Capturing from Conversations

Answers worked out in conversation are easy to lose. When a Q&A or troubleshooting session settles something worth keeping, ask the Agent to capture it. What you bring becomes a source, the Agent plans which pages it affects, and the normal update flow turns it into reviewed knowledge.

flowchart LR
  qa["Q&A or discussion"] --> keep{"Worth keeping?"}
  keep -->|yes| src["Save as a source<br/>summary / document / code / MR"]
  src --> plan["Agent proposes<br/>affected pages"]
  plan --> upd["Update and Review"]
  upd --> kb["Approved knowledge"]

What can become a source

You have	It becomes
The conclusions of this conversation	A session summary, optionally linked to commits or merge requests
Text you paste or an excerpt you hold	A note
A new document	A document or Lark source, captured
Specific code or a merge request	An update of the registered repository to the merged commit

That explanation of the retry limits is right. Save this conversation as a
session summary linked to the merge request, and update the affected pages.

What the Agent does

1. Saves the material. A conversation is summarized — the problem, confirmed decisions, reasons and open questions — never stored as a transcript.
2. Proposes a plan: which existing pages to revise and whether a new page is warranted. A new topic goes through structure review first.
3. Runs the update with the saved source in the affected pages' scope, and brings the changed pages to Knowledge Review.

Saving a source alone starts no knowledge work, so you can also save now and update later. A summary that arrives after the code was already processed can still update a page; the code version is reused, not re-parsed.

When a question finds a gap

When querying shows that the knowledge is missing or out of date, the query Skill suggests an update with the affected page, the proposed change and the source scope — and starts nothing until you accept. On acceptance it hands the evidence to the production entry in the selected Context knowledge workspace. See Asking Questions.

Boundaries

• Only material you supply or authorize is captured; the Agent does not scan its own history.
• A linked merge request records an association; it does not prove the change was merged or tested.
• Saving a conversation or approving knowledge does not approve any upstream change discussed in it.

从对话中采集

在对话中得出的答案很容易丢失。当一次问答或排障过程确定了值得保留的内容时，请让 Agent 采集它。你提供的内容会成为来源，Agent 会规划它影响哪些页面，再由常规的更新流程把它变成经过审核的知识。

flowchart LR
  qa["问答或讨论"] --> keep{"值得保留吗？"}
  keep -->|是| src["保存为来源<br/>摘要 / 文档 / 代码 / MR"]
  src --> plan["Agent 提出<br/>受影响的页面"]
  plan --> upd["更新与审核"]
  upd --> kb["已批准的知识"]

哪些内容可以成为来源

你手头有	它会成为
本次对话的结论	一份会话摘要，可选择关联提交或合并请求
你粘贴的文本或手头的摘录	一条笔记
一份新文档	一个已采集的文档或飞书来源
特定代码或一个合并请求	把已登记的仓库更新到合并后的提交

That explanation of the retry limits is right. Save this conversation as a
session summary linked to the merge request, and update the affected pages.

Agent 会做什么

1. 保存材料。对话会被总结为问题、已确认的决定、原因和未决问题，而不会以完整记录的形式保存。
2. 提出计划：修订哪些现有页面，以及是否需要新建页面。新主题会先经过结构审核。
3. 将保存的来源纳入受影响页面的范围并运行更新，然后把改动的页面提交到知识审核。

仅保存来源不会启动任何知识工作，所以你也可以现在保存、稍后再更新。即使摘要在代码处理完之后才到达，也仍然可以更新页面；代码版本会被复用，不会重新解析。

当提问发现缺口时

当查询表明知识缺失或已过时，查询 Skill 会建议一次更新，并给出受影响的页面、建议的改动和来源范围；在你接受之前，它不会启动任何操作。接受后，它会把证据交给所选 Context 知识工作区中的生产入口。请参阅提问。

边界

• 只采集你提供或授权的材料；Agent 不会扫描自己的历史记录。
• 关联的合并请求只记录一种关联关系，并不能证明改动已经合并或经过测试。
• 保存对话或批准知识，并不代表批准其中讨论的任何上游改动。