Skip to content

Reading Review Reports ​

Nothing the Agent writes becomes knowledge until it passes Knowledge Review. The Agent opens one HTML report for the whole current batch of candidate pages and waits for your answer. This report decides both the scope and the quality of what gets published, so it is worth reading with three questions in mind.

What is in the report ​

PartWhat it shows
Change overviewThe batch at a glance, with the existing navigation and the expected file tree of the Context knowledge workspace
NavigationWhere each page will sit; pages without a placement are flagged so the Agent can bind them before asking you
PagesEach candidate's title, path, summary and full rendered content, including diagrams and source citations. New pages are marked New; changed pages show the changed blocks next to the previous text
Revision notesA note box per page; typing in it marks the page for revision

Opening a page marks it READ in your browser. Read counts and notes are local conveniences, not approval. The revision counter lists every page with notes, and Copy exports them as plain text with article IDs and titles for pasting into the conversation.

Three things to judge ​

  1. Structure — Are the topics, directories and pages divided the way readers will look for them?
  2. Content — Do the pages carry the intended purpose, depth and style? Can a reader actually complete the promised task with them?
  3. Scope — Is related material missing that should be included, or is something here that should be dropped?

Citations and placements alone do not prove a page is useful. Spot-check a few pages against their sources, especially defaults, conditions and steps.

Answering ​

mermaid
flowchart LR
  rep["Review report"] --> a{"Your reply"}
  a -->|approve| ok["Approved into knowledge/"]
  a -->|notes only| fix["Agent repairs"]
  fix --> rep2["Updated report"]
  rep2 --> a
  a -->|notes + approve after fixing| fix2["Agent repairs and verifies"]
  fix2 --> ok

Reply in the conversation; no review code or magic phrase is needed.

You sayWhat happens
Approve, or continueThe presented batch is approved
Revision notes onlyThe Agent repairs those pages, checks the result and shows an updated report
Notes plus "approve after fixing"The Agent repairs, verifies the fix, and approves only the agreed pages without asking again
"Fix it and show me"Editing only; it waits for your decision on the new report

Silence, an ignored card or read progress is never approval. Rejecting a page leaves it out of this batch; it does not retire an already approved article.

Taking part of a batch now ​

You can approve the pages that are ready while others go back for repair. Approval alone does not build; to see the approved pages published before the rest are finished, ask for an early delivery. The Agent runs close and build for them, as long as they do not link to pages still pending, and then returns to the remaining review.

After approval ​

Approval writes the pages to knowledge/. Close then rebuilds knowledge/structure.yaml and verifies the approved knowledge without rewriting its prose, and build produces the package and site. Nothing is published before approval.

阅读审核报告 ​

Agent 写的任何内容,在通过知识审核之前都不会成为知识。Agent 会为当前整批候选页面打开一份 HTML 报告,并等待你的答复。这份报告同时决定了发布内容的范围和质量,因此值得带着三个问题来阅读。

报告里有什么 ​

部分展示内容
变更概览本批次的概况,包括现有导航和 Context 知识工作区的预期文件树
导航每个页面将被放在哪里;没有位置的页面会被标出,以便 Agent 在询问你之前先为它们绑定位置
页面每个候选页面的标题、路径、摘要和完整的渲染内容,包括图表和来源引用。新页面标记为 New;改动的页面会在旧文本旁显示改动的块
修订意见每个页面都有一个意见框;在其中输入内容即把该页面标记为待修订

打开某个页面后,浏览器会将其标记为 READ。已读计数和意见只是本地的便利功能,并不代表批准。修订计数器会列出所有带意见的页面,Copy 会把它们连同文章 ID 和标题导出为纯文本,方便粘贴到对话中。

需要判断的三件事 ​

  1. 结构:主题、目录和页面的划分方式,是否符合读者查找内容的习惯?
  2. 内容:页面是否承载了预期的目的、深度和风格?读者真的能借助它们完成页面所承诺的任务吗?
  3. 范围:是否缺少应当纳入的相关材料,或者包含了应当去掉的内容?

仅有引用和位置并不能证明页面有用。请对照来源抽查几个页面,尤其是默认值、条件和步骤。

答复 ​

mermaid
flowchart LR
  rep["审核报告"] --> a{"你的答复"}
  a -->|批准| ok["批准进入 knowledge/"]
  a -->|仅修订意见| fix["Agent 修复"]
  fix --> rep2["更新后的报告"]
  rep2 --> a
  a -->|意见 + 修复后批准| fix2["Agent 修复并验证"]
  fix2 --> ok

直接在对话中答复即可,不需要审核码或特定口令。

你说会发生什么
批准,或继续当前呈现的批次获得批准
仅给出修订意见Agent 修复这些页面,检查结果并展示更新后的报告
修订意见加上“修复后批准”Agent 修复并验证修复结果,然后只批准约定的页面,不再重复询问
“修好再给我看”只做编辑;它会等待你对新报告做出决定

沉默、忽略卡片或阅读进度都不等于批准。拒绝某个页面只会把它排除在本批次之外,不会让已批准的文章退役。

先接收批次中的一部分 ​

你可以先批准已经就绪的页面,同时把其他页面退回修复。仅批准并不会触发构建;如果想在其余页面完成之前先看到已批准页面发布,请要求提前交付。只要这些页面没有链接到仍在等待中的页面,Agent 就会为它们执行 close 和 build,然后回到剩余的审核。

批准之后 ​

批准会把页面写入 knowledge/。随后 close 会重建 knowledge/structure.yaml 并验证已批准的知识,但不会改写其正文;build 则生成知识包和站点。批准之前不会发布任何内容。