Working with Agents

You drive Context through your Agent, not through the CLI. Invoke /c4a:context (Cursor: /c4a-context), describe the knowledge in normal language, and answer the decisions it brings back. The Agent reads the workflow's current Route from the CLI, runs the returned commands and writes the pages; the CLI validates sources and page identities, derives paths and owns Review, knowledge/ and dist/.

Ordinary coding or design conversations never start Context on their own; invoke it explicitly, or continue a conversation where it is already active.

Brief it well

The more the first message settles, the fewer questions follow. Cover:

• Readers and questions — who will use the knowledge and what they need to answer.
• Sources — repositories, modules, document folders or links, and what to skip.
• Output — a package for Agents, a website, or both, and the language.

/c4a:context Initialize a Context knowledge workspace in ./kb and build knowledge for
../billing-service and ../docs/billing. Readers are on-call engineers who need
to diagnose failed invoices. Skip the legacy-export module. Investigate first,
then give me the work-start report.

Answers you have already given are reused for the rest of the task, not asked again at every step.

What it asks you, and what it decides

You decide	The Agent decides
Which sources and modules are in scope	Which Indexer Skills to use
Which readers and questions matter	Topics, page boundaries and organization within the approved plan
Whether two subjects are the same knowledge owner	Wording and structure of each page
Whether a custom Indexer or executable extension is acceptable	Mechanical steps returned by the Route
Whether the outline and the candidate pages are approved	—

Even in fully managed mode the Agent asks when missing purpose or scope would change the outcome. Before indexing a large boundary it identifies obsolete or irrelevant areas with you from the actual material, rather than guessing from names.

Following progress

Production is the long phase — roughly a minute per article task on average. The Agent reports progress against the task total, and you can ask for a local preview at any time. Transition reports stay short: what changed, the current state and the next decision.

Look at the work-start report and the first changed pages early. If something is off — tone, a counterexample, a style rule — say so. The earlier a correction lands, the better the final result.

Adding material or changing course

Say it in the conversation:

Add ../docs/refunds to the sources, re-plan the affected part and give me a
new work-start report.

New material within the confirmed purpose becomes a plan amendment that keeps completed work; a new purpose or source still goes back through your confirmation. To correct a page, see Revising a Page.

Interruptions and resuming

Workflow state lives in the knowledge workspace, not in the conversation. After a closed session or a crash, open a new conversation in the same knowledge workspace and invoke /c4a:context; the Agent reads the current Route and continues. Accepted tasks are never resubmitted, and deleting temporary runtime files is not a way to restart — it does not authorize replaying accepted work.

Several Agents in parallel

When your host can coordinate child Agents, the work can be split across them. The CLI decides which batches are eligible; each worker writes only its assigned drafts, and one coordinator submits everything to Context. Without that capability the Agent uses a single batch — no simulated workers. Parallelism never waives a decision that is yours.

Git and publishing

Completing Context work does not authorize Git operations. The Agent stages, commits, pushes, publishes or deploys only when you ask.

与 Agent 协作

你通过 Agent 驱动 Context，而不是直接操作 CLI。调用 /c4a:context（Cursor 中为 /c4a-context），用自然语言描述你需要的知识，并回答它带回来的决策问题。Agent 从 CLI 读取工作流当前的 Route，执行返回的命令并撰写页面；CLI 负责校验来源和页面身份、推导路径，并掌管知识审核、knowledge/ 和 dist/。

普通的编码或设计对话不会自行启动 Context；你需要显式调用它，或在 Context 已经激活的对话中继续。

把任务交代清楚

第一条消息确定的内容越多，后续的问题就越少。请覆盖以下几点：

• 读者和问题：谁会使用这些知识，他们需要回答哪些问题。
• 来源：仓库、模块、文档文件夹或链接，以及需要跳过的内容。
• 产出：面向 Agent 的知识包、网站，还是两者都要，以及使用的语言。

/c4a:context Initialize a Context knowledge workspace in ./kb and build knowledge for
../billing-service and ../docs/billing. Readers are on-call engineers who need
to diagnose failed invoices. Skip the legacy-export module. Investigate first,
then give me the work-start report.

你已经给出的回答会在整个任务中复用，不会在每一步重复询问。

哪些会问你，哪些由它决定

由你决定	由 Agent 决定
哪些来源和模块在范围内	使用哪些 Indexer 技能
哪些读者和问题是重点	在已批准的计划内划分主题、页面边界和组织方式
两个主题是否属于同一个知识归属	每个页面的措辞和结构
是否接受自定义 Indexer 或可执行扩展	Route 返回的机械步骤
大纲和候选页面是否批准	—

即使在全托管模式下，如果缺失的目的或范围会改变结果，Agent 也会询问你。在对一个大范围建立索引之前，它会根据实际材料与你一起识别过时或无关的部分，而不是凭名称猜测。

跟进进度

生产是耗时最长的阶段，平均每个文章任务大约一分钟。Agent 会对照任务总数汇报进度，你也可以随时要求本地预览。阶段切换时的汇报保持简短：发生了什么变化、当前状态以及下一个决策。

尽早查看开工报告和最先改动的页面。如果有不对的地方，比如语气、某个反例或某条风格规则，直接指出。纠正越早，最终结果越好。

补充材料或调整方向

直接在对话中说明：

Add ../docs/refunds to the sources, re-plan the affected part and give me a
new work-start report.

在已确认目的范围内的新材料会成为一次计划修订，已完成的工作会保留；新的目的或来源仍需回到你这里确认。要修正某个页面，请参阅修订页面。

中断与恢复

工作流状态保存在知识工作区中，而不是对话里。会话关闭或崩溃后，在同一个知识工作区中打开新对话并调用 /c4a:context，Agent 会读取当前 Route 并继续。已接受的任务永远不会重新提交；删除临时运行时文件也不是重新开始的方式，这样做并不会授权重放已接受的工作。

多个 Agent 并行

当你的宿主能够协调子 Agent 时，工作可以拆分给它们。CLI 决定哪些批次可以并行；每个 worker 只撰写分配给它的草稿，由一个协调者统一提交给 Context。如果宿主不具备这种能力，Agent 会使用单个批次，不会模拟 worker。并行永远不会免除本应由你做出的决定。

Git 与发布

完成 Context 工作并不代表授权进行 Git 操作。只有在你要求时，Agent 才会暂存、提交、推送、发布或部署。