Closes the order-of-events race in #861 root cause 3: when the engine
bursts events, the dispatch loop can pull `MessageComplete` off the
channel ahead of `ThinkingComplete`. Today's `MessageComplete` reads
`app.last_reasoning.take()` to attach the reasoning block to the
assistant message in `api_messages`. If `ThinkingComplete` has not
fired yet, `last_reasoning` is `None` and the thinking content is
dropped — DeepSeek V4 then returns HTTP 400 on the next turn because
it requires `reasoning_content` replay for assistant messages that
carry tool calls.
Adds a defensive head-of-handler drain in `MessageComplete`: when
`streaming_thinking_active_entry.is_some()`, finalize the active
thinking entry and stash the reasoning buffer into `last_reasoning`
before the existing body runs. The drain is a no-op in the normal
case where `ThinkingComplete` arrived first (the entry has already
been cleared), so this branch is order-independent.
Adds `message_complete_drain_preserves_thinking_when_thinking_complete_lost`
which exercises the head-of-handler invariant directly: with a
thinking entry still active and `last_reasoning` empty, the drain
must move the buffer into `last_reasoning` before downstream reads.
Refs #861 (RC3). RC1 and RC2 are already addressed by the existing
`finalize_current_streaming_thinking` plumbing in
`apply_engine_error_to_app` and `start_streaming_thinking_block`;
RC4 (streaming-time truncation affordance) is left out of this PR
to keep the scope on the data-loss path.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
LinQ
741f36d2be