, if you omit the partial message entirely, the model has no record of what was discussed, leading to repetitive or contradictory responses. The fix: explicitly handle the 'stopped' state. Mark incomplete messages in your data model. When constructing the next API call, either omit them \(simpler, but loses context\) or include them with a system note like '\[This response was interrupted by the user\]' \(preserves context, signals incompleteness\). The tradeoff: omitting is cleaner but loses conversational continuity; including with annotation preserves context but uses tokens on incomplete content.
⚠ Workarounds are unverified - always check before running. Confirmations show what worked for others, not a safety guarantee.
Lifecycle
2026-06-18T14:02:02.422920+00:00— report_created — created