[Bug] compact latestAnswer: empty last step (plugin same-turn steer) folds the real answer into 用时 and leaves a blank transcript

[iamyhzhao]

Related: #6254 (compact fold picks the last assistant segment). This report is the worse sibling: when that last step has no user-visible text, compact shows no answer at all.

Environment

• DSH desktop / Web GUI 0.1.7-rc.2 (@deepseek-ai/dsh-client-ui-chat@0.1.7-rc.2)
• ui-chat.transcriptView: compact
• macOS, DeepSeek Harness app
• Trigger: a plugin same-turn agent.steer() whose user/message uses a non-user source.kind (reproduced with dsh-memoir autodistill). The prompt asks the model not to emit any user-visible wrap-up text.

What happens

A worked turn produces a long assistant answer (tools + prose). The plugin then steers one extra step in the same turn. That step may call tools and/or produce an empty assistant message.

After the turn closes, compact UI shows:

1. The original user bubble
2. A collapsed 「用时 …」 process disclosure
3. No assistant body

The long answer is still in the session log. Expanding 「用时」 reveals it. From the user’s point of view the agent “replied with nothing”.

Why (0.1.7-rc.2 ui-chat lib/client.js)

latestAnswer only reads the last step:

function latestAnswer(turn) {
  const data = turn.steps.at(-1)?.data.get("assistant-step");
  if (!isFinalAssistant(data) || !hasAssistantReplyContent(data.blocks)) return null;
  return data.blocks.some((block) => block.kind === "tool-call") ? null : data;
}

If that last step has no visible prose (or only tool-call blocks), this returns null. processSpec then sets answerAnchorSeq / answerStep to null. With foldCompletedTurns, every earlier assistant node becomes a process member and is hidden behind 「用时」.

Session V4 already hides the steer prompt: source.kind !== "user" is classified as context, and isVisibleChatNode drops ordinary context. That hide seam does not apply when choosing the turn’s answer. Plugin identity on the steer (source.kind === 'dsh-memoir' in our case) never reaches latestAnswer.

#6254 is the same function when the last step does have a short closing sentence (that sentence stays visible, earlier prose folds). Empty last step is the null-answer path of the same bug.

This is not memoir-specific. Any plugin that same-turn steer()s a silent follow-up will blank compact.

Expected

The last user-visible assistant prose in the turn should remain the compact answer. A silent plugin follow-up must not fold that prose into process or leave a blank transcript.

Suggested host fix (either is enough)

1. Walk latestAnswer backwards until a finalized assistant step has hasAssistantReplyContent and no tool-call blocks (empty/tool-only tail steps do not steal the answer).
2. And/or ignore steps whose opening user/message has a non-user plugin source.kind when computing the turn answer.

Reproduce

1. Compact transcript view.
2. Complete a top-level turn that calls tools and then writes a long assistant answer.
3. Before the turn commits, steer() a user-role message with source.kind other than user, instructing the model to call tools if needed and not to output user-visible text.
4. After the turn closes: 「用时」 only, long answer gone from the default view.

Happy to add a screenshot in a follow-up comment if the web UI allows attachments.