Summary / 概述

EN: Fixes #1727. When a model streams a turn containing only a reasoning/thinking block — empty content, no tool_calls (e.g. gpt-oss via ollama's harmony→OpenAI shim, which maps the answer into reasoning_content) — the engine persisted nothing, emitted no event, and finished the turn silently. The UI sat idle: spinner never cleared, no reply, no error. It looked hung. This PR surfaces a status notice instead of silence.

中文： 修复 #1727。当模型本回合只返回了 reasoning/thinking 块——content 为空、无 tool_calls（例如 ollama 的 harmony→OpenAI 适配把答案塞进 reasoning_content 的 gpt-oss）——引擎既不持久化任何消息、也不发出任何事件，回合静默结束。界面就此卡住：spinner 不消失、无回复、无报错，看起来像挂死。本 PR 用一条 status 提示替代这种静默。

Root cause / 根因

EN: In crates/tui/src/core/engine/turn_loop.rs, the assistant‑message persist site is an if‑only branch gated on has_sendable_assistant_content (true only for Text/ToolUse). A thinking‑only stream produces neither, so the if is skipped — no message, no event — and if tool_uses.is_empty() then falls straight through to break, ending the turn with zero events emitted. The maintainer's v0.8.40 audit (#1736) confirms #1727 was not shipped in v0.8.39 and is release‑blocking, and that this exact status branch is still missing.

中文： 在 crates/tui/src/core/engine/turn_loop.rs 中，assistant 消息的持久化处是一个仅有 if 的分支，条件为 has_sendable_assistant_content（仅 Text/ToolUse 为真）。thinking‑only 流两者皆无，于是该 if 被跳过——既无消息也无事件——随后 if tool_uses.is_empty() 直接走到 break，回合在零事件下结束。维护者的 v0.8.40 审计（#1736）确认 #1727 在 v0.8.39 并未随版本发布、属于发布阻断项，且这一 status 分支至今缺失。

Fix / 修复

Minimum‑surface‑area (the issue author's "option 1"): add an else if arm at the same persist site that, only on a clean end (tool_uses empty, turn_error.is_none(), not cancelled), warns and emits an Event::status(...): "Model returned reasoning but no answer or tool call; turn ended without output. Send a follow‑up to retry." No assistant message is persisted (prior behavior preserved). No retry/re‑prompt/placeholder policy is added (explicitly out of scope). The guard ordering is extracted into a tiny pure helper should_emit_thinking_only_status(...) so it is unit‑testable and matches the existing should_hold_turn_for_subagents style. Reuses the exact surrounding patterns: Event::status, crate::logging::warn, self.cancel_token.is_cancelled(), turn_error.is_none().

中文： 最小改动面（issue 作者的“方案一”）：在同一持久化处新增 else if 分支，仅在干净结束时（tool_uses 为空、turn_error.is_none()、未取消）warn 并发出 Event::status(...)：“模型返回了推理但没有答案或工具调用；本回合无输出结束，可发送追问重试。” 不持久化任何 assistant 消息（保持原行为）；不加入重试/重提示/占位策略（明确超出范围）。守卫顺序被抽到一个纯函数小助手 should_emit_thinking_only_status(...)，便于单测，并与既有 should_hold_turn_for_subagents 风格一致。复用周边既有写法。

Scope / 改动边界

• The existing if has_sendable_assistant_content branch is untouched; normal text/tool turns persist and render exactly as before.
• No double‑notice: guarded by turn_error.is_none() (stream errors still surface their own error) and !cancelled.
• 既有 if 分支不动；正常文本/工具回合行为不变。受 turn_error.is_none()、!cancelled 守卫，不会重复提示。

Testing / 测试

New unit test thinking_only_turn_emits_status_only_on_clean_end pins the decision: notify only on clean end; suppressed when tool uses pending / an error already shown / cancelled.

cargo test -p deepseek-tui --bin deepseek-tui -- thinking_only_turn_emits_status_only_on_clean_end   # ok
cargo clippy -p deepseek-tui   # clean (no warnings/errors)

Honest test limitation: the test pins the extracted pure decision, not the full async handle_deepseek_turn loop — consistent with every existing turn‑loop test in this module (they all pin pure helpers; an end‑to‑end test would need a mock client/session/channel, outside a surgical fix). The else if wiring is covered by inspection.

新增单测 thinking_only_turn_emits_status_only_on_clean_end，与本模块既有“纯函数式”测试风格一致；端到端回合测试需 mock 客户端/会话/通道，超出本次外科式修复范围，已如实说明。

Refs #1727, #1736.