[chengyongru]

Background

Streaming channels intentionally skip the final outbound package when it is marked as a StreamedResponseEvent, because that content is expected to have already arrived through deltas. That suppression mechanism originated in #2365.

Empty-response recovery later made a non-streaming finalization reachable after a stream had already been closed (#2733), and the silent retries added in #2889 amplify the path into multiple stream_end events. #3613 fixed the analogous drop for error/tool-error responses, but empty finalization could still be marked as streamed despite emitting no answer delta.

Root cause

AgentLoop treated the presence of an on_stream callback as proof that the final response content had been streamed. During empty-response finalization, the callback exists, but the final answer comes from chat_with_retry and never passes through the delta callback. ChannelManager then skips the final package, leaving the session populated while streaming clients see no answer.

Fix

• Track whether the most recently completed answer stream segment actually emitted content.
• Emit StreamedResponseEvent only when that final segment delivered content.
• Let non-streaming recovery output remain a regular outbound message so ChannelManager calls channel.send().
• Keep the existing streamed tool-call regression honest by making its fake provider emit real deltas.

Regression coverage

The new test reproduces the production sequence:

• two empty streaming responses
• three StreamEndEvents
• zero answer deltas
• a successful non-streaming finalization

It asserts that the final answer is emitted as a regular outbound message rather than StreamedResponseEvent.

Verification

• pytest tests/agent -q — 1383 passed, 2 skipped
• ruff check nanobot/ — passed
• bun run test — 706 passed
• bun run build — passed
• Real gateway + headless Playwright with a local OpenAI-compatible fake provider:
  • request sequence: streaming empty, streaming empty, non-streaming finalization
  • final answer rendered in WebUI
  • final answer remained visible after reload
  • public webui-thread route returned exactly one assistant final answer