feat: 新增用户补充指引功能，含队列管理与 UI 交互

[xinggitxing]

概述

新增"补充指引"机制，允许用户在 LLM 正在处理提示词时（不中断当前轮次）发送额外指令。这实现了"边执行边引导"的交互模式，用户可输入指引文本进入队列，在下次 LLM API 调用前自动注入。

动机

此前当 LLM 忙碌时（例如在多步骤任务中执行工具调用），用户只能等待完成或直接中断智能体来发送新指令。本功能引入了一个非侵入式的补充消息队列：用户可以在智能体工作时输入并提交指引内容，系统会
将这些消息作为 system 消息在下次 LLM API 调用前注入（跳过总结阶段）。

核心变更

会话层（src/session.ts）

• 新增 PendingSupplementary 类型和 MAX_SUPPLEMENTARY_QUEUE（每会话上限 10 条）
• MessageMeta.isSupplementary 标记用于消息渲染
• 新增方法：addSupplementaryMessage、cancelSupplementaryMessage、countPendingSupplementary、listPendingSupplementary、isInSummaryPhase
• 私有方法 flushSupplementaryMessages()：将队列条目转换为 system 角色的 SessionMessage[]，内容前缀为 [User Supplementary Guidance]\n
• 补充消息在 activateSession() 中每次 LLM 调用前注入，但仅在 非总结阶段（上次 LLM 响应包含 tool_calls 时）
• isSummarizing 标记跟踪 LLM 返回是否包含 tool_calls

UI — App（src/ui/App.tsx）

• 新增状态：supplementaryCount、supplementaryList、isSummarizing
• onSupplementaryStatusChanged 回调将队列同步到 React 状态
• 忙碌结束逻辑自动将未刷新的补充文本回填到 PromptInput
• PromptInput 始终保持挂载（通过 height: 0 隐藏），防止视图切换时缓冲区丢失

UI — PromptInput（src/ui/PromptInput.tsx）

• 补充消息列表 UI，支持上下导航、退格取消、回车提交
• isSummarizing 时阻止输入，显示"等待总结完成..."
• 忙碌但 非 总结时，回车提交内容作为补充指引
• Prompt 草稿合并逻辑：将现有缓冲区文本与补充草稿合并

UI — MessageView（src/ui/components/MessageView/index.tsx）

• 为 message.meta?.isSupplementary 新增渲染分支：黄色边框的 [Supplementary Guidance] 信息块

测试（src/tests/session.test.ts）

新增 10 个测试用例：

• addSupplementaryMessage 队列添加并返回 ID
• addSupplementaryMessage 队列满时返回 null
• cancelSupplementaryMessage 移除特定消息
• cancelSupplementaryMessage 空会话返回 false
• flushSupplementaryMessages 返回正确角色和前缀的 system 消息
• 补充队列按会话隔离
• isInSummaryPhase 初始返回 false
• 补充列表返回副本（不可变）
• onSupplementaryStatusChanged 在添加和取消时被调用

文件变更

文件	变更量
src/session.ts	+100 行
src/tests/session.test.ts	+166 行（10 个测试）
src/ui/App.tsx	+121 / -33 行
src/ui/PromptInput.tsx	+112 / -33 行
src/ui/components/MessageView/index.tsx	+15 行

测试结果

• ✅ npm run check — 类型检查、代码规范、格式检查全部通过
• ✅ 10 个新增测试用例全部通过

[xinggitxing]

Issues #113 【TUI】不打断任务的情况下，补充发送引导信息。

[mulyawansentosa]

Implemented the "supplementary guidance" mechanism described in this issue (and #113):
<PR_LINK>

Two layers, so a new instruction never has to wait for the previous one:

1. Mid-turn injection. The prompt is queued on the session and, right before every LLM call of
   that same turn, it is appended to the conversation as a user message (with images and
   selected skills). The model reads it together with the work it already did and can revise or
   supersede the earlier instructions:

   user:      install nginx
   assistant: <tool_calls>
   tool:      <result>
   user:      stop, use apache instead      ← guidance, read by the next LLM call
   assistant: <revised plan>
   

2. Steering a streaming answer. Injection can only happen at a request boundary, so a long
   text answer would still block it. When a prompt arrives while an answer is streaming, that one
   request is aborted, the text produced so far is kept as an assistant message, and the loop
   immediately issues a new request carrying the guidance:

   user:      explain how the deploy works
   assistant: The deploy runs in three stag…       ← cut short
   user:      no, just tell me the rollback step   ← read straight away
   

   Steering is opt-in: Ctrl+Enter steers a single message, and steerMode: "interrupt" in
   settings.json makes plain Enter steer too ("queue" is the default). Tools are never
   interrupted: steering only targets the model stream, so a running command finishes and the
   guidance is injected at the next request boundary.

If a turn would otherwise end while guidance is still queued, the loop stays alive for one more
iteration, so pending guidance is never silently dropped or attached to a later, unrelated prompt.
Guidance is not injected while the session waits for the user (an ask_user_question answer or a
permission decision), so it can never be consumed as that answer. Esc still interrupts the turn.

Notes compared to the design sketched here:

1. The injected message uses role: "user" with meta.isSupplementary, not a system reminder.
   A user message is the strongest signal for superseding earlier instructions, it keeps the
   prompt verbatim in the transcript, and it carries image attachments natively. system remains
   an option if you prefer the softer "guidance" framing.
2. The implementation is split into two commits (injection, then steering) so the turn-boundary
   behaviour can be reviewed on its own.

New tests: 9 core tests (queue semantics, the actual request payloads of a turn that receives
guidance mid-flight, the aborted stream with its follow-up request, and the no-op case) plus 4 Ink
integration tests for the input wiring. npm run check and npm test (333 CLI + 392 core + 59
VS Code tests) pass.

中文说明：AI 正在回复时，Enter 不再被拒绝，也不必等到本轮结束。实现分两层：(1) 消息进入会话队列，
在本轮每次 LLM 调用前作为 user 消息（含图片与 skills）注入，模型能结合已完成的工作修改甚至推翻
之前的指令；(2) 如果模型正在流式输出回答，则该次请求会被中断，已生成内容保留为 assistant 消息
（标记 meta.interrupted），并立即发起携带新指令的新请求。正在执行的命令不会被中断，steering 只作用
于模型输出流。steering 默认是 opt-in：Ctrl+Enter 立即转向；steerMode 默认 "queue"，设为
"interrupt" 后普通 Enter 也会截断当前输出。