@deepseek-ai/dsh-subagent 0.1.5-rc.2 → 0.1.6-alpha.2

package/README.i18n.yaml

@@ -2,5 +2,5 @@
 # side as of the last confirmed-consistent state. Both languages carry equal authority;
 # after editing either side, bring the other along and re-record with:
 #   pnpm run verify-translation-pairing --write packages/subagent/subagent/README.md
-README.md: 5f9668fd7a14cdfca33ab72a217c07372a24502f
-README.zh.md: 563ed06237088c4bc250ca3c4f5a863b6ffe8239
+README.md: 4634f0a33f71709668732b7070a2bccded3db5c0
+README.zh.md: d3d8a597f23153b0c82e663d98142a9966f48c22

package/README.md

@@ -42,6 +42,18 @@ Mount the service with a provider and the delegation tool. The provider register
 
 An agent that calls the tool gets the child's final answer as the tool result. Mounting the service alone changes nothing: nothing can delegate until a provider and a tool are composed.
 
+### Delegation settings
+
+The limits section on the **Plugins → Subagent** page edits the Host’s `subagent` settings section. User values override this plugin's composition; reset removes the user override. `maxDepth` defaults to `1` and supplies the delegation tools' depth when their own configuration omits it. An explicit tool depth, including `provider-managed`, takes precedence. Depth `0` disables delegation through tools inheriting this setting; depth `1` permits direct children only. Changes apply on the next delegation attempt. Direct service callers continue to supply their own optional request depth.
+
+### Continuable capacity
+
+Set `maxActiveSubagents` on the host `dsh-subagent` plugin to limit live children sharing uninterrupted continuable parent links. It defaults to `8` and accepts positive safe integers. A non-continuable parent starts a separate pool and does not consume a slot; continuable descendants inherit that pool. Fresh creation and cold resume reserve before reconstructing the Agent, and cleanup returns the slot after handle disposal. A waiting parent, pending inbox work, and an Activation being stopped still occupy slots. Messages to a resident child reuse its slot. One-shot and external-provider runs are outside this limit. Pool inheritance does not cross a one-shot parent; its continuable children share a separate pool. Depth remains the delegation tool's separate policy.
+
+The current `maxActiveSubagents` value is sampled before every new or cold-resumed Activation. Raising it admits more children in existing trees; lowering it leaves resident children running and refuses further admissions until usage is below the limit.
+
+At capacity, creation or cold resume rejects with `ACTIVATION_LIMIT_REACHED` (browser prompts receive `subagent/delivery-unavailable`): wait for a child to finish or continue using the existing agents. Admission does not queue, because a parent waiting for descendants must not wait for its own occupied slot. Slots are process-local and do not constrain cumulative Session history or token usage.
+
 ### One-shot and continuable children
 
 One-shot children run once and settle with a single result, plus an optional structured output and a safe diagnostic on failure. A start request may override the child Agent's provider, model, reasoning effort, and output-token limit through `agentOptions`; every requested option requires the provider's matching capability. Continuable children keep a durable session and accept later messages in order: the caller receives a stable child id, sends adjacent-Agent messages, and can interrupt the current turn without destroying the child. The tool row's `backgroundMode` picks the shape (`one-shot` by default, or `continuable` on providers that support it).
@@ -90,7 +102,7 @@ This section explains how the service is built and where the observable behavior
 
 ### One-shot flow
 
-A request is validated against the provider's advertised capabilities, a durable descriptor is snapshotted, and the provider builds the child. Both in-process providers advertise `agentOptions`: child creation merges requested fields over the provider, model, and reasoning effort in the parent's latest logged request, falls back to creation options before the first request, and retains the configured token limit. A route change without an explicit effort clears the inherited route-owned effort so the selected model resolves its default. DSH SDK also advertises this capability and publishes immutable `agentRouteDefaults`, which supply its instance provider/model defaults before exact-route preflight; `start()` still owns direct callers and the output cap. ACP, Codex, and Claude Code reject agent-route overrides rather than silently ignoring them. On success the run is published and ownership transfers to the caller; on failure the provider rolls back every unpublished resource. The result carries the child's final output, an optional structured value, a stop reason, and an optional safe diagnostic.
+A request is validated against the provider's advertised capabilities, a durable descriptor is snapshotted, and the provider builds the child. Both in-process providers advertise `agentOptions`: child creation merges requested fields over the provider, model, and reasoning effort in the parent's latest logged request, falls back to creation options before the first request, and retains the configured token limit. They also snapshot delegated permission state before the first await: an Auto or Full access parent gives the child the same `permission/preset` identity, while the existing sandbox override and approval-policy pin continue to apply. Recording both identities prevents an older same-bundle fork value from winning. Auto then reviews every supported child call independently: ordinary project-local work is low risk and allowed, medium-risk work requires explicit action, exact-target and scope authorization from the existing creation prompt or an authenticated human/direct-parent message, without conflicting human limits, while high-risk work is always denied. The reviewer derives that context from `parentSession` and existing messages; delegation adds no parent call metadata, delegation records, review receipt, or Session format. A route change without an explicit effort clears the inherited route-owned effort so the selected model resolves its default. DSH SDK also advertises `agentOptions` but runs a separate child runtime, so it does not inherit Auto; ACP, Codex, and Claude Code likewise retain their own permission systems after the parent delegation call passes review. On success the run is published and ownership transfers to the caller; on failure the provider rolls back every unpublished resource. The result carries the child's final output, an optional structured value, a stop reason, and an optional safe diagnostic.
 
 ### Continuable flow
 
@@ -118,6 +130,7 @@ Read these pages when the package-level contract is not enough. They move from t
 - [Subagent capability seam](../../../.agents/notes/implemented/feature/2026-06-21-subagent-capability-seam.md) — the design record for the delegation capability family.
 - [Continuable subagents](../../../.agents/notes/implemented/feature/2026-07-28-continuable-subagent-conversations.md) — durable children that accept follow-up turns.
 - [In-process spawn backend](../subagent-spawn-in-process/README.md) — the simplest provider to compose.
+- [Auto review](../../experimental/auto-review/README.md) — the current-session authorization mode inherited only by in-process DSH children.
 - [Out-of-process ACP backend](../subagent-acp/README.md) — children with their own runtime over the Agent Client Protocol.
 - [tool-subagent-control README](../tool-subagent-control/README.md) — the follow-up, interrupt, and listing surface.
 
@@ -130,11 +143,11 @@ Read these pages when the package-level contract is not enough. They move from t
 
 #### What the model sees
 
-One user-role parent message opening with the outcome — `Background subagent <child-id> finished and will do no further work unless you send it more.`, or the matching line for a child that was stopped, ran out of room, declined, or failed — followed by `Its closing message:` and the child's final assistant content, or `It left no closing message.` when it produced none. This runtime-owned notice is distinct from model-authored parent/child messages, which use `sendMessage()` and `AgentMessageSource`; delegation schemas and model controls belong to the Consumer packages.
+One user-role parent message opening with the outcome — `Background subagent <child-id> finished and will do no further work unless you send it more.`, or the matching line for a child that was stopped, ran out of room, declined, or failed — followed by `Its closing message:` and the nonempty text blocks from the child's final assistant output, preserving their content and order. Reasoning and other nontext blocks are excluded; when no nonempty text remains, the notice says `It left no closing message.` This runtime-owned notice is distinct from model-authored parent/child messages, which use `sendMessage()` and `AgentMessageSource`; delegation schemas and model controls belong to the Consumer packages.
 
 #### Token effect
 
-One notice per settled Activation in the parent's request, sized by the child's final message. A child that sends its own message and then settles costs the parent both.
+One notice per settled Activation in the parent's request, sized by the child's final text. A child that sends its own message and then settles costs the parent both.
 
 #### KV Cache effect
 
@@ -173,6 +186,7 @@ These limits define when the seam is a poor fit or needs special operational car
 - **Wake gap during cancellation convergence** — a follow-up accepted after an interrupt signal but before the driver becomes idle stays queued until another waking send.
 - **Pending injected context retains an Activation** — settlement conservatively treats every Inbox occurrence as unfinished. Context parked after the Agent becomes idle keeps the child and its live ancestors resident until a waking delivery claims it, a queue mutation removes it, or manager teardown discards it.
 - **Process-local residency** — the Activation inbox and ownership graph do not coordinate two harness processes; concurrent access to one persistence store needs a durable mailbox and cross-process lease protocol.
+- **Saved settlement notices are not rewritten** — a saved user-role notice containing reasoning still fails DeepSeek Messages serialization while it remains in the parent's request history.
 - **No replay of accepted-but-unlogged messages** — a crash can lose an accepted prompt that never reached the child's session log; the lost message is not replayed automatically.
 - **No durable parent mailbox** — child-to-parent messages require a resident continuable child and live direct parent, and provide acceptance identity rather than exactly-once delivery.
 - **Lifecycle events are observe-only** — a run-affecting `subagent/end` continuation or decision API waits for a concrete consumer.

package/README.zh.md

@@ -1,5 +1,5 @@
 ---
-description: "面向用户与维护者的 subagent 委派 seam，用于选择提供方后端、组装委派工具或排查子 agent 运行问题。"
+description: "面向用户与维护者的 subagent 委派 seam，用于选择提供方后端、组装委派工具或排查子 agent（智能体）运行问题。"
 kind: "package-reference"
 ---
 
@@ -9,7 +9,7 @@ kind: "package-reference"
 
 ## 概述
 
-使用 `dsh-subagent` 把工作委派给具名子 agent、收集结果，并跨轮次继续受支持的子级对话。一个组合可以并排提供进程内、ACP、SDK、Codex 或 Claude Code 子级。需要单个结果时选择一次性子级；需要后续消息与中断能力时选择可继续子级。你还可以检查可用子级及其模式、活动状态与血缘，而无需加载或恢复它们。启用时需要至少一个受支持的子级后端和一个委派工具。
+使用 `dsh-subagent` 把工作委派给具名子 agent、收集结果，并跨轮次继续受支持的子级对话。一个组合可以并排提供进程内、ACP（Agent Client Protocol）、SDK、Codex 或 Claude Code 子级。需要单个结果时选择一次性子级；需要后续消息与中断能力时选择可继续子级。你还可以检查可用子级及其模式、活动状态与谱系，而无需加载或恢复它们。启用时需要至少一个受支持的子级后端和一个委派工具。
 
 ## 目录
 
@@ -42,17 +42,29 @@ kind: "package-reference"
 
 调用该工具的 agent 会把子 agent 的最终答案作为工具结果收到。只挂载服务本身不会改变任何行为：在组合出提供方和工具之前，什么都不能委派。
 
+### 委派设置
+
+**插件 → Subagent** 页面的限制部分编辑 Host 的 `subagent` 设置分节。用户值覆盖本插件的组合配置；恢复默认会删除用户覆盖。`maxDepth` 默认为 `1`，在委派工具自身未配置深度时提供默认值。工具显式指定的深度（包括 `provider-managed`）优先。深度 `0` 禁止继承此设置的工具委派；深度 `1` 只允许直接子代理。修改在下一次委派时生效。直接调用服务的调用方仍自行提供可选的请求深度。
+
+### 可续接子代理容量
+
+在 Host 的 `dsh-subagent` 插件上设置 `maxActiveSubagents`，限制通过连续可续接父子关系共享名额的存活子代理数。默认值为 `8`，接受正安全整数。非可续接父代理建立独立的池，自身不占名额；可续接后代继承该池。新建和冷恢复在重建 Agent 前预占名额，清理在 handle 释放后归还名额。等待后代的父代理、有待处理收件箱内容的代理以及正在停止的 Activation 仍占名额。向驻留子代理发送消息复用其名额。一次性和外部提供方运行不受此限制。池的继承不会跨越一次性父代理；其可续接子代理共享独立的池。深度仍由委派工具的独立策略决定。
+
+每次新建或冷恢复 Activation 前都会读取当前 `maxActiveSubagents`。调高后已有树可接纳更多子代理；调低后驻留子代理继续运行，使用量降至上限以下前拒绝新接纳。
+
+容量耗尽时，新建或冷恢复以 `ACTIVATION_LIMIT_REACHED` 拒绝（浏览器消息返回 `subagent/delivery-unavailable`）：等待子代理完成，或继续使用现有代理。接纳不会排队，避免等待后代的父代理又等待自己占用的名额。名额仅存在于当前进程，不限制累计 Session 历史或 token 用量。
+
 ### 一次性与可继续子级
 
-一次性子 agent 只运行一次，并以单个结果结算，可附带可选的结构化输出与失败时的安全诊断。启动请求可以通过 `agentOptions` 覆盖子 Agent 的提供方、模型、推理等级与输出 token 上限；每个请求的选项都要求提供方声明对应能力。可继续子 agent 保留持久会话并按顺序接受后续消息：调用方收到稳定的子 agent id、发送相邻 Agent 消息，并可中断当前轮次而不销毁子 agent。工具行的 `backgroundMode` 选择形态（默认 `one-shot`，或在支持的提供方上使用 `continuable`）。
+一次性子 agent 只运行一次，并以单个结果结算，可附带可选的结构化输出与失败时的安全诊断。启动请求可以通过 `agentOptions` 覆盖子 Agent 的提供方、模型、推理强度与输出 token 上限；每个请求的选项都要求提供方声明对应能力。可继续子 agent 保留持久会话并按顺序接受后续消息：调用方收到稳定的子 agent id、发送相邻 Agent 消息，并可中断当前轮次而不销毁子 agent。工具行的 `backgroundMode` 选择形态（默认 `one-shot`，或在支持的提供方上使用 `continuable`）。
 
 ### 消息、中断与发现
 
-每个确切在线 Agent 都可以对直接可继续 child 使用 `sendMessage()`；驻留的可继续 child 还可以对自己的直接 parent 使用它。正在工作的目标通过 Steer 在最近 step 接收 Agent 消息；空闲目标启动轮次，且只有直接 child 可以冷恢复。parent 也可以随时中断正在运行的后代或列举自己的子级。浏览器发出的继续执行 prompt 会独立选择 Queue 或 Steer，并且可以携带图片部分：Host 先通过附件存储完成整批图片的准入与持久化，子级 inbox 才接受这条消息；当子级声明的模型不接受图片输入时拒绝投递。发现覆盖两种形态：服务列举直接子级与完整后代树——模式、活动状态与血缘——直接读取在线会话状态与可选持久化，不加载任何子 agent。
+每个确切在线 Agent 都可以对直接可继续 child 使用 `sendMessage()`；驻留的可继续 child 还可以对自己的直接 parent 使用它。正在工作的目标通过 Steer 在最近 step 接收 Agent 消息；空闲目标启动轮次，且只有直接 child 可以冷恢复。parent 也可以随时中断正在运行的后代或列举自己的子级。浏览器发出的继续执行提示词会独立选择 Queue 或 Steer，并且可以携带图片部分：Host 先通过附件存储完成整批图片的准入与持久化，子级 inbox 才接受这条消息；当子级声明的模型不接受图片输入时拒绝投递。发现覆盖两种形态：服务列举直接子级与完整后代树——模式、活动状态与谱系——直接读取在线会话状态与可选持久化，不加载任何子 agent。
 
 ### 失败与恢复
 
-需要所选提供方不具备的能力的请求会在启动时响亮失败，而不会被静默忽略。失败的子 agent 运行会返回停止原因，提供方后端还会附加安全诊断；被取消的请求以 `aborted` 结算。子 agent 相互隔离：崩溃或行为异常的子 agent 无法破坏父级会话。
+需要所选提供方不具备的能力的请求会在启动时明确报错，而不会被静默忽略。失败的子 agent 运行会返回停止原因，提供方后端还会附加安全诊断；被取消的请求以 `aborted` 结算。子 agent 相互隔离：崩溃或行为异常的子 agent 无法破坏父级会话。
 
 -----
 
@@ -90,20 +102,20 @@ kind: "package-reference"
 
 ### 一次性流程
 
-请求先对照提供方声明的能力进行校验，随后对持久化描述符做快照，再由提供方构建子 agent。两个进程内提供方都声明 `agentOptions`：创建子级时把请求字段叠加到父级最新已记录请求的提供方、模型与推理等级之上；父级还没有请求时回退到创建选项，并保留配置的 token 上限。更改路由而不显式指定推理等级时，会清除继承的路由自有等级，使所选模型解析自己的默认值。DSH SDK 也声明该能力并公开不可变的 `agentRouteDefaults`，使其实例持有的提供方／模型默认值在确切路由预检前成为基线；`start()` 仍负责直接调用方与输出上限。ACP、Codex 与 Claude Code 会拒绝 agent 路由覆盖，而不是静默忽略。成功时运行被发布、所有权转移给调用方；失败时提供方回滚每个尚未发布的资源。结果携带子 agent 的最终输出、可选的结构化值、停止原因与可选的安全诊断。
+请求先对照提供方声明的能力进行校验，随后对持久化描述符做快照，再由提供方构建子 agent。两个进程内提供方都声明 `agentOptions`：创建子级时把请求字段叠加到父级最新已记录请求的提供方、模型与推理强度之上；父级还没有请求时回退到创建选项，并保留配置的 token 上限。它们还会在第一次 await 前快照委派权限状态：Auto 或 Full access 父级让子级获得相同的 `permission/preset` 身份，而既有沙箱覆盖与审批策略固定仍然生效；同时记录这两个身份可防止 fork 中更早的同旋钮组合身份胜出。Auto 随后会独立审查 child 的每个受支持调用：普通项目内工作为低风险并直接允许；中风险工作必须在既有创建 prompt 或已核验的 human／直接父级消息中获得动作、准确目标和范围的明确授权，且不与 human 限制冲突；高风险工作始终拒绝。reviewer 从 `parentSession` 与既有消息派生这份上下文；委派不会新增父 call metadata、委派记录、review receipt 或 Session format。更改路由而不显式指定推理强度时，会清除继承的路由自有强度，使所选模型解析自己的默认值。DSH SDK 也声明 `agentOptions`，但会运行独立子运行时，因此不继承 Auto；ACP、Codex 与 Claude Code 同样在父级委派调用通过审查后保留各自的权限系统。成功时运行被发布、所有权转移给调用方；失败时提供方回滚每个尚未发布的资源。结果携带子 agent 的最终输出、可选的结构化值、停止原因与可选的安全诊断。
 
 ### 可继续流程
 
-管理器预留 child 身份、解析持久化描述符、创建（或冷恢复）child、把它安装进 Activation 并提交提示词。模型编写的消息通过固定 Steer 调度跨一条 parent/child 边；浏览器人类 prompt 通过内部适配器选择 Queue 或 best-effort Steer，其他 host 协议仍可保留 Queue 以创建独立轮次。Session queue command 仅根据 child 自身的 continuable descriptor 准入在线 subagent-owned Agent。Settlement 会等待 Agent 活动结束、Inbox 为空且没有所拥有子级，再在准入开放时 flush 最终 Session 状态。管理器随后在 child lock 内重新验证 wake generation、Session 序号、Inbox 与所拥有子级；`Agent.runMaintenance()` 的同步 task 入口会占用 idle 阶段，并在同一个 JavaScript turn 内关闭私有 subagent Inbox，然后才 dispose handle。直接 child 不存在 Activation 时会从持久化会话冷恢复。当驻留 Activation 结算时，管理器会在 parent 自身的轮次流中告知该 child 的直接 parent。
+管理器预留 child 身份、解析持久化描述符、创建（或冷恢复）child、把它安装进 Activation 并提交提示词。模型编写的消息通过固定 Steer 调度跨一条 parent/child 边；浏览器人类 prompt 通过内部适配器选择 Queue 或 best-effort Steer，其他 host 协议仍可保留 Queue 以创建独立轮次。Session queue command 仅根据 child 自身的 continuable descriptor 准入在线 subagent-owned Agent。Settlement 会等待 Agent 活动结束、Inbox 为空且没有所拥有子级，再在准入开放时 flush 最终 Session 状态。管理器随后在 child lock 内重新验证 wake generation、Session 序号、Inbox 与所拥有子级；`Agent.runMaintenance()` 的同步 task 入口会占用 idle 阶段，并在同一个 JavaScript turn 内关闭私有 subagent Inbox，然后才释放句柄。直接 child 不存在 Activation 时会从持久化会话冷恢复。当驻留 Activation 结算时，管理器会在 parent 自身的轮次流中告知该 child 的直接 parent。
 
-本地子级创建成功时，父 Session 追加一条 `subagent/catalog` 事实。一次性创建在 provider 返回后记录；可继续创建在初始 inbox 准入后、返回子级 id 前记录。失败会释放子级，不发布补偿性目录事件。一次性目录追加失败时会处理 run 的结果拒绝，并保留目录错误；资源释放失败会单独记录。`subagentCatalog` projection 排除 fork 继承的事实，通过 Session 观察和客户端快照中的 `projections.values.subagentCatalog` 暴露直接子级列表。无效的自身 catalog payload（包括不支持的版本）会使 projection 恢复失败。其不可变存储和检查点校验使用 [`dsh-chunked-list`](../../util/chunked-list/README.zh.md)。其视图对 D 条事实以 O(D) 时间保留父目录事件顺序。[父目录决策](../../../.agents/notes/implemented/architecture/2026-09-01-parent-owned-subagent-catalog.zh.md) 说明排序、持久化成本和替代方案。
+本地子级创建成功时，父 Session 追加一条 `subagent/catalog` 事实。一次性创建在提供方返回后记录；可继续创建在初始 inbox 准入后、返回子级 id 前记录。失败会释放子级，不发布补偿性目录事件。一次性目录追加失败时会处理 run 的结果拒绝，并保留目录错误；资源释放失败会单独记录。`subagentCatalog` projection 排除 fork 继承的事实，通过 Session 观察和客户端快照中的 `projections.values.subagentCatalog` 暴露直接子级列表。无效的自身 catalog payload（包括不支持的版本）会使 projection 恢复失败。其不可变存储和检查点校验使用 [`dsh-chunked-list`](../../util/chunked-list/README.zh.md)。其视图对 D 条事实以 O(D) 时间保留父目录事件顺序。[父目录决策](../../../.agents/notes/implemented/architecture/2026-09-01-parent-owned-subagent-catalog.zh.md) 说明排序、持久化成本和替代方案。
 
 ### 所有权与不变式
 
 - **发布即边界**——发布前提供方拥有设置并须在失败时回滚；发布后调用方拥有运行并须 dispose（资源释放）它。
 - **注册受 effect 作用域约束**——移除提供方会阻止新启动，但绝不撤销已接受的运行。
 - **Agent 消息权限基于确切相邻关系**——`sendMessage()` 要求确切在线 sender；每个 sender 都可以指定直接可继续 child，只有具备驻留可继续 Activation 的 sender 可以指定自己的直接 parent。
-- **描述符仅进日志**——它是会话事件，不进入模型历史，并跨压缩（compaction）保留；可继续描述符会显式记录解析后的子级提供方、模型与推理等级，用于冷恢复。
+- **描述符仅进日志**——它是会话事件，不进入模型历史，并跨压缩（compaction）保留；可继续描述符会显式记录解析后的子级提供方、模型与推理强度，用于冷恢复。
 
 </details>
 
@@ -118,6 +130,7 @@ kind: "package-reference"
 - [Subagent 能力 seam](../../../.agents/notes/implemented/feature/2026-06-21-subagent-capability-seam.zh.md)——委派能力家族的设计记录。
 - [可继续的 subagent](../../../.agents/notes/implemented/feature/2026-07-28-continuable-subagent-conversations.zh.md)——接受后续轮次的持久子级。
 - [进程内 spawn 后端](../subagent-spawn-in-process/README.zh.md)——最容易组合的提供方。
+- [Auto review](../../experimental/auto-review/README.zh.md)——只有进程内 DSH 子级继承的当前会话授权模式。
 - [进程外 ACP 后端](../subagent-acp/README.zh.md)——经 Agent Client Protocol 拥有自有运行时的子级。
 - [tool-subagent-control README](../tool-subagent-control/README.zh.md)——后续消息、中断与列举面。
 
@@ -130,11 +143,11 @@ kind: "package-reference"
 
 #### 模型看到什么
 
-一条用户角色的父级消息，开头是结果本身——`Background subagent <child-id> finished and will do no further work unless you send it more.`，或子级被停止、耗尽额度、拒绝任务或失败时的对应句子——随后是 `Its closing message:` 与子级的最终 assistant 内容；若子级没有产出内容，则是 `It left no closing message.`。这条由 runtime 生成的通知与模型编写的父子消息相互独立；后者使用 `sendMessage()` 与 `AgentMessageSource`。委派 schema 与模型控制工具归 Consumer 包所有。
+一条用户角色的父级消息，开头是结果本身——`Background subagent <child-id> finished and will do no further work unless you send it more.`，或子级被停止、耗尽额度、拒绝任务或失败时的对应句子——随后是 `Its closing message:` 与子级最终 assistant 输出中的非空文本块，保留原始内容与顺序。推理与其他非文本块不会进入通知；若没有剩余的非空文本，通知会写明 `It left no closing message.`。这条由运行时生成的通知与模型编写的父子消息相互独立；后者使用 `sendMessage()` 与 `AgentMessageSource`。委派 schema 与模型控制工具归消费方包所有。
 
 #### Token 影响
 
-父级请求中，每个已结算的 Activation 一条通知，长度取决于子级的最终消息。如果子级先发送自己的消息再结算，父级请求会同时承担两者。
+父级请求中，每个已结算的 Activation 一条通知，长度取决于子级的最终文本。如果子级先发送自己的消息再结算，父级请求会同时承担两者。
 
 #### KV Cache 影响
 
@@ -171,8 +184,9 @@ You are a delegated subagent: your permission scope was fixed when you were star
 - **仅允许相邻模型消息**——`sendMessage()` 要求确切在线 sender；每个 sender 都可以指定直接可继续 child，只有具备驻留可继续 Activation 的 sender 可以指定自己的直接 parent。浏览器提示使用独立的人类 Queue 或 Steer 控制路径。
 - **child 到 parent 的投递要求直接 parent 保持在线**——服务没有持久 parent mailbox；parent 缺失时会拒绝消息，而非接受无法唤醒的工作。
 - **取消收敛期间存在唤醒缺口**——中断信号发出后、driver 进入 idle 前被接受的后续消息会保持排队，直到另一条唤醒发送到达。
-- **待处理的注入 context 会保留 Activation**——settlement 会保守地把每个 Inbox occurrence 都视为未完成。Agent 进入 idle 后停放的 context 会让 child 及其在线祖先继续驻留，直到唤醒投递将其 claim、queue 变更将其移除，或 manager teardown 将其丢弃。
+- **待处理的注入上下文会保留 Activation**——settlement 会保守地把每个 Inbox occurrence 都视为未完成。Agent 进入 idle 后停放的上下文会让 child 及其在线祖先继续驻留，直到唤醒投递将其 claim、queue 变更将其移除，或 manager teardown 将其丢弃。
 - **驻留仅限进程内**——Activation inbox 与所有权图不会在两个 harness 进程之间协调；对单个持久化存储的并发访问需要持久化邮箱与跨进程租约协议。
+- **已保存的结算通知不会被改写**——若用户角色的已保存通知含有推理块，只要它仍在父级请求历史中，DeepSeek Messages 序列化就会失败。
 - **不回放已接受但未记录的消息**——崩溃可能丢失从未写入子会话日志、已被接受的提示词；丢失的消息不会自动回放。
 - **没有持久化 parent mailbox**——child 到 parent 的消息要求驻留的可继续 child 与在线直接 parent，提供的是接受标识，不保证恰好一次投递。
 - **生命周期事件只供观察**——影响运行的 `subagent/end` 延续或决策接口仍需等待具体消费方。

package/lib/index.js

@@ -1,9 +1,10 @@
+import z from "@deepseek-ai/schemastery";
 import { scopeTarget } from "@deepseek-ai/dsh-scope";
 import { assertObjectJsonSchema } from "@deepseek-ai/dsh-tools";
 import { canonicalClientTimeZone } from "@deepseek-ai/dsh-util-time";
 import { Remote, RemoteError, TypertRemoteService } from "@deepseek-ai/dsh-typert-protocol";
 import { AttachmentError } from "@deepseek-ai/dsh-attachment";
-import { z } from "zod";
+import { z as z$1 } from "zod";
 import { HarnessError, ReasoningEffortId, boundContextSummary, contentHasImage, createUserMessage, errorChain, joinAssistantStreamText } from "@deepseek-ai/dsh-llm";
 import { randomUUID } from "node:crypto";
 import { foldConsumedWork } from "@deepseek-ai/dsh-agent";
@@ -35,19 +36,19 @@ var SubagentError = class extends HarnessError {
 *
 * @module @deepseek-ai/dsh-subagent
 */
-const SESSION_ID_SCHEMA = z.string().min(1);
+const SESSION_ID_SCHEMA = z$1.string().min(1);
 const CONTROL_ID_SCHEMAS = {
-	"subagent.list": z.object({ parentSessionId: SESSION_ID_SCHEMA }),
-	"subagent.prompt": z.object({
+	"subagent.list": z$1.object({ parentSessionId: SESSION_ID_SCHEMA }),
+	"subagent.prompt": z$1.object({
 		parentSessionId: SESSION_ID_SCHEMA,
 		childSessionId: SESSION_ID_SCHEMA,
-		mode: z.literal("continuable"),
-		delivery: z.enum(["queue", "steer"])
+		mode: z$1.literal("continuable"),
+		delivery: z$1.enum(["queue", "steer"])
 	}),
-	"subagent.interrupt": z.object({
+	"subagent.interrupt": z$1.object({
 		parentSessionId: SESSION_ID_SCHEMA,
 		childSessionId: SESSION_ID_SCHEMA,
-		mode: z.literal("continuable")
+		mode: z$1.literal("continuable")
 	})
 };
 /**
@@ -113,6 +114,7 @@ function rejectPrompt(error, childSessionId, signal) {
 		case "UNAUTHORIZED": throw new RemoteError("subagent/unauthorized", "subagent does not belong to this parent", { childSessionId }, { cause: error });
 		case "DRAINING":
 		case "ACTIVATION_CLOSING":
+		case "ACTIVATION_LIMIT_REACHED":
 		case "CONTINUATION_UNAVAILABLE":
 		case "PERSISTENCE_UNAVAILABLE": throw new RemoteError("subagent/delivery-unavailable", "subagent follow-up is temporarily unavailable", { childSessionId }, { cause: error });
 		default: break;
@@ -554,17 +556,20 @@ function applyChildComposition(childCtx, parent, composition) {
 	if (composition.toolFilter !== void 0) childCtx.tools.restrict(composition.toolFilter);
 }
 /**
-* Capture the policy to seed into one delegation. Call synchronously before
+* Capture the permission state to seed into one delegation. Call synchronously before
 * the child start's first await: a later parent switch belongs to the
-* parent's future, not to this child. Only the parent session's explicit
-* sandbox override is captured — never deployment defaults or one-shot
-* grants — and the approval policy is pinned to `'never'` regardless of the
-* parent's own policy.
+* parent's future, not to this child. Auto and Full access identities are
+* inherited only through the in-process DSH path so either can replace a stale
+* same-bundle fork value. Only the parent session's explicit sandbox override
+* is captured — never deployment defaults or one-shot grants — and the approval
+* policy is pinned to `'never'` regardless of the parent's own policy.
 * @param parent - the delegating parent agent.
 * @returns the sandbox override (or `undefined` without one) and the approval pin.
 */
 function captureDelegatedPolicyOverrides(parent) {
+	const preset = parent.ctx.get("permissionPresets")?.current(parent.session);
 	return {
+		permissionPreset: preset === "auto" || preset === "danger-full-access" ? preset : void 0,
 		sandboxMode: parent.ctx.get("sandboxPolicy")?.overrideOf(parent.session),
 		approvalPolicy: parent.ctx.get("approval") === void 0 ? void 0 : "never"
 	};
@@ -587,6 +592,7 @@ function appendDelegatedPolicyOverrides(childSession, overrides) {
 		policy: overrides.approvalPolicy,
 		source: "delegation"
 	});
+	if (overrides.permissionPreset !== void 0) childSession.append("permission/preset", { preset: overrides.permissionPreset });
 }
 //#endregion
 //#region lib/types/continuation-messages.js
@@ -653,24 +659,25 @@ function settlementSummary(childId, stopReason) {
 	}
 }
 /**
-* Build the runtime-owned settlement notice delivered to a child's parent.
+* Build the runtime-owned settlement notice from the child's nonempty closing text.
 * @param childId - durable child session id named in the notice.
 * @param terminal - recorded terminal state for the settled Activation.
 * @returns the durable user-message representation delivered to the parent.
 */
 function createSettlementMessage(childId, terminal) {
 	const summary = settlementSummary(childId, terminal.stopReason);
+	const closingText = (terminal.output ?? []).flatMap((block) => block.type === "text" && block.text.length > 0 ? [block] : []);
 	return createUserMessage({
 		content: [{
 			type: "text",
 			text: summary
-		}, ...terminal.output === void 0 ? [{
+		}, ...closingText.length === 0 ? [{
 			type: "text",
 			text: "It left no closing message."
 		}] : [{
 			type: "text",
 			text: "Its closing message:"
-		}, ...terminal.output]],
+		}, ...closingText]],
 		source: {
 			kind: "subagent-settled",
 			form: "notice",
@@ -747,6 +754,19 @@ var SubagentInbox = class {
 *
 * @module @deepseek-ai/dsh-subagent/continuation-activation
 */
+/** Process-local slots shared through uninterrupted continuable parent links. */
+var ActivationPool = class {
+	slots = /* @__PURE__ */ new Set();
+	/** Reserve before reconstruction; the returned release also tolerates unpublished rollback. */
+	reserve(capacity) {
+		if (this.slots.size >= capacity) throw new SubagentError(`subagent limit reached (active child limit: ${capacity}); wait for an existing child to finish or complete this work with the current agents`, "ACTIVATION_LIMIT_REACHED");
+		const slot = Symbol();
+		this.slots.add(slot);
+		return () => {
+			this.slots.delete(slot);
+		};
+	}
+};
 /** Serialize each durable child's delivery, release, and disposal. */
 var ChildLock = class {
 	tails = /* @__PURE__ */ new Map();
@@ -770,8 +790,11 @@ var ChildLock = class {
 var ContinuableActivationRegistry = class {
 	ctx;
 	observeActivation;
+	maxActiveSubagents;
 	/** Child session id → its live Activation. Process-local, never durable. */
 	resident = /* @__PURE__ */ new Map();
+	/** Root identities retain their pool across child settlement without retaining dead roots. */
+	rootPools = /* @__PURE__ */ new WeakMap();
 	/** Materializations admitted before drain, tracked through publication or rollback. */
 	materializations = /* @__PURE__ */ new Set();
 	/** Per-child serializer shared by delivery, release, and disposal. */
@@ -791,9 +814,10 @@ var ContinuableActivationRegistry = class {
 	* @param ctx - context providing Agents, Sessions, and teardown ownership.
 	* @param observeActivation - build the lifecycle observer for one residency epoch.
 	*/
-	constructor(ctx, observeActivation) {
+	constructor(ctx, observeActivation, maxActiveSubagents) {
 		this.ctx = ctx;
 		this.observeActivation = observeActivation;
+		this.maxActiveSubagents = maxActiveSubagents;
 		const scope = ctx.plugin(function activationOwner() {});
 		this.ownerCtx = scope.ctx;
 		ctx.on("agent/disposed", ({ agent }) => {
@@ -974,14 +998,20 @@ var ContinuableActivationRegistry = class {
 	*/
 	materialize(inputs) {
 		this.assertAdmitting(inputs.parent);
-		const settled = Promise.withResolvers();
+		inputs.signal.throwIfAborted();
 		const lineage = this.liveLineage(inputs.parent);
+		const pool = this.resident.get(inputs.parent.id)?.pool ?? this.rootPool(inputs.parent);
+		const releaseSlot = pool.reserve(this.maxActiveSubagents());
+		const settled = Promise.withResolvers();
 		const materialization = {
 			lineage,
 			settled: settled.promise
 		};
 		this.materializations.add(materialization);
-		return this.materializeTracked(inputs, lineage).finally(() => {
+		return this.materializeTracked(inputs, lineage, pool, releaseSlot).catch((error) => {
+			releaseSlot();
+			throw error;
+		}).finally(() => {
 			this.materializations.delete(materialization);
 			settled.resolve();
 		});
@@ -1056,8 +1086,17 @@ var ContinuableActivationRegistry = class {
 		const lineage = this.liveLineage(agent);
 		for (const [root, members] of this.closingScopes) if (members.has(agent) || lineage.includes(root)) return root;
 	}
+	/** Resolve a root's pool once; descendants inherit their resident parent's pool directly. */
+	rootPool(parent) {
+		let pool = this.rootPools.get(parent);
+		if (pool === void 0) {
+			pool = new ActivationPool();
+			this.rootPools.set(parent, pool);
+		}
+		return pool;
+	}
 	/** Perform one tracked materialization through publication or rollback. */
-	async materializeTracked(inputs, parentLineage) {
+	async materializeTracked(inputs, parentLineage, pool, releaseSlot) {
 		const { childId, provider, parent, create } = inputs;
 		inputs.signal.throwIfAborted();
 		const setup = (childCtx, child) => {
@@ -1085,6 +1124,8 @@ var ContinuableActivationRegistry = class {
 			setup
 		});
 		const activation = {
+			pool,
+			releaseSlot,
 			childId,
 			parentSession: parent.id,
 			provider,
@@ -1123,6 +1164,7 @@ var ContinuableActivationRegistry = class {
 				await activation.handle.dispose();
 			} finally {
 				this.resident.delete(activation.childId);
+				activation.releaseSlot();
 				this.releaseOwnership(activation.childId);
 			}
 		});
@@ -1236,6 +1278,7 @@ var ContinuableActivationRegistry = class {
 		if (failures.length === 1) failure = failures[0];
 		else if (failures.length > 1) failure = new SubagentError(`subagent "${childId}" activation teardown failed at ${failures.length} boundaries: ` + failures.map((item) => errorChain(item)).join("; "), "ACTIVATION_TEARDOWN_FAILED", { cause: new AggregateError(failures) });
 		this.resident.delete(childId);
+		activation.releaseSlot();
 		this.notifySettlement(activation, activation.observer.terminal(failure));
 		this.releaseOwnership(childId);
 		activation.observer.settle(failure);
@@ -1427,23 +1470,23 @@ function foldSubagentDescriptor(events) {
 	if (event === void 0) return void 0;
 	return parseSubagentDescriptor(event.data);
 }
-const sessionIdSchema = z.string();
-const oneShotCatalogSchema = z.object({
-	version: z.literal(0),
+const sessionIdSchema = z$1.string();
+const oneShotCatalogSchema = z$1.object({
+	version: z$1.literal(0),
 	childId: sessionIdSchema,
-	childCreatedAt: z.number().int().nonnegative(),
-	mode: z.literal("one-shot"),
-	label: z.string().optional()
+	childCreatedAt: z$1.number().int().nonnegative(),
+	mode: z$1.literal("one-shot"),
+	label: z$1.string().optional()
 }).strict();
-const continuableCatalogSchema = z.object({
-	version: z.literal(0),
+const continuableCatalogSchema = z$1.object({
+	version: z$1.literal(0),
 	childId: sessionIdSchema,
-	childCreatedAt: z.number().int().nonnegative(),
-	mode: z.literal("continuable"),
-	label: z.string()
+	childCreatedAt: z$1.number().int().nonnegative(),
+	mode: z$1.literal("continuable"),
+	label: z$1.string()
 }).strict();
-const eventDataSchema = z.union([oneShotCatalogSchema, continuableCatalogSchema]);
-const viewSchema = z.array(z.union([oneShotCatalogSchema.omit({
+const eventDataSchema = z$1.union([oneShotCatalogSchema, continuableCatalogSchema]);
+const viewSchema = z$1.array(z$1.union([oneShotCatalogSchema.omit({
 	version: true,
 	childId: true,
 	childCreatedAt: true
@@ -1458,8 +1501,8 @@ const viewSchema = z.array(z.union([oneShotCatalogSchema.omit({
 	id: sessionIdSchema,
 	createdAt: continuableCatalogSchema.shape.childCreatedAt
 })]));
-const stateSchema = z.object({
-	inheritedEventCount: z.number().int().nonnegative(),
+const stateSchema = z$1.object({
+	inheritedEventCount: z$1.number().int().nonnegative(),
 	head: chunkedListSchema(eventDataSchema).optional()
 }).strict();
 /**
@@ -1628,10 +1671,10 @@ var SubagentContinuationManager = class {
 	ctx;
 	host;
 	activations;
-	constructor(ctx, host) {
+	constructor(ctx, host, maxActiveSubagents) {
 		this.ctx = ctx;
 		this.host = host;
-		this.activations = new ContinuableActivationRegistry(ctx, (provider, childId, parent) => host.observeActivation(provider, childId, parent));
+		this.activations = new ContinuableActivationRegistry(ctx, (provider, childId, parent) => host.observeActivation(provider, childId, parent), maxActiveSubagents);
 	}
 	/**
 	* Start one continuable background child and resolve at initial inbox acceptance.
@@ -2317,12 +2360,12 @@ function sessionQueryCode(error) {
 *
 * @module @deepseek-ai/dsh-subagent/projection
 */
-const activeIntervalSchema = z.object({
-	since: z.number().int().nonnegative(),
-	through: z.number().int().nonnegative()
+const activeIntervalSchema = z$1.object({
+	since: z$1.number().int().nonnegative(),
+	through: z$1.number().int().nonnegative()
 }).strict();
-const projectionSchema = z.object({
-	settledMs: z.number().int().nonnegative(),
+const projectionSchema = z$1.object({
+	settledMs: z$1.number().int().nonnegative(),
 	active: activeIntervalSchema.optional()
 }).strict().transform(({ settledMs, active }) => ({
 	settledMs,
@@ -2338,11 +2381,11 @@ const projectionSchema = z.object({
 */
 const subagentTimingProjectionDefinition = {
 	key: "subagentTiming",
-	stateSchema: z.object({
-		settledMs: z.number().int().nonnegative(),
+	stateSchema: z$1.object({
+		settledMs: z$1.number().int().nonnegative(),
 		active: activeIntervalSchema.optional(),
-		pendingTurnStart: z.number().int().nonnegative().optional(),
-		descriptorSeen: z.boolean()
+		pendingTurnStart: z$1.number().int().nonnegative().optional(),
+		descriptorSeen: z$1.boolean()
 	}).strict(),
 	init: () => ({
 		descriptorSeen: false,
@@ -2401,17 +2444,17 @@ const subagentTimingProjectionDefinition = {
 	},
 	stateVersion: 2
 };
-const identityValueSchema = z.discriminatedUnion("mode", [z.object({
-	mode: z.literal("one-shot"),
-	label: z.string().optional(),
-	seq: z.number().int().nonnegative().max(Number.MAX_SAFE_INTEGER).transform(SessionSeq)
-}).strict(), z.object({
-	mode: z.literal("continuable"),
-	label: z.string(),
-	seq: z.number().int().nonnegative().max(Number.MAX_SAFE_INTEGER).transform(SessionSeq)
+const identityValueSchema = z$1.discriminatedUnion("mode", [z$1.object({
+	mode: z$1.literal("one-shot"),
+	label: z$1.string().optional(),
+	seq: z$1.number().int().nonnegative().max(Number.MAX_SAFE_INTEGER).transform(SessionSeq)
+}).strict(), z$1.object({
+	mode: z$1.literal("continuable"),
+	label: z$1.string(),
+	seq: z$1.number().int().nonnegative().max(Number.MAX_SAFE_INTEGER).transform(SessionSeq)
 }).strict()]);
 const identitySchema = identityValueSchema.nullable();
-const identityStateSchema = z.object({ identity: identityValueSchema.optional() }).strict();
+const identityStateSchema = z$1.object({ identity: identityValueSchema.optional() }).strict();
 /** Interpret one `subagent/descriptor` event's identity; no value when the payload cannot be trusted. */
 function descriptorIdentity(event) {
 	let descriptor;
@@ -2841,7 +2884,12 @@ let SubagentRuntime = (() => {
 				value: _metadata
 			});
 		}
-		providers = (__runInitializers(this, _instanceExtraInitializers), /* @__PURE__ */ new Map());
+		static Config = z.object({
+			maxDepth: z.number().step(1).min(0).max(Number.MAX_SAFE_INTEGER).default(1),
+			maxActiveSubagents: z.number().step(1).min(1).max(Number.MAX_SAFE_INTEGER).default(8)
+		});
+		settingsSource = __runInitializers(this, _instanceExtraInitializers);
+		providers = /* @__PURE__ */ new Map();
 		continuations;
 		/**
 		* The contained lifecycle-edge publisher. Built here because scoped dispatch
@@ -2849,14 +2897,27 @@ let SubagentRuntime = (() => {
 		* composes into the carrier.
 		*/
 		emitLifecycle;
-		constructor(ctx) {
+		constructor(ctx, config) {
 			super(ctx, "subagents");
+			assertSubagentMaxDepth(config.maxDepth);
+			this.settingsSource = () => config;
+			ctx.inject(["settings"], (settingsCtx) => {
+				settingsCtx.settings.installSection(ctx, "subagent", SubagentRuntime.Config, config, {
+					validate: (value) => {
+						assertSubagentMaxDepth(value.maxDepth);
+					},
+					setSource: (source) => {
+						this.settingsSource = source;
+					},
+					onChange: () => {}
+				});
+			});
 			this.emitLifecycle = createLifecycleEmitter(this.ctx, (parent) => scopeTarget(this, parent));
 			ctx.inject(["agents"], (childCtx) => {
 				const manager = new SubagentContinuationManager(childCtx, {
 					prepareContinuable: (name, request) => this.prepareContinuable(name, request),
 					observeActivation: (provider, childId, parent) => this.observeActivation(provider, childId, parent)
-				});
+				}, () => this.settingsSource().maxActiveSubagents);
 				this.continuations = manager;
 				childCtx.effect(() => () => {
 					/* v8 ignore else -- one injected binding owns the slot until its fiber disposes. */
@@ -2870,6 +2931,15 @@ let SubagentRuntime = (() => {
 			});
 		}
 		/**
+		* Resolve a delegation tool's depth policy against the current user setting.
+		* @param configured - Explicit tool limit, or provider-managed for external delegation.
+		* @returns The numeric limit, or undefined when the provider owns depth enforcement.
+		*/
+		resolveMaxDepth(configured) {
+			if (configured === "provider-managed") return void 0;
+			return configured ?? this.settingsSource().maxDepth;
+		}
+		/**
 		* Establish one durable continuable child and deliver its initial prompt.
 		* Resolves when the child's inbox accepts that prompt, without waiting for the
 		* turn to start or for the message to reach the Session log; any earlier
@@ -2900,12 +2970,12 @@ let SubagentRuntime = (() => {
 		}
 		/**
 		* Deliver one host-protocol message to a direct continuable child.
-		* Symbol-keyed so host adapters can preserve their own provenance without
+		* Symbol-keyed so host adapters can preserve their own source descriptors without
 		* widening the public Service Definition or impersonating an Agent sender.
 		* @param parent - exact live direct parent authorizing delivery.
 		* @param childId - durable direct-child session id.
 		* @param content - host-authored content to deliver.
-		* @param source - durable host-protocol provenance.
+		* @param source - durable host-protocol source descriptor.
 		* @param signal - caller cancellation before inbox acceptance.
 		* @param delivery - Queue as a distinct turn or Steer at the nearest step.
 		* @returns the accepted message's inbox id.
@@ -3029,6 +3099,7 @@ let SubagentRuntime = (() => {
 		* nearest step and retains the Agent loop's best-effort fallback semantics.
 		* Image parts are admitted and persisted through the attachment store
 		* before delivery, and the child's model must accept image input.
+		* Cold resume at capacity rejects with `subagent/delivery-unavailable`.
 		* @param request - durable address, delivery, minted identity, content, and optional browser zone.
 		* @param signal - carrier cancellation, owning the call until inbox acceptance.
 		* @returns the accepted message's inbox identity.