Merge remote-tracking branch 'origin/master' into fix/subagent-empty-terminal-message-output

This commit is contained in:
Hypatia May
2026-08-10 21:59:42 +08:00
459 changed files with 6806 additions and 4766 deletions

View File

@@ -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/2026-08-08-npm-backed-git-repository-plugin-preparation.md
2026-08-08-npm-backed-git-repository-plugin-preparation.md: 958b932f82f4da3cf63aa911260411855e514409
2026-08-08-npm-backed-git-repository-plugin-preparation.zh.md: d2256e0eae303c371371b9b5ba1967105aa61834
# pnpm run verify-translation-pairing --write .agents/notes/implemented/bug-fix/2026-07-30-bounded-overwrite-diff-basis.md
2026-07-30-bounded-overwrite-diff-basis.md: 7a09934bd1798059de43a092f338d37aa9ccbd9a
2026-07-30-bounded-overwrite-diff-basis.zh.md: 1d6bdd1068d119aae859132d9f8216ca29d0dc11

View File

@@ -0,0 +1,31 @@
# Agent Note: Bound overwrite contextual-diff bases at the provider
Status: implemented
English | [中文](2026-07-30-bounded-overwrite-diff-basis.zh.md)
## Problem
`dsh-fs-local` returned the complete prior file in `FsWriteOutcome.before` so consumers could build a contextual overwrite diff. That presentation-only pre-read was unbounded: a large overwrite could allocate the entire prior file, and checking an earlier path stat alone could not enforce a limit because an external process could replace or grow the file between the stat and the read. A large replacement also made the contextual hunk approach the replacement size even when the prior file was small. This closes the deferred bound recorded by [result-time applied-hunk diffs](../../archived/architecture/2026-07-02-result-time-applied-hunk-diffs.md).
## Decision
`LocalFileSystem.Config.diffBasisMaxBytes` is a positive safe-integer deployment setting no greater than the runtime's Buffer-allocation and string-decoding limits, with a 10 MiB default. An overwrite supplies `before` only when the UTF-8 replacement is strictly below that limit and the prior file opened for the basis also ends below it. The prior read opens a descriptor, checks that descriptor, and reads at most the configured byte count in cancellation-aware chunks; reaching the boundary returns `null`. A size change after descriptor stat also returns `null`, even if the final size remains below the limit, because a partial prefix would be an incorrect diff basis. Binary or invalid UTF-8 prior content likewise returns `null`, as does any descriptor-phase errno — a prior file deleted or made unreadable between the caller's preflight and the basis open cannot fail a write the caller already committed to; only cancellation and non-errno faults propagate. These outcomes do not block the atomic write.
The local provider owns this decision because `before` is its optional, best-effort basis: it can avoid acquiring prior content that the configured pair limit has already made ineligible. `tool-fs` continues to own diff computation, retention, and presentation. The setting is independent of `tool-fs.readStreamMinSize`; read routing and overwrite presentation are different policies and need not share a value.
`before: null` asks consumers to use their existing whole-file fallback. The limit bounds only the extra prior-content acquisition and eligibility for a contextual pair. It does not bound the caller-owned replacement, the returned `after` value, or a consumer's fallback rendering.
## Alternatives considered
**Keep a hardcoded threshold equal to the read tool's streaming threshold.** Rejected because the read threshold is deployment-configurable and consumer-owned. Two same-valued constants would create an unenforced cross-package coupling, while the overwrite basis is itself a deployment memory/presentation choice.
**Gate only the prior side in the provider and cap new-content diffing in `tool-fs`.** Rejected because it would acquire prior text even when the provider's configured pair limit already excludes the replacement, and it would split one `before` eligibility rule across two plugins. Consumers remain free to impose additional output limits.
**Trust the initial `probe()` size before using an ordinary whole-file read.** Rejected because that size can become stale before the read. The descriptor reader must enforce the bound on the object it actually reads.
**Stream a contextual diff for arbitrarily large pairs.** Rejected for this bug fix because the current filesystem seam returns complete `before`/`after` strings and the current diff implementation consumes them. A streaming diff would require a separate cross-package protocol and presentation design.
## Consequences
Deployments can tune the extra overwrite-basis cost without changing read routing. At or above the exclusive limit, overwrites still succeed and remain visible through the whole-file fallback, but lose contextual hunks. Below the limit, the provider can still hold almost `diffBasisMaxBytes` of prior text in addition to the caller's replacement. The bounded descriptor read adds an open/stat/read sequence for eligible overwrites, while preventing a stale path probe from turning that sequence into an unbounded allocation.

View File

@@ -0,0 +1,31 @@
# Agent Note: 在提供方限制覆写上下文 diff 基础
Status: implemented
[English](2026-07-30-bounded-overwrite-diff-basis.md) | 中文
## Problem
`dsh-fs-local` 会在 `FsWriteOutcome.before` 中返回完整旧文件,供消费方生成覆写上下文 diff。这个仅用于展示的预读没有上限:大文件覆写可能分配整个旧文件;而仅检查较早的路径 stat 也无法真正实施上限,因为外部进程可以在 stat 与读取之间替换文件或扩大文件。即使旧文件很小,大替换内容也会使上下文 hunk 接近替换内容本身的大小。本改动关闭了 [result-time applied-hunk diff](../../archived/architecture/2026-07-02-result-time-applied-hunk-diffs.md) 中记录的暂缓上限事项。
## Decision
`LocalFileSystem.Config.diffBasisMaxBytes` 是一个不超过运行时 Buffer 分配和字符串解码上限的正安全整数部署配置,默认 10 MiB。只有当 UTF-8 替换内容严格低于该上限,且为生成基础而打开的旧文件最终也低于该上限时,覆写才提供 `before`。旧文件读取会打开文件描述符、检查该描述符,并按可响应取消的分块最多读取配置的字节数;一旦到达边界便返回 `null`。描述符 stat 后发生大小变化时同样返回 `null`,即使最终大小仍低于上限,因为部分前缀会成为错误的 diff 基础。旧内容为二进制或无效 UTF-8 时也返回 `null`;描述符阶段的任何 errno 同样如此——旧文件在调用方预检之后、基础读取打开之前被删除或变得不可读,不能让调用方已经提交的写入失败;只有取消和非 errno 故障会继续向上传播。这些结果都不会阻止原子写入。
本地提供方拥有该决策,因为 `before` 是它提供的可选、尽力而为的基础:当配置的成对上限已使替换内容不合格时,它可以避免获取旧内容。`tool-fs` 继续拥有 diff 计算、保留与展示。该配置独立于 `tool-fs.readStreamMinSize`;读取路由与覆写展示是不同策略,无需共享数值。
`before: null` 要求消费方使用既有的整文件回退。该上限只限制额外获取旧内容的成本,以及上下文内容对是否合格;它不限制调用方持有的替换内容、返回的 `after` 值或消费方的回退渲染。
## Alternatives considered
**保留一个与读取工具流式阈值相等的硬编码阈值。** 否决,因为读取阈值可由部署配置,且归消费方所有。两个同值常量会形成无法强制的一致性耦合,而覆写基础本身也是部署层面的内存与展示选择。
**提供方只限制旧内容一侧,并在 `tool-fs` 中限制新内容 diff。** 否决,因为当提供方配置的成对上限已经排除替换内容时,这仍会获取旧文本;同时会把同一条 `before` 合格规则拆到两个插件中。消费方仍可自由施加额外的输出限制。
**信任初次 `probe()` 的大小,再执行普通整文件读取。** 否决,因为该大小可能在读取前变旧;描述符读取必须对它真正读取的对象实施上限。
**为任意大的内容对流式生成上下文 diff。** 本次缺陷修复不采用,因为当前文件系统 seam 返回完整的 `before`/`after` 字符串,当前 diff 实现也消费这两个字符串。流式 diff 需要独立的跨包协议与展示设计。
## Consequences
部署可以调整额外的覆写基础成本,而不改变读取路由。达到或超过排他上限时,覆写仍会成功,并通过整文件回退保持可见,但不再提供上下文 hunk。低于上限时,除调用方的替换内容外,提供方仍可能持有接近 `diffBasisMaxBytes` 的旧文本。对于合格覆写,有上限的描述符读取会增加一次 open/stat/read 序列,同时防止陈旧路径探测把该序列变成无上限分配。

View File

@@ -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/2026-08-06-host-backed-web-preferences.md
2026-08-06-host-backed-web-preferences.md: d56a8d2e330b214a1922997e3cc7165fd0fb31e4
2026-08-06-host-backed-web-preferences.zh.md: 593646fe0845c20fb09cb7d115e6fa558226506e

View File

@@ -0,0 +1,43 @@
# Agent Note: Persist Web user preferences through Host settings
Status: implemented
English | [中文](2026-08-06-host-backed-web-preferences.zh.md)
## Problem
The Web Appearance, Language, and busy-Enter preferences lived in browser `localStorage`. Browser storage is scoped to an origin, so reopening `dsh web` on another port selected a different partition and lost choices even though both processes used the same DSH home. These are user-level product preferences; session selection, drafts, disclosure state, and other transient browser state remain page-local.
The first theme implementation moved only Appearance to Host settings but awaited its initial RPC before providing `ThemeService`. A slow or unavailable settings request therefore suspended the assembled page. It also subscribed after the read, could miss an invalidation in that window, did not carry namespace revisions on writes, and allowed queued writes from a disposed plugin to reach the Host.
## Decision
The owning Host halves register three schemas: optional `locale.preference` (`zh` or `en`, where absence delegates to the browser), `ui-theme.preference` (`light`, `dark`, or `system`, default `system`), and `ui-conversation.busyEnter` (`queue` or `steer`, default `queue`). The local settings provider stores explicit choices in `$DSH_HOME/settings.yaml`, which resolves to `~/.dsh/settings.yaml` under the default home. The API proxy explicitly exposes all three namespaces beside the other Web settings; registration alone never crosses that configuration boundary.
The client runtime provides one `bindSettingsScope` lifecycle per namespace — the browser mirror of the Host-side settings owner seam. It installs `settings/changed` and `connection/reset` listeners before starting a background initial read, so no settings transport can block plugin activation and an invalidation cannot fall into a read-before-subscribe gap, and it publishes a snapshot store (status, section value, revision, writability, host/memory mode) the domain service subscribes to. The default decoder validates each incoming section against the namespace's own serialized wire schema, rehydrated through dsh-client-schema-form, so domains carry no hand-written wire guards. Domain services take the scope as an ordinary constructor collaborator, publish their provisional defaults immediately—browser-derived locale, system theme, and Queue—then adopt an accepted Host section without writing it back; a service constructed without a scope (standalone dictionary or policy fixtures) simply stays process-local.
User changes update the live service synchronously and queue a `settings.mutate` path operation through `scope.set`. The scope serializes gestures, sends the latest known namespace revision as `expectedRevision`, records every successful revision, and lets only the latest write settlement republish live state. A rejected or failed latest write reloads Host state. Disposal rejects new work, skips queued operations, suppresses publication by the in-flight operation, and waits for that operation to settle before the plugin reaches quiescence.
Remote browsers cannot call the loopback-only configuration API, so their preferences remain process-local. Dynamic third-party theme ids remain in-process extensions outside the built-in Host schema; removing one resets the live registry without replacing the last durable built-in preference.
## Alternatives considered
**Keep `localStorage` and copy values between ports.** One origin cannot enumerate another origin's storage, and a Host relay would recreate the settings service around a browser-specific format.
**Mirror Host settings into `localStorage`.** A second authority requires boot and invalidation conflict rules while retaining the partition that caused the defect. The Host document is the sole durable source.
**Await the initial read to avoid a provisional render.** Configuration availability is not a prerequisite for drawing the page. A background read may cause one live convergence, but it keeps failure isolated and preserves the existing browser/system/default fallbacks.
**Give every domain its own settings controller.** The concurrency, revision, failure, invalidation, and disposal rules are identical; copying them already produced lifecycle drift in the theme implementation. Domain-owned schemas keep product policy out of the shared runtime.
**A per-field preference controller with paired sync/persist callbacks.** The first shared lifecycle synchronized one scalar field through a domain `sync` callback while the service wrote back through an injected `persist` callback. The mutual callbacks forced two-phase construction — a defaulted no-op writer later replaced via `bindPersistence` — every additional field of a namespace would have carried its own controller and whole-document read, and each domain re-declared a hand-written guard the registered wire schema already expresses. The namespace scope publishes a snapshot the service subscribes to and accepts writes directly, so the callback pair and the second construction phase do not exist.
**Move every `localStorage` entry into settings.** Current session, drafts, panel disclosure, trajectory display state, and similar entries are browser-instance state rather than user configuration. Promoting them would synchronize transient navigation state across tabs and ports without a product contract.
## Consequences
Appearance, Language, and busy-Enter choices follow the DSH user home across reloads, ports, and loopback origins. Direct edits to `settings.yaml` converge through the existing invalidation stream, while legacy `dsh.theme`, `dsh.locale`, and `dsh.conversation.busyEnter` entries are neither read nor written.
Boot may briefly show the domain default before the background read settles. A transient read failure keeps that default or the last good in-process value; reconnect retries. A write rejection can visibly restore the durable preference after the immediate local change.
Focused unit coverage pins schema registration, listener-before-read ordering, nonblocking activation, schema-validated section acceptance, revisioned ordered writes, stale-response containment, failure recovery, disposal quiescence, and remote memory mode. The namespace-granular scope also carries multi-field sections, so later configuration surfaces can ride the same lifecycle instead of hand-rolling describe/mutate synchronization. The keyless Web settings scenario writes all three preferences through the UI, verifies the YAML document and empty legacy storage, reloads, and boots another Host on a distinct port against the same DSH home.

View File

@@ -0,0 +1,43 @@
# Agent Note: 通过 Host settings 持久化 Web 用户偏好
Status: implemented
[English](2026-08-06-host-backed-web-preferences.md) | 中文
## 问题
Web 的 Appearance、Language 和繁忙态 Enter 偏好原本存在浏览器 `localStorage` 中。浏览器存储以 origin 为作用域,因此换一个端口重新打开 `dsh web` 会选中另一个存储分区并丢失选择,即使两个进程使用同一个 DSH home。这些是用户级产品偏好;会话选择、草稿、折叠展开状态和其他瞬态浏览器状态仍保留在页面内。
第一版主题实现只把 Appearance 移入 Host settings,但会在提供 `ThemeService` 之前等待初始 RPC。缓慢或不可用的 settings 请求因而会挂起组装后的页面。该实现还在读取后才建立订阅,可能错过此窗口内的失效通知;它写入时不携带 namespace revision,并且允许已释放插件所排队的写入到达 Host。
## 决策
各领域所属的 Host half 注册三份 schema:可选的 `locale.preference`(`zh` 或 `en`,缺失时交由浏览器决定)、`ui-theme.preference`(`light`、`dark` 或 `system`,默认为 `system`),以及 `ui-conversation.busyEnter`(`queue` 或 `steer`,默认为 `queue`)。本地 settings 提供方将显式选择存入 `$DSH_HOME/settings.yaml`,在使用默认 home 时,该路径解析为 `~/.dsh/settings.yaml`。API 代理会显式暴露这三个 namespace,与其他 Web settings 并列;仅注册它们,绝不会跨越该配置边界。
客户端运行时为每个 namespace 提供一份 `bindSettingsScope` 生命周期——即 Host 侧 settings owner seam 的浏览器镜像。它在开始后台初始读取之前安装 `settings/changed` 和 `connection/reset` 监听器,因此任何 settings 传输都不会阻塞插件激活,失效通知也不会掉入先读取、后订阅的空档;它还会发布一个供领域服务订阅的快照 store(状态、分节值、revision、可写性、host/内存模式)。默认解码器会对照该 namespace 自身的序列化 wire schema(经 dsh-client-schema-form 还原)校验每个传入分节,因此各领域无需携带手写的 wire 校验器。领域服务把 scope 当作普通的构造函数协作者接收,立即发布各自的暂定默认值:由浏览器派生的 locale、系统主题和 Queue;随后采纳已获接受的 Host 分节,但不将其写回;不带 scope 构造的服务——独立词典或政策 fixture(测试前置数据)——则仅停留在进程本地。
用户变更会同步更新实时服务,并经 `scope.set` 将一项 `settings.mutate` 路径操作排入队列。scope 会串行处理手势,以最新已知 namespace revision 作为 `expectedRevision` 发送,记录每次成功写入的 revision,并且只允许最新写入的结算结果重新发布实时状态。最新写入被拒或失败时,scope 会重新加载 Host 状态。插件释放会拒绝新工作、跳过已排队操作、抑制运行中操作发布状态,并等待该操作结算后才让插件达到完全停稳。
远程浏览器无法调用仅限回环请求的配置 API,因此其偏好仅保留在进程内。动态第三方主题 id 仍是内置 Host schema 之外的进程内扩展;移除其中一个会重置实时注册表,但不会替换上一个持久化的内置偏好。
## 曾考虑的替代方案
**保留 `localStorage`,并在不同端口间复制值。** 一个 origin 无法枚举另一个 origin 的存储,而 Host 中继会围绕浏览器特有格式重新实现一套 settings 服务。
**将 Host settings 镜像到 `localStorage`。** 第二个权威来源会要求另外定义启动与失效时的冲突规则,同时依然保留造成该缺陷的分区。Host settings 文档是唯一的持久化真源。
**等待初始读取,以避免暂定渲染。** 绘制页面不以配置可用为前置条件。后台读取可能引发一次实时收敛,但它会隔离失败,并保留既有的浏览器/系统/默认回落路径。
**让每个领域拥有自己的 settings 控制器。** 并发、revision、失败、失效与释放规则完全一致;此前的主题实现已因复制这些规则产生生命周期漂移。由领域持有 schema,可以避免把产品政策放入共享运行时。
**带成对 sync/persist 回调的逐字段偏好控制器。** 第一版共享生命周期经领域提供的 `sync` 回调同步单个标量字段,服务则经注入的 `persist` 回调写回。这对相互依赖的回调迫使构造分两阶段完成——写入器先默认为无操作,稍后经 `bindPersistence` 替换——namespace 每新增一个字段,本都得再携带一个自己的控制器和一次全文档读取,且每个领域都重新声明了一个已注册 wire schema 本已表达的手写校验器。namespace scope 发布一份供服务订阅的快照并直接接受写入,因此这对回调与第二个构造阶段都不存在。
**把每个 `localStorage` 条目都移入 settings。** 当前会话、草稿、面板展开状态、trajectory 显示状态和类似条目属于浏览器实例状态,而非用户配置。将它们提升为设置,会在没有产品契约的情况下,跨标签页和端口同步短暂导航状态。
## 后果
Appearance、Language 和繁忙态 Enter 选择会跟随 DSH 用户 home,跨越重新加载、端口与回环 origin。直接编辑 `settings.yaml` 所产生的变更会通过现有失效流收敛,而旧的 `dsh.theme`、`dsh.locale` 和 `dsh.conversation.busyEnter` 条目既不会被读取,也不会被写入。
启动时可能会在后台读取结算前短暂显示领域默认值。短暂的读取失败会保留该默认值或上一个正确的进程内值;重连时会重试。写入被拒时,界面可能会在本地值立即变化后明显恢复为持久化偏好。
聚焦的单元测试覆盖 schema 注册、先监听后读取的顺序、非阻塞激活、经 schema 校验的分节接受、携带 revision 的有序写入、陈旧响应隔离、故障恢复、释放时完全停稳,以及远程端仅内存模式。以 namespace 为粒度的 scope 也承载多字段分节,因此后续的配置表面可以沿用同一份生命周期,而不必手搭 describe/mutate 同步。无密钥 Web settings 场景通过 UI 写入全部三项偏好,校验 YAML 文档并确认旧 `localStorage` 为空,重新加载,再使用同一个 DSH home 在不同端口上启动另一个 Host。

View File

@@ -1,48 +0,0 @@
# Agent Note: npm-backed preparation makes GitHub repository Plugins self-contained
Status: implemented
English | [中文](2026-08-08-npm-backed-git-repository-plugin-preparation.zh.md)
## Problem
The repository Plugin authoring contract requires `scripts.prepack` to invoke `dsh-plugin-prepare`. Supplying that executable from the running DSH installation made a source package appear valid even when its own manifest could not obtain the helper. It therefore did not prove the behavior users need after `@deepseek-ai/dsh-repository-plugin` is published: an ordinary Git-hosted npm package must be installable and preparable from only its declared dependencies.
A selectable `.dsh-plugin` inside a pnpm workspace has a second isolation requirement. pnpm prepares a Git-hosted package by running the repository's preferred package manager before packing the selected subdirectory. A nested `pnpm install` can join the containing workspace; when the root lockfile does not list `.dsh-plugin` as an importer, pnpm can report success without installing dependencies declared only by that package. Its TypeScript build or prepare command then fails, or a pre-generated artifact hides the missing dependency.
The checked-in headless fixture mounts an already prepared wrapper. It proves runtime composition, not GitHub acquisition, npm resolution, or package-owned preparation.
## Decision
The `.dsh-plugin` package declares `@deepseek-ai/dsh-repository-plugin` as an ordinary development dependency and invokes its published `dsh-plugin-prepare` executable from `scripts.prepack`. The package may declare any other build and runtime dependencies and run arbitrary compilation before the helper. The repository Plugin package marks its Cordis and DSH peers optional so a helper-only development install resolves only the helper's actual `zod` runtime dependency; an application composition still supplies the peers used by the package's Cordis entry.
DSH does not materialize or prepend a prepare executable. `RepositoryCache` supplies only a transaction-owned `pnpm` wrapper: the outer install runs the pinned pnpm entry directly, while pnpm's hard-coded Git-package `pnpm install` reinvokes the same entry with `--ignore-workspace`. The selected package therefore owns dependency resolution even beneath another pnpm lockfile, and normal package-manager lifecycle `PATH` construction exposes `node_modules/.bin/dsh-plugin-prepare`. The temporary pnpm wrapper disappears after the child settles. The repository remains trusted package-manager input: all dependency and lifecycle code executes under the existing trust contract.
The Node 24 consumer lane passes an exact source derived from the pull request head repository and SHA. It uses the existing private DeepSeek Harness repository rather than creating another repository per run. A job-scoped Git configuration gives the read-only job token access to that exact private source and rewrites pnpm's SSH fallback to authenticated HTTPS.
The built-entry acceptance also creates an in-process npm registry. It stages the current built `@deepseek-ai/dsh-repository-plugin` as a publication artifact by removing `private`, replacing workspace protocols with the release version, and packing the declared files. The registry serves the resulting packument and tarball, while a job-local npm config directs only the `@deepseek-ai` scope to it. The real built `dsh run` child then fetches the exact Git source; that package resolves the helper through npm, type-checks and bundles a TypeScript Cordis entry and MCP server, prepares the adjacent skill, and loads all three contributions. A deliberately failing host `PATH` command proves the lifecycle selected the dependency-local executable. The acceptance also requires registry resolution and inspects the immutable prepared cache, so restoring a host-injected helper cannot satisfy it.
## Alternatives considered
**Inject `dsh-plugin-prepare` from the running DSH installation.** Rejected because it lets an incomplete repository manifest pass and tests a host-only path that npm consumers cannot reproduce.
**Publish the source fixture itself to npm.** Rejected because the product contract is specifically that the DSH Plugin remains Git-hosted; only the reusable preparation helper is an npm dependency.
**Create a new private GitHub repository in every CI run.** Rejected because the pull request repository at its exact head SHA is already a real authenticated private Git remote. Per-run repository mutation would add credentials, cleanup, and eventual-consistency failure modes without changing the acquisition path.
**Prepare after `RepositoryCache` installs the selected package.** Rejected because pnpm's packed subdirectory no longer contains sibling source assets referenced by paths such as `../skills`; preparation must happen before packlist.
**Clone GitHub repositories in DSH and bypass pnpm's Git fetcher.** Rejected because it would duplicate ref resolution, subdirectory selection, dependency installation, packlist behavior, and cache integrity already owned by the pinned package manager.
## Consequences
- A repository author can commit a `.dsh-plugin` package, TypeScript source, skills, and MCP definitions to GitHub without publishing that Plugin package to npm. The package must declare the published preparation dependency.
- Private GitHub sources use the host's standard Git authentication. CI proves that path with a temporary read-only configuration rather than persistent runner credentials.
- `prepack`, not `prepare`, is part of the authoring format. It may contain arbitrary package-owned build steps but must invoke the dependency-provided helper; missing dependency or lifecycle metadata fails before a cache generation is usable.
- A selected package in a pnpm repository installs from its own manifest rather than an enclosing workspace. It cannot rely on workspace-only hoisting; ordinary registry and relative `file:` dependencies remain package-owned inputs.
- Exact source strings identify immutable cache generations; a changed ref or source configuration selects another generation.
- Package dependencies, compilation, preparation, and the trusted `dsh.entry` contribution remain owned by the repository package and the [trusted-code decision](../architecture/2026-08-08-trusted-repository-package-code.md).
## Testing
`packages/boot/app-boot/tests/repository-cache.spec.ts` runs a package excluded from its source repository's root pnpm lockfile through a local Git subpath and requires relative `file:` dependencies to provide both its build command and `dsh-plugin-prepare`; it also proves that visible environment survives while credential-shaped variables are scrubbed. `packages/self-modification/repository-plugin/tests/repository-plugin.spec.ts` pins helper-bearing `prepack` metadata and preparation output. `examples/headless-agent/tests/keyless-smoke.e2e.ts` keeps the checked-in prepared fixture on that source contract. `apps/cli/tests/github-repository-plugin.built.e2e.ts` is the product acceptance: simulated published helper package, job-local npm registry, fresh DSH home, exact authenticated private GitHub source, actual built `dsh run`, package-owned TypeScript build, real MCP execution, code-entry transformation, mock LLM request observation, and prepared cache inspection.

View File

@@ -1,48 +0,0 @@
# Agent Note: 基于 NPM 的准备机制使 GitHub repository 插件自包含
状态:已实现
[English](2026-08-08-npm-backed-git-repository-plugin-preparation.md) | 中文
## 问题
repository 插件创作约定要求 `scripts.prepack` 调用 `dsh-plugin-prepare`。如果由正在运行的 DSH 安装提供该可执行文件,即使源包自身的 manifest(元数据清单)无法取得辅助程序,它也会显得有效。因此,这并未证明 `@deepseek-ai/dsh-repository-plugin` 发布后用户所需的行为:普通 Git 托管 NPM 包必须只依靠自身声明的依赖即可安装和准备。
pnpm workspace 内可选择的 `.dsh-plugin` 还有另一项隔离要求。pnpm 会在打包所选子目录前运行仓库首选的包管理器,以准备 Git 托管包。嵌套的 `pnpm install` 可能加入外层 workspace;当根 lockfile 未把 `.dsh-plugin` 列为 importer 时,pnpm 可能报告成功,却未安装仅由该包声明的依赖。随后,其 TypeScript 构建或准备命令会失败;也可能因为存在预生成产物,依赖缺失被掩盖。
签入仓库的 headless fixture(测试前置数据)挂载的是已准备好的包装层。它证明运行时组合,而不证明 GitHub 获取、NPM 解析或包自有准备。
## 决策
`.dsh-plugin` 包将已发布的 `@deepseek-ai/dsh-repository-plugin` 声明为普通开发依赖,并在 `scripts.prepack` 中调用其已发布的 `dsh-plugin-prepare` 可执行文件。该包可以声明其他任意构建依赖与运行时依赖,并在辅助程序前执行任意编译。repository 插件包把 Cordis 与 DSH 对等依赖(peer dependency)标为可选,因此仅为使用辅助程序而进行的开发安装只会解析辅助程序实际依赖的 `zod` 运行时依赖;应用组合仍会提供该包 Cordis 入口所使用的对等依赖。
DSH 不会生成准备阶段可执行文件,也不会将其前置到 `PATH`。`RepositoryCache` 只提供一个由事务持有的 `pnpm` 包装脚本:外层安装直接运行锁定的 pnpm 入口,而 pnpm 为 Git 包硬编码的 `pnpm install` 会以 `--ignore-workspace` 重新调用同一入口。因此,即使位于另一个 pnpm lockfile 之下,所选包仍自行负责依赖解析,正常的包管理器生命周期 `PATH` 构造会暴露 `node_modules/.bin/dsh-plugin-prepare`。临时 pnpm 包装脚本会在子进程结算后消失。repository 仍是受信任的包管理器输入:所有依赖与生命周期代码都按既有信任约定执行。
Node 24 消费方 CI 任务会传入从 PR(Pull Request)head 仓库与 SHA 派生的精确源。它复用现有私有 DeepSeek Harness 仓库,而不会为每次运行新建仓库。作业作用域的 Git 配置允许只读作业 token 访问该精确私有源,并把 pnpm 的 SSH 回退改写为已认证 HTTPS。
构建入口验收还会创建一个进程内 NPM 注册表。它通过移除 `private`、将 workspace protocol 替换为发布版本并打包声明的文件,把当前已构建的 `@deepseek-ai/dsh-repository-plugin` 暂存为发布产物。注册表会提供由此生成的 `packument` 与 tarball,作业本地 NPM 配置则只把 `@deepseek-ai` scope 指向它。实际构建的 `dsh run` 子进程随后获取精确 Git 源;该包通过 NPM 解析辅助程序,对 TypeScript Cordis 入口和 MCP server 进行类型检查与打包,准备相邻的 skill(技能),并加载全部三类贡献。一个刻意设为失败的宿主 `PATH` 命令可以证明,该生命周期选中的是依赖内的可执行文件。验收还要求经过注册表解析并检查不可变的已准备缓存,因此恢复宿主注入的辅助程序也无法通过。
## 考虑过的替代方案
**从正在运行的 DSH 安装注入 `dsh-plugin-prepare`。** 拒绝,因为这会让 manifest 不完整的 repository 包通过,并测试 NPM 消费方无法复现的纯宿主路径。
**把源 fixture 本身发布到 NPM。** 拒绝,因为产品约定明确要求 DSH 插件仍托管在 Git;只有可复用的准备辅助程序是 NPM 依赖。
**在每次 CI 运行中创建新的私有 GitHub 仓库。** 拒绝,因为 PR 仓库的精确 head SHA 已是经过认证的真实私有 Git remote。每次运行的仓库变更会增加凭据、清理和最终一致性失败模式,却不改变获取路径。
**在 `RepositoryCache` 安装所选包后再准备。** 拒绝,因为 pnpm 打包后的子目录不再包含 `../skills` 等路径所引用的同仓库相邻资源;准备必须在生成 packlist 前完成。
**在 DSH 中克隆 GitHub 仓库并绕过 pnpm 的 Git 获取器。** 拒绝,因为这会重复实现已由锁定包管理器负责的 ref 解析、子目录选择、依赖安装、packlist 行为和缓存完整性。
## 后果
- 仓库作者可以把 `.dsh-plugin` 包、TypeScript 源码、skill 与 MCP 定义提交到 GitHub,而无需把该插件包发布到 NPM。该包必须声明已发布的准备依赖。
- 私有 GitHub 源使用宿主的标准 Git 认证。CI 使用临时的只读配置而非运行器上的持久凭据来验证该路径。
- 创作格式使用 `prepack` 而不是 `prepare`。其中可以包含任意包自有构建步骤,但必须调用依赖提供的辅助程序;依赖或生命周期元数据缺失时,会在缓存 generation 可用前失败。
- pnpm 仓库中的所选包按自身 manifest 安装,而不继承外层 workspace。它不能依赖仅由 workspace 提升而可见的包;普通注册表依赖和相对 `file:` 依赖仍是包自有输入。
- 精确源字符串标识不可变缓存 generation;改变 ref 或源配置会选择另一个 generation。
- 包依赖、编译、准备和受信任的 `dsh.entry` 贡献仍由 repository 包和[受信任代码决策](../architecture/2026-08-08-trusted-repository-package-code.md)负责。
## 测试
`packages/boot/app-boot/tests/repository-cache.spec.ts` 会通过本地 Git 子路径运行一个未列入源仓库根 pnpm lockfile 的包,并要求相对 `file:` 依赖同时提供构建命令与 `dsh-plugin-prepare`;该测试还证明可见环境变量得以保留,而名称符合凭据模式的变量会被清除。`packages/self-modification/repository-plugin/tests/repository-plugin.spec.ts` 锁定包含辅助命令的 `prepack` 元数据与准备输出。`examples/headless-agent/tests/keyless-smoke.e2e.ts` 使签入仓库的已准备 fixture 继续符合该源格式约定。`apps/cli/tests/github-repository-plugin.built.e2e.ts` 是产品验收测试:模拟发布的辅助程序包、作业本地 NPM 注册表、全新 DSH 主目录、精确且经过认证的私有 GitHub 源、实际构建的 `dsh run`、包自有 TypeScript 构建、真实 MCP 执行、代码入口转换、mock LLM(大语言模型)请求观测,以及已准备缓存检查。

View File

@@ -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/2026-08-10-child-agents-join-their-parent-preset.md
2026-08-10-child-agents-join-their-parent-preset.md: 4534004ad54df69822872b9595a29443fc3a990b
2026-08-10-child-agents-join-their-parent-preset.zh.md: bdf9928bea4b75e2915c8adf5c15f8a01c6583e4

View File

@@ -0,0 +1,59 @@
# Agent Note: Child agents join their parent's preset composition
Status: implemented
English | [中文](2026-08-10-child-agents-join-their-parent-preset.zh.md)
## Problem
Tool and prompt-section visibility is inherited along `dsh-scope`'s parent chain, and an agent's scope key is minted with no parent. [Per-session agent presets](../architecture/2026-08-03-per-session-agent-presets.md) moved every model-facing row onto the agent plane and made `AgentPresets.mount()` the one thing that binds that parent link — from the api-proxy's session create, resume, and fork paths. The two in-process subagent drivers compose their children through `applyChildComposition()`, which installed only the per-child persona and tool filter, so a child's scope chain had length one and its registry view resolved the global layer alone.
That layer is now empty in any deployment with a preset roster: the web-app patch layer disables every host-plane tool row. A one-shot child therefore reached the model with zero tools, a continuable child with only the host-plane `report`, and neither carried its parent's persona, workspace context, plan-mode section, or skill catalog. The fork path had already been given the same treatment for the same reason; delegation had not.
The child's durable header compounded it. `childSessionMeta()` recorded no preset, so a cold read of a child session resolved the deployment default — a tool set the child never ran under, which is exactly what the model-visible ⟺ logged rule exists to prevent.
## Decision
`AgentPresets.composeFrom(agentCtx, parentCtx)` joins one agent to the standing composition another already runs on, and returns the preset id joined. It locates the parent's mount through `standingMountFor()` — the agent's key is parented to its preset's standing key, the same relation `serviceForAgent()` reads — and binds the child's key to that same standing key, keeping the binding under the roster's sole re-link authority. A parent that joined no preset yields no join and no error, which is the rosterless deployment: its model-facing rows sit in the host composition, where the child already resolves them through the global layer.
This is a bind, not a mount, and both differences are load-bearing. The child gets its parent's exact generation, so a composition file edited since the parent started cannot hand the child a different one than its parent's history was produced under, and a preset deleted since cannot fail a child whose parent keeps running. It is also synchronous, which is what lets the child creation windows use it — both in-process drivers compose inside a synchronous `setup`.
`applyChildComposition(childCtx, parent, composition)` takes the parent and performs the join before applying the child's own registrations. The parameter is the point: it makes composing a child without the join unrepresentable at the call sites, rather than leaving each new driver to remember a second step. `childSessionMeta()` records the joined id through `AgentPresets.composedPreset()`, read from the parent's live scope chain rather than its header, because a parent that switched preset while blank runs on the newer composition while its header still names the older one.
`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: `ToolRegistry` 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.
## Alternatives considered
**Re-mount the parent's preset by id in the child's setup.** Rejected on both semantics and mechanics. It re-reads the roster and re-stats the composition file, so an edit since the parent started forks the child onto a different generation, and a preset deleted since fails the child while its parent runs on. `mount()` is also asynchronous, which the synchronous creation windows cannot accept without restructuring both drivers.
**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.
**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.
**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.
## Testing
`packages/preset/agent-presets/tests/mount.spec.ts` covers the join against real fixture compositions: the child sees its parent's tools and prompt sections, no second generation is mounted, the join survives the parent's disposal (a background child outliving its parent), the reported id matches, a parent without a preset joins nothing, and an unscoped context is refused.
`packages/core/tools/tests/scoped.spec.ts` covers the restriction rule directly: a child's filter removes a tool it inherited from an ancestor scope, the child's own registrations survive its own filter, and an ancestor's restriction still reaches every scope nested inside it.
`packages/subagent/subagent-inprocess/tests/preset-inheritance.spec.ts` asserts the model-visible result through `startInProcessRun()` on a host composition carrying no model-facing rows: the schemas in the child's own request, its parent's prompt section, the recorded header preset, a `toolFilter` applied over the inherited preset tools, and a parent that switched preset while blank — to a DIFFERENT preset, so the assertion distinguishes reading the parent's live scope chain from reading its creation header.
The assembled-transcript layer is the shipped Web composition's e2e rather than a keyless snapshot. Every runnable example this repo ships composes no preset roster, so the defect is not observable in the snapshot harness at all: a snapshot scenario would first need an example that mounts a roster AND delegates. The Web e2e boots the real `base` + `web-app` patch layers with both shipped presets, which is the assembled evidence the testing policy asks for; the Web browser lane's subagent goldens carry the visible consequence, since a child that records its preset now shows the preset badge its parent shows.
## Consequences
Delegation now costs a scope-parent bind per child and nothing else — no extra plugin instances, no roster read, no failure mode. A child's capabilities are exactly its parent's, minus whatever its own `toolFilter` removes; a per-subagent preset ("agent types") remains unbuilt and would be a new request field rather than a change to this join.
`applyChildComposition()` changed shape, so any future out-of-tree in-process driver must supply the parent. That is the intended cost: the previous signature let a caller compose a capability-less child and get no error.
A cold-resumed continuable child joins its parent's CURRENT composition rather than the one its own header records. The window is narrow — the parent must create the child, stay blank, switch preset, and only then wake it, since a resident child never re-joins and a one-shot child never resumes — and the alternative is worse: resolving the child's own recorded id would re-read the roster and hand back the preset-deleted failure mode this join exists to avoid. The child's header still records what it started under, so the divergence is observable rather than silent.
`ToolRegistry` now reads a restriction's exempt set as "what this scope registers itself" rather than "the global layer", which changes one documented behavior beyond delegation: a tool an ANCESTOR scope contributes is now subject to a descendant's filter, where before only global-layer tools were. Nothing else on the chain loses its exemption — a scope's own registrations stay outside its own filter, which is the property the delegation runtime depends on.

View File

@@ -0,0 +1,59 @@
# Agent Note: Child agents join their parent's preset composition
Status: implemented
[English](2026-08-10-child-agents-join-their-parent-preset.md) | 中文
## Problem
工具与提示段的可见性沿 `dsh-scope` 的父链继承,而 agent 的 scope key 铸造出来时没有父。[逐会话 agent preset](../architecture/2026-08-03-per-session-agent-presets.md) 把所有面向模型的行搬到了 agent 平面,并让 `AgentPresets.mount()` 成为绑定那条父链的唯一途径——调用点在 api-proxy 的会话创建、恢复与 fork 路径上。两个进程内 subagent 驱动通过 `applyChildComposition()` 组装子 agent,而它只安装了逐子 agent 的 persona 与工具限制,于是子 agent 的 scope 链长度为一,其注册表视图只能解析到全局层。
在任何配置了 preset roster 的部署里,那一层现在是空的:web-app 补丁层禁用了全部宿主平面工具行。因此一次性子 agent 抵达模型时工具为零,可继续子 agent 只剩宿主平面的 `report`,两者都不带父方的 persona、工作区上下文、plan-mode 段与技能目录。fork 路径此前已因同一理由做过相同处理;委派没有。
子 agent 的持久化 header 让问题更进一步。`childSessionMeta()` 不记录任何 preset,于是冷读一个子会话解析到的是部署默认值——一套该子 agent 从未运行过的工具集,而这正是"模型可见 ⟺ 已记录"规则要杜绝的情形。
## Decision
`AgentPresets.composeFrom(agentCtx, parentCtx)` 让一个 agent 加入另一个 agent 已在运行的常驻组装,并返回所加入的 preset id。它通过 `standingMountFor()` 定位父方的挂载——agent 的 key 认父到其 preset 的常驻 key,正是 `serviceForAgent()` 读取的同一关系——再把子 agent 的 key 绑到同一个常驻 key 上,绑定句柄仍归 roster 独有的重链权威持有。未加入任何 preset 的父方不产生加入、也不报错,那就是无 roster 的部署:它面向模型的行位于宿主组装中,子 agent 已经能通过全局层解析到它们。
这是认父而非挂载,两处差别都要紧。子 agent 拿到的是父方那个确切的代际,因此父方启动后被编辑过的组装文件不可能把与父方历史所产出时不同的另一个代际交给它,此后被删除的 preset 也不可能让一个父方仍在运行的子 agent 失败。它还是同步的,这正是子 agent 创建窗口能够使用它的前提——两个进程内驱动都在同步的 `setup` 中完成组装。
`applyChildComposition(childCtx, parent, composition)` 接收父方,并在应用子 agent 自身注册之前完成加入。这个参数正是要点所在:它让"组装子 agent 却不做该加入"在各调用点无法表达,而不是把第二个步骤留给每个新驱动去记住。`childSessionMeta()` 通过 `AgentPresets.composedPreset()` 记录所加入的 id,该值从父方**活着的** scope 链读取而不是从其 header 读取,因为在空白期切换过 preset 的父方运行在更新的那份组装上,而它的 header 仍写着旧的那个。
`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:`ToolRegistry` 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。
## Alternatives considered
**在子 agent 的 setup 里按 id 重新挂载父方的 preset。** 语义与机制两方面都不成立而被否决。它会重读 roster 并重新 stat 组装文件,因此父方启动后的一次编辑就会把子 agent 分叉到另一个代际,而此后被删除的 preset 会让子 agent 失败、父方却照常运行。`mount()` 还是异步的,同步的创建窗口无法在不重构两个驱动的前提下接受它。
**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。
**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。
**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。
## Testing
`packages/preset/agent-presets/tests/mount.spec.ts` 用真实 fixture 组装覆盖该加入:子 agent 看到父方的工具与提示段、不会挂载出第二个代际、加入在父方 dispose 后依然成立(活得比父方久的后台子 agent)、上报的 id 一致、没有 preset 的父方不产生加入、以及无 scope 的上下文被拒绝。
`packages/core/tools/tests/scoped.spec.ts` 直接覆盖该限制规则:子 agent 的过滤器能移除它从祖先作用域继承来的工具、子 agent 自身的注册在自己的过滤器下存活、祖先的限制仍作用于其内嵌套的每个作用域。
`packages/subagent/subagent-inprocess/tests/preset-inheritance.spec.ts` 在一个不含任何面向模型行的宿主组装上,通过 `startInProcessRun()` 断言模型可见的结果:子 agent 自身请求中的 schema、父方的提示段、记录下来的 header preset、施加在继承来的 preset 工具之上的 `toolFilter`,以及在空白期切换过 preset 的父方——切换到**另一个** preset,这样断言才能区分"读父方活 scope 链"与"读父方创建 header"。
组装记录这一层用的是真实 shipped Web 组装的 e2e,而不是无密钥快照。本仓库所有可运行 example 都不组装 preset roster,因此该缺陷在快照 harness 里根本不可观察:要做快照场景,得先有一个既挂载 roster 又发起委派的 example。Web e2e 启动的是真实的 `base` + `web-app` 补丁层与两个 shipped preset,这正是测试政策要求的组装证据;Web 浏览器 lane 的 subagent golden 承载了可见后果——记录了 preset 的子 agent 现在会显示与其父方相同的 preset 徽标。
## Consequences
委派现在的成本是每个子 agent 一次 scope 认父,再无其他——没有额外的插件实例、没有 roster 读取、没有新的失败模式。子 agent 的能力恰好等于父方的能力,减去它自己的 `toolFilter` 所移除的部分;逐 subagent 的 preset("agent 类型")仍未构建,那会是一个新的请求字段,而不是对这次加入的改动。
`applyChildComposition()` 的形态变了,因此将来任何仓库外的进程内驱动都必须提供父方。这是刻意付出的代价:此前的签名允许调用方组装出一个毫无能力的子 agent 而不报任何错。
冷恢复的可继续子 agent 加入的是父方**当前**的组装,而不是它自己 header 所记录的那份。窗口很窄——父方必须先建子、保持空白、切换 preset,之后才唤醒它;驻留中的子 agent 不会重新加入,一次性子 agent 也不会恢复——而替代方案更糟:按子 agent 自己记录的 id 解析会重读 roster,把这次认父刻意规避掉的"preset 已删除"失败模式又请回来。子 agent 的 header 仍记录它启动时的那份,因此这处分歧是可观察的而非静默的。
`ToolRegistry` 现在把限制的豁免集合读作"该作用域自己注册的东西"而不是"全局层",这在委派之外改变了一处既有行为:**祖先**作用域贡献的工具现在会受后代过滤器约束,而此前只有全局层的工具会。链上其余部分的豁免不变——作用域自身的注册仍在自己的过滤器之外,这正是委派运行时所依赖的性质。