功能需求：Kimi Work 会话内支持对 Agent 回复的可视化批注与审阅反馈 || Functional requirements: Kimi Work session supports visual annotations and review feedback on Agent responses

功能需求：Kimi Work 会话内支持对 Agent 回复的可视化批注与审阅反馈

提交对象：Kimi Work / Kimi Code 团队
提交渠道（备选）：Kimi 客户端「账号 → 设置 → 意见反馈」；code@moonshot.ai；Kimi CLI GitHub Issues
起草日期：2026-09-14

一句话需求

希望 Kimi Work 支持对 Agent 的任意一条回复（尤其是计划、报告、方案类长回复）进行逐段可视化批注，批注结果以结构化形式返回给 Agent 用于修订，而不是只能靠用户在输入框里用文字描述修改意见。

用户场景与痛点

1. Agent 产出长回复（方案分析、计划、报告草稿）后，用户想对其中某几段提出修改意见。目前只能打字描述"第三部分第二点改成……"，定位不精确、沟通成本高。
2. 在计划/方案确认环节，用户希望能有 Approve / Reject with feedback 的闸门式交互：批准后 Agent 继续执行，驳回时附带逐条批注。
3. 这类"人在回路的审阅层"是长程 Agent 工作流（Goal Mode、多 Agent 编排）的关键缺失环节——任务越长，纯文字纠偏的成本越高。

期望的功能形态（按优先级）

1. 回复批注入口：每条 Agent 回复上提供"批注"按钮，打开一个批注视图，支持选中文本添加评论、整体批准/驳回。
2. 结构化回传：批注结果（引用的原文位置 + 评论内容 + 批准/驳回决定）作为结构化数据注入对话，Agent 能据此精确定位并修订。
3. （可选）开放会话消息读取接口：若短期内不做内置 UI，希望 Kimi Work 能导出或暴露会话消息的可读格式（如标准 transcript JSONL），便于第三方审阅工具接入。

参考实现

开源工具 Plannotator（本地 localhost 批注 UI + hook 集成）已为 Claude Code、Codex CLI、OpenCode、Gemini CLI、Kimi Code CLI 实现了上述能力：

• plannotator last：对宿主 CLI 会话中最新一条助手消息打开批注 UI，批注结果回传 stdout；
• 计划审批通过宿主 hook 在退出 plan mode 时自动弹出。

现状佐证：在 Kimi Work 桌面端会话中运行 plannotator last 报错
No rendered assistant message found in session logs——因为桌面端会话存储
（~/Library/Application Support/kimi-desktop/daimon-share/.../wire.jsonl）
的格式与 Kimi Code CLI 的会话日志格式不同，第三方工具无法读取。

这说明需求真实存在且已有成熟范式：只要 Kimi Work 提供内置批注 UI，或将消息导出为与 Kimi Code CLI 一致的可读格式，即可打通。

用户价值

• 大幅降低长回复的纠偏成本，提升 Goal Mode / 长程任务的可用性；
• 让"审阅—批准—继续"成为一等交互，与 Kimi Work 的权限审批体系（三档授权）形成互补；
• 开放消息格式还能带动第三方审阅/批注工具生态。

---

Functional requirement: Kimi Work session supports visual annotations and review feedback on Agent responses.

Submit to: Kimi Work / Kimi Code team
Submission channel (optional): Kimi client "Account → Settings → Feedback"; code@moonshot.ai; Kimi CLI GitHub Issues
Drafting date: 2026-09-14

One sentence requirement

It is hoped that Kimi Work will support paragraph-by-paragraph visual annotation for any of the Agent's replies (especially long replies such as plans, reports, and plans), and the annotation results will be returned to the Agent in a structured form for revision, instead of relying on the user to describe modification opinions in text in the input box.

User scenarios and pain points

1. After the Agent generates a long reply (program analysis, plan, report draft), the user wants to make modifications to certain paragraphs. At present, we can only type the description "Change the second point of the third part to...". The positioning is inaccurate and the communication cost is high.
2. In the plan/solution confirmation process, users hope to have a gate-type interaction of Approve / Reject with feedback: after approval, the Agent will continue to execute, and when rejecting, it will be accompanied by one-by-one comments.
3. This type of "human-in-the-loop review layer" is the key missing link in long-range Agent workflows (Goal Mode, multi-Agent orchestration) - the longer the task, the higher the cost of pure text correction.

Desired functional form (by priority)

1. Reply Batch Inlet: Each Agent reply provides a "Comment" button to open a comment view, which supports adding comments to selected text and overall approval/rejection.
2. Structured postback: The annotation result (original text location of the quote + comment content + approval/rejection decision) is injected into the conversation as structured data, and the Agent can accurately locate and revise it accordingly.
3. (Optional) Open session message reading interface: If there is no built-in UI in the short term, we hope that Kimi Work can export or expose the readable format of session messages (such as standard transcript JSONL) to facilitate the access of third-party review tools.

Reference implementation

The open source tool Plannotator (local localhost annotation UI + hook integration) has implemented the above capabilities for Claude Code, Codex CLI, OpenCode, Gemini CLI, Kimi Code CLI:

• plannotator last: Open the annotation UI for the latest assistant message in the host CLI session, and the annotation result is returned to stdout;
• Plan approval pops up automatically when exiting plan mode through host hook.

Situation evidence: Running plannotator last in a Kimi Work desktop session reports an error
No rendered assistant message found in session logs - because the desktop session is stored
(~/Library/Application Support/kimi-desktop/daimon-share/.../wire.jsonl)
The format is different from the session log format of Kimi Code CLI and cannot be read by third-party tools.

This shows that the demand is real and there is a mature paradigm: as long as Kimi Work provides a built-in annotation UI, or exports messages into a readable format consistent with Kimi Code CLI, it can be solved.

User value

• Significantly reduce the correction cost of long replies and improve the availability of Goal Mode/long-distance tasks;
• Make "review-approve-continue" a first-class interaction, complementing Kimi Work's permission approval system (three levels of authorization);
• Open message formats can also drive an ecology of third-party review/annotation tools.

Start by inspecting the Kimi Work session storage path mentioned in the issue, especially the wire.jsonl files, and compare it with the Kimi Code CLI session format. Use the existing plannotator last behavior and Kimi Code CLI integration as reference. Done should mean a defined, testable path for visual annotations and structured review feedback, or a readable session export if that scope is chosen.