Context for AI

Appearance

Scheduled Updates ​

Context itself never polls for changes. To keep knowledge current without anyone asking, schedule an Agent run: at a fixed time, the Agent checks the registered sources for changes since the last processed baseline, prepares an update for what changed and stops where a person must decide. Use whatever scheduler your Agent environment provides; the instructions are what matter.

mermaid
flowchart LR
  timer["Scheduled trigger"] --> check["Agent checks upstream<br/>documents and code"]
  check -->|no change| report["Short report, stop"]
  check -->|changes| upd["Update plan and<br/>changed pages"]
  upd --> review["Knowledge Review<br/>waits for a person"]

Write the task instructions ​

A scheduled prompt runs unattended, so it must say what a person would otherwise answer:

  • Scope — which Context knowledge workspace and which sources or modules to check.
  • No change — what to do when nothing changed, such as a one-line report.
  • When to stop — which situations need a person: Knowledge Review, new sources, failed reads.
  • Delivery — whether to build, and whether anything may be published.
text
/c4a:context In ./kb, check billing-service and the billing docs for changes
since the last processed version. If nothing changed, report "no change" and
stop. Otherwise prepare the update and stop at Knowledge Review. Do not
publish or commit.

Delegating review — or not ​

By default the run stops at Knowledge Review and a person reviews the changed pages. If the automation is trusted to review on its own, put the policy block in the instructions it actually delivers to the Agent:

yaml
CONTEXT_RUN_POLICY:
  knowledge_review: delegate

Delegation applies to that task only, and it does not widen the source scope, authorize publishing or bypass Gates that never delegate. See Managed Mode and Review Policy.

What each run does ​

Each run follows the ordinary update flow: the Agent acquires the new fixed versions with its own Git and document tools, submits the update, revises only affected pages and brings them to review. Unchanged module trees produce a "no knowledge change" conclusion instead of new pages. A source that cannot be read is reported as a gap; it is never treated as a change or as a reason to remove a page.

定时更新 ​

Context 本身从不轮询变更。要在无人提出请求的情况下保持知识最新，可以安排一次定时的 Agent 运行：在固定时间，Agent 检查已登记的来源自上次已处理基线以来的变更，为发生变化的部分准备更新，并在需要人来决定的地方停下。使用你的 Agent 环境提供的任意调度器即可；关键在于任务指令。

mermaid
flowchart LR
  timer["定时触发"] --> check["Agent 检查上游<br/>文档和代码"]
  check -->|无变化| report["简短报告，停止"]
  check -->|有变化| upd["更新计划和<br/>变更的页面"]
  upd --> review["知识审核<br/>等待人来决定"]

编写任务指令 ​

定时提示词在无人值守时运行，因此必须写明原本需要人来回答的内容：

  • 范围——检查哪个 Context 知识工作区，以及哪些来源或模块。
  • 无变化——什么都没变时怎么做，例如输出一行报告。
  • 何时停下——哪些情况需要人：知识审核、新来源、读取失败。
  • 交付——是否构建，以及是否允许发布任何内容。
text
/c4a:context In ./kb, check billing-service and the billing docs for changes
since the last processed version. If nothing changed, report "no change" and
stop. Otherwise prepare the update and stop at Knowledge Review. Do not
publish or commit.

委托审核——或者不委托 ​

默认情况下，运行会在知识审核处停下，由人审核变更的页面。如果信任该自动化自行审核，请把策略代码块放进它实际交给 Agent 的指令中：

yaml
CONTEXT_RUN_POLICY:
  knowledge_review: delegate

委托只对该任务生效，并且不会扩大来源范围、授权发布或绕过永不委托的 Gate。参见全托管模式与审核策略。

每次运行做什么 ​

每次运行都遵循普通的更新流程：Agent 用自己的 Git 和文档工具获取新的固定版本，提交更新，只修订受影响的页面，并将它们送去审核。未变化的模块树会得出“无知识变更”的结论，而不是生成新页面。无法读取的来源会作为缺口报告；它绝不会被视为变更，也不会成为删除页面的理由。