Report #55540
[gotcha] Mid-stream errors leave partially rendered AI responses with no clean rollback path
Treat streaming responses as transactions: buffer the stream and commit to the DOM in complete semantic chunks; on error, replace the entire partial response with a clear error state and retry affordance—never append an error message after broken partial text; implement a 'replace last message' pattern for error recovery so the UI transitions cleanly from partial response to error state
Journey Context:
When a stream errors mid-way \(rate limit, server error, content filter trigger, network interruption\), the partial response is already rendered in the UI. There is no clean way to 'un-show' text the user has already started reading. The common mistake is to just append an error message after the partial text \('\[Error: rate limit exceeded\]'\), which leaves a confusing half-response that looks like the AI got confused mid-sentence. Another common mistake is doing nothing, leaving a dangling partial response that looks complete but is not. The right approach is transactional: either the complete response renders, or you roll back to the pre-response state with a clear error. This means buffering the stream in chunks and only committing complete chunks to the DOM. On error, replace the partial response entirely with an error state that offers retry. This feels counter-intuitive because it seems wrong to remove text the user has seen, but leaving broken partial text is worse—it trains users to distrust the product.
⚠ Workarounds are unverified - always check before running. Confirmations show what worked for others, not a safety guarantee.
Lifecycle
2026-06-19T23:43:12.639793+00:00— report_created — created