Preparing a Workspace

Preparing a Context knowledge workspace leaves it with usable registered sources and no unfinished task, while keeping approved knowledge and configuration. Ask for it explicitly — for example "prepare this workspace for the next task". A clear request authorizes the task-state cleanup it describes; the Agent asks only about an unresolved target or a loss you have not agreed to. The switch that governs this is context.gate.workspace_reset_restore.

flowchart LR
  scope["Establish the scope"] --> finish_old["End the old task<br/>task prepare preview, then apply"]
  finish_old --> sources["Restore usable sources"]
  sources --> ready["Verify and stop"]

When you only want to capture documents

Capturing named documents and pausing is not preparation and does not start article production. The Agent registers the sources and stops after reporting what was captured; existing tasks, drafts, approved content and registrations are untouched, and no task resume is needed. Production resumes only when you ask for knowledge.

1. Establish the scope

The Agent locates the workspace with context entry, then looks at current work, running processes, Git changes and source registrations, and tells you which unfinished drafts and queued maintenance will be discarded. If a writer is still running, it waits for the result first. A broken status is investigated, not taken as proof that the workspace is empty.

Registered sources are not an active task. Preparation does not repair invalid snapshots or change which sources are registered, and leaving a source out of one task never deletes its registration or evidence.

2. End the old task

context task prepare --format json

The command first previews exactly which Context-owned task files — production, Review and maintenance state — would be removed. Once you have agreed to discard that scope, the Agent runs the returned apply command with its plan digest. If interrupted, it reruns the preview and uses the resume command it returns.

Removed	Always kept
Unfinished production stages, drafts, Review state and queued maintenance	Approved pages, source records and snapshots, project configuration, repository checkouts, other .tmp files and existing outputs

Other caches are optional cleanup. Keep reports, unique material, unknown files and modified checkouts unless you agree to lose them; never delete locks, transaction records or active tool directories. Keep repository checkouts used by registered sources — removing them forces a costly clone on the next update — and keep source-region baselines, which let updates tell moved text from changed text.

Prepare after a successful delivery rather than straight after close: recording the version, building and retries may still need the current task. Delivery cleanup already removes completed drafts and Review state.

3. Restore usable sources

Source	How it is restored
Repositories	context source recovery-plan --format json lists what is missing; a matching local checkout is reused, or context source restore --input <file> --format json clones it when authorized. Modules sharing a remote and commit are cloned once, at the registered commit — never a newer one. A dirty checkout you supplied is never reset.
Lark documents	The stored body and required attachments are checked against the source records and repaired with the document tools. If the current remote differs from the old snapshot, that is a source update to decide, not a recovery.
Notes, session summaries, local files	Formal source content is kept and its paths checked. Lost originals are never reconstructed from generated knowledge.

4. Verify and stop

The workspace is ready when the preview shows no task files, no writer is running, the configuration loads, and the selected sources resolve to their registered versions and paths. The Agent reports any blockers with their next action and then stops — it does not start writing or recapture everything. To go back to an older version instead, see Restoring a Version.

准备知识工作区

准备 Context 知识工作区后，它会处于已登记来源可用、没有未完成任务的状态，同时保留已批准的知识和配置。你需要明确提出请求——例如“为下一个任务准备这个知识工作区”。清晰的请求即授权了其中描述的任务状态清理；Agent 只会在目标不明确，或涉及你尚未同意的损失时才询问。控制这一行为的开关是 context.gate.workspace_reset_restore。

flowchart LR
  scope["确定范围"] --> finish_old["结束旧任务<br/>task prepare 先预览再应用"]
  finish_old --> sources["恢复可用来源"]
  sources --> ready["验证并停止"]

只想采集文档时

采集指定文档后暂停不属于准备，也不会启动文章生产。Agent 登记来源，报告采集结果后即停止；现有任务、草稿、已批准内容和登记信息都不受影响，也不需要 task resume。只有当你请求知识时，生产才会继续。

1. 确定范围

Agent 用 context entry 定位知识工作区，然后查看当前工作、运行中的进程、Git 变更和来源登记情况，并告诉你哪些未完成的草稿和排队中的维护会被丢弃。如果仍有写入进程在运行，它会先等待其结果。status 出错时会被排查，而不会被当作知识工作区为空的证据。

已登记的来源不是活动任务。准备不会修复无效快照，也不会改变登记了哪些来源；某个来源没有纳入一次任务，也绝不会删除它的登记信息或证据。

2. 结束旧任务

context task prepare --format json

该命令首先预览将被删除的、确切属于 Context 的任务文件——生产、Review 和维护状态。在你同意丢弃这一范围后，Agent 会带上计划摘要运行返回的 apply 命令。如果中途被打断，它会重新运行预览，并使用其返回的恢复命令。

删除	始终保留
未完成的生产阶段、草稿、Review 状态和排队中的维护	已批准页面、来源记录和快照、项目配置、仓库检出、其他 .tmp 文件和已有产出

其他缓存的清理是可选的。除非你同意舍弃，否则保留报告、独有材料、未知文件和已修改的检出；绝不删除锁、事务记录或活动中的工具目录。保留已登记来源所用的仓库检出——删掉它们会让下次更新不得不重新进行代价高昂的克隆——并保留来源区段基线，更新时靠它区分文本是被移动还是被修改。

应在交付成功之后再准备，而不是 close 之后马上准备：记录版本、构建和重试可能仍需要当前任务。交付清理已经会删除已完成的草稿和 Review 状态。

3. 恢复可用来源

来源	恢复方式
仓库	context source recovery-plan --format json 列出缺失的内容；匹配的本地检出会被复用，或在获得授权时由 context source restore --input <file> --format json 克隆。共享同一远程和提交的模块只克隆一次，并检出到登记的提交——绝不会是更新的提交。你提供的有未提交修改的检出绝不会被重置。
飞书文档	对照来源记录检查已存储的正文和必需的附件，并用文档工具修复。如果当前远程内容与旧快照不同，这属于需要决定的来源更新，而不是恢复。
笔记、会话摘要、本地文件	保留正式的来源内容并检查其路径。丢失的原文绝不会根据生成的知识重建。

4. 验证并停止

当预览显示没有任务文件、没有写入进程在运行、配置可以加载，且所选来源都解析到其登记的版本和路径时，知识工作区就准备好了。Agent 会报告所有阻塞项及其下一步操作，然后停止——它不会开始编写，也不会重新采集全部内容。如果想回到旧版本，参见恢复版本。