Merge branch 'origin/master' into feat/toolcard-image-result

2026-09-01 14:45:24 +08:00
commit a23c3dd64e
@@ -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 .agents/notes/implemented/architecture/2026-06-11-dev-invariants-over-deep-readonly.md
-2026-06-11-dev-invariants-over-deep-readonly.md: 66980f1ee09c6112f72786d6c3a147aadbc57f6c
-2026-06-11-dev-invariants-over-deep-readonly.zh.md: 576e53e0b27e65f6fa071ff649509223a7bc30ff
+2026-06-11-dev-invariants-over-deep-readonly.md: 7e5f55e8910797bd46674050ea5eb8abbca4aef7
+2026-06-11-dev-invariants-over-deep-readonly.zh.md: c1413e9c89284312af41b6c115b7e50a99d1b5e3
@@ -22,7 +22,7 @@ Responsibility is split between an always-on storage boundary and optional devel

 `Session` accepts an event only after one recursive pass has materialized a lossless JSON snapshot. That pass rejects unsupported values and produces the exact detached record that enters the log, so validation and storage cannot observe different values from a stateful getter or retain caller-owned nested references.

-The accepted event and all of its descendants are deep-frozen before publication. `append()` returns that owned frozen event, `session/event` observers receive the same record, and `session.events` returns a frozen array snapshot. A previously returned array does not grow after a later append. Seed records pass through the same validation, snapshot, and freeze boundary before construction succeeds.
+The accepted event and all of its descendants are deep-frozen before publication. `append()` returns that owned frozen event, and `session/event` observers and `eventAt(seq)` receive the same record. `snapshotEvents(fromSeq?, toSeqExclusive?)` returns a frozen array snapshot; a previously returned array does not grow after a later append. `seq` and `eventAt()` avoid array materialization when a caller needs only the current length or one event. Seed records pass through the same validation, snapshot, and freeze boundary before construction succeeds.

 This guarantee belongs in `Session`, not in an optional listener, because every composition relies on trustworthy history. A production deployment, a focused test, or a custom embedding receives the same storage semantics whether or not development support plugins are registered.

@@ -32,7 +32,7 @@ This guarantee belongs in `Session`, not in an optional listener, because every

 ### Package-owned invariant companions check relationships

-`dsh-invariants` registers the configurable `ctx.invariants` service and contains no product checks. Every package publishes a `./invariant` ownership companion; `dsh-session`, `dsh-agent`, `dsh-scope`, and `dsh-agent-loop` currently add the rules that require trace state or observation of another seam: monotonic sequence numbers, turn and step nesting, tool-call/result pairing, legal agent-status transitions, subject-correct scoped dispatch, and equality between a loop-built request and the request reconstructed from its session-log prefix. Global enablement and package-name regex filters belong to the service ([package-owned invariant service](2026-07-19-package-owned-invariant-service.md)).
+`dsh-invariants` registers the configurable `ctx.invariants` service and contains no product checks. A package publishes a `./invariant` ownership companion only for an independently observable runtime relationship; packages without one omit the companion and record the reason in their README. `dsh-session`, `dsh-agent`, `dsh-scope`, and `dsh-agent-loop` provide the initial rules that require trace state or observation of another seam: monotonic sequence numbers, turn and step nesting, tool-call/result pairing, legal agent-status transitions, subject-correct scoped dispatch, and equality between a loop-built request and the request reconstructed from its session-log prefix. Global enablement and package-name regex filters belong to the service ([package-owned invariant service](2026-07-19-package-owned-invariant-service.md); [omission decision](../simplification/2026-08-28-omit-unneeded-invariant-companions.md)).

 When the session companion attaches to an existing or seeded session, it replays the immutable log to rebuild trace state. The service gives each contribution a disposable child fiber, so hot reload is safe in the middle of a turn without giving diagnostics ownership of session storage.

@@ -48,12 +48,12 @@ Freezing history only when an invariants plugin is installed would make the core

 ### Clone only when deriving messages

-Detaching `deriveMessages()` would protect the most common request path but leave other readers of `session.events`, append return values, and session-event observers able to mutate durable history. The log must protect its own boundary; derived projections are an additional isolation boundary, not a substitute.
+Detaching `deriveMessages()` would protect the most common request path but leave other readers of `snapshotEvents()`, `eventAt()`, append return values, and session-event observers able to mutate durable history. The log must protect its own boundary; derived projections are an additional isolation boundary, not a substitute.

 ## Consequences

 - Every accepted live or seeded session event is detached from caller-owned inputs and deeply immutable before any observer can receive it.
- `session.events` exposes stable immutable snapshots instead of the private growing array.
+- `snapshotEvents()` exposes stable immutable snapshots instead of the private growing array; `seq` and `eventAt()` serve scalar reads without copying that array.
 - Request-side mutation cannot reach stored history through derived messages.
 - Development builds can enable relational assertions without changing storage behavior, and disposing or filtering a companion does not weaken log immutability.
 - `dsh-invariants` configures global enablement plus package allow/block regex lists; each check remains owned and tested by its product package.
@@ -22,7 +22,7 @@ TypeScript readonly 类型不是充分的运行时边界。它们在程序运行

 `Session` 仅在一次递归遍历完成无损 JSON 快照的物化之后才接受事件。该遍历拒绝不支持的值，并产出进入日志的已分离的确切记录，因此验证与存储不会从有状态的 getter 观察到不同的值，也不会保留调用方拥有的嵌套引用。

-被接受的事件及其所有后代在发布前被深度冻结。`append()` 返回由 Session 拥有的冻结事件，`session/event` 观察者接收同一记录，`session.events` 返回冻结的数组快照。先前返回的数组不会因后续 append 而增长。种子记录在构造成功前经过相同的验证、快照与冻结边界。
+被接受的事件及其所有后代在发布前被深度冻结。`append()` 返回由 Session 拥有的冻结事件，`session/event` 观察者和 `eventAt(seq)` 接收同一记录。`snapshotEvents(fromSeq?, toSeqExclusive?)` 返回冻结的数组快照；先前返回的数组不会因后续 append 而增长。调用方只需要当前长度或单个事件时，`seq` 和 `eventAt()` 不会物化数组。种子记录在构造成功前经过相同的验证、快照与冻结边界。

 此保证属于 `Session` 而非可选监听器，因为每种组合都依赖可信的历史。无论是否注册了开发支持插件，生产部署、聚焦测试或自定义嵌入都获得相同的存储语义。

@@ -32,7 +32,7 @@ TypeScript readonly 类型不是充分的运行时边界。它们在程序运行

 ### 包拥有的不变式配套插件检查关系

-`dsh-invariants` 注册可配置的 `ctx.invariants` 服务，本身不包含产品检查。每个包发布一个 `./invariant` 所有权配套插件；`dsh-session`、`dsh-agent`、`dsh-scope` 和 `dsh-agent-loop` 目前添加需要跟踪状态或观察另一个 seam 的规则：单调递增的序列号、轮次与步骤嵌套、工具调用/结果配对、合法的 agent（智能体）状态转换、主体正确的作用域分发，以及循环构建的请求与从其会话日志前缀重建的请求之间的相等性。全局启用和包名 regex 过滤器归该服务所有（见[包拥有的不变式服务](2026-07-19-package-owned-invariant-service.zh.md)）。
+`dsh-invariants` 注册可配置的 `ctx.invariants` 服务，本身不包含产品检查。只有拥有可独立观察的运行时关系时，包才发布 `./invariant` 所有权配套插件；没有该关系的包会省略 companion 并在 README 中记录原因。`dsh-session`、`dsh-agent`、`dsh-scope` 和 `dsh-agent-loop` 提供首批需要跟踪状态或观察另一个 seam 的规则：单调递增的序列号、轮次与步骤嵌套、工具调用/结果配对、合法的 agent（智能体）状态转换、主体正确的作用域分发，以及循环构建的请求与从其会话日志前缀重建的请求之间的相等性。全局启用和包名 regex 过滤器归该服务所有（见[包拥有的不变式服务](2026-07-19-package-owned-invariant-service.zh.md)与[省略决策](../simplification/2026-08-28-omit-unneeded-invariant-companions.zh.md)）。

 当会话配套插件附加到已有会话或以种子记录初始化的会话时，它回放不可变日志以重建跟踪状态。服务为每项贡献提供一个可 dispose（资源释放）的子 fiber，因此轮次中途热重载是安全的，同时不赋予诊断逻辑对会话存储的所有权。

@@ -48,12 +48,12 @@ TypeScript readonly 类型不是充分的运行时边界。它们在程序运行

 ### 仅在派生消息时克隆

-分离 `deriveMessages()` 能保护最常见的请求路径，但 `session.events` 的其他读取者、append 返回值和会话事件观察者仍能修改持久历史。日志必须保护自身的边界；派生投影是额外的隔离边界，而非替代品。
+分离 `deriveMessages()` 能保护最常见的请求路径，但 `snapshotEvents()`、`eventAt()` 的其他读取者、append 返回值和会话事件观察者仍能修改持久历史。日志必须保护自身的边界；派生投影是额外的隔离边界，而非替代品。

 ## 后果

 - 每个被接受的实时或种子会话事件在任何观察者接收之前，都已从调用方拥有的输入中分离并深度不可变。
- `session.events` 暴露稳定的不可变快照，而非持续增长的私有数组。
+- `snapshotEvents()` 暴露稳定的不可变快照，而非持续增长的私有数组；`seq` 和 `eventAt()` 为标量读取提供无需复制数组的路径。
 - 请求侧的修改无法通过派生消息触及已存储的历史。
 - 开发构建可以启用关系断言而不改变存储行为；dispose 或过滤一个配套插件不会削弱日志不可变性。
 - `dsh-invariants` 配置全局启用状态以及包名允许/阻止 regex 列表；每项检查仍由其产品包拥有并测试。
@@ -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 .agents/notes/implemented/architecture/2026-07-19-package-invariant-runtime-contracts.md
-2026-07-19-package-invariant-runtime-contracts.md: b5799a37a61244193b46db6ea4ae15f306d144b2
-2026-07-19-package-invariant-runtime-contracts.zh.md: e6035bbabba7188017746c57c5b6a48761710658
+2026-07-19-package-invariant-runtime-contracts.md: a1b635cc40844f1846c04e203dbb842d1c7328ed
+2026-07-19-package-invariant-runtime-contracts.zh.md: d315c440f1c100911386f57d6f82d5f16be631ca
@@ -14,20 +14,20 @@ Some packages genuinely own no continuously observable relation. Pure utilities,

 ## Decision

-### Registration is exhaustive; assertions must be meaningful
+### Published assertions must be meaningful

-Every workspace package publishes a separately built `./invariant` companion and registers its exact npm package name. A companion does one of two things:
+A workspace package publishes a separately built `./invariant` companion only when it owns an independently observable runtime relationship. A published companion:

- installs a package-owned check over an event stream or relevant mutable data structure and reports violations through its bound `fail(message)` reporter; or
- uses an empty installer whose declaration has an owner-specific `No runtime invariant:` comment explaining why the package has no plausible runtime relation to observe.
+- installs a package-owned check over an event stream or relevant mutable data structure and reports violations through its bound `fail(message)` reporter; and
+- registers the package's exact npm name while keeping diagnostics outside the root entrypoint.

-The empty form is an explicit architectural conclusion, not a generated placeholder. A future package change that introduces mutable state or an event protocol must replace the explanation with the corresponding check.
+When no plausible relationship exists, the package omits the companion and publication wiring and records its package-specific reason in the README. A future change that introduces an independently observable relationship must replace the explanation with the corresponding check. The omission mechanics and current audit are owned by the [omit-unneeded-companions decision](../simplification/2026-08-28-omit-unneeded-invariant-companions.md).

 The central `dsh-invariants` service owns only configuration, registration uniqueness, child-fiber lifecycle, rollback, disposal, and package-attributed failure. It exposes no generic plugin-shape, service-shape, or startup-assertion helpers and imports no product package.

-### Implemented checks
+### Representative implemented checks

-The current 103-package workspace has 21 executable companions and 82 justified empty companions.
+Published companions are enumerated mechanically by `verify-package-invariants`; the current audit count is recorded in the [omit-unneeded-companions decision](../simplification/2026-08-28-omit-unneeded-invariant-companions.md). The table below samples representative runtime relationships rather than listing every companion.

 | Owner | Runtime relationship |
 |---|---|
@@ -57,13 +57,13 @@ Session-backed companions validate existing durable events when they load, using

 ### Repository gate and tests

-`verify-package-invariants` discovers every workspace package and enforces companion source, exact-name registration, named-only Loader shape, `./invariant` exports, publication files, dependencies, TypeScript references, and bundle entries. Its AST rule rejects generated markers, default exports, and unexplained empty installers. A non-empty installer must accept and use the failure reporter, and registration must pass that checked local `install` function. The gate deliberately does not infer semantic quality from method names or helper calls.
+`verify-package-invariants` discovers every workspace package. It accepts clean omission, rejects stale or partial companion wiring, and enforces exact-name registration, named-only Loader shape, `./invariant` exports, publication files, dependencies, TypeScript references, and bundle entries for published companions. Its AST rule rejects generated markers, default exports, and empty installers. Every installer must accept and use the failure reporter, and registration must pass that checked local `install` function. The gate deliberately does not infer semantic quality from method names or helper calls.

-Vitest mounts `InvariantRegistry` with `{ enabled: true }` for every package test topology and loads the owning companion. The invariant subpath path mapping resolves source companions instead of stale built output. Focused suites cover every executable companion's valid and invalid observations, and the exhaustive topology runs every source companion through the real Loader namespace normalization. After the structural gate validates each publication map, an artifact gate stages its manifest-declared `lib/` files, imports the compiled `./invariant` self-reference under plain Node, and repeats that Loader-shape check, so a companion that imports an undeclared runtime chunk fails before release. Tests that synthesize event streams must produce a valid surrounding lifecycle unless the test is intentionally asserting a violation.
+Vitest mounts `InvariantRegistry` with `{ enabled: true }` for every package test topology and loads the owning companion when one is published. The invariant subpath path mapping resolves source companions instead of stale built output. Focused suites cover every published companion's valid and invalid observations, and the exhaustive topology runs every source companion through real Loader namespace normalization. After the structural gate validates each publication map, an artifact gate stages its manifest-declared `lib/` files, imports the compiled `./invariant` self-reference under plain Node, and repeats that Loader-shape check, so a companion that imports an undeclared runtime chunk fails before release. Tests that synthesize event streams must produce a valid surrounding lifecycle unless the test is intentionally asserting a violation.

 ## Alternatives considered

- **Keep generated empty companions.** Rejected because an unexplained placeholder can survive after a package gains a meaningful runtime relation.
+- **Keep explained empty companions.** Rejected because source, publication, dependency, and test wiring are disproportionate machinery for a negative conclusion that belongs in the package README.
 - **Require an assertion from every package.** Rejected because method-presence, plugin-shape, and fixed-example assertions duplicate stronger type, load, and unit-test contracts without checking runtime consistency.
 - **Keep generic shape helpers in the service.** Rejected because they blur compile-time API validation with runtime invariants and encourage centrally defined product assumptions.
 - **Move the product checks into the service.** Rejected because product vocabulary, dependencies, tests, and change ownership belong with the package that emits the data.
@@ -71,8 +71,8 @@ Vitest mounts `InvariantRegistry` with `{ enabled: true }` for every package tes

 ## Consequences

- Every package has visible ownership and publication wiring, but only packages with a plausible runtime relation add listeners or trace state.
- Empty companions remain reviewable decisions with package-specific explanations and fail the gate if the explanation is removed.
+- Packages with a plausible runtime relation have visible ownership and publication wiring; packages without one record the omission reason in their README.
+- Empty companions fail the gate, and partial omission wiring fails before build or release.
 - Type declarations, Cordis loadability, plugin metadata, service method APIs, and pure algebra remain covered by their owning compile, load, unit, or integration gates.
 - Runtime failures identify the owning npm package and point to an inconsistent observation rather than restating a required API shape.
 - The original selection, blocklist precedence, duplicate ownership, rollback, disposal, and HMR service contracts remain unchanged.
@@ -14,20 +14,20 @@ Status: implemented

 ## 决策

-### 注册必须全覆盖；断言必须有意义
+### 已发布的断言必须有意义

-每个 workspace 包都发布单独构建的 `./invariant` companion，并用完整 npm 包名注册。companion 只能采用以下两种形式之一：
+只有拥有可独立观察的运行时关系时，workspace 包才发布单独构建的 `./invariant` companion。已发布 companion 必须：

- 安装包自有的事件流或相关可变数据结构检查，并通过绑定的 `fail(message)` 报告器报告违规；或
- 使用空安装器，并在其声明前写一条该包专属的 `No runtime invariant:` 注释，说明为什么该包没有合理的运行时关系可供观测。
+- 安装包自有的事件流或相关可变数据结构检查，并通过绑定的 `fail(message)` 报告器报告违规；并且
+- 用该包的准确 npm 包名注册，同时保持诊断逻辑不进入根入口。

-空形式是明确的架构结论，不是生成占位符。如果后续包变更引入可变状态或事件协议，就必须用相应检查替换该说明。
+没有合理关系时，包会省略 companion 与发布接线，并在 README 中记录该包的具体原因。如果后续变更引入可独立观察的关系，就必须用相应检查替换该说明。省略机制与当前审计由[省略不必要 companion 的决策](../simplification/2026-08-28-omit-unneeded-invariant-companions.zh.md)负责。

 中央 `dsh-invariants` 服务只负责配置、注册唯一性、子 fiber 生命周期、回滚、dispose（资源释放）和归属到包的失败。它不暴露通用插件形状、服务形状或启动断言 helper，也不导入产品包。

-### 已实施的检查
+### 已实施检查示例

-当前 103 个包的 workspace 包含 21 个可执行 companion 和 82 个有理由的空 companion。
+已发布 companion 由 `verify-package-invariants` 机械枚举；当前审计数量记录在[省略不必要 companion 的决策](../simplification/2026-08-28-omit-unneeded-invariant-companions.zh.md)中。下表仅展示有代表性的运行时关系，不会逐项列出所有 companion。

 | 所有者 | 运行时关系 |
 |---|---|
@@ -57,13 +57,13 @@ Status: implemented

 ### 仓库门禁与测试

-`verify-package-invariants` 发现每个 workspace 包，并强制 companion 源文件、完整名称注册、仅含具名 export 的 Loader 形状、`./invariant` export、发布文件、依赖、TypeScript reference 和 bundle entry 完整。其 AST 规则拒绝生成标记、默认导出和没有解释的空安装器。非空安装器必须接收并使用失败报告器，注册时还必须传入该经检查的本地 `install` 函数。门禁不会通过方法名或 helper 调用推断语义质量。
+`verify-package-invariants` 发现每个 workspace 包。它接受完整省略，拒绝陈旧或不完整的 companion 接线，并对已发布 companion 强制完整名称注册、仅含具名 export 的 Loader 形状、`./invariant` export、发布文件、依赖、TypeScript reference 和 bundle entry 完整。其 AST 规则拒绝生成标记、默认导出和空 installer。每个 installer 都必须接收并使用失败报告器，注册时还必须传入该经检查的本地 `install` 函数。门禁不会通过方法名或 helper 调用推断语义质量。

-Vitest 为每个包测试拓扑使用 `{ enabled: true }` 挂载 `InvariantRegistry`，并加载所有者 companion。不变量 subpath 的 path mapping 会解析源 companion，而不是陈旧的构建输出。聚焦 suite 覆盖每个可执行 companion 的有效和无效观测；穷举拓扑通过真实 Loader 命名空间归一化运行每个源 companion。结构门禁验证每个包的发布映射后，产物门禁会暂存其 manifest（元数据清单）声明的 `lib/` 文件，在 plain Node 下导入已编译的 `./invariant` 自引用，并重复执行该 Loader 形状检查；这样，若 companion 导入未声明的运行时分片，门禁就会在发布前失败。合成事件流的测试必须构造有效的外围生命周期，除非测试本身就是在断言违规。
+Vitest 为每个包测试拓扑使用 `{ enabled: true }` 挂载 `InvariantRegistry`，并在所有者发布 companion 时加载它。不变量 subpath 的 path mapping 会解析源 companion，而不是陈旧的构建输出。聚焦 suite 覆盖每个已发布 companion 的有效和无效观测；穷举拓扑通过真实 Loader 命名空间归一化运行每个源 companion。结构门禁验证每个包的发布映射后，产物门禁会暂存其 manifest（元数据清单）声明的 `lib/` 文件，在 plain Node 下导入已编译的 `./invariant` 自引用，并重复执行该 Loader 形状检查；这样，若 companion 导入未声明的运行时分片，门禁就会在发布前失败。合成事件流的测试必须构造有效的外围生命周期，除非测试本身就是在断言违规。

 ## 考虑过的替代方案

- **保留生成的空 companion。** 拒绝，因为包获得有意义的运行时关系后，没有解释的占位符仍可能继续存在。
+- **保留带说明的空 companion。** 拒绝，因为只为表达 README 可以直接记录的否定结论而保留源码、发布、依赖与测试接线，成本过高。
 - **要求每个包都执行断言。** 拒绝，因为方法存在性、插件形状和固定示例断言会重复更强的类型、加载和单元测试约定，却没有检查运行时一致性。
 - **在服务中保留通用形状 helper。** 拒绝，因为这会混淆编译期 API 验证和运行时不变量，并鼓励在中央定义产品假设。
 - **把产品检查移入服务。** 拒绝，因为产品词汇、依赖、测试和变更所有权应归属于产生这些数据的包。
@@ -71,8 +71,8 @@ Vitest 为每个包测试拓扑使用 `{ enabled: true }` 挂载 `InvariantRegis

 ## 后果

- 每个包都有可见的所有权与发布 wiring，但只有具备合理运行时关系的包才会增加 listener 或 trace 状态。
- 空 companion 是带包专属说明、可评审的决策；删除说明后门禁会失败。
+- 拥有合理运行时关系的包具有可见的所有权与发布 wiring；没有该关系的包会在 README 中记录省略原因。
+- 空 companion 会让门禁失败，不完整的省略接线也会在构建或发布前失败。
 - 类型声明、Cordis 可加载性、插件 metadata、服务方法 API 和纯代数继续由所属的编译、加载、单元或集成门禁覆盖。
 - 运行时失败会标明所属 npm 包，并指出不一致的观测，而不是复述必要的 API 形状。
 - 原有 selection、blocklist 优先级、重复所有权、回滚、dispose 和 HMR（热模块替换）服务约定保持不变。
@@ -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 .agents/notes/implemented/architecture/2026-07-19-package-owned-invariant-service.md
-2026-07-19-package-owned-invariant-service.md: 88fb870eac6f72307fa2bfeaa714fe59a6a9130d
-2026-07-19-package-owned-invariant-service.zh.md: 46b7789034c4a359fed95744e403f0a3a4dee6de
+2026-07-19-package-owned-invariant-service.md: b955c99a2576b6b2f8208a16ad2792af181c3468
+2026-07-19-package-owned-invariant-service.zh.md: 4fc0fb5d615753c0c057e927f59359847fc1328f
@@ -10,7 +10,7 @@ Runtime invariant checks span session traces, agent state, scoped dispatch, and

 Deployments that opt into diagnostics need more than presence or absence of one plugin. Such a composition carries the known invariant contributions while permitting a global off switch and package-selective diagnostics. Selection must remain stable when a package loads later or reloads under HMR, and disabled contributions must not allow two plugins to claim the same package name silently.

-Package ownership must also be exhaustive. Without a mechanical repository rule, a new package can omit the companion, dependency, or publication wiring and remain invisible to diagnostics until a maintainer notices the gap.
+Published ownership must be mechanically complete. Without a repository rule, a package can expose a partial companion, dependency, or publication map and remain broken until a maintainer notices the gap; packages that publish none must keep their reason reviewable in the README.

 ## Decision

@@ -18,7 +18,7 @@ Package ownership must also be exhaustive. Without a mechanical repository rule,

 `@deepseek-ai/dsh-invariants` is a product-independent Cordis service plugin that registers `ctx.invariants`. It owns configuration, registration uniqueness, child-fiber lifecycle, and package-attributed failures. It imports no session, agent, scope, or agent-loop package and contains none of their checks.

-Every workspace package publishes a `./invariant` companion plugin that registers its exact full npm name. A companion checks a meaningful event or mutable-data relationship when its owner has one; otherwise it carries an owner-specific explanation for its empty installer. Generated ownership placeholders and synthetic API-shape assertions are forbidden by the follow-up [runtime-contract Agent Note](2026-07-19-package-invariant-runtime-contracts.md). Package root entrypoints do not import or register diagnostics implicitly, so loading a root package does not change runtime checking or require the invariant service.
+A workspace package publishes a `./invariant` companion plugin only when it owns an independently observable event or mutable-data relationship. The companion registers its exact full npm name. Packages without such a relationship omit the companion and publication wiring and record the reason in their README; generated placeholders, empty installers, and synthetic API-shape assertions are forbidden by the [runtime-contract Agent Note](2026-07-19-package-invariant-runtime-contracts.md) and [omission decision](../simplification/2026-08-28-omit-unneeded-invariant-companions.md). Package root entrypoints do not import or register diagnostics implicitly, so loading a root package does not change runtime checking or require the invariant service.

 ### Configuration and selection

@@ -64,9 +64,9 @@ The former functional-plugin entry point and one-argument `InvariantError` const
 | `@deepseek-ai/dsh-scope/invariant` | `@deepseek-ai/dsh-scope` | scoped-event carrier presence and subject consistency |
 | `@deepseek-ai/dsh-agent-loop/invariant` | `@deepseek-ai/dsh-agent-loop` | model-request reconstruction |

-These four owners supplied the initial stateful checks. The follow-up runtime-contract decision adds checks for seventeen more owners with real event or mutable-data relationships and records justified empty companions for the rest. Every companion is a separately bundled `./invariant` export with its own declarations and Loader-safe namespace plugin shape; the service package's own companion imports its local service type to avoid a self-dependency.
+These four owners supplied the initial stateful checks. Later owners add companions for real event or mutable-data relationships, while packages without one omit the companion and document why. Every published companion is a separately bundled `./invariant` export with its own declarations and Loader-safe namespace plugin shape.

-`verify-package-invariants` discovers every workspace package and rejects missing companion source, generated markers, unexplained empty installers, non-empty installers that omit or ignore the reporter, foreign or unresolved registration names, missing `./invariant` exports or published files, missing invariant peer/development dependencies and project references, and bundle overrides that omit the companion entry.
+`verify-package-invariants` discovers every workspace package, accepts clean omission, and rejects partial companion wiring, generated markers, empty installers, installers that omit or ignore the reporter, foreign or unresolved registration names, missing `./invariant` exports or published files, missing invariant peer/development dependencies and project references, and bundle overrides that omit a published companion entry.

 ### Scoped-event semantic map

@@ -84,7 +84,7 @@ Service tests cover defaults, global disablement, allow/block selection, blockli

 Composition tests cover standard-spine forwarding and generated SDK entries. Loader tests preserve each companion namespace, while built plain-Node smokes exercise the compiled subpath exports. The scoped-event freshness gate reruns its semantic Program analysis.

-Every Vitest configuration loads a test host that mounts an explicitly enabled service before an ordinary Cordis root's first plugin and adds the current test package's companion. One exhaustive topology mounts all package companions once; focused service and owner tests construct their own invariant topology so they can exercise disablement, filtering, rollback, and reload without duplicate ownership. Gate tests also execute every companion's `apply` function and verify that it calls `register` with its manifest name, rather than accepting source text alone.
+Every Vitest configuration loads a test host that mounts an explicitly enabled service before an ordinary Cordis root's first plugin and adds the current test package's companion when one exists. One exhaustive topology mounts all published companions once; focused service and owner tests construct their own invariant topology so they can exercise disablement, filtering, rollback, and reload without duplicate ownership. Gate tests also execute every published companion's `apply` function and verify that it calls `register` with its manifest name, rather than accepting source text alone.

 ## Alternatives considered

@@ -96,10 +96,10 @@ Every Vitest configuration loads a test host that mounts an explicitly enabled s
 ## Consequences

 - Product packages own and test their relational assertions while the service stays product-independent.
- Every package pays the publication and dependency cost of a companion; only owners with a meaningful runtime relationship add listener or trace-state cost.
+- Only owners with a meaningful runtime relationship pay the publication, dependency, listener, or trace-state cost of a companion; other packages record the omission reason in their README.
 - Compositions that mount the diagnostics can disable all checks or select package names without changing their plugin tree.
 - Explicit companion entries make diagnostic cost and ownership visible in Cordis config and package exports.
- One selected executable contribution adds one child fiber and its listener/state cost; a selected empty contribution has no listener or trace-state cost, while filtered registrations retain only name ownership.
+- One selected contribution adds one child fiber and its listener/state cost, while filtered registrations retain only name ownership.
 - Regex sources are deployment configuration and remain fixed until the service reloads.
- Ordinary Vitest roots install the owning test package's selected companion; one exhaustive topology pays the full child-fiber cost once for repository-wide registration coverage.
+- Ordinary Vitest roots install the owning test package's selected companion when published; one exhaustive topology pays the full child-fiber cost once for repository-wide registration coverage.
 - Session storage validation, snapshotting, freezing, cited source-event validation, and surface acceptance remain always on and are not affected by invariant selection.
@@ -10,7 +10,7 @@ Status: implemented

 选择启用诊断的部署还需要比“是否加载一个插件”更细的控制。这类组合会携带已知的不变式贡献，同时允许全局关闭或按包选择诊断。包稍后加载或在 HMR（热模块替换）下重载时，选择结果必须保持稳定；被过滤的贡献也不能让两个插件静默占用同一个包名。

-包所有权还必须覆盖完整。若没有机械化的仓库规则，新包可能遗漏伴随插件、依赖或发布配置，并一直不会进入诊断范围，直到维护者发现这一缺口。
+已发布的包所有权必须机械完整。若没有仓库规则，包可能暴露不完整的 companion、依赖或发布映射，并一直保持损坏，直到维护者发现；不发布 companion 的包则必须在 README 中保留可评审的原因。

 ## 决策

@@ -18,7 +18,7 @@ Status: implemented

 `@deepseek-ai/dsh-invariants` 是与产品无关的 Cordis 服务插件，注册 `ctx.invariants`。它只负责配置、注册唯一性、子 fiber 生命周期和带包归属的失败；不导入 session、agent、scope 或 agent-loop 包，也不包含这些包的检查。

-工作区内的每个包都发布 `./invariant` 伴随插件，注册自己完整且准确的 npm 包名。如果所有者具备有意义的事件或可变数据关系，companion 就检查该关系；否则空 installer 必须携带该所有者专属的说明。后续的[运行时约定 Agent Note](2026-07-19-package-invariant-runtime-contracts.zh.md) 禁止生成的所有权占位符和合成 API 形状断言。包的根入口不会隐式导入或注册诊断，因此加载根包不会改变运行时检查，也不要求不变式服务存在。
+只有拥有可独立观察的事件或可变数据关系时，工作区包才发布 `./invariant` 伴随插件；该 companion 会注册自己完整且准确的 npm 包名。没有该关系的包会省略 companion 与发布接线，并在 README 中记录原因；[运行时约定 Agent Note](2026-07-19-package-invariant-runtime-contracts.zh.md) 与[省略决策](../simplification/2026-08-28-omit-unneeded-invariant-companions.zh.md)禁止生成占位符、空 installer 和合成 API 形状断言。包的根入口不会隐式导入或注册诊断，因此加载根包不会改变运行时检查，也不要求不变式服务存在。

 ### 配置与选择

@@ -64,9 +64,9 @@ blocklist 匹配优先于 allowlist 匹配。每个条目都是区分大小写
 | `@deepseek-ai/dsh-scope/invariant` | `@deepseek-ai/dsh-scope` | 作用域事件载体的存在性与主体一致性 |
 | `@deepseek-ai/dsh-agent-loop/invariant` | `@deepseek-ai/dsh-agent-loop` | 模型请求重建 |

-这四个所有者提供了首批有状态检查。后续运行时约定决策为另外十七个确有事件或可变数据关系的所有者增加检查，并为其余包记录有理由的空 companion。每个伴随入口都是单独打包的 `./invariant` export，具有独立声明和对 Loader 安全的命名空间插件形态；服务包自身的伴随插件导入本地服务类型，避免形成自依赖。
+这四个所有者提供了首批有状态检查。后续所有者会为真实事件或可变数据关系增加 companion，没有该关系的包则省略 companion 并记录原因。每个已发布伴随入口都是单独打包的 `./invariant` export，具有独立声明和对 Loader 安全的命名空间插件形态。

-`verify-package-invariants` 会发现每个工作区包，并拒绝缺失的伴随插件源码、生成标记、没有解释的空 installer、缺少或不使用失败报告器的非空 installer、外部或无法解析的注册名、缺失的 `./invariant` export 或发布文件、缺失的不变式对等依赖（peer dependency）、开发依赖及项目引用，以及遗漏伴随入口的自定义构建配置。
+`verify-package-invariants` 会发现每个工作区包，接受完整省略，并拒绝不完整的 companion 接线、生成标记、空 installer、缺少或不使用失败报告器的 installer、外部或无法解析的注册名、缺失的 `./invariant` export 或发布文件、缺失的不变式对等依赖（peer dependency）、开发依赖及项目引用，以及遗漏已发布伴随入口的自定义构建配置。

 ### 作用域事件语义映射

@@ -84,7 +84,7 @@ Workspace 约束识别独立的不变式 bundle；包 exports、项目引用、

 组合测试覆盖标准主干转发和生成的 SDK 条目。Loader 测试固定每个伴随命名空间，构建后的纯 Node 冒烟测试覆盖编译子路径 export。作用域事件新鲜度门禁会重新执行语义 Program 分析。

-每个 Vitest 配置都会加载测试宿主；在普通 Cordis 根上下文启动第一个插件之前，宿主会挂载显式启用的服务，并添加当前测试包的伴随插件。一个完整拓扑会一次挂载所有包的伴随插件；服务与所有者的聚焦测试自行构建不变式拓扑，从而在不发生重复所有权冲突的前提下覆盖关闭、过滤、回滚与重载。门禁测试还会执行每个伴随插件的 `apply` 函数，并验证它调用 `register` 时使用 manifest（元数据清单）中的包名，而不是只检查源码文本。
+每个 Vitest 配置都会加载测试宿主；在普通 Cordis 根上下文启动第一个插件之前，宿主会挂载显式启用的服务，并在当前测试包存在伴随插件时添加它。一个完整拓扑会一次挂载所有已发布伴随插件；服务与所有者的聚焦测试自行构建不变式拓扑，从而在不发生重复所有权冲突的前提下覆盖关闭、过滤、回滚与重载。门禁测试还会执行每个已发布伴随插件的 `apply` 函数，并验证它调用 `register` 时使用 manifest（元数据清单）中的包名，而不是只检查源码文本。

 ## 考虑过的替代方案

@@ -96,10 +96,10 @@ Workspace 约束识别独立的不变式 bundle；包 exports、项目引用、
 ## 后果

 - 产品包拥有并测试自己的关系断言，服务保持与产品无关。
- 每个包都承担 companion 的发布与依赖成本；只有具备有意义运行时关系的所有者才增加 listener 或 trace 状态成本。
+- 只有具备有意义运行时关系的所有者才承担 companion 的发布、依赖、listener 或 trace 状态成本；其他包在 README 中记录省略原因。
 - 挂载诊断的组合无需改变插件树即可关闭全部检查或按包名选择。
 - 显式伴随条目让诊断成本和所有权在 Cordis 配置与包 export 中可见。
- 每个选中的可执行贡献增加一个子 fiber 及其 listener/状态成本；选中的空贡献不增加 listener 或 trace 状态成本，被过滤注册则只保留包名占用。
+- 每个选中贡献增加一个子 fiber 及其 listener/状态成本，被过滤注册则只保留包名占用。
 - 正则表达式源属于部署配置，在服务重载前保持固定。
- 普通 Vitest 根上下文会安装当前测试包中被选中的伴随插件；一个完整拓扑只支付一次全部子 fiber 成本，用于覆盖整个仓库的注册。
+- 当前测试包发布伴随插件时，普通 Vitest 根上下文会安装其中被选中的伴随插件；一个完整拓扑只支付一次全部子 fiber 成本，用于覆盖整个仓库的注册。
 - 会话存储验证、快照、冻结、引用的源事件验证与 surface 接受规则始终启用，不受不变式选择影响。
@@ -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 .agents/notes/implemented/architecture/2026-07-25-web-client-session-scope-and-provide-channel.md
-2026-07-25-web-client-session-scope-and-provide-channel.md: b4566d70c79607bbf736ee02e3e37a79c2391232
-2026-07-25-web-client-session-scope-and-provide-channel.zh.md: 1d1acf00fa6a1efc868c3613715a5ff781e0323a
+2026-07-25-web-client-session-scope-and-provide-channel.md: feffb1ac0c5bb91c33e61ea18583202e96bad34d
+2026-07-25-web-client-session-scope-and-provide-channel.zh.md: 5d368b4fd810264cb1b451d583de51e1d4cb232b
@@ -61,7 +61,7 @@ Session instances share the scope's lifecycle; liveness eligibility = host-liste

 A session "materialized but with no first prompt" is governed by the summary-derived bit `blank` (a derived column, not a header field; SessionHeader stays immutable):

- The host criterion: `session.events.length === 0` (zero log events = no user message yet). A live session reads `summarize()` straight from memory; a cold session is always `false` — the JSONL provider's lazy-create contract guarantees a never-appended Session never enters `persistence.list()`, so blank never touches disk.
+- The host criterion: `session.seq === 0` (zero log events = no user message yet). A live session reads `summarize()` straight from memory; a cold session is always `false` — the JSONL provider's lazy-create contract guarantees a never-appended session never enters `persistence.list()`, so blank never touches disk.
 - The wire carries it in two places: the required `SessionSummary.blank` column, and the required `blank` field on the `host/session-added` frame (always true at creation, letting other tabs enter the same blank-session state into their mirrors).
 - The client mirror only lowers, never raises (monotonic), flipped from three sources, all reusing existing wire signals:
  - The sender's own tab: the **successful response** to the first `prompt()` flips false (acceptance proves the user/message is already in the host log — this flip is confirmation, not optimism; `onEngaged` synchronously updates the list mirror, converting the current `New Session` row in place to an ordinary title, adding no list row). A rejected first prompt keeps the session blank: aligned with host authority, still shown as `New Session`, keeping its connectWorkspace reuse eligibility while it remains a Workspace member.
@@ -61,7 +61,7 @@ Session 实例与 scope 同生命周期，存活资格 = host listed（一个判

 「实体化但无首条提示词」的会话经 summary 派生位 `blank` 治理（派生列而非 header 字段，SessionHeader 保持不可变）：

- host 判据：`session.events.length === 0`（零日志事件 = 尚无用户消息）。live 会话 `summarize()` 内存直读；cold 会话恒 `false`——JSONL provider 的 lazy-create 约定保证 never-appended Session 不进入 `persistence.list()`，所以 blank 从不落盘。
+- host 判据：`session.seq === 0`（零日志事件 = 尚无用户消息）。live 会话 `summarize()` 内存直读；cold 会话恒 `false`——JSONL provider 的 lazy-create 约定保证 never-appended 会话不进入 `persistence.list()`，所以 blank 从不落盘。
 - wire 承载两处：`SessionSummary.blank` 必填列；`host/session-added` 帧必填 `blank` 字段（创建时恒 true，供别的 tab 按同一空会话状态入镜像）。
 - client 镜像只降不升（单调），三来源翻转，全部复用既有 wire 信号：
  - 发送方本地：首次 `prompt()` 的**成功响应**翻 false（受理即证明用户消息已入 host 日志——此点翻转是确证而非乐观；`onEngaged` 同步更新列表镜像，当前 `New Session` 行原地转为普通标题，不新增列表行）。首条提示词被拒则会话保持 blank：与 host 权威对齐、继续显示为 `New Session`、在仍为该工作区成员时保持 connectWorkspace 复用资格。
@@ -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 .agents/notes/implemented/architecture/2026-07-26-packed-chunk-rows-by-default.md
-2026-07-26-packed-chunk-rows-by-default.md: 14da6b3cbe650e80118e7c960c96bf618acd1e48
-2026-07-26-packed-chunk-rows-by-default.zh.md: f62a8e52a67adc960ac3150552594b4f061e6205
+2026-07-26-packed-chunk-rows-by-default.md: bd4b3b9f773afbf6aa7e88d51b6e842d6634c222
+2026-07-26-packed-chunk-rows-by-default.zh.md: eafe6632150aadd74395f4d0f09d064fb703a03d
@@ -18,7 +18,7 @@ Reading is unconditional and layout-blind. Packed, unpacked, and mixed files loa

 ### Logical events and physical rows

-The JSONL packing path stays at the `dsh-session` storage seam through `packChunkRuns()` and `decodeStorageRecord()`. The encoder recognizes exact delta-event shapes, preserves unrecognized events verbatim, and packs only runs of at least three. A packed row is encoding vocabulary, not a `SessionEventMap` member: it never enters `Session.events` or fires `session/event`. The [packed session-history transport decision](2026-08-15-packed-session-history-transport.md) reuses this vocabulary for a bounded lossless wire interval without changing those event semantics.
+The JSONL packing path stays at the `dsh-session` storage seam through `packChunkRuns()` and `decodeStorageRecord()`. The encoder recognizes exact delta-event shapes, preserves unrecognized events verbatim, and packs only runs of at least three. A packed row is encoding vocabulary, not a `SessionEventMap` member: it never enters the Session log or fires `session/event`. The [packed session-history transport decision](2026-08-15-packed-session-history-transport.md) reuses this vocabulary for a bounded lossless wire interval without changing those event semantics.

 The JSONL backend packs each durable append batch. Raw `compression: 'none'` and default Zstandard framing carry the same logical storage records; selecting raw mode for reviewable fixtures does not disable packing. Repository replay readers and normalizers decode the shared row format instead of maintaining snapshot-specific codecs.

@@ -18,7 +18,7 @@ JSONL 存储 seam 可以在不改变逻辑日志的情况下减少这部分封

 ### 逻辑事件与物理行

-JSONL 打包路径保留在 `dsh-session` 的存储 seam，并通过 `packChunkRuns()` 和 `decodeStorageRecord()` 实现。编码器识别精确的增量事件形态，原样保留无法识别的事件，并且只打包至少包含 3 个事件的连续段。打包行属于编码词汇，不是 `SessionEventMap` 成员：它绝不会进入 `Session.events`，也不会触发 `session/event`。[打包会话历史传输决策](2026-08-15-packed-session-history-transport.zh.md)会为有界的无损协议区间复用该词汇，而不改变这些事件语义。
+JSONL 打包路径保留在 `dsh-session` 的存储 seam，并通过 `packChunkRuns()` 和 `decodeStorageRecord()` 实现。编码器识别精确的增量事件形态，原样保留无法识别的事件，并且只打包至少包含 3 个事件的连续段。打包行属于编码词汇，不是 `SessionEventMap` 成员：它绝不会进入 Session 日志，也不会触发 `session/event`。[打包会话历史传输决策](2026-08-15-packed-session-history-transport.zh.md)会为有界的无损协议区间复用该词汇，而不改变这些事件语义。

 JSONL 后端会打包每个持久追加批次。原始模式 `compression: 'none'` 与默认 Zstandard 帧承载相同的逻辑存储记录；为使 fixture 便于评审而选择原始模式，不会禁用打包。仓库中的回放读取器和规范化器会解码共享行格式，而不维护快照专用编解码器。

@@ -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 .agents/notes/implemented/architecture/2026-07-31-code-runtime-portable-identifier-seam.md
-2026-07-31-code-runtime-portable-identifier-seam.md: 2011b0f6bc8209e628227ddf486aa1143a63688a
-2026-07-31-code-runtime-portable-identifier-seam.zh.md: 36af33366d004fedc6b1077a937d6519de743638
+2026-07-31-code-runtime-portable-identifier-seam.md: e4cf236f62407c9fda42a3e2cdcc5d3ef02a1f92
+2026-07-31-code-runtime-portable-identifier-seam.zh.md: 63fc89eb0d674a381ce7a5a626bd51d5f8b234d3
@@ -25,7 +25,7 @@ The constants live in the Service Definition even though the worker is the only

 ## Scope

-This decision delivers only the Service Definition extension and the worker's adoption of it. The `py-types` renderer and PTC mode language dispatch are owned by the [language-dispatch note](../feature/2026-07-31-ptc-language-dispatch.md); a Python backend does not exist yet. The Service Definition README keeps its worker-only wording for that reason: linking to a `dsh-code-runtime-python` README that does not exist would break the dead-link gate.
+This decision delivers the Service Definition extension and the worker-thread backend's adoption of it. The `py-types` renderer and PTC mode language dispatch are owned by the [language-dispatch note](../feature/2026-07-31-ptc-language-dispatch.md). The private experimental CPython subprocess backend (`dsh-experimental-code-runtime-python`) adopts the same portable-identifier contract.

 `RESERVED_BINDING_GLOBALS` encodes the Python bootstrap's concrete design ahead of the backend itself: it seeds exactly `__builtins__`/`__name__` and wraps the program under `__dsh_main__`. A Python backend that seeds any additional module global (`__doc__`, `__loader__`, `__spec__`, `__file__`, `__package__`, …) MUST widen this set in the same change, exactly as adding a language widens `PORTABLE_RESERVED_WORDS` — a name the bootstrap seeds but the set omits is the portability split this contract exists to prevent.

@@ -25,7 +25,7 @@ Service Definition 同时把可移植标识符子集收窄为 `[A-Za-z_][A-Za-z0

 ## Scope

-本决策只交付 Service Definition 扩展与 worker 对它的采用。`py-types` 渲染器与 PTC mode 的语言分发归[语言分发 note](../feature/2026-07-31-ptc-language-dispatch.zh.md) 所有；Python 后端尚不存在。Service Definition README 因此保留仅描述 worker 的措辞：链接到一个不存在的 `dsh-code-runtime-python` README 会破坏死链 gate。
+本决策交付 Service Definition 扩展与 worker-thread 后端对它的采用。`py-types` 渲染器与 PTC mode 的语言分发归[语言分发 note](../feature/2026-07-31-ptc-language-dispatch.zh.md)所有。私有的实验性 CPython 子进程后端（`dsh-experimental-code-runtime-python`）采用同一 portable-identifier 契约。

 `RESERVED_BINDING_GLOBALS` 先于后端本身编码了 Python bootstrap 的具体设计：它恰好 seed `__builtins__`/`__name__`，并把程序包装在 `__dsh_main__` 之下。任何 seed 额外模块 global（`__doc__`、`__loader__`、`__spec__`、`__file__`、`__package__` 等）的 Python 后端必须在同一改动中扩宽此集合，正如新增一门语言即扩宽 `PORTABLE_RESERVED_WORDS`——bootstrap 会 seed 却不在集合中的名称，正是本约定要防止的可移植性分裂。

@@ -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 .agents/notes/implemented/architecture/2026-07-31-code-runtime-python-fd3-protocol.md
-2026-07-31-code-runtime-python-fd3-protocol.md: 5572fe58cb1dd8832ff9405670afc7f80a20362c
-2026-07-31-code-runtime-python-fd3-protocol.zh.md: 6254e94a7b48b38edfbe23a6ea0b994d04ac21f4
+2026-07-31-code-runtime-python-fd3-protocol.md: cd8a42b509598d4782fc7c0637839e0dfd06f289
+2026-07-31-code-runtime-python-fd3-protocol.zh.md: a6454c17dc23e3f6385fe2dc3b46eabdb241faff
@@ -2,13 +2,15 @@

 Status: implemented

+The CPython code runtime now lives at `packages/experimental/code-runtime-python` (private, npm name `@deepseek-ai/dsh-experimental-code-runtime-python`); promotion to a released package follows the experimental-packages decision.
+
 English | [中文](2026-07-31-code-runtime-python-fd3-protocol.zh.md)

 ## Problem

-`@deepseek-ai/dsh-code-runtime-python` owns the wire protocol intended for a CPython code-runtime provider. Such a provider runs each model program in a fresh `python3 -I` subprocess and bridges binding calls and completion values over the child's fd 3. The host cannot trust that channel: model code has full access to fd 3 and can forge any frame, so every inbound frame is hostile input that the host must validate and rebuild before reading. The protocol also has to carry lossless JSON without the depth limit `JSON.stringify` and `json.dumps` impose, because the seam's `CodeJsonValue` is depth-unbounded.
+`@deepseek-ai/dsh-experimental-code-runtime-python` owns the wire protocol intended for a CPython code-runtime provider. Such a provider runs each model program in a fresh `python3 -I` subprocess and bridges binding calls and completion values over the child's fd 3. The host cannot trust that channel: model code has full access to fd 3 and can forge any frame, so every inbound frame is hostile input that the host must validate and rebuild before reading. The protocol also has to carry lossless JSON without the depth limit `JSON.stringify` and `json.dumps` impose, because the seam's `CodeJsonValue` is depth-unbounded.

-The package ships the protocol independently from a runtime implementation. It exports no `PythonCodeRuntime`, subprocess path, or Python-side JSON codec; those remain work for a future provider. The protocol builds on the [portable identifier seam](2026-07-31-code-runtime-portable-identifier-seam.md).
+The private experimental package contains both the protocol and runtime implementation: `PythonCodeRuntime` (the plugin's default export), the `python3 -I` subprocess path, and the Python-side JSON codec all live in `@deepseek-ai/dsh-experimental-code-runtime-python`. The protocol builds on the [portable identifier seam](2026-07-31-code-runtime-portable-identifier-seam.md).

 ## Decision

@@ -20,24 +22,24 @@ The package ships the protocol independently from a runtime implementation. It e

 `py/protocol.py` mirrors the message shapes as `TypedDict`s and re-declares the two surfaces both sides EXECUTE against — `PROTOCOL_FD = 3` and `log_truncation_marker` — with byte-identical text.

-The package remains independently buildable with protocol-only exports. `check-workspace-constraints` reads every `packages/<group>/<pkg>/package.json` unconditionally, while the coverage and invariant-topology checks exercise the package as soon as its directory exists.
+The package ships the runtime alongside the protocol; it remains independently buildable. `check-workspace-constraints` reads every `packages/<group>/<pkg>/package.json` unconditionally, while the coverage and invariant-topology checks exercise the package as soon as its directory exists.

 ## Wire contract

-Frames are JSON-lines on fd 3, one object per line, leaving stdout/stderr free for the program's own output. Child → host: `boot-ack`, `call`, `log`, `done`. Host → child: `boot` (first frame), `run` (after `boot-ack`), and one `reply` per `call`. The `log` frame's `truncated` flag marks the frame that IS the child ledger's own truncation marker, so the host stops capturing at the same point the child did instead of inferring it from its own budget. `done.error.kind` is one of `exception`, `invalid-output`, `output-limit`; wall/CPU budgets, aborts, and substrate death are observed host-side, not carried as frames.
+Frames are JSON-lines on fd 3, one object per line, leaving stdout/stderr free for the program's own output. Child → host: `boot-ack`, `call`, `log`, `done`. Host → child: `boot` (first frame), `run` (after `boot-ack`), and one `reply` per `call`. The `log` frame's `truncated` flag marks the frame that IS the child ledger's own truncation marker, so the host stops capturing at the same point the child did instead of inferring it from its own budget. The `log` frame's `open` flag marks an unterminated line committed by an explicit flush: the host holds it and appends the next frame to the same entry, so an explicit flush followed by more text reads back as one line rather than a fake newline. The one exception is truncation: when a later over-budget frame trips the ledger, the already-billed prefix is committed as its own entry and the truncation marker follows it (marker last, no re-charge). The merged entry's wire cost is billed exactly once, split incrementally across its fragments on both sides (O(k) for k fragments, never a re-walk of the whole hold): the FIRST fragment pays the full JSON-string cost plus the separator, each continuation and the closing frame pay only their content; the host's exact-cost caps are `logBudget - 1` for a first fragment (the ledger's reserved byte, matching `admit`) and `logBudget + 2` for a continuation or closing frame (billed without the two quotes), and `jsonStringCostUpTo` returns `undefined` below a 2-byte cap; the child keys its split billing off `_open_started` alone, so a closing frame bills as the merged tail. `done.error.kind` is one of `exception`, `invalid-output`, `output-limit`; wall/CPU budgets, aborts, and substrate death are observed host-side, not carried as frames.

 ## Mirror alignment

-`py/protocol.py` and `src/protocol.ts` agree that `LogMessage` carries `truncated`, `DoneMessage.error` carries `kind`, and `Namespace` may carry `errorClass`. `tests/protocol-mirror.e2e.ts` spawns a real `python3` and asserts `PROTOCOL_FD`, `log_truncation_marker`, and each `TypedDict`'s required and optional wire field sets against `src/protocol.ts`. A renamed or dropped field, or a required/optional mismatch, fails the test. Field *types* are not compared across the language boundary; review and a future provider's real-subprocess suite own that gap.
+`py/protocol.py` and `src/protocol.ts` agree that `LogMessage` carries `truncated`, `DoneMessage.error` carries `kind`, and `Namespace` may carry `errorClass`. `tests/protocol-mirror.e2e.ts` spawns a real `python3` and asserts `PROTOCOL_FD`, `log_truncation_marker`, and each `TypedDict`'s required and optional wire field sets against `src/protocol.ts`. A renamed or dropped field, or a required/optional mismatch, fails the test. Field *types* are not compared across the language boundary; review and the runtime's real-subprocess suite (`runtime.spec.ts`) own that gap.

 ## Alternatives considered

-**Require a future Python JSON codec (`_encode_json_plain` / `_decode_json_plain`) to live in `py/protocol.py` for cross-side symmetry with `protocol.ts`.** Rejected. The repository's "prefer symmetry for parallel values" rule points at genuinely parallel values; these are not. The host-side codec in `protocol.ts` validates hostile input and is self-contained. A child-side codec would produce trusted output and belong with bootstrap-owned emission and cost accounting; forcing only its entry points into `protocol.py` would couple the vocabulary mirror to runtime internals or create an import cycle. `protocol.py` remains a pure wire-vocabulary mirror. No Python codec ships in this package.
+**Require a future Python JSON codec (`_encode_json_plain` / `_decode_json_plain`) to live in `py/protocol.py` for cross-side symmetry with `protocol.ts`.** Rejected. The repository's "prefer symmetry for parallel values" rule points at genuinely parallel values; these are not. The host-side codec in `protocol.ts` validates hostile input and is self-contained. A child-side codec would produce trusted output and belong with bootstrap-owned emission and cost accounting; forcing only its entry points into `protocol.py` would couple the vocabulary mirror to runtime internals or create an import cycle. `protocol.py` remains a pure wire-vocabulary mirror; the codec (`_encode_json_plain` / `_decode_json_plain`) lives in `bootstrap.py` with the runtime it serves.

 **Keep the protocol files outside a buildable package until a runtime ships.** Rejected: the workspace-constraint, coverage, and invariant-topology checks require every directory under `packages/<group>/<pkg>` to be a buildable package, and the protocol has independent tests and a public wire vocabulary.

 ## Consequences

-Bought: the fd-3 protocol and its hostile-input codec form a self-contained, fully unit-covered layer, with an executing guard against TypeScript/Python field-set drift. A future runtime can consume a reviewed wire contract.
+Bought: the fd-3 protocol and its hostile-input codec form a self-contained, fully unit-covered layer, with an executing guard against TypeScript/Python field-set drift. The runtime built on it (`bootstrap.py`) consumes the reviewed wire contract.

-Cost: the package name denotes a Python runtime family while `src/index.ts` exports only the protocol vocabulary. The mirror e2e compares field names and required/optional status across the two sides but not field types; comparing type declarations across TypeScript and Python has no mechanical equivalent, so review and the future runtime's real-subprocess suite retain that responsibility.
+Cost: the package name denotes a Python runtime family and `src/index.ts` exports the full `PythonCodeRuntime` implementation, so the protocol vocabulary is only one part of the package surface. The mirror e2e compares field names and required/optional status across the two sides but not field types; comparing type declarations across TypeScript and Python has no mechanical equivalent, so review and the runtime's real-subprocess suite retain that responsibility.
@@ -2,13 +2,15 @@

 Status: implemented

+CPython 代码运行时现在位于 `packages/experimental/code-runtime-python`（私有，npm 名 `@deepseek-ai/dsh-experimental-code-runtime-python`）；提升为发布包遵循 experimental-packages 决策。
+
 [English](2026-07-31-code-runtime-python-fd3-protocol.md) | 中文

 ## Problem

-`@deepseek-ai/dsh-code-runtime-python` 负责供 CPython code-runtime 提供方使用的 wire protocol。这样的提供方会在全新的 `python3 -I` 子进程中运行每个模型程序，并通过子进程 fd 3 桥接 binding 调用与完成值。Host 不能信任这条通道：模型代码可以完全访问 fd 3 并伪造任意帧，因此 host 必须把每个入站帧视为敌意输入，先校验并重建后才能读取。协议还必须承载无深度限制的 lossless JSON，因为 seam 的 `CodeJsonValue` 深度无界，而 `JSON.stringify` 和 `json.dumps` 都有递归深度限制。
+`@deepseek-ai/dsh-experimental-code-runtime-python` 负责供 CPython code-runtime 提供方使用的 wire protocol。这样的提供方会在全新的 `python3 -I` 子进程中运行每个模型程序，并通过子进程 fd 3 桥接 binding 调用与完成值。Host 不能信任这条通道：模型代码可以完全访问 fd 3 并伪造任意帧，因此 host 必须把每个入站帧视为敌意输入，先校验并重建后才能读取。协议还必须承载无深度限制的 lossless JSON，因为 seam 的 `CodeJsonValue` 深度无界，而 `JSON.stringify` 和 `json.dumps` 都有递归深度限制。

-该包独立交付协议，不包含 runtime 实现。它不导出 `PythonCodeRuntime`、子进程路径或 Python 侧 JSON codec；这些属于未来提供方。协议建立在[可移植标识符 seam](2026-07-31-code-runtime-portable-identifier-seam.zh.md)之上。
+这个私有实验包同时包含协议与 runtime 实现：`PythonCodeRuntime`（插件的默认导出）、`python3 -I` 子进程路径与 Python 侧 JSON codec 都在 `@deepseek-ai/dsh-experimental-code-runtime-python` 中。协议建立在[可移植标识符 seam](2026-07-31-code-runtime-portable-identifier-seam.zh.md)之上。

 ## Decision

@@ -20,24 +22,24 @@ Status: implemented

 `py/protocol.py` 用 `TypedDict` 镜像消息形状，并重新声明两侧都会 EXECUTE 的两个面——`PROTOCOL_FD = 3` 与 `log_truncation_marker`——文本逐字节一致。

-该包只导出协议，同时保持独立可构建。`check-workspace-constraints` 会无条件读取每个 `packages/<group>/<pkg>/package.json`，coverage 与 invariant-topology 检查则会在包目录存在时立即覆盖该包。
+该包随协议一起交付 runtime，同时保持独立可构建。`check-workspace-constraints` 会无条件读取每个 `packages/<group>/<pkg>/package.json`，coverage 与 invariant-topology 检查则会在包目录存在时立即覆盖该包。

 ## Wire contract

-帧是 fd 3 上的 JSON-lines，每行一个对象，让 stdout/stderr 空出给程序自己的输出。Child → host：`boot-ack`、`call`、`log`、`done`。Host → child：`boot`（首帧）、`run`（在 `boot-ack` 之后）、以及每个 `call` 对应一个 `reply`。`log` 帧的 `truncated` 标志标记那个本身就是子进程 ledger 截断标记的帧，使 host 在与子进程相同的点停止捕获，而不是从自己的预算去推断。`done.error.kind` 是 `exception`、`invalid-output`、`output-limit` 之一；wall/CPU 预算、abort、substrate 死亡都在 host 侧观测，不作为帧携带。
+帧是 fd 3 上的 JSON-lines，每行一个对象，让 stdout/stderr 空出给程序自己的输出。Child → host：`boot-ack`、`call`、`log`、`done`。Host → child：`boot`（首帧）、`run`（在 `boot-ack` 之后）、以及每个 `call` 对应一个 `reply`。`log` 帧的 `truncated` 标志标记那个本身就是子进程 ledger 截断标记的帧，使 host 在与子进程相同的点停止捕获，而不是从自己的预算去推断。`log` 帧的 `open` 标志标记由显式 flush 提交的未结束行：宿主持有它并把下一个帧追加到同一条目，因此显式 flush 后接更多文本读回为一行而不是假换行。唯一例外是截断：当后续超预算帧触发账本时，已计费的前缀作为独立条目先提交，截断 marker 跟在后面（marker 保持末位，无重复计费）。合并条目的线上成本恰好计费一次，在两侧按片段增量分摊（k 个片段 O(k)，绝不对整个持有重走）：首片段付完整 JSON 字符串成本加分隔符，每个续接与闭合帧只付内容；宿主精确成本 cap 是首片段 `logBudget - 1`（账本预留字节，与 `admit` 一致）、续接或闭合帧 `logBudget + 2`（不含两个引号计费），且 `jsonStringCostUpTo` 在低于 2 字节 cap 时返回 `undefined`；子进程按 `_open_started` 单独键控拆分计费，因此闭合帧按合并尾部计费。`done.error.kind` 是 `exception`、`invalid-output`、`output-limit` 之一；wall/CPU 预算、abort、substrate 死亡都在 host 侧观测，不作为帧携带。

 ## Mirror alignment

-`py/protocol.py` 与 `src/protocol.ts` 一致规定：`LogMessage` 携带 `truncated`，`DoneMessage.error` 携带 `kind`，`Namespace` 可以携带 `errorClass`。`tests/protocol-mirror.e2e.ts` 启动真实 `python3`，对照 `src/protocol.ts` 断言 `PROTOCOL_FD`、`log_truncation_marker` 以及每个 `TypedDict` 的必填和可选 wire 字段集。字段改名、删除或必填／可选性不一致都会使测试失败。字段*类型*不跨语言边界比较；这项缺口由评审和未来提供方的真实子进程套件负责。
+`py/protocol.py` 与 `src/protocol.ts` 一致规定：`LogMessage` 携带 `truncated`，`DoneMessage.error` 携带 `kind`，`Namespace` 可以携带 `errorClass`。`tests/protocol-mirror.e2e.ts` 启动真实 `python3`，对照 `src/protocol.ts` 断言 `PROTOCOL_FD`、`log_truncation_marker` 以及每个 `TypedDict` 的必填和可选 wire 字段集。字段改名、删除或必填／可选性不一致都会使测试失败。字段*类型*不跨语言边界比较；这项缺口由评审和 runtime 的真实子进程套件（`runtime.spec.ts`）负责。

 ## Alternatives considered

-**要求未来的 Python JSON codec（`_encode_json_plain` / `_decode_json_plain`）放进 `py/protocol.py`，以便与 `protocol.ts` 跨侧对称。**拒绝。仓库的 “prefer symmetry for parallel values” 规则指向真正平行的值；这两者不是。`protocol.ts` 中的 host 侧 codec 校验敌意输入且自包含。Child 侧 codec 会产出受信任输出，应与 bootstrap 拥有的发出逻辑和成本核算放在一起；只把入口强塞进 `protocol.py` 会让 vocabulary 镜像耦合 runtime 内部实现，或制造 import 环。`protocol.py` 保持纯 wire-vocabulary 镜像。本包尚未交付 Python codec。
+**要求未来的 Python JSON codec（`_encode_json_plain` / `_decode_json_plain`）放进 `py/protocol.py`，以便与 `protocol.ts` 跨侧对称。**拒绝。仓库的 “prefer symmetry for parallel values” 规则指向真正平行的值；这两者不是。`protocol.ts` 中的 host 侧 codec 校验敌意输入且自包含。Child 侧 codec 会产出受信任输出，应与 bootstrap 拥有的发出逻辑和成本核算放在一起；只把入口强塞进 `protocol.py` 会让 vocabulary 镜像耦合 runtime 内部实现，或制造 import 环。`protocol.py` 保持纯 wire-vocabulary 镜像；codec（`_encode_json_plain`／`_decode_json_plain`）与它所服务的 runtime 一起位于 `bootstrap.py`。

 **在 runtime 交付前把协议文件放在不可构建的包外。**拒绝：workspace-constraint、coverage 与 invariant-topology 检查要求 `packages/<group>/<pkg>` 下的每个目录都是可构建包，而协议本身拥有独立测试与公开 wire vocabulary。

 ## Consequences

-收获：fd-3 协议及其敌意输入 codec 构成自包含、unit 全覆盖的一层，并由执行中的 guard 防止 TypeScript／Python 字段集漂移。未来 runtime 可以直接消费经过评审的 wire contract。
+收获：fd-3 协议及其敌意输入 codec 构成自包含、unit 全覆盖的一层，并由执行中的 guard 防止 TypeScript／Python 字段集漂移。基于它构建的 runtime（`bootstrap.py`）消费经过评审的 wire contract。

-代价：包名表示 Python runtime 家族，而 `src/index.ts` 只导出协议 vocabulary。mirror e2e 会比较两侧字段名与必填／可选状态，但不比较字段类型；跨 TypeScript 与 Python 比较类型声明没有机械等价物，因此评审与未来 runtime 的真实子进程套件继续负责这项检查。
+代价：包名表示 Python runtime 家族，而 `src/index.ts` 导出完整的 `PythonCodeRuntime` 实现，协议 vocabulary 只是包表面的一部分。mirror e2e 会比较两侧字段名与必填／可选状态，但不比较字段类型；跨 TypeScript 与 Python 比较类型声明没有机械等价物，因此评审与 runtime 的真实子进程套件继续负责这项检查。
@@ -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 .agents/notes/implemented/architecture/2026-08-05-large-session-jsonl-restore-pipeline.md
-2026-08-05-large-session-jsonl-restore-pipeline.md: eab53c683880ef7095233ed8122e532eb5add547
-2026-08-05-large-session-jsonl-restore-pipeline.zh.md: 2cd0d2ca20074d6adb0735db08071638ae2ced88
+2026-08-05-large-session-jsonl-restore-pipeline.md: 309d9dc6bdb5c3160f3e6e76a8318915df58fe59
+2026-08-05-large-session-jsonl-restore-pipeline.zh.md: 28acd3ebe804dca22a0619c257ff3ad9c09500a9
@@ -41,7 +41,7 @@ Borrowed seeds used by ordinary creation and fork paths still take a JSON snapsh
 - **Concatenate all plaintext before scanning** — rejected because it retains the compressed input, complete plaintext, whole-log UTF-8 string, line metadata, and parsed rows at the same time, and it rescans a torn-frame prefix.
 - **Implement a streaming JSON parser** — rejected because JSONL already provides record boundaries; native newline search plus `JSON.parse` removes the large intermediates without owning another parser or changing JSON semantics.
 - **Use a shared `WeakSet` while freezing restored events** — rejected because JSON materialization cannot produce cycles, and the set adds a lookup per object while retaining the complete graph during traversal.
- **Skip validation or freezing for restored values** — rejected because durable storage is a runtime boundary and `Session.events` promises immutable accepted history. The optimized path specializes those operations around stronger ownership facts instead of removing them.
+- **Skip validation or freezing for restored values** — rejected because durable storage is a runtime boundary and Session read methods promise immutable accepted history. The optimized path specializes those operations around stronger ownership facts instead of removing them.

 ## Consequences

@@ -41,7 +41,7 @@ Zstandard 结构扫描器会在解码前识别完整帧范围。系统单独解
 - **扫描前拼接全部明文**：不予采纳，因为该方案会同时保留压缩输入、完整明文、整份日志的 UTF-8 字符串、行元数据和解析记录，并会重新扫描撕裂帧前缀。
 - **实现流式 JSON 解析器**：不予采纳，因为 JSONL 已提供记录边界；使用原生换行搜索与 `JSON.parse` 就能移除大型中间结构，无需自行维护另一套解析器或改变 JSON 语义。
 - **冻结恢复事件时共享一个 `WeakSet`**：不予采纳，因为 JSON 物化不可能产生循环引用，而该集合会对每个对象增加一次查找，并在遍历期间保留完整对象图。
- **跳过恢复值的校验或冻结**：不予采纳，因为持久存储属于运行时边界，而 `Session.events` 承诺已接受历史不可变。优化路径利用更强的所有权事实特化这些操作，而不是将其移除。
+- **跳过恢复值的校验或冻结**：不予采纳，因为持久存储属于运行时边界，而 Session 读取方法承诺已接受历史不可变。优化路径利用更强的所有权事实特化这些操作，而不是将其移除。

 ## 后果

@@ -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 .agents/notes/implemented/architecture/2026-08-10-fork-children-stay-one-shot.md
-2026-08-10-fork-children-stay-one-shot.md: ba1e99c4d78d14230a2199cd7fb3c3eb9d3ad754
-2026-08-10-fork-children-stay-one-shot.zh.md: 0d4e214be952f5a4d96c296f38a5b0dbfbd85c86
+2026-08-10-fork-children-stay-one-shot.md: 0edf6ca9347dcf57313365e570daed22589f69fd
+2026-08-10-fork-children-stay-one-shot.zh.md: e832846d4018cf328e02c693bf21fe47a401d320
@@ -1,4 +1,4 @@
-# Agent Note: Forked children stay one-shot
+# Agent Note: Forked children preserve the parent request prefix

 Status: implemented

@@ -6,44 +6,36 @@ English | [中文](2026-08-10-fork-children-stay-one-shot.zh.md)

 ## Problem

-Fork's only difference from spawn is that the child Session is seeded with the parent's completed-turn prefix ([subagent-fork-in-process](../../../../packages/subagent/subagent-fork-in-process/README.md)). That seed costs real tokens — the inherited history is re-sent in every child request — and its one concrete payoff is provider-side prefix reuse: under the same provider and model, a child request whose leading bytes are identical to the parent's re-prefills none of the shared span. Anything a child scope adds *ahead* of the inherited history spends that payoff, because reuse stops at the first differing byte.
+Fork differs from spawn by seeding the child Session with the parent's completed-turn prefix. That seed costs tokens, and its intended payoff is provider-side prefix reuse: under the same provider and model, a child request whose leading bytes match the parent's does not prefill the shared span again. A child-only system-prompt section or tool schema ahead of the inherited history defeats that payoff.

-The child-scoped `report` return channel is now the largest such addition, and since [the report obligation](../feature/2026-08-06-continuable-child-report-obligation.md) it is two deltas rather than one: the `report` tool schema and the `tool:report` system-prompt section. Both live in the request head — the system block and the tool block precede every message — so a continuable forked child invalidates reuse before the first inherited turn and re-prefills the whole transcript it was forked to reuse. That composition pays fork's duplication cost and collects none of its benefit, while the parent still holds a reusable prefix the child could have shared.
+The earlier shipped composition avoided this mismatch by keeping forked children one-shot. That restriction was a consequence of the former child-only return tool, not an intrinsic property of continuable fork.

 ## Decision

-Every shipped composition inherits the fork delegation tool's `backgroundMode: one-shot` from the [base bundle](../../../../packages/bundle/base/cordis.patch.yml). The base bundle leaves `run_in_background` available because it also mounts the task service needed to settle background work.
+The model-facing `send_message` tool is registered globally for every Agent in a composition. A continuable forked child therefore receives the same tool name, description, schema, and ordering as its parent. Its initial task is appended after the inherited Session seed, and the task includes the direct parent id plus guidance to return results with `send_message({ agent_id, message })` when that tool is visible to the child.

-One-shot children — foreground and background alike — are created through `SubagentRuntime.start()`, which never enters the continuable activation-setup registry, so neither `report` nor its prompt section is installed. A forked one-shot child's system prompt and tool schemas therefore equal its parent's, apart from the `persona` and `toolFilter` deltas a deployment opts into per delegation tool.
+The base and headless compositions retain one-shot fork as their conservative lifecycle policy. The `cordis`, `standard`, and `ptc` CLI presets may bind fork to the continuable lifecycle because that binding no longer inserts child-only request-head fields. `ForkInProcessProvider.prepareContinuable()` and `ctx.subagents.startContinuable()` remain the implementation seam for those presets.

-`spawn` keeps `backgroundMode: continuable`. Continuable children and the report obligation ship unchanged for the provider whose child starts with no inherited prefix to protect, so this decision costs the report channel nothing.
-
-### The restriction is composition, not code
-
-`ForkInProcessProvider.prepareContinuable` stays implemented and `ctx.subagents.startContinuable()` still accepts `fork`; only the shipped `cordis.yml` rows changed. `tool-subagent` knows both the provider's `inheritsParentContext` and its own `backgroundMode` at mount, so a load-time rejection of the pair was available and is deliberately not added: the pair is not wrong in general. It is wrong only while a child-scope delta precedes inherited history, and the package that creates that delta — [`dsh-tool-subagent-report`](../../../../packages/subagent/tool-subagent-report/README.md) — is separately installable and, by its own design, invisible to `tool-subagent`. A deployment that omits the report package can run continuable forked children with the prefix intact. Encoding one roster's consequence as a delegation-tool invariant would make the tool assert something it cannot observe.
-
-The reintroduction condition is recorded as a `TODO(fork-continuable-prefix-reuse)` marker on `prepareContinuable` itself, the one method the shipped compositions do not call, and tracked as issue #2124: continuable fork reopens when a child's system prompt and tool schemas can match its parent's byte for byte.
+Byte-identical prefix reuse is qualified by explicit deployment choices. A fork delegation that applies a child persona or `toolFilter` may still change the request head. In particular, filtering out `send_message` removes both the schema and the return guidance from the child; the runtime does not bypass an explicit allow-list.

 ## Alternatives considered

-**Reject `inheritsParentContext` + `continuable` at mount.** A loud load-time failure would prevent silent reintroduction, which is what the configuration change cannot do. Rejected because the delegation tool cannot see the report package and the combination is legitimate without it; the invariant would be false for a deployment that never installs a child-scope delta, and `tool-subagent` would be asserting a fact owned by the roster.
+**Keep every fork one-shot.** This preserves the prefix but unnecessarily gives up durable, multi-turn forked children after the child-only schema difference is gone.

-**Stop mounting the fork provider at all.** This was the broader form of the restriction. Rejected because foreground fork *is* the prefix-reusing case and is untouched by the report channel, so a full ban gives up the capability without buying anything the one-shot binding does not already buy — and would leave no shipped composition exercising session seeding.
+**Install a child-only return alias.** A recipient-free alias would make child calls shorter, but it would recreate a tool-schema and prompt delta before inherited history and duplicate the adjacent-Agent operation.

-**Ship continuable forked children and accept the loss.** Rejected because the loss is total rather than marginal: reuse breaks ahead of the inherited history, so the child pays full prefill on a transcript it duplicated for the sole purpose of not paying it. A deployment that wants a long-lived child with no inherited context already has `spawn`.
+**Add the return instruction to the system prompt.** This would place child-only bytes ahead of inherited messages. Appending it to the initial user task preserves the inherited prefix and keeps the parent id next to the task that needs it.

-**Make `report` visible to every Agent.** A global registration would restore byte-identical prefixes by giving parent and child the same schema and section. Rejected because roots, one-shot children, remote children, and agentless callers would advertise a tool with no derivable recipient, and execution-time rejection would make schema visibility disagree with authority — the scope-local decision the [report tool Agent Note](../feature/2026-07-30-continuable-subagent-report-tool.md) already settled.
-
-**Install the child-scope deltas after the inherited history.** Rejected as unrepresentable: the system prompt and the tool schemas are request-head structures in every provider's wire format, so no ordering within them can place a child-only addition behind the message list.
+**Ignore an explicit child `toolFilter`.** Structural return tools previously bypassed the child allow-list. Rejected because a declared tool restriction must determine both schema visibility and guidance; hidden authority would make the model-facing roster inaccurate.

 ## Consequences

- No shipped composition creates a continuable forked child; `subagent_fork` returns a result to its caller's turn, and `send_message` addresses only spawned children.
- A forked child's request prefix stays byte-identical to its parent's unless the deployment configures `persona` or `toolFilter` on the fork delegation tool, so the token cost of seeding buys provider-side reuse again.
- The fork provider's continuable path has no production caller and no assembled-composition coverage. It keeps its package-level tests, and the seam still accepts it, so a bundle or `--patch` overlay can reintroduce it with no code change and no warning.
- `subagent_fork`'s model-visible schema changes: the continuable background wording is replaced by the one-shot task wording in the base bundle, and disappears entirely from the two examples. The affected keyless snapshot tool-schema sidecars are re-recorded in the same change.
- The report obligation's reach narrows to spawned children in shipped deployments. Its default `next-step` scheduling, authority model, and coverage remain independent of fork composition.
+- Parent and continuable-fork child expose byte-identical ordered tool schemas when the delegation does not request a persona or tool filter.
+- The inherited Session seed precedes the child's initial task and return guidance.
+- The base and headless profiles keep one-shot fork, while selected CLI presets exercise continuable fork without a child-only request-head addition.
+- A child sends zero or more messages to its direct parent explicitly; its final answer is not implicitly copied. The manager-owned settlement notice remains unconditional and separate.
+- Keyless snapshots and package tests pin schema equality, inherited-history ordering, parent-id guidance, and child-to-parent delivery through the same `send_message` operation used in the other direction.

 ### Accepted risks

-The constraint lives in three configuration files and a code comment, not in a gate. A future bundle row or profile patch can set `backgroundMode: continuable` on a fork tool and silently reintroduce the prefix loss; nothing fails loud. That is the accepted cost of not encoding one roster's consequence into `tool-subagent`.
+Provider-side prefix reuse still depends on the selected provider and model and on the absence of explicit persona or tool-filter differences. The harness proves equality of its assembled request-head inputs, not a provider's cache behavior.
@@ -1,49 +1,41 @@
-# Agent Note: fork 出的 child 保持 one-shot
+# Agent Note：Fork child 保留 parent 请求前缀

-Status: implemented
+状态：已实现

 [English](2026-08-10-fork-children-stay-one-shot.md) | 中文

 ## 问题

-fork 与 spawn 的唯一区别是 child 的 Session 会以 parent 已完成轮次的前缀作为初始内容（见 [subagent-fork-in-process](../../../../packages/subagent/subagent-fork-in-process/README.zh.md)）。这份初始内容有实打实的 token 成本——继承的历史会在 child 的每次请求中重新发送——而它唯一确定的回报是提供方侧的前缀复用：在提供方与模型相同的前提下，起始字节与 parent 逐字节相同的 child 请求，无需为这段共享区间重新预填充。任何由 child 作用域添加在继承历史*之前*的内容都会消耗掉这份回报，因为复用在第一个不同字节处即告停止。
+fork 与 spawn 的差异在于：fork 会用 parent 已完成轮次的前缀作为 child Session 的种子。该种子会消耗 token，其预期收益是提供方侧的前缀复用：使用相同提供方和模型时，如果 child 请求的开头字节与 parent 相同，共享区段就无需再次预填充。任何位于继承历史之前、仅属于 child 的系统提示词 section 或工具 schema 都会破坏这项收益。

-作用域局部的 `report` 返回通道现在是此类添加中最大的一项，而自[report 义务](../feature/2026-08-06-continuable-child-report-obligation.zh.md)起它是两项而非一项增量：`report` 工具 schema，以及 `tool:report` 系统提示词 section。两者都位于请求头部——系统块与工具块先于所有消息——因此一个可继续的 fork child 会在第一条继承轮次之前就使复用失效，并重新预填充它当初 fork 就是为了复用的整份 transcript（文本记录）。这种组合付出了 fork 的复制成本却收不到它的收益，而 parent 手上仍握着一份 child 本可共享的可复用前缀。
+先前的随附组合通过把 fork child 保持为 one-shot 来避开这种不匹配。该限制源于原来的 child-only 返回工具，并非可继续 fork 的固有属性。

 ## 决策

-所有交付组合都从 [base bundle](../../../../packages/bundle/base/cordis.patch.yml)继承 fork 委派工具的 `backgroundMode: one-shot`。base bundle 保留 `run_in_background`，因为它也挂载了结算后台工作所需的 task 服务。
+面向模型的 `send_message` 工具在组合中的每个 Agent 上全局注册。因此，可继续 fork child 获得与 parent 相同的工具名称、描述、schema 和顺序。其初始任务追加在继承的 Session 种子之后；当 child 可以看到该工具时，任务还会包含直接 parent id，以及使用 `send_message({ agent_id, message })` 返回结果的指引。

-one-shot child——前台与后台皆然——经由 `SubagentRuntime.start()` 创建，该路径从不进入可继续的 activation setup 注册表，因此 `report` 与它的提示词 section 都不会被安装。于是一个 fork 出的 one-shot child 的系统提示词与工具 schema 与其 parent 相同，只差部署逐个委派工具主动选择的 `persona` 与 `toolFilter` 增量。
+base 与 headless 组合保留 one-shot fork 作为其保守生命周期策略。`cordis`、`standard` 和 `ptc` CLI preset 可以把 fork 绑定为可继续生命周期，因为该绑定不再插入 child-only 请求头字段。`ForkInProcessProvider.prepareContinuable()` 与 `ctx.subagents.startContinuable()` 仍是这些 preset 使用的实现 seam。

-`spawn` 保持 `backgroundMode: continuable`。对于 child 起步时本就没有继承前缀需要保护的那个提供方，可继续 child 与 report 义务随附行为不变，因此本决策没有让 report 通道付出任何代价。
+逐字节相同的前缀复用受显式部署选择约束。配置 child persona 或 `toolFilter` 的 fork 委派仍可能改变请求头。尤其是过滤掉 `send_message` 时，child 会同时失去该 schema 与返回指引；运行时不会绕过显式 allow-list。

-### 该限制在于组合，不在于代码
+## 考虑过的替代方案

-`ForkInProcessProvider.prepareContinuable` 仍然实现完好，`ctx.subagents.startContinuable()` 也仍接受 `fork`；改动的只有随附的 `cordis.yml` 行。`tool-subagent` 在挂载时同时知道提供方的 `inheritsParentContext` 与自身的 `backgroundMode`，因此一个加载期拒绝该组合的检查是可行的，而这里刻意不加：该组合并非普遍错误。它只在某个 child 作用域增量位于继承历史之前时才是错的，而产生该增量的包——[`dsh-tool-subagent-report`](../../../../packages/subagent/tool-subagent-report/README.zh.md)——是独立安装的，并且按其自身设计对 `tool-subagent` 不可见。一个不安装 report 包的部署可以在前缀完好的前提下运行可继续的 fork child。把某一份插件清单的后果写成委派工具的不变量，会让该工具断言它无法观察到的事实。
+**让所有 fork 保持 one-shot。** 这能保留前缀，但 child-only schema 差异消失后，继续放弃持久且多轮的 fork child 已无必要。

-重新开放的条件记录为 `prepareContinuable` 方法上的 `TODO(fork-continuable-prefix-reuse)` 标记——随附组合不调用这个方法——并由 issue #2124 跟踪：当 child 的系统提示词与工具 schema 能与其 parent 逐字节一致时，可继续 fork 即可重新开放。
+**安装 child-only 返回别名。** 无需填写接收方的别名可以缩短 child 调用，但会在继承历史之前重新产生工具 schema 与提示词差异，并重复相邻 Agent 操作。

-## 备选方案
+**把返回指令放进系统提示词。** 这会在继承消息之前加入 child-only 字节。将其追加到初始用户任务，可以保留继承前缀，并让 parent id 紧邻需要它的任务。

-**在挂载时拒绝 `inheritsParentContext` 与 `continuable` 的组合。** 一次响亮的加载期失败可以阻止悄然的重新引入，而配置改动做不到这一点。否决的原因是委派工具看不到 report 包，且在没有它时该组合是合法的；对于从不安装任何 child 作用域增量的部署，这个不变量是假的，而 `tool-subagent` 会去断言一件由插件清单拥有的事实。
-
-**干脆不挂载 fork 提供方。** 这是该限制更彻底的形式。否决的原因是前台 fork *正是*复用前缀的那种情形，且不受 report 通道影响，因此全面禁用会在不换来任何 one-shot 绑定尚未换来的东西的同时放弃该能力——并且随附组合将没有任何一个演练 session 初始内容。
-
-**照常随附可继续的 fork child 并接受这份损失。** 否决的原因是这份损失是全额而非边际的：复用在继承历史之前就已中断，于是 child 为一份自己复制过来、目的恰恰是不必付费的 transcript 付了全额预填充。想要一个没有继承上下文的长期 child 的部署，本来就有 `spawn`。
-
-**让 `report` 对每个 Agent 可见。** 全局注册会通过让 parent 与 child 拥有相同的 schema 与 section 来恢复逐字节相同的前缀。否决的原因是根 agent、one-shot child、远端 child 与无 agent 调用方都会宣告一件推导不出收件方的工具，而执行期拒绝会让 schema 可见性与权限彼此矛盾——这正是[report 工具 Agent Note](../feature/2026-07-30-continuable-subagent-report-tool.zh.md)已经定下的作用域局部决策。
-
-**把 child 作用域增量安装到继承历史之后。** 否决的原因是它无法表达：在每个提供方的协议格式中，系统提示词与工具 schema 都是请求头部结构，因此它们内部的任何排序都无法把仅属于 child 的添加放到消息列表之后。
+**忽略显式 child `toolFilter`。** 结构性返回工具过去会绕过 child allow-list。否决该方案，因为声明的工具限制必须同时决定 schema 可见性与指引；隐藏权限会让面向模型的工具清单失真。

 ## 后果

- 没有任何随附组合会创建可继续的 fork child；`subagent_fork` 把结果返回给调用方的轮次，而 `send_message` 只寻址 spawn 出的 child。
- 除非部署在 fork 委派工具上配置了 `persona` 或 `toolFilter`，fork child 的请求前缀与其 parent 逐字节相同，因此初始内容的 token 成本重新换来了提供方侧的复用。
- fork 提供方的可继续路径没有生产调用方，也没有整体组装层面的覆盖。它保留自己的包内测试，seam 也仍然接受它，因此某个组合包或 `--patch` 覆盖层可以无需改动代码、也不会有任何警告地把它重新引入。
- `subagent_fork` 面向模型的 schema 发生变化：base 组合包中可继续的后台措辞被 one-shot 的 task 措辞取代，在两个示例中则完全消失。受影响的无密钥快照工具 schema 伴随文件在同一次改动中重新记录。
- 在随附部署中，report 义务的覆盖范围收窄到 spawn 出的 child。它的 `next-step` 默认调度、权限模型与覆盖仍独立于 fork 组合。
+- 未请求 persona 或工具过滤时，parent 与可继续 fork child 暴露逐字节相同且顺序一致的工具 schema。
+- 继承的 Session 种子位于 child 初始任务及其返回指引之前。
+- base 与 headless profile 保持 one-shot fork；选定的 CLI preset 会在没有 child-only 请求头增量的前提下使用可继续 fork。
+- child 显式向直接 parent 发送零条或多条消息；最终回答不会被隐式复制。管理器负责的结算通知仍然无条件执行，并且与 Agent 消息分离。
+- keyless snapshot 与包测试固定 schema 相等性、继承历史顺序、parent-id 指引，以及通过同一个 `send_message` 操作完成的 child-to-parent 投递。

-### 已接受的风险
+### 已接受风险

-该限制存在于三个配置文件与一处代码注释中，而不在门禁里。未来某个组合包行或 profile 补丁可以在 fork 工具上设置 `backgroundMode: continuable`，从而悄然重新引入前缀损失；没有任何东西会失败得很响亮。这就是不把某一份插件清单的后果写入 `tool-subagent` 所接受的代价。
+提供方侧前缀复用仍取决于选定的提供方和模型，以及是否不存在显式 persona 或工具过滤差异。harness 证明的是自己组装出的请求头输入相等，而不是提供方的缓存行为。
@@ -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 .agents/notes/implemented/architecture/2026-08-10-remote-event-delivery.md
-2026-08-10-remote-event-delivery.md: 4b9c2f224e36fb97f798492c999726eb55bad214
-2026-08-10-remote-event-delivery.zh.md: 4ca1f8c244df4e985d1223b6c468b450be943abf
+2026-08-10-remote-event-delivery.md: 1706cc525bd7707874353a6658ec2f52597044a9
+2026-08-10-remote-event-delivery.zh.md: ac79642e90791892870c8920e02867faea180b07
@@ -194,4 +194,4 @@ The few required Client symbols are mirrored on the test side: `scaffold.ts` exp
 - **Test-side mirrors can drift.** No mechanism compares mirrored Client constants under `apps/web/tests` with their source. Drift instead produces a selector mismatch. `apps/web/tests/README.md` records the review rule; a grep-level gate is deliberately omitted.
 - **Capabilities deliberately omitted.** Payload projection and redaction are unsupported, scopes other than Agent are unsupported, and ordinary notifications are not replayed. Recoverable state needs a query, cursor, or opening baseline; a waterfall is replayed only while its original Host invocation remains pending.
 - **Some Client packages remain in the Host graph.** Twelve projects, including `connection`, `runtime`, and `ui-slots`, remain reachable through unsplit `directory-picker-browse`/`-native` and `api/gateway → client/connection`. They compile and no longer pull in API Remotes' Client face, so this change does not split them. Direct `dsh-client-runtime/client` imports in two chat e2e files rely on Runtime's current presence in that graph rather than a general guarantee.
- **The invariant companion intentionally has no runtime check.** A prior revision asserted delivery form on the live event bus, coupling the companion to the allowlist and causing Rolldown to emit a third bundle chunk omitted by the mechanically derived publication list. The Host-face `TypertForwardableEventEntry` assertion already rejects those mismatches at compile time, so the companion is an explained empty installer.
+- **The package intentionally publishes no invariant companion.** A prior revision asserted delivery form on the live event bus, coupling diagnostics to the allowlist and causing Rolldown to emit a third bundle chunk omitted by the mechanically derived publication list. The Host-face `TypertForwardableEventEntry` assertion already rejects those mismatches at compile time, and the package README records why no independent runtime relation remains.
@@ -196,4 +196,4 @@ Client 要求首项是带非空 `clientId` 与 `host.home` 的 `ready`；后续
 - **测试侧镜像值可能漂移**：没有任何机制核对 `apps/web/tests` 中镜像的 client 常量与其源；安全网只是漂移会让选择器失配。规则写在 `apps/web/tests/README.md`，由 review 守；grep 级门禁经评估后刻意不做。
 - **放弃的能力**：不支持投影或脱敏载荷，不支持 Agent 以外的 Scope，也不为普通通知提供重放。需要可靠恢复的状态必须拥有查询、cursor 或 opening baseline；waterfall 只重放仍处于同一次 Host 调用生命周期内的 pending request。
 - **仍有 client 包留在 host 图里**：12 个工程（`connection`、`runtime`、`ui-slots` 等）经未拆分的 `directory-picker-browse`/`-native` 与 `api/gateway → client/connection` 仍可达 host 图。它们都能编译且不再牵连 api/remotes 的 client face，因此没有阻塞本次改动；拆分那些包能减少几个，但经评估后不做。两个 chat e2e 直接引 `dsh-client-runtime/client` 依赖 `runtime` 本来就在图里——属偶然而非保证。
- **invariant companion 不做运行期检查**：早先的修订曾在活事件总线上断言投递形状（`thisArg === null`、`mode === 'emit'`），这让 companion 与名单值耦合，并使 rolldown 把它提成第三个 bundle chunk——而机械推导的发布文件清单并不携带它。host 面的 `TypertForwardableEvent` 断言在编译期已拒绝这两种偏离，因此该 companion 是一个带说明的空 installer。
+- **本包不发布 invariant companion**：早先的修订曾在活事件总线上断言投递形状（`thisArg === null`、`mode === 'emit'`），这让诊断逻辑与名单值耦合，并使 rolldown 把它提成第三个 bundle chunk——而机械推导的发布文件清单并不携带它。Host 面的 `TypertForwardableEventEntry` 断言已在编译期拒绝这些偏离，包 README 也记录了不再存在独立运行时关系的原因。
@@ -1,6 +1,6 @@
 # Bilingual-pair consistency record (docs/i18n/README.md): the git blob hash of each
 # 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 .agents/notes/implemented/feature/2026-07-30-continuable-subagent-report-tool.md
-2026-07-30-continuable-subagent-report-tool.md: 07d17f18f318a86070d9b8612512fa3c2a3815e2
-2026-07-30-continuable-subagent-report-tool.zh.md: 9cbf76a9f3f50d9b1a84b573621cd6fc2a1611c4
+#   pnpm run verify-translation-pairing --write .agents/notes/implemented/architecture/2026-08-21-session-log-read-intent.md
+2026-08-21-session-log-read-intent.md: 81cdd770886b24588052d9860e42e4a07650fb03
+2026-08-21-session-log-read-intent.zh.md: e5d0425981e6425323260661bee42281b1e96c98
@@ -0,0 +1,32 @@
+# Agent Note: Session log reads state their materialization cost
+
+Status: implemented
+
+English | [中文](2026-08-21-session-log-read-intent.zh.md)
+
+## Problem
+
+An all-purpose `Session.events` accessor hides an array-sized copy behind every read after an append. The full frozen snapshot can be cached, but streaming invalidates that cache for every new event, so a caller that only needs the log length or one event can repeatedly copy millions of references. Making the return value immutable does not require every read intent to pay that cost; event immutability is owned separately by the [source-owned session immutability decision](2026-06-11-dev-invariants-over-deep-readonly.md).
+
+## Decision
+
+`Session` exposes three cost-specific read operations. `seq` reads the current length in constant time, `eventAt(seq)` reads one event in constant time, and `snapshotEvents(fromSeq?, toSeqExclusive?)` explicitly materializes a frozen array for consumers that need array operations or a stable materialized range. Sequence parameters are non-negative log positions, not `Array.prototype.slice` offsets from the end. Recurring domain state such as the selected agent preset is read from a Session projection instead of rescanning live history.
+
+The complete current snapshot is cached until append because repeated whole-log consumers can share the same immutable array. A range snapshot copies only the selected references and is not cached: arbitrary range caching would retain unbounded arrays and require an eviction policy. Previously returned snapshots remain stable because accepted events are immutable and a snapshot array never grows after append.
+
+Recurring domain-state reads use [session projections](2026-08-19-session-projection-state-and-client-views.md) when the required value can be maintained incrementally. Raw-log snapshots remain appropriate for persistence, export, replay, and consumers whose output is the event sequence itself. This API makes materialization visible but does not attempt to eliminate every full-log fold in the same change.
+
+## Alternatives considered
+
+**Keep a cached `events` array accessor.** This preserves ordinary array syntax but makes scalar and indexed reads appear cheap while an append can turn either into a whole-log copy.
+
+**Return a custom immutable cut with array-like traversal operations.** A captured length could provide a stable constant-time cut over the growing log, but the abstraction would reimplement selected array semantics and keep expanding as callers request more operations. Explicit indexed reads, explicit materialization, and projections cover the shipped intents with a smaller public API.
+
+**Cache every materialized range or maintain an incremental chunked snapshot.** Range caching needs retention and eviction rules, while a chunked public representation changes consumers and serialization for a cost that many callers avoid through indexed reads or projections. These representations remain options if measured full-snapshot consumers justify them.
+
+## Consequences
+
+- Length and single-event reads do not copy the log.
+- Full and ranged snapshots retain an explicit linear cost proportional to the selected event count.
+- Consumers choose between raw history and incrementally maintained state at the call site.
+- The public API does not emulate an array; callers materialize only when they need array operations.
@@ -0,0 +1,32 @@
+# Agent Note: 会话日志读取显式表达物化成本
+
+Status: implemented
+
+[English](2026-08-21-session-log-read-intent.md) | 中文
+
+## 问题
+
+通用的 `Session.events` 访问器会在每次 append 后的首次读取中隐藏一次与数组大小成正比的复制。完整冻结快照可以缓存，但流式输出的每个新事件都会使缓存失效，因此只需要日志长度或单个事件的调用方也可能反复复制数百万个引用。返回值不可变并不要求每种读取意图都承担该成本；事件不可变性由[源端拥有的会话不可变性决策](2026-06-11-dev-invariants-over-deep-readonly.zh.md)另行负责。
+
+## 决策
+
+`Session` 提供三种成本不同的读取操作。`seq` 以常数时间读取当前长度，`eventAt(seq)` 以常数时间读取一个事件，`snapshotEvents(fromSeq?, toSeqExclusive?)` 则为需要数组操作或稳定物化区间的消费方显式物化冻结数组。序列参数是非负日志位置，不是从末尾计算的 `Array.prototype.slice` 偏移量。所选 agent preset 这类重复读取的领域状态来自会话投影，而不是反复扫描活跃日志。
+
+当前完整快照会缓存到下一次 append，使重复读取整个日志的消费方可以共享同一个不可变数组。区间快照只复制所选引用且不缓存：缓存任意区间会保留数量无界的数组，并要求额外的淘汰策略。先前返回的快照保持稳定，因为已接受的事件不可变，且快照数组不会在 append 后增长。
+
+当所需值可以增量维护时，重复的领域状态读取使用[会话投影](2026-08-19-session-projection-state-and-client-views.zh.md)。持久化、导出、回放，以及输出本身就是事件序列的消费方仍适合使用原始日志快照。此 API 使物化成本显式可见，但不会在同一项改动中消除所有完整日志折叠。
+
+## 曾考虑的替代方案
+
+**保留缓存的 `events` 数组访问器。** 这保留了普通数组语法，但会让标量读取和索引读取看似廉价，而一次 append 就可能使其中任何一次读取变成完整日志复制。
+
+**返回提供类数组遍历操作的自定义不可变 cut。** 捕获长度可以在持续增长的日志上提供稳定且常数时间的 cut，但随着调用方要求更多操作，该抽象会不断重新实现所选数组语义。显式索引读取、显式物化和投影以更小的公开 API 覆盖了已经交付的读取意图。
+
+**缓存每个物化区间或维护增量分片快照。** 区间缓存需要保留与淘汰规则，而公开的分片表示会改变消费方和序列化方式；许多调用方已通过索引读取或投影避开这项成本。若实测的完整快照消费方证明有必要，仍可考虑这些表示。
+
+## 后果
+
+- 长度读取和单事件读取不会复制日志。
+- 完整快照和区间快照仍有与所选事件数量成正比的显式线性成本。
+- 消费方在调用点选择原始历史或增量维护的状态。
+- 公开 API 不模拟数组；调用方只在需要数组操作时物化。
@@ -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 .agents/notes/implemented/architecture/2026-08-25-sparse-first-party-prompt-section-orders.md
-2026-08-25-sparse-first-party-prompt-section-orders.md: ffa2e6a4f602178007a6702938dfd71a2f85cbaa
-2026-08-25-sparse-first-party-prompt-section-orders.zh.md: d26ac18b075a2f0ebccccdbd072c7c066c2fe1b0
+2026-08-25-sparse-first-party-prompt-section-orders.md: 3cfdb58af3c576b3e46449a763db1c811073517e
+2026-08-25-sparse-first-party-prompt-section-orders.zh.md: 31cbb98eab9eb786ac9d854b76df4c18efb88563
@@ -24,7 +24,7 @@ The allocation preserves the established first-party sequence except for two del
 | Work modes | `plan:policy` 500, `team:policy` 600 |
 | Invocation prelude | `tools:ptc-only` 800, `context:file-reference` 900 |
 | Local tools | `tool:bash` 1000, `tool:pwsh` 1010, `tool:read` 1100, `tool:write` 1200, `tool:edit` 1300, `tool:glob` 1400, `tool:grep` 1500, `tool:jobs` 1600, `tool:pty` 1700 |
-| Higher-level tools | `tool:web_search` 2000, `tool:web_fetch` 2100, `tool:lsp` 2200, `tool:session-query` 2300, `tool:goal` 2400, `tool:cordis` 2500, `tool:workflow` 2600, `tool:ralph` 2700, continuable-subagent guidance 2800, `tool:report` 2900 |
+| Higher-level tools | `tool:web_search` 2000, `tool:web_fetch` 2100, `tool:lsp` 2200, `tool:session-query` 2300, `tool:goal` 2400, `tool:cordis` 2500, `tool:workflow` 2600, `tool:ralph` 2700, continuable-subagent guidance 2800 |
 | Generated protocol | `tools:sdk` 5000 |
 | Final-output obligations | deliverable file references 9000, `tool:structured_output` 9900 |

@@ -24,7 +24,7 @@ Status: implemented
 | 工作模式 | `plan:policy` 500、`team:policy` 600 |
 | 调用前置说明 | `tools:ptc-only` 800、`context:file-reference` 900 |
 | 本地工具 | `tool:bash` 1000、`tool:pwsh` 1010、`tool:read` 1100、`tool:write` 1200、`tool:edit` 1300、`tool:glob` 1400、`tool:grep` 1500、`tool:jobs` 1600、`tool:pty` 1700 |
-| 高层工具 | `tool:web_search` 2000、`tool:web_fetch` 2100、`tool:lsp` 2200、`tool:session-query` 2300、`tool:goal` 2400、`tool:cordis` 2500、`tool:workflow` 2600、`tool:ralph` 2700、可继续运行的 subagent 指导 2800、`tool:report` 2900 |
+| 高层工具 | `tool:web_search` 2000、`tool:web_fetch` 2100、`tool:lsp` 2200、`tool:session-query` 2300、`tool:goal` 2400、`tool:cordis` 2500、`tool:workflow` 2600、`tool:ralph` 2700、可继续运行的 subagent 指导 2800 |
 | 生成协议 | `tools:sdk` 5000 |
 | 最终输出义务 | 可交付文件引用 9000、`tool:structured_output` 9900 |

@@ -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 .agents/notes/implemented/architecture/2026-08-26-inspector-execution-realms-and-protocol-planes.md
-2026-08-26-inspector-execution-realms-and-protocol-planes.md: e8bff0661d2d0c86c216b0a18e2feb7a2c786709
-2026-08-26-inspector-execution-realms-and-protocol-planes.zh.md: 2db6307d1bfc536ac5e8b0f5d6f03e4cfe334989
+2026-08-26-inspector-execution-realms-and-protocol-planes.md: 1c274bda4cb10458ddc9ecec846b6355fb811c2f
+2026-08-26-inspector-execution-realms-and-protocol-planes.zh.md: 49fde96a3b0acc5d9621b1ca4cfd1546ca970931
@@ -16,7 +16,7 @@ The [cross-realm CDP inspector decision](2026-08-23-cross-realm-cdp-inspector.md

 Top-level source directories identify execution ownership. `client/` contains only browser Client code, `host/` only Host Node-main-thread code, `worker/` only Worker-thread code, and `shared/` code that is safe in every environment. A module that executes in the Worker on behalf of a Client belongs under `worker/`, not `client/`.

-The repository-required `src/index.ts` and `src/invariant.ts` discovery entries are the only root-level source exceptions. They expose the Host package entry and its service type or register the invariant companion, contain no Inspector runtime implementation, and remain at fixed paths for repository tooling.
+The repository-required `src/index.ts` discovery entry is the only root-level source exception in this package. It exposes the Host package entry and its service type, contains no Inspector runtime implementation, and remains at a fixed path for repository tooling. The package publishes no invariant companion because its relationships fail in their owning wire, generation, Worker, or CDP operations, as recorded in the README.

 ```text
 src/
@@ -16,7 +16,7 @@ Inspector 包的代码运行在三个 JavaScript 环境中：浏览器 Client、

 顶层源码目录标识执行归属。`client/` 只包含浏览器 Client 代码，`host/` 只包含 Host Node 主线程代码，`worker/` 只包含 Worker thread 代码，`shared/` 只包含在所有环境中都安全的代码。即使某个模块代表 Client，只要它实际在 Worker 中执行，就仍属于 `worker/`，而不是 `client/`。

-仓库要求的 `src/index.ts` 与 `src/invariant.ts` 发现入口是仅有的源码根目录例外。它们暴露 Host package entry 及其 service type，或注册 invariant companion，不包含 Inspector 运行时实现，并为仓库工具保留在固定路径。
+仓库要求的 `src/index.ts` 发现入口是本包唯一的源码根目录例外。它暴露 Host package entry 及其 service type，不包含 Inspector 运行时实现，并为仓库工具保留在固定路径。本包不发布 invariant companion，因为其关系会在所属 wire、generation、Worker 或 CDP 操作中失败，原因记录在 README 中。

 ```text
 src/
@@ -1,6 +1,6 @@
 # Bilingual-pair consistency record (docs/i18n/README.md): the git blob hash of each
 # 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 .agents/notes/implemented/bug-fix/2026-08-17-subagent-report-settlement-ordering.md
-2026-08-17-subagent-report-settlement-ordering.md: 30dfab5e96a7cea2ef6d4f03f480d17a86c5e775
-2026-08-17-subagent-report-settlement-ordering.zh.md: 658eb18e3a8cb40734136af32c6c62faef066a6e
+#   pnpm run verify-translation-pairing --write .agents/notes/implemented/architecture/2026-08-27-adjacent-agent-steer-messaging.md
+2026-08-27-adjacent-agent-steer-messaging.md: 31c5a4fc8c0e807ff8ab869561d0acd32eccc060
+2026-08-27-adjacent-agent-steer-messaging.zh.md: c49da882a5beb7a7e8379d15a0628fb5c617e040
@@ -0,0 +1,82 @@
+# Agent Note: Adjacent Agents share one Steer send_message operation
+
+Status: implemented
+
+English | [中文](2026-08-27-adjacent-agent-steer-messaging.zh.md)
+
+## Problem
+
+Continuable Agents originally used direction-specific model controls. A parent called `send_message({ subagent_id, message })`, which delegated to a FIFO `followup` service operation. A child instead received a child-scoped `report({ output })` tool, a `tool:report` system-prompt section, and deployment-selected quiet or waking delivery. The tools described one adjacent-Agent operation through different schemas, service paths, provenance, and scheduling.
+
+A continuable child owns its own Session, so its parent does not automatically receive the child's transcript, tool output, or reasoning. The return path must therefore remain explicit and repeatable: a child may send progress before it finishes, remain available after sending, or fail before it can cooperate. Turning every final assistant message into an implicit result would conflate turn completion with model-selected communication and would not cover abnormal endings.
+
+The child-only tool and system-prompt section also preceded every inherited fork turn. They made a continuable fork child's request head differ from its parent's before the history that fork exists to reuse, forcing the provider to prefill the entire copied transcript again.
+
+## Decision
+
+`SubagentRuntime.sendMessage(sender, targetId, content, { signal })` is the only public model-authored message operation. The continuation manager accepts only the exact live sender and a target on one adjacent edge:
+
+- parent to direct continuable child, authorized by the child's durable `SessionHeader.parentSession`;
+- resident continuable child to its exact live direct parent, authorized by the child's Activation.
+
+Siblings, self-targets, ancestors beyond one edge, stale Agent objects, unknown targets, and one-shot children are not alternate routes. The operation has no caller-supplied source, delivery mode, offline parent mailbox, or provider dispatch.
+
+Every accepted message uses `Agent.steer()`. A running target receives it at the nearest step boundary; an idle target starts a turn. An absent direct child is cold-resumed through the existing continuation lifecycle before the same Steer delivery. The manager retains waking-send accounting so a continuation-managed target cannot settle between synchronous inbox insertion and driver admission.
+
+Every direction uses one durable source. The service derives `senderSessionId` from the authorized Agent and frames the model-visible content as `Agent <sender-id> sent a message:`, so attribution cannot diverge from authority.
+
+```ts
+import type { SessionId } from '@deepseek-ai/dsh-session'
+
+interface AgentMessageSource {
+  readonly kind: 'agent-message'
+  readonly form: 'relay'
+  readonly senderSessionId: SessionId
+}
+```
+
+### One model tool and one return instruction
+
+The globally registered model tool is direction-neutral and has one fixed schema:
+
+```ts
+interface SendMessageInput {
+  readonly agent_id: string
+  readonly message: string
+}
+```
+
+Parents and children inherit the same definition in the same registry order. The standard definition carries a process-stable internal identity that a scoped same-name tool does not satisfy. A child `toolFilter` may explicitly remove the inherited tool, and a scoped replacement may provide different semantics; neither case receives the standard call instruction. When the standard tool remains visible, the continuation manager appends the JSON-encoded direct parent id and the instruction to send one self-contained result before finishing, plus earlier actionable findings, to the child's initial user task. For a fork child this task follows the inherited completed-turn prefix; no child-only system-prompt section or tool schema precedes that prefix.
+
+The instruction is guidance, not settlement enforcement. Sending does not end the child's turn, zero or several calls remain mechanically valid, and the runtime never rejects a child for staying silent. The manager-owned `subagent-settled` notice remains unconditional and separately attributed because it records how an Activation ended and preserves terminal output when the child cannot cooperate.
+
+Human browser prompts are not model-authored Agent messages. The remote prompt path keeps a private Queue delivery so each human prompt remains a distinct turn. Interrupt behavior and settlement delivery remain independent.
+
+### Complete removal and reintroduction condition
+
+The standalone `@deepseek-ai/dsh-tool-subagent-report` package, `report` schema, `tool:report` prompt section, `reportDelivery` configuration, report-specific message source, catalog entries, composition rows, and supported-behavior snapshots are absent. The unified tool gives up the recipient-free child shortcut and the old ability for a structural return tool to survive an explicit child allow-list. Those capabilities return only if a concrete use case requires semantics that an adjacent `agent_id` and fixed Steer cannot express; reintroducing them requires a distinct model operation and prefix-cost evidence, not an alias over `sendMessage()`.
+
+## Alternatives considered
+
+**Keep `followup` and add child-to-parent routing.** The name promises a later turn and inherits `Agent.followup()` semantics. It would obscure the chosen nearest-step behavior and preserve a parent-centric name for a direction-neutral capability.
+
+**Keep a recipient-free `report` wrapper over `sendMessage()`.** This preserves a convenient child shortcut and lets a scope-local registration survive global tool filtering. It loses because the separate schema and prompt duplicate one operation, make parent and child request heads differ, and let equivalent directions drift again.
+
+**Make `report` global.** Roots, one-shot children, remote children, and agentless callers cannot derive a report recipient. Advertising it globally would make schema visibility disagree with authority, while `send_message` already makes the recipient explicit.
+
+**Turn every child final message into an implicit send.** A long-lived child may have nothing useful to send in one turn and several findings in another. Automatic delivery would merge model-authored communication with the runtime's settlement account and could not replace the unconditional notice on errors, cancellation, or token exhaustion.
+
+**Rely only on the tool description.** A tool description helps after the model considers that tool; the failure mode is a child that believes it is finished without considering any return call. Initial-task guidance reaches that decision without changing the inherited system or tool prefix.
+
+**Keep quiet delivery as deployment policy.** A quiet model-authored message can be accepted while an idle target never reads it. Fixed Steer gives both directions one delivery meaning and preserves accepted order with later settlement notices.
+
+## Consequences
+
+- Model consumers expose one `send_message({ agent_id, message })` definition to parents and children, with no model-selected Queue versus Steer parameter.
+- The continuation manager remains the sole owner of adjacency authorization, residency, cold resume, waking admission, and teardown races.
+- Accepted messages may extend a running target's current turn; messages waiting together share next-step FIFO ordering.
+- Caller cancellation owns work only until inbox acceptance and does not retract an accepted message or dispose the target.
+- The initial task carries JSON-encoded dynamic parent addressing after a fork prefix, while the request-head system prompt and tool ordering remain reusable.
+- Human prompts, settlement notices, QueueDock, and the base bundle's one-shot fork policy remain separate decisions.
+
+This decision consolidates and removes the fully superseded report-tool and child-report-obligation records. It supersedes the `followup` naming choice in [Intent-named subagent continuation operations](../simplification/2026-07-27-intent-named-subagent-continuation-operations.md) and retains the accepted-order guarantee in [Child Agent messages precede their settlement notices](../bug-fix/2026-08-17-subagent-message-settlement-ordering.md).
@@ -0,0 +1,82 @@
+# Agent Note: 相邻 Agent 共享一个 Steer send_message 操作
+
+Status: implemented
+
+[English](2026-08-27-adjacent-agent-steer-messaging.md) | 中文
+
+## 问题
+
+可继续 Agent 最初使用方向专属的模型控制。parent 调用 `send_message({ subagent_id, message })`，委托给 FIFO `followup` 服务操作。child 则获得 child 作用域的 `report({ output })` 工具、`tool:report` 系统提示词 section，以及由部署选择的静默或唤醒投递。两个工具用不同 schema、服务路径、来源与调度描述同一个相邻 Agent 操作。
+
+可继续 child 拥有自己的 Session，因此 parent 不会自动收到 child 的 transcript（文本记录）、工具输出或推理。返回路径必须保持显式且可重复：child 可以在结束前发送进度、发送后仍保持可用，也可能在来得及配合前失败。把每条最终 assistant 消息变成隐式结果会混淆轮次完成与模型选择的通信，而且无法覆盖异常结束。
+
+child 专属工具与系统提示词 section 还位于每个继承 fork 轮次之前。它们让可继续 fork child 的请求头在 fork 旨在复用的历史之前就与 parent 不同，迫使提供方重新预填充整份复制 transcript。
+
+## 决策
+
+`SubagentRuntime.sendMessage(sender, targetId, content, { signal })` 是唯一公开的模型编写消息操作。继续执行管理器只接受确切在线 sender 与一条相邻边上的目标：
+
+- parent 到直接可继续 child，由 child 的持久化 `SessionHeader.parentSession` 授权；
+- 驻留的可继续 child 到其确切在线直接 parent，由 child 的 Activation 授权。
+
+sibling、自身目标、超过一条边的 ancestor、陈旧 Agent 对象、未知目标与一次性 child 都不是替代路由。该操作没有调用方提供的 source、投递模式、离线 parent mailbox 或提供方分发。
+
+每条被接受的消息都使用 `Agent.steer()`。运行中目标在最近 step 边界接收消息；空闲目标启动轮次。缺失的直接 child 会先通过现有继续执行生命周期冷恢复，再接受同一 Steer 投递。管理器保留唤醒发送记账，因此受继续执行管理的目标不会在同步 inbox 插入与 driver 准入之间结算。
+
+两个方向使用同一种持久来源。服务从已授权 Agent 推导 `senderSessionId`，并把模型可见内容组装为 `Agent <sender-id> sent a message:`，因此来源信息不会偏离权限。
+
+```ts
+import type { SessionId } from '@deepseek-ai/dsh-session'
+
+interface AgentMessageSource {
+  readonly kind: 'agent-message'
+  readonly form: 'relay'
+  readonly senderSessionId: SessionId
+}
+```
+
+### 一个模型工具与一条返回指令
+
+全局注册的模型工具与方向无关，并使用一个固定 schema：
+
+```ts
+interface SendMessageInput {
+  readonly agent_id: string
+  readonly message: string
+}
+```
+
+parent 与 child 以相同注册表顺序继承相同定义。标准定义携带进程稳定的内部身份，同名的作用域工具不满足该身份。child `toolFilter` 可以显式移除继承的工具，作用域替代工具也可以提供不同语义；两种情况都不会收到标准调用指令。当标准工具仍可见时，继续执行管理器会把经过 JSON 编码的直接 parent id、结束前发送一份自包含结果的指令，以及更早发送可操作发现的指令追加到 child 初始用户任务。对 fork child 而言，该任务位于继承的已完成轮次前缀之后；没有 child 专属系统提示词 section 或工具 schema 位于此前缀之前。
+
+该指令是指导，不是结算强制。发送不会结束 child 轮次，机制仍允许零次或多次调用，runtime 绝不会因 child 保持沉默而拒绝它。由管理器负责的 `subagent-settled` 通知仍无条件发送并采用独立来源，因为它记录 Activation 如何结束，并在 child 无法配合时保留终态输出。
+
+浏览器中的人类提示不是模型编写的 Agent 消息。远程提示路径保留私有 Queue 投递，使每条人类提示保持为独立轮次。中断行为与结算投递保持独立。
+
+### 完整移除与重新引入条件
+
+独立的 `@deepseek-ai/dsh-tool-subagent-report` 包、`report` schema、`tool:report` 提示词 section、`reportDelivery` 配置、report 专属消息来源、目录项、组合行和受支持行为快照均已不存在。统一工具放弃了无需接收方的 child 快捷方式，也放弃了让结构性返回工具绕过显式 child allow-list 的旧能力。只有具体用例需要相邻 `agent_id` 与固定 Steer 无法表达的语义时，这些能力才会重新出现；重新引入需要独立的模型操作与前缀成本证据，而非 `sendMessage()` 之上的别名。
+
+## 考虑过的替代方案
+
+**保留 `followup` 并添加 child 到 parent 路由。** 该名称承诺后续轮次并继承 `Agent.followup()` 语义。它会掩盖选定的最近 step 行为，并为方向无关能力保留以 parent 为中心的名称。
+
+**保留 `sendMessage()` 之上无需接收方的 `report` 包装层。** 这会保留便利的 child 快捷方式，并让作用域局部注册绕过全局工具过滤。它落选是因为独立 schema 与提示词重复一个操作、使 parent 与 child 请求头不同，并允许等价方向再次漂移。
+
+**让 `report` 全局可见。** 根 Agent、一次性 child、远程 child 与无 Agent 调用方无法推导 report 接收方。全局宣传它会让 schema 可见性与权限不一致，而 `send_message` 已显式给出接收方。
+
+**把每条 child 最终消息变成隐式发送。** 长期运行的 child 可能在某个轮次没有值得发送的内容，在另一个轮次却有多条发现。自动投递会混合模型编写通信与 runtime 结算说明，而且无法替代错误、取消或 token 耗尽时的无条件通知。
+
+**只依赖工具描述。** 工具描述会在模型考虑该工具后提供帮助；失败模式是 child 认为自己已经完成而根本没有考虑返回调用。初始任务指导能触及该决策，又不会改变继承的系统或工具前缀。
+
+**保留静默投递作为部署策略。** 静默的模型编写消息可能被接受，但空闲目标永远不会读取它。固定 Steer 为两个方向提供一种投递含义，并保持与后续结算通知的接受顺序。
+
+## 后果
+
+- 模型 Consumer 向 parent 与 child 公开一个 `send_message({ agent_id, message })` 定义，不提供模型选择的 Queue 与 Steer 参数。
+- 继续执行管理器仍是相邻关系授权、驻留、冷恢复、唤醒准入与拆卸竞态的唯一所有者。
+- 被接受的消息可以延长运行中目标的当前轮次；一起等待的消息共享 next-step FIFO 顺序。
+- 调用方取消只在 inbox 接受前掌管工作，不会撤回已接受消息或 dispose（资源释放）目标。
+- 初始任务在 fork 前缀之后携带经过 JSON 编码的动态 parent 地址，而请求头系统提示词与工具顺序保持可复用。
+- 人类提示、结算通知、QueueDock 与 base bundle 的一次性 fork 策略仍是独立决策。
+
+本决策合并并删除了已完全被取代的 report 工具与 child report 义务记录。它取代[按意图命名的 subagent 继续执行操作](../simplification/2026-07-27-intent-named-subagent-continuation-operations.zh.md)中的 `followup` 命名选择，并保留[Child Agent 消息先于其结算通知](../bug-fix/2026-08-17-subagent-message-settlement-ordering.zh.md)中的接受顺序保证。
@@ -0,0 +1,6 @@
+# Bilingual-pair consistency record (docs/i18n/README.md): the git blob hash of each
+# 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 .agents/notes/implemented/architecture/2026-08-31-viewport-activated-syntax-highlighting.md
+2026-08-31-viewport-activated-syntax-highlighting.md: 050c4e50240f9e1d1eb4f2708df6bc93c3a1faf1
+2026-08-31-viewport-activated-syntax-highlighting.zh.md: abd02161ae00d8bdd31863826565175e1d8fddf9
@@ -0,0 +1,41 @@
+# Agent Note: Viewport-activated syntax highlighting
+
+Status: implemented
+
+English | [中文](2026-08-31-viewport-activated-syntax-highlighting.zh.md)
+
+## Problem
+
+A long conversation mounts code fences and read cards far outside the visible viewport. Eager highlighting tokenizes every supported block and creates every token span during that mount, so main-thread work and DOM size scale with the whole rendered history rather than the code the reader can see. The [Shiki selection](../process/2026-07-26-web-syntax-highlighting-shiki.md) and [incremental streaming highlighter](../feature/2026-08-20-web-streaming-fence-highlight.md) bound initialization and repeated prefix work, but neither avoids the first per-block tokenization for unseen history.
+
+## Decision
+
+`useViewportHighlighting` owns one shared `IntersectionObserver` for syntax-highlightable `CodeBlock` and `ReadBlock` instances. A supported block renders its existing plain-text arm until its root first intersects the viewport. An absent or unsupported language never registers with the observer. A browser without `IntersectionObserver` activates highlighting after mount so the capability still works.
+
+The first intersecting entry removes its target from the observer and activates that component for the rest of its lifetime. Leaving the viewport never returns it to plain text. This one-way transition avoids repeated tokenization, token-DOM construction, and visual churn while scrolling. The shared observer disconnects when no inactive registered blocks remain.
+
+`CodeBlock` gates both settled `highlightToHtml` calls and streaming `StreamingHighlightSession` creation. It starts an activated stream from the current accumulated source, then retains the existing incremental tokenizer and React line caches. `ReadBlock` gates `highlightLines` while retaining its line rows and gutter. The plain and highlighted arms keep the same source text, code font, padding, wrapping, and line height; Shiki's color, bold, italic, and underline token styles remain unchanged.
+
+The module-level Shiki singleton warm-up remains eager. Viewport activation defers code-block content tokenization and token-span construction, not the fixed boot-grammar warm-up or the plain content DOM.
+
+## Testing
+
+The focused jsdom test replaces the process-global `IntersectionObserver`, mounts several code surfaces, and proves that non-intersecting and unsupported blocks remain plain, intersecting blocks share one observer, leaving the viewport does not remove highlighting, and an activated block continues to highlight changed source. It also covers read-card activation and observer disposal. Each test restores the global and unmounts every component, so the module-level registry cannot leak registrations into another case.
+
+Existing component tests run without `IntersectionObserver` and therefore cover the immediate fallback together with the established Shiki output, font styles, streaming caches, and plain-language behavior. Browser geometry is not measured by this unit suite; geometry stability relies on the unchanged shared typography and box styles of the plain and highlighted arms.
+
+## Alternatives considered
+
+**Deactivate highlighting when a block leaves the viewport.** This can reclaim token DOM from blocks already viewed, but scrolling repeatedly rebuilds the same token tree, discards streaming caches, and changes visible presentation at both viewport edges. One-way activation pays the cost at most once per mounted block.
+
+**Drop bold and italic token styles to make every token use identical font metrics.** This weakens syntax presentation, especially for highlighted Markdown, and is unnecessary for the chosen lifecycle: both render arms already use the same code font and fixed line height. Shiki's existing token styles remain intact.
+
+**Pin a measured pixel height during activation.** A fixed measurement becomes stale when a streaming fence grows or responsive wrapping changes, and can clip content or introduce an inner vertical scrollbar. The plain arm stays in normal flow instead of adding measurement state.
+
+**Virtualize complete code blocks or retain only a token window.** This can also bound DOM after a reader has visited every block, but it changes selection, copy, scroll anchoring, and streaming-cache ownership. Viewport activation removes unseen work without changing those behaviors.
+
+## Consequences
+
+Supported code that is never viewed incurs no content tokenization and creates no token spans. The first viewport intersection pays the normal synchronous highlight cost; a lazily imported grammar may keep the block plain until its existing load notification arrives. Activated blocks retain their highlighted DOM when scrolled away, so memory use grows with blocks the reader has visited rather than shrinking with the current viewport.
+
+The optimization is local to presentation. Markdown parsing, Shiki grammar selection and styling, stream-tail tokenization, copy text, and settled output remain unchanged.
@@ -0,0 +1,41 @@
+# Agent Note: 视口激活的语法高亮
+
+Status: implemented
+
+[English](2026-08-31-viewport-activated-syntax-highlighting.md) | 中文
+
+## 问题
+
+长对话会挂载远在可见视口之外的代码围栏和读取卡片。预先高亮会在挂载时 tokenize 每个受支持的块并创建全部 token span，使主线程工作量和 DOM 大小随完整渲染历史增长，而不是随读者能看到的代码增长。[Shiki 选型](../process/2026-07-26-web-syntax-highlighting-shiki.zh.md)与[流式增量高亮](../feature/2026-08-20-web-streaming-fence-highlight.zh.md)分别约束初始化成本和重复处理前缀的成本，但都无法避免不可见历史首次发生的逐块 tokenize。
+
+## 决策
+
+`useViewportHighlighting` 为可进行语法高亮的 `CodeBlock` 和 `ReadBlock` 实例持有一个共享的 `IntersectionObserver`。受支持的块在根元素首次与视口相交之前渲染既有的纯文本臂。语言缺失或不受支持时不会向 observer 注册。浏览器不提供 `IntersectionObserver` 时会在挂载后激活高亮，使该能力仍然可用。
+
+首个相交条目会从 observer 中移除自己的目标，并在该组件余下的生命周期里保持激活。离开视口不会恢复为纯文本。这种单向转换避免滚动时反复 tokenize、创建 token DOM 和发生视觉切换。不存在尚未激活的注册块时，共享 observer 会断开。
+
+`CodeBlock` 同时控制定稿态的 `highlightToHtml` 调用和流式 `StreamingHighlightSession` 的创建。流式块激活时从当前累积源码开始，随后保留既有的增量 tokenizer 与 React 行缓存。`ReadBlock` 控制 `highlightLines`，同时保留其行和行号槽。纯文本臂与高亮臂使用相同的源码文本、代码字体、内边距、换行规则和行高；Shiki 现有的 token 颜色、粗体、斜体和下划线样式保持不变。
+
+模块级 Shiki 单例仍然预先预热。视口激活延迟的是代码块内容的 tokenize 与 token span 创建，不是固定的启动语法预热或纯文本内容 DOM。
+
+## Testing
+
+聚焦的 jsdom 测试替换进程全局 `IntersectionObserver`，挂载多个代码表面，并证明未相交和不受支持的块保持纯文本、相交块共用一个 observer、离开视口不会移除高亮，且已激活的块会继续高亮变化后的源码。测试还覆盖读取卡片激活和 observer 释放。每个测试都会恢复全局值并卸载全部组件，因此模块级注册表不会把注册项泄漏到其他用例。
+
+既有组件测试在没有 `IntersectionObserver` 的环境中运行，因此同时覆盖立即回退路径，以及既有 Shiki 输出、字体样式、流式缓存和纯文本语言行为。该单元测试套件不测量浏览器几何尺寸；几何稳定性依赖纯文本臂与高亮臂不变的共享字体排印和盒模型样式。
+
+## 曾考虑的替代方案
+
+**块离开视口时停用高亮。** 这可以回收已经看过的块所占用的 token DOM，但滚动会反复重建同一棵 token 树、丢弃流式缓存，并在视口两端改变可见呈现。单向激活使每个已挂载块最多支付一次成本。
+
+**移除粗体和斜体 token 样式，使所有 token 使用完全相同的字体度量。** 这会削弱语法呈现，尤其影响高亮后的 Markdown；所选生命周期也不需要这一取舍，因为两个渲染臂已经使用相同的代码字体和固定行高。Shiki 现有 token 样式保持不变。
+
+**在激活时锁定测量得到的像素高度。** 当流式围栏增长或响应式换行变化时，固定测量会陈旧，并可能裁切内容或引入内部纵向滚动条。纯文本臂继续处于正常文档流中，不增加测量状态。
+
+**虚拟化完整代码块，或仅保留一个 token 窗口。** 这也能在读者访问所有块之后限制 DOM，但会改变文本选择、复制、滚动锚定与流式缓存的归属。视口激活消除不可见工作，同时不改变这些行为。
+
+## 后果
+
+从未进入视口的受支持代码不会发生内容 tokenize，也不会创建 token span。首次与视口相交时支付普通的同步高亮成本；若语法采用懒加载，代码块可能继续保持纯文本，直到既有的加载通知到达。已激活的块滚出视口后会保留高亮 DOM，因此内存占用随读者访问过的块增长，而不会随当前视口缩减。
+
+该优化仅作用于呈现层。Markdown 解析、Shiki 语法选择与样式、流式尾行 tokenize、复制文本和定稿输出均保持不变。
@@ -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 .agents/notes/implemented/bug-fix/2026-07-29-human-transcript-append-origin.md
-2026-07-29-human-transcript-append-origin.md: 72cafb5a1a149bcdb73d885d4e4fc4ac01e48cd2
-2026-07-29-human-transcript-append-origin.zh.md: 79fcc56b7b60082ebd2946d70419a659ff5687b7
+2026-07-29-human-transcript-append-origin.md: 00f4cf09139bdba4c0fffcdc476b698d098154bc
+2026-07-29-human-transcript-append-origin.zh.md: 631ba71d841033c19a9daa21e646468a0e366966
@@ -8,7 +8,7 @@ English | [中文](2026-07-29-human-transcript-append-origin.zh.md)

 The terminal and the host history gateway both treated the model-visible surface as the human transcript. A successful compaction replaces a surface range with one checkpoint node, so the moment that replacement landed the terminal dropped every message it shadowed — conversation the user had already read — and re-ran that destructive rebuild on any later replacement. The same confusion reached pagination: `maxMessages` counted every `user/message` and `assistant/message` in the window, so a model-only replacement copy consumed a page slot the human never filled, and the cut could land between a compaction's log-only `compaction/summary` event and the replacement that cites it.

-Nothing was lost from the log. `Session.events` still held every original message and full tool result; the surface only decides what the model is sent next. The defect was entirely in the projection.
+Nothing was lost from the log. `Session.snapshotEvents()` still returned every original message and full tool result; the surface only decides what the model is sent next. The defect was entirely in the projection.

 ## Decision

@@ -8,7 +8,7 @@ Status: implemented

 终端与宿主历史网关都把模型可见的 surface 当作 transcript（文本记录）。一次成功的压缩（compaction）会用一个检查点节点替换一段 surface 范围，因此该替换一落地，终端就丢弃了它所遮蔽的每条消息——那些是用户已经读过的对话——并在此后任何替换到来时重新执行这次破坏性重建。同样的混淆也波及分页：`maxMessages` 统计窗口内的每个 `user/message` 和 `assistant/message`，于是仅供模型使用的替换副本占用了一个人类从未填充的页面额度，而切分点还可能落在压缩的仅日志 `compaction/summary` 事件与引用它的替换之间。

-日志本身没有丢失任何内容。`Session.events` 仍保存着每条原始消息和完整的工具结果；surface 只决定接下来发送给模型的内容。缺陷完全在投影层。
+日志本身没有丢失任何内容。`Session.snapshotEvents()` 仍返回每条原始消息和完整的工具结果；surface 只决定接下来发送给模型的内容。缺陷完全在投影层。

 ## 决策

@@ -0,0 +1,6 @@
+# Bilingual-pair consistency record (docs/i18n/README.md): the git blob hash of each
+# 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 .agents/notes/implemented/bug-fix/2026-07-31-code-runtime-python-settlement-fixes.md
+2026-07-31-code-runtime-python-settlement-fixes.md: 26a5947b6602d56dc291f2f2e692743522580b12
+2026-07-31-code-runtime-python-settlement-fixes.zh.md: f2e7955a40b00f5b08017f44d2cf0529b2b31bfb
@@ -0,0 +1,151 @@
+# Agent Note: Settlement, framing, and lifecycle fixes in the CPython backend
+
+Status: implemented
+
+English | [中文](2026-07-31-code-runtime-python-settlement-fixes.zh.md)
+
+## Problem
+
+The CPython subprocess backend for PTC mode, built on the [fd-3 frame protocol](../architecture/2026-07-31-code-runtime-python-fd3-protocol.md), resolves every program outcome as a `CodeRunResult`, rejects `run()` only for seam misuse, and disposes to quiescence so no subprocess that stays in the child's process group outlives the fiber (a descendant that escapes the group with `setsid()` is the documented exception — see the package README's Known Limitations). A sequence of review passes surfaced defects that broke those contracts in ways unit coverage did not catch — each hid behind a `/* v8 ignore */`, a captured-callable that read as a fix but was not, a memory effect invisible through the seam, a load-time bound that double-counted, a process-group escalation that a survivor could outlast, a cross-event-loop completion that silently deadlocked, a synchronous throw outside the settlement path, or a transport boundary rendered as a log boundary. Most behavioral fixes ship with a test that fails without them; eleven do not, and are called out as such — the chunked frame read (a syscall-count improvement with no cross-platform-deterministic failure), the confirmed-empty finalize (its only seam-observable effect, a frozen heartbeat, freezes the instant SIGKILL is delivered, which the pre-fix finalize-on-delivery code also produced, and the discriminating probe is the signal-0 check the Alternatives reject as cross-environment-unreliable), the shared stdout/stderr budget (its only seam-observable difference is which entry boundary a mid-stream flush lands on, and that depends on the relative arrival timing of two independent OS pipes, which `os.sched_yield` does not make deterministic; the per-pipe memory bounds it strengthens ARE covered by the single-pipe flood tests), the `flush_line` join-clear-push reorder (it lowers the settlement-flush peak from three copies to two, but the 12x load gate already covers the three-copy newline path, so every gate-admitted config stays within the address space under both orders and no seam-observable difference exists — the memory effect is inside the Python child, unmeasurable through the seam like the shared-budget case), pacing binding replies (its in-tree case only asserts the framed replies still round-trip; the peak it removes lives inside the host's fd-3 writable buffer, unseen through the seam, so the 32.0 MiB → 0.0 MiB reduction is measurable only out-of-tree), dropping a late binding resolution before snapshot (its three assertions all also hold pre-fix, because `sendReply` already dropped after-settlement values — just later than the snapshot), the done-value TOCTOU pre-encoding (a concurrent mutation racing the encode cannot be deterministically constructed through the seam — its daemon-mutation regression only asserts the result is never a `worker-exit`, which is probabilistic and non-discriminating, so under the existing no-fail-before-with-a-reason precedent it is registered as no-fail-before), the stray-UTF-8 budget-flush retention (a budget flush landing exactly on a multibyte boundary is not schedulable through the seam; it is cross-referenced as v8-ignored), and the late-rejection settled guard (a rejection arriving after the run has already settled cannot be deterministically constructed from the seam), and the log-fragment seal (a 25 M single-character drip that would OOM is not deterministically constructible in CI; the in-tree case only asserts it completes and truncates), and the unknown-binding preview cap (the whole-target `JSON.stringify` peak is a transient allocation inside the reply path — its only seam-observable trace is peak memory under a forged near-ceiling `global`/`name`, not measurable through the seam; the in-tree case only asserts the run completes).
+
+## Decision
+
+Independent corrections, each in the package that owns the defect.
+
+### The unknown-binding preview is escaped from a 1 KiB prefix
+
+The unknown-binding reply built its message with `JSON.stringify` over the WHOLE capped target (`global` + `.` + `name`, each up to `maxValueBytes` code units) — the escaped form could reach ~6x the input under control-heavy fields, a multi-hundred-MB spike near the `maxValueBytes` ceiling that no hostile-peer bound would have admitted. The preview is now escaped from a 1 KiB prefix of the target (enough to identify the binding); `capMessage` still enforces the reply budget.
+
+### A merged open-log entry is billed once, split across its fragments
+
+An explicit `flush()` of an unterminated line emits a `log` frame with `open: true`, and the host appends the next frame to the SAME entry (`print('a', end='', flush=True); print('b')` reads back as one `'ab'` entry, not a fake newline). The split-billing arithmetic — first fragment pays quotes+content+separator, continuations and the closing frame pay content only, host caps `logBudget - 1`/`logBudget + 2`, the sub-2-byte walk guard, the child's `_open_started` keying — is stated once, in the [fd-3 protocol note's wire-contract section](../architecture/2026-07-31-code-runtime-python-fd3-protocol.md).
+
+### Boot-write failure no longer rejects run()
+
+In [`src/index.ts`](../../../../packages/experimental/code-runtime-python/src/index.ts) the fd-3 boot-frame write is the last statement of `run()`'s synchronous setup. Its `catch` calls `finish()`, and `finish()` reads `wallTimer` and `onAbort` and — through `settle()` — `live`. Those bindings are `const` and were declared AFTER the boot-write, so on a synchronous write failure `finish()` touched them in their temporal dead zone and threw a `ReferenceError`. That escaped the Promise executor and REJECTED `run()`, violating the seam's "outcomes resolve" contract: the caller saw a thrown error instead of the `worker-exit` the catch constructs. The boot-write block is now emitted after `wallTimer`, `onAbort`, and `live` are initialized, and the `/* v8 ignore */` that had hidden the branch from coverage is removed so the catch is measured.
+
+### Log capture is serialized against settlement
+
+In [`py/bootstrap.py`](../../../../packages/experimental/code-runtime-python/py/bootstrap.py) the settlement `flush_out()`/`flush_err()` on the main coroutine read and clear each stream's `_pending` list and mutate the shared `LogBuffer` ledger. Model code may start daemon threads whose `print`/`write` mutate the same state concurrently. Capturing the bound method (`out_stream.flush_line`) fixed only WHICH callable settlement invokes, not what it reads mid-flight: an interleaved flush could join a `_pending` list being mutated under it, corrupting the ledger and costing the `done` frame — stranding the run to the wall clock. `LogBuffer` now owns one re-entrant lock shared by both streams; `_LogStream.write` and `flush_line`, and `LogBuffer.push`, take it, so the whole read-modify-write is atomic across threads.
+
+### Fd-3 residual is copied, not viewed
+
+Also in `src/index.ts`, after the newline loop over a `Buffer.concat` of the pending fd-3 chunks, the leftover partial line was carried forward as the `subarray` VIEW it was sliced to. A view keeps the entire concat backing allocation alive, so a large frame followed by a tiny trailing fragment pinned a whole frame's worth of memory while `pendingBytes` — set to the fragment's length — reported far less than was retained. The residual is now detached into a fresh right-sized `Buffer` via the exported `detachResidual` helper, letting the concat allocation be collected and keeping `pendingBytes` an honest measure.
+
+### Output-cap load bound is parse-cap minus envelope, not divided by six
+
+The load-time check that rejects a `maxLogBytes`/`maxValueBytes` larger than one fd-3 frame can carry divided the frame ceiling by six for worst-case escape expansion. But both budgets are metered in ALREADY-ESCAPED serialized bytes — the host log ledger charges the serialized cost via `jsonStringCostUpTo` (which walks to the cap without allocating the escaped copy), `checkDoneValue` measures the escaped form, and the producing-side `_cap_message` also caps by serialized cost — so a payload admitted under the cap occupies at most `cap + envelope` on the wire; escaping is inside the charge and must not be multiplied in again. The bound is now `FRAME_PARSE_CAP_BYTES - FRAME_ENVELOPE_BYTES` (the receive path rejects raw frames past the 64 MiB parse cap before decoding — the run settles as a worker-exit — so a budget must not exceed what an honest child's frame can carry through that parser), and the unused `MAX_JSON_ESCAPE_EXPANSION` constant is gone. The old bound was not unsafe — it under-admitted — but it silently forbade legitimate large caps. The same load check also rejects a NON-INTEGER `maxLogBytes`/`maxValueBytes`: the child reads each budget through `int(...)`, which floors a float, so `maxLogBytes: 3.5` would truncate at 3 bytes child-side while the host meters the fraction — the two sides enforcing different public config. Rejecting the float at load keeps them in step, matching the worker backend.
+
+### Same-group survivors are reaped before the fiber goes quiescent
+
+A model program can leave a descendant in the child's OWN process group (no `setsid`, so `kill(-pid)` reaches it) that ignores SIGTERM but releases the inherited stdout/stderr/fd-3 pipes. The leader then exits, its `close` fires because the pipes drained, and settlement runs while that descendant is still alive. `kill()` arms an `unref`'d SIGKILL timer after SIGTERM; the fix is that `settle()` no longer resolves the run's `finished` promise — nor drops the run from `live` — immediately when an escalation is in flight. Instead, when `killing` is set and the process group is not yet empty (`process.kill(-pid, 0)` does not throw ESRCH), it polls the group on a REF'd timer, bounded by `graceMs + CLOSE_REAP_MARGIN_MS`, and both drops the run from `live` and resolves `finished` only once the group has emptied. The ref'd poll is the load-bearing part: it keeps the host event loop alive until the SIGKILL has actually reaped the group, so even a short-lived host — a one-shot headless run, a config subprocess — cannot exit and reparent the survivor to init. Deferring the `live` removal is what makes a `dispose()` racing a just-resolved `run()` still await the survivor: dropping the run from `live` at settlement (before the reap) would let teardown snapshot an empty set and return while the descendant lived. In the normal case (the leader was the only member) the first probe returns ESRCH and settlement finalizes with zero added latency. `teardown()` awaits each run's `finished`, so disposal is genuinely quiescent, matching its JSDoc — including for a run that already resolved.
+
+Settlement also CANCELS the SIGKILL timer the moment the group is confirmed empty (the normal path, and when the poll sees the survivor gone). Leaving it armed would expose a PID-reuse hazard: a `kill(-pid)` left pending for up to `graceMs` after the leader was reaped could hit a RECYCLED pgid once the kernel reused the leader's pid, SIGKILLing an unrelated group (`killGroup` swallowing ESRCH does not help — the danger is precisely the kill that SUCCEEDS against a reused group). Clearing it on the empty probe bounds the reuse window to only the genuine-survivor case, where the group cannot be empty to reuse.
+
+The window the cleared timer cannot cover is closed by an IDENTITY check inside `killGroup`. Every signal it sends is a raw `process.kill(-child.pid, sig)`, which — unlike `child.kill()` — has no handle guard, so it would reach a recycled pgid during the interval between the leader being reaped and `close` firing (measured at 3039 ms with a pipe-holding descendant). The leader's start time is therefore read once at spawn (`/proc/<pid>/stat` field 22) and re-read before each signal, with two rulings: a reading that is PRESENT AND DIFFERENT means the number now belongs to another process, so the signal is withheld; an ABSENT reading means the leader was already reaped, which is the ordinary case for every escalation — its `/proc` entry is gone while the group it led can still hold the survivor this teardown exists to reap — so the signal proceeds. Absent is also the constant reading on a platform with no `/proc`, where the guard is inert and the pre-existing behavior stands. Reading absent as a mismatch is not hypothetical: the first version did, which withheld the grace SIGKILL and the poll deadline's SIGKILL, and the three same-group heartbeat cases went red on the Linux coverage lane while passing on Darwin, where the reader always returns undefined.
+
+The reap poll also handles a host event loop BLOCKED past both timers. If a synchronous computation holds the loop from before the poll was scheduled until after its deadline, both the poll timer and the grace-window SIGKILL timer are overdue when the loop resumes, and Node runs the earlier-scheduled poll first — so the grace SIGKILL may never have fired. The deadline branch therefore sends SIGKILL ITSELF (idempotent if the timer already ran) rather than cancelling the unfired escalation, then grants ONE more `CLOSE_REAP_MARGIN_MS` and keeps polling until the group is confirmed empty, because finalizing on mere signal delivery would declare quiescence while the group is still dying. The outer bound on the wait is therefore `graceMs + 2 * CLOSE_REAP_MARGIN_MS`. A final hard bound finalizes if that extra margin elapses with the group still non-empty; that branch carries a `/* v8 ignore */` because it is reachable only where a SIGKILL'd survivor lingers as a zombie and is never `wait()`'d — a container whose PID 1 does not reap orphans — which cannot be built deterministically across CI platforms. The ignore's reason states that environment dependence rather than claiming the branch cannot run, cross-referencing the Alternatives entry that rejected the signal-0 reap assertion for the same reason.
+
+### RLIMIT clamps against the inherited soft limit, not only the hard
+
+In [`py/bootstrap.py`](../../../../packages/experimental/code-runtime-python/py/bootstrap.py) `_clamped` bounded a requested `(soft, hard)` rlimit pair by the inherited HARD limit alone. A deployment that inherited a soft limit below the requested one — say inherited `(100, 200)`, requested `(150, 160)` — got back `(150, 160)`, RAISING the effective soft from 100 to 150: for `RLIMIT_AS` that loosens the memory ceiling, for `RLIMIT_CPU` it defers SIGXCPU, both violating "strictest of configured and inherited". `_clamped` now clamps each side against its own inherited counterpart (`RLIM_INFINITY` imposing no ceiling), then pins soft under hard so `setrlimit` never sees an inverted pair. The settlement-time CPU recheck (`die_if_cpu_exhausted`) follows the same rule: it compares spent CPU against the EFFECTIVE clamped `cpu_soft`, not the configured `cpuSeconds`, so a program that traps SIGXCPU and burns past a stricter inherited soft before returning is reported as a timeout rather than a false success. The recheck restores SIG_DFL BEFORE unblocking a program-masked SIGXCPU (`pthread_sigmask(SIG_UNBLOCK, ...)`, captured at import): a program that installed a custom handler AND masked the signal would otherwise have that pending handler run at the unblock — in model code, able to re-mask or raise — so the disposition must already be SIG_DFL when the signal is released; with SIG_DFL first the pending signal kills inside the kernel with no bytecode window, and the `kill` re-raise is the fallback for the never-pending case. The SIGXCPU diagnostic no longer names the configured `cpuSeconds` as the effective budget — under a stricter inherited soft that number is wrong — and instead reports that CPU time was exhausted at "at most the configured N seconds", which holds whichever limit fired.
+
+### Concurrent binding replies are paced against fd 3
+
+`sendReply` ignored `proto.write`'s `false` return, so a program resolving several large values in one `asyncio.gather` round encoded every reply in the same turn and queued all of them in fd 3's writable buffer. Binding resolution carries no seam-level byte cap to bound that, and the failure kills the HOST process rather than failing the run: measured on a 64 KiB-highWaterMark pipe, eight 4 MiB replies buffered 32.0 MiB at once. Replies now go through a queue that encodes and writes one frame at a time, awaiting `drain` when the pipe is full, which measured a 0.0 MiB peak for the same shape. The encode happens inside the loop so a queued reply the run no longer needs is dropped by the `settled` check without ever being serialized. The same `settled` predicate also guards the reply callback AFTER `await fn(...)` but BEFORE `snapshotJsonValue`, so a wide value that resolves after settlement is dropped before its width is walked — the host does not expand a late value for a run whose outcome is already fixed.
+
+Pacing changes nothing the model can observe. The child matches each reply to its `call` by id from a pump that reads fd 3 continuously, so arrival order was never observable, and the bindings themselves still run concurrently -- only the host's peak memory and the flush timing change. That is also why serializing is not a narrowing of the seam's concurrency contract, which was the reason this was first deferred; that reasoning was wrong.
+
+### Binding replies complete on the calling loop's thread
+
+Also in `py/bootstrap.py`, a binding reply Future is created on the loop that ran `dispatch`. When the model calls a binding from a worker THREAD via `asyncio.run(tools.x(...))`, that Future belongs to the thread's loop, not the main loop where `_pump_replies` reads the reply. `asyncio.Future` is not thread-safe: completing it from another thread does not wake its own loop, so the direct `set_result`/`set_exception` left the awaiting thread stranded and the run degraded to a wall-clock timeout. Each pending entry now records its Future's loop alongside the Future, and `_pump_replies` completes it via that loop's `call_soon_threadsafe`. The shared `pending`/`next_id` state is guarded by a `threading.Lock` held across the id claim, the fd-3 write, and the counter advance, so concurrent callers cannot interleave frames out of the id order the host requires. `call_soon_threadsafe` onto a loop that has already CLOSED (the worker thread finished and abandoned its call before the reply arrived) raises `RuntimeError`; that schedule is wrapped so the moot reply is dropped rather than letting the exception end the pump task and strand every later reply.
+
+### The blocking frame reader reads in chunks, not byte by byte
+
+`ProtocolChannel.read_frame` — used for the `boot` and `run` handshake frames — read through `FileIO.readline()` on the unbuffered (`buffering=0`) fd, which issues one `os.read(1)` per byte. The `run` frame arrives AFTER `RLIMIT_CPU` is in force, so a legitimate multi-megabyte program burned seconds of CPU in millions of single-byte syscalls before `ast.parse` ran — potentially exhausting the budget on the read alone. It now reads in `_READ_CHUNK_BYTES` chunks into the same `_pending` residual buffer the async reader already uses (the wrapping `os.fdopen` object is gone; both readers call `os.read(self._fd, ...)` directly), so the read cost is trivial and read-ahead past a newline is preserved for the next frame. Both readers track a running scan offset (`find(b"\n", scanned)`) so a large frame accumulated across many chunks is scanned once, not re-scanned from index 0 per chunk — a chunked rescan would have replaced the byte-at-a-time cost with an O(N²) memchr cost on the same large-frame path.
+
+### Synchronous spawn failure resolves worker-exit, not reject
+
+Also in `src/index.ts`, `spawn` is called before the settlement Promise executor exists. Node defers only a fixed set of spawn errnos (EACCES, EAGAIN, EMFILE, ENFILE, ENOENT) to an asynchronous `error` event, which the settlement path already turns into a `worker-exit`; every other errno throws SYNCHRONOUSLY from `spawn`. A `pythonBin` longer than the platform PATH_MAX passes the load-time validation (non-empty, no NUL) but makes `spawn` throw `ENAMETOOLONG` here — outside the executor — so `run()` REJECTED instead of resolving, violating resolve-don't-reject, and left this run's just-materialized staging directory on disk since only `settle()` removes it. The `spawn` call and the fd-3 narrowing are now wrapped: a synchronous throw removes the staging directory and resolves the same `worker-exit` class (`python spawn error: …`) the async `error` event produces.
+
+### Interpreter selection and the child environment settle at load
+
+`pythonBin` resolves once at plugin load to an executable absolute path and is version-probed under the same scrubbed environment used for runs. The provider requires CPython 3.10 or newer and retains that exact path, so a later `PATH` or working-directory change cannot switch interpreters; an explicit path that is not an executable regular file, an unresolved basename, or an unsupported interpreter fails before `ctx.codeRuntime` registers. The synchronous probe has a fixed five-second deadline and sends `SIGKILL` at that deadline, so a wrapper that ignores `SIGTERM` cannot block plugin load. Each probe and run receives only `TMPDIR`: macOS system Python needs it to avoid emitting a startup warning into captured stderr, while credentials, `PATH`, `HOME`, and every other ambient host value remain unavailable to model code. If the validated executable disappears after activation, the ordinary spawn settlement still resolves `worker-exit`.
+
+### Stray pipe output is aggregated by line, not by transport chunk
+
+Also in `src/index.ts`, native stdout/stderr bytes (C-extension writes, `os.write` past the pipe buffer) were pushed to `logs` one entry per Node `data` chunk. `logs` entries are joined with `\n` downstream (PTC mode), so a single newline-free write larger than one pipe read — arriving as several `data` chunks — read back with model-visible newlines inserted at arbitrary transport boundaries. Capture now accumulates raw `Buffer` chunks (the same shape as the fd-3 reader, and for the same reasons: a string `+=` accumulator re-copies the whole residual per chunk and scanning it from index 0 each chunk is a second quadratic — both O(N²) on a large newline-free write), splits on the raw `0x0a` byte, and admits one entry per complete line. A newline never appears inside a UTF-8 multibyte sequence, so decoding each split line is safe without a streaming decoder. Three separate bounds keep the residual from exhausting host memory, each mirroring the fd-3 reader: the fragment list SEALS into finished blocks past `MAX_PENDING_CHUNKS` so a program pacing single-byte `os.write`s cannot accumulate millions of live Buffer objects (whose per-object overhead no byte count sees); the residual is flushed when the COMBINED running SERIALIZED cost of both pipes — tracked through `accrueStrayCost`, which decodes UTF-8 structurally across chunks so a byte that renders as U+FFFD is charged the three bytes that replacement character serializes to — would cross the budget, so a control-char or illegal-UTF-8 flood flushes at a fraction of the raw bytes rather than accumulating a full budget's worth of raw bytes first, and stdout and stderr are metered together rather than each against the full budget (which would let both retain nearly a budget's worth at once, doubling the peak); and once the ledger has truncated, buffering stops so nothing accumulates for output that can never be admitted. `accrueStrayCost` charging illegal bytes their U+FFFD width is the fix for a byte that never begins a valid sequence (0x80–0xC1, 0xF5–0xFF), a multibyte sequence that breaks before completing, or a structurally-complete but ILLEGAL sequence: `toString('utf8')` renders each of those bytes as its own U+FFFD (3 bytes), so it validates each lead's first-continuation range (WHATWG: `E0`→A0-BF, `ED`→80-9F, `F0`→90-BF, `F4`→80-8F, others 80-BF) and charges 3 per byte of any sequence outside it. Charging the raw 1 undercounted a `b"\xff"` flood threefold, and charging only the structural width undercounted a CESU-8 surrogate (`ED A0 80`) or overlong (`E0 80 80`) threefold just as cheaply, letting the residual grow to a full budget's worth of raw bytes before flushing and, near a large `maxLogBytes`, expand toward a ~1 GiB peak in the flush's concat plus `toString`. The per-entry charge on the admitted string is metered by SERIALIZED cost through `jsonStringCostUpTo`, which walks the string to the cap and stops — the previous `Buffer.byteLength(JSON.stringify(text))` allocated the whole escaped form first, so a near-budget control-char-dense line under a large `maxLogBytes` could momentarily allocate over a gigabyte just to measure it. `jsonStringCostUpTo` (the string-walking function, reached by a forged `log` frame whose text `JSON.parse` produced) charges a LONE surrogate the full six escaped bytes (`\uXXXX` under ES2019 well-formed `JSON.stringify`), not the three bytes `Buffer.byteLength` reports for its U+FFFD rendering, so a `\ud800` flood is not undercharged by half; `accrueStrayCost` walks raw bytes and never sees a surrogate as such — a CESU-8-encoded surrogate reaches it as three bytes its per-lead range check rejects, each charged 3 (total 9), matching what `toString('utf8')` renders. The residual is flushed on the pipe's `end` and also explicitly in the `closeDeadline` handler before it destroys the streams: a `setsid` escapee holding the pipes open forces settlement through that path without an `end`, so a final newline-free `os.write(1, …)` the leader emitted before exiting would otherwise be dropped from `logs`.
+
+### An incompatible output-budget/addressSpaceMb pair is rejected at load
+
+The child ([`py/bootstrap.py`](../../../../packages/experimental/code-runtime-python/py/bootstrap.py)) builds, charges, and frames a `maxLogBytes` log entry or a `maxValueBytes` completion value under `RLIMIT_AS`, and both ledgers trigger on CHARACTER count against a serialized-BYTE budget. An astral character is one character but four bytes of CPython `str` storage and four UTF-8 bytes, and the heaviest path holds THREE such copies at once: a single `sys.stdout.write(line + "\n")` keeps the caller's `text` argument (alive for the whole `write` call, ~4×), the line slice handed to `LogBuffer.push` (~4×), and the `text.encode("utf-8")` copy `_push_locked` takes to charge and ship it (~4×) — a peak of ~12× the budget. The settlement `flush_line` path holds only two (its `"".join(...)` and that encode copy — it drops the pending chunks before pushing), so the newline path is the binding worst case. When a budget approaches `addressSpaceMb`, a LEGITIMATE near-budget output breaches the address space during that build-and-encode and dies as `worker-exit` instead of truncating (log) or failing as `output-limit` (value). Metering every child write against the address space at runtime is the wrong fix: an exact serialized-cost check on the hot path is either a full `encode` (the allocation being avoided) or a per-character Python loop (which burns the CPU budget — a 10 MB legitimate write hits SIGXCPU under `cpuSeconds: 1`). Both trade one resource bound for another. Instead [`src/index.ts`](../../../../packages/experimental/code-runtime-python/src/index.ts) rejects the incompatible pair at LOAD: each budget times `OUTPUT_BUDGET_WORST_CASE_ADDRESS_SPACE_MULTIPLE` (twelve — the three simultaneous ~4× copies of the newline path) must fit the address space LEFT after a fixed `INTERPRETER_BASELINE_BYTES` reservation for the interpreter's own footprint, with a `>=` so a budget whose worst-case peak exactly equals that room is rejected (that peak plus the reserved baseline is the whole address space, the `RLIMIT_AS` edge). `flush_line` was also made to drop the pending chunks BEFORE its push, matching the newline path's join-clear-push order, so it holds at most the join and its encode rather than three copies. The baseline is reserved SEPARATELY from the multiple because it is a fixed cost, not one that scales with the budget: folding it into the multiple would leave a budget sized right at `addressSpaceMb / 12` admitted while its peak plus the interpreter still overran. Both `maxLogBytes` and `maxValueBytes` are gated symmetrically; the value path's build-and-encode is the same shape. The check runs on every platform, not just where `RLIMIT_AS` is enforced: the incompatibility is a property of the config values, so a Linux deployment OOMs regardless of the host that assembled the config, and a uniform load-time rejection is the fail-loud contract (Darwin skips only the runtime `setrlimit`). This eliminates the class at the config seam rather than patching the write path, so `_LogStream` keeps its original character-count buffering (a valid lower bound on serialized cost, memory-safe once the budget fits the address space). The value path enforces the same discipline in a second place: `_check_done_value` (the byte meter) and `_encode_json_plain` (the frame encoder) walk in O(DEPTH), not O(width). Each container pushes ONE cursor frame that pulls its children one at a time rather than one traversal tuple or stack entry per child — a flat `[0] * 6_000_000` serializes to ~12 MB but a per-element walk allocates ~400 MB of bookkeeping (~28× the serialized size, far past the 12× the gate reserves), so a value the meter admits could OOM on the walk's own frames. With the cursor, the only width-proportional allocation is the output string the meter already bounded.
+
+The host gate validates against the CONFIGURED `addressSpaceMb`, but a launch environment can inherit a STRICTER `RLIMIT_AS` (a `ulimit -v` wrapper below `addressSpaceMb`), which the bootstrap's `_clamped` correctly lowers the EFFECTIVE limit to — leaving the budgets sized for a ceiling the child never gets. So `bootstrap.py` re-checks both budgets against the effective clamped soft limit after applying it, mirroring the host gate's multiple and baseline, and raises at boot (caught by the setrlimit-phase handler and reported as `exception`, the same class as any other resource-limit-application failure) rather than letting a near-budget output OOM mid-run. The two child constants are kept in step with the host's by the shared reasoning, not a wire field.
+
+One residual write-path copy is fixed alongside, independent of the config gate: `_LogStream.write`'s newline branch buffered the whole unterminated tail after the last newline (`text[pos:]`) into `_pending` before the flush trigger could bound it, so an early newline followed by a huge tail (`"\n" + "A" * 30 MiB`) made a second full copy of the model's own string — the `RLIMIT_AS` death the path exists to avoid, and one the config gate does not cover because the tail can far exceed `maxLogBytes`. The tail is now sliced to a `remaining + 4`-character prefix (anything past `remaining` characters cannot be admitted, the char count being a lower bound on the serialized cost), which the flush trigger then rejects with the marker.
+
+### The completion value and error are pre-encoded at their validation point
+
+In [`py/bootstrap.py`](../../../../packages/experimental/code-runtime-python/py/bootstrap.py), `_done_with_value` now returns the whole terminal frame as a PRE-ENCODED JSON string on the success path: the admitted value is serialized once here, at the validation point inside `_run`'s `try`, as `'{"type": "done", "value": ' + _encode_json_plain(value) + "}"`. The program can keep mutating a returned list/dict from a daemon thread or signal handler after it returns, so a second traversal held at a later point would be a TOCTOU — a concurrent mutation into a non-JSON type would let that later encode throw outside the settlement handler and downgrade a settled run host-side to `worker-exit`. Serializing once, inside the `try` that wraps this call, closes the window: if a concurrent mutation makes the encode throw, the exception handler classifies it as an `exception`, and once the string is produced the frame is written verbatim with no further touching of the live value. `_run` binds the `_done_with_value` ENTRY NAME into a local (`done_with_value_bound`) before the program runs, and `_done_with_value` itself binds `_check_done_value` and `_encode_json_plain` as DEF-TIME default arguments — so a `__main__` rebind of the entry name or those two names after model execution cannot rewrite a legitimate success into an `exception`. The log ledgers (host `logBudget` and child `_remaining`) start ONE byte below the budget, reserving the serialized outer-array envelope (two brackets and n-1 commas over n entries' separators), so a result that exactly exhausts the ledger still serializes within the configured cap; the truncation marker remains envelope, not payload. Once a ledger has truncated, the host clears both stray pipes' buffered output wholesale (every later byte would be no-op'd by `admit`, so retaining it would spend host memory on output that can never be admitted); the child runs `-u` so `sys.__stdout__`/`sys.__stderr__` writes are visible to stray capture immediately, and the settlement flush still drains the original std streams before the done frame (a guard against a buffered wrapper surviving a `sys.__stdout__ = boom` rebind). The constructor rejects a `maxLogBytes` below 64 (the smallest budget with one byte of room for the truncation marker's own serialized form); `maxValueBytes` keeps only the positive-integer requirement, since a completion can be a single byte and the done-frame envelope is seam protocol cost. The marker remains envelope, so a truncated run with admitted entries serializes to at most `maxLogBytes + marker + envelope` (recorded in the package README). A rebind of a transitive dep the encoder reaches (e.g. `_dump_scalar`/`_dump_string`/`json`/`io` — a non-exhaustive set) can still make the encode throw and downgrade a success to an `exception`, which is registered as an accepted residual in the package README.
+
+`send_done` (a local function inside `_run`) writes the pre-encoded string through a BOUND `channel.write_encoded`, and encodes a dict error frame through a bound `_encode_json_plain` before writing it — it never calls `channel.send_sync`, whose body re-resolves `self.write_encoded` and the module-level `_encode_json_plain` at call time. `_encode_json_plain` and `channel.write_encoded` are bound into locals before the program runs, for the same reason `flush_out`/`flush_err`/`safe_model_traceback` are: the program runs as `__main__`, so `import __main__; __main__.ProtocolChannel.send_sync = boom` or `__main__._encode_json_plain = boom` would otherwise re-resolve the send/encode to a rebranded callable at call time and, when that replacement raises, skip the `done` frame and downgrade a settled verdict to a host-side `worker-exit`.
+
+### A budget flush retains an unfinished trailing multibyte sequence
+
+Also in [`src/index.ts`](../../../../packages/experimental/code-runtime-python/src/index.ts), `flushStray(stray, retainPartialTail)` withholds an unfinished multibyte tail from the decode on the BUDGET-triggered flush (the combined-cost threshold in `captureStray`): when the residual ends on a partial UTF-8 lead sequence (`stray.utf8.expected > 0`), the leading byte plus the continuations consumed so far (≤3 bytes) are detached from the frame as the new residual, and only the complete prefix is admitted and decoded. Nothing is admitted when the whole residual is a single unfinished sequence, so a legal, un-finished character is never rendered as U+FFFD in a released, un-truncated entry, and no bogus empty entry is pushed. The withheld tail is re-accrued from a FRESH `stray.utf8` state — metering it against the post-flush `expected > 0` state would charge the carried lead byte as an illegal continuation — so the next chunk continues the walk correctly and the pipe's cost/UTF-8 state is rebuilt over the retained tail. The `end`/`closeDeadline` paths pass `false` and decode the FULL residual unchanged, because there a trailing incomplete sequence is real truncated input and the U+FFFD is the honest render.
+
+### A late binding rejection returns before formatting the error
+
+Also in `src/index.ts`, the binding-rejection catch branch now checks `settled` and returns BEFORE formatting `messageOf(error)`. A rejection that arrives after `maxWallMs`, an abort, or dispose has already settled the run would otherwise have `messageOf(error)` run hostile `toString`/`message` getters — spending host heap and time on a run whose outcome is already fixed — before `sendReply` peeks at `settled`. Dropping the framed reply early spares that waste. The running loop's otherwise-mostly-linear reply drain also reads by a head cursor into the queue array instead of `shift()`ing each entry, so a large `asyncio.gather` of wide bindings awaiting fd 3's `drain` drains in linear time rather than O(n²) from repeated re-slicing.
+
+### A newline-free drip seals its fragments; the CPU soft limit is kept below the hard; the done frame falls back to a fixed literal; the reply queue clears consumed slots
+
+In [`py/bootstrap.py`](../../../../packages/experimental/code-runtime-python/py/bootstrap.py), `_LogStream` now seals the pending-fragment list past a cap: a newline-free drip of one character per `write` would otherwise accumulate one list slot (and one str object) per call, and under a large `maxLogBytes` a 25 M single-character flood OOMs on its own accounting (plus the same-size list `_push_bounded_prefix` then builds) before the byte budget is reached. Past `_PENDING_MAX_CHUNKS` the current fragments are joined into ONE block moved to a `_pending_blocks` list (the character count is unchanged), bounding the live fragment count exactly as the host-side `captureStray` seal does; the join is only the ≤cap current fragments, never the whole accumulated buffer, so a large drip stays O(B) rather than re-copying the growing block O(B²/cap) times. The HOST-side open hold mirrors the same seal: a budget-sized single-character open flood (`print('x', end='', flush=True)` in a loop is an honest-child path) would otherwise accumulate one fragment array slot plus string object header per frame — ~30× overhead the byte cap cannot see, up to ~2 GB of host auxiliary heap near the `maxLogBytes` load ceiling. Past `MAX_PENDING_CHUNKS` the held fragments coalesce into `openSealed`; the closing-frame merge, `truncateLogs`, and the `finish` residual all read sealed + current fragments and clear the seal.
+
+`_clamped` also lowers a clamped RLIMIT_CPU soft limit that EQUALS the hard by one unit (when the hard is at least 2). A `ulimit -t N` sets both, and with soft == hard the kernel checks the hard limit in the same tick and SIGKILLs a busy loop directly, so SIGXCPU is never delivered — and the host classifies a CPU overrun ONLY on `signal === 'SIGXCPU'`, so a definite budget exhaustion would be misreported as a `worker-exit`. Lowering the soft one unit gives SIGXCPU a window to fire, so the overrun is reported as a timeout. This is scoped to RLIMIT_CPU (a one-byte soft differential on RLIMIT_AS would only misalign the child's applied limit with the host budget gate, with no signal to preserve). The `hard >= 2` guard leaves a `hard == 1` blind spot — a 1-second dual limit cannot lower the soft to 0, so a definite overrun there is still reported as `worker-exit`.
+
+`send_done` wraps its encode+write in a try and, on any throw from a rebound transitive name (`_dump_scalar`/`os`), writes a fixed pre-encoded done frame via the `_run`-local bound `_os_write`/`_memoryview`/`_FALLBACK_DONE_FRAME` — so a settled `exception` verdict is never downgraded to a `worker-exit`, and the host still gets a verdict. The reply queue's head-cursor drain clears each consumed slot so a wide written payload is released immediately, bounding host memory to the current backlog under sustained fd-3 backpressure. The exception classes the settlement-path `except` clauses catch are likewise bound before any model code runs: `_BaseException` is a `_run` LOCAL and a closure cell in `_make_failure_reporter`; `_RuntimeError`, `_BindingRejection`, `str`, and `bool` are DEF-TIME default arguments of `_pump_replies` (a body-local `X = X` binding is too late — the model's top-level statements run before the pump's first step). A rebind of `__main__.BaseException` cannot make a program exception escape the handler and lose the `done` frame; a rebind of `__main__.RuntimeError` (or `_BindingRejection`/`str`/`bool`) cannot make a closed-loop scheduling failure escape the pump catch and strand every later reply to the wall clock.
+
+## Testing
+
+- `tests/runtime.spec.ts` rejects absent, non-executable, non-CPython, pre-3.10, and unresponsive interpreter configurations at load; changes `PATH` after activation to prove the resolved executable is frozen; removes that executable after activation to preserve the late `worker-exit` path; and asserts a running program sees `TMPDIR` but not `PATH`, `HOME`, or `DEEPSEEK_API_KEY`. The native-output case pins each source stream's order without requiring a total order across independent channels, and the Darwin resource-limit cases state or skip the platform-specific `RLIMIT_AS` behavior.
+- `snapshots/session/ptc-python-turn` replaces the headless PTC worker provider with the private Python provider through the real Loader, replays a Python `run_code` program over real bash bindings, and pins the Python SDK prompt, tool schema, dispatch events, captured log, and completion value.
+- `tests/boot-write-failure.spec.ts` mocks `spawn` so the fd-3 pipe throws on the boot write — the one path a real subprocess cannot be coerced into — and asserts `run()` resolves a `worker-exit` rather than rejecting. A sibling case makes the mocked `spawn` throw SYNCHRONOUSLY and asserts `run()` still resolves a `worker-exit` and removes its staging directory, keyed off the exact bootstrap path the mocked `spawn` received in its argv so a sibling worker's concurrent staging cannot flake it. Both are isolated in this spec so the real-subprocess suite is untouched.
+- `tests/residual-detach.spec.ts` unit-tests `detachResidual`: the carried copy equals the residual, owns a backing store sized to its own length (fixture kept above Node's Buffer pool threshold), and does not share the source frame's `ArrayBuffer`.
+- `tests/runtime.spec.ts` — the output-cap case asserts the `parse-cap - envelope` bound (67108800) and its message. A daemon-thread case drives four threads emitting unterminated writes through settlement's flush. A native-write case writes 200 KiB with no newline via `os.write` under a raised `maxLogBytes` and asserts it reads back as EXACTLY one log entry (proving stray output is aggregated by line, not split at pipe-chunk boundaries); a companion writes `b"one\ntwo\nthree"` and asserts three entries (proving real newlines still delimit). A newline-free-flood case writes 2 MiB under a 4 KiB `maxLogBytes` and asserts the capture ends at the truncation marker and stays under budget (proving the residual is bounded by the ledger, not buffered whole); a NUL-flood companion writes 4000 newline-free NULs under the same budget and asserts truncation (proving the residual is charged by SERIALIZED cost, ~6× raw, measured without allocating the escaped copy); an illegal-UTF-8 case paces single-byte `\xff` writes under a 3072-byte budget with `Buffer.concat` wrapped to measure the peak merged buffer, asserting it stays under 2048 (charged at the U+FFFD width 3 the residual flushes near 1024 raw bytes; a raw-byte undercount would let it reach ~3072, so the bound discriminates); a CESU-8/overlong case paces the structurally-well-formed but illegal `ED A0 80` one byte at a time and asserts the same peak bound (charged at the true 9 per sequence it flushes early; charging the structural width 3 triples the peak, so reverting the per-lead range check turns it red); a broken-multibyte case writes a 3-byte lead then a fresh ASCII byte in separate chunks and asserts both a captured `A` and a U+FFFD (exercising `accrueStrayCost`'s cross-chunk broken-sequence branch); a post-truncation case writes a 108-byte payload (under the smallest PIPE_BUF, so one atomic write) whose first line exhausts a 64-byte budget and asserts the second line is dropped (exercising the post-truncation admit no-op in one `data` callback, no v8-ignore); a short-escape case writes a line mixing a tab, quote, backslash, a `\uXXXX` control, a multibyte character, and ASCII, asserting it round-trips verbatim (exercising every branch of `jsonStringCostUpTo`); a reassembly case writes a payload spanning every valid multibyte lead class (E0-range, plain 3-byte, F0, and F4) past the pipe buffer and asserts it round-trips with no U+FFFD (exercising `accrueStrayCost`'s per-lead ranges and cross-chunk reassembly); a lone-surrogate case forges an fd-3 `log` frame flooding 1000 `\ud800` escapes under a 4 KiB budget and asserts truncation (the count sits in the window where charging 3 bytes would admit and 6 bytes truncates, proving the surrogate is charged its full escaped width); a stray-sealing case paces 60000 single-byte newline-free `os.write(1, …)` calls under a raised budget with `Buffer.concat` wrapped to measure copy volume, asserting the trickle coalesces to one entry and the cumulative copy stays under a measured 256 KiB threshold (the sealed shape copies ~120 KB, the re-merge shape ~538 KB, so reverting the seal to a re-merge turns the assertion red — proving the fragment list seals into blocks past `MAX_PENDING_CHUNKS`). A closeDeadline-flush case has the leader write a newline-free diagnostic then spawn a `setsid` orphan holding the pipes open, and asserts the diagnostic survives in `logs` (proving the residual is flushed before the deadline destroys the streams). The same-group reap case spawns a SIGTERM-ignoring same-group descendant that releases the pipes and bumps a heartbeat file; the test asserts the heartbeat STOPS after the grace-window SIGKILL — an assertion robust whether the killed descendant is reaped or lingers as a zombie, so it holds where PID 1 does not wait() orphans. A dispose-after-resolve case asserts `dispose()` of a completed run with a same-group survivor returns only after the survivor stops executing (proving the run stays in `live` until its group is reaped), with an `expect(afterDispose).toBeGreaterThan(0)` guard so the frozen-heartbeat assertion cannot pass vacuously when the file was never written. A deadline case busy-blocks the event loop past both timers and asserts the survivor's heartbeat freezes (proving the poll's deadline arm sends SIGKILL itself rather than cancelling the unfired escalation). The cross-loop case runs a binding from a worker thread's own `asyncio.run` loop while the main coroutine yields with `await asyncio.sleep`, asserting the reply round-trips instead of timing out; a companion case abandons a thread's call so its loop closes, then answers it before a later binding — asserting the pump survives the closed-loop `call_soon_threadsafe` (host-gated ordering makes it deterministic, fail-before hangs the later binding to the wall clock). The inherited-soft-limit case runs the interpreter through a `ulimit -S -t` wrapper that sets a CPU soft limit below `cpuSeconds` and asserts the appliLine truncated
+
+## Alternatives considered
+
+**Leave the boot-write `/* v8 ignore */` and fix only the ordering.** Rejected: the ignore is what let the TDZ regression ship uncaught. Removing it makes the catch a measured branch, so per-file 100% coverage now proves the failure path is exercised.
+
+**Fix the flush race by capturing more bound methods.** Rejected: this is the approach that already failed. Binding a callable fixes reference resolution, not concurrent access to the mutable state the callable reads. Only mutual exclusion over the shared ledger closes the race.
+
+**Guard the residual with a size threshold (copy only large frames).** Rejected: the branch runs once per newline-bearing read, the copy is bounded by the residual's own length (always a partial line), and a threshold adds a tunable and a second code path for no measurable saving. An unconditional right-sized copy is simpler and always correct.
+
+**Assert the residual memory effect through the seam.** Rejected: the retained allocation is not observable through `CodeRunResult`, so a black-box test could not distinguish fixed from unfixed. Extracting `detachResidual` makes the backing-store invariant a deterministic unit test instead.
+
+**Reap the same-group survivor with a fire-and-forget `unref`'d SIGKILL timer alone.** Rejected: an `unref`'d timer does not keep the host alive, so a host that exits within the grace window (a one-shot run, a config subprocess) never fires the SIGKILL and the survivor is reparented to init — the same "no subprocess outlives the fiber" violation in a different shape, and `teardown`'s "await each child's exit" JSDoc would be false. Awaiting the group's death on a ref'd poll keeps the host alive exactly long enough to reap, at zero cost in the common empty-group case.
+
+**Assert the reap with `process.kill(pid, 0)` throwing ESRCH.** Rejected: a SIGKILL'd process lingers as a zombie until its parent `wait()`s it, and in a container whose PID 1 does not reap orphans the signal-0 probe keeps succeeding, so the assertion would false-fail cross-environment. A heartbeat file that stops advancing detects "no longer executing," which a reaped process and a zombie both satisfy.
+
+**Complete the cross-loop Future with a plain `set_result` and rely on the GIL.** Rejected: the GIL serializes bytecode but does not make `asyncio.Future` cross-loop-safe — completing a Future from a thread other than its loop's does not schedule its callbacks or wake the loop. `call_soon_threadsafe` on the owning loop is the documented mechanism.
+
+**Leave the SIGKILL timer armed after settlement (the earlier same-group fix).** Rejected: an unref'd timer left to fire up to `graceMs` after the leader was reaped can `kill(-pid)` a RECYCLED pgid, striking an unrelated group; the danger is the kill that succeeds, which `killGroup`'s ESRCH swallow cannot prevent. Clearing the timer once the group is confirmed empty bounds the reuse window to the genuine-survivor case, where the group is not empty to reuse.
+
+**Clamp rlimits by the inherited hard limit only.** Rejected: that silently RAISES an inherited soft limit stricter than the request, loosening the very containment the clamp exists to preserve. Clamping each side against its own inherited bound (then pinning soft under hard) keeps the strictest of configured and inherited on both.
+
+**Bill the host-side `capMessage` backstop by serialized cost, matching the child's `_cap_message`.** Rejected: the two caps guard different things. `_cap_message`'s output re-crosses fd 3 as a JSON string, so its escaped width is what the frame ceiling bounds — serialized billing is required there. `capMessage`'s output goes straight into `CodeRunResult.error.message` and never re-crosses a frame-bounded channel, so the honest measure of what it retains is the raw byte length of the model-visible string. An honest child has already capped by serialized cost and raw length ≤ serialized cost, so a well-formed message passes unchanged; a forged control-heavy message could serialize to ~6× its raw length, but since it travels no capped channel, billing it by that inflated wire width would truncate a legitimately-sized diagnostic for no containment gain. Each side's JSDoc documents the split and points at the other.
+
+**Push stray pipe output one entry per `data` chunk.** Rejected: `logs` entries are joined with `\n` downstream, so a transport chunk boundary would become a model-visible newline — a single native write split across pipe reads would read back with spurious line breaks. Aggregating by real newline (raw-chunk buffer + split on `0x0a`) matches the child's line-granular `log` frames; the ledger still bounds a newline-free flood by admitting-and-truncating the residual when it would cross the budget.
+
+**Enforce the fd-3 frame ceiling per-frame (split before the counter check) to avoid a batch-edge false reject.** Rejected: the ceiling check reads the byte counter BEFORE any `Buffer.concat`, precisely so a hostile program cannot force ~2× the 64 MiB frame cap of host memory (the counter and the join are a second copy of everything held). Splitting first to bill a single frame would `Buffer.concat` an over-ceiling frame before rejecting it, reintroducing that doubling — two regression tests assert the pre-concat order for exactly this reason. The batch-edge false reject the per-frame order would fix (a legitimate near-cap frame whose newline-bearing chunk also carries the next frame's leading bytes nudging the counter over the ceiling for one pipe read) is reachable only when `maxLogBytes`/`maxValueBytes` is configured within one pipe read of the 64 MiB cap — orders of magnitude past the 32/64 KiB defaults. The memory-safety bound against hostile input at any config takes precedence over a false reject reachable only at a pathological near-ceiling config; the counter's over-count and this trade-off are documented at the check.
+
+**Flush the two stray pipes in residual-arrival order when the combined budget crosses.** Rejected: stdout and stderr are independent OS streams whose `data` events already interleave nondeterministically with each other and with the child's own fd-3 `log` frames. The seam's `CodeRunResult.logs` JSDoc reads "in order", which the surrounding text scopes to program-emission order WITHIN a stream — ordering ACROSS concurrent streams is inherently best-effort here, since no host-side flush order can reconstruct the true interleaving the kernel already lost, so preserving a residual's arrival order at the flush buys nothing. A fixed drain order is as valid as any. Tracking a per-residual arrival tick to drain the earlier pipe first would add a branch whose two sides fire only on the relative timing of two OS pipes, which `os.sched_yield` does not make deterministic, so the branch could not be covered without a flaky test — cost with no observable contract benefit.
+
+**Meter the child log ledger against the address space at runtime instead of rejecting the config at load.** Rejected: an exact serialized-cost check on every child write is either a full `encode` — the very allocation an oversized write cannot afford, which the ledger's cheap pre-check exists to avoid — or a per-character Python loop, which burns the CPU budget (a 10 MB legitimate write hits SIGXCPU under `cpuSeconds: 1`). Each runtime approach trades the memory bound for another resource bound on the hot path. The address-space breach is a property of the `maxLogBytes`/`addressSpaceMb` pair, not of any particular write, so rejecting the incompatible pair once at load eliminates the whole class without any per-write cost and keeps `_LogStream`'s original character-count buffering, which is memory-safe once the budget fits the address space.
+
+## Consequences
+
+Interpreter misconfiguration fails before the service is published, every run uses the executable selected at load, and the child receives only `TMPDIR`, removing macOS startup noise without exposing host credentials. The private provider remains absent from shipped profiles while a keyless Loader snapshot pins its source-checkout PTC composition.
+
+The seam's resolve-don't-reject contract holds on the boot-write path and the synchronous-spawn-failure path, both with measured coverage, and neither strands a staging directory. Log capture is thread-safe at the cost of one re-entrant lock acquisition per write and flush, and stray native output is delimited by its own newlines rather than by transport chunks. Fd-3 residual memory is bounded by the actual retained bytes, and both frame readers scan an accumulating frame once rather than quadratically. The output caps admit every value a frame can carry and reject a non-integer budget at load. Disposal is genuinely quiescent against a same-group survivor — bounded by `graceMs + 2 * CLOSE_REAP_MARGIN_MS`, zero-cost when the group is already empty, with the SIGKILL timer cleared once the group empties so a stale kill cannot strike a recycled pgid — RLIMIT enforcement keeps the strictest of configured and inherited on both soft and hard (and the SIGXCPU diagnostic no longer names a budget the host cannot guarantee), bindings called from model-created threads complete instead of timing out, and the handshake frame reader no longer burns the CPU budget on a large program. Every behavioral fix carries a test that fails without it, except the eleven called out in the Problem section — the chunked frame read (a syscall-count improvement), the confirmed-empty finalize (whose only seam-observable effect freezes at signal delivery, which the pre-fix code also produced), the shared stdout/stderr budget (whose only seam-observable difference turns on nondeterministic cross-pipe arrival timing), the `flush_line` reorder (whose lowered peak stays within what the 12x gate already admits, so no config behaves differently), pacing binding replies (the 32.0 MiB → 0.0 MiB peak reduction lives inside the host's fd-3 writable buffer, unmeasurable through the seam), dropping a late binding resolution before snapshot (its three assertions all hold pre-fix, so it is not a fail-before case), the done-value TOCTOU pre-encoding (its concurrent-mutation race is not deterministically constructible through the seam, and the daemon-mutation regression's only assertion is probabilistic), the stray-UTF-8 budget-flush retention (a budget flush landing on a multibyte boundary is not schedulable through the seam — v8-ignored), and the late-rejection settled guard (a rejection arriving after settlement is not deterministically constructible from the seam), and the log-fragment seal (its 25 M-scale OOM is not deterministically constructible in CI), and the unknown-binding preview cap (its whole-target `JSON.stringify` peak is a transient allocation inside the reply path, unmeasurable through the seam) — so a future regression on the rest goes red.
@@ -0,0 +1,151 @@
+# Agent Note: CPython 后端中的结算、分帧与生命周期修复
+
+Status: implemented
+
+[English](2026-07-31-code-runtime-python-settlement-fixes.md) | 中文
+
+## Problem
+
+用于 PTC mode 的 CPython 子进程后端建立在 [fd-3 帧协议](../architecture/2026-07-31-code-runtime-python-fd3-protocol.zh.md)之上，把每个程序结果都 resolve 成一个 `CodeRunResult`，仅在 seam 被误用时才 reject `run()`，并且会 dispose 到完全停稳，从而没有任何留在子进程自己进程组内的子进程存活得比 fiber 更久（一个用 `setsid()` 逃出该进程组的后代是有文档记载的例外——见该包 README 的 Known Limitations）。一连串审查暴露出一些缺陷，它们以单元测试覆盖率无法捕获的方式破坏了这些契约：每一个都藏在一处 `/* v8 ignore */` 之后、一个读起来像修复但实际并非修复的捕获可调用对象之后、一处透过 seam 不可见的内存效应之后、一处重复计数的加载期上界之后、一处存活者能够熬过的进程组升级之后、一处静默死锁的跨事件循环完成之后、一处位于结算路径之外的同步抛出之后，或者一处被当作日志边界处理的传输边界之后。大多数行为修复都附带一个在缺少它时会失败的测试；有十一处没有，并被如此标注——分块读取帧（一处系统调用次数的改进，没有可跨平台确定性断言的失败）、确认为空后的收尾（它唯一透过 seam 可观测的效应，即一个冻结的心跳，会在 SIGKILL 被投递的瞬间冻结，而修复前"投递即收尾"的代码也会产生同样的结果，用于区分的探测手段是 Alternatives 以跨环境不可靠为由否决的 signal-0 检查）、共享的 stdout／stderr 预算（它唯一透过 seam 可观测的差异，是一次流中冲刷落在哪条条目边界上，而这取决于两条相互独立的 OS 管道的相对到达时机，`os.sched_yield` 并不能使其确定；它所强化的按管道计的内存界限确实由单管道洪泛测试覆盖），`flush_line` 的 join-清空-push 重排序（它把结算期冲刷的峰值从三份副本降到两份，但 12× 加载门本就覆盖了换行路径的三副本峰值，因此每个被门放行的配置在两种顺序下都落在地址空间之内、不存在透过 seam 可观测的差异——该内存效应在 Python 子进程内部，与共享预算那处一样无法透过 seam 度量），节流 binding 回复（它的树内用例只断言分帧后的回复仍能完整往返；它所移除的峰值位于宿主 fd-3 可写缓冲内部、透过 seam 不可见，因此 32.0 MiB → 0.0 MiB 的降幅只能在树外度量），以及在快照之前丢弃迟到的 binding 解析（它的三条断言在修复前同样成立，因为 `sendReply` 本就丢弃结算之后的值——只是比快照晚），完成值的 TOCTOU 预编码（与编码竞态的并发变异无法透过 seam 确定性构造——它的 daemon 变异回归只断言结果永不为 `worker-exit`，这种断言是概率性的、不具有判别力，因此按既有的"无 fail-before 测试且有理由"先例登记为无 fail-before）、stray UTF-8 预算冲刷的扣留（正好落在多字节边界上的预算冲刷无法透过 seam 调度；它被交叉标注为 v8-ignored），以及结算后到达的迟到拒绝的 settled 先查（在运行已经结算之后才到达的拒绝无法从 seam 确定性构造），以及日志分片封存（25 M 级单字符滴灌 OOM 无法在 CI 确定性构造；树内用例只断言其完成并截断），以及 unknown-binding 预览上限（完整 target 的 `JSON.stringify` 峰值是回复路径内的一次瞬时分配——它唯一透过 seam 可观测的痕迹是伪造近上限 `global`／`name` 时的峰值内存，无法透过 seam 度量；树内用例只断言运行完成）。
+
+## Decision
+
+若干处相互独立的修正，各自位于拥有对应缺陷的包中。
+
+### unknown-binding 预览从 1 KiB 前缀转义
+
+unknown-binding 回复用 `JSON.stringify` 对完整的限幅 target（`global` 加 `.` 加 `name`，各最多 `maxValueBytes` 个 code unit）构造消息——在控制字符密集字段下转义形式可达输入的约 6 倍，在 `maxValueBytes` 上限附近产生数亿字节峰值，这是任何敌意对等方边界都不会放行的。预览现在从 target 的 1 KiB 前缀转义（足以辨识 binding）；`capMessage` 仍执行回复预算。
+
+### 合并的 open 日志条目只计费一次，按片段分摊
+
+未结束行的显式 `flush()` 发出带 `open: true` 的 `log` 帧，宿主把下一个帧追加到同一条目（`print('a', end='', flush=True); print('b')` 读回为一条 `'ab'` 条目而不是假换行）。拆分计费算术——首片段付引号加内容加分隔符、续接与闭合帧只付内容、宿主 cap `logBudget - 1`／`logBudget + 2`、低于 2 字节的 walk guard、子进程的 `_open_started` 键控——只登记一次，见 [fd-3 协议 note 的 wire-contract 段](../architecture/2026-07-31-code-runtime-python-fd3-protocol.zh.md)。
+
+### Boot-write failure no longer rejects run()
+
+在 [`src/index.ts`](../../../../packages/experimental/code-runtime-python/src/index.ts) 中，fd-3 引导帧写入是 `run()` 同步初始化阶段的最后一条语句。它的 `catch` 会调用 `finish()`，而 `finish()` 读取 `wallTimer` 和 `onAbort`，并通过 `settle()` 读取 `live`。这些绑定是 `const`，且声明在引导写入之后，因此在同步写入失败时，`finish()` 会在它们处于暂时性死区（temporal dead zone）时访问它们，从而抛出一个 `ReferenceError`。该错误逃出了 Promise executor 并 reject 了 `run()`，违反了 seam 的"结果一律 resolve"契约：调用方看到的是一个被抛出的错误，而不是 catch 构造的 `worker-exit`。现在引导写入代码块被放到 `wallTimer`、`onAbort` 和 `live` 初始化之后，并且那处曾把该分支从覆盖率中隐藏的 `/* v8 ignore */` 已被移除，从而使该 catch 被纳入度量。
+
+### Log capture is serialized against settlement
+
+在 [`py/bootstrap.py`](../../../../packages/experimental/code-runtime-python/py/bootstrap.py) 中，主协程上的结算 `flush_out()`／`flush_err()` 会读取并清空各个流的 `_pending` 列表，并修改共享的 `LogBuffer` 账本。模型代码可能启动一些 daemon 线程，其 `print`／`write` 会并发地修改同一状态。捕获绑定方法（`out_stream.flush_line`）只解决了结算调用哪个可调用对象的问题，而没有解决它在执行途中读取什么的问题：一次交错的 flush 可能拼接一个正在其下被修改的 `_pending` 列表，从而破坏账本并丢失 `done` 帧，使该次运行一直拖到墙钟超时。现在 `LogBuffer` 持有一把由两个流共享的可重入锁；`_LogStream.write` 和 `flush_line`，以及 `LogBuffer.push`，都会获取该锁，因此整个读-改-写过程在多线程间是原子的。
+
+### Fd-3 residual is copied, not viewed
+
+同样在 `src/index.ts` 中，在对待处理 fd-3 分片的 `Buffer.concat` 结果按换行符做循环之后，剩余的不完整行被以它被切出的 `subarray` 视图形式向前传递。视图会使整个 concat 的底层分配保持存活，因此一个大帧后面跟着一个极小的尾部片段，会钉住整整一帧大小的内存，而 `pendingBytes`（被设为该片段的长度）报告的值远小于实际保留的内存。现在，残余数据通过导出的 `detachResidual` 辅助函数被分离到一个大小恰当的新 `Buffer` 中，从而让 concat 分配得以被回收，并使 `pendingBytes` 成为一个诚实的度量值。
+
+### Output-cap load bound is ceiling minus envelope, not divided by six
+
+那处在加载期拒绝比单个 fd-3 帧所能承载更大的 `maxLogBytes`／`maxValueBytes` 的检查，会把帧上限除以六以应对最坏情况下的转义膨胀。但这两项预算都是以已转义的序列化字节来计量的：宿主日志账本通过 `jsonStringCostUpTo` 按序列化开销计费（它走到上限而不分配转义后的副本），`checkDoneValue` 度量的是转义后的形式，而生产侧的 `_cap_message` 同样按序列化开销设上限，因此一个在上限之内被放行的载荷在传输时最多占用 `cap + envelope`；转义已经包含在计费之内，不能再被乘一次。现在该上界为 `FRAME_PARSE_CAP_BYTES - FRAME_ENVELOPE_BYTES`（接收路径在解码前拒绝原始长度超过 64 MiB parse cap 的帧——本次运行以 worker-exit 结算——因此预算不得超过诚实子进程的帧能穿过该解析器的值），未使用的 `MAX_JSON_ESCAPE_EXPANSION` 常量已被删除。旧的上界并非不安全（它是放行不足），但它静默地禁止了合法的大上限。这同一处加载检查还会拒绝一个非整数的 `maxLogBytes`／`maxValueBytes`：子进程通过 `int(...)` 读取每一项预算，而 `int(...)` 会对浮点数向下取整，因此 `maxLogBytes: 3.5` 会在子进程侧截断在 3 字节，而宿主却把小数部分也计入——两侧因此强制着不同的公开配置。在加载期拒绝该浮点数使两侧保持一致，与 worker 后端相符。
+
+### Same-group survivors are reaped before the fiber goes quiescent
+
+模型程序可能在子进程自己的进程组里（没有 `setsid`，因此 `kill(-pid)` 能到达它）留下一个后代，它忽略 SIGTERM，但释放了继承而来的 stdout／stderr／fd-3 管道。随后 leader 退出，由于管道已被抽空，它的 `close` 触发，于是结算在那个后代仍存活时运行。`kill()` 在 SIGTERM 之后装设一个 `unref` 的 SIGKILL 定时器；本次修复是，当有一次升级正在进行时，`settle()` 既不立即 resolve 该次运行的 `finished` promise，也不立即把该运行从 `live` 中移除。取而代之的是，当 `killing` 被置位且进程组尚未为空时（`process.kill(-pid, 0)` 不抛出 ESRCH），它在一个 ref 的定时器上轮询该进程组，以 `graceMs + CLOSE_REAP_MARGIN_MS` 为界，仅当进程组已清空后才把该运行从 `live` 移除并 resolve `finished`。这个 ref 的轮询是承重部分：它让宿主事件循环保持存活，直到 SIGKILL 真正回收了该进程组，因此即使是一个短命的宿主（一次性的 headless 运行、一个配置子进程）也无法退出并把存活者 reparent 给 init。把 `live` 的移除推迟，正是让一个与刚返回的 `run()` 竞争的 `dispose()` 仍会 await 该存活者的原因：若在回收之前就把运行从 `live` 移除，teardown 会快照到一个空集合并在后代仍存活时返回。在正常情况下（leader 是唯一成员），第一次探测返回 ESRCH，结算以零附加延迟完成收尾。`teardown()` 会 await 每次运行的 `finished`，因此 dispose 是真正完全停稳的，与其 JSDoc 相符——包括对一个已经 resolve 的运行也是如此。
+
+结算还会在进程组被确认为空的那一刻取消 SIGKILL 定时器（正常路径，以及轮询看到存活者已消失时）。让它继续处于装设状态会暴露一个 PID 复用隐患：一个在 leader 被回收后仍挂起长达 `graceMs` 的 `kill(-pid)`，可能在内核复用了 leader 的 pid 之后击中一个被回收（recycled）的 pgid，从而 SIGKILL 掉一个无关的进程组（`killGroup` 吞掉 ESRCH 并无帮助——危险恰恰是那次针对被复用进程组成功执行的 kill）。在空进程组探测时清除它，把复用窗口收窄到只剩真正存在存活者的情形，此时进程组不可能为空以供复用。
+
+被清除的定时器覆盖不到的那段窗口，由 `killGroup` 内部的**身份校验**封死。它发出的每个信号都是裸 `process.kill(-child.pid, sig)`——与 `child.kill()` 不同，它没有 handle 守卫——因此在 leader 被回收到 `close` 触发之间的那段间隔里（实测有一个持有管道的后代时可达 3039 毫秒），信号会打到一个被复用的 pgid 上。所以 leader 的启动时刻在 spawn 时读取一次（`/proc/<pid>/stat` 第 22 字段），并在每次发信号前重读，有两条裁定：读数**存在且不同**意味着该数字现在属于另一个进程，于是扣下信号；读数**缺失**意味着 leader 已被回收，而这正是每次升级的常态——它的 `/proc` 条目已消失，而它曾领导的进程组仍可能持有本次 teardown 要回收的那个存活者——于是信号照常发出。缺失也是无 `/proc` 平台上的恒定读数，那里守卫处于惰性状态、保持原有行为。把缺失读作身份不符并非假想：第一版就是这样做的，它扣下了宽限期的 SIGKILL 与轮询截止分支的 SIGKILL，导致三个同组心跳用例在 Linux coverage lane 上变红，而在 Darwin 上因读取器恒返回 undefined 而通过。
+
+回收轮询还会处理宿主事件循环被阻塞、越过两个定时器的情形。如果一次同步计算从轮询被调度之前一直占住事件循环、直到越过它的截止时间，那么当事件循环恢复时，轮询定时器和宽限窗口的 SIGKILL 定时器都已逾期，而 Node 会先运行更早调度的轮询——因此宽限窗口的 SIGKILL 可能从未触发。为此截止时间分支会自己发送 SIGKILL（若定时器已运行则该操作幂等），而不是取消尚未触发的升级，随后再额外给予一个 `CLOSE_REAP_MARGIN_MS`，并持续轮询直到进程组被确认为空，因为仅凭信号投递就收尾会在进程组仍在消亡时宣告完全停稳。因此等待的外层上界为 `graceMs + 2 * CLOSE_REAP_MARGIN_MS`。若这段额外余量耗尽而进程组仍非空，一个最终的硬性上界会收尾；该分支带有一处 `/* v8 ignore */`，因为它仅在一个被 SIGKILL 的存活者作为僵尸进程滞留且从未被 `wait()`——一个 PID 1 不回收孤儿进程的容器——时才可达，而这无法在各 CI 平台上确定性地构造出来。该 ignore 的理由陈述的是这种环境依赖性，而不是声称该分支不可能运行，并交叉引用 Alternatives 中以同样理由否决 signal-0 回收断言的那一条。
+
+### RLIMIT clamps against the inherited soft limit, not only the hard
+
+在 [`py/bootstrap.py`](../../../../packages/experimental/code-runtime-python/py/bootstrap.py) 中，`_clamped` 仅用继承而来的 HARD 限制来约束一个请求的 `(soft, hard)` rlimit 对。一个继承了低于请求值的软限制的部署——比如继承 `(100, 200)`、请求 `(150, 160)`——会拿回 `(150, 160)`，把有效软限制从 100 抬高到 150：对 `RLIMIT_AS` 而言这放松了内存上限，对 `RLIMIT_CPU` 而言它推迟了 SIGXCPU，两者都违反了"取配置值与继承值中最严格者"。现在 `_clamped` 用每一侧各自继承而来的对应值来约束该侧（`RLIM_INFINITY` 不施加任何上限），随后把 soft 钉在 hard 之下，因此 `setrlimit` 绝不会看到一个倒置的对。结算时的 CPU 复查（`die_if_cpu_exhausted`）遵循同一规则：它把已消耗的 CPU 与实际生效的、被夹紧的 `cpu_soft` 比较，而不是与配置的 `cpuSeconds` 比较，因此一个捕获 SIGXCPU、在返回前消耗超过更严格的继承软限制的程序会被报告为 timeout，而非误判为成功。 复查会在解除程序屏蔽的 SIGXCPU（`pthread_sigmask(SIG_UNBLOCK, ...)`，import 期捕获）之前先恢复 SIG_DFL：一个既安装自定义 handler 又屏蔽信号的程序，否则会在 unblock 的瞬间让那个挂起的 handler 以模型代码身份运行（可重新屏蔽或抛出），所以信号被释放时处置必须已是 SIG_DFL；SIG_DFL 在前时，挂起的信号在内核内直接致死、无字节码窗口，而 `kill` 重投递是给从未挂起情形的兜底。SIGXCPU 诊断不再把配置的 `cpuSeconds` 说成实际生效的预算——在一个更严格的继承软限制之下那个数字是错的——而是报告 CPU 时间是在"至多配置的 N 秒"处被耗尽，这一表述无论哪个限制先触发都成立。
+
+### 并发 binding 回复对 fd 3 做节流
+
+`sendReply` 忽略了 `proto.write` 的 `false` 返回值，因此一个在一轮 `asyncio.gather` 中解析多个大值的程序，会把每条回复都在同一个 turn 内编码、并全部排入 fd 3 的可写缓冲。binding 回复在 seam 层没有字节上限可以约束它，而且这个失败杀掉的是**宿主进程**而不是让本次运行失败：在 highWaterMark 为 64 KiB 的管道上实测，八条 4 MiB 回复会同时缓冲 32.0 MiB。现在回复走一个队列，一次编码并写出一帧，管道写满时等待 `drain`——同样形状实测峰值为 0.0 MiB。编码放在循环内部，因此一条运行已不再需要的排队回复会被 `settled` 检查丢弃，根本不会被序列化。同一个 `settled` 谓词还在 `await fn(...)` 之后、`snapshotJsonValue` 之前守护回复回调，因此一个在结算之后才 resolve 的宽值会在走完它的宽度之前被丢弃——宿主不会为一个结果已定的 run 展开一个迟到的宽值。
+
+节流不改变任何模型可观测的行为。子进程通过一个持续读取 fd 3 的 pump、按 id 把每条回复匹配到它自己的 `call`，因此到达顺序从来不可观测，而各 binding 本身仍然并发执行——只有宿主的峰值内存与冲刷时延改变。这也正是为什么串行化并不构成对 seam 并发契约的收窄，而那恰是最初推迟此项的理由；那个理由是错的。
+
+### Binding replies complete on the calling loop's thread
+
+同样在 `py/bootstrap.py` 中，一个绑定回复 Future 是在运行 `dispatch` 的那个事件循环上创建的。当模型通过 `asyncio.run(tools.x(...))` 从一个工作线程调用某个绑定时，该 Future 属于该线程的事件循环，而不是 `_pump_replies` 读取回复的主事件循环。`asyncio.Future` 不是线程安全的：从另一个线程完成它并不会唤醒它自己的事件循环，因此直接的 `set_result`／`set_exception` 会让那个正在等待的线程被搁置，该次运行退化为墙钟超时。现在每个待处理条目都会在记录 Future 的同时记录其 Future 所属的事件循环，`_pump_replies` 通过该事件循环的 `call_soon_threadsafe` 来完成它。共享的 `pending`／`next_id` 状态由一把 `threading.Lock` 保护，该锁跨越 id 认领、fd-3 写入和计数器推进这三步持有，因此并发调用方无法以违反宿主所要求的 id 顺序来交错帧。对一个已经关闭的事件循环（工作线程已结束、在回复到达前放弃了它的调用）调用 `call_soon_threadsafe` 会抛出 `RuntimeError`；该调度被包裹起来，使这个已无意义的回复被丢弃，而不是让异常终结 pump 任务并搁置此后的每一个回复。
+
+### The blocking frame reader reads in chunks, not byte by byte
+
+`ProtocolChannel.read_frame`（用于 `boot` 和 `run` 握手帧）过去通过在无缓冲（`buffering=0`）fd 上的 `FileIO.readline()` 读取，这会为每个字节发起一次 `os.read(1)`。`run` 帧在 `RLIMIT_CPU` 生效之后才到达，因此一个合法的数兆字节程序会在 `ast.parse` 运行之前，在数以百万计的单字节系统调用中烧掉数秒 CPU——有可能仅在读取这一步就耗尽预算。现在它以 `_READ_CHUNK_BYTES` 为单位分块读取，写入异步读取器已经使用的那同一个 `_pending` 残余缓冲区（包裹用的 `os.fdopen` 对象已被移除；两个读取器都直接调用 `os.read(self._fd, ...)`），因此读取开销微不足道，并且越过换行符的预读也为下一帧保留了下来。两个读取器都跟踪一个持续推进的扫描偏移（`find(b"\n", scanned)`），使一个跨多个分块累积起来的大帧只被扫描一次，而不是每来一个分块就从索引 0 重新扫描——分块式重扫会把逐字节的开销换成同一大帧路径上 O(N²) 的 memchr 开销。
+
+### Synchronous spawn failure resolves worker-exit, not reject
+
+同样在 `src/index.ts` 中，`spawn` 是在结算 Promise 的 executor 存在之前被调用的。Node 只把一组固定的 spawn errno（EACCES、EAGAIN、EMFILE、ENFILE、ENOENT）推迟为一个异步的 `error` 事件，而结算路径已经把它转成一个 `worker-exit`；其余每一个 errno 都会从 `spawn` 同步抛出。一个长度超过平台 PATH_MAX 的 `pythonBin` 能通过加载期校验（非空、无 NUL），却会让 `spawn` 在此处抛出 `ENAMETOOLONG`——在 executor 之外——因此 `run()` 会 reject 而不是 resolve，违反了"只 resolve、不 reject"，并且由于只有 `settle()` 才会移除本次运行刚物化出来的暂存目录，它会把该目录留在磁盘上。现在 `spawn` 调用和 fd-3 收窄被包裹起来：一次同步抛出会移除暂存目录，并 resolve 与异步 `error` 事件所产生的同一类 `worker-exit`（`python spawn error: …`）。
+
+### 解释器选择与子进程环境在加载期固定
+
+`pythonBin` 在插件加载期解析为一个可执行绝对路径，并在与运行时相同的受限环境中完成版本探测。提供方要求 CPython 3.10 或更高版本并保留该确切路径，因此后续 `PATH` 或工作目录变化不能切换解释器；不是可执行普通文件的显式路径、无法解析的裸名或不受支持的解释器都会在 `ctx.codeRuntime` 注册前失败。同步探测有固定的五秒期限，并在期限到达时发送 `SIGKILL`，因此忽略 `SIGTERM` 的包装脚本不能阻塞插件加载。每次探测与运行只接收 `TMPDIR`：macOS 系统 Python 需要它来避免向被捕获的 stderr 发出启动警告，而凭证、`PATH`、`HOME` 与其他宿主环境值均不会进入模型代码。若已校验的可执行文件在激活后消失，普通 spawn 结算仍 resolve 为 `worker-exit`。
+
+### Stray pipe output is aggregated by line, not by transport chunk
+
+同样在 `src/index.ts` 中，原生 stdout／stderr 字节（C 扩展写入、越过管道缓冲区的 `os.write`）过去每来一个 Node `data` 分片就被推入 `logs` 一条条目。`logs` 条目在下游（PTC mode）会用 `\n` 拼接，因此一次大于单次管道读取、且不含换行符的写入——它以若干个 `data` 分片到达——回读时会在任意传输边界处被插入模型可见的换行符。现在捕获会累积原始 `Buffer` 分片（与 fd-3 读取器同一形态，出于同样的原因：一个字符串 `+=` 累加器会为每个分片重新复制整份残余数据，而每个分片都从索引 0 扫描它则是第二重平方——在一次大的不含换行符的写入上二者都是 O(N²)），在原始的 `0x0a` 字节处切分，并为每个完整行准入一条条目。换行符绝不会出现在一个 UTF-8 多字节序列内部，因此对每个切出的行做解码无需流式解码器即可安全进行。三条相互独立的界限使残余数据不至于耗尽宿主内存，每一条都与 fd-3 读取器相对应：分片列表在越过 `MAX_PENDING_CHUNKS` 后会封存（SEAL）为已完成的块，因此一个以单字节 `os.write` 控速的程序无法累积起数以百万计的存活 Buffer 对象（其逐对象开销是任何字节计数都看不到的）；当两个管道合并（COMBINED）的持续推进序列化（SERIALIZED）开销——通过 `accrueStrayCost` 跟踪，它跨分片按结构解码 UTF-8，因此一个渲染为 U+FFFD 的字节会被计入该替换字符序列化后的三个字节——将要越过预算时，残余数据会被冲刷，因此一场控制字符或非法 UTF-8 的洪泛会在原始字节的一小部分处就冲刷，而不是先累积起满满一个预算份额的原始字节，并且 stdout 与 stderr 是合并计量的，而不是各自对照完整预算（那样会让两者同时各保留将近一个预算份额，使峰值翻倍）；而一旦账本已经截断，缓冲便停止，从而不会为永远无法被准入的输出累积任何内容。`accrueStrayCost` 按 U+FFFD 宽度对非法字节计费，正是针对一个从不作为合法序列开头的字节（0x80–0xC1、0xF5–0xFF）、一个在完成前断裂的多字节序列，或一个结构完整但非法（ILLEGAL）的序列的修复：`toString('utf8')` 会把其中每一个这样的字节都渲染为它自己的 U+FFFD（3 字节），因此它校验每个前导字节的首个后续字节范围（WHATWG：`E0`→A0-BF、`ED`→80-9F、`F0`→90-BF、`F4`→80-8F，其余为 80-BF），并对任何落在该范围之外的序列按每字节 3 计费。按原始的 1 计费会把一场 `b"\xff"` 洪泛少计三倍，而只按结构宽度计费同样廉价地把一个 CESU-8 代理项（`ED A0 80`）或过长编码（`E0 80 80`）少计三倍，让残余数据在冲刷前增长到满满一个预算份额的原始字节，并且在一个较大的 `maxLogBytes` 附近，在冲刷的 concat 加 `toString` 中膨胀到约 1 GiB 的峰值。被准入字符串的每条条目计费通过 `jsonStringCostUpTo` 按序列化开销计量，它把字符串走到上限即停止——先前的 `Buffer.byteLength(JSON.stringify(text))` 会先分配出整份转义后的形式，因此在一个较大的 `maxLogBytes` 之下，一行接近预算、控制字符密集的内容，仅仅为了度量它就可能瞬时分配超过一 GB。`jsonStringCostUpTo`（走字符串的那个函数，由一个伪造的、其文本经 `JSON.parse` 产生的 `log` 帧到达）给一个孤立（LONE）代理项计满六个转义字节（在 ES2019 良构 `JSON.stringify` 下为 `\uXXXX`），而不是 `Buffer.byteLength` 为其 U+FFFD 渲染所报告的三个字节，因此一场 `\ud800` 洪泛不会被少计一半；`accrueStrayCost` 走原始字节，从不把一个代理项当作代理项看到——一个 CESU-8 编码的代理项到达它时是三个字节，被它的逐前导字节范围检查所拒绝，每个计 3（共 9），与 `toString('utf8')` 所渲染的相符。该残余数据会在管道 `end` 时冲刷，也会在 `closeDeadline` 处理器销毁流之前被显式冲刷：一个持有管道不放的 `setsid` 逃逸者会迫使结算在没有 `end` 的情况下走那条路径，因此 leader 在退出前发出的最后一次不含换行符的 `os.write(1, …)` 否则会从 `logs` 中被丢弃。
+
+### An incompatible output-budget/addressSpaceMb pair is rejected at load
+
+子进程（[`py/bootstrap.py`](../../../../packages/experimental/code-runtime-python/py/bootstrap.py)）在 `RLIMIT_AS` 之下构建、计费并分帧一条 `maxLogBytes` 的日志条目或一个 `maxValueBytes` 的完成值，而两个账本都是按字符计数、对照一个序列化字节预算触发的。一个星芒面字符是一个字符，但占 CPython `str` 存储的四个字节以及四个 UTF-8 字节，且最重的路径峰值时有三份这样的副本同时存活：一次 `sys.stdout.write(line + "\n")` 会持有调用方的 `text` 实参（在整个 `write` 调用期间存活，约 4 倍）、交给 `LogBuffer.push` 的行切片（约 4 倍）、以及 `_push_locked` 为计费和发送而取的 `text.encode("utf-8")` 副本（约 4 倍）——峰值约为预算的 12 倍。结算期的 `flush_line` 路径只持有两份（它的 `"".join(...)` 与那份 encode 副本——它在 push 之前先丢弃 pending 分块），因此换行路径才是起约束作用的最坏情况。当一项预算逼近 `addressSpaceMb` 时，一次合法的、接近预算的输出会在那次构建加编码期间突破地址空间，并作为 `worker-exit`（日志）而不是截断而终止，或作为 `output-limit`（值）而失败。在运行时对每次子进程写入按地址空间计量是错误的修复：热路径上一次精确的序列化开销检查，要么是一次完整的 `encode`（正是要避免的那次分配），要么是一个逐字符的 Python 循环（它会烧掉 CPU 预算——一次 10 MB 的合法写入会在 `cpuSeconds: 1` 之下触发 SIGXCPU）。两者都是拿一种资源界限换另一种。取而代之，[`src/index.ts`](../../../../packages/experimental/code-runtime-python/src/index.ts) 在加载期（LOAD）拒绝这个不兼容的组合：每项预算乘以 `OUTPUT_BUDGET_WORST_CASE_ADDRESS_SPACE_MULTIPLE`（十二——换行路径三份同时存在的约 4 倍副本）必须放得进为解释器自身占用预留一份固定的 `INTERPRETER_BASELINE_BYTES` 之后剩下的地址空间，并用一个 `>=`，使得一项其最坏情况峰值恰好等于那片余量的预算也会被拒绝（该峰值加上预留的基线正好是整个地址空间，即 `RLIMIT_AS` 边界）。`flush_line` 也被改为在 push 之前先丢弃 pending 分块，与换行路径的 join-清空-push 顺序一致，使它至多只持有 join 及其 encode 副本，而非三份副本。该基线是与倍数分开（SEPARATELY）预留的，因为它是一项固定开销，而非随预算伸缩的开销：把它折进倍数会让一项恰好为 `addressSpaceMb / 12` 大小的预算被放行，而其峰值加上解释器仍会越界。`maxLogBytes` 和 `maxValueBytes` 都被对称地门控；值路径的构建加编码是同一形态。该检查在每个平台上都运行，而不仅在强制 `RLIMIT_AS` 的平台上：这种不兼容是那些配置值的属性，因此一个 Linux 部署无论由哪个宿主组装配置都会 OOM，而一致的加载期拒绝正是 fail-loud 契约（Darwin 仅跳过运行时的 `setrlimit`）。这在配置 seam 处消除了这一类问题，而不是给写入路径打补丁，因此 `_LogStream` 保留它原有的按字符计数的缓冲（一个对序列化开销有效的下界，一旦预算放进地址空间就是内存安全的）。值路径在第二处施加同样的纪律：`_check_done_value`（字节计量器）与 `_encode_json_plain`（帧编码器）都以 O(DEPTH) 而非 O(width) 遍历。每个容器只压入一个游标帧、逐个拉取子元素，而不是每个子元素一个遍历元组或栈条目——一个扁平的 `[0] * 6_000_000` 序列化后约 12 MB，但逐元素遍历会分配约 400 MB 的簿记（约为序列化尺寸的 28 倍，远超门预留的 12 倍），于是一个被计量器放行的值可能因遍历自身的帧而 OOM。改用游标后，唯一与宽度成正比的分配就是计量器已界定的输出字符串。
+
+宿主门控是对照配置的（CONFIGURED）`addressSpaceMb` 校验的，但一个启动环境可能继承一个更严格的（STRICTER）`RLIMIT_AS`（一个低于 `addressSpaceMb` 的 `ulimit -v` 包装层），而 bootstrap 的 `_clamped` 会正确地把有效（EFFECTIVE）限制降到该值——从而让这些预算是按一个子进程永远得不到的上限来定尺寸的。因此 `bootstrap.py` 在应用该有效被夹紧的软限制之后，会对照它重新检查两项预算，镜像宿主门控的倍数与基线，并在引导期抛出（被 setrlimit 阶段的处理器捕获，并作为 `exception` 上报——与任何其他资源限制应用失败同属一类），而不是任由一次接近预算的输出在运行途中 OOM。这两个子进程侧常量与宿主侧的保持一致，靠的是共享的推理，而不是一个 wire 字段。
+
+在此之外还一并修复了一处残余写入路径的复制，它与配置门控相互独立：`_LogStream.write` 的换行分支会在冲刷触发器能够对其设界之前，先把最后一个换行符之后整个未结束的尾部（`text[pos:]`）缓冲进 `_pending`，因此一个早出现的换行符后跟一个巨大的尾部（`"\n" + "A" * 30 MiB`）会对模型自身的字符串再做一份完整副本——正是这条路径存在所要规避的那次 `RLIMIT_AS` 死亡，而且是配置门控无法覆盖的一次，因为该尾部可能远超 `maxLogBytes`。现在该尾部被切到一个 `remaining + 4` 字符的前缀（超过 `remaining` 字符的任何内容都无法被准入，因为字符计数是序列化开销的下界），随后冲刷触发器会用标记将它拒绝。
+
+### 完成值与错误在其校验点处预编码
+
+在 [`py/bootstrap.py`](../../../../packages/experimental/code-runtime-python/py/bootstrap.py) 中，`_done_with_value` 现在会在成功路径上把整个终止帧作为一个已预编码的 JSON 字符串返回：被准入的值在这里、在 `_run` 的 `try` 之内的校验点处恰好序列化一次，即 `'{"type": "done", "value": ' + _encode_json_plain(value) + "}"`。程序在返回之后仍可能从 daemon 线程或信号处理器继续变异它返回的 list／dict，因此在一个更晚的点上做第二次遍历会构成一次 TOCTOU——如果一次变异让一个并发变异的后续编码在结算处理器之外抛出，就会把一次已结算的运行在宿主侧降级成 `worker-exit`。在这里、在包裹该调用的 `try` 之内恰好序列化一次，就关上了这个窗口：如果一次并发变异导致编码抛出，异常处理器会把它如实分类为 `exception`；一旦字符串产生出来，该帧就会被逐字写走、不再触碰任何活对象。`_run` 在程序运行前把 `_done_with_value` 的入口名绑成局部（`done_with_value_bound`），而 `_done_with_value` 自身把 `_check_done_value` 与 `_encode_json_plain` 绑定为 def 期默认参数——因此模型执行后对入口名或这两个名字的 `__main__` 重绑无法把一个合法成功改写为 `exception`。日志账本（宿主 `logBudget` 与子进程 `_remaining`）从预算低 1 字节起算，预留序列化外层数组的外壳（两条括号与 n-1 个逗号，覆盖 n 条目的分隔符），因此恰好耗尽账本的结果序列化后仍在配置上限之内。 一旦账本已截断，宿主会整体清空两条 stray 管道的缓冲输出（之后的每个字节都会被 `admit` 变成 no-op，保留它只会把宿主内存花在永远无法准入的输出上）；子进程以 `-u` 运行，使 `sys.__stdout__`/`sys.__stderr__` 的写入对 stray 捕获立即可见，而结算 flush 仍在 done 帧前排空原始 std 流（防御 `sys.__stdout__ = boom` 重绑后残留的缓冲包装）。构造器拒绝低于 64 的 `maxLogBytes`（能为截断标记自身序列化形式留出一字节余量的最小预算）；`maxValueBytes` 只保留正整数要求，因为完成值可以只有一字节、且 done 帧外壳是 seam 协议成本。标记仍是 envelope，因此带已放行条目的截断运行序列化后至多为 `maxLogBytes + marker + envelope`（已记录在包 README）。但编码器到达的一个传递依赖（例如 `_dump_scalar`/`_dump_string`/`json`/`io`——非穷举清单）重绑仍可让编码抛出、把成功降级为 `exception`，这在包 README 中被登记为已接受残余。
+
+`send_done`（`_run` 内部的一个局部函数）通过绑定的 `channel.write_encoded` 写出已预编码的字符串，并在写之前用绑定的 `_encode_json_plain` 编码一个 dict 错误帧——它绝不经 `channel.send_sync`，因为后者的函数体会在调用时刻重新解析 `self.write_encoded` 和模块级的 `_encode_json_plain`。`_encode_json_plain` 与 `channel.write_encoded` 在程序运行前就被绑定进局部变量，理由与 `flush_out`／`flush_err`／`safe_model_traceback` 被绑定相同：程序以 `__main__` 运行，因此 `import __main__; __main__.ProtocolChannel.send_sync = boom` 或 `__main__._encode_json_plain = boom` 本会在调用时刻把发送／编码重新解析成被替换的可调用对象，当该替换抛出时跳过 `done` 帧、把已结算的结论降级成宿主侧的 `worker-exit`。
+
+### 预算触发的冲刷会扣留下一个未完成的多字节尾序列
+
+同样在 [`src/index.ts`](../../../../packages/experimental/code-runtime-python/src/index.ts) 中，`flushStray(stray, retainPartialTail)` 在预算触发的冲刷（`captureStray` 中的合并成本阈值）上会把一个未完成的多字节尾部从解码中扣留：当残余以部分 UTF-8 前导序列结束（`stray.utf8.expected > 0`）时，前导字节加上迄今已消耗的续字节（≤3 字节）会从该帧中分离出来作为新的残余，只有完整的前缀被准入并解码。当整个残余就是单个未完序列时什么都不准入，因此一个合法、未完成的字符绝不会在一个被放行、未截断的条目里被渲染成 U+FFFD，也不会推进一条虚假的空条目。被扣留的尾部会从一个全新的 `stray.utf8` 状态重新累计——若用冲刷后 `expected > 0` 的状态来计量它，会把那些被承载下来的前导字节当作非法续字节计费——从而下一个分块能正确地继续推进，且该管道在保留的尾部之上重建其成本／UTF-8 状态。`end`／`closeDeadline` 路径传入 `false`，原样解码整个残余，因为在那里一个不完整的尾序列是真实的坏输入，U+FFFD 才是如实呈现。
+
+### 迟到的 binding 拒绝在格式化错误之前就返回
+
+同样在 `src/index.ts` 中，binding 拒绝的 catch 分支现在会在格式化 `messageOf(error)` **之前**检查 `settled` 并返回。一次在 `maxWallMs`、abort 或 dispose 已经把该运行结算之后才到达的拒绝，本会让 `messageOf(error)` 在这之前运行敌意的 `toString`／`message` getter——为一个结局已定的运行花费宿主堆与时间——然后 `sendReply` 才去窥探 `settled`。及早丢弃这一条已分帧的回复省下了这笔开销。运行中那条本就大致线性的回复排空改用队头游标按数组下标读取、而非 `shift()` 逐项弹出，因此一大轮等待 fd 3 的 `drain` 的宽 binding 的 `asyncio.gather` 会以线性时间排空，而不是因反复切片退化成 O(n²)。
+
+### 无换行滴灌会封存其分片；CPU 软限制保持在硬限制之下；done 帧回退到固定字面量；回复队列清空已消费槽位
+
+在 [`py/bootstrap.py`](../../../../packages/experimental/code-runtime-python/py/bootstrap.py) 中，`_LogStream` 现在会在待处理分片列表越过一个上限时封存它：无换行、每次 `write` 一个字符的滴灌会每次调用累积一个 list 槽位（以及一个 str 对象），在一个大的 `maxLogBytes` 下，25 M 次单字符洪泛会在字节预算达到之前，于其自身记账上 OOM（加上 `_push_bounded_prefix` 随后构造的同规模列表）。越过 `_PENDING_MAX_CHUNKS` 后，当前分片被 join 成一个块并移入 `_pending_blocks` 列表（字符数不变），把存活的碎片数量限制在宿主侧 `captureStray` 封存所做的同等水平；该 join 只针对 ≤cap 的当前分片，从不针对整个累积缓冲，因此大的滴灌保持 O(B)，而不是以 O(B²/cap) 次反复复制不断增长的块。宿主侧 open hold 镜像同样的封存：预算内的单字符 open 洪泛（`print('x', end='', flush=True)` 循环是诚实子进程可达路径）否则会为每帧累积一个片段数组槽位加字符串对象头——约 30× 字节计数看不到的开销，在 `maxLogBytes` 装载上限附近最高约 2 GB 宿主辅助堆。越过 `MAX_PENDING_CHUNKS` 后持有的片段并入 `openSealed`；闭合帧合并、`truncateLogs` 与 `finish` 残段都读取 sealed 加当前片段并清空封存。
+
+`_clamped` 还会把钳制出的、与硬限制相等的 RLIMIT_CPU 软限制降低一个单位（当硬限制至少为 2 时）。`ulimit -t N` 会同时设置两者，而当 soft == hard 时，内核会在同一 tick 检查硬限制并直接 SIGKILL 一个忙循环，因此 SIGXCPU 永远不会送达——而宿主只在 `signal === 'SIGXCPU'` 时把 CPU 超限分类为超时，所以一次确定的预算耗尽会被误报为 `worker-exit`。把软限制降低一个单位给 SIGXCPU 一个触发窗口，因此超限会被报告为超时。这仅限定于 RLIMIT_CPU（在 RLIMIT_AS 上的一字节软差异只会让子进程实际应用的限制与宿主预算门失步，没有需要保留的信号）。`hard >= 2` 守卫留下了 `hard == 1` 盲区——一个 1 秒的双限制无法把软限制降到 0，因此那里的确定超限仍被报告为 `worker-exit`。
+
+`send_done` 将其 encode+write 包进 try，任何来自被重绑的传递名（`_dump_scalar`/`os`）的抛出都会写入一条固定的预编码 done 帧，经由 `_run` 局部绑定的 `_os_write`/`_memoryview`/`_FALLBACK_DONE_FRAME`——因此一个已结算的 `exception` 判决绝不会被降级为 `worker-exit`，宿主仍会拿到一个判决。回复队列的头游标排空会清除每个已消费槽位，因此一个已写出的宽 payload 会被立即释放，把宿主内存限制在持续的 fd-3 背压下的当前积压量。结算路径各 `except` 子句所捕获的异常类同样在任何模型代码运行之前绑定：`_BaseException` 是 `_run` 的局部与 `_make_failure_reporter` 的闭包单元；`_RuntimeError`、`_BindingRejection`、`str` 与 `bool` 是 `_pump_replies` 的 def 期默认参数（函数体内的 `X = X` 绑定太晚——模型顶层语句会先于泵体首步执行）。重绑 `__main__.BaseException` 无法让程序异常逃出处理器、丢失 `done` 帧；重绑 `__main__.RuntimeError`（或 `_BindingRejection`/`str`/`bool`）无法让闭环调度失败逃出泵的捕获、把每条后续回复搁浅到墙钟超时。
+
+## Testing
+
+- `tests/runtime.spec.ts` 在加载期拒绝缺失、不可执行、非 CPython、低于 3.10 或无响应的解释器配置；在激活后更改 `PATH`，证明已解析的可执行文件保持固定；在激活后删除该文件，保留迟到的 `worker-exit` 路径；并断言运行中的程序能看到 `TMPDIR`，但看不到 `PATH`、`HOME` 或 `DEEPSEEK_API_KEY`。原生输出用例分别固定每个来源流内的顺序，而不要求独立通道之间存在总顺序；Darwin 资源限制用例明确说明或跳过平台特有的 `RLIMIT_AS` 行为。
+- `snapshots/session/ptc-python-turn` 通过真实 Loader 把 headless PTC worker 提供方替换为私有 Python 提供方，经真实 bash binding 重放一个 Python `run_code` 程序，并固定 Python SDK prompt、tool schema、dispatch event、捕获日志与完成值。
+- `tests/boot-write-failure.spec.ts` 对 `spawn` 做 mock，使 fd-3 管道在引导写入时抛出异常（这是真实子进程无法被迫进入的唯一路径），并断言 `run()` resolve 出一个 `worker-exit` 而非 reject。一个同级用例让被 mock 的 `spawn` 同步抛出，并断言 `run()` 仍然 resolve 出一个 `worker-exit`，且会移除它的暂存目录——以被 mock 的 `spawn` 在其 argv 中收到的确切引导路径为准，因此一个同级 worker 的并发暂存不会让它变得不稳定。两者都被隔离在这个 spec 中，因此真实子进程测试套件不受影响。
+- `tests/residual-detach.spec.ts` 对 `detachResidual` 做单元测试：向前传递的副本与残余数据相等、拥有一个大小与其自身长度一致的底层存储（fixture 保持在 Node 的 Buffer 池阈值之上），并且不与源帧的 `ArrayBuffer` 共享。
+- `tests/runtime.spec.ts`：output-cap 用例断言 `parse-cap - envelope` 上界（67108800）及其消息。一个 daemon 线程用例驱动四个线程穿过结算的 flush 发出未结束的写入。一个 native-write 用例在抬高后的 `maxLogBytes` 之下，通过 `os.write` 写入 200 KiB 且不含换行符，断言它回读时恰好是一条日志条目（证明散逸输出是按行聚合的，而不是在管道分片边界处被切开）；一个配套用例写入 `b"one\ntwo\nthree"`，断言得到三条条目（证明真正的换行符仍然起分隔作用）。一个 newline-free-flood 用例在一个 4 KiB 的 `maxLogBytes` 之下写入 2 MiB，断言捕获终止于截断标记且保持在预算之内（证明残余数据受账本约束，而不是被整体缓冲）；一个 NUL-flood 配套用例在同一预算之下写入 4000 个不含换行符的 NUL，断言发生截断（证明残余数据是按序列化开销计费的，约为原始的 6 倍，且在度量时不分配转义后的副本）；一个 illegal-UTF-8 用例在一个 3072 字节的预算之下控速发出单字节 `\xff` 写入，并对 `Buffer.concat` 做包装以度量峰值合并缓冲区，断言它保持在 2048 之下（按 U+FFFD 宽度 3 计费时残余数据在约 1024 原始字节处冲刷；一次原始字节的少计会让它达到约 3072，因此该界限具有区分力）；一个 CESU-8/overlong 用例把结构良构但非法的 `ED A0 80` 一次一个字节地控速发出，断言同样的峰值界限（按每序列真实的 9 计费时它提前冲刷；按结构宽度 3 计费会使峰值增至三倍，因此把逐前导字节范围检查回退会使它变红）；一个 broken-multibyte 用例在分开的分片里先写入一个 3 字节的前导字节、再写入一个新的 ASCII 字节，断言同时捕获到一个 `A` 和一个 U+FFFD（覆盖 `accrueStrayCost` 的跨分片断裂序列分支）；一个 post-truncation 用例写入一个 108 字节的载荷（小于最小的 PIPE_BUF，因此是一次原子写入），其首行耗尽一个 64 字节的预算，断言第二行被丢弃（覆盖单次 `data` 回调中的截断后准入空操作，无需 v8-ignore）；一个 short-escape 用例写入一行混合了制表符、引号、反斜杠、一个 `\uXXXX` 控制字符、一个多字节字符和 ASCII 的内容，断言它原样完成往返（覆盖 `jsonStringCostUpTo` 的每一条分支）；一个 reassembly 用例写入一个跨越每个合法多字节前导字节类别（E0 范围、普通 3 字节、F0 和 F4）、越过管道缓冲区的载荷，断言它原样完成往返且不含 U+FFFD（覆盖 `accrueStrayCost` 的逐前导字节范围与跨分片重组）；一个 lone-surrogate 用例在一个 4 KiB 预算之下伪造一个以 1000 个 `\ud800` 转义洪泛的 fd-3 `log` 帧，断言发生截断（该计数正落在计 3 字节会放行、计 6 字节则截断的窗口内，证明该代理项是按其完整转义宽度计费的）；一个 stray-sealing 用例在抬高后的预算之下控速发出 60000 次单字节、不含换行符的 `os.write(1, …)` 调用，并对 `Buffer.concat` 做包装以度量复制量，断言这股细流合并为一条条目、且累积复制量保持在一个实测的 256 KiB 阈值之下（封存后的形态复制约 120 KB，重新合并的形态复制约 538 KB，因此把封存回退成重新合并会使该断言变红——证明分片列表在越过 `MAX_PENDING_CHUNKS` 后封存为块）。一个 closeDeadline-flush 用例让 leader 写入一段不含换行符的诊断，随后 spawn 一个持有管道不放的 `setsid` 孤儿进程，断言该诊断在 `logs` 中存留下来（证明残余数据在截止时间销毁流之前被冲刷）。same-group 回收用例 spawn 一个忽略 SIGTERM 的同进程组后代，它释放管道并递增一个心跳文件；该测试断言在宽限窗口的 SIGKILL 之后心跳停止：无论被杀死的后代是被回收还是作为僵尸进程滞留，这个断言都成立，因此它在 PID 1 不 wait() 孤儿进程的环境下同样成立。一个 dispose-after-resolve 用例断言，对一个已完成、且存在同进程组存活者的运行调用 `dispose()`，只有在该存活者停止执行之后才返回（证明该运行会一直留在 `live` 中，直到它的进程组被回收），并带有一个 `expect(afterDispose).toBeGreaterThan(0)` 守卫，使得当心跳文件从未被写入时，冻结心跳的断言不会被空洞地通过。一个 deadline 用例忙阻塞事件循环越过两个定时器，断言该存活者的心跳冻结（证明轮询的截止时间分支自身发送 SIGKILL，而不是取消尚未触发的升级）。cross-loop 用例在主协程通过 `await asyncio.sleep` 让出时，从一个工作线程自己的 `asyncio.run` 事件循环运行一个绑定，断言该回复完成往返而不是超时；一个配套用例放弃某个线程的调用，使其事件循环关闭，Line truncated
+
+## Alternatives considered
+
+**保留引导写入处的 `/* v8 ignore */`，只修复顺序。** 已否决：正是那处 ignore 让这个 TDZ 回归得以未被发现地进入代码库。移除它使该 catch 成为被度量的分支，因此按文件计的 100% 覆盖率现在能证明该失败路径确实被执行。
+
+**通过捕获更多绑定方法来修复 flush 竞态。** 已否决：这正是已经失败过的做法。绑定一个可调用对象解决的是引用解析，而不是对该可调用对象所读取的可变状态的并发访问。只有对共享账本施加互斥才能消除该竞态。
+
+**用大小阈值来保护残余数据（只复制大帧）。** 已否决：该分支在每次包含换行符的读取时运行一次，复制的规模受残余数据自身长度约束（始终是一个不完整行），而阈值会引入一个可调参数和第二条代码路径，却换不来任何可度量的节省。无条件地做大小恰当的复制更简单，且始终正确。
+
+**通过 seam 断言残余数据的内存效应。** 已否决：被保留的分配透过 `CodeRunResult` 不可观测，因此黑盒测试无法区分已修复与未修复。转而抽取出 `detachResidual`，把底层存储的不变量变成一个确定性的单元测试。
+
+**仅用一个发后不理的 `unref` SIGKILL 定时器来回收同进程组存活者。** 已否决：`unref` 的定时器不会让宿主保持存活，因此一个在宽限窗口内退出的宿主（一次性运行、一个配置子进程）永远不会触发 SIGKILL，存活者被 reparent 给 init，这是同一个"没有子进程存活得比 fiber 更久"的违规换了个形态，而且 `teardown` 的"await 每个子进程退出"的 JSDoc 会变为不实。在一个 ref 的轮询上 await 进程组的消亡，让宿主恰好保持存活足够长以完成回收，在常见的空进程组情形下代价为零。
+
+**用 `process.kill(pid, 0)` 抛出 ESRCH 来断言回收。** 已否决：一个被 SIGKILL 的进程会作为僵尸进程滞留，直到它的父进程 `wait()` 它，而在一个 PID 1 不回收孤儿进程的容器里，signal-0 探测会持续成功，因此该断言会在跨环境时误报失败。一个停止推进的心跳文件检测的是"不再执行"，而被回收的进程和僵尸进程都满足这一点。
+
+**用一个普通的 `set_result` 完成跨事件循环的 Future 并依赖 GIL。** 已否决：GIL 序列化字节码，但并不使 `asyncio.Future` 跨事件循环安全：从一个并非其事件循环所属的线程完成一个 Future，不会调度它的回调，也不会唤醒该事件循环。在拥有该 Future 的事件循环上调用 `call_soon_threadsafe` 才是有文档记载的机制。
+
+**在结算之后让 SIGKILL 定时器继续处于装设状态（早先的同进程组修复）。** 已否决：一个被留待在 leader 被回收后长达 `graceMs` 才触发的 `unref` 定时器，可能 `kill(-pid)` 一个被回收（recycled）的 pgid，击中一个无关的进程组；危险是那次成功执行的 kill，而 `killGroup` 吞掉 ESRCH 无法阻止它。在进程组被确认为空后清除该定时器，把复用窗口收窄到真正存在存活者的情形，此时进程组不为空以供复用。
+
+**只用继承而来的硬限制来约束 rlimit。** 已否决：那会静默地抬高一个比请求更严格的继承软限制，放松了该约束本应保持的那种收束。用每一侧各自继承而来的界来约束该侧（随后把 soft 钉在 hard 之下），在 soft 和 hard 两者上都保持配置值与继承值中的最严格者。
+
+**按序列化开销对宿主侧的 `capMessage` 兜底做计费，与子进程的 `_cap_message` 相符。** 已否决：这两处上限守护的是不同的东西。`_cap_message` 的输出会作为一个 JSON 字符串再次穿过 fd 3，因此帧上限约束的是它转义后的宽度——那里必须按序列化计费。`capMessage` 的输出直接进入 `CodeRunResult.error.message`，绝不会再次穿过一个受帧上限约束的通道，因此对它所保留内容的诚实度量是模型可见字符串的原始字节长度。一个诚实的子进程已经按序列化开销设过上限，而原始长度 ≤ 序列化开销，因此一条格式良好的消息会原样通过；一条伪造的、控制字符密集的消息可能序列化到其原始长度约 6 倍，但由于它不经过任何受上限约束的通道，按那个被抬高的传输宽度对它计费只会截断一条尺寸合法的诊断，而换不来任何收束上的收益。每一侧的 JSDoc 都记录了这一区分，并指向另一侧。
+
+**每来一个 `data` 分片就把散逸的管道输出推入一条条目。** 已否决：`logs` 条目在下游会用 `\n` 拼接，因此一个传输分片边界会变成一个模型可见的换行符——一次被拆散在多次管道读取中的原生写入会带着无端的换行回读。按真正的换行符聚合（原始分片缓冲 + 在 `0x0a` 处切分）与子进程的按行粒度的 `log` 帧相符；账本仍然通过在残余数据将要越过预算时把它准入并截断，来约束一场不含换行符的洪泛。
+
+**逐帧强制 fd-3 帧上限（在计数器检查之前先切分）以避免一次批次边缘的误拒。** 已否决：帧上限检查在任何 `Buffer.concat` 之前读取字节计数器，正是为了让一个敌意程序无法迫使宿主内存达到 64 MiB 帧上限的约 2 倍（计数器与那次拼接是所持全部内容的第二份副本）。先切分以对单个帧计费，会在拒绝一个超上限的帧之前就 `Buffer.concat` 它，从而重新引入那种翻倍——正是出于这个原因，有两个回归测试断言了先计数后拼接的顺序。逐帧顺序本会修复的那次批次边缘误拒（一个合法的接近上限的帧，其携带换行符的分片同时也带上了下一帧的起始字节，在一次管道读取中把计数器推过上限）只有当 `maxLogBytes`／`maxValueBytes` 被配置到距 64 MiB 帧上限一次管道读取以内时才可达——比 32／64 KiB 的默认值高出好几个数量级。在任何配置下都抵御敌意输入的内存安全边界，优先于一个仅在病态的接近上限配置下才可达的误拒；计数器的超额计数与这一权衡都记录在该检查处。
+
+**当合并预算被越过时，按残余数据到达顺序冲刷两个散逸管道。** 已否决：stdout 与 stderr 是相互独立的 OS 流，它们的 `data` 事件本就彼此之间、以及与子进程自己的 fd-3 `log` 帧之间不确定地交错。seam 的 `CodeRunResult.logs` JSDoc 写着「in order」，周围的文字把它限定为一条流之内的程序发出顺序——跨并发流的顺序在这里本质上是尽力而为，因为没有任何宿主侧的冲刷顺序能够重建内核已经丢失的真实交错，因此在冲刷处保留一条残余数据的到达顺序换不来任何东西。一个固定的排空顺序与任何顺序一样有效。跟踪一个逐残余数据的到达计次以先排空较早的管道，会增加一个分支，它的两侧只在两个 OS 管道的相对时机上才触发，而 `os.sched_yield` 并不使之具有确定性，因此该分支无法在不写一个不稳定测试的情况下被覆盖——有成本却没有可观测的契约收益。
+
+**在运行时按地址空间对子进程日志账本计量，而不是在加载期拒绝该配置。** 已否决：对每次子进程写入做一次精确的序列化开销检查，要么是一次完整的 `encode`——正是一次超大写入负担不起、而账本那处廉价预检本就为规避它而存在的那次分配——要么是一个逐字符的 Python 循环，它会烧掉 CPU 预算（一次 10 MB 的合法写入会在 `cpuSeconds: 1` 之下触发 SIGXCPU）。每一种运行时做法都是在热路径上拿内存界限换另一种资源界限。地址空间的突破是 `maxLogBytes`／`addressSpaceMb` 组合的属性，而不是任何一次具体写入的属性，因此在加载期一次性拒绝这个不兼容的组合，能在没有任何逐次写入代价的情况下消除整类问题，并保留 `_LogStream` 原有的按字符计数的缓冲，它一旦预算放进地址空间就是内存安全的。
+
+## Consequences
+
+解释器误配置会在服务发布前失败，每次运行都使用加载期选定的可执行文件，并且子进程只接收 `TMPDIR`，从而消除 macOS 启动噪声而不暴露宿主凭证。私有提供方仍不进入已发布 profile，同时由一个 keyless 真实 Loader 快照固定其源码检出 PTC 组合。
+
+seam 的"只 resolve、不 reject"契约在引导写入路径和同步 spawn 失败路径上都得以成立，两者的覆盖率都是被度量的，且两者都不会遗留一个暂存目录。日志捕获是线程安全的，代价是每次写入和 flush 都要获取一次可重入锁，并且散逸的原生输出由它自己的换行符来分隔，而不是由传输分片来分隔。fd-3 残余数据的内存受实际保留的字节数约束，并且两个帧读取器都以一次而非平方级的方式扫描一个不断累积的帧。输出上限放行一个帧所能承载的每一个值，并在加载期拒绝一个非整数的预算。dispose 面对同进程组存活者是真正完全停稳的（以 `graceMs + 2 * CLOSE_REAP_MARGIN_MS` 为界，在进程组已为空时代价为零，并且一旦进程组清空就清除 SIGKILL 定时器，从而一次滞留的 kill 无法击中一个被回收的 pgid），RLIMIT 强制在 soft 和 hard 两者上都保持配置值与继承值中的最严格者（并且 SIGXCPU 诊断不再把一个宿主无法保证的预算说出来），并且从模型创建的线程调用的绑定会完成而不是超时，而且握手帧读取器不再在一个大程序上烧掉 CPU 预算。每处行为修复都附带一个在缺少它时会失败的测试，除了 Problem 一节点出的那十一处——分块读取帧（一处系统调用次数的改进）、确认为空后的收尾（它唯一透过 seam 可观测的效应会在信号投递时冻结，而修复前的代码也会产生同样的结果）、共享的 stdout／stderr 预算（它唯一透过 seam 可观测的差异取决于不确定的跨管道到达时机），`flush_line` 重排序（它降低后的峰值仍落在 12× 门本就放行的范围内，因此没有配置会有不同表现）、节流 binding 回复（32.0 MiB → 0.0 MiB 的峰值降幅位于宿主 fd-3 可写缓冲内部、透过 seam 不可度量），在快照之前丢弃迟到的 binding 解析（它的三条断言在修复前同样成立，因此不是一处缺代码即变红的用例），完成值的 TOCTOU 预编码（它的并发变异竞态无法透过 seam 确定性构造，而 daemon 变异回归的唯一断言是概率性的），stray UTF-8 预算冲刷的扣留（落在多字节边界上的预算冲刷无法透过 seam 调度——v8-ignored），以及结算后到达的迟到拒绝的 settled 先查（结算之后才到达的拒绝无法从 seam 确定性构造），以及日志分片封存（其 25 M 规模的 OOM 无法在 CI 中确定性构造），以及 unknown-binding 预览上限（其完整 target 的 `JSON.stringify` 峰值是回复路径内的一次瞬时分配，无法透过 seam 度量）——因此其余各处未来若发生回归都会变红。
@@ -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 .agents/notes/implemented/bug-fix/2026-08-05-context-meter-blind-to-compaction.md
-2026-08-05-context-meter-blind-to-compaction.md: daf11b41236943d5fb35e9363c128ce470c7f4e9
-2026-08-05-context-meter-blind-to-compaction.zh.md: 6ea2f8814c4731fdf7c08d8b867cd7086ceccdbf
+2026-08-05-context-meter-blind-to-compaction.md: a23e51771c74e350fd3fb04e3a23ab3cdfb6092e
+2026-08-05-context-meter-blind-to-compaction.zh.md: 12fd41af0d8c8864a120f26d91148543f74b2e2b
@@ -27,7 +27,7 @@ This reverses the "the ring, header, and bar length stay provider-exact" half of

 ## Alternatives considered

-**Project `measure().totalTokens` instead.** The measurement service already composes exactly this (`baseline` anchor plus signed `surfaceDeltaTokens`), and it reacts correctly — measured at 4383 → 304 across the same compaction. But it is a service over private replay state, not a pure fold, and a projection cannot call it. Reproducing its anchor inside a `ProjectionDefinition` needs `_estimateProviderAssistant`'s random access to the chunk events cited by seq (`session.events[seq]`), which `apply(state, event)` does not have. Anchoring on the sampled surface total is the same idea reachable from a pure per-event fold.
+**Project `measure().totalTokens` instead.** The measurement service already composes exactly this (`baseline` anchor plus signed `surfaceDeltaTokens`), and it reacts correctly — measured at 4383 → 304 across the same compaction. But it is a service over private replay state, not a pure fold, and a projection cannot call it. Reproducing its anchor inside a `ProjectionDefinition` needs `_estimateProviderAssistant`'s random access to the chunk events cited by seq (`session.eventAt(seq)`), which `apply(state, event)` does not have. Anchoring on the sampled surface total is the same idea reachable from a pure per-event fold.

 **Emit a synthetic usage record at the end of compaction.** Would move `pressureTokens` itself, but the only usage compaction holds is the summarization request's own — a different prompt entirely. Recording it as the conversation's prompt size would be a lie in the durable log rather than in one display.

@@ -27,7 +27,7 @@ AFTER  compact:  ring=4%  header=~4227/100000   rows=[system 18, tools 0, messag

 ## 备选方案

-**改为投影 `measure().totalTokens`。** 测量服务本来就合成了正是这个量（`baseline` 锚点加有符号的 `surfaceDeltaTokens`），而且反应正确——同一次压缩前后实测为 4383 → 304。但它是一个建立在私有重放状态上的服务，不是纯折叠，投影无法调用它。要在 `ProjectionDefinition` 内部复现它的锚点，需要 `_estimateProviderAssistant` 对按 seq 引用的分片事件进行随机访问（`session.events[seq]`），而 `apply(state, event)` 拿不到。以取样时的表层总量作为锚点，是同一个思路在纯逐事件折叠中可达的版本。
+**改为投影 `measure().totalTokens`。** 测量服务本来就合成了正是这个量（`baseline` 锚点加有符号的 `surfaceDeltaTokens`），而且反应正确——同一次压缩前后实测为 4383 → 304。但它是一个建立在私有重放状态上的服务，不是纯折叠，投影无法调用它。要在 `ProjectionDefinition` 内部复现它的锚点，需要 `_estimateProviderAssistant` 对按 seq 引用的分片事件进行随机访问（`session.eventAt(seq)`），而 `apply(state, event)` 拿不到。以取样时的表层总量作为锚点，是同一个思路在纯逐事件折叠中可达的版本。

 **在压缩结束时补写一条合成的用量记录。** 这确实能推动 `pressureTokens` 本身，但压缩手上唯一的用量是摘要请求自己的用量——那是完全另一个提示词。把它记成本对话的提示词规模，等于把谎言写进持久日志，而不只是写进某一处展示。

@@ -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 .agents/notes/implemented/bug-fix/2026-08-10-child-agents-join-their-parent-preset.md
-2026-08-10-child-agents-join-their-parent-preset.md: 5a3d0d1c5e2b492b10a0274204c10e244c37196f
-2026-08-10-child-agents-join-their-parent-preset.zh.md: ce5d69404d4f8b6a9879668d2ed0cb1c9d133a40
+2026-08-10-child-agents-join-their-parent-preset.md: 5321d99bad35cf26ebc0e7842dc68c2775bc384b
+2026-08-10-child-agents-join-their-parent-preset.zh.md: f0d5b59c3787ef69c2c983da0c8ec0ff4a43e6e7
@@ -22,7 +22,7 @@ This is a bind, not a mount, and both differences are load-bearing. The child ge

 `dsh-subagent` reaches the roster through `ctx.get('agentPresets')` with a type-only import and an optional peer dependency — the documented opportunistic-consumption pattern it already uses for `sandboxPolicy` and `approval`.

-Giving the child its parent's tools exposed a second defect the same agent-plane move introduced: `ToolRuntime` exempted SCOPED registrations from a restriction and filtered only the global layer, so once every model-facing row became an ancestor contribution, a child's `toolFilter` stopped constraining anything — and, with the global layer empty, `restrict()` rejected every name it was given as unknown, failing the child outright. The exempt set is the tools a scope registers ITSELF, not the tools that happen to live in the global layer; reading it the second way held only while those two sets coincided. `view()` now filters everything a scope inherits — the global layer and every ancestor layer — and exempts only its own. The own-layer exemption is load-bearing rather than incidental: the delegation runtime registers a child's `report` and structured-output tools into the child's own layer, and a filter naming the capabilities the child may use must not strip the machinery it answers through.
+Giving the child its parent's tools exposed a second defect the same agent-plane move introduced: `ToolRuntime` exempted SCOPED registrations from a restriction and filtered only the global layer, so once every model-facing row became an ancestor contribution, a child's `toolFilter` stopped constraining anything — and, with the global layer empty, `restrict()` rejected every name it was given as unknown, failing the child outright. The exempt set is the tools a scope registers ITSELF, not the tools that happen to live in the global layer; reading it the second way held only while those two sets coincided. `view()` now filters everything a scope inherits — the global layer and every ancestor layer — and exempts only its own. The own-layer exemption is load-bearing rather than incidental: the delegation runtime registers a child's structured-output tool into the child's own layer, and a filter naming the capabilities the child may use must not strip the machinery it answers through.

 ## Alternatives considered

@@ -30,11 +30,11 @@ Giving the child its parent's tools exposed a second defect the same agent-plane

 **Bind the child's key to the PARENT's key rather than to the standing mount.** Rejected because it changes what a child inherits: the parent's own scope layer carries its per-agent restrictions, which would then intersect into every descendant, and a child outliving its parent would hang off a disposed agent's key. Joining the standing mount gives the child its parent's composition and nothing else.

-**Extend the continuable activation setup registry to cover one-shot children.** Rejected because that registry's contribution type is synchronous `(childCtx) => () => void` with per-installation revocation, modelling deployment capabilities that come and go, while a preset join is a one-time bind with no revocation of its own. Widening it would have made the omission possible again for any driver that skipped the registry.
+**Introduce one shared child-setup registry for both drivers.** Rejected because a synchronous, revocable contribution models deployment capabilities that come and go, while a preset join is a one-time bind with no revocation of its own. Routing composition through an optional registry would make the omission possible again for any driver that skipped it.

 **Let `dsh-subagent` import `resolveSessionPreset` and mount by the resolved id.** Rejected because it makes the preset roster a hard module edge for a package that must work without one, and it lands back on the remount semantics above.

-**Filter every layer on the chain, including the scope's own.** Rejected because it makes a per-child capability filter delete that child's reporting and structured-output tools, which the delegation runtime registers into the child's own layer — an `allow` naming the capabilities a child may use would leave it unable to answer at all.
+**Filter every layer on the chain, including the scope's own.** Rejected because it makes a per-child capability filter delete that child's structured-output tool, which the delegation runtime registers into the child's own layer — an `allow` naming the capabilities a child may use would leave it unable to produce the requested result.

 **Leave the durable header alone and fix only the live join.** Rejected because the live child and the same child read cold would then disagree about which composition produced its history — the same class of defect, moved rather than fixed.

@@ -22,7 +22,7 @@ Status: implemented

 `dsh-subagent` 以类型级导入加可选 peer 依赖的方式，通过 `ctx.get('agentPresets')` 触达 roster——这正是它对 `sandboxPolicy` 与 `approval` 已在使用的、有明确文档的机会性消费模式。

-把父方的工具交给子 agent 之后，暴露出同一次 agent 平面搬迁引入的第二个缺陷：`ToolRuntime` 把**作用域级**注册排除在限制之外、只过滤全局层，因此当所有面向模型的行都变成祖先贡献之后，子 agent 的 `toolFilter` 就不再约束任何东西——而且全局层为空时，`restrict()` 会把收到的每个名字都判为未知并直接让子 agent 创建失败。豁免集合应当是作用域**自己注册**的工具，而不是恰好位于全局层的工具；后一种读法只在这两个集合重合时才成立。`view()` 现在过滤作用域继承来的一切——全局层与每个祖先层——只豁免它自己那层。这条自身层豁免是承重的而非顺带的：委派运行时把子 agent 的 `report` 与结构化输出工具注册进子 agent 自己那层，而一个只点名子 agent 可用能力的过滤器绝不能把它回报所依赖的机制一并剥掉。
+把父方的工具交给子 agent 之后，暴露出同一次 agent 平面搬迁引入的第二个缺陷：`ToolRuntime` 把**作用域级**注册排除在限制之外、只过滤全局层，因此当所有面向模型的行都变成祖先贡献之后，子 agent 的 `toolFilter` 就不再约束任何东西——而且全局层为空时，`restrict()` 会把收到的每个名字都判为未知并直接让子 agent 创建失败。豁免集合应当是作用域**自己注册**的工具，而不是恰好位于全局层的工具；后一种读法只在这两个集合重合时才成立。`view()` 现在过滤作用域继承来的一切——全局层与每个祖先层——只豁免它自己那层。这条自身层豁免是承重的而非顺带的：委派运行时把子 agent 的结构化输出工具注册进子 agent 自己那层，而一个只点名子 agent 可用能力的过滤器绝不能把它产出请求结果所依赖的机制一并剥掉。

 ## 考虑过的替代方案

@@ -30,11 +30,11 @@ Status: implemented

 **把子 agent 的 key 绑到**父方的** key 而不是常驻挂载上。** 否决，因为这改变了子 agent 继承的内容：父方自己的 scope 层携带其逐 agent 限制，那些限制会就此与每个后代求交，而活得比父方久的子 agent 会挂在一个已 dispose 的 agent key 上。加入常驻挂载给到子 agent 的是父方的组装，仅此而已。

-**扩展可继续 activation setup 注册表以覆盖一次性子 agent。** 否决，因为该注册表的贡献类型是同步的 `(childCtx) => () => void` 并带有逐次安装的撤销，建模的是会来会走的部署能力，而 preset 加入是一次性认父、自身没有撤销可言。扩展它反而会让任何绕过该注册表的驱动重新具备遗漏的可能。
+**为两个驱动引入一份共享 child setup 注册表。** 否决，因为同步且可撤销的贡献建模的是会来会走的部署能力，而 preset 加入是一次性认父、自身没有撤销可言。让组合经由可选注册表完成，反而会让任何绕过它的驱动重新具备遗漏的可能。

 **让 `dsh-subagent` 导入 `resolveSessionPreset` 并按解析出的 id 挂载。** 否决，因为这会给一个必须在没有 roster 时也能工作的包引入硬模块边，而且最终仍落回上述的重新挂载语义。

-**过滤链上的每一层，包括作用域自身那层。** 否决，因为那会让逐子 agent 的能力过滤器把该子 agent 的回报与结构化输出工具一并删掉——它们由委派运行时注册进子 agent 自己那层——于是一个点名"子 agent 可用哪些能力"的 `allow` 会让它彻底无法回报。
+**过滤链上的每一层，包括作用域自身那层。** 否决，因为那会让逐子 agent 的能力过滤器删掉该子 agent 的结构化输出工具——它由委派运行时注册进子 agent 自己那层——于是一个点名“子 agent 可用哪些能力”的 `allow` 会让它无法产出所请求的结果。

 **只修活着的加入，不动持久化 header。** 否决，因为那样活着的子 agent 与冷读同一个子 agent 会对"哪份组装产出了这段历史"给出不同答案——同一类缺陷，只是被搬了个地方而不是被修掉。

@@ -0,0 +1,6 @@
+# Bilingual-pair consistency record (docs/i18n/README.md): the git blob hash of each
+# 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 .agents/notes/implemented/bug-fix/2026-08-17-subagent-message-settlement-ordering.md
+2026-08-17-subagent-message-settlement-ordering.md: cdc996643c84c5f50a3bd1836e82645660dc8c57
+2026-08-17-subagent-message-settlement-ordering.zh.md: 1143da1560e4969dcc4f6a0c6d5ca18060b56191
@@ -0,0 +1,44 @@
+# Agent Note: Child Agent messages precede their settlement notices
+
+Status: implemented
+
+English | [中文](2026-08-17-subagent-message-settlement-ordering.zh.md)
+
+## Problem
+
+A continuable child can send selected content and later produce an unconditional manager-authored settlement notice. If those two messages enter queues with different claim priority, the later settlement notice can reach the parent model before the earlier child message. The first step of a turn claims the complete `next-step` batch before one `next-turn` message, so mixing a FIFO later-turn send with a next-step settlement reverses causal order. [Issue #2600](https://github.com/deepseek-harness/deepseek-harness/issues/2600) records the defect.
+
+The child instruction says to send a finding whenever it changes what the parent should do next. Deferring that message to a later turn contradicts its scheduling meaning and separates causally ordered messages across queues with different claim priority.
+
+## Decision
+
+Every model-authored adjacent-Agent message uses fixed Steer delivery through `SubagentRuntime.sendMessage()`. A running parent reads the child message at its nearest safe step boundary and an idle parent starts a turn. There is no quiet or next-turn model delivery option.
+
+The continuation manager retains `sendWaking()` and `admitWaking()` around messages delivered to resident continuable parents. Their purpose is waking-send admission accounting: the receiving Activation remains live between synchronous inbox insertion and the microtask that observes the wake.
+
+### Ordering across parent states
+
+A running parent receives an accepted child message and the child's later settlement notice in the same `next-step` FIFO. If the parent becomes idle before settlement arrives, it has already claimed the child message; settlement may then open a later turn without reversing observed order.
+
+During parent maintenance, the child message occupies `next-step` and latches a wake, while settlement may occupy `next-turn` because maintenance reports idle status. The initial claim still takes next-step input before the queued turn. Waking input submitted after cancellation follows the core Agent's cancellation convergence rather than bypassing it.
+
+### Verification
+
+The control-tool suite holds a parent inside an active model request, submits child messages, settles the child, and verifies sender identity, Steer admission, FIFO batching, and preservation after settlement. Continuation coverage pins waking admission accounting for a resident continuable parent and keeps the runtime-owned settlement source distinct from `agent-message`.
+
+The keyless continuable-subagent snapshot uses the shipped fixed delivery. Its child-visible tool schema is the same as the parent's, and the accepted child message precedes the later settlement notice without a scheduling overlay.
+
+## Alternatives considered
+
+**Offer quiet delivery.** A quiet message can remain unread after an idle parent parks. It also gives equivalent model-authored messages different liveness semantics and reopens deployment-dependent ordering.
+
+**Offer next-turn delivery.** A later next-step settlement notice can still overtake it. Preserving message-before-settlement would require a cross-queue ordering barrier, and no current model operation requires later-turn isolation strongly enough to own that mechanism.
+
+**Move settlement notices to `next-turn`.** Settlement batching uses the next-step queue so several children finishing together cost one parent step instead of one turn each. Moving settlement would increase latency and model work to retain an unnecessary message scheduling mode.
+
+## Consequences
+
+- A child message may extend an open parent turn. It never interrupts the active model request or tool execution; the agent loop admits it only at a step boundary.
+- Messages accepted together share one next-step batch, preserving FIFO order and limiting turn amplification.
+- Model callers cannot choose a delivery mode, so ordering and wake behavior do not vary by deployment or call.
+- A child-to-parent send still requires the direct parent to remain live; the service provides no durable parent mailbox.
@@ -0,0 +1,44 @@
+# Agent Note: Child Agent 消息先于其结算通知
+
+Status: implemented
+
+[English](2026-08-17-subagent-message-settlement-ordering.md) | 中文
+
+## 问题
+
+可继续 child 可以发送选中内容，之后还会产生一条由管理器编写且无条件投递的结算通知。如果这两条消息进入领取优先级不同的队列，较晚的结算通知可能先于较早的 child 消息到达 parent 模型。一个轮次的第一个 step 会先领取完整 `next-step` 批次，再领取一条 `next-turn` 消息，因此混用 FIFO 后续轮次发送与 next-step 结算会颠倒因果顺序。[Issue #2600](https://github.com/deepseek-harness/deepseek-harness/issues/2600)记录了该缺陷。
+
+child 指令要求在发现会改变 parent 下一步动作时发送该发现。把这条消息推迟到后续轮次既违背其调度含义，也会把具有因果顺序的消息拆到领取优先级不同的队列。
+
+## 决策
+
+每条模型编写的相邻 Agent 消息都通过 `SubagentRuntime.sendMessage()` 使用固定 Steer 投递。运行中的 parent 在最近安全 step 边界读取 child 消息，空闲 parent 则启动一个轮次。模型没有静默或 next-turn 投递选项。
+
+继续执行管理器在投递到驻留可继续 parent 的消息周围保留 `sendWaking()` 与 `admitWaking()`。它们负责唤醒发送准入记账：接收方 Activation 会在同步 inbox 插入与观察到唤醒的微任务之间保持在线。
+
+### 不同 parent 状态下的顺序
+
+运行中的 parent 在同一条 `next-step` FIFO 中接收已接受的 child 消息与该 child 随后的结算通知。如果 parent 在结算到达前变为空闲，它已经领取 child 消息；结算随后可以开启后续轮次，而不会颠倒观察顺序。
+
+parent 处于 maintenance 时，child 消息占用 `next-step` 并锁存一次唤醒，而结算可能因 maintenance 报告空闲状态而占用 `next-turn`。初始领取仍会先取 next-step 输入，再取排队轮次。取消后提交的唤醒输入遵循核心 Agent 的取消收敛，而不会绕过它。
+
+### 验证
+
+控制工具测试套件让 parent 保持在活跃模型请求中，提交 child 消息、结算 child，并验证 sender 身份、Steer 准入、FIFO 批处理与结算后保留。继续执行覆盖固定驻留可继续 parent 的唤醒准入记账，并让 runtime 所有的结算来源与 `agent-message` 保持不同。
+
+无密钥可继续 subagent 快照使用随附的固定投递。其 child 可见工具 schema 与 parent 相同，且已接受的 child 消息先于之后的结算通知，无需调度 overlay。
+
+## 考虑过的替代方案
+
+**提供静默投递。** 空闲 parent 停驻后可能永远不读取静默消息。它还会让等价的模型编写消息具有不同存活语义，并重新引入依赖部署的顺序。
+
+**提供 next-turn 投递。** 较晚的 next-step 结算通知仍可能越过它。保持消息先于结算需要跨队列顺序屏障，而当前没有模型操作对后续轮次隔离的需求强到足以承担该机制。
+
+**把结算通知移到 `next-turn`。** 结算批处理使用 next-step 队列，使多个 child 同时结束只消耗 parent 的一个 step，而不是每个 child 一个轮次。移动结算会为了保留不必要的消息调度模式而增加延迟与模型工作。
+
+## 后果
+
+- child 消息可以延长开放的 parent 轮次。它绝不会中断活跃模型请求或工具执行；agent loop 只在 step 边界接纳它。
+- 一起被接受的消息共享一个 next-step 批次，保持 FIFO 顺序并限制轮次放大。
+- 模型调用方不能选择投递模式，因此顺序与唤醒行为不会随部署或调用而变化。
+- child 到 parent 的发送仍要求直接 parent 保持在线；服务不提供持久 parent mailbox。
@@ -1,44 +0,0 @@
-# Agent Note: Subagent reports precede their settlement notices
-
-Status: implemented
-
-English | [中文](2026-08-17-subagent-report-settlement-ordering.zh.md)
-
-## Problem
-
-A continuable child can explicitly report selected content and later produce an unconditional manager-authored settlement notice. Report delivery used `Agent.followup()` and entered the parent's `next-turn` queue, while settlement delivery to a running parent used `Agent.steer()` and entered `next-step`. The first step of a turn claims the complete `next-step` batch before one `next-turn` message, so the later settlement notice could reach the model before the earlier report. The assembled report scenario required `reportDelivery: quiet` to avoid that nondeterministic interleaving. [Issue #2600](https://github.com/deepseek-harness/deepseek-harness/issues/2600) records the defect.
-
-The report tool tells a child to report whenever a finding changes what its parent should do next. Deferring that message to a later turn contradicted the tool's scheduling meaning and separated causally ordered messages across queues with different claim priority.
-
-## Decision
-
-`SubagentReportDelivery` is `'quiet' | 'next-step'`, and `next-step` is the default. Next-step delivery calls `parent.steer()`, so a running parent reads the report at its nearest safe step boundary and an idle parent starts a turn. Quiet delivery continues to call `parent.inject()` and enters the same queue without waking an idle parent.
-
-The continuation manager retains `sendWaking()` and `admitWaking()` around next-step reports delivered to resident continuable parents. Their purpose is waking-send admission accounting, independent of whether the message targets a step or a turn: the receiving Activation remains live between synchronous inbox insertion and the microtask that observes the wake.
-
-### Ordering across parent states
-
-A running parent receives an accepted report and the child's later settlement notice in the same `next-step` FIFO. If the parent becomes idle before settlement arrives, it has already claimed the report; settlement may then open a later turn without reversing the observed order.
-
-During parent maintenance, the report occupies `next-step` and latches a wake, while settlement may occupy `next-turn` because maintenance reports idle status. The initial claim still takes next-step input before the queued turn. Waking input submitted after cancellation is redirected by `Agent.send()` to `next-turn`, so report and settlement follow the core agent's cancellation convergence rather than bypassing it.
-
-### Verification
-
-The report package holds a parent inside an active model request, submits a child report, settles that child, and asserts the pending parent batch is ordered `subagent-report`, then `subagent-settled`, with no queued later turn. Separate coverage pins repeated reports as one FIFO next-step batch, idle-parent wakeup, and waking admission accounting for a continuable parent.
-
-The assembled ACP report scenario uses the shipped default. Its scheduling fence keeps the child behind the parent's delegation turn and holds the parent in maintenance until settlement follows the report. The report latches the wake while the settlement notice queues a turn; when maintenance ends, the parent claims next-step input before next-turn input and observes both notices in causal order without a quiet-delivery overlay.
-
-## Alternatives considered
-
-**Keep the `wakeup` name but change its implementation to `steer()`.** The existing public description defined `wakeup` as one later parent turn. Reusing the value for a different inbox target would leave configuration unable to state the behavior it selects. The pre-release configuration instead names `next-step` directly.
-
-**Expose `quiet | next-step | next-turn`.** A next-turn report still permits a later next-step settlement notice to overtake it. Preserving report-before-settlement would require a cross-queue ordering barrier, and no current deployment requires next-turn isolation strongly enough to own that mechanism.
-
-**Move settlement notices to `next-turn`.** Settlement batching deliberately uses the next-step queue so several children finishing together cost one parent step instead of one turn each. Moving settlement would increase latency and model work to retain a report scheduling mode with no current consumer.
-
-## Consequences
-
- A report may extend an open parent turn. It never interrupts the active model request or tool execution; the agent loop admits it only at a step boundary.
- Reports accepted together share one next-step batch, preserving FIFO order and reducing the turn amplification of the former one-turn-per-report behavior.
- The `wakeup` configuration value is rejected rather than retained as an alias. This repository has no external pre-release compatibility promise for Cordis configuration.
- `quiet` remains the deployment escape for reports that must not wake a parked parent, with the existing risk that no model reads them until another waking input arrives.
@@ -1,44 +0,0 @@
-# Agent Note: Subagent report 先于其结算通知
-
-Status: implemented
-
-[English](2026-08-17-subagent-report-settlement-ordering.md) | 中文
-
-## 问题
-
-可继续 child 可以显式上报选中内容，之后还会产生一条由管理器撰写且无条件投递的结算通知。报告投递曾使用 `Agent.followup()` 并进入 parent 的 `next-turn` 队列，而面向运行中 parent 的结算投递使用 `Agent.steer()` 并进入 `next-step`。一个轮次的第一个 step 会先领取完整 `next-step` 批次，再领取一条 `next-turn` 消息，因此较晚的结算通知可能先于较早的报告到达模型。整体组装的报告场景必须使用 `reportDelivery: quiet`，才能避开这种不确定交错。[Issue #2600](https://github.com/deepseek-harness/deepseek-harness/issues/2600)记录了该缺陷。
-
-report 工具要求 child 在发现会改变 parent 下一步动作的信息时上报。把这条消息推迟到后续轮次，既违背了工具的调度含义，也让具有因果顺序的消息分散到领取优先级不同的队列中。
-
-## 决策
-
-`SubagentReportDelivery` 为 `'quiet' | 'next-step'`，默认值为 `next-step`。Next-step 投递调用 `parent.steer()`，因此运行中的 parent 会在最近的安全 step 边界读取报告，空闲 parent 则会启动一个轮次。静默投递继续调用 `parent.inject()`，进入同一队列但不唤醒空闲 parent。
-
-对于投递到驻留可继续 parent 的 next-step 报告，继续执行管理器会保留外围的 `sendWaking()` 与 `admitWaking()`。它们负责唤醒发送的准入记账，与消息面向 step 还是 turn 无关：接收方 Activation 在同步插入 inbox 与观察该唤醒的微任务之间保持在线。
-
-### 不同 parent 状态下的顺序
-
-运行中的 parent 会在同一个 `next-step` FIFO 中接收已接受的报告和该 child 稍后的结算通知。若 parent 在结算到达前变为空闲，它已经领取了报告；结算随后可以开启一个更晚的轮次，而不会反转观察顺序。
-
-parent 处于 maintenance 时，报告占据 `next-step` 并锁存一次唤醒，而结算可能因为 maintenance 呈现空闲状态而占据 `next-turn`。首次领取仍会先取 next-step 输入，再取排队轮次。取消后提交的唤醒输入会由 `Agent.send()` 重定向到 `next-turn`，因此报告和结算会遵循核心 agent 的取消收敛，而不会绕过它。
-
-### 验证
-
-report 包把 parent 保持在一个活动模型请求中，提交 child 报告，再让该 child 结算，并断言等待中的 parent 批次按 `subagent-report`、`subagent-settled` 排序，且没有排队的后续轮次。独立覆盖还会固定重复报告形成一个 FIFO next-step 批次、空闲 parent 唤醒，以及可继续 parent 的唤醒准入记账。
-
-整体组装的 ACP 报告场景使用随附默认值。调度围栏让 child 等到 parent 的委派轮次之后，并让 parent 保持 maintenance，直至结算跟在报告之后到达。报告会锁存唤醒，结算通知则排入后续轮次；maintenance 结束时，parent 先领取 next-step 输入、再领取 next-turn 输入，因此无需静默投递 overlay 也能按因果顺序观察两条通知。
-
-## 备选方案
-
-**保留 `wakeup` 名称，但把其实现改为 `steer()`。** 既有公开描述把 `wakeup` 定义为一个后续 parent 轮次。让该值复用于不同的 inbox 目标，会使配置无法准确说明自己选择的行为。预发布配置因此直接使用 `next-step` 名称。
-
-**暴露 `quiet | next-step | next-turn`。** Next-turn 报告仍可能被稍后的 next-step 结算通知超越。要保住报告先于结算，需要跨队列顺序屏障；当前没有任何部署对 next-turn 隔离的需求强到足以承担该机制。
-
-**把结算通知移到 `next-turn`。** 结算批处理刻意使用 next-step 队列，使多个一起结束的 child 只花费 parent 的一个 step，而不是各自一个轮次。移动结算会增加延迟和模型工作量，只为保留一个没有当前消费方的报告调度模式。
-
-## 后果
-
- 报告可能延长已打开的 parent 轮次。它绝不会打断活动模型请求或工具执行；agent loop 只会在 step 边界准入它。
- 一起接受的报告会共享一个 next-step 批次，保持 FIFO 顺序，并减少原先每份报告各占一个轮次所造成的轮次放大。
- `wakeup` 配置值会被拒绝，而不是保留为别名。本仓库对预发布 Cordis 配置不作外部兼容承诺。
- 对于不得唤醒停驻 parent 的报告，`quiet` 仍是部署退路，同时保留既有风险：在另一条唤醒输入到达之前，没有模型会读取这些报告。
@@ -0,0 +1,6 @@
+# Bilingual-pair consistency record (docs/i18n/README.md): the git blob hash of each
+# 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 .agents/notes/implemented/bug-fix/2026-08-29-code-runtime-python-call-backlog-and-binding-metadata-snapshot.md
+2026-08-29-code-runtime-python-call-backlog-and-binding-metadata-snapshot.md: afdad301a4853754184b75668d167e71420c2480
+2026-08-29-code-runtime-python-call-backlog-and-binding-metadata-snapshot.zh.md: 48d6e0748979b09053aa51a94788fdd95d997179
@@ -0,0 +1,70 @@
+# Agent Note: Bound in-flight binding calls, snapshot binding metadata, compact the reply queue, and meter wide completions with cursors in the CPython backend
+
+Status: implemented
+
+English | [中文](2026-08-29-code-runtime-python-call-backlog-and-binding-metadata-snapshot.zh.md)
+
+## Problem
+
+A further review round on the CPython subprocess backend (packages/experimental/code-runtime-python) surfaced seven findings on the binding-dispatch, validation, completion-metering, and frame-parse paths. First, the reply-backlog cap counts only RESOLVED calls — `pendingReplies` grows after the binding's `await` resolves — so a child flooding calls against a binding whose promise never settles accumulates one async closure per frame until the wall clock without ever tripping the cap. Second, `validateBindings` reads `errorClass.name`, `errorClass.memberNameProperty`, and `namespace.global` several times and retains the original errorClass object for the boot frame, whose `JSON.stringify` re-reads it after validation: a getter that returns a valid value during validation and then throws or returns a conflicting value at stringify time turns the seam-misuse rejection into a worker-exit, or injects a different name than validation approved. Third, `replyQueue` never shrinks mid-drain: the drain loop clears consumed slots to `undefined` but leaves `length` (and the backing store) growing, so a child that reads replies just fast enough to keep the drain alive but never empty grows the array linearly with cumulative throughput. Fourth, the completion meter `checkDoneValue` pushes every member of an open container onto an explicit work stack, so a wide completion value near the frame cap (millions of members) copies that many references onto the stack — O(width) auxiliary memory on top of the already-parsed value — OOMing the host after the parse succeeded. Fifth, a done frame processed in the SAME data event as more than 1024 call frames settles the run before the post-macrotask call-backlog check runs (which no-ops once settled), so a child could finish successfully while leaving the outstanding closures behind. Sixth, the child's `_dump_string` folds a spelled-out surrogate pair into its astral code point, so two DIFFERENT Python dict keys — `"\ud83d\ude00"` and `"\U0001f600"` — encode to the SAME JSON member and the host's `JSON.parse` silently drops one of them, violating the lossless-JSON promise for completions and binding arguments. Seventh, the load gate bounds the CHILD's build-and-encode under `RLIMIT_AS` but not the HOST's `JSON.parse`: a legitimately configured wide completion near the frame cap (e.g. a 3-million-key dict under a 50 MiB budget) materializes several times its raw bytes in the host's property storage, so a constrained host heap (e.g. `--max-old-space-size=256`) dies with a process-level OOM during the parse — before `checkDoneValue` (which only sees the already-parsed value) could reject it.
+
+## Decision
+
+### In-flight binding calls are capped at 1024, checked once per macrotask after the microtasks drain
+
+`case 'call'` counts the outstanding binding calls before dispatch (`pendingCalls`) and releases the slot in the async body's `finally`, covering the reply-written, resolution-rejected, and settled-drop exits. The data handler schedules ONE post-batch check per macrotask via `setImmediate` (deduped by a flag): it runs after the current macrotask's microtasks, so it sees the TRUE outstanding count — the live count is inflated by the batch's own frames (the finallys have not run yet), and a per-event snapshot is stale when flowing mode fires several `data` events within one macrotask before any microtask drains. When the count passes `MAX_PENDING_REPLIES` — strictly greater, so exactly 1024 outstanding calls are allowed — the run settles as a `worker-exit` with a call-backlog message. The check no-ops once `settled`, so a `done` or `log` frame wins over the cap: a program that returns with binding calls it started but never awaited still completes with its value. The `done` handler independently re-checks the count before accepting the frame, closing the window where a done in the SAME batch as a flood would settle the run before the post-macrotask check could fire. This is a count bound, not a byte bound.
+
+### Binding metadata is snapshotted into plain values before validation and the boot frame
+
+`validateBindings` reads `namespace.global`, `errorClass.name`, and `errorClass.memberNameProperty` each exactly once into a plain local, validates the copies, and stores a plain `{ name, memberNameProperty }` object in the bindings map. The boot frame serializes that stored copy, so validation and the boot frame see identical values regardless of getter state; a stateful getter cannot change or throw between the two stages.
+
+### The reply queue compacts its consumed prefix mid-drain
+
+`drainReplies` compacts the consumed prefix (`replyQueue.splice(0, head); head = 0`) once `head` reaches `MAX_PENDING_REPLIES`. The splice is O(head) once per bound of consumed frames — amortized O(1) per reply — bounding the backing store to O(backlog + bound) for a drain that never empties.
+
+### The completion meter walks wide values with one cursor per nesting level
+
+`checkDoneValue` now holds one cursor per OPEN container (a values iterator for the root and arrays, an entries iterator for objects whose key escapes are metered when the entry is reached), the same shape `hasNonLosslessNumber` and the child's `_check_done_value` already use. The byte budget still bounds the walk: each member is metered as its cursor yields it, and the width lower-bound checks bail an over-budget container before the cursor descends. The auxiliary state is O(depth), not O(width), so a wide completion near the frame cap meters exactly instead of copying millions of references. `encodeJsonPlain` keeps its per-container task stack, which is O(width) but holds only references while the encoded output is itself O(total bytes) — same-order as its result, so the exemption is documented in its comment.
+
+### The frame parse cap is bounded by the host's heap
+
+The raw-byte frame cap does not protect the host process: `JSON.parse` of a wide-object frame materializes several times the raw bytes in property storage. The WORST shape is a dict of many short unique keys, which forces V8's dictionary-mode property storage plus one interned string per key — measured 6.4x for a 3,000,000-key frame (~31 MB raw) on a 1 GiB heap, trending up with key count (a flat unique-key array is ~4x, a repeated-key dict ~3x); a 256 MiB heap OOMs on that frame outright. The effective cap each instance enforces is `min(protocol cap, floor((heap_size_limit - HOST_PARSE_BASELINE_BYTES) / HOST_PARSE_WORST_CASE_MULTIPLE))` with a 16x multiple — ~2.5x over the measured worst shape — derived from the host's configured heap limit (`--max-old-space-size` honored via `v8.getHeapStatistics().heap_size_limit`). A default Node heap (~4 GiB) never binds; a constrained host lowers the cap and the load gate rejects any budget whose frame could not be parsed safely, failing loud at load instead of OOMing the host mid-parse. The child's `RLIMIT_AS` gate is a separate resource and stays unchanged.
+
+### Dict keys that fold to one JSON member are rejected as non-lossless
+
+The child's `_dump_string` folds a spelled-out surrogate pair into its astral code point so the host's UTF-16 strings (where the two code units and the single character are the SAME string) meter at the same cost. Python can hold both spellings as distinct keys, so a dict containing `"\ud83d\ude00"` and `"\U0001f600"` would emit two members with the same JSON key and the host's `JSON.parse` would silently drop one. Both lossless-JSON walks (`_lossless_json_violation` for binding arguments, `_check_done_value` for completions) now track each dict's combined keys in a per-dict seen-set — O(keys), the same order as the dict itself — and reject a collision as non-lossless before any encoding.
+
+## Testing
+
+- `tests/runtime.spec.ts` — a hostile child floods 5000 sequential calls against a binding that never settles (`await new Promise(() => {})`); the run settles as `worker-exit` with the call-backlog message long before `maxWallMs`. Verified fail-before: without the cap the run times out at the wall clock.
+- `tests/runtime.spec.ts` — a legitimate `asyncio.gather` of 1025 instant calls completes with all 1025 results: the post-macrotask check sees the count after the finallys drained, where a per-frame check could trip on the 1025th frame of a single 64 KiB read.
+- `tests/runtime.spec.ts` — a program that schedules 1024 slow bindings (still pending) and returns `"done"` completes with its value: the check no-ops once the done frame settles the run, and the strict threshold allows exactly 1024 outstanding calls. Verified fail-before: an unconditional event-boundary check failed this exact case.
+- `tests/runtime.spec.ts` — a single 62 KiB write of 1025 compact calls against a never-settling binding settles as `worker-exit` long before `maxWallMs`, even though no further frames ever arrive: the per-macrotask check fires after the batch. Verified fail-before: a per-event admission snapshot never re-checks without further frames and the run waited out the wall clock.
+- `tests/runtime.spec.ts` — a single write of 1025 compact calls PLUS a done frame in the same batch settles as `worker-exit`: the done handler re-checks the count before accepting the frame, where the post-macrotask check would no-op after the done settled the run. Verified fail-before: without the done re-check the run completed successfully with the outstanding closures left behind.
+- `tests/runtime.spec.ts` — a burst of 1300 instant calls whose frames split across pipe reads completes with all results: the check runs after all of a macrotask's finallys, where a per-event snapshot could see a stale in-flight count when flowing mode fires several events before any microtask drains.
+- `tests/runtime.spec.ts` — a completion value and binding arguments whose dict contains both `"\ud83d\ude00"` and `"\U0001f600"` as keys are rejected as non-lossless (invalid-output / a lossless-JSON call rejection): the two spellings fold to one JSON member, which the host's JSON.parse would silently collapse. Verified fail-before: without the collision check both round-tripped with one key dropped.
+- `tests/protocol.spec.ts` — `hostFrameParseCeiling` derives the effective parse cap from a simulated heap: the protocol cap binds on a default heap, a ~304 MiB host limit yields a 15 MiB cap, and a tiny heap leaves almost no parse room.
+- `tests/runtime.spec.ts` — a child node with a 128 MiB old space rejects a 50 MiB completion budget at load (`maxValueBytes must not exceed`), where the address-space gate alone would admit it. Verified fail-before: with the heap bound ignored the budget loaded.
+- `tests/runtime.spec.ts` — a child node with a 128 MiB old space builds a wide-unique-key dict whose frame is AT the derived cap and parses it, surviving. Verified fail-before: with the parse multiple at 8 the derived cap doubles and the same subprocess OOMs during the parse.
+- The suite's temp fixtures (`dsh-bad-bin-`, `dsh-fake-bin-`, `dsh-rlimit-*`, `dsh-staging-`, heartbeat dirs, wrapper scripts) are now registered and removed after each test, so repeated runs do not accumulate `dsh-*` artifacts in the shared tmpdir.
+- Two namespace-shape tests — `errorClass.name`/`errorClass.memberNameProperty` and `namespace.global` exposed through getters that throw or change on a second read; the run boots and completes, and each field is read exactly once (asserted). Verified fail-before: without the snapshot, the errorClass getter threw inside validation and the global getter injected a different name, failing the program with `NameError`.
+- `tests/runtime.spec.ts` — a child floods calls whose replies exceed the writable high-water mark, blocking the first drain write; the resumed drain consumes a backlog past the compaction bound while a second wave of calls is still pending, and the child reads fd 3 itself (blocking the reply pump) to verify all 1524 replies arrive. No fixed sleep: the child's reads pace at the drain's delivery rate, and the host finishes pushing a wave within milliseconds, so the queue is always full at the splice; newlines are counted per chunk (each reply carries exactly one), never by re-scanning the accumulated total, which would be O(n²). Verified fail-before: a splice that removed pending frames dropped the second wave and the run hung to the wall clock.
+- `tests/protocol.spec.ts` — a 2,000,000-element array and a 100,000-key object meter at their exact serialized size, reject one byte under, and still find a `-0` tail element, pinning the cursor walk's breadth behavior.
+
+## Alternatives considered
+
+**Pause the fd-3 read side instead of counting in-flight calls.** Rejected: pausing reads would also stall processing of `done` and `log` frames the child may send after its last call, changing settlement timing; a count cap is deterministic and matches the existing frame-cap pattern.
+
+**Check the in-flight count per frame.** Rejected: the finallys run on the microtask queue, which drains only when the macrotask ends, so a single event carrying more than `MAX_PENDING_REPLIES` legitimate call frames would trip a per-frame check even though every binding settled immediately.
+
+**Check the count at call admission against a per-event snapshot.** Rejected twice: an unconditional event-boundary check reclassifies a `done` frame as worker-exit when a program returns with calls it never awaited, and a snapshot that refreshes per `data` event is stale when flowing mode fires several events within one macrotask (a legitimate burst whose second chunk carries more in-flight calls than the cap would be killed). Checking the true count once per macrotask, after the microtasks drain, is chunking-independent on both axes, and the strict threshold lets exactly `MAX_PENDING_REPLIES` outstanding calls complete normally.
+
+**Read metadata once but keep the original errorClass object.** Rejected: the boot frame's `JSON.stringify` re-invokes the getters; only a plain stored copy guarantees both stages read the same values.
+
+**Rely on the drain's `finally` reset for queue memory.** Rejected: the reset runs only when the drain ends; a drain that never empties keeps growing. Mid-drain compaction bounds the backing store while the drain is alive.
+
+**Keep the completion meter's explicit member stack.** Rejected: the byte budget bounds the WALK but not the stack's reference count, which is O(width) — a wide value near the frame cap copies millions of references and can OOM the host after the parse succeeded; the per-level cursor shape keeps O(depth) state.
+
+## Consequences
+
+In-flight binding closures are bounded like the reply backlog, so a child flooding calls against a never-settling binding fails the run early instead of accumulating closures until the wall clock, while a legitimate large concurrent gather is unaffected (the cap is checked after the microtask queue drains). The boot frame serializes exactly the metadata validation approved, regardless of getter state. The reply queue's backing store stays bounded during sustained partial drains; the compaction is internal memory hygiene with no observable behavior change. The completion meter keeps its exact byte accounting with O(depth) auxiliary state, so a wide completion value meters without a host OOM.
@@ -0,0 +1,70 @@
+# Agent Note: 在 CPython 后端限制在途 binding 调用、快照 binding 元数据、压缩回复队列并用游标计量宽完成值
+
+Status: implemented
+
+[English](2026-08-29-code-runtime-python-call-backlog-and-binding-metadata-snapshot.md) | 中文
+
+## Problem
+
+对 CPython 子进程后端（packages/experimental/code-runtime-python）的又一轮评审在 binding 分发、校验、完成值计量与帧解析路径上浮出七项发现。其一，回复积压上限只计数已解析的调用——`pendingReplies` 在 binding 的 `await` 解析后才增长——因此向 promise 永不结算的 binding 洪泛调用的子进程会每个帧累积一个异步闭包直到墙钟，却始终不触发该上限。其二，`validateBindings` 多次读取 `errorClass.name`、`errorClass.memberNameProperty` 与 `namespace.global`，并把原始 errorClass 对象保留到引导帧，其 `JSON.stringify` 在校验后重读该对象：getter 在校验时返回合法值、在序列化时抛错或返回冲突值，会把 seam 误用拒绝变成 worker-exit，或注入一个未经校验批准的名字。其三，`replyQueue` 在排空进行中从不收缩：排空循环把已消费槽位清成 `undefined`，但 `length`（及其后备存储）继续增长，因此以恰好能让排空持续存活却永不排空的速率读取回复的子进程，会让数组随累计吞吐量线性增长。其四，完成值计量器 `checkDoneValue` 把开放容器的每个成员压入显式工作栈，因此接近帧上限的宽完成值（数百万成员）会把同等数量的引用复制进栈——在已解析值之上再占 O(width) 辅助内存——在解析成功后 OOM 终止宿主。其五，与超过 1024 个调用帧处于同一 data 事件的 done 帧会在批后调用积压检查运行之前结算运行（该检查在 settled 后为空操作），因此子进程可以「成功」完成却留下未结算闭包。其六，子端 `_dump_string` 把拼写出来的代理项对折叠成星面码点，因此两个不同的 Python 字典键——`"\ud83d\ude00"` 与 `"\U0001f600"`——编码成同一个 JSON 成员，宿主的 `JSON.parse` 会静默丢弃其一，违反完成值与 binding 参数的 lossless JSON 承诺。其七，加载门限制的是子端在 `RLIMIT_AS` 下的构建与编码，而非宿主的 `JSON.parse`：合法配置的接近帧上限宽完成值（如 50 MiB 预算下的 300 万键字典）会在宿主属性存储中物化其原始字节的若干倍，因此受限堆宿主（如 `--max-old-space-size=256`）会在解析期间以进程级 OOM 死亡——早于 `checkDoneValue`（它只能看到已解析值）拒绝它。
+
+## Decision
+
+### 在途 binding 调用限制为 1024，每个宏任务在微任务排空后检查一次
+
+`case 'call'` 在分发前对在途 binding 调用计数（`pendingCalls`），并在异步体的 `finally` 中释放槽位，覆盖回复已写入、解析被拒绝与结算后丢弃三种出口。data 处理器经 `setImmediate`（用标志去重）为每个宏任务调度一次批后检查：它在该宏任务的微任务之后运行，因此看到的是真实在途计数——实时计数被本批自身帧抬高（`finally` 尚未运行），而按 `data` 事件刷新的快照在 flowing 模式下同一宏任务内连续发出多个事件（微任务尚未排空）时会过期。计数**严格大于** `MAX_PENDING_REPLIES`（恰好 1024 个在途调用允许）时，运行以带 call-backlog 消息的 `worker-exit` 结算。检查在 `settled` 后为空操作，因此 `done` 或 `log` 帧优先于上限：启动了 binding 调用却未 await 就返回的程序仍以其值正常完成。`done` 处理器在接受帧之前独立复查计数，封住与洪泛同批的 done 会在批后检查触发前结算运行的窗口。这是计数上限而非字节上限。
+
+### binding 元数据在校验与引导帧之前快照为纯值
+
+`validateBindings` 把 `namespace.global`、`errorClass.name` 与 `errorClass.memberNameProperty` 各恰好读取一次到普通局部变量，对副本做校验，并在 bindings 映射中存入普通 `{ name, memberNameProperty }` 对象。引导帧序列化该存储副本，因此无论 getter 处于何种状态，校验与引导帧看到的都是相同的值；有状态的 getter 无法在两个阶段之间改变或抛错。
+
+### 回复队列在排空进行中压缩已消费前缀
+
+`drainReplies` 在 `head` 达到 `MAX_PENDING_REPLIES` 时压缩已消费前缀（`replyQueue.splice(0, head); head = 0`）。该 splice 为 O(head)，每消费一上限的帧执行一次——均摊到每条回复为 O(1)——使永不排空的排空把后备存储限制在 O(积压 + 上限)。
+
+### 完成值计量器每层持一个游标遍历宽值
+
+`checkDoneValue` 现在为每个开放容器持一个游标（根与数组用 values 迭代器，对象用 entries 迭代器——key 的转义字节在该 entry 到达时计量），与 `hasNonLosslessNumber` 及子端 `_check_done_value` 已用的形态一致。字节预算仍然限制遍历：每个成员在游标产出时计量，宽度下界检查会在游标下降之前拒绝超预算容器。辅助状态为 O(depth) 而非 O(width)，因此接近帧上限的宽完成值精确计量，而不是复制数百万引用。`encodeJsonPlain` 保留其每容器任务栈——该栈为 O(width) 但只持有引用，而编码输出本身即 O(total bytes)，与结果同量级，豁免已在注释中说明。
+
+### 帧解析上限受宿主堆约束
+
+原始字节帧上限并不保护宿主进程：`JSON.parse` 一个宽对象帧会在属性存储中物化其原始字节的若干倍。**最坏形态是大量短唯一键的字典**——迫使 V8 进入字典模式属性存储并为每个键内化一个字符串——1 GiB 堆上 3,000,000 键帧（约 31 MB 原始）实测 6.4 倍且随键数上升（平铺唯一键数组约 4 倍、重复键字典约 3 倍）；256 MiB 堆直接在该帧上 OOM。每个实例执行的有效上限为 `min(协议上限, floor((heap_size_limit - HOST_PARSE_BASELINE_BYTES) / HOST_PARSE_WORST_CASE_MULTIPLE))`，系数为 16——实测最坏形态的约 2.5 倍安全余量——由宿主配置的堆上限推导（`--max-old-space-size` 经 `v8.getHeapStatistics().heap_size_limit` 生效）。默认 Node 堆（约 4 GiB）永不收紧；受限宿主会降低上限，加载门拒绝任何帧无法被安全解析的预算，在加载期响亮失败而非在解析中途 OOM 宿主。子端的 `RLIMIT_AS` 门是另一资源，保持不变。
+
+### 折叠为同一 JSON 成员的字典键按非 lossless 拒绝
+
+子端 `_dump_string` 把拼写出来的代理项对折叠成星面码点，使宿主的 UTF-16 字符串（两个码元与单个字符是同一字符串）按相同成本计量。Python 可以把两种拼写作为不同键持有，因此包含 `"\ud83d\ude00"` 与 `"\U0001f600"` 的字典会发出两个同键成员，宿主的 `JSON.parse` 会静默丢弃其一。两条 lossless-JSON 遍历（binding 参数的 `_lossless_json_violation` 与完成值的 `_check_done_value`）现在用每字典 seen 集跟踪合并后的键——O(keys)，与字典本身同量级——在编码前把冲突判为非 lossless。
+
+## Testing
+
+- `tests/runtime.spec.ts`——敌意子进程向永不结算的 binding（`await new Promise(() => {})`）洪泛 5000 个连续调用；运行在远早于 `maxWallMs` 时以带 call-backlog 消息的 `worker-exit` 结算。已实测失败前置：没有该上限时运行在墙钟处超时。
+- `tests/runtime.spec.ts`——合法的 `asyncio.gather` 并发 1025 个即时调用并全部完成：批后检查看到的是 `finally` 排空后的计数，而逐帧检查可能被单次 64 KiB 读取中的第 1025 帧误触发。
+- `tests/runtime.spec.ts`——程序调度 1024 个慢 binding（仍未结算）并返回 `"done"` 时以其值正常完成：done 帧结算运行后检查为空操作，且严格阈值允许恰好 1024 个在途调用。已实测失败前置：无条件的事件边界检查恰好在该用例上失败。
+- `tests/runtime.spec.ts`——单次 62 KiB 写入的 1025 个紧凑调用对抗永不结算的 binding，在远早于 `maxWallMs` 时以 `worker-exit` 结算，即使之后不再有帧到达：每宏任务检查在该批之后触发。已实测失败前置：按事件刷新的接纳快照在没有后续帧时永不复查，运行等到墙钟。
+- `tests/runtime.spec.ts`——单次写入的 1025 个紧凑调用**外加同批 done 帧**以 `worker-exit` 结算：done 处理器在接受帧之前复查计数，而批后检查会在 done 结算运行后空操作。已实测失败前置：没有 done 复查时运行成功完成并留下未结算闭包。
+- `tests/runtime.spec.ts`——1300 个即时调用的突发（帧跨管道读取拆分）全部完成：检查在该宏任务的所有 `finally` 之后运行，而按事件快照在 flowing 模式同宏任务内多个事件、微任务未排空时会看到过期的在途计数。
+- `tests/runtime.spec.ts`——字典同时含 `"\ud83d\ude00"` 与 `"\U0001f600"` 两个键的完成值与 binding 参数按非 lossless 拒绝（invalid-output / lossless-JSON 调用拒绝）：两种拼写折叠为一个 JSON 成员，宿主的 JSON.parse 会静默折叠。已实测失败前置：没有碰撞检查时两者都以丢键 round-trip。
+- `tests/protocol.spec.ts`——`hostFrameParseCeiling` 从模拟堆推导有效解析上限：默认堆上协议上限约束，约 304 MiB 宿主上限得出 15 MiB 上限，极小堆几乎不留下解析空间。
+- `tests/runtime.spec.ts`——128 MiB old space 的子 node 在加载期拒绝 50 MiB 完成值预算（`maxValueBytes must not exceed`），而地址空间门单独会放行。已实测失败前置：忽略堆上限时该预算正常加载。
+- `tests/runtime.spec.ts`——128 MiB old space 的子 node 构造帧恰在推导上限处的宽唯一键字典并解析，存活。已实测失败前置：解析系数回退到 8 时推导上限翻倍，同一子进程在解析中 OOM。
+- 套件的临时 fixture（`dsh-bad-bin-`、`dsh-fake-bin-`、`dsh-rlimit-*`、`dsh-staging-`、heartbeat 目录、wrapper 脚本）现登记并在每个测试后移除，重复运行不再在共享 tmpdir 累积 `dsh-*` 工件。
+- 两个 namespace 形态测试——`errorClass.name`/`errorClass.memberNameProperty` 与 `namespace.global` 经由第二次读取即抛错或改变的 getter 暴露；运行正常引导并完成，且每个字段恰好读取一次（已断言）。已实测失败前置：没有快照时，errorClass getter 在校验内抛错，global getter 注入不同名字，程序以 `NameError` 失败。
+- `tests/runtime.spec.ts`——子进程洪泛回复超过可写高水位线的调用，阻塞第一次排空写入；恢复的排空在第二波调用仍待发时消费超过压缩上限的积压，子进程直接读取 fd 3（阻塞回复泵）验证全部 1524 条回复送达。无固定睡眠：子进程的读取以排空的投递速率节流，宿主在毫秒内完成一波推送，因此压缩点队列必然已满；换行按块计数（每条回复恰好一个），绝不重扫累计总量——那会是 O(n²)。已实测失败前置：移除待发帧的 splice 会丢掉第二波回复，运行挂到墙钟。
+- `tests/protocol.spec.ts`——2,000,000 元素数组与 100,000 键对象以精确序列化大小计量、少一个字节即拒绝，并仍能发现尾部的 `-0`，钉住游标遍历的广度行为。
+
+## Alternatives considered
+
+**暂停 fd-3 读侧而非计数在途调用。** 拒绝：暂停读取也会让子进程在最后一个调用后可能发送的 `done` 与 `log` 帧处理停滞，改变结算时机；计数上限是确定性的，且与既有帧上限模式一致。
+
+**逐帧检查在途计数。** 拒绝：`finally` 在微任务队列上运行，微任务只在宏任务结束时排空，因此单个事件携带超过 `MAX_PENDING_REPLIES` 个合法调用帧时，即使每个 binding 都立即结算，逐帧检查也会误触发。
+
+**在调用接纳处对照按事件刷新的快照检查。** 两次拒绝：无条件的事件边界检查会把程序返回未 await 调用时的 `done` 帧改判为 worker-exit；按 `data` 事件刷新的快照在 flowing 模式同一宏任务内多个事件时过期（第二块携带超过上限的在途调用的合法突发会被误杀）。每宏任务在微任务排空后检查真实计数，在两个轴上都不依赖分块；严格阈值让恰好 `MAX_PENDING_REPLIES` 个在途调用正常完成。
+
+**只读取一次元数据但保留原始 errorClass 对象。** 拒绝：引导帧的 `JSON.stringify` 会重新调用 getter；只有存入普通副本才能保证两个阶段读到相同的值。
+
+**依赖排空的 `finally` 重置来回收队列内存。** 拒绝：重置只在排空结束时运行；永不排空的排空会持续增长。排空进行中的压缩在排空存活期间限制后备存储。
+
+**保留完成值计量器的显式成员栈。** 拒绝：字节预算限制遍历本身，但不限制栈的引用数——那是 O(width)——接近帧上限的宽值会复制数百万引用，在解析成功后 OOM 宿主；每层游标形态保持 O(depth) 状态。
+
+## Consequences
+
+在途 binding 闭包与回复积压一样受限，向永不结算的 binding 洪泛调用的子进程会让运行提前失败，而不是把闭包累积到墙钟；合法的并发大 gather 不受影响（上限在微任务队列排空后检查）。引导帧序列化校验批准的元数据，与 getter 状态无关。回复队列的后备存储在持续的部分排空期间保持有界；压缩是内部内存卫生，无可观察的行为变化。完成值计量器以 O(depth) 辅助状态保持精确的字节核算，宽完成值不再因计量本身 OOM 宿主。
@@ -0,0 +1,6 @@
+# Bilingual-pair consistency record (docs/i18n/README.md): the git blob hash of each
+# 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 .agents/notes/implemented/bug-fix/2026-08-29-code-runtime-python-load-and-dispatch-hardening.md
+2026-08-29-code-runtime-python-load-and-dispatch-hardening.md: 3d64f96420cd337fc8c7bb44e02f912e1868deed
+2026-08-29-code-runtime-python-load-and-dispatch-hardening.zh.md: 64772eb14a09585b1ee0ab10ffcf1298b77b0a35
@@ -0,0 +1,41 @@
+# Agent Note: Load-time pythonBin validation, binding snapshot, and reply-drain settle in the CPython backend
+
+Status: implemented
+
+English | [中文](2026-08-29-code-runtime-python-load-and-dispatch-hardening.zh.md)
+
+## Problem
+
+Review of the CPython subprocess backend (packages/experimental/code-runtime-python) surfaced four non-blocking findings that a long-running host could still misbehave under: an explicit `pythonBin` path bypassed the load-time configuration checks, a throwing binding member accessor could escape the fd-3 data callback and terminate the host, the reply drain could hang forever waiting for a `drain` event that a destroyed pipe never emits, and two leak assertions diffed a global tmpdir in a way a parallel vitest worker could false-positive on.
+
+## Decision
+
+### An explicit pythonBin must be an executable regular file at load
+
+`resolvePythonBin` returned an absolute or slash-containing `pythonBin` verbatim, so a missing, non-executable, or directory path passed the constructor's load checks (which only rejected empty/NUL values and unresolvable basenames) and surfaced only at the first `run()` as a misleading `worker-exit`. The explicit-path branch now validates with the same `accessSync(X_OK)` + `statSync().isFile()` checks the PATH branch uses (a directory passes `X_OK`, so the regular-file requirement is the deciding half), resolving relative explicit paths against the host CWD first — the same place `spawn` would have looked. A failing explicit path makes `resolvePythonBin` return `undefined`, and the load check now distinguishes the two failure classes in its message: `is not an executable regular file` for an explicit path, `does not resolve on PATH` for a basename.
+
+### Binding callables are snapshotted during validation
+
+`namespace.functions` is caller-supplied, so its members may be exposed through getters or a Proxy. Reading one of them inside the fd-3 `data` callback — `record[message.name]` — threw OUTSIDE the dispatcher's try and terminated the host (an `uncaughtException` handler, if installed, would only let the run degrade to the wall clock). `validateBindings` now reads every member into a plain own-property record during run()'s synchronous validation segment, so a throwing accessor becomes the seam-misuse rejection run() already reserves for malformed bindings. The snapshot is also the single key set the boot frame advertises AND dispatch reads, so a getter whose keys differ between reads cannot desynchronize the child's allowed names from what the host will actually call. The record is null-prototype (`Object.create(null)`): the seam contract treats member names like `__proto__` or `constructor` as ordinary own properties, and a plain `{}` assignment of `__proto__` hits the prototype setter instead of creating the own property, dropping the name from the boot frame and making a call to it fail with `KeyError`.
+
+### The reply drain settles on a destroyed pipe
+
+`drainReplies` awaited `once(proto, 'drain')` after a full-buffer write; a pipe destroyed under the wait (child exited, close-deadline teardown) never emits `drain` again, and `events.once` rejects only on `error`, not on `close` — the await could hang forever, leaving `draining` true and the unconsumed queue (and any wide payloads it still holds) pinned with the closure. The wait now listens for `drain`, `close`, and `error` together, removing all three listeners whichever wins, and the drain loop short-circuits on `proto.destroyed` before the next write, so the `finally` clears the queue and resets `draining`.
+
+## Testing
+
+- `tests/runtime.spec.ts` — the load-rejection cases cover a missing absolute path, a non-executable regular file, a directory, and a slash-containing relative path, each asserting the `is not an executable regular file` message; a positive case keeps an absolute interpreter path loading and running. A case with a getter that throws on read asserts `run()` rejects as seam misuse; a companion with a counting getter asserts the accessor is read exactly once (the snapshot), proving dispatch and the boot frame share the snapshot. The spawn-failure case now stages an executable wrapper, loads the runtime, deletes the wrapper, and asserts the run still resolves `worker-exit` (a load-time-valid path can still fail at run time; the old fixture used a path that is now rejected at load).
+- `tests/boot-write-failure.spec.ts` — a fake child backpressures every fd-3 write and destroys the pipe while the host waits for `drain`; the run settles on the wall clock instead of hanging on the drain wait.
+- The two staging-leak cases assert the exact paths this test file staged (recorded by the mocked `mkdtempSync`) are gone, instead of diffing a global tmpdir that a sibling worker could perturb.
+
+## Alternatives considered
+
+**Leave the explicit-path branch unvalidated and let the first run() report it.** Rejected: a missing, non-executable, or directory interpreter path is a self-contained configuration error that the caller can fix without running a program, and the empty/NUL and basename checks already set the precedent that these fail at load. The run-time `worker-exit` it produced was also indistinguishable from a substrate failure, so the caller could not tell a configuration mistake from an environment problem.
+
+**Guard the member access inside the dispatch path instead of snapshotting.** Rejected: a try around `record[message.name]` would still read the getter on EVERY call, repeating its side effects and allowing its key set to differ between the boot frame's advertisement and dispatch. Snapshotting once, during validation, converts the throw into the seam-misuse rejection run() already reserves and fixes the key set to one record.
+
+**Extend the drain wait with a timeout.** Rejected: a timeout would settle the wait while the pipe might still be alive, dropping a queued reply that a still-open pipe could have taken. Listening for `close`/`error` settles exactly when the pipe is gone, which is the only case where `drain` can never arrive.
+
+## Consequences
+
+Load now rejects a self-contained configuration error earlier (an explicit interpreter path that is not an executable regular file), matching the basename treatment. Binding member accessors are read once, at validation, so a getter's side effects cannot repeat per call. A destroyed fd-3 pipe no longer strands the reply drain. The leak assertions are immune to concurrent staging by sibling workers.
@@ -0,0 +1,41 @@
+# Agent Note: CPython 后端的加载期 pythonBin 校验、binding 快照与回复排空结算
+
+Status: implemented
+
+[English](2026-08-29-code-runtime-python-load-and-dispatch-hardening.md) | 中文
+
+## Problem
+
+对 CPython 子进程后端（packages/experimental/code-runtime-python）的评审浮出四项非阻断发现，在长驻宿主上仍可能表现异常：显式 `pythonBin` 路径绕过加载期配置校验；抛错的 binding 成员访问器可能逃出 fd-3 data 回调并终止宿主；回复排空可能永远等待一个已销毁管道不会再发出的 `drain` 事件；两处泄漏断言对全局 tmpdir 做差集，并行 vitest worker 可能误报。
+
+## Decision
+
+### 显式 pythonBin 在加载期必须是可执行的普通文件
+
+`resolvePythonBin` 对绝对路径或含斜杠的 `pythonBin` 原样返回，因此不存在、不可执行或指向目录的路径能通过构造器的加载期检查（只拒绝空串/NUL 值与无法解析的裸名），直到首次 `run()` 才以误导性的 `worker-exit` 暴露。显式路径分支现在复用 PATH 分支所用的 `accessSync(X_OK)` + `statSync().isFile()` 检查（目录也能通过 `X_OK`，因此普通文件要求是起决定作用的一半），先把相对显式路径解析到宿主 CWD——与 `spawn` 会查找的位置相同。失败的显式路径使 `resolvePythonBin` 返回 `undefined`，加载检查现在在消息中区分两类失败：显式路径报 `is not an executable regular file`，裸名报 `does not resolve on PATH`。
+
+### binding 可调用对象在校验期被快照
+
+`namespace.functions` 由调用方提供，其成员可能通过 getter 或 Proxy 暴露。在 fd-3 `data` 回调中读取其中一个成员——`record[message.name]`——会在分发器 try 之外抛出并终止宿主（即使安装了 `uncaughtException` 处理器，运行也只会退化到墙钟超时）。`validateBindings` 现在在 run() 的同步校验段把每个成员读入一个普通自有属性记录，因此抛错的访问器变成 run() 为畸形 binding 预留的 seam-misuse 拒绝。该快照同时是 boot 帧宣告与分发读取的同一份键集，因此键随读取变化的 getter 无法让子进程被允许的名字与宿主实际调用的名字失步。记录采用无原型构造（`Object.create(null)`）：seam 契约把 `__proto__`、`constructor` 之类的成员名当作普通自有属性，普通 `{}` 对 `__proto__` 的赋值会命中原型 setter 而非创建自有属性，使该名字从 boot 帧消失、对其的调用以 `KeyError` 失败。
+
+### 回复排空在管道已销毁时结算
+
+`drainReplies` 在缓冲区满写入后 `await once(proto, 'drain')`；在等待期间被销毁的管道（子进程退出、close 截止时间拆卸）永远不会再发出 `drain`，而 `events.once` 只在 `error` 时拒绝、不在 `close` 时结算——该 await 可能永远挂起，使 `draining` 保持 true，未消费的队列（及其仍持有的宽 payload）随闭包滞留。等待现在同时监听 `drain`、`close` 与 `error`，任一事件胜出即移除全部三个监听器；排空循环在下一次写入前用 `proto.destroyed` 短路，因此 `finally` 会清空队列并复位 `draining`。
+
+## Testing
+
+- `tests/runtime.spec.ts`——加载拒绝用例覆盖不存在的绝对路径、不可执行的普通文件、目录与含斜杠的相对路径，各自断言 `is not an executable regular file` 消息；一个正向用例让绝对解释器路径通过加载并运行。一个 getter 在读取时抛错的用例断言 `run()` 以 seam misuse 拒绝；一个配套用例用计数 getter 断言访问器恰好被读取一次（快照），证明分发与 boot 帧共享快照。spawn 失败用例现在先暂存一个可执行 wrapper、加载 runtime、删除 wrapper，再断言运行仍 resolve 为 `worker-exit`（加载期合法的路径仍可能在运行期失败；旧 fixture 用的路径现在在加载期就被拒绝）。
+- `tests/boot-write-failure.spec.ts`——一个 fake child 让每次 fd-3 写入都背压，并在宿主等待 `drain` 时销毁管道；运行在墙钟上结算，而不是挂在排空等待上。
+- 两处暂存泄漏用例断言本测试文件暂存的确切路径（由被 mock 的 `mkdtempSync` 记录）已消失，而不是对可能被同级 worker 扰动的全局 tmpdir 做差集。
+
+## Alternatives considered
+
+**让显式路径分支不做校验，由首次 run() 报告。** 已拒绝：不存在、不可执行或指向目录的解释器路径是调用方无需运行程序即可修复的自包含配置错误，且空串/NUL 与裸名检查已确立这些应在加载期失败的先例。它产生的运行期 `worker-exit` 也与子进程故障无法区分，调用方无法分辨配置错误与环境问题。
+
+**在分发路径内守卫成员访问，而非快照。** 已拒绝：在 `record[message.name]` 周围加 try 仍会在每次调用时读取 getter，重复其副作用，并允许其键集在 boot 帧宣告与分发之间不一致。在校验期快照一次，把抛错转化为 run() 已预留的 seam-misuse 拒绝，并把键集固定为同一份记录。
+
+**给排空等待加超时。** 已拒绝：超时会在管道可能仍存活时结算等待，丢弃一个仍可被存活的管道接收的排队回复。监听 `close`/`error` 恰好在管道消失时结算，这是 `drain` 永远不会到达的唯一情形。
+
+## Consequences
+
+加载期现在更早地拒绝一个自包含配置错误（非可执行普通文件的显式解释器路径），与裸名的处理一致。binding 成员访问器在校验期被读取一次，getter 的副作用不会逐次调用重复。已销毁的 fd-3 管道不再搁浅回复排空。泄漏断言对同级 worker 的并发暂存免疫。
@@ -0,0 +1,6 @@
+# Bilingual-pair consistency record (docs/i18n/README.md): the git blob hash of each
+# 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 .agents/notes/implemented/bug-fix/2026-08-29-code-runtime-python-reply-backlog-and-surrogate-count.md
+2026-08-29-code-runtime-python-reply-backlog-and-surrogate-count.md: 5ae31f669e2e207bc2f496d11ca3464f032783f1
+2026-08-29-code-runtime-python-reply-backlog-and-surrogate-count.zh.md: 799178dd54ceddd9b80b11e94d723282a037d398
@@ -0,0 +1,33 @@
+# Agent Note: Bound the reply backlog and count lone surrogates without a match list in the CPython backend
+
+Status: implemented
+
+English | [中文](2026-08-29-code-runtime-python-reply-backlog-and-surrogate-count.zh.md)
+
+## Problem
+
+A further review round on the CPython subprocess backend (packages/experimental/code-runtime-python) surfaced two unbounded-allocation findings. First, `replyQueue` had no bound: a child that never reads fd 3 keeps the reply pipe full forever, so the drain loop waits on `drain` while every call frame it keeps sending resolves a binding and queues another reply — the backlog (and the binding results it pins) grows until the wall clock. Second, `_json_str_cost` counted lone surrogates with `_SURROGATE.findall(folded)`, which materializes one single-character string per surrogate: a surrogate-dense completion value near the budget (each surrogate serializes to six bytes, so a budget-sized value holds millions of them) allocates millions of objects before the meter returns, defeating the meter's own contract of counting without building.
+
+## Decision
+
+### The reply backlog is capped at 1024 pending frames
+
+`sendReply` now counts pending replies separately from the consumed slots the drain loop clears, and settles the run as a `worker-exit` with a reply-queue message before pushing when the backlog reaches `MAX_PENDING_REPLIES`. The counter is decremented as the drain writes each frame and reset when the drain finishes, so it measures only replies the host still holds. This mirrors the frame cap's treatment of an oversized inbound frame: a child that stops participating in the protocol fails the run early instead of growing host memory until the wall clock. It is a count bound, not a byte bound — binding results carry no seam-level byte cap, so the bound limits how many are retained, not how large any one is.
+
+### Lone surrogates are counted by length difference, not by a match list
+
+`_json_str_cost` computed `lone = len(_SURROGATE.findall(folded))`, building a list of one single-character string per lone surrogate. The count is now the length difference between `folded` and `without = _SURROGATE.sub("", folded)`: after pair-combining, every remaining surrogate is lone and exactly one code point, so the number removed is the count, and the `without` string is needed by the meter anyway. The meter returns the identical byte cost with no per-surrogate objects.
+
+## Testing
+
+- `tests/runtime.spec.ts` — a hostile child floods 5000 sequential valid call frames and never reads fd 3; the run settles as `worker-exit` with the reply-queue message long before `maxWallMs`, proving the backlog cap fires instead of a wall-clock timeout. A surrogate-dense completion of 3,000,000 lone surrogates pins the boundary at scale: 18,000,002 serialized bytes succeed at an 18,000,002 budget and report `output-limit` one byte under, proving the meter counts every surrogate exactly (the len-diff is verified equal to the old findall count across lone-high, lone-low, paired, astral, and mixed cases).
+
+## Alternatives considered
+
+**Pause the fd-3 read side while waiting for drain instead of capping the queue.** Rejected: pausing reads would also stall processing of `done` and `log` frames the child may send after its last call, changing settlement timing; a count cap is deterministic and matches the existing frame-cap pattern.
+
+**Keep findall and rely on the character-count lower bound.** Rejected: the lower bound admits a string by CHARACTER count while each surrogate serializes to six bytes, so a budget-sized surrogate-dense string passes it and reaches the meter; the match list is exactly the allocation the meter exists to avoid.
+
+## Consequences
+
+A child that stops consuming its replies now fails the run as a `worker-exit` once 1024 replies are retained, bounding host memory without a wall-clock wait. The completion-value meter counts lone surrogates with no per-surrogate allocation, keeping its documented counting-without-building contract for surrogate-dense values.
@@ -0,0 +1,33 @@
+# Agent Note: 在 CPython 后端限制回复积压并改用长度差计数孤立代理项
+
+Status: implemented
+
+[English](2026-08-29-code-runtime-python-reply-backlog-and-surrogate-count.md) | 中文
+
+## Problem
+
+对 CPython 子进程后端（packages/experimental/code-runtime-python）的又一轮评审浮出两项无界分配发现。其一，`replyQueue` 没有上限：从不读取 fd 3 的子进程让回复管道永远占满，排空循环只能等待 `drain`，而它持续发送的每个调用帧都会解析一个 binding 并入队一条回复——积压（连同其钉住的 binding 结果）一直增长到墙钟。其二，`_json_str_cost` 用 `_SURROGATE.findall(folded)` 计数孤立代理项，每个代理项物化一个单字符字符串：接近预算的代理项密集完成值（每个代理项序列化为六个字节，预算大小的值可容纳数百万个）会在计量返回前分配数百万个对象，违背计量器自身「计数而不构建」的契约。
+
+## Decision
+
+### 回复积压限制为 1024 个待发帧
+
+`sendReply` 现在把待发回复数与排空循环已清空的槽位分开计数，当积压达到 `MAX_PENDING_REPLIES` 时，在入队前以带回复队列消息的 `worker-exit` 结算运行。计数器在排空写入每帧时递减、排空结束时重置，因此只度量宿主仍持有的回复。这与帧上限对超大入站帧的处理一致：停止参与协议的子进程让运行提前失败，而不是让宿主内存增长到墙钟。这是计数上限而非字节上限——binding 结果在 seam 层没有字节上限，因此该上限限制保留的数量，而非单个结果的大小。
+
+### 孤立代理项改用长度差计数，而非匹配列表
+
+`_json_str_cost` 原先计算 `lone = len(_SURROGATE.findall(folded))`，为每个孤立代理项构建一个单字符字符串的列表。现在计数改为 `folded` 与 `without = _SURROGATE.sub("", folded)` 的长度差：配对合并后，剩余的每个代理项都是孤立且恰好一个码点，因此被移除的数量即计数，而 `without` 字符串本就是计量需要的。计量器返回完全相同的字节成本，且不产生任何按代理项计的对象。
+
+## Testing
+
+- `tests/runtime.spec.ts`——敌意子进程洪泛 5000 个连续合法调用帧且从不读取 fd 3；运行在远早于 `maxWallMs` 时以带回复队列消息的 `worker-exit` 结算，证明积压上限先于墙钟超时触发。3,000,000 个孤立代理项的代理项密集完成值在规模上钉住边界：18,000,002 个序列化字节在 18,000,002 预算下成功、少一个字节时报 `output-limit`，证明计量器精确计数每个代理项（长度差在孤立高、孤立低、配对、星面和混合用例下与旧 findall 计数逐一相等，已实测验证）。
+
+## Alternatives considered
+
+**在等待 drain 时暂停 fd-3 读侧而非限制队列。** 拒绝：暂停读取也会让子进程在最后一个调用后可能发送的 `done` 与 `log` 帧处理停滞，改变结算时机；计数上限是确定性的，且与既有帧上限模式一致。
+
+**保留 findall 并依赖字符计数下界。** 拒绝：下界按字符数放行字符串，而每个代理项序列化为六个字节，因此预算大小的代理项密集字符串能通过下界并进入计量器；匹配列表正是计量器要避免的分配。
+
+## Consequences
+
+停止消费回复的子进程现在会在保留 1024 条回复时以 `worker-exit` 结算运行，无需等待墙钟即可限制宿主内存。完成值计量器对孤立代理项的计数不再产生按代理项计的分，保持其对代理项密集值「计数而不构建」的既有契约。
@@ -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 .agents/notes/implemented/feature/2026-07-21-continuable-background-subagents.md
-2026-07-21-continuable-background-subagents.md: b2a3a8c53db5ae2860ed5cc6edccadfcd417e7fa
-2026-07-21-continuable-background-subagents.zh.md: 24cc09e621731cbb54f4d081d232f0598220d418
+2026-07-21-continuable-background-subagents.md: e1adf8efa19cce8e0632170886770b8d12bc9e3d
+2026-07-21-continuable-background-subagents.zh.md: f62c314320a1dace43f3028b68a1bb7d65203e84
@@ -57,7 +57,7 @@ The continuation manager does not serialize two callers that race a stopped chil

 ### Model-facing `send_message`

-The model receives one `send_message(subagent_id, message)` tool backed by `SubagentRuntime.followup()`, matching the intent verb on `Agent`. The service operation owns steer-or-resume orchestration and is distinct from the run's `SubagentRun.steer?()`, which only delivers to an already active run. The tool performs no lifecycle routing of its own. It attributes the follow-up as `{ kind: 'coordinator', senderSessionId: parent.id }` and forwards `{ source, signal }`; the service requires both facts in one options object. The source crosses both live steering and cold resume, while cancellation owns only a pending live-delivery wait because a cold-resume Task returns immediately and owns its later cancellation. The child model still receives ordinary user-role content, while the durable source prevents model-generated follow-ups from being classified as direct human input. A human adapter instead supplies `{ kind: 'user' }` and its interaction signal. The tool lives in the separately loaded `@deepseek-ai/dsh-tool-subagent-control` package so provider-bound `@deepseek-ai/dsh-tool-subagent` instances can continue registering distinct delegation tools for spawn, fork, or ACP without registering duplicate global control tools.
+The model receives one global `send_message(agent_id, message)` tool backed by `SubagentRuntime.sendMessage()`. The exact live sender may name only its direct parent or direct continuable child; the service owns adjacency checks, cold resume, fixed Steer scheduling, and durable `{ kind: 'agent-message', senderSessionId }` attribution. Cancellation owns work only until inbox acceptance. A human adapter remains separate because browser-authored input carries user provenance and request identity rather than Agent authority. The tool lives in the separately loaded `@deepseek-ai/dsh-tool-subagent-control` package so provider-bound `@deepseek-ai/dsh-tool-subagent` instances can continue registering distinct delegation tools for spawn, fork, or ACP without registering duplicate global control tools.

 - If the child has a running Task and live-steering capability, the service calls `run.steer(message, source)` and returns the existing Job id; it creates no Task of its own.
 - If the child has no running Task, `send_message` creates a fresh Task, cold-resumes the durable session with the message, and returns the new Job id.
@@ -57,7 +57,7 @@ durable child Session

 ### 面向模型的 `send_message`

-模型获得一个由 `SubagentRuntime.followup()` 支撑的 `send_message(subagent_id, message)` 工具，与 `Agent` 上的意图动词一致。该服务操作负责在 steering 与恢复之间编排；它不同于 run 的 `SubagentRun.steer?()`，后者只能向已活跃的 run 发送消息。工具本身不执行生命周期路由。该工具将后续消息的来源标记为 `{ kind: 'coordinator', senderSessionId: parent.id }`，并转发 `{ source, signal }`；服务要求在一个选项对象中同时提供这两项信息。来源会贯穿在线 steering 和 cold resume 两条路径，而取消只控制尚未完成的在线投递等待，因为 cold resume Task 会立即返回，并自行负责后续取消。child 模型收到的仍是普通的 user role 内容，而持久化的来源信息可防止模型生成的后续消息被归类为直接用户输入。用户适配器则提供 `{ kind: 'user' }` 及其交互信号。该工具位于单独加载的 `@deepseek-ai/dsh-tool-subagent-control` 包中，因此按提供方绑定的 `@deepseek-ai/dsh-tool-subagent` 实例可以继续为 spawn、fork 或 ACP 注册不同的委派工具，而不会重复注册全局控制工具。
+模型获得一个由 `SubagentRuntime.sendMessage()` 支撑的全局 `send_message(agent_id, message)` 工具。确切在线 sender 只能指定其直接 parent 或直接可继续 child；服务负责相邻关系检查、冷恢复、固定 Steer 调度，以及持久化 `{ kind: 'agent-message', senderSessionId }` 来源信息。取消只负责 inbox 接受前的工作。用户适配器保持分离，因为浏览器编写的输入携带用户来源信息与请求身份，而不是 Agent 权限。该工具位于单独加载的 `@deepseek-ai/dsh-tool-subagent-control` 包中，因此按提供方绑定的 `@deepseek-ai/dsh-tool-subagent` 实例可以继续为 spawn、fork 或 ACP 注册不同的委派工具，而不会重复注册全局控制工具。

 - 如果 child 存在运行中的 Task 并支持在线消息，服务会调用 `run.steer(message, source)` 并返回现有 job id；它不会创建新 Task。
 - 如果 child 没有运行中的 Task，`send_message` 会创建新 Task，使用该消息从持久化存储恢复会话，并返回新的 job id。
@@ -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 .agents/notes/implemented/feature/2026-07-28-continuable-subagent-conversations.md
-2026-07-28-continuable-subagent-conversations.md: f456bacbf775bf914b47051e19639811e2385f65
-2026-07-28-continuable-subagent-conversations.zh.md: ea6df53026b6a76d8e100fa4fef0aebcd8fcd7cf
+2026-07-28-continuable-subagent-conversations.md: 1efb18ab69cfe779a7de85f288a6e3aa7caaca83
+2026-07-28-continuable-subagent-conversations.zh.md: 7cb1b4a4cff1fc7f6fc41662d05693917084b63a
@@ -44,7 +44,7 @@ Cold resume does not dispatch through a subagent provider. The continuation mana

 `SubagentProvider.start()` and `SubagentRun` remain exclusively on the unchanged one-shot path. A continuable Activation directly owns its `AgentHandle` and never creates, wraps, or retains a `SubagentRun`; `SubagentRun.steer?()` is therefore absent.

-`ctx.subagents.followup(parent, childId, content, { source, signal })` remains the sole parent-to-child continuation-message operation. The exact live parent Agent authorizes delivery; cold resume checks that authority before reconstruction and every path checks it again in the final no-await inbox-admission span, so a parent unregistered or replaced during materialization cannot authorize delivery. `source` records who supplied the admitted message and grants no authority. The model-facing `send_message` tool keeps only its stable `subagent_id` and `message` fields and always submits a follow-up turn. Both start and follow-up return the accepted `MessageId`, and neither reports how the manager materialized the Activation.
+`ctx.subagents.sendMessage(sender, targetId, content, { signal })` is the sole model-authored continuation-message operation. The exact live sender authorizes delivery to its direct parent or direct continuable child; cold resume checks direct-child authority before reconstruction and every path checks again in the final no-await inbox-admission span, so an Agent unregistered or replaced during materialization cannot authorize delivery. The service derives durable `agent-message` provenance from that sender. The model-facing `send_message` tool keeps only `agent_id` and `message` and uses fixed Steer scheduling. Both start and send return the accepted `MessageId`, and neither reports how the manager materialized the Activation.

 For start and follow-up, the caller signal owns lookup, materialization, and admission only until inbox acceptance. After the operation returns its `MessageId`, the manager owns the Activation independently; later caller cancellation does not cancel the accepted turn or dispose the child.

@@ -111,15 +111,13 @@ Top-level teardown is host-owned rather than represented as another Activation.

 The activation-owner scope exists because ordinary Cordis owner effects unwind in reverse registration order, which cannot express the dynamic child graph. Manager initialization registers the private scope's structural disposer first and its drain disposer afterward, so reverse unwind invokes the drain before releasing that scope; merely registering a cleanup effect on the same scope as later Agent handles would allow structural handle disposal to bypass child-first ordering. Each materialization registers its barrier participant and snapshots its exact live ancestry before starting the inner transaction, then remains tracked until it installs an Activation or fully rolls back. The Activation retains weak membership of that ancestry, so an intermediate Agent may leave the registry without hiding a still-live descendant from its host root. Each Activation installs one memoized disposal promise before cancellation or recursive callbacks, allowing scoped host shutdown, global manager unload, child release, and normal settlement to converge without double release. Cancellation propagates top-down before slow descendant cleanup; handle release remains child-first. Sibling branches drain independently; one disposal failure is recorded but does not prevent the manager from attempting the remaining selected handles, and the aggregate drain reports failure after all selected branches settle. Durable child Sessions survive this process-local teardown.

-### Report delivery extension
+### Adjacent-Agent messaging

-The optional child-scoped `report(output)` tool was added later without changing Activation residency or adding another queue. It can be called zero or multiple times per turn, derives the live direct parent rather than accepting a recipient, and selects quiet injection or a waking parent follow-up through deployment config. The [report-tool Agent Note](2026-07-30-continuable-subagent-report-tool.md) owns its authority, acknowledgement, setup-contribution, and delivery contracts.
+The shared `sendMessage(sender, targetId, content, options)` service operation adds no second queue. It accepts an exact live sender, permits only its direct parent or direct continuable child, and uses fixed Steer scheduling through the Agent inbox. The global `send_message({ agent_id, message })` tool exposes that same operation in both directions; the child's initial task identifies its direct parent when the tool is visible. The [adjacent-Agent messaging Agent Note](../architecture/2026-08-27-adjacent-agent-steer-messaging.md) owns its schema, authority, attribution, and prompt placement.

-### Deferred steering
+### Fixed Steer scheduling

-This version exposes no subagent steering operation. Parent continuation messages always open later FIFO turns, so the continuation layer stores no current-turn controller and adds no controller-aware Agent admission contract.
-
-A later host UI may expose separate **Steer** and **Follow up** actions. Host steering would be strict and live-only: it may call the existing Agent steering path only while the Activation accepts a next step, must reject otherwise, and must never fall back to queueing or cold resume. Exposing parent steering to a model-facing tool remains a separate design.
+Every accepted Agent message uses `Agent.steer()`. A running target claims it at the nearest step boundary; an idle or cold-resumed target starts a turn. The continuation layer does not expose a caller-selectable quiet, next-turn, or follow-up mode.

 ### Authority and recorded sender identity

@@ -44,7 +44,7 @@ inbox 接受消息前发生任何失败，操作都会在不返回任何 id 的

 `SubagentProvider.start()` 和 `SubagentRun` 只保留在不变的 one-shot 路径上。可继续激活直接持有自身的 `AgentHandle`，绝不创建、包装或保留 `SubagentRun`；因此，`SubagentRun.steer?()` 不存在。

-`ctx.subagents.followup(parent, childId, content, { source, signal })` 仍是唯一的从 parent 到 child 的继续执行消息操作。确切的在线 parent Agent 授权投递；冷恢复会在重建前检查该权限，每条路径还会在最终无 await 的 inbox 准入区间再次检查，因此在物化期间被注销或替换的 parent 无法授权投递。`source` 记录谁提供了获准消息，不赋予任何权限。面向模型的 `send_message` 工具只保留稳定的 `subagent_id` 和 `message` 字段，并始终提交一个 follow-up 轮次。start 和 follow-up 都返回已接受的 `MessageId`，两者都不报告管理器如何物化激活。
+`ctx.subagents.sendMessage(sender, targetId, content, { signal })` 是唯一由模型编写的继续执行消息操作。确切在线 sender 授权向其直接 parent 或直接可继续 child 投递；冷恢复会在重建前检查直接 child 权限，每条路径还会在最终无 await 的 inbox 准入区间再次检查，因此在物化期间被注销或替换的 Agent 无法授权投递。服务从该 sender 推导持久化 `agent-message` 来源信息。面向模型的 `send_message` 工具只保留 `agent_id` 和 `message`，并使用固定 Steer 调度。start 与 send 都返回已接受的 `MessageId`，两者都不报告管理器如何物化 Activation。

 对于 start 和 follow-up，调用方 signal 只在 inbox 接受消息前持有查找、物化和准入。操作返回 `MessageId` 后，管理器会独立持有该激活；调用方之后的取消不会取消已接受的轮次，也不会 dispose child。

@@ -111,15 +111,13 @@ Agent inbox 是唯一队列。每条继续执行消息都使用 `Agent.followup(

 activation-owner 作用域之所以存在，是因为普通 Cordis owner effect 按注册逆序撤销，无法表达动态 child 图。管理器初始化时先注册私有作用域的结构化 disposer，再注册自身的 drain disposer，使逆序撤销先执行 drain、再释放该作用域；如果只在与后续 Agent handle 相同的作用域上注册 cleanup effect，结构化 handle dispose 就可能绕过 child-first 顺序。每个物化过程都会在启动内部事务前注册其屏障参与项，并对其确切的在线祖先建立快照，然后保持跟踪，直到安装 Activation 或完全回滚。Activation 会保留其在这组祖先中的弱成员关系，因此中间 Agent 即使离开注册表，也不会让仍在线的后代脱离宿主根节点的可见范围。每个 Activation 都会在取消或递归回调前安装一个记忆化的 dispose promise，使限定作用域的宿主关闭、全局管理器卸载、child 释放和正常结算能够汇合，而不会重复释放。取消会在等待缓慢的后代清理之前自顶向下传播；handle 释放仍是 child-first。同级分支独立 drain；系统会记录单次 dispose 失败，但仍会尝试其余选中 handle，聚合 drain 则在所有选中分支结算后报告失败。这次进程内拆卸不会销毁持久化 child 会话。

-### 报告投递扩展
+### 相邻 Agent 消息

-后来添加的可选 child 作用域 `report(output)` 工具不会改变 Activation 驻留状态，也不会增加另一条队列。它每轮可调用零次或多次，不允许指定接收方，而是推导在线的直接 parent；投递采用静默注入还是唤醒 parent follow-up，由部署配置选择。[report 工具 Agent Note](2026-07-30-continuable-subagent-report-tool.zh.md)规定其权限、确认、设置贡献和投递约定。
+共享的 `sendMessage(sender, targetId, content, options)` 服务操作不会增加第二条队列。它接收确切在线 sender，只允许其直接 parent 或直接可继续 child，并通过 Agent inbox 使用固定 Steer 调度。全局 `send_message({ agent_id, message })` 工具在两个方向暴露同一个操作；当 child 可以看到该工具时，其初始任务会标明直接 parent。[相邻 Agent 消息 Agent Note](../architecture/2026-08-27-adjacent-agent-steer-messaging.zh.md)规定其 schema、权限、来源信息与提示词位置。

-### 延后的 steering（中途引导）
+### 固定 Steer 调度

-本版本不暴露 subagent steering 操作。parent 的继续执行消息始终开启后续 FIFO 轮次，因此继续执行层不存储当前轮次控制方，也不新增能够感知控制方的 Agent 准入约定。
-
-后续宿主 UI 可以分别暴露 **Steer** 和 **Follow up** 操作。宿主 steering 必须严格且仅限在线使用：只有当激活接受下一步骤时，它才能调用现有的 Agent steering 路径；其他情况必须拒绝，而且绝不能转为排队或冷恢复。是否通过面向模型的工具暴露 parent steering 仍需单独设计。
+每条已接受的 Agent 消息都使用 `Agent.steer()`。运行中的目标会在最近的 step 边界领取消息；空闲或冷恢复的目标会启动一个轮次。继续执行层不暴露由调用方选择的 quiet、next-turn 或 follow-up 模式。

 ### 权限与已记录的发送方身份

@@ -1,118 +0,0 @@
-# Agent Note: Continuable subagent report tool
-
-Status: implemented
-
-English | [中文](2026-07-30-continuable-subagent-report-tool.zh.md)
-
-## Problem
-
-Continuable in-process subagents can receive later parent messages, retain descendants, settle, and cold-resume, but the base lifecycle gives them no way to send selected content back to their direct parent. Their complete output already remains reconstructable from the durable child Session, so the missing capability is explicit delivery rather than result storage.
-
-Treating every final assistant message as an implicit result would conflate turn completion with reporting. A long-lived child may have nothing useful to report in one turn, may report progress several times in another, and must remain available after reporting. Recipient authority, quiet versus next-step delivery, acknowledgement, durability, and retry behavior therefore need one explicit contract.
-
-## Decision
-
-Add the independently installed `@deepseek-ai/dsh-tool-subagent-report` package. It contributes an ordinary model-facing `report` tool to each continuable in-process child Activation. The mechanism accepts zero or multiple calls in a turn; the child is separately instructed to call it once before finishing ([the report obligation](2026-08-06-continuable-child-report-obligation.md)). Success neither concludes the turn, settles the Activation, nor prevents later parent follow-ups, and finishing a turn never reports automatically.
-
-The feature is a collaboration control, not a result-bearing execution wrapper. It adds no Task, `SubagentRun`, result promise, Activation state, delivery queue, or replay path.
-
-### Model-facing contract
-
-`report` accepts exactly `{ output: string }` and returns exactly `{ messageId: string }`. It accepts no child id, recipient id, or delivery mode. `exec.agent` binds the tool call to the reporting child, the service derives the sole recipient from durable `parentSession`, and deployment config owns scheduling.
-
-`messageId` is the stable `MessageId` of the user-role message accepted into the parent's inbox. It is not a read receipt, parent-log acknowledgement, turn-completion receipt, or persistence flush.
-
-The description states that reporting is required before finishing, repeatable, direct-parent-only, and non-terminal. It warns that a failed tool result may still follow an accepted send because a later `tools/post-execute` failure can replace the result. Without an idempotency key, stronger wording would encourage duplicate retries after ambiguous failure.
-
-The tool uses generic rendering with no locations. Its acknowledgement includes `messageId`. Scope-local registration keeps presentation and execution aligned: roots, one-shot children, remote providers, sibling scopes, and agentless execution neither see nor execute `report`. It installs after the child's global `toolFilter`, so a delegation allow-list cannot accidentally remove the structural return channel; deployments that require no return channel omit the package.
-
-### Service authority
-
-The subagent seam exposes `ctx.subagents.reportFrom(child, content, { delivery, signal }): Promise<MessageId>`. The exact live child Agent is the sender credential. The continuation manager accepts only an Activation whose `handle.agent === child`, derives its direct parent from the child's durable header, and requires that id to resolve to a live parent Agent in the final synchronous authorization-and-send span. The API accepts no caller-selected recipient, ancestor, or sender fields.
-
-Roots, one-shot children, forged objects, stale Agents, and same-id replacements fail with `UNAUTHORIZED`. A closing child Activation fails with `ACTIVATION_CLOSING`; manager drain and pre-acceptance cancellation retain their existing lifecycle errors. A missing or send-rejecting direct parent fails with `PARENT_UNAVAILABLE` and `direct parent is not live; report was not delivered`. Failure returns no id, cold-resumes no parent, writes no offline mailbox, and mutates no absent-parent Session.
-
-Nested reporting crosses exactly one edge. A grandchild reports to its direct child parent, never to the top-level coordinator. That intermediate child may explicitly report a derived update later.
-
-### Delivery policy
-
-The package validates `reportDelivery: 'quiet' | 'next-step'`; the default is `next-step` ([ordering decision](../bug-fix/2026-08-17-subagent-report-settlement-ordering.md)).
-
-Quiet delivery calls `parent.inject()`. It adds model-visible next-step context without waking an idle parent; a running parent stages the report for the next safe log position.
-
-Next-step delivery calls `parent.steer()`. It wakes a parked parent and joins a running parent's nearest step boundary. When that parent is itself a continuable Activation, the send uses the manager's existing admission accounting so the parent cannot settle between synchronous inbox insertion and the admission microtask. Reports share the next-step FIFO with a later settlement notice, preserving their accepted causal order.
-
-Both modes frame one user-role message as `Background subagent <child-id> reported:` followed by the exact `output`. The durable message source is `{ kind: 'subagent-report', senderSessionId: child.id }`. Normal Agent ordering governs concurrent sends; the subagent layer creates no second queue.
-
-### Acknowledgement and recovery
-
-Success means the exact live parent synchronously accepted the message. The context becomes reconstructable only when it reaches its normal log boundary; a next-step delivery has woken the parent, while quiet delivery may remain pending. The inbox message id remains separate from the returned stable message id.
-
-The first version provides no durable mailbox, idempotency key, delivery receipt, retry protocol, or exactly-once claim. A process failure can leave the caller uncertain, and retry after an unknown outcome may duplicate a report. The durable child transcript remains the recovery source when the parent is unavailable.
-
-### Composition and lifecycle
-
-The subagent seam adds `registerContinuableSetup(contribution): () => void`, backed by `SubagentActivationSetupRegistry`. Each synchronous contribution receives the unpublished child context and returns the disposer for its installation. The continuation manager first applies base child composition, then current contributions in registration order through the same setup closure used for fresh creation and cold resume.
-
-The registry owns registration, per-child installation records, setup rollback, child-scope cleanup, and immediate revocation. Applying a batch returns the Agent setup commit that revalidates provisioning after every setup await and immediately before Agent publication. A throwing or concurrently revoked contribution therefore rejects before either Agent or Session publication and rolls back the batch. New registrations affect a resident child only on its next Activation; removing a registration first closes it to new setup and then revokes every provisioning or resident installation immediately. Registration disposal and child-context disposal are idempotent and attempt every release before aggregating failures.
-
-This seam keeps the continuation manager unaware of tool names. The report package installs only `report` and its child-scoped guidance section; `@deepseek-ai/dsh-tool-subagent-control` independently installs parent-side `send_message` and `list_agents`. A deployment can install either direction, both, or neither. Providers remain data-only, durable descriptors do not snapshot report availability or delivery mode, and cold resume uses the deployment's current contributions and policy.
-
-### Snapshot coverage
-
-The ACP snapshot harness adds `waitForSubagentTurnEnd`, selecting the Nth harvested child by the same order as `session.N.jsonl`. It waits for a closed child turn containing a request header so a continuable child's earlier descriptor-seed turn cannot satisfy the boundary. This lets the assembled scenario wait for the child-side report without inventing a parent-visible signal.
-
-The authored snapshot starts a continuable child, executes the real scope-local `report` tool, and observes default next-step delivery before the manager's later settlement notice. A snapshot-only maintenance fence holds the parent until both messages are pending, proving next-step input is claimed before queued next-turn input when the parent resumes. It declares child pins `1`, so the otherwise non-global `report` schema and the child's own prompt are checked against `tool-schemas.1.expected.json` and `system-prompt.1.expected.md` while the root keeps the class pins. The generated tool catalog separately mints a child scope to include the same scope-local schema.
-
-## Alternatives considered
-
-### Automatically deliver every final answer
-
-Automatic delivery cannot represent zero reports, progress reports, or several selected updates. It also couples reporting to settlement and can duplicate content already reported explicitly.
-
-### Always wake the parent
-
-Waking on every report creates unsolicited turns and can cascade through nested subagents. Quiet delivery was chosen as the default on the assumption that the parent had another reason to read its context. [The report obligation](2026-08-06-continuable-child-report-obligation.md) supersedes that choice: a parked background coordinator has no such reason, so waking is the default and this paragraph now records why `quiet` still exists.
-
-### Let the child choose the delivery mode
-
-Giving the model a mode argument grants it control over scheduler pressure and makes behavior deployment-dependent. The child chooses content and timing; deployment config chooses whether that content wakes the parent.
-
-### Register a global tool
-
-A global `report` would advertise an unusable capability to roots, one-shot children, remote children, and agentless callers. Execution-time rejection would make schema visibility disagree with authority.
-
-### Combine both directions in the control package
-
-`send_message` and `report` have different audiences, scopes, configuration, and lifecycle. Independent packages let deployments grant either direction without implying the other.
-
-### Persist an offline parent mailbox
-
-Mutating or cold-resuming an absent parent requires a new durable addressing, authorization, conflict, acknowledgement, and replay protocol. Requiring a live direct parent keeps the first version on the existing Agent send path.
-
-### Reintroduce a Task or result promise
-
-A result-bearing wrapper makes one report or one turn appear terminal and recreates the lifetime mismatch that continuable Activations removed. Explicit repeatable sends need no intermediate execution object.
-
-### Validate setup after Agent creation
-
-A post-creation revocation check can reject the Activation only after the Agent and Session have been published. Disposing the returned handle removes the live objects but cannot delete persistence through the current seam, leaving a resumable child that the continuation manager said was never established. Returning an `AgentSetupCommit` instead lets the Agent factory perform the same mutable-state check synchronously at its publication boundary.
-
-## Consequences
-
- A continuable in-process child exposes exactly one scope-local `report` schema only while the report package's contribution is installed; unrelated Agents never expose it.
- The tool returns the parent message's stable `MessageId`; its inbox occurrence is not a separate public identity.
- Only the exact resident child may report, and only to the exact live direct parent derived from durable lineage. The service has no recipient parameter or offline fallback.
- Next-step delivery is the validated default: it wakes an idle parent or extends a running parent's turn at the nearest step boundary. Quiet delivery never wakes an idle parent.
- Child cancellation or disposal after parent acceptance does not retract the report. Before acceptance, child disposal, drain, parent loss, or caller cancellation rejects the operation.
- Fresh and resumed Activations compose current setup contributions before publication. Grants wait for the next Activation; revocation is immediate for resident children.
- Unit coverage pins visibility, allow-list behavior, both delivery modes, stable message and sender identities, nested routing, invalid senders, absent parents, cancellation, drain, revocation races, and the absence of Jobs or implicit final reporting.
- The keyless assembled snapshot proves the real child tool, default next-step ordering before settlement, and durable parent framing.
-
-### Accepted risks
-
-The acceptance boundary is weaker than durable end-to-end delivery. A crash can leave the result ambiguous, and retries may duplicate reports.
-
-Next-step delivery can amplify model work when nested children report frequently. Reports waiting together share one step, and deployment ownership through `reportDelivery` bounds but does not remove that risk.
-
-Registry presence is the parent liveness signal. A host-owned parent whose `AgentHandle.dispose()` has started but has not yet unwound its scope can still accept and append a report that it will not act on in this process. Closing that gap requires an Agent-level disposal-start signal rather than subagent-layer inference.
@@ -1,118 +0,0 @@
-# Agent Note: 可继续 subagent 报告工具
-
-Status: implemented
-
-[English](2026-07-30-continuable-subagent-report-tool.md) | 中文
-
-## 问题
-
-可继续的进程内 subagent 能够接收 parent 后续发来的消息、保留后代、结算并冷恢复，但基础生命周期无法让它们将选中内容发送给直接 parent。child 的完整输出已可从持久化会话中重建，因此缺失的能力是显式投递，而非结果存储。
-
-如果将每条 assistant 最终消息都视为隐式结果，就会混淆轮次完成与报告。长期运行的 child 可能在某个轮次中无内容可报告，也可能在另一个轮次多次报告进展，而且报告后必须仍可继续工作。因此，接收方权限、静默投递与 next-step 投递、确认、持久性和重试行为都需要一份显式约定。
-
-## 决策
-
-新增可独立安装的 `@deepseek-ai/dsh-tool-subagent-report` 包。它会向每个可继续进程内 child Activation 贡献一个普通的面向模型 `report` 工具。机制本身接受一个轮次中调用零次或多次；child 会另行被要求在结束前调用一次（见[报告义务](2026-08-06-continuable-child-report-obligation.zh.md)）。调用成功既不会结束该轮次或结算 Activation，也不会阻止 parent 之后继续 follow-up；完成轮次也绝不会自动报告。
-
-该功能是协作控制，不是承载结果的执行包装层。它不新增 Task、`SubagentRun`、结果 promise、Activation 状态、投递队列或回放路径。
-
-### 面向模型的约定
-
-`report` 只接受 `{ output: string }`，也只返回 `{ messageId: string }`。它不接受 child id、接收方 id 或投递模式。`exec.agent` 将工具调用绑定到发送报告的 child；服务从持久化 `parentSession` 中推导唯一接收方，调度则由部署配置决定。
-
-`messageId` 是已接受进入 parent inbox 的用户角色消息所对应的稳定 `MessageId`。它不是已读回执、parent 日志确认、轮次完成回执或持久化 flush。
-
-工具描述会明确报告操作在结束前必须执行、可重复、仅限直接 parent 且不会结束轮次。它还会警告：发送被接受后，后续 `tools/post-execute` 失败可能替换工具结果，因此工具结果失败时内容仍可能已经送达。没有幂等键时，更强的表述会诱导调用方在结果不明确的失败后重复重试。
-
-该工具使用不带 location 的通用渲染，其确认中包含 `messageId`。作用域局部注册使呈现与执行保持一致：root、one-shot child、远程提供方、同级作用域和无 agent（智能体）执行既不能看到，也不能执行 `report`。它会在 child 的全局 `toolFilter` 之后安装，因此委派 allow-list 不会意外移除这条结构性返回通道；不需要返回通道的部署不安装该包。
-
-### 服务权限
-
-subagent seam 暴露 `ctx.subagents.reportFrom(child, content, { delivery, signal }): Promise<MessageId>`。确切的在线 child Agent 是发送方凭据。继续执行管理器只接受 `handle.agent === child` 的 Activation，从 child 的持久化 header 中推导其直接 parent，并要求该 id 在最终的同步授权与发送区间解析为一个在线 parent Agent。该 API 不接受由调用方选择的接收方、祖先或发送方字段。
-
-root、one-shot child、伪造对象、陈旧 Agent 和同 id 替换对象都以 `UNAUTHORIZED` 失败。正在关闭的 child Activation 以 `ACTIVATION_CLOSING` 失败；管理器 drain 和接受前取消保留既有的生命周期错误。直接 parent 不存在或拒绝接受时，以 `PARENT_UNAVAILABLE` 和 `direct parent is not live; report was not delivered` 失败。失败不返回 id，不冷恢复 parent，不写入离线邮箱，也不会修改缺失 parent 的会话。
-
-嵌套报告恰好跨越一条边。grandchild 会向其直接 child parent 报告，绝不会直接向顶层 coordinator 报告。中间 child 可以稍后显式报告自己归纳的更新。
-
-### 投递策略
-
-该包会校验 `reportDelivery: 'quiet' | 'next-step'`，默认值为 `next-step`（见[顺序决策](../bug-fix/2026-08-17-subagent-report-settlement-ordering.zh.md)）。
-
-静默投递调用 `parent.inject()`。它会添加模型可见的 next-step 上下文，但不唤醒空闲 parent；运行中的 parent 会把报告暂存到下一个安全日志位置。
-
-Next-step 投递调用 `parent.steer()`。它会唤醒停驻的 parent，并加入运行中 parent 最近的 step 边界。当该 parent 本身也是可继续 Activation 时，发送会使用管理器现有的准入记账，防止 parent 在同步插入 inbox 与准入微任务之间结算。报告与稍后的结算通知共享 next-step FIFO，从而保持其被接受时的因果顺序。
-
-两种模式都会将一条用户角色消息封装为 `Background subagent <child-id> reported:`，后面跟随完全原样的 `output`。持久化消息来源为 `{ kind: 'subagent-report', senderSessionId: child.id }`。并发发送的顺序由 Agent 的常规规则决定；subagent 层不会创建第二条队列。
-
-### 确认与恢复
-
-成功表示确切的在线 parent 已同步接受该消息。上下文只有到达正常日志边界后才可重建；next-step 投递已经唤醒 parent，而静默投递可能继续等待。inbox 消息 id 不会成为另一个公开身份。
-
-首个版本不提供持久化邮箱、幂等键、投递回执、重试协议或恰好一次保证。进程故障可能让调用方无法确定结果，在结果未知时重试则可能重复报告。parent 不可用时，持久化 child transcript（文本记录）仍是恢复来源。
-
-### 组合与生命周期
-
-subagent seam 新增 `registerContinuableSetup(contribution): () => void`，由 `SubagentActivationSetupRegistry` 支撑。每个同步贡献都会接收尚未发布的 child 上下文，并返回其安装的 disposer。继续执行管理器首先应用基础 child 组合，然后通过同一个用于首次创建与冷恢复的设置闭包，按注册顺序应用当前贡献。
-
-注册表负责注册、每个 child 的安装记录、设置回滚、child 作用域清理和立即撤销。应用一个批次会返回 Agent setup 提交对象，用于在每次 setup 的 await 结算后以及紧邻 Agent 发布前重新校验配置状态。因此，某项贡献抛出异常或被并发撤销时，会在 Agent 与会话发布前拒绝操作并回滚该批次。新注册项只会在驻留 child 的下一个 Activation 生效；移除注册项时，会先将它对新设置关闭，再立即撤销为正在预配置或驻留的每个 child 安装的实例。注册 dispose（资源释放）与 child 上下文 dispose 都是幂等的，两者都会先尝试每项释放，再聚合失败。
-
-该 seam 使继续执行管理器无需知道工具名。report 包只安装 `report` 及其 child 作用域指引 section；`@deepseek-ai/dsh-tool-subagent-control` 则独立安装 parent 侧的 `send_message` 和 `list_agents`。部署时可安装任一方向、同时安装两者或两者均不安装。提供方仍只负责数据，持久化描述符不会对 report 可用性或投递模式建立快照，冷恢复则使用部署当前的贡献与策略。
-
-### 快照覆盖
-
-ACP（Agent Client Protocol）快照 harness 新增 `waitForSubagentTurnEnd`，按与 `session.N.jsonl` 相同的顺序选择第 N 个已收集 child。它会等待一个包含请求 header 的已闭合 child 轮次，以防可继续 child 早期播种描述符的轮次错误满足该边界。这样，整体组装的场景无需伪造 parent 可见信号，就能等待 child 侧报告。
-
-手写快照会启动一个可继续 child，执行真实的作用域局部 `report` 工具，并观察默认 next-step 投递先于管理器稍后的结算通知。一个仅用于快照的 maintenance 围栏会保持 parent，直至两条消息都处于待领取状态，从而证明 parent 恢复时先领取 next-step 输入、再领取排队的 next-turn 输入。它声明 child pin `1`，因此本不属于全局的 `report` schema 与该 child 自身的提示词会分别与 `tool-schemas.1.expected.json` 和 `system-prompt.1.expected.md` 比对，root 则继续使用类别 pin。生成的工具目录会另外铸造一个 child 作用域，以收录同一个作用域局部 schema。
-
-## 曾考虑的替代方案
-
-### 自动投递每个最终回答
-
-自动投递无法表示零次报告、进展报告或多次精选更新。它还会将报告与结算耦合，并可能重复投递已显式报告的内容。
-
-### 始终唤醒 parent
-
-每次报告都唤醒 parent 会产生未经请求的轮次，还可能沿嵌套 subagent 级联扩散。当初选择静默投递作为默认值，前提是 parent 还有别的理由去读自己的上下文。[报告义务](2026-08-06-continuable-child-report-obligation.zh.md)取代了该选择：已经停驻的后台协调者并没有这样的理由，因此唤醒成为默认值，而本段现在记录的是 `quiet` 为何仍然保留。
-
-### 允许 child 选择投递模式
-
-向模型提供 mode 参数会赋予其控制调度器压力的能力，并使行为依赖部署。child 只决定内容和时机；该内容是否唤醒 parent，由部署配置决定。
-
-### 注册全局工具
-
-全局 `report` 会向 root、one-shot child、远程 child 和无 agent 调用方公布一项无法使用的能力。到执行时才拒绝，会使 schema 可见性与权限不一致。
-
-### 将两个方向合并到 control 包
-
-`send_message` 与 `report` 的受众、作用域、配置和生命周期各不相同。独立的包可让部署授予任意一个方向，而不暗示也授予另一个方向。
-
-### 持久化离线 parent 邮箱
-
-修改或冷恢复不在线的 parent，需要一套新的持久化寻址、权限、冲突、确认和回放协议。要求直接 parent 在线，可以让首个版本继续使用现有 Agent 发送路径。
-
-### 重新引入 Task 或结果 promise
-
-承载结果的包装层会让一次报告或一个轮次看似具有终止性，并重新引入可继续 Activation 已经移除的生命周期不匹配。显式、可重复的发送无需中间执行对象。
-
-### 在 Agent 创建后校验 setup
-
-创建完成后的撤销检查只能在 Agent 与会话均已发布后拒绝 Activation。对返回的 handle 执行 dispose 会移除实时对象，但当前 seam 无法删除持久化内容，因此会留下一个仍可恢复的 child，而继续执行管理器却判定它从未建立。改为返回 `AgentSetupCommit`，Agent 工厂便可在自身的发布边界同步执行同一项可变状态检查。
-
-## 影响
-
- 只有安装 report 包贡献时，可继续进程内 child 才会恰好暴露一个作用域局部 `report` schema；无关 Agent 永远不会暴露该 schema。
- 工具返回 parent 消息的稳定 `MessageId`；其 inbox 中的出现不会成为另一个公开身份。
- 只有确切的驻留 child 才能报告，且只能报告给根据持久化谱系推导的确切在线直接 parent。服务不接受接收方参数，也不提供离线 fallback。
- Next-step 投递是校验后的默认模式：它会唤醒空闲 parent，或在最近的 step 边界延长运行中 parent 的轮次。静默投递绝不会唤醒空闲 parent。
- parent 接受后取消或 dispose child 不会撤回报告。接受前，child dispose、drain、parent 丢失或调用方取消都会拒绝操作。
- 新建和恢复的 Activation 都会在发布前组合当前设置贡献。新授权等待下一个 Activation 才生效，而已驻留 child 的授权撤销立即生效。
- 单元覆盖固定可见性、allow-list 行为、两种投递模式、稳定的消息与发送方身份、嵌套路由、无效发送方、缺失的 parent、取消、drain、撤销竞争，以及不存在 Task 或隐式最终报告。
- 无密钥整体组装快照证明真实 child 工具、默认 next-step 顺序先于结算，以及持久化 parent 封装。
-
-### 已接受的风险
-
-该接受边界弱于持久化端到端投递。崩溃可能导致结果不明，重试则可能重复报告。
-
-嵌套 child 频繁报告时，next-step 投递可能放大模型工作量。一起等待的报告会共享一个 step，通过 `reportDelivery` 交由部署所有者控制也会限制该风险，但无法完全消除。
-
-注册表中的存在性就是 parent 在线信号。宿主拥有的 parent 如果已开始 `AgentHandle.dispose()` 但尚未完成其作用域清理，仍可能接受并追加一条本进程不会再处理的报告。要弥合这个缺口，需要 Agent 层面的 dispose 开始信号，不能由 subagent 层推断。
@@ -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 .agents/notes/implemented/feature/2026-08-05-context-form-vocabulary.md
-2026-08-05-context-form-vocabulary.md: 912842a9d59fd49491a53f987d172d64d4c0e101
-2026-08-05-context-form-vocabulary.zh.md: 0d4e1d44e4d22b8bfcb6c73bb015bc3240381bd3
+2026-08-05-context-form-vocabulary.md: be792256fa3375dd2621d647d6d3074b1a6d6b18
+2026-08-05-context-form-vocabulary.zh.md: f648dd1071128b879e8c36e9fba1646b6c70cda1
@@ -39,7 +39,7 @@ That move also relocates catalog **identity**: the republish digest now covers t

 Both readers are **all-or-nothing**: one unreadable entry disqualifies the record rather than being dropped, because a body that replaces the model-facing text must not present a confident but incomplete account of what the model read. The row's form marker reports what actually rendered, not what was declared.

-The producer side validates the same durable data with the same posture. `catalogHistory` reads `source.entries` out of `agent.session.events`, which on resume or fork is a persistence seed whose validation only guarantees a source object with a non-empty `kind` — no per-kind field is checked. An unreadable catalog is therefore skipped as "not this plugin's record", the posture the replaced content digest had; throwing there would fail every later step of that Session at the latest, least diagnosable point.
+The producer side validates the same durable data with the same posture. `catalogHistory` reads `source.entries` out of `agent.session.snapshotEvents()`, which on resume or fork is a persistence seed whose validation only guarantees a source object with a non-empty `kind` — no per-kind field is checked. An unreadable catalog is therefore skipped as "not this plugin's record", the posture the replaced content digest had; throwing there would fail every later step of that session at the latest, least diagnosable point.

 Everything else — including a form this UI version does not present, a form absent from the source, and a `catalog` whose entries are unusable — renders the **opaque** body: the model-facing text with its real line breaks, then the remaining source data as fields. Opaque is the documented default; the contract assigns these unsupported cases to it. A resumed, forked, or foreign log must render whether or not its producer is mounted here, which is also why the classification lives in the durable source rather than in a client-side table keyed by producer.

@@ -39,7 +39,7 @@ Status: implemented

 两个读取器都是**全有或全无**：一条不可读的条目即判定整条记录不可用，而不是把它丢掉——会替换掉面向模型文本的内容区，不得给出自信但残缺的「模型读到了什么」。行上的形态标记报告的是实际渲染出的形态，而非声明的形态。

-生产方一侧对同一份持久数据采取同样的姿态。`catalogHistory` 从 `agent.session.events` 读 `source.entries`，而恢复或 fork 时它来自持久化 seed，seed 验证只保证来源是带非空 `kind` 的对象，不校验任何 kind 特有字段。因此不可读的目录被当作「不是本插件的记录」跳过——正是被替换掉的内容 digest 原有的姿态；在那里抛错会让该 Session 此后每一步都在最晚、最难定位的点失败。
+生产方一侧对同一份持久数据采取同样的姿态。`catalogHistory` 从 `agent.session.snapshotEvents()` 读 `source.entries`，而恢复或 fork 时它来自持久化 seed，seed 验证只保证来源是带非空 `kind` 的对象，不校验任何 kind 特有字段。因此不可读的目录被当作「不是本插件的记录」跳过——正是被替换掉的内容 digest 原有的姿态；在那里抛错会让该会话此后每一步都在最晚、最难定位的点失败。

 其余一切——包括本 UI 版本不呈现的形态、来源未声明形态、以及条目不可用的 `catalog`——一律渲染 **opaque** 内容区：按真实换行展示面向模型的文本，其后把剩余来源数据列成字段。opaque 是文档规定的默认；约定要求这些不支持的情况使用它。恢复的、fork 的、外部写入的日志，无论其生产方是否挂载在此处都必须渲染得出来——这同样是分类信息必须落在持久来源里、而不是落在客户端以生产方为键的表里的原因。

@@ -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 .agents/notes/implemented/feature/2026-08-05-durable-web-schedule.md
-2026-08-05-durable-web-schedule.md: 2a07d8257df6e940b316749a9b96a13abaf201dc
-2026-08-05-durable-web-schedule.zh.md: c3a37f69ab74e5ededb7ca89c45ad9acfd029251
+2026-08-05-durable-web-schedule.md: bac4a5cfd8965032dad2cf689ca42b1f8e5da1e3
+2026-08-05-durable-web-schedule.zh.md: 266a90d3a56caccda56be9fe648c059939b8155f
@@ -60,7 +60,7 @@ Dispatch records queue admission, not model completion or user receipt. Framing

 ### Read-only Web catalog

-The Schedule overlay enables the otherwise-disabled [`dsh-client-ui-schedule`](../../../../packages/client/ui-schedule/README.md) client together with the Host service. The complete active projection also feeds [`dsh-client-ui-workspace`](../../../../packages/client/ui-workspace/README.md). This note owns that opt-in read-only presentation boundary: the projection is current active state, not a dispatch or delivery receipt, so ordinary Assistant turns remain the delivery presentation.
+The Schedule overlay enables the otherwise-disabled [`dsh-client-ui-schedule`](../../../../packages/client/ui-schedule/README.md) client together with the Host service. The complete active projection also feeds [`dsh-client-ui-workspace`](../../../../packages/client/ui-workspace/README.md). This note owns that opt-in read-only presentation boundary: the projection is current active state, not a dispatch or delivery receipt, so ordinary Assistant turns remain the delivery presentation. The catalog is a fixed `document.body` portal whose left edge follows the trigger when space permits and shifts left to retain a 16px viewport margin near the right edge. `useAnchoredPosition` owns measurement and resize, captured-scroll, panel-resize, and cleanup behavior; Schedule supplies the trigger and portal refs, bottom placement, a 5px gap, and the existing inside/outside dismissal boundary without adding a general popover abstraction.

 ## Alternatives considered

@@ -80,7 +80,7 @@ The Schedule overlay enables the otherwise-disabled [`dsh-client-ui-schedule`](.

 ## Verification

-Package tests pin strict replay, one-shot and Every transitions, creation-anchor arithmetic, latest-only catch-up, multi-record batching, fork suffixes, id reuse, offset and local-calendar profiles, IANA validation, daylight-saving gaps and overlaps, time bounds, timer segmentation, wall-clock movement, overdue admission, fixed framing, enqueue and append failures, barrier recovery, projection registration and restoration, registration rollback, and quiescent disposal at per-file 100% coverage. A property test compares Every calculation and replay across varied intervals and skipped spans. A production JSONL restart test proves one overdue reminder dispatches through the real Agent lifecycle and does not redispatch after another restart. Focused client suites own catalog and sidebar behavior. Keyless assembled Web scenarios retain ordinary After/At/Every delivery evidence plus one Schedule-catalog smoke for overlay reachability, the current header catalog, ordinary/search alarms, narrow dark layout, and one live empty update.
+Package tests pin strict replay, one-shot and Every transitions, creation-anchor arithmetic, latest-only catch-up, multi-record batching, fork suffixes, id reuse, offset and local-calendar profiles, IANA validation, daylight-saving gaps and overlaps, time bounds, timer segmentation, wall-clock movement, overdue admission, fixed framing, enqueue and append failures, barrier recovery, projection registration and restoration, registration rollback, and quiescent disposal at per-file 100% coverage. A property test compares Every calculation and replay across varied intervals and skipped spans. A production JSONL restart test proves one overdue reminder dispatches through the real Agent lifecycle and does not redispatch after another restart. Focused client suites own catalog and sidebar behavior, including the body portal, spacious left alignment, portal-inside pointer handling, outside dismissal, Escape, live empty, and timer cleanup. The shared primitive suite owns the positioning hook's resize, captured-scroll, panel-resize, and cleanup lifecycle. Keyless assembled Web scenarios retain ordinary After/At/Every delivery evidence plus one 900×900 Schedule-catalog smoke for overlay reachability, fixed portal placement, right-edge clamping, width and overflow, ordinary/search alarms, narrow dark layout, a light-theme browser screenshot, and one live empty update.

 ## Consequences

@@ -60,7 +60,7 @@ dispatch 记录的是队列准入，而不是模型完成或用户收到提醒

 ### 只读 Web 目录

-Schedule overlay 会把默认禁用的 [`dsh-client-ui-schedule`](../../../../packages/client/ui-schedule/README.zh.md) client 与 Host 服务一同启用。完整活动 projection 也会交给 [`dsh-client-ui-workspace`](../../../../packages/client/ui-workspace/README.zh.md)。本 Note 拥有这条 opt-in 只读呈现边界：该 projection 表示当前活动状态，而非 dispatch 或交付回执，因此普通 Assistant 轮次仍是交付呈现。
+Schedule overlay 会把默认禁用的 [`dsh-client-ui-schedule`](../../../../packages/client/ui-schedule/README.zh.md) client 与 Host 服务一同启用。完整活动 projection 也会交给 [`dsh-client-ui-workspace`](../../../../packages/client/ui-workspace/README.zh.md)。本 Note 拥有这条 opt-in 只读呈现边界：该 projection 表示当前活动状态，而非 dispatch 或交付回执，因此普通 Assistant 轮次仍是交付呈现。目录是挂到 `document.body` 的 fixed portal；空间足够时左边缘跟随触发按钮，靠近视口右侧时向左避让并保留 16px 边距。`useAnchoredPosition` 拥有测量以及 resize、捕获阶段 scroll、面板 resize 与清理行为；Schedule 提供触发器与 portal ref、bottom 放置、5px 间距和既有内外 dismissal 边界，不增加通用 popover 抽象。

 ## 已考虑的替代方案

@@ -80,7 +80,7 @@ Schedule overlay 会把默认禁用的 [`dsh-client-ui-schedule`](../../../../pa

 ## 验证

-包测试以逐文件 100% coverage 固定严格回放、一次性与 Every 状态转换、创建锚点运算、只追赶最新一次、多记录批处理、fork 后缀、id 复用、偏移量与本地日历 profile、IANA 校验、夏令时缺口与重叠、时间边界、timer 分段、墙钟变化、overdue 准入、固定 framing、入队与 append 失败、barrier 恢复、projection 注册与恢复、注册 rollback 和完全停稳的 dispose。属性测试会在不同间隔与跳过跨度下比较 Every 计算与回放。production JSONL restart 测试证明一条 overdue 提醒会经过真实 Agent 生命周期 dispatch，并且再次 restart 后不会重复 dispatch。聚焦 client suite 拥有目录与侧边栏行为。无密钥组装 Web 场景保留普通 After／At／Every 交付证据，再由一个 Schedule 目录 smoke 覆盖 overlay 可达性、当前 header 目录、普通／搜索闹钟、窄屏暗色布局与一次 live empty 更新。
+包测试以逐文件 100% coverage 固定严格回放、一次性与 Every 状态转换、创建锚点运算、只追赶最新一次、多记录批处理、fork 后缀、id 复用、偏移量与本地日历 profile、IANA 校验、夏令时缺口与重叠、时间边界、timer 分段、墙钟变化、overdue 准入、固定 framing、入队与 append 失败、barrier 恢复、projection 注册与恢复、注册 rollback 和完全停稳的 dispose。属性测试会在不同间隔与跳过跨度下比较 Every 计算与回放。production JSONL restart 测试证明一条 overdue 提醒会经过真实 Agent 生命周期 dispatch，并且再次 restart 后不会重复 dispatch。聚焦 client suite 拥有目录与侧边栏行为，包括 body portal、空间充足时的左对齐、portal 内指针处理、外部 dismissal、Escape、live empty 与 timer 清理。共享 primitive suite 拥有定位 hook 的 resize、捕获阶段 scroll、面板 resize 与清理生命周期。无密钥组装 Web 场景保留普通 After／At／Every 交付证据，再由一个 900×900 Schedule 目录 smoke 覆盖 overlay 可达性、fixed portal 定位、右侧钳制、宽度与 overflow、普通／搜索闹钟、窄屏暗色布局、浅色主题浏览器截图与一次 live empty 更新。

 ## 后果

@@ -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 .agents/notes/implemented/feature/2026-08-05-feedback-gated-session-telemetry.md
-2026-08-05-feedback-gated-session-telemetry.md: ce07e63681a659fb9bf758a24301c531f3f27e39
-2026-08-05-feedback-gated-session-telemetry.zh.md: 9657c7dc9243377233f20b265569829b06162fb5
+2026-08-05-feedback-gated-session-telemetry.md: 9dab135029bd4a5dd209157b48b6f7dca4040224
+2026-08-05-feedback-gated-session-telemetry.zh.md: 824aa1a04f3854163fdd16f4db7a7593540976e0
@@ -18,7 +18,7 @@ Session telemetry originally has one mounted behavior: every accepted record ent

 The generic telemetry coordinator owns `live` and `on-demand` capture. Live capture projects, clones, redacts, and hands each event to the backend on the session firehose. On-demand capture registers no continuous capture listeners; `captureSession(session, throughSeq)` reads the canonical log from the handoff cursor through an inclusive boundary, then projects, clones, redacts, and hands over that prefix. The cursor advances only for handed-over records. The [buffer-free replay decision](../simplification/2026-08-06-buffer-free-feedback-telemetry.md) owns why the on-demand path uses the canonical log instead of copied records.

-Mode resolution is a closed, fail-before-setup check: an unknown direct-construction value fails before transport configuration is read. Only `FULL` exposes the public service's `emit()` path to the SDK pipeline. `FEEDBACK_ONLY` gives its on-demand coordinator a private backend capability; its listener passes an event to `captureSession()` only when the exact `feedback/record` object is already stored at `session.events[event.seq]`. `Session.append` commits that object before publishing `session/event`, so replay includes the feedback but cannot extend past its boundary. `DISABLED` creates neither the capability nor the SDK pipeline and does not inspect exporter configuration.
+Mode resolution is a closed, fail-before-setup check: an unknown direct-construction value fails before transport configuration is read. Only `FULL` exposes the public service's `emit()` path to the SDK pipeline. `FEEDBACK_ONLY` gives its on-demand coordinator a private backend capability; its listener passes an event to `captureSession()` only when `session.eventAt(event.seq)` returns that exact `feedback/record` object. `Session.append` commits that object before publishing `session/event`, so replay includes the feedback but cannot extend past its boundary. `DISABLED` creates neither the capability nor the SDK pipeline and does not inspect exporter configuration.

 ## Alternatives considered

@@ -18,7 +18,7 @@ Status: implemented

 通用遥测协调器拥有 `live` 与 `on-demand` 捕获。实时捕获在会话 firehose 上投影、深拷贝、脱敏每个事件，并将其交给后端。按需捕获不注册持续捕获监听器；`captureSession(session, throughSeq)` 从 handoff 游标起读取权威日志，直至含边界的指定序列号，然后投影、深拷贝、脱敏并交接该前缀。游标只为已交接记录推进。[无缓冲回放决策](../simplification/2026-08-06-buffer-free-feedback-telemetry.zh.md)说明了按需路径为何使用权威日志而非记录副本。

-模式解析采用封闭式检查，并在设置前失败：通过直接构造传入未知值时，会在读取传输配置前失败。只有 `FULL` 向 SDK 流水线开放公共服务的 `emit()` 路径。`FEEDBACK_ONLY` 向其按需协调器提供私有后端能力；其监听器向 `captureSession()` 传递事件的唯一条件，是该事件与那个 `feedback/record` 对象身份完全相同，且该对象已存储于 `session.events[event.seq]`。`Session.append` 在发布 `session/event` 前已提交该对象，因此回放包含该反馈，但不会越过其边界。`DISABLED` 既不创建该能力，也不创建 SDK 流水线，并且不检查导出器配置。
+模式解析采用封闭式检查，并在设置前失败：通过直接构造传入未知值时，会在读取传输配置前失败。只有 `FULL` 向 SDK 流水线开放公共服务的 `emit()` 路径。`FEEDBACK_ONLY` 向其按需协调器提供私有后端能力；其监听器向 `captureSession()` 传递事件的唯一条件，是 `session.eventAt(event.seq)` 返回完全相同的 `feedback/record` 对象。`Session.append` 在发布 `session/event` 前已提交该对象，因此回放包含该反馈，但不会越过其边界。`DISABLED` 既不创建该能力，也不创建 SDK 流水线，并且不检查导出器配置。

 ## 考虑过的替代方案

@@ -1,60 +0,0 @@
-# Agent Note: The continuable child return channel is an obligation
-
-Status: implemented
-
-English | [中文](2026-08-06-continuable-child-report-obligation.zh.md)
-
-## Problem
-
-A continuable background child owns its own Session, so nothing it writes there reaches the agent that started it. [The report tool](2026-07-30-continuable-subagent-report-tool.md) gave that child a return channel and then presented it as one option among several: the schema said "call this zero or more times", nothing in the child's prompt asked it to call the tool at all, and the accepted default scheduling (`quiet`) added the report to a parked parent's next request without waking it.
-
-Each of those choices is defensible alone. Together they made the return channel unusable as a delegation contract. A child that finished its work, wrote its answer into its own transcript, and stopped left the parent with nothing; a child that did report reached a parent that had already parked and would not read the report until something unrelated woke it. External reports of parents busy-polling `list_agents`, re-sending messages to settled children, and abandoning `subagent` for `workflow` all reduce to the same missing guarantee.
-
-## Decision
-
-The return channel is an instruction the child receives, not a capability it may discover. The report package installs two scope-local registrations into every continuable in-process child, and one disposer revokes both:
-
- the `report` tool, whose description now states that the child calls it once before finishing with a self-contained final result, and earlier for progress that changes what the parent should do next;
- a `tool:report` system-prompt section at first-party order 2900 carrying the same obligation in the child's own voice, so a child that never reads tool descriptions closely still receives it.
-
-`reportDelivery` defaults to `next-step`. An accepted report wakes a parked parent driver or joins a running parent's nearest step boundary, matching the instruction to report findings that change the parent's next action. `quiet` remains available for deployments that prefer unread reports over model-work amplification. The [report/settlement ordering decision](../bug-fix/2026-08-17-subagent-report-settlement-ordering.md) owns the scheduling rationale.
-
-### Why the section and the description both exist
-
-They address different failure modes. The tool description is read when the model is already considering `report`; the prompt section is read when it is deciding whether it is finished. The obligation belongs at both points because the failure this fixes — a child that simply stops — happens at the second one.
-
-The section is registered on the child's own scope, the same mechanism [child composition](../../../../packages/subagent/subagent/src/child-agent.ts) already uses for a shadowing persona, so the parent and every sibling see neither the tool nor the guidance. `installReportTool` rolls the section back if tool registration fails, and its returned disposer attempts both revocations before surfacing cleanup failures.
-
-### Instruction, not enforcement
-
-Nothing rejects a child that never reports. No runtime path inspects whether a report was sent, and `report` still accepts zero or many calls per turn. The change is model-facing wording plus a scheduling default; the service authority, acknowledgement, and recovery contracts are unchanged.
-
-That boundary is deliberate: prompt text can only reach a child that is still running its own loop. A child stopped by an error, a token ceiling, cancellation, or teardown never gets the chance to comply, which is why the runtime keeps its own account of settlement rather than trusting this instruction ([manager-owned settlement delivery](2026-08-06-manager-owned-subagent-settlement-delivery.md)).
-
-### Snapshot coverage
-
-The assembled ACP `subagent-report` scenario exercises the shipped default: the child reports while the parent is in maintenance, the later settlement notice queues behind it, and the resumed parent claims the next-step report before next-turn settlement. Because the child's scope composes a prompt the class pin cannot describe, the snapshot harness has `pinsChildSystemPrompts`, the exact counterpart of `pinsChildToolSchemas`: it moves one child fixture's prompt into `system-prompt.<n>.expected.md`, leaves every other request-header field to the class pin, requires the sidecar exactly when declared, and rejects a sidecar identical to that class pin so a redundant copy cannot drift.
-
-## Alternatives considered
-
-**Keep `quiet` as the default and rely on the prompt alone.** This was the shipped position, and it supersedes nothing on its own: a report the parent never reads is indistinguishable from a report never sent. The [report-tool note's](2026-07-30-continuable-subagent-report-tool.md) rejection of always-waking assumed the parent had another reason to look at its context; a parked background coordinator does not. Turn amplification is the real cost, and it is now the reason `quiet` still exists rather than the reason it is the default.
-
-**Let the child choose the delivery mode per call.** Unchanged from the original rejection: the model would own scheduler pressure, and behavior would vary per call rather than per deployment.
-
-**Put the obligation only in the tool description.** A description is read while choosing among tools. The child this change targets is not choosing a tool; it believes it is done. Prompt guidance is the surface that reaches that decision.
-
-**Enforce the obligation at settlement by rejecting a silent child.** There is nothing to reject: by the time settlement is observable the child's loop is over, and failing its teardown would destroy work rather than deliver it. Delivering the terminal facts unconditionally from the runtime is the answer to that case, and it belongs to the continuation manager, not to this package.
-
-## Consequences
-
- Every continuable in-process child with this package loaded carries one extra prompt section and a longer `report` description in every request; no other Agent's request changes.
- The default deployment wakes the parent once per accepted report. A nested tree that reports frequently consumes extra parent requests, while reports waiting together share one step; `quiet` is the documented escape.
- `installReportTool` requires `ctx.systemPrompt` in the child scope, so the package declares `systemPrompt` in `inject` and fails at load rather than at the next child materialization.
- Unit coverage pins the new default, two load-bearing instruction phrases, the section's child-only scope against both the parent and a sibling, and rollback or revocation of both registrations.
- Three assembled ACP scenarios with continuable children pin the complete instruction text through the new sidecar; a future change to any child-scoped section fails those scenarios instead of passing silently.
-
-### Accepted risks
-
-Next-step delivery by default amplifies model work in deep trees. The deployment owns that through `reportDelivery`; reports waiting together share one step, and one accepted report causes at most one wake.
-
-A child can still finish without reporting, and this change cannot detect it. Only the runtime's own [settlement account](2026-08-06-manager-owned-subagent-settlement-delivery.md) closes that case.
@@ -1,60 +0,0 @@
-# Agent Note: 可继续 child 的返回通道是一项义务
-
-Status: implemented
-
-[English](2026-08-06-continuable-child-report-obligation.md) | 中文
-
-## 问题
-
-可继续后台 child 拥有自己的 Session，因此它写在那里的任何内容都不会到达启动它的 agent。[report 工具](2026-07-30-continuable-subagent-report-tool.zh.md)为该 child 提供了一条返回通道，却把它呈现为若干选项之一：schema 里写着「可调用零次或多次」，child 的提示词中没有任何地方要求它调用该工具，而已采纳的默认调度（`quiet`）会把报告加入已停驻 parent 的下一次请求，却不唤醒它。
-
-这些选择单独看都站得住脚。合在一起，它们让这条返回通道无法作为委派契约使用。一个完成工作、把答案写进自己 transcript（文本记录）随后停止的 child，会让 parent 一无所获；而确实上报了的 child，面对的是一个已经停驻、要等到别的事件把它唤醒才会读到报告的 parent。外部反馈中的 parent 忙轮询 `list_agents`、反复向已结算 child 发送消息、以及放弃 `subagent` 改用 `workflow`，都可归结为同一处缺失的保证。
-
-## 决策
-
-返回通道是 child 收到的一条指令，而不是它需要自行发现的能力。report 包会向每个可继续进程内 child 安装两项作用域局部注册，并由同一个 disposer 撤销两者：
-
- `report` 工具，其描述现在说明 child 要在结束前调用一次并给出自足的最终结果，并在部分进展会改变 parent 下一步动作时提前调用；
- 一个 first-party order 为 2900 的 `tool:report` 系统提示词 section，用 child 自己的语气承载同一条义务，使从不细读工具描述的 child 仍能收到它。
-
-`reportDelivery` 的默认值为 `next-step`。一条被接受的报告会唤醒停驻的 parent driver，或加入运行中 parent 最近的 step 边界，与发现会改变 parent 下一步动作时上报的指令一致。对于宁可让报告无人阅读也要避免模型工作量放大的部署，`quiet` 依旧可用。[报告与结算顺序决策](../bug-fix/2026-08-17-subagent-report-settlement-ordering.zh.md)负责调度理由。
-
-### 为什么 section 与描述同时存在
-
-两者针对不同的失效模式。工具描述是在模型已经在考虑 `report` 时被读到的；提示词 section 是在它判断自己是否已经完成时被读到的。这条义务必须同时出现在两处，因为本次修复的失效——child 直接停下——发生在第二处。
-
-该 section 注册在 child 自己的作用域上，与[child 组合](../../../../packages/subagent/subagent/src/child-agent.ts)为遮蔽式 persona 已经使用的机制相同，因此 parent 与所有同级都看不到该工具与该指引。工具注册失败时，`installReportTool` 会回滚该 section；它返回的 disposer 会先尝试撤销两项注册，再抛出清理失败。
-
-### 是指令，不是强制
-
-没有任何东西会拒绝一个从不上报的 child。没有任何运行时路径会检查是否发送过报告，`report` 仍接受一个轮次中调用零次或多次。本次改动是面向模型的措辞加上一个调度默认值；服务权限、确认与恢复契约都保持不变。
-
-这条边界是刻意划定的：提示词文本只能到达仍在运行自身循环的 child。被错误、token 上限、取消或拆卸终止的 child 根本没有机会遵守，因此运行时会自己记录结算这件事，而不是信任这条指令（见[由管理器负责的结算投递](2026-08-06-manager-owned-subagent-settlement-delivery.zh.md)）。
-
-### 快照覆盖
-
-整体组装的 ACP `subagent-report` 场景演练随附的默认行为：child 在 parent 处于 maintenance 时上报，稍后的结算通知排在其后，而恢复的 parent 会先领取 next-step 报告、再领取 next-turn 结算。由于该 child 的作用域组合出类别 pin 无法描述的提示词，快照 harness 提供 `pinsChildSystemPrompts`，它与 `pinsChildToolSchemas` 完全对称：把一个 child fixture 的提示词移入 `system-prompt.<n>.expected.md`，其余请求 header 字段仍归类别 pin 所有，要求 sidecar 恰好在声明时存在，并拒绝与该类别 pin 完全相同的 sidecar，使冗余副本无法悄悄漂移。
-
-## 备选方案
-
-**保留 `quiet` 作为默认值，只依赖提示词。** 这曾是随附的立场，而它本身什么也没有解决：一条 parent 从不阅读的报告，与一条从未发送的报告无法区分。[report 工具 Agent Note](2026-07-30-continuable-subagent-report-tool.zh.md)对「始终唤醒」的否决，前提是 parent 还有别的理由去查看自己的上下文；已停驻的后台协调者并没有。轮次放大才是真正的代价，而它现在是 `quiet` 仍然保留的理由，而不是它作为默认值的理由。
-
-**让 child 按调用选择投递模式。** 与最初的否决相同：模型将掌握调度压力，行为也会随调用而非随部署变化。
-
-**只把义务写在工具描述里。** 描述是在从多个工具中选择时被读到的。本次改动针对的 child 并不在选择工具，它认为自己已经做完了。提示词指引才是能触及该判断的界面。
-
-**在结算时拒绝沉默的 child，以此强制该义务。** 没有什么可以拒绝：当结算可被观察时 child 的循环已经结束，让它的拆卸失败只会毁掉工作而不会送达结果。由运行时无条件投递终止事实才是这一情形的答案，而它属于继续执行管理器，不属于本包。
-
-## 后果
-
- 加载本包后，每个可继续进程内 child 的每次请求都会多出一个提示词 section 和一段更长的 `report` 描述；其他任何 Agent 的请求都不变。
- 默认部署会为每条被接受的报告唤醒 parent 一次。频繁上报的嵌套树会消耗额外的 parent 请求，而一起等待的报告会共享一个 step；`quiet` 是有文档记载的退路。
- `installReportTool` 需要 child 作用域中的 `ctx.systemPrompt`，因此本包在 `inject` 中声明 `systemPrompt`，从而在加载时失败，而不是等到下一次 child 物化时。
- 单元覆盖固定了新默认值、两处关键指令措辞、该 section 相对 parent 与同级均仅限 child 的作用域，以及两项注册在安装回滚或撤销时的清理。
- 三个带可继续 child 的整体组装 ACP 场景通过新的 sidecar 逐字固定完整的 child 提示词；今后任何对 child 作用域 section 的改动都会让这些场景失败，而不是悄悄通过。
-
-### 已接受的风险
-
-默认 next-step 投递会在深层树中放大模型工作量。部署通过 `reportDelivery` 掌握该取舍；一起等待的报告会共享一个 step，且每条被接受的报告至多产生一次唤醒。
-
-child 仍可能不上报就结束，本次改动无法检测这一点。只有运行时自己的[结算记账](2026-08-06-manager-owned-subagent-settlement-delivery.zh.md)才能补上这一情形。
@@ -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 .agents/notes/implemented/feature/2026-08-06-manager-owned-subagent-settlement-delivery.md
-2026-08-06-manager-owned-subagent-settlement-delivery.md: 27daa6d5150950efb50bf23dea945498651d2c09
-2026-08-06-manager-owned-subagent-settlement-delivery.zh.md: 1e3b866204d54e26b87061153fe97ffaed369141
+2026-08-06-manager-owned-subagent-settlement-delivery.md: 7dfa05c247ee0efb71963e057731c7ff6a5a1989
+2026-08-06-manager-owned-subagent-settlement-delivery.zh.md: 6bd8836d966c85ee0ef08ec19bbf96a4f72b0dd4
@@ -8,7 +8,7 @@ English | [中文](2026-08-06-manager-owned-subagent-settlement-delivery.zh.md)

 Continuable background delegation was the one asynchronous operation a model could start but could not reach the end of. Every other shape has a retrieval primitive or a return value: a background bash command and a one-shot background subagent both settle through a Task that `job_output(wait: true)` can block on, a workflow and a foreground subagent return their result to the caller. A continuable background child returned only its durable id, and nothing existed that a parent could wait on or would be handed.

-[The report obligation](2026-08-06-continuable-child-report-obligation.md) closed the cooperative half of that gap by instructing the child to report before it finishes. Instruction cannot close the rest. A child stopped by a token ceiling, a model failure, cancellation, or teardown never reaches the point where it could comply — not rarely, but never — and those are precisely the endings a waiting parent most needs to hear about. The observable downstream symptoms were parents busy-polling `list_agents`, re-sending messages to children that had already settled, and deployments abandoning `subagent` for `workflow` because a workflow at least returns something.
+Child-authored messages close the cooperative half of that gap: a child can send progress and a final handoff to its direct parent. Model choice cannot close the rest. A child stopped by a token ceiling, a model failure, cancellation, or teardown may never send such a message, and those are precisely the endings a waiting parent most needs to hear about. The observable downstream symptoms were parents busy-polling `list_agents`, re-sending messages to children that had already settled, and deployments abandoning `subagent` for `workflow` because a workflow at least returns something.

 The signal already existed. `subagent/end` has carried `stopReason` and `lastAssistantMessage` since continuable Activations shipped. What was missing was any consumer that turned it into context the parent's model could see.

@@ -20,7 +20,7 @@ When a resident Activation settles, `notifySettlement()` resolves the child's du

 ### Provenance

-The notice carries `{ kind: 'subagent-settled', form: 'notice', summary, senderSessionId }`. It is deliberately not the existing `subagent-report` kind. A report is content the child chose; this is the runtime stating what became of the child. Merging them would credit the child with words it never wrote, and would make a durable log unable to distinguish "the child said it was done" from "the harness observed that it stopped". The `notice` form also gives a UI the collapsed one-line presentation this message wants, where `relay` would present it as correspondence.
+The notice carries `{ kind: 'subagent-settled', form: 'notice', summary, senderSessionId }`. It is deliberately not the `agent-message` kind used by `send_message`. An Agent message is content the child chose; this is the runtime stating what became of the child. Merging them would credit the child with words it never wrote, and would make a durable log unable to distinguish "the child said it was done" from "the harness observed that it stopped". The `notice` form also gives a UI the collapsed one-line presentation this message wants, where `relay` presents Agent correspondence.

 ### Two ordering rules, and why the manager owns them

@@ -56,13 +56,13 @@ Both matter past the notice: `subagent/end` carries `stopReason` to the jsonrpc

 ### Snapshot coverage

-Three assembled ACP scenarios cover the notice: a child that never reports, a child that reports first, and a child driven through several follow-up turns. All three needed an explicit fence. The notice arrives once the child's teardown finishes, which races whatever the parent is already doing, so each scenario holds the child behind the parent's spawn turn and then waits for the parent turn the notice opens (`waitForTurnStart` at that turn, then `waitForTurnEnd`) before the script continues. Waiting for a turn the run is not fenced to produce is not coverage: it is a timeout when the notice lands in the turn already running instead.
+Three assembled ACP scenarios cover the notice: a child that sends no message, a child that sends a message first, and a child driven through several Agent-message turns. All three need an explicit fence. The notice arrives once the child's teardown finishes, which races whatever the parent is already doing, so each scenario holds the child behind the parent's spawn turn and then waits for the parent turn the notice opens (`waitForTurnStart` at that turn, then `waitForTurnEnd`) before the script continues. Waiting for a turn the run is not fenced to produce is not coverage: it is a timeout when the notice lands in the turn already running instead.

 `subagent-continuable` is the one that pins a failure. Its child's last turn dies on the forced durability checkpoint without entering a step, so that transcript is where the stop-reason rule above is visible end to end: the notice says the child *failed*, carries the earlier `SECOND_OK` as its last content rather than as a result, and the parent's own acknowledgement turn reaches the ACP client.

-A keyless headless Loader snapshot covers the user-visible path end to end. Its replay parent omits `run_in_background` to exercise the continuable background default, never calls `list_agents`, `send_message`, or Task tools, consumes the manager-authored `subagent-settled` notice, and produces its final answer. The child never calls `report`, so the transcript cannot pass through the cooperative report path. A test-only Loader fence holds the parent's post-spawn request until the real manager notice enters its inbox, removing platform scheduling from the transcript without synthesizing the notice.
+A keyless headless Loader snapshot covers the user-visible path end to end. Its replay parent omits `run_in_background` to exercise the continuable background default, never calls `list_agents`, `send_message`, or Task tools, consumes the manager-authored `subagent-settled` notice, and produces its final answer. The child sends no Agent message, so the transcript depends only on the runtime notice. A test-only Loader fence holds the parent's post-spawn request until the real manager notice enters its inbox, removing platform scheduling from the transcript without synthesizing the notice.

-The `subagent-report` scenario uses the default next-step report delivery. A snapshot-only fence holds the child until the parent's spawn turn ends, then holds the parent in maintenance until settlement follows the report. The resumed parent claims the next-step report before the queued next-turn settlement. The [report/settlement ordering decision](../bug-fix/2026-08-17-subagent-report-settlement-ordering.md) owns this cross-state ordering.
+The `subagent-send-message` scenario holds the child until the parent's spawn turn ends, then holds the parent in maintenance until settlement follows the child-authored message. The resumed parent claims the next-step Agent message before the queued next-turn settlement. The [message/settlement ordering decision](../bug-fix/2026-08-17-subagent-message-settlement-ordering.md) owns this cross-state ordering.

 The refusal and interruption wordings are pinned verbatim in unit tests rather than in a replayed transcript: producing them needs a rejecting policy plugin or a cancellation fenced at a step boundary, which the keyless assemblies do not otherwise carry, and the assembled scenarios already pin the notice pathway itself end to end.

@@ -72,7 +72,7 @@ The refusal and interruption wordings are pinned verbatim in unit tests rather t

 **Attach an external `subagent/end` listener.** Rejected on three counts above — no parent in the payload, a disposed child handle, and an ordering the listener cannot influence. A listener would also have to be strictly synchronous to beat the release, and nothing at that seam enforces it, so the correct version would be correct only by accident.

-**Deliver only when the child did not report.** This was the first design. It needs per-Activation bookkeeping, still misses the child that reported progress and then died before its result, and — decisively — makes the parent-facing promise conditional. "Usually you are told" is not a contract a tool description can state, and a model that cannot rely on the notice will poll anyway.
+**Deliver only when the child sent no message.** This was the first design. It needs per-Activation bookkeeping, still misses the child that sent progress and then died before its result, and — decisively — makes the parent-facing promise conditional. "Usually you are told" is not a contract a tool description can state, and a model that cannot rely on the notice will poll anyway.

 **Make delivery configurable.** A deployment switch would return the model-facing text to "usually", which is the failure this change exists to remove. Protocol constants and safety invariants stay fixed; this is one of them.

@@ -87,8 +87,8 @@ The refusal and interruption wordings are pinned verbatim in unit tests rather t
 - `Activation` carries `parentSession` and `announced`. The first exists because the child handle is disposed before delivery; the second is what keeps a rolled-back materialization silent.
 - `foldConsumedWork()` replaces `dsh-session`'s `findLastMessageTurnEnd()` and moves to `dsh-agent`, which owns the inbox marker it reads; the one-shot in-process path folds the same answer and does not classify a cut-short one-shot child as `completed`.
 - Unit coverage pins the unconditional contract, each terminal reason, idle and busy scheduling, the batch, the maintenance regression, the pre-release ordering, a parent that is gone, and a rejected send that must not fail teardown.
- Three ACP scenarios use an explicit settlement fence, and `subagent-report` pins the default report-before-settlement next-step order.
- A keyless headless Loader snapshot pins background start → manager-authored settlement notice → final parent answer with no polling or child `report` call.
+- Three ACP scenarios use an explicit settlement fence, and `subagent-send-message` pins the Agent-message-before-settlement next-step order.
+- A keyless headless Loader snapshot pins background start → manager-authored settlement notice → final parent answer with no polling or child-authored message.

 ### Accepted risks

@@ -100,4 +100,4 @@ Stop-reason attribution is a best effort over the log's existing splice vocabula

 Turn amplification is real for deep or wide trees, and it is not configurable by design. The step-boundary batch bounds it for simultaneous settlement but not for children that settle apart.

-Reports and their later settlement notices are ordered through the parent's next-step FIFO. Independent settlements from sibling children retain their actual delivery order rather than a synthetic sibling ordering.
+Agent messages and their later settlement notices are ordered through the parent's next-step FIFO. Independent settlements from sibling children retain their actual delivery order rather than a synthetic sibling ordering.
@@ -8,7 +8,7 @@ Status: implemented

 可继续后台委派是模型唯一一种能够发起、却无法抵达终点的异步操作。其他每一种形态都有取回原语或返回值：后台 bash 命令与一次性后台 subagent 都通过 Task 结算，`job_output(wait: true)` 可以阻塞等待；workflow 与前台 subagent 会把结果返回给调用方。可继续后台 child 只返回它持久化的 id，而父级既没有可等待的对象，也不会被交付任何东西。

-[报告义务](2026-08-06-continuable-child-report-obligation.zh.md)通过要求 child 在结束前上报，补上了这一缺口中协作的那一半。指令无法补上其余部分。被 token 上限、模型失败、取消或拆卸终止的 child 永远走不到能够遵守的那一步——不是很少，而是从不——而这些恰恰是等待中的父级最需要被告知的结束方式。可观察到的下游症状包括：父级忙轮询 `list_agents`、向已经结算的 child 反复发送消息，以及部署放弃 `subagent` 转用 `workflow`，因为 workflow 至少会返回点什么。
+由 child 编写的消息补上了这一缺口中协作的那一半：child 可以向直接 parent 发送进度与最终交接。模型选择无法补上其余部分。被 token 上限、模型失败、取消或拆卸终止的 child 可能永远不会发送这种消息，而这些恰恰是等待中的父级最需要被告知的结束方式。可观察到的下游症状包括：父级忙轮询 `list_agents`、向已经结算的 child 反复发送消息，以及部署放弃 `subagent` 转用 `workflow`，因为 workflow 至少会返回点什么。

 信号本身早就存在。自可继续 Activation 发布以来，`subagent/end` 就一直携带 `stopReason` 与 `lastAssistantMessage`。缺的是把它变成父级模型能看到的上下文的那个消费者。

@@ -20,7 +20,7 @@ Status: implemented

 ### 来源信息

-该通知携带 `{ kind: 'subagent-settled', form: 'notice', summary, senderSessionId }`，刻意不复用既有的 `subagent-report` kind。上报是 child 选择的内容；这条消息则是运行时在陈述这个 child 后来怎样了。把两者合并会把 child 从未写过的话算到它头上，也会让持久化日志无法区分「child 说它做完了」和「harness 观察到它停下了」。`notice` 形态还为 UI 提供了这条消息想要的折叠单行呈现，而 `relay` 会把它呈现为往来信件。
+该通知携带 `{ kind: 'subagent-settled', form: 'notice', summary, senderSessionId }`，刻意不复用 `send_message` 使用的 `agent-message` kind。Agent 消息是 child 选择的内容；这条消息则是运行时在陈述这个 child 后来怎样了。把两者合并会把 child 从未写过的话算到它头上，也会让持久化日志无法区分「child 说它做完了」和「harness 观察到它停下了」。`notice` 形态还为 UI 提供了这条消息想要的折叠单行呈现，而 `relay` 会把 Agent 往来消息呈现为通信。

 ### 两条顺序规则，以及为什么归管理器所有

@@ -56,13 +56,13 @@ Status: implemented

 ### 快照覆盖

-三个整体组装的 ACP 场景覆盖该通知：一个从不上报的 child、一个先上报的 child，以及一个被多轮 follow-up 驱动的 child。三者都需要显式栅栏。通知在 child 拆卸完成后才到达，会与父级当时正在做的事竞争，因此每个场景都会把 child 保持到父级启动轮次结束，再等待该通知开启的那个父级轮次（先 `waitForTurnStart` 到该轮次，再 `waitForTurnEnd`），然后脚本才继续。等待一个运行并未被栅栏保证会产生的轮次不算覆盖：一旦通知落进已经在跑的那个轮次，它就是一次超时。
+三个整体组装的 ACP 场景覆盖该通知：一个不发送消息的 child、一个先发送消息的 child，以及一个被多轮 Agent 消息驱动的 child。三者都需要显式栅栏。通知在 child 拆卸完成后才到达，会与父级当时正在做的事竞争，因此每个场景都会把 child 保持到父级启动轮次结束，再等待该通知开启的那个父级轮次（先 `waitForTurnStart` 到该轮次，再 `waitForTurnEnd`），然后脚本才继续。等待一个运行并未被栅栏保证会产生的轮次不算覆盖：一旦通知落进已经在跑的那个轮次，它就是一次超时。

 `subagent-continuable` 是其中固定失败结局的那个。它的 child 最后一个轮次在被强制的持久化检查点上死亡，且未进入任何 step，因此该 transcript 正是上面那条终止原因规则的端到端可见之处：通知说该 child **失败**，把此前的 `SECOND_OK` 作为它最后产出的内容而非结果携带，而父级自己的确认轮次会到达 ACP 客户端。

-另有一个无密钥的 headless Loader 快照端到端覆盖用户可见路径。其重放父级省略 `run_in_background` 以覆盖可继续后台默认路径，从不调用 `list_agents`、`send_message` 或 Task 工具，消费管理器写入的 `subagent-settled` 通知，并给出最终答案。child 从不调用 `report`，因此该 transcript 不可能经由协作式上报路径通过。一个仅用于测试的 Loader 栅栏会把父级启动后的请求保持到真实管理器通知进入其 inbox 为止，从 transcript 中排除平台调度差异，但不会伪造该通知。
+另有一个无密钥的 headless Loader 快照端到端覆盖用户可见路径。其重放父级省略 `run_in_background` 以覆盖可继续后台默认路径，从不调用 `list_agents`、`send_message` 或 Task 工具，消费管理器写入的 `subagent-settled` 通知，并给出最终答案。child 不发送 Agent 消息，因此该 transcript 只依赖运行时通知。一个仅用于测试的 Loader 栅栏会把父级启动后的请求保持到真实管理器通知进入其 inbox 为止，从 transcript 中排除平台调度差异，但不会伪造该通知。

-`subagent-report` 场景使用默认 next-step 报告投递。一个仅用于快照的围栏会让 child 等到 parent 的派生轮次结束，随后让 parent 保持 maintenance，直至结算跟在报告之后到达。恢复的 parent 会先领取 next-step 报告、再领取排队的 next-turn 结算。[报告与结算顺序决策](../bug-fix/2026-08-17-subagent-report-settlement-ordering.zh.md)负责说明这种跨状态顺序。
+`subagent-send-message` 场景会让 child 等到 parent 的派生轮次结束，随后让 parent 保持 maintenance，直至结算跟在 child 编写的消息之后到达。恢复的 parent 会先领取 next-step Agent 消息、再领取排队的 next-turn 结算。[消息与结算顺序决策](../bug-fix/2026-08-17-subagent-message-settlement-ordering.zh.md)负责说明这种跨状态顺序。

 拒绝与中断两种措辞在单元测试中逐字钉死，而不进入重放 transcript：触发它们需要一个会拒绝的策略插件、或一次在 step 边界被栅栏卡住的取消，而无密钥组装本身并不携带这些；通知通路本身已由整体组装场景端到端钉住。

@@ -72,7 +72,7 @@ Status: implemented

 **挂一个外部 `subagent/end` listener。** 因上文三点被否决——payload 里没有父级、child handle 已被 dispose，以及 listener 无法影响的顺序。listener 还必须严格同步才能抢在释放之前，而该 seam 上没有任何东西强制这一点，因此正确的版本只能靠碰巧正确。

-**仅在 child 没有上报时投递。** 这是最初的设计。它需要按 Activation 记账，仍会漏掉「报了进度、随后在给出结果前死掉」的 child，而且最关键的是：它让面向父级的承诺变成有条件的。「通常你会被告知」不是工具描述能陈述的契约，而无法依赖该通知的模型无论如何都会去轮询。
+**仅在 child 没有发送消息时投递。** 这是最初的设计。它需要按 Activation 记账，仍会漏掉「发送了进度、随后在给出结果前死掉」的 child，而且最关键的是：它让面向父级的承诺变成有条件的。「通常你会被告知」不是工具描述能陈述的契约，而无法依赖该通知的模型无论如何都会去轮询。

 **把投递做成可配置。** 部署开关会把面向模型的文本重新变回「通常」，而这正是本次改动要消除的失效。协议常量与安全不变量保持固定；这就是其中之一。

@@ -87,8 +87,8 @@ Status: implemented
 - `Activation` 携带 `parentSession` 与 `announced`。前者存在是因为 child handle 在投递前已被 dispose；后者让被回滚的物化保持静默。
 - `foldConsumedWork()` 取代 `dsh-session` 的 `findLastMessageTurnEnd()`，并迁移到 `dsh-agent`——它拥有该 fold 所读取的 inbox 标记；一次性 in-process 路径折叠同一个答案，不会把被中途切断的一次性 child 归类为 `completed`。
 - 单元覆盖固定了无条件约定、每种终止原因、空闲与繁忙两种调度、批量语义、维护期回归、释放前顺序、父级已消失，以及一次不得让拆卸失败的发送被拒。
- 三个 ACP 场景使用显式的结算围栏，`subagent-report` 固定默认的报告先于结算的 next-step 顺序。
- 一个无密钥的 headless Loader 快照固定了「后台启动 → 管理器写入的结算通知 → 父级最终答案」路径，其中没有轮询，也没有 child `report` 调用。
+- 三个 ACP 场景使用显式的结算围栏，`subagent-send-message` 固定 Agent 消息先于结算的 next-step 顺序。
+- 一个无密钥的 headless Loader 快照固定了「后台启动 → 管理器写入的结算通知 → 父级最终答案」路径，其中没有轮询，也没有 child 编写的消息。

 ### 已接受的风险

@@ -100,4 +100,4 @@ Status: implemented

 对于深或宽的树，轮次放大是真实存在的，而且按设计不可配置。step 边界的批量语义只能约束同时结算的情形，无法约束分散结算的 child。

-报告与其稍后的结算通知通过 parent 的 next-step FIFO 排序。来自同级 child 的独立结算保留其实际投递顺序，不会虚构同级间的顺序。
+Agent 消息与其稍后的结算通知通过 parent 的 next-step FIFO 排序。来自同级 child 的独立结算保留其实际投递顺序，不会虚构同级间的顺序。
@@ -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 .agents/notes/implemented/feature/2026-08-10-continuable-subagent-policy-inheritance.md
-2026-08-10-continuable-subagent-policy-inheritance.md: a33a211747cf19cd465c56b8bda6ebfe1d7b6e06
-2026-08-10-continuable-subagent-policy-inheritance.zh.md: af43bbdc10c926b5255e8e5af98f7ea950b60091
+2026-08-10-continuable-subagent-policy-inheritance.md: e567c0557be45ba9e0e0515d742bd6f99a5f10e7
+2026-08-10-continuable-subagent-policy-inheritance.zh.md: 5c9cc677e9154d5b14214979283c00fd28a4c2df
@@ -6,7 +6,7 @@ English | [中文](2026-08-10-continuable-subagent-policy-inheritance.zh.md)

 ## Problem

-The one-shot in-process driver has seeded parent sandbox/approval overrides into its children since the [in-process policy-inheritance decision](2026-07-25-subagent-policy-inheritance.md), but the continuable path never did: `SubagentContinuationManager` materialization applied only child composition and the activation setup registry. The default bundle wires both delegation tools as `backgroundMode: continuable`, so in a default deployment every background child silently fell back to deployment defaults — a parent switched to `danger-full-access` produced children stuck at `workspace-write` whose every out-of-workspace operation raised an approval prompt, and a parent's unattended `'never'` approval stance reverted to prompting ([dsh-external/issues#334](https://github.com/dsh-external/issues/issues/334)).
+The one-shot in-process driver has seeded parent sandbox/approval overrides into its children since the [in-process policy-inheritance decision](2026-07-25-subagent-policy-inheritance.md), but the continuable path never did: `SubagentContinuationManager` materialization applied only child composition. The default bundle wires both delegation tools as `backgroundMode: continuable`, so in a default deployment every background child silently fell back to deployment defaults — a parent switched to `danger-full-access` produced children stuck at `workspace-write` whose every out-of-workspace operation raised an approval prompt, and a parent's unattended `'never'` approval stance reverted to prompting ([dsh-external/issues#334](https://github.com/dsh-external/issues/issues/334)).

 ## Decision

@@ -16,7 +16,7 @@ The capture/append pair moved from the one-shot driver into the seam's shared ch

 ## Alternatives considered

- **An activation-setup-registry contribution** (`registerContinuableSetup`) — rejected: a contribution receives only the child context, so it cannot capture the parent's overrides at the delegation boundary; the registry applies on cold resume as well as fresh creation, which would re-append or re-capture; and nothing ties a contribution's capture to the start call's synchronous prefix, so the pre-await capture guarantee would be lost.
+- **A generic child-setup contribution** — rejected: a contribution receives only the child context, so it cannot capture the parent's overrides at the delegation boundary; applying it on cold resume as well as fresh creation would re-append or re-capture; and nothing ties its capture to the start call's synchronous prefix, so the pre-await capture guarantee would be lost.
 - **Re-capturing the parent's overrides at cold resume** — rejected: a resumed child would silently change policy with the parent's later switches, breaking the snapshot-at-delegation semantic and making effective policy depend on resume timing instead of the child's own log. A parent that wants a resumed child under new policy re-delegates.
 - **Importing the one-shot driver's inline logic from the continuation manager** — rejected: the Service Definition package cannot depend on its own provider package, and duplicating the capture/append pair in `continuation.ts` invites drift; `child-agent.ts` already holds every other shared composition step.
 - **Seeding the events into the descriptor seed turn** — rejected: the capture value is not known when the seed is assembled for every caller, and the one-shot precedent already establishes unpublished-setup appends as the ordering that places inherited facts after fork history with `firstLiveSeq` intact.
@@ -6,7 +6,7 @@ Status: implemented

 ## 问题

-自[进程内策略继承决策](2026-07-25-subagent-policy-inheritance.zh.md)以来，一次性进程内驱动器一直会把父级的沙箱／审批覆盖项注入其子级，但可继续路径从未这样做：`SubagentContinuationManager` 的物化只应用子级组合与 Activation（激活）设置注册表。默认组合包把两个委派工具都配置为 `backgroundMode: continuable`，因此在默认部署中，每个后台子 agent（智能体）都静默回退到部署默认值：切换到 `danger-full-access` 的父级产出的子 agent 卡在 `workspace-write`，每次工作区外操作都会触发审批提示；父级无人值守的 `'never'` 审批立场也退回为发起提示的行为（[dsh-external/issues#334](https://github.com/dsh-external/issues/issues/334)）。
+自[进程内策略继承决策](2026-07-25-subagent-policy-inheritance.zh.md)以来，一次性进程内驱动器一直会把父级的沙箱／审批覆盖项注入其子级，但可继续路径从未这样做：`SubagentContinuationManager` 的物化只应用子级组合。默认组合包把两个委派工具都配置为 `backgroundMode: continuable`，因此在默认部署中，每个后台子 agent（智能体）都静默回退到部署默认值：切换到 `danger-full-access` 的父级产出的子 agent 卡在 `workspace-write`，每次工作区外操作都会触发审批提示；父级无人值守的 `'never'` 审批立场也退回为发起提示的行为（[dsh-external/issues#334](https://github.com/dsh-external/issues/issues/334)）。

 ## 决策

@@ -16,7 +16,7 @@ Status: implemented

 ## 考虑过的替代方案

- **一项 Activation 设置注册表贡献**（`registerContinuableSetup`）：不予采纳。贡献只接收子级上下文，因此无法在委派边界捕获父级的覆盖项；该注册表在冷恢复与全新创建时都会应用，会导致重复追加或重复捕获；而且没有任何机制把贡献的捕获绑定到 start 调用的同步前缀，await 前捕获的保证会因此丢失。
+- **一项通用 child 设置贡献**：不予采纳。贡献只接收子级上下文，因此无法在委派边界捕获父级的覆盖项；在冷恢复与全新创建时都应用它会导致重复追加或重复捕获；而且没有任何机制把它的捕获绑定到 start 调用的同步前缀，await 前捕获的保证会因此丢失。
 - **在冷恢复时重新捕获父级覆盖项**：不予采纳。恢复的子 agent 会随父级后续切换静默改变策略，这会破坏委派时快照的语义，并让生效策略取决于恢复时机而非子级自身的日志。希望恢复的子 agent 采用新策略的父级应重新委派。
 - **让继续执行管理器导入一次性驱动器的内联逻辑**：不予采纳。Service Definition 包不能依赖自己的提供方包，而在 `continuation.ts` 中复制捕获／追加这对函数会招致偏差；`child-agent.ts` 已经承载其余每个共享组合步骤。
 - **把这些事件写入描述符种子轮次**：不予采纳。种子为每个调用方组装时，捕获值尚不可知；而且一次性路径的先例已经确立：在未发布的设置阶段追加，才是把继承事实排在 fork 历史之后、同时保持 `firstLiveSeq` 不变的顺序。
@@ -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 .agents/notes/implemented/feature/2026-08-11-background-first-continuable-delegation.md
-2026-08-11-background-first-continuable-delegation.md: 59232ae8821ef4a093fd610ecbbb39690316ce6c
-2026-08-11-background-first-continuable-delegation.zh.md: d39f129e0ceb4aa75ba91c6860fb017c69f0fbe6
+2026-08-11-background-first-continuable-delegation.md: a3dcd75c820742eda622b04ce6077b444f66846b
+2026-08-11-background-first-continuable-delegation.zh.md: f0cbdca1c1cba7400e6f9926453fb640bde575bb
@@ -8,7 +8,7 @@ English | [中文](2026-08-11-background-first-continuable-delegation.zh.md)

 A continuable child already has a durable id, independent turns, follow-up messaging, and a manager-owned settlement notice. Treating an omitted `run_in_background` as foreground makes that lifecycle depend on the model restating `true` on every call. It also obscures the useful scheduling test: the parent should wait only when its next action requires the child's result.

-The child-scoped `report` prompt requires a self-contained final report, while [manager-owned settlement delivery](2026-08-06-manager-owned-subagent-settlement-delivery.md) independently sends the run outcome and closing message. A completed child can therefore wake its parent with a final report and again with settlement. Background-first scheduling must preserve both deliveries: the child-authored handoff remains mandatory guidance, while the manager-authored notice covers every terminal path regardless of model compliance.
+The child's initial task tells it how to address its direct parent with the shared `send_message` tool, while [manager-owned settlement delivery](2026-08-06-manager-owned-subagent-settlement-delivery.md) independently sends the run outcome and closing message. A child may send progress or a final handoff before settlement. Background-first scheduling preserves both: Agent-authored messages remain explicit model choices, while the manager-authored notice covers every terminal path regardless of model compliance.

 ## Decision

@@ -20,9 +20,9 @@ The model-facing text divides responsibility by location:
 - the `run_in_background` parameter states the lifecycle-specific default and when to override it;
 - a `tool:<toolName>` system-prompt section tells the model to start independent delegations together, continue useful work while they run, and choose foreground only when the next action depends on the result. The section renders only when that tool remains visible in the assembly scope, so a child tool restriction removes the schema and its guidance together.

-The [continuable child report obligation](2026-08-06-continuable-child-report-obligation.md) remains unchanged: the child prompt requires one self-contained final report and earlier reports for findings that change the parent's next action. Manager-owned settlement remains unconditional and does not inspect whether a report arrived. The two messages may repeat final content, but they retain distinct authors and purposes: `report` is the child's explicit handoff, while settlement records how the run ended and preserves terminal output when the child cannot cooperate. `reportDelivery` remains deployment scheduling policy with `next-step` as its default, preserving report-before-settlement order through the parent inbox.
+The child receives its direct parent id and return guidance in the initial task after any inherited fork seed. It may call `send_message` zero or more times, including for findings that change the parent's next action and for a self-contained final handoff. Manager-owned settlement remains unconditional and does not inspect whether an Agent message arrived. The two messages may repeat final content, but they retain distinct authors and purposes: `send_message` carries content the child chose, while settlement records how the run ended and preserves terminal output when the child cannot cooperate. Both use the Agent inbox and fixed Steer scheduling; the accepted child message precedes the later settlement notice.

-The keyless headless `subagent-settlement` scenario omits `run_in_background`, receives the immediate child id, and reaches the final parent answer through the manager-authored settlement notice even though its fixture deliberately does not call `report`. Package tests separately pin explicit `false` as foreground, the parent scheduling text, and the child's mandatory-report prompt.
+The keyless headless `subagent-settlement` scenario omits `run_in_background`, receives the immediate child id, and reaches the final parent answer through the manager-authored settlement notice even though its fixture deliberately sends no child-authored message. Package tests separately pin explicit `false` as foreground, the parent scheduling text, and the child's parent-id return guidance.

 ## Alternatives considered

@@ -32,14 +32,14 @@ The keyless headless `subagent-settlement` scenario omits `run_in_background`, r

 **Change only the prompt.** Prompt preference without runtime resolution still turns an omitted argument into foreground. The model must be able to rely on the advertised default rather than reproduce it perfectly on every tool call.

-**Suppress settlement after a final report arrives.** Conditional settlement reintroduces per-Activation bookkeeping and loses the unconditional runtime guarantee when a child reports progress and then fails. Settlement remains unconditional even when the resulting message overlaps a final report.
+**Suppress settlement after a final Agent message arrives.** Conditional settlement reintroduces per-Activation bookkeeping and loses the unconditional runtime guarantee when a child sends progress and then fails. Settlement remains unconditional even when the resulting message overlaps a final handoff.

-**Use `report` only for progress before settlement.** This removes duplicate final content but also removes the explicit child-authored handoff from the child prompt. The final-report obligation remains, and runtime settlement remains its independent fallback and terminal record.
+**Reserve `send_message` for progress before settlement.** This removes duplicate final content but makes the shared adjacent-Agent operation depend on message purpose. The child may explicitly hand off a final result, while runtime settlement remains its independent fallback and terminal record.

 ## Consequences

 - An ordinary continuable call is non-blocking without spelling `run_in_background: true`; serialized delegation is an explicit `false` choice.
 - Independent subagent calls in one assistant message overlap under the tool loop's concurrency-safe dispatch, while dependent foreground calls can still be issued one at a time.
 - Parent guidance, tool schema, runtime resolution, and settlement delivery state the same default.
- A compliant child reports one self-contained final result and may report important findings earlier. Every Activation also produces an unconditional settlement notice, so a completed run may deliver overlapping final content twice.
+- A child may send one self-contained final result and important findings earlier. Every Activation also produces an unconditional settlement notice, so a completed run may deliver overlapping final content twice.
 - One-shot background Jobs and disabled-background tool instances retain their existing behavior.
@@ -8,7 +8,7 @@ Status: implemented

 可继续 child 已经具备持久化 id、独立轮次、后续消息以及由管理器负责的结算通知。如果把省略的 `run_in_background` 视为前台，模型就必须在每次调用时重复写出 `true`，才能得到这套生命周期。这样也会掩盖真正有用的调度判断：只有当 parent 的下一步动作需要 child 结果时，parent 才应等待。

-child 作用域的 `report` 提示词要求发送自包含的最终报告，而[由管理器负责的结算投递](2026-08-06-manager-owned-subagent-settlement-delivery.zh.md)会独立发送本次运行的结束结果与收尾消息。已完成的 child 因而可能先用最终报告唤醒 parent，再用结算通知唤醒一次。后台优先调度会保留两次投递：由 child 编写的交接仍是强制提示词指引，由管理器生成的通知则不依赖模型是否遵循指令，覆盖每种终止路径。
+child 的初始任务会告诉它如何使用共享的 `send_message` 工具向直接 parent 发送消息，而[由管理器负责的结算投递](2026-08-06-manager-owned-subagent-settlement-delivery.zh.md)会独立发送本次运行的结束结果与收尾消息。child 可以在结算前发送进度或最终交接。后台优先调度会保留两者：由 Agent 编写的消息仍是模型的显式选择，由管理器生成的通知则不依赖模型是否遵循指令，覆盖每种终止路径。

 ## 决策

@@ -20,9 +20,9 @@ child 作用域的 `report` 提示词要求发送自包含的最终报告，而[
 - `run_in_background` 参数说明具体生命周期的默认值以及何时覆盖；
 - `tool:<toolName>` 系统提示词 section 会告诉模型同时启动相互独立的委派、在它们运行时继续有用工作，并且仅当下一步动作依赖结果时选择前台。只有当该工具在组装作用域中仍可见时才会渲染这个 section，因此子级工具限制会同时移除 schema 与对应指引。

-[可继续 child 上报义务](2026-08-06-continuable-child-report-obligation.zh.md)保持不变：child 提示词要求发送一份自包含的最终报告，并在发现会改变 parent 下一步动作的信息时提前报告。由管理器负责的结算仍然无条件执行，不检查报告是否已经到达。这两条消息可能重复最终内容，但作者和用途不同：`report` 是 child 的显式交接，结算则记录本次运行如何结束，并在 child 无法配合时保留终止输出。`reportDelivery` 仍是部署调度策略，默认值为 `next-step`，通过 parent inbox 保持报告先于结算的顺序。
+child 会在继承的 fork 种子之后，从初始任务获得直接 parent id 与返回指引。它可以调用零次或多次 `send_message`，包括发送会改变 parent 下一步动作的发现，以及自包含的最终交接。由管理器负责的结算仍然无条件执行，不检查 Agent 消息是否已经到达。这两条消息可能重复最终内容，但作者和用途不同：`send_message` 携带 child 自己选择的内容，结算则记录本次运行如何结束，并在 child 无法配合时保留终止输出。两者都使用 Agent inbox 与固定 Steer 调度；已接受的 child 消息先于后续结算通知。

-无密钥 headless `subagent-settlement` 场景省略 `run_in_background`，收到立即返回的 child id；尽管 fixture（测试前置数据）有意不调用 `report`，它仍通过管理器生成的结算通知到达 parent 最终答案。包测试另行固定了显式 `false` 的前台语义、parent 调度文本以及 child 的强制报告提示词。
+无密钥 headless `subagent-settlement` 场景省略 `run_in_background`，收到立即返回的 child id；尽管 fixture（测试前置数据）有意不发送 child 编写的消息，它仍通过管理器生成的结算通知到达 parent 最终答案。包测试另行固定了显式 `false` 的前台语义、parent 调度文本以及 child 的 parent-id 返回指引。

 ## 考虑过的替代方案

@@ -32,14 +32,14 @@ child 作用域的 `report` 提示词要求发送自包含的最终报告，而[

 **只修改提示词。** 如果运行时解析不变，提示词偏好仍会让省略参数的调用进入前台。模型必须能够依赖公布的默认值，而不是在每次工具调用中完美复述它。

-**最终报告到达后抑制结算通知。** 条件结算会重新引入每次 Activation 的记账，并且当 child 先报告进度、随后失败时丢掉无条件运行时保证。即使生成的消息与最终报告重叠，结算仍然无条件执行。
+**最终 Agent 消息到达后抑制结算通知。** 条件结算会重新引入每次 Activation 的记账，并且当 child 先发送进度、随后失败时丢掉无条件运行时保证。即使生成的消息与最终交接重叠，结算仍然无条件执行。

-**只用 `report` 发送结算前的进度。** 这样可以消除重复的最终内容，但也会从 child 提示词中移除由 child 编写的显式交接。最终报告义务保持不变，运行时结算则继续作为它的独立后备和终止记录。
+**只用 `send_message` 发送结算前的进度。** 这样可以消除重复的最终内容，但会让共享的相邻 Agent 操作依赖消息用途。child 可以显式交接最终结果，运行时结算则继续作为它的独立后备和终止记录。

 ## 后果

 - 普通可继续调用无需写出 `run_in_background: true` 即为非阻塞；串行委派需要显式选择 `false`。
 - 同一条 assistant 消息中的独立 subagent 调用会在工具循环的并发安全分发下重叠执行；有依赖的前台调用仍可逐个发出。
 - parent 指引、工具 schema、运行时解析和结算投递陈述同一个默认值。
- 遵循指令的 child 会发送一份自包含的最终结果，也可以更早报告重要发现。每次 Activation 还会产生无条件结算通知，因此已完成的运行可能两次投递相互重叠的最终内容。
+- child 可以发送一份自包含的最终结果，也可以更早发送重要发现。每次 Activation 还会产生无条件结算通知，因此已完成的运行可能两次投递相互重叠的最终内容。
 - 一次性后台 Task 与禁用后台的工具实例保留现有行为。
@@ -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 .agents/notes/implemented/feature/2026-08-20-web-streaming-fence-highlight.md
-2026-08-20-web-streaming-fence-highlight.md: ccc961da1febba3611e3e558087b586c6f1f474b
-2026-08-20-web-streaming-fence-highlight.zh.md: e5d29545bf659ed572d052f1212d68d344d49b5a
+2026-08-20-web-streaming-fence-highlight.md: aa44a8c7ea91ff4913d92449d62a6ebf9406e922
+2026-08-20-web-streaming-fence-highlight.zh.md: 88bae36928ac8d00960e529addeb4f30240e3579