dsh-plugin-dev-kb 1.3.0 → 1.4.0

package/kb/extra/subsystems/browser-use.md

@@ -0,0 +1,67 @@
+# Browser use
+
+English | [中文](browser-use.zh.md)
+
+Browser use lets a model inspect and operate web pages through a configured backend. DSH owns the task loop; the provider supplies browser operations and keeps browser state across turns of one live Session.
+
+## Choose a provider
+
+Mount [`dsh-browser-use`](../../packages/browser-use/browser-use/README.md) and one provider in the same composition. The providers are experimental public npm packages and require explicit activation. Their initial browser engine is Chromium.
+
+| Provider | Integration |
+|---|---|
+| [Playwright MCP](../../packages/experimental/browser-use-playwright-mcp/README.md) | Playwright's browser-control MCP tools |
+| [Chrome DevTools MCP](../../packages/experimental/browser-use-chrome-devtools-mcp/README.md) | Chrome DevTools inspection and control through MCP |
+| [Stagehand](../../packages/experimental/browser-use-stagehand-native/README.md) | Native browser operations with AI-assisted actions, observation, and extraction |
+
+The shared service registers only a name and rejects any second provider, including another instance with the same name. It has no common browser-operation methods, browser resources, or model-controlled selector. Provider configuration in a profile or preset selects launch or attachment for that activation.
+
+## Session ownership
+
+A launched browser belongs to the exact live Agent and Session that uses it. Calls across turns reuse that browser. Disposing the Session runtime closes its launched resources; reloading or forking a Session starts fresh browser state. Browser profiles and login state are not restored from the Session log.
+
+An attached browser remains externally owned. The provider reserves it for one Session within that provider instance, preserves its existing browser state, and rejects simultaneous attachment by another Session. Teardown disconnects and leaves the external browser running. Separate DSH processes and other clients remain outside this reservation.
+
+Provider shutdown stops tool admission and waits for owned work and resource cleanup before releasing the shared provider registration. Cancellation cannot undo a browser action already delivered.
+
+## MCP initialization
+
+An MCP provider initializes one client for each live Agent created after the provider loads. The existing serial `agent/created` event awaits connection and discovery before creation or resume completes and queued input runs. The client remains with the Session across turns. Startup failure or cancellation rejects creation or resume and triggers client cleanup.
+
+If an attachment is busy, that activation continues without the browser and does not retry on later turns. After release, a newly created or resumed activation can acquire it. Loading or reloading the provider does not adopt already active Sessions; the [shared runtime](../../packages/experimental/browser-use-runtime/README.md) owns these initialization rules.
+
+## Tools and recorded results
+
+Provider tools use the normal DSH execution pipeline and Session log. The providers own their tool schemas, result rendering, image support, configuration, and upstream limitations; the shared service adds no model-visible content. Stagehand's AI-assisted operations use its explicitly configured native model while DSH retains the task loop. DSH model routing, credential reuse, underlying inference request/response capture, and integration into Session usage accounting are deferred; returned SDK data and metadata remain ordinary logged tool results.
+
+Browser MCP connections also expose [resources and server instructions](mcp.md). Resource calls addressed to a browser server use its Session queue and reject other Sessions; server instructions are assembled only for its owning Session.
+
+The [decision record](../../.agents/notes/implemented/architecture/2026-09-12-browser-use-provider-registration.md) explains the registration-only service and per-Session ownership.
+
+<!-- BEGIN GENERATED cordis-surface (gen-cordis-catalog.ts) — do not edit between markers -->
+
+<a id="cordis-surface"></a>
+
+## Cordis API
+
+Generated from source by `scripts/gen-cordis-catalog.ts` (verified fresh by `pnpm run verify-cordis-catalog` in doc-sync; regenerate with `pnpm run gen-cordis-catalog`) — the language sides differ only in locale-specific paired document paths. Signature blocks use a `ts cordis-catalog` fence and keep the original source JSDoc; dispatch modes are defined in the [primer](../cordis-primer.md#dispatch-modes), and the framework-inherited `ctx` API lives in [cordis-api/inherited.md](../cordis-api/inherited.md).
+
+<a id="ctxbrowseruse--browseruseregistry"></a>
+
+### `ctx.browserUse` — `BrowserUseRegistry`
+
+Owns one optional provider registration in the shared browser-use service.
+
+```ts cordis-catalog
+/**
+ * Reserve the sole provider slot until the contribution is disposed.
+ * A second registration fails even when it repeats the current name. Providers
+ * must stop their tools and await owned work before releasing this registration.
+ * @param name - provider-owned name used in registration diagnostics.
+ * @returns the effect disposer for this exact registration.
+ */
+register(name: BrowserUseProviderName): () => Promise<void>
+```
+
+Source: [`packages/browser-use/browser-use/src/index.ts`](../../packages/browser-use/browser-use/src/index.ts)
+<!-- END GENERATED cordis-surface -->

package/kb/extra/subsystems/browser-use.zh.md

@@ -0,0 +1,67 @@
+# 浏览器操作
+
+[English](browser-use.md) | 中文
+
+浏览器操作让模型通过配置的后端检查与操作网页。DSH 拥有任务循环；提供方提供浏览器操作，并在一个实时 Session 的多个轮次之间保留浏览器状态。
+
+## 选择提供方
+
+在同一组合中挂载 [`dsh-browser-use`](../../packages/browser-use/browser-use/README.zh.md) 和一个提供方。这些提供方是实验性公共 npm 包，需要显式激活。它们最初支持的浏览器引擎是 Chromium。
+
+| 提供方 | 集成方式 |
+|---|---|
+| [Playwright MCP](../../packages/experimental/browser-use-playwright-mcp/README.zh.md) | Playwright 的浏览器控制 MCP 工具 |
+| [Chrome DevTools MCP](../../packages/experimental/browser-use-chrome-devtools-mcp/README.zh.md) | 通过 MCP 进行 Chrome DevTools 检查与控制 |
+| [Stagehand](../../packages/experimental/browser-use-stagehand-native/README.zh.md) | 原生浏览器操作，支持 AI（人工智能）辅助的动作、观测与提取 |
+
+共享服务只注册名称，并拒绝任何第二次提供方注册，包括同名实例。它不包含通用浏览器操作方法、浏览器资源或模型控制的选择器。Profile 或 preset 中的提供方配置为此次激活选择启动或附加模式。
+
+## Session 所有权
+
+启动的浏览器属于使用它的确切实时 Agent 与 Session。跨轮次的调用复用该浏览器。Session 运行时释放时关闭其启动的资源；重新加载或 fork Session 时创建全新浏览器状态。浏览器 profile 和登录状态不会从 Session 日志恢复。
+
+附加的浏览器仍归外部所有。提供方在该提供方实例内将浏览器保留给一个 Session，保留现有浏览器状态，并拒绝另一个 Session 同时附加。清理会断开连接并保持外部浏览器运行。独立 DSH 进程与其他客户端不受此保留约束。
+
+提供方关闭时先停止接收工具调用，并等待自有工作与资源清理完成，再释放共享提供方注册。取消无法撤销已交付的浏览器操作。
+
+## MCP 初始化
+
+MCP 提供方为其加载后创建的每个活动 Agent 初始化一个客户端。现有的串行 `agent/created` 事件等待连接和发现结束后，创建或恢复才完成，排队输入才开始运行。客户端跨轮次归 Session 所有。启动失败或取消会拒绝创建或恢复，并触发客户端清理。
+
+如果附加连接已被占用，本次激活不使用浏览器，但继续运行，后续轮次不会重试。连接释放后，新创建或恢复的激活可以获取它。加载或重新加载提供方不会接管已经活动的 Session；[共享运行时](../../packages/experimental/browser-use-runtime/README.zh.md)拥有这些初始化规则。
+
+## 工具与记录结果
+
+提供方工具使用常规 DSH 执行管线与 Session 日志。提供方拥有自己的工具 schema、结果渲染、图像支持、配置和上游限制；共享服务不添加模型可见内容。Stagehand 的 AI 辅助操作使用其显式配置的原生模型，DSH 保留任务循环。DSH 模型路由、凭据复用、底层推理请求/响应捕获，以及与 Session 用量计量的集成均属暂缓工作；返回的 SDK 数据和元数据仍作为普通工具结果记录。
+
+浏览器 MCP 连接还提供[资源和服务器指令](mcp.zh.md)。发往浏览器服务器的资源调用使用其 Session 队列，并拒绝其他 Session 的请求；服务器指令只会组装到所属 Session 的提示词中。
+
+[决策记录](../../.agents/notes/implemented/architecture/2026-09-12-browser-use-provider-registration.zh.md)解释只注册名称的服务与按 Session 管理的所有权。
+
+<!-- BEGIN GENERATED cordis-surface (gen-cordis-catalog.ts) — do not edit between markers -->
+
+<a id="cordis-surface"></a>
+
+## Cordis API
+
+Generated from source by `scripts/gen-cordis-catalog.ts` (verified fresh by `pnpm run verify-cordis-catalog` in doc-sync; regenerate with `pnpm run gen-cordis-catalog`) — the language sides differ only in locale-specific paired document paths. Signature blocks use a `ts cordis-catalog` fence and keep the original source JSDoc; dispatch modes are defined in the [primer](../cordis-primer.zh.md#dispatch-modes), and the framework-inherited `ctx` API lives in [cordis-api/inherited.md](../cordis-api/inherited.md).
+
+<a id="ctxbrowseruse--browseruseregistry"></a>
+
+### `ctx.browserUse` — `BrowserUseRegistry`
+
+Owns one optional provider registration in the shared browser-use service.
+
+```ts cordis-catalog
+/**
+ * Reserve the sole provider slot until the contribution is disposed.
+ * A second registration fails even when it repeats the current name. Providers
+ * must stop their tools and await owned work before releasing this registration.
+ * @param name - provider-owned name used in registration diagnostics.
+ * @returns the effect disposer for this exact registration.
+ */
+register(name: BrowserUseProviderName): () => Promise<void>
+```
+
+Source: [`packages/browser-use/browser-use/src/index.ts`](../../packages/browser-use/browser-use/src/index.ts)
+<!-- END GENERATED cordis-surface -->

package/kb/extra/subsystems/claude-code-mods.md

@@ -0,0 +1,146 @@
+# Claude Code mods: compatibility
+
+English | [中文](claude-code-mods.zh.md)
+
+Every way a mod behaves differently through this bridge than under Claude Code, as of Claude Code 2.1.287 ([mods reference](https://code.claude.com/docs/en/plugins/mods/reference), its `claude-code.d.ts`) and the [Getting started with Claude Code mods](https://claude.dev/blog/getting-started-with-claude-code-mods/) post of 2026-10-01. A row names the Claude Code behavior, what this bridge does, and the reason or the follow-up. Anything not listed here behaves as the reference describes; a hook on an event this host never raises is named in a warning when the mod loads, and a `$` call outside the served table rejects with `no implementation for <namespace>.<method>`.
+
+## Packaging and loading
+
+| Claude Code | This bridge | Reason or follow-up |
+|---|---|---|
+| A plugin directory with `.claude-plugin/plugin.json` and `hooks/hooks.json` naming one hooks module; `claude --plugin-dir`, marketplaces, `/plugin install` | A mod is a plugin: `defineMod({ name, version, root, userConfig, register })` wraps the hooks module and mounts through cordis.yml after the bridge; `plugin.json` and `hooks.json` are not read | Mods join the one composition mechanism every DSH plugin uses; the example directories keep their manifests so they also run under `claude --plugin-dir` |
+| `userConfig` schema in `plugin.json` with typed defaults; options validated at load | `userConfig` is the defaults object of the `defineMod` spec; the plugin's cordis.yml `config` overlays it and is validated as a record of strings, numbers, booleans, and string lists | No manifest is read |
+| Tiers (`user`, `project`, `managed`) and `next.to(e, tier)` | Every mod loads as `user`; `tier()` in a test is accepted and ignored; `next.to` rejects | Managed settings are not a DSH concept |
+| Hot reload of the hooks module on edit; `claude plugin validate`, `claude plugin test`, types generation from `types/index.d.ts` | None of these commands; a remount re-runs `register` on the same evaluated module, so module-level variables keep their values | Follow-up only on demand; the test kit's `createModTestKit` replaces `claude plugin test` inside Vitest |
+| A hooks module may be TypeScript | A `.ts` hooks module loads only where the launcher transpiles (the source launch does; a built install does not) | DSH ships plain Node |
+| Mods run in-process with `$` as their only access to the host | The hooks module runs in-process with Node's globals, no access rule, and the process's full authority: `$.env` reads and writes the harness environment, `$.http.fetch` reaches any URL, `$.fs` and `$.tool.call` act as the session | No sandbox is applied to mods; mount only mods you would run as a plugin |
+| `import type { … } from 'claude-code'`, `declare module 'claude-code' { interface PluginState }` | The `claude-code` module name resolves only inside this repository's test setup, to the bridge's types; a mod outside it imports `@deepseek-ai/dsh-experimental-claude-code-mods` | The type names are the bridge's own |
+| `settings hooks` in `hooks.json` run beside mod hooks | Not run; mount `@deepseek-ai/dsh-hooks-claude-code` for them | Different bridge |
+
+## Events
+
+| Event | Claude Code | This bridge |
+|---|---|---|
+| `session.start` | Raised once per session, before the first prompt | Raised for a root agent from `agent/created`, awaited before its first turn; not raised for subagents |
+| `session.end` | Raised once per session | Raised for a root agent from `agent/disposed`; `$.state` is readable until the hooks settle |
+| `prompt.submit` | `e.text` is the typed prompt; `context` entries become one block each after the prompt as typed; a prompt a plugin submitted carries `origin: { kind: 'plugin', name }` | Same; `e.text` joins the text blocks of the human's own (`user`-sourced) messages in the claimed batch, a batch of injected context alone raises nothing, and a prompt `$.prompt.submit` queued carries the submitting mod as its origin |
+| `turn.start`, `turn.complete` | Per turn; subagent turns carry `agentId` | Same; `turn.complete`'s `{ text }` reaches the host log, not a line under the answer; `durationMs` counts from `turn/start` |
+| `tool.call` | Runs before Claude Code's permission check; `next({ ...e, command })` rewrites the arguments the tool runs with, and `next({ ...e, tool })` reroutes the call | Runs around `tools/execute`, after the harness permission decision; a hook (or `.catch` handler) that rewrites arguments or names another tool is skipped with a report — the argument case names the [pre-tool input rewrite proposal](../../.agents/notes/proposed/feature/2026-06-30-pre-tool-input-rewrite.md) — because the logged call runs as logged. `{ deny }`, `{ result }`, observing, and rewriting the result or `isError` after `next` work as described |
+| `command.run` | Runs for a command the mod registered | Same |
+| `ui.render` | Raised for every render site (`AbovePrompt`, `Pane`, `ToolUse`, `AssistantMessage`, `Spinner`, …) | Raised for `AbovePrompt` only, once per redraw of a session's band; `Pane` hooks run only through the test kit's `$.ui.mount({ component: 'Pane' })` |
+| `tool.check`, `tool.describe`, `tool.list`, `tool.register` | Shape the tool list and permissions | Not raised; `$.tool.register` and `$.tool.list` are served as `$` calls |
+| `turn.step` | Per model step | Not raised |
+| Every other reference event: the other `prompt.*` events, `skill.prompt`, `attribution.text`, `command.describe`, `config.*`, the other `session.*` events, `agent.*`, `plugin.register`, `engine.create`, `telemetry.*`, `ui.resolve`, `ui.press`, `ui.input`, `ui.select`, `ui.focus`, `ui.scroll`, `ui.close`, `ui.message` | Raised by the engine | Registered, never raised, named in a warning at load |
+| `classic.*` | Classic settings hooks as events | Refused at `register`, like any unknown event name |
+| `<namespace>.<method>` op events (`fs.read`, `store.set`, …) | A later-loaded mod's `$` call reaches the mods loaded before it | Same; `tool.call` arrives through the tool pipeline instead |
+
+## The chain
+
+| Claude Code | This bridge | Reason |
+|---|---|---|
+| A hook may call `next` more than once; each call runs the chain beneath again | `next` runs once: a second call returns the first run's result | One tool execution per logged call |
+| A hook that answers while its `next` is still running | The run beneath is awaited before the hook's answer settles; a failure beneath after the answer is reported | The event settles only when nothing it started is running |
+| 10 s own-time budget, 1 s for `.catch`, time inside `next` and `$` calls excluded except `$.clock.sleep` | Same, configurable (`hookTimeoutMs`, `catchTimeoutMs`) | — |
+| Not returning a result object is a failure | Same; `null` is accepted as an answer (a surface drawn empty) | `ui.render` with nothing to draw |
+| A result's unknown fields are dropped and wrong types fail the hook | A `prompt.submit` result with wrong types keeps the original text and only the string context lines | Lenient on the one event where partial results are common |
+
+## `$` members
+
+| Namespace | Claude Code | This bridge |
+|---|---|---|
+| `$.plugin` | `name`, `root`, `tier` | `name`, `root` from the `defineMod` spec |
+| `$.ui` | `resolve`, `invalidate`, `open`, `close`, `panes`, `focus`, `log`, `toast`, `status`, `notice`, `ask`, `blit`, `copy`, `press`, `input`, `select` | `resolve`, `invalidate`, `open`, `close`, `panes`, `log`, `toast`, `status`, `ask`; `open` answers `{ isPlaced: false }` with a reason and redraws the band; `log`, `toast`, `status` reach the host log; the rest reject |
+| `$.command` | `register`, `run`, `list` | Same; `run` of a command nobody answers rejects with the registering mod named |
+| `$.tool` | `register`, `call`, `list` | Same; a registered tool is `mcp__<plugin>__<tool>`; `call` runs the harness pipeline and injects a tool's deferred contexts into the session |
+| `$.prompt` | `submit`, `compose` | `submit`; the message is `user`-sourced, framed `Message from the "<plugin>" mod:` unless `asUser`, and the `prompt.submit` it raises names the mod as origin |
+| `$.session` | `id`, `cwd`, `root`, `repo`, `model`, `turns`, `messages`, `usage`, `version`, `send`, `append`, `authorize`, `compact`, `surfaces` | `id`, `cwd`, `root`, `model`, `turns`, `messages`, `usage`, `version`; `cwd` and `root` both report the session workspace; `usage.context.window` is `0` and `percent` absent until the token meter knows the route's window; `rateLimits` is empty; `version` names this bridge |
+| `$.state` | Per session, survives hot reload | Per session, in memory; a read during `ui.render` subscribes the band to the slot |
+| `$.store` | Per plugin, durable | Same, in the `claude_code_mods` storage domain, 4 MiB per plugin |
+| `$.clock` | `now`, `sleep`, `after`, `every`; timers die with the plugin | Same; a timer also dies with the session whose event scheduled it, and `sleep` rejects when the event is cancelled |
+| `$.fs` | `read`, `write`, `list`, `exists`, `stat`, `ancestors` | All but `ancestors`; 4 MiB per file; `stat.mtimeMs` is `0` |
+| `$.process` | `run`, `spawn` | `run` only, argv without a shell, `processTimeoutMs` default |
+| `$.http` | `fetch` | Same, bodies up to 4 MiB |
+| `$.env` | `get`, `set` | Same; the process environment is shared by every session and plugin |
+| `$.model`, `$.agent`, `$.config`, `$.settings`, `$.mcp`, `$.audio`, `$.telemetry`, `$.turn` | Served | Reject with `no implementation`; `$.model.complete` waits on a logged side-request event |
+
+## Drawing
+
+| Claude Code | This bridge |
+|---|---|
+| `Box`, `Text`, `Button`, `Input`, `Spacer`, and the host's own element kinds | `Box`, `Text`, `Button` from `$.ui.resolve(e)`; the tree is validated, and a `Button`'s `onPress` stays in the host behind a per-drawing action id |
+| `Box` props: `flexDirection`, `padding*`, `gap`, `border*`, `width`, `height`, `overflow`, `scroll` | `flexDirection`, `padding`, `paddingX`, `paddingY`, `gap`, `border`, `borderColor`; the others serialize and the Web band ignores them |
+| `Text` props: `color`, `bold`, `dimColor`, `italic`, `underline`, `wrap`, `truncate` | `color` (the terminal palette names), `bold`, `dimColor`, `italic`, `underline`; the Web band wraps |
+| `Button` props: `label`, `hotkey`, `disabled`, `onPress`; a bare digit arms the hotkey | `label`, `hotkey` (shown as a hint), `disabled`, `onPress` (a click) |
+| The band is one instance; `hasSurvey`, `isWorking`, `maxRows`, `bodyColumns` reflect the terminal | One instance per session, consulted in mod load order; `bodyColumns` and `viewport.columns` are the bridge's `bandColumns` (default `120`), `maxRows` its `bandRows` (default `10`), `hasSurvey` and `isWorking` are `false`; the Web band wraps instead of scrolling |
+| The band redraws on the host's frame and when a `$.state` value a render read changes | Redraws after `session.start`, when a `$.state` value the last drawing read is written, after `$.ui.open`, `$.ui.close`, `$.ui.invalidate`, after each `tool.call` chain and `turn.complete` settle, and after a press; a mod that changes module-level state at other times draws it at the next trigger. A press runs `onPress` without a time limit |
+| `Pane` with `focus`, `scroll`, docked beside the transcript | No pane: `$.ui.open` answers `{ isPlaced: false }` so a mod falls back to the band; the test kit mounts `Pane` hooks directly |
+| Hotkeys, `TextInput`, focus, scrolling, `ui.press`, `ui.input`, `ui.select` | Not served; follow-up |
+
+## Testing
+
+| `claude-code/testing` | This bridge |
+|---|---|
+| `describe`, `test(name, ($, on) => …)`, `expect`, `mock`, `tier` | Inside this repository the same module name provides them over Vitest; `test` infers the mod from the test file's location (`<mod>/tests/*.test.ts` → `<mod>/index.ts`) and `defineModTests` overrides it. Outside, `createModTestKit` and `mock` come from `@deepseek-ai/dsh-experimental-claude-code-mods/testing` |
+| `on` stubs the engine's answers; the kit answers `ui.*` itself | Same; `ui.open` answers `{ isPlaced: false }`, `ui.invalidate` and `ui.close` succeed; `session.cwd`, `process.run`, `fs.*`, `store.*`, `env.*` need a stub (`mock.store`, `mock.env`, `mock.clock` answer whole namespaces) |
+| `$.ui.mount({ plugin, surface, component, props })` → `find`, `findAll`, `press`, `unmount` | Same, plus `tree()` and `text()`; every read renders afresh through every loaded mod's hooks; `plugin` and `surface` are accepted for source compatibility |
+| 5 s hook budget in tests | Same (`budgetMs`) |
+
+## The example mods
+
+| Mod | Source | Through this bridge |
+|---|---|---|
+| Token Weather | The post's module, type contract, and test, unchanged | Runs as published; the test's three `as any` casts are dropped because the kit is typed. Draws in the band from `$.session.usage` through the token meter |
+| Blast Radius | The post's `tool.call` hook unchanged; `classify`, `measure`, and the trees completed to the post's description | Holds the command with real `sleep` polls, draws the report in the band (`isPlaced: false`), and Proceed or Cancel decides; load it before Token Weather |
+| Replay Theater | The post's five hooks unchanged; `stepsFor`, the diff, `openReplay`, the pane and the hint completed | Records Edit and Write calls, hints in the band after the turn, and steps through the diffs in the band because no pane is placed |
+| `diff` ([anthropics/claude-code](https://github.com/anthropics/claude-code/tree/main/mods/diff), read at `52c7644`) | TypeScript, ~570 files, hooks `command.run`, `prompt.submit`, `session.start`, `tool.call`, `ui.close`, `ui.focus`, `ui.render`, `ui.scroll` | Not runnable: needs `$.settings.read`, `$.telemetry.*`, `$.session.messages` over a pane, `ui.focus` and `ui.scroll` events, and a placed `Pane`. Assessed by reading; the repository's terms do not allow a fixture copy |
+| `agents-md` (same repository) | Hooks `agent.spawn`, `prompt.context`, `session.start`, `tool.call` | Not runnable: `prompt.context` and `agent.spawn` are not raised and `$.fs.ancestors`, `$.telemetry.*` are not served |
+| `sec-default` (same repository) | Hooks `tool.check`, `prompt.compose`, `settings.read`, `classic.*`, `plugin.register`, … | Not runnable: managed-tier policy over events this host does not raise |
+| `telemetry` (same repository) | Hooks `engine.create`, `telemetry.*`, `session.start`, `session.end`; calls `$.session.repo`, `$.settings.read` | Not runnable: `telemetry.*` and `engine.create` are not raised and the two `$` members are not served |
+
+<!-- BEGIN GENERATED cordis-surface (gen-cordis-catalog.ts) — do not edit between markers -->
+
+<a id="cordis-surface"></a>
+
+## Cordis API
+
+Generated from source by `scripts/gen-cordis-catalog.ts` (verified fresh by `pnpm run verify-cordis-catalog` in doc-sync; regenerate with `pnpm run gen-cordis-catalog`) — the language sides differ only in locale-specific paired document paths. Signature blocks use a `ts cordis-catalog` fence and keep the original source JSDoc; dispatch modes are defined in the [primer](../cordis-primer.md#dispatch-modes), and the framework-inherited `ctx` API lives in [cordis-api/inherited.md](../cordis-api/inherited.md).
+
+<a id="ctxclaudecodemods--claudecodemods"></a>
+
+### `ctx.claudeCodeMods` — `ClaudeCodeMods`
+
+The bridge service: loaded mods, their hooks, and the harness listeners that raise their events. Mods join through ClaudeCodeMods.add, which defineMod calls when a mod plugin mounts. The Remote methods let a Client draw each session's band above the prompt and press its buttons.
+
+```ts cordis-catalog
+/**
+ * Watch the band above the prompt of one session: the current drawing, then
+ * every redraw, until the Client stops watching.
+ * @param agent - the session's agent, resolved by the Gateway.
+ * @param signal - carrier cancellation.
+ * @returns the band's snapshots.
+ */
+@Remote({ mode: 'stream' }) watchBand(agent: Agent, signal: AbortSignal): AsyncIterable<SurfaceSnapshot>
+
+/**
+ * Press a button of the band's current drawing: runs the mod's `onPress` and redraws.
+ * @param agent - the session's agent, resolved by the Gateway.
+ * @param generation - the drawing the Client saw.
+ * @param actionId - the button's action id in that drawing.
+ * @returns the snapshot after the press.
+ */
+@Remote pressBand(agent: Agent, generation: number, actionId: string): Promise<SurfaceSnapshot>
+
+/**
+ * Load one mod beneath every mod loaded before it: run its `register`,
+ * keep its hooks, and report which of its events this host never raises.
+ * @param definition - the mod as its plugin defined it.
+ * @returns the disposer that removes the mod's hooks, closes its timers, and releases its registrations.
+ * @throws Error when the name is invalid or taken, or when `register` throws (Claude Code's `hooks module did not load` wording).
+ */
+async add(definition: ModDefinition): Promise<() => Promise<void>>
+```
+
+Types: [Agent](core.md)
+
+Source: [`packages/experimental/claude-code-mods/src/index.ts`](../../packages/experimental/claude-code-mods/src/index.ts)
+<!-- END GENERATED cordis-surface -->

package/kb/extra/subsystems/claude-code-mods.zh.md

@@ -0,0 +1,146 @@
+# Claude Code 模组：兼容性
+
+[English](claude-code-mods.md) | 中文
+
+模组经本桥接运行时与在 Claude Code 下行为不同的每一处，基准为 Claude Code 2.1.287（[模组参考](https://code.claude.com/docs/en/plugins/mods/reference)及其 `claude-code.d.ts`）与 2026-10-01 的 [Getting started with Claude Code mods](https://claude.dev/blog/getting-started-with-claude-code-mods/) 一文。每行给出 Claude Code 的行为、本桥接的做法，以及原因或后续工作。未列出的一切均按参考文档描述的方式运行；对本宿主永不触发的事件注册钩子会在模组加载时以警告点名，服务表之外的 `$` 调用以 `no implementation for <namespace>.<method>` 拒绝。
+
+## 打包与加载
+
+| Claude Code | 本桥接 | 原因或后续 |
+|---|---|---|
+| 一个插件目录，含 `.claude-plugin/plugin.json` 与指定一个钩子模块的 `hooks/hooks.json`；`claude --plugin-dir`、市场、`/plugin install` | 模组就是插件：`defineMod({ name, version, root, userConfig, register })` 包装钩子模块，并在桥接之后经 cordis.yml 挂载；不读取 `plugin.json` 与 `hooks.json` | 模组加入每个 DSH 插件都使用的同一组合机制；示例目录保留其清单，因此也能在 `claude --plugin-dir` 下运行 |
+| `plugin.json` 中带类型默认值的 `userConfig` schema；选项在加载时校验 | `userConfig` 是 `defineMod` 规格的默认值对象；插件的 cordis.yml `config` 覆盖其上，并按字符串、数字、布尔与字符串列表的记录校验 | 不读取清单 |
+| 层级（`user`、`project`、`managed`）与 `next.to(e, tier)` | 每个模组都以 `user` 加载；测试中的 `tier()` 被接受并忽略；`next.to` 拒绝 | 托管设置不是 DSH 的概念 |
+| 编辑时热重载钩子模块；`claude plugin validate`、`claude plugin test`、由 `types/index.d.ts` 生成类型 | 这些命令都没有；重新挂载会在同一已求值模块上重新运行 `register`，模块级变量保留其值 | 仅按需跟进；测试工具包的 `createModTestKit` 在 Vitest 中替代 `claude plugin test` |
+| 钩子模块可以是 TypeScript | `.ts` 钩子模块只在启动器转译时可加载（源码启动可以；构建后的安装不行） | DSH 交付的是纯 Node |
+| 模组在进程内运行，`$` 是其访问宿主的唯一途径 | 钩子模块以 Node 全局对象在进程内运行，没有访问规则，并拥有进程的全部权限：`$.env` 读写 harness 环境变量，`$.http.fetch` 可访问任意 URL，`$.fs` 与 `$.tool.call` 以会话的身份行动 | 不对模组施加沙箱；只挂载你愿意作为插件运行的模组 |
+| `import type { … } from 'claude-code'`、`declare module 'claude-code' { interface PluginState }` | `claude-code` 模块名只在本仓库的测试设置中解析到桥接的类型；仓库外的模组导入 `@deepseek-ai/dsh-experimental-claude-code-mods` | 类型名是桥接自己的 |
+| `hooks.json` 中的设置钩子与模组钩子并行运行 | 不运行；为它们挂载 `@deepseek-ai/dsh-hooks-claude-code` | 不同的桥接 |
+
+## 事件
+
+| 事件 | Claude Code | 本桥接 |
+|---|---|---|
+| `session.start` | 每会话触发一次，在首个提示词之前 | 由根智能体的 `agent/created` 触发，在其首轮之前等待完成；不为子智能体触发 |
+| `session.end` | 每会话触发一次 | 由根智能体的 `agent/disposed` 触发；钩子结束前 `$.state` 可读 |
+| `prompt.submit` | `e.text` 是输入的提示词；`context` 条目各成一块跟在输入的提示词之后；插件提交的提示词带 `origin: { kind: 'plugin', name }` | 相同；`e.text` 拼接已认领批次中人类自己（`user` 来源）消息的文本块，仅含注入上下文的批次不触发，`$.prompt.submit` 排入的提示词以提交的模组为来源 |
+| `turn.start`、`turn.complete` | 每轮触发；子智能体的轮带 `agentId` | 相同；`turn.complete` 的 `{ text }` 到达宿主日志，而非答案下方的一行；`durationMs` 从 `turn/start` 起算 |
+| `tool.call` | 在 Claude Code 的权限检查之前运行；`next({ ...e, command })` 改写工具实际运行的参数，`next({ ...e, tool })` 把调用改派给另一工具 | 包裹 `tools/execute` 运行，在 harness 权限决定之后；改写参数或改名工具的钩子（或 `.catch` 处理器）会被跳过并报告——参数情形指向[工具前输入改写提案](../../.agents/notes/proposed/feature/2026-06-30-pre-tool-input-rewrite.zh.md)——因为已写入日志的调用按原样运行。`{ deny }`、`{ result }`、观察以及在 `next` 之后改写结果或 `isError` 均按描述工作 |
+| `command.run` | 为模组注册的命令运行 | 相同 |
+| `ui.render` | 为每个渲染位点触发（`AbovePrompt`、`Pane`、`ToolUse`、`AssistantMessage`、`Spinner`……） | 只为 `AbovePrompt` 触发，每次会话横幅重绘一次；`Pane` 钩子只经测试工具包的 `$.ui.mount({ component: 'Pane' })` 运行 |
+| `tool.check`、`tool.describe`、`tool.list`、`tool.register` | 塑造工具列表与权限 | 不触发；`$.tool.register` 与 `$.tool.list` 作为 `$` 调用服务 |
+| `turn.step` | 每个模型步骤 | 不触发 |
+| 参考中的其他所有事件：其他 `prompt.*` 事件、`skill.prompt`、`attribution.text`、`command.describe`、`config.*`、其他 `session.*` 事件、`agent.*`、`plugin.register`、`engine.create`、`telemetry.*`、`ui.resolve`、`ui.press`、`ui.input`、`ui.select`、`ui.focus`、`ui.scroll`、`ui.close`、`ui.message` | 由引擎触发 | 可注册，永不触发，在加载时以警告点名 |
+| `classic.*` | 经典设置钩子作为事件 | 在 `register` 时拒绝，与任何未知事件名相同 |
+| `<namespace>.<method>` 操作事件（`fs.read`、`store.set`……） | 更晚加载的模组的 `$` 调用到达在它之前加载的模组 | 相同；`tool.call` 则经由工具流水线到达 |
+
+## 链
+
+| Claude Code | 本桥接 | 原因 |
+|---|---|---|
+| 钩子可多次调用 `next`；每次调用都再次运行下层链 | `next` 只运行一次：第二次调用返回第一次运行的结果 | 每条写入日志的调用只执行一次工具 |
+| 钩子在其 `next` 仍在运行时作答 | 下层运行会在钩子答案落定前被等待；答案之后的下层失败会被报告 | 事件只在其启动的一切都结束后落定 |
+| 自身时间预算 10 秒，`.catch` 为 1 秒，`next` 与 `$` 调用内的时间不计，`$.clock.sleep` 除外 | 相同，可配置（`hookTimeoutMs`、`catchTimeoutMs`） | — |
+| 不返回结果对象视为失败 | 相同；`null` 被接受为答案（绘制为空的表面） | 无可绘制内容的 `ui.render` |
+| 结果中的未知字段被丢弃，类型错误使钩子失败 | 类型错误的 `prompt.submit` 结果保留原文本，并只取字符串形式的 context 行 | 对这一常见部分结果的事件从宽处理 |
+
+## `$` 成员
+
+| 命名空间 | Claude Code | 本桥接 |
+|---|---|---|
+| `$.plugin` | `name`、`root`、`tier` | 来自 `defineMod` 规格的 `name`、`root` |
+| `$.ui` | `resolve`、`invalidate`、`open`、`close`、`panes`、`focus`、`log`、`toast`、`status`、`notice`、`ask`、`blit`、`copy`、`press`、`input`、`select` | `resolve`、`invalidate`、`open`、`close`、`panes`、`log`、`toast`、`status`、`ask`；`open` 带原因回答 `{ isPlaced: false }` 并重绘横幅；`log`、`toast`、`status` 到达宿主日志；其余拒绝 |
+| `$.command` | `register`、`run`、`list` | 相同；`run` 一个无人回答的命令时拒绝，并点名注册它的模组 |
+| `$.tool` | `register`、`call`、`list` | 相同；注册的工具为 `mcp__<plugin>__<tool>`；`call` 运行 harness 流水线，并把工具推迟的上下文注入会话 |
+| `$.prompt` | `submit`、`compose` | `submit`；消息为 `user` 来源，除非 `asUser` 否则套上 `Message from the "<plugin>" mod:` 框架，它触发的 `prompt.submit` 以该模组为来源 |
+| `$.session` | `id`、`cwd`、`root`、`repo`、`model`、`turns`、`messages`、`usage`、`version`、`send`、`append`、`authorize`、`compact`、`surfaces` | `id`、`cwd`、`root`、`model`、`turns`、`messages`、`usage`、`version`；`cwd` 与 `root` 都报告会话工作区；在 token 计量器得知路由窗口前 `usage.context.window` 为 `0` 且无 `percent`；`rateLimits` 为空；`version` 指明本桥接 |
+| `$.state` | 按会话，热重载后保留 | 按会话，内存中；`ui.render` 期间的读取让横幅订阅该槽位 |
+| `$.store` | 按插件，持久 | 相同，在 `claude_code_mods` 存储域中，每插件 4 MiB |
+| `$.clock` | `now`、`sleep`、`after`、`every`；定时器随插件消亡 | 相同；定时器还随安排它的事件所属会话消亡，事件取消时 `sleep` 拒绝 |
+| `$.fs` | `read`、`write`、`list`、`exists`、`stat`、`ancestors` | 除 `ancestors` 外全部；每文件 4 MiB；`stat.mtimeMs` 为 `0` |
+| `$.process` | `run`、`spawn` | 仅 `run`，不经 shell 的 argv，默认 `processTimeoutMs` |
+| `$.http` | `fetch` | 相同，正文最多 4 MiB |
+| `$.env` | `get`、`set` | 相同；进程环境由所有会话与插件共享 |
+| `$.model`、`$.agent`、`$.config`、`$.settings`、`$.mcp`、`$.audio`、`$.telemetry`、`$.turn` | 已服务 | 以 `no implementation` 拒绝；`$.model.complete` 等待一个写入日志的副请求事件 |
+
+## 绘制
+
+| Claude Code | 本桥接 |
+|---|---|
+| `Box`、`Text`、`Button`、`Input`、`Spacer` 以及宿主自己的元素种类 | 来自 `$.ui.resolve(e)` 的 `Box`、`Text`、`Button`；树经校验，`Button` 的 `onPress` 留在宿主中，位于按绘制分配的动作 id 之后 |
+| `Box` 属性：`flexDirection`、`padding*`、`gap`、`border*`、`width`、`height`、`overflow`、`scroll` | `flexDirection`、`padding`、`paddingX`、`paddingY`、`gap`、`border`、`borderColor`；其余会序列化，Web 横幅忽略 |
+| `Text` 属性：`color`、`bold`、`dimColor`、`italic`、`underline`、`wrap`、`truncate` | `color`（终端调色板名）、`bold`、`dimColor`、`italic`、`underline`；Web 横幅会换行 |
+| `Button` 属性：`label`、`hotkey`、`disabled`、`onPress`；单个数字键触发热键 | `label`、`hotkey`（显示为提示）、`disabled`、`onPress`（点击） |
+| 横幅是一个实例；`hasSurvey`、`isWorking`、`maxRows`、`bodyColumns` 反映终端 | 每会话一个实例，按模组加载顺序咨询；`bodyColumns` 与 `viewport.columns` 取桥接的 `bandColumns`（默认 `120`），`maxRows` 取 `bandRows`（默认 `10`），`hasSurvey` 与 `isWorking` 为 `false`；Web 横幅换行而非滚动 |
+| 横幅随宿主帧重绘，并在渲染读取过的 `$.state` 值变化时重绘 | 在 `session.start` 之后、上次绘制读取过的 `$.state` 值被写入时、在 `$.ui.open`、`$.ui.close`、`$.ui.invalidate` 之后、在每条 `tool.call` 链与 `turn.complete` 结束后以及按下之后重绘；在其他时刻改变模块级状态的模组在下一次触发时绘出。按下运行 `onPress` 时没有时间限制 |
+| 带 `focus`、`scroll`、停靠在对话旁的 `Pane` | 没有面板：`$.ui.open` 回答 `{ isPlaced: false }`，模组回退到横幅；测试工具包直接挂载 `Pane` 钩子 |
+| 热键、`TextInput`、焦点、滚动、`ui.press`、`ui.input`、`ui.select` | 不服务；后续工作 |
+
+## 测试
+
+| `claude-code/testing` | 本桥接 |
+|---|---|
+| `describe`、`test(name, ($, on) => …)`、`expect`、`mock`、`tier` | 在本仓库内同一模块名基于 Vitest 提供它们；`test` 从测试文件位置推断模组（`<mod>/tests/*.test.ts` → `<mod>/index.ts`），`defineModTests` 可覆盖。仓库外则从 `@deepseek-ai/dsh-experimental-claude-code-mods/testing` 获取 `createModTestKit` 与 `mock` |
+| `on` 为引擎的答案打桩；工具包自行回答 `ui.*` | 相同；`ui.open` 回答 `{ isPlaced: false }`，`ui.invalidate` 与 `ui.close` 成功；`session.cwd`、`process.run`、`fs.*`、`store.*`、`env.*` 需要桩（`mock.store`、`mock.env`、`mock.clock` 回答整个命名空间） |
+| `$.ui.mount({ plugin, surface, component, props })` → `find`、`findAll`、`press`、`unmount` | 相同，另有 `tree()` 与 `text()`；每次读取都经所有已加载模组的钩子重新渲染；`plugin` 与 `surface` 为源码兼容而接受 |
+| 测试中钩子预算 5 秒 | 相同（`budgetMs`） |
+
+## 示例模组
+
+| 模组 | 来源 | 经本桥接 |
+|---|---|---|
+| Token Weather | 文中的模块、类型契约与测试，原样 | 按发布原样运行；测试中的三处 `as any` 转换因工具包带类型而去掉。通过 token 计量器从 `$.session.usage` 在横幅中绘制 |
+| Blast Radius | 文中的 `tool.call` 钩子原样；`classify`、`measure` 与树按文中描述补全 | 以真实的 `sleep` 轮询持有命令，在横幅中绘制报告（`isPlaced: false`），由 Proceed 或 Cancel 决定；请加载在 Token Weather 之前 |
+| Replay Theater | 文中的五个钩子原样；`stepsFor`、差异、`openReplay`、面板与提示补全 | 记录 Edit 与 Write 调用，轮结束后在横幅中提示，并因无面板放置而在横幅中逐步浏览差异 |
+| `diff`（[anthropics/claude-code](https://github.com/anthropics/claude-code/tree/main/mods/diff)，于 `52c7644` 阅读） | TypeScript，约 570 个文件，钩住 `command.run`、`prompt.submit`、`session.start`、`tool.call`、`ui.close`、`ui.focus`、`ui.render`、`ui.scroll` | 不可运行：需要 `$.settings.read`、`$.telemetry.*`、面板上的 `$.session.messages`、`ui.focus` 与 `ui.scroll` 事件以及已放置的 `Pane`。以阅读方式评估；该仓库的条款不允许复制夹具 |
+| `agents-md`（同一仓库） | 钩住 `agent.spawn`、`prompt.context`、`session.start`、`tool.call` | 不可运行：`prompt.context` 与 `agent.spawn` 不触发，`$.fs.ancestors`、`$.telemetry.*` 不服务 |
+| `sec-default`（同一仓库） | 钩住 `tool.check`、`prompt.compose`、`settings.read`、`classic.*`、`plugin.register`…… | 不可运行：基于本宿主不触发的事件的托管层策略 |
+| `telemetry`（同一仓库） | 钩住 `engine.create`、`telemetry.*`、`session.start`、`session.end`；调用 `$.session.repo`、`$.settings.read` | 不可运行：`telemetry.*` 与 `engine.create` 不触发，两个 `$` 成员不服务 |
+
+<!-- BEGIN GENERATED cordis-surface (gen-cordis-catalog.ts) — do not edit between markers -->
+
+<a id="cordis-surface"></a>
+
+## Cordis API
+
+Generated from source by `scripts/gen-cordis-catalog.ts` (verified fresh by `pnpm run verify-cordis-catalog` in doc-sync; regenerate with `pnpm run gen-cordis-catalog`) — the language sides differ only in locale-specific paired document paths. Signature blocks use a `ts cordis-catalog` fence and keep the original source JSDoc; dispatch modes are defined in the [primer](../cordis-primer.zh.md#dispatch-modes), and the framework-inherited `ctx` API lives in [cordis-api/inherited.md](../cordis-api/inherited.md).
+
+<a id="ctxclaudecodemods--claudecodemods"></a>
+
+### `ctx.claudeCodeMods` — `ClaudeCodeMods`
+
+The bridge service: loaded mods, their hooks, and the harness listeners that raise their events. Mods join through ClaudeCodeMods.add, which defineMod calls when a mod plugin mounts. The Remote methods let a Client draw each session's band above the prompt and press its buttons.
+
+```ts cordis-catalog
+/**
+ * Watch the band above the prompt of one session: the current drawing, then
+ * every redraw, until the Client stops watching.
+ * @param agent - the session's agent, resolved by the Gateway.
+ * @param signal - carrier cancellation.
+ * @returns the band's snapshots.
+ */
+@Remote({ mode: 'stream' }) watchBand(agent: Agent, signal: AbortSignal): AsyncIterable<SurfaceSnapshot>
+
+/**
+ * Press a button of the band's current drawing: runs the mod's `onPress` and redraws.
+ * @param agent - the session's agent, resolved by the Gateway.
+ * @param generation - the drawing the Client saw.
+ * @param actionId - the button's action id in that drawing.
+ * @returns the snapshot after the press.
+ */
+@Remote pressBand(agent: Agent, generation: number, actionId: string): Promise<SurfaceSnapshot>
+
+/**
+ * Load one mod beneath every mod loaded before it: run its `register`,
+ * keep its hooks, and report which of its events this host never raises.
+ * @param definition - the mod as its plugin defined it.
+ * @returns the disposer that removes the mod's hooks, closes its timers, and releases its registrations.
+ * @throws Error when the name is invalid or taken, or when `register` throws (Claude Code's `hooks module did not load` wording).
+ */
+async add(definition: ModDefinition): Promise<() => Promise<void>>
+```
+
+Types: [Agent](core.zh.md)
+
+Source: [`packages/experimental/claude-code-mods/src/index.ts`](../../packages/experimental/claude-code-mods/src/index.ts)
+<!-- END GENERATED cordis-surface -->

package/kb/extra/subsystems/computer-use.md

@@ -0,0 +1,56 @@
+# Computer use
+
+English | [中文](computer-use.zh.md)
+
+Computer use lets a model observe and operate the local desktop through a configured provider. The shared DSH capability is called **computer use**; **Cua Driver** names the upstream implementation.
+
+## Choose a provider
+
+Mount [`dsh-computer-use`](../../packages/computer-use/computer-use/README.md) and one provider in the same composition. Both Cua Driver providers are experimental public npm packages and require explicit activation.
+
+| Provider | Runtime |
+|---|---|
+| [Cua Driver MCP](../../packages/experimental/computer-use-cua-driver-mcp/README.md) | An already installed `cua-driver` executable connected through MCP |
+| [Cua Driver native](../../packages/experimental/computer-use-cua-driver-native/README.md) | The platform-native runtime installed with the npm dependency |
+
+Each provider supplies its upstream tool catalog. The shared service registers only a name and rejects any second provider, including another instance with the same name. It has no common desktop-operation methods or model-controlled selector.
+
+## Lifetime and desktop sharing
+
+A provider retains its registration while it shuts down its tools and owned resources. Startup failure releases the attempted registration. The MCP provider keeps its registration during reconnects.
+
+One registered provider does not reserve a desktop for a Session. Callers coordinate complete observe, act, and verify workflows across Sessions and separate DSH processes. A cancelled call cannot undo input that the desktop already received.
+
+## Results and platform requirements
+
+Tools use the normal execution pipeline and Session log. Image-capable model routes with an attachment store receive durable screenshots; unsupported image routes receive the existing MCP image diagnostic. Provider READMEs own installation, permission, and platform limitations.
+
+The [decision record](../../.agents/notes/implemented/architecture/2026-09-12-computer-use-provider-registration.md) explains the registration-only service and the two Cua Driver integrations.
+
+<!-- BEGIN GENERATED cordis-surface (gen-cordis-catalog.ts) — do not edit between markers -->
+
+<a id="cordis-surface"></a>
+
+## Cordis API
+
+Generated from source by `scripts/gen-cordis-catalog.ts` (verified fresh by `pnpm run verify-cordis-catalog` in doc-sync; regenerate with `pnpm run gen-cordis-catalog`) — the language sides differ only in locale-specific paired document paths. Signature blocks use a `ts cordis-catalog` fence and keep the original source JSDoc; dispatch modes are defined in the [primer](../cordis-primer.md#dispatch-modes), and the framework-inherited `ctx` API lives in [cordis-api/inherited.md](../cordis-api/inherited.md).
+
+<a id="ctxcomputeruse--computeruseregistry"></a>
+
+### `ctx.computerUse` — `ComputerUseRegistry`
+
+Owns one optional provider registration in the shared computer-use service.
+
+```ts cordis-catalog
+/**
+ * Reserve the sole provider slot until the contribution is disposed.
+ * A second registration fails even when it repeats the current name. Providers
+ * must stop their tools and await owned work before releasing this registration.
+ * @param name - provider-owned name used in registration diagnostics.
+ * @returns the effect disposer for this exact registration.
+ */
+register(name: ComputerUseProviderName): () => Promise<void>
+```
+
+Source: [`packages/computer-use/computer-use/src/index.ts`](../../packages/computer-use/computer-use/src/index.ts)
+<!-- END GENERATED cordis-surface -->

package/kb/extra/subsystems/computer-use.zh.md

@@ -0,0 +1,56 @@
+# 计算机操作
+
+[English](computer-use.md) | 中文
+
+计算机操作让模型通过配置的提供方观察并操作本地桌面。DSH 的共享能力称为 **computer use（计算机操作）**；**Cua Driver** 是上游实现的名称。
+
+## 选择提供方
+
+在同一组合中挂载 [`dsh-computer-use`](../../packages/computer-use/computer-use/README.zh.md) 和一个提供方。两个 Cua Driver 提供方都是公开发布到 npm 的实验性包，均需显式启用。
+
+| 提供方 | 运行时 |
+|---|---|
+| [Cua Driver MCP](../../packages/experimental/computer-use-cua-driver-mcp/README.zh.md) | 通过 MCP 连接已安装的 `cua-driver` 可执行文件 |
+| [Cua Driver 原生](../../packages/experimental/computer-use-cua-driver-native/README.zh.md) | 随 npm 依赖安装的平台原生运行时 |
+
+各提供方提供上游工具目录。共享服务只注册名称，并拒绝任何第二个提供方，包括使用相同名称的另一个实例。服务不包含通用桌面操作方法或模型控制的选择器。
+
+## 生命周期和桌面共享
+
+提供方在关闭工具和自有资源期间保留注册。启动失败会释放此次尝试的注册。MCP 提供方在重连期间保留注册。
+
+一个已注册的提供方不会为某个 Session 预留桌面。调用方负责协调跨 Session 和独立 DSH 进程的完整观察、操作和验证流程。取消调用无法撤销桌面已收到的输入。
+
+## 结果和平台要求
+
+工具使用常规执行流程和 Session 日志。支持图像的模型路由在挂载附件存储时接收持久化截图；不支持图像的路由接收现有 MCP 图像诊断。提供方 README 负责说明安装、权限和平台限制。
+
+[决策记录](../../.agents/notes/implemented/architecture/2026-09-12-computer-use-provider-registration.zh.md)解释只负责注册的服务和两个 Cua Driver 集成。
+
+<!-- BEGIN GENERATED cordis-surface (gen-cordis-catalog.ts) — do not edit between markers -->
+
+<a id="cordis-surface"></a>
+
+## Cordis API
+
+Generated from source by `scripts/gen-cordis-catalog.ts` (verified fresh by `pnpm run verify-cordis-catalog` in doc-sync; regenerate with `pnpm run gen-cordis-catalog`) — the language sides differ only in locale-specific paired document paths. Signature blocks use a `ts cordis-catalog` fence and keep the original source JSDoc; dispatch modes are defined in the [primer](../cordis-primer.zh.md#dispatch-modes), and the framework-inherited `ctx` API lives in [cordis-api/inherited.md](../cordis-api/inherited.md).
+
+<a id="ctxcomputeruse--computeruseregistry"></a>
+
+### `ctx.computerUse` — `ComputerUseRegistry`
+
+Owns one optional provider registration in the shared computer-use service.
+
+```ts cordis-catalog
+/**
+ * Reserve the sole provider slot until the contribution is disposed.
+ * A second registration fails even when it repeats the current name. Providers
+ * must stop their tools and await owned work before releasing this registration.
+ * @param name - provider-owned name used in registration diagnostics.
+ * @returns the effect disposer for this exact registration.
+ */
+register(name: ComputerUseProviderName): () => Promise<void>
+```
+
+Source: [`packages/computer-use/computer-use/src/index.ts`](../../packages/computer-use/computer-use/src/index.ts)
+<!-- END GENERATED cordis-surface -->