Turn can end with a reasoning-only assistant message and no visible output #58686

Summary

A turn can end with nothing shown to the user: the final assistant message contains only a
reasoning part, finish = "stop", no text, no tool call. The spinner stops and the session
sits idle — from the user's side the agent "silently dropped out" mid-work. Measured 6 times
in 2.5 days on one machine (2% of turns), each followed by the user waiting 8–58 minutes
before sending a nudge by hand.

Environment

• opencode version: 2.0.26
• OS: Windows 10.0.26300 (win32 x64)
• Install/channel: latest
• Model on the affected turns: deepseek-v4.1-flash via opencode-go (per message record)
• Active plugins: @cortexkit/opencode-magic-context@0.48.0

Reproduction

Not deterministic; it is a provider/model outcome that the harness accepts as a finished
turn. To observe it, run long multi-step turns on a reasoning model and watch for a turn that
ends with reasoning only:

session_message -> assistant: finish = "stop", rawFinish = "stop",
                  content = [ { "type": "reasoning" } ]   // no text, no tool part
session_message -> idle                                    // turn closed

Expected Behavior

Either the turn does not close without an answer (keep generating / retry once), or the
session clearly reports that the model returned no answer.

Actual Behavior

The turn closes silently. The user sees the spinner stop and no output; the conversation only
continues when they send another message themselves.

Measured on this machine (487 turn ends, 2.5 days): 6 occurrences, all in one long
session (a build agent working through a checklist). Gaps until the user's next message:
58.3, 29.8, 19.7, 8.1 minutes and two cases with no follow-up at all.

Additional Context

Why the harness cannot tell this apart from a normal end

finish = "stop" is what the provider reported (rawFinish is also "stop"), so from the
runner's point of view the model finished its answer — it just had nothing to say. There is
no check of the form "the final assistant message has no text and no tool call".

2.1.4 does not add such a check

Searched packages/ai/src and packages/core/src/session in 2.1.4 for empty-response
handling: the only related code is history normalisation (omits empty assistant turns,
filters whitespace-only text) — no retry and no user-visible notice when the final message
has no content. After upgrading to 2.1.4 no new empty turn was observed in the following
hour; the sample is too short to conclude anything, and the code path is unchanged.

Platform scope

Not OS-specific: the outcome comes from the provider and the runner accepts it as a finished
turn. The measurements below are from one Windows machine, but nothing in the path depends on
the platform.

Suggested fix direction

Treat "final assistant message has no text, no tool call, no pending tool" as an anomaly:
either retry the step once, or append a visible notice (Модель вернула пустой ответ — продолжить?). Even a one-line marker would turn a silent stop into a known state.

Cost observed

Six silent stops, ~2 hours of combined waiting where the agent had already stopped and
nobody knew.