Residents watching a streamed 311 answer can tap “wrong street” while tokens are still arriving. How should the conversation flow be developed?

Residents watching a streamed 311 answer must be able to interrupt when they see an early mistake such as the wrong street. Dynamic streaming conversation flows should expose mid-stream feedback that cancels or redirects generation so the agent can correct course before finishing a wrong civic answer. That design reduces wasted tokens, shortens time-to-correction, and matches how people naturally interrupt a clerk who starts down the wrong address. The conversation state should accept the correction signal, halt or revise the in-flight stream, and continue with the corrected street or topic without discarding the whole session context when possible.

Ignoring taps until the full answer finishes treats streaming as one-way broadcast and forces residents to absorb incorrect guidance before any fix. Forcing a brand-new chat for every mid-answer correction discards useful context and punishes users for catching errors early. Disabling streaming so only one complete block arrives after all tools finish removes progressive feedback and first-token responsiveness that public-facing municipal agents need under latency SLOs.

Exam caveat: mid-stream correction is a conversation-control capability, not a claim that streaming alone guarantees factual correctness; tools and policies still ground the revised answer. Operational check: while a 311 stream is emitting the wrong street name, send a “wrong street” interrupt, verify generation stops or redirects, confirm the corrected street appears in the continued reply, and ensure residents are not required to open a new chat solely to fix that mid-stream error.