Merge master settings checkpoint into plugin management UI

.agents/notes/implemented/architecture/2026-06-11-dev-invariants-over-deep-readonly.i18n.yaml

@@ -2,5 +2,5 @@
 # side as of the last confirmed-consistent state. Both languages carry equal authority;
 # after editing either side, bring the other along and re-record with:
 #   pnpm run verify-translation-pairing --write .agents/notes/implemented/architecture/2026-06-11-dev-invariants-over-deep-readonly.md
-2026-06-11-dev-invariants-over-deep-readonly.md: 9adb741db1e26e3a9f750fed60524351159cfeaa
-2026-06-11-dev-invariants-over-deep-readonly.zh.md: d352eca985b6594ec5d10c9b73c2ecc18ea458b5
+2026-06-11-dev-invariants-over-deep-readonly.md: 0784960397f4d198e136e284ef9232b4ed3e2c2c
+2026-06-11-dev-invariants-over-deep-readonly.zh.md: 7183813130560bf82426c8719570b668ba86b7e5

.agents/notes/implemented/architecture/2026-06-11-dev-invariants-over-deep-readonly.md

@@ -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, 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.
+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` reads the current length without materializing an array. Synchronous historical readers are deprecated under the [event-read policy](2026-09-09-deprecate-synchronous-session-event-reads.md). 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.
 
@@ -53,7 +53,7 @@ Detaching `deriveMessages()` would protect the most common request path but leav
 ## Consequences
 
 - Every accepted live or seeded session event is detached from caller-owned inputs and deeply immutable before any observer can receive it.
-- `snapshotEvents()` exposes stable immutable snapshots instead of the private growing array; `seq` and `eventAt()` serve scalar reads without copying that array.
+- Existing `snapshotEvents()` and `eventAt()` callers retain immutable read results while their migration is deferred; `seq` reads the log length without copying the 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.

.agents/notes/implemented/architecture/2026-06-11-dev-invariants-over-deep-readonly.zh.md

@@ -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。
+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。`seq` 无需物化数组即可读取当前长度。同步历史读取方法按[事件读取策略](2026-09-09-deprecate-synchronous-session-event-reads.zh.md)弃用。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。
 
@@ -53,7 +53,7 @@ TypeScript readonly 类型不是充分的运行时边界。它们在程序运行
 ## 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。
-- `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。
+- 现有 `snapshotEvents()` 和 `eventAt()` 调用方在暂缓迁移期间仍获得不可变的读取结果；`seq` 无需复制数组即可读取日志长度。
 - 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。

.agents/notes/implemented/architecture/2026-07-05-prompt-variables-and-tool-guidance-ownership.i18n.yaml

@@ -2,5 +2,5 @@
 # side as of the last confirmed-consistent state. Both languages carry equal authority;
 # after editing either side, bring the other along and re-record with:
 #   pnpm run verify-translation-pairing --write .agents/notes/implemented/architecture/2026-07-05-prompt-variables-and-tool-guidance-ownership.md
-2026-07-05-prompt-variables-and-tool-guidance-ownership.md: 35bb7c6fabc85ae6f93bdbb67e13910eea627ca3
-2026-07-05-prompt-variables-and-tool-guidance-ownership.zh.md: 53d69cf45c02f6534334561b626d2c2ae6087c05
+2026-07-05-prompt-variables-and-tool-guidance-ownership.md: 8f5542cb09930e62fd1e26960166fbdd6d3f6745
+2026-07-05-prompt-variables-and-tool-guidance-ownership.zh.md: 348ae36cf6549d412adab1d3cdb5cc03f0badd01

.agents/notes/implemented/architecture/2026-07-05-prompt-variables-and-tool-guidance-ownership.md

@@ -26,7 +26,7 @@ The assembled system prompt had four defects, all of one family: facts the harne
 
 ### Prompt variables
 
-Plugins register `{{name}}` values through `ctx.systemPrompt.variable(name, provider)`. Assembly resolves them into the waterfall-visible variable map. Rendering rejects unknown own-property references, registered providers that return `undefined`, malformed complete references, and unbalanced references that still contain a closing `}}`; a lone unmatched `{{` remains prose, and substituted values are not rescanned. Registration rejects invalid or duplicate variable names, and section names are unique.
+Plugins register `{{name}}` values through `ctx.systemPrompt.variable(name, provider)`. Assembly resolves them into the waterfall-visible variable map. Rendering rejects unknown own-property references, registered providers that return `undefined`, malformed complete references, and unbalanced references that still contain a closing `}}`; a lone unmatched `{{` remains prose, and substituted values are not rescanned. Registration rejects invalid or duplicate variable names, and section names are unique. Sections may set `interpolate: false` to preserve generated documents literally; `tools:sdk` does so because tool descriptions and schemas may document their own `{{…}}` syntax.
 
 `dsh-agent-loop` registers the two built-ins, both pure projections of the context agent: `model` (= `options.model`) and `cwd` (= `session.header.cwd`). The example personas write `powered by the {{model}} model` — the model name is stated once, in the `model:` config key. `{{cwd}}` is demonstrated in the ACP example only: every ACP session carries the client's cwd, while config-pre-created stdio agents have none (a persona claiming `{{cwd}}` there fails the turn — by design). The variables stay on the loop plugin (unlike the sections below): they are runtime facts of the agents THIS loop drives, and a replacement loop supplies its own.
 
@@ -47,7 +47,7 @@ Per-tool semantics and selection guidance live in tool descriptions. Prompt sect
 - **The loop composes an identity line itself** — hardcodes model-facing prose in the one package that must stay thin ("plugins, not loop changes"), and outside the section pipeline it would be a second composition path. (The identity DOES ship as a code literal — but as an ordinary section registered by `dsh-system-prompt`, whose `system-prompt/assemble` waterfall remains the escape valve for a deployment that must drop it.)
 - **Inject the model name via the `agent/request` waterfall** — prompt text would be composed in two places and the earlier rendered persona could disagree with the final routed header. The request plugin that owns late routing must also own any earlier prompt claim about that model.
 - **Hand-write the model name in each persona** — duplicates the `model:` key one line above and silently lies after a config edit; the exact disease this decision cures.
-- **Lenient interpolation (leave unknown refs verbatim, or substitute empty)** — a typo ships `{{modle}}` (or a hole) to the model and nobody notices until transcript review.
+- **Lenient interpolation (leave unknown refs verbatim, or substitute empty)** — a typo ships `{{modle}}` (or a hole) to the model and nobody notices until transcript review. Leaving only unknown names unchanged would still substitute registered names inside tool documentation.
 - **Per-instance subagent wording in config** — returns model-facing prose to every deployment × instance, reviving the hand-written-guidance-in-leaf-YAML drift. **Keying wording off the provider NAME** — `providerName` is itself config, so a renamed provider silently gets the wrong words.
 - **Resolving the provider at `apply` time (a load-order requirement)** and **section-only subagent wording (lazily resolved at assemble)** — the alternatives to the provider-lifecycle events; both rejected in [the provider-lifecycle-events Agent Note](../../archived/architecture/2026-07-05-subagent-provider-lifecycle-events.md).
 
@@ -60,7 +60,7 @@ Per-tool semantics and selection guidance live in tool descriptions. Prompt sect
 
 - The tui-agent prompt renders identity, persona with the interpolated model, then fs/shell/web guidance through one assembly path.
 - Fork and fresh subagent descriptions reflect whether the provider inherits completed conversation turns; the tool appears, disappears, and is reworded with provider lifecycle changes.
-- Unknown, valueless, malformed, or unbalanced variable references name the section and throw; duplicate section, variable, and tool registrations also throw.
+- In interpolated sections, unknown, valueless, malformed, or unbalanced variable references name the section and throw; duplicate section, variable, and tool registrations also throw.
 - Snapshot replay is prompt-independent: it keys recorded chunk streams by turn and step without comparing the outgoing request.
 
 ## Consequences
@@ -69,4 +69,4 @@ Per-tool semantics and selection guidance live in tool descriptions. Prompt sect
 - `{{model}}` reflects `AgentOptions.model` at assembly time. A plugin that switches models in the `agent/request` waterfall makes the prompt's claim stale for that step, and one that SUPPLIES the model there (options.model unset — the loop's documented fallback) leaves the variable valueless at render, failing a `{{model}}` persona before the waterfall runs. Both have the same remedy, and it is the ownership rule itself: the plugin that owns the late-bound model fact states it early on the `system-prompt/assemble` waterfall (`assembly.variables['model'] = …`) — one owner, both statements; a loop test pins the supply path end-to-end. Accepted.
 - While a bound provider is absent (not yet activated, unloaded, mid-HMR-reload), the subagent tool does not exist and a model request in that window simply lacks it. That is the honest state — the alternative was a registered tool whose description or execution could not be trusted.
 - Strictness means a persona can fail a turn at render (e.g. `{{cwd}}` on a cwd-less session). The failure is contained — the turn ends `error`, the loop survives — and it is an authoring error we WANT loud.
-- No escape syntax for a literal `{{name}}` in prompt prose yet; add one if a real prompt ever needs it.
+- Inline escapes remain unsupported in interpolated text; literal sections need no escaping. PTC unit tests cover both modes and runtime languages, and the recorded `ptc-turn` scenario preserves tool-template examples in the model-visible prompt.

.agents/notes/implemented/architecture/2026-07-05-prompt-variables-and-tool-guidance-ownership.zh.md

@@ -26,7 +26,7 @@ Status: implemented
 
 ### 提示词变量
 
-插件通过 `ctx.systemPrompt.variable(name, provider)` 注册 `{{name}}` 值。组装过程将它们解析到 waterfall 可见的变量映射中。渲染阶段拒绝以下情况：引用未知的自有属性、已注册的提供方返回 `undefined`、格式错误的完整引用、以及仍包含闭合 `}}` 的不平衡引用；孤立的未匹配 `{{` 保留为行文，替换后的值不会被重新扫描。注册阶段拒绝无效或重复的变量名，section 名称也必须唯一。
+插件通过 `ctx.systemPrompt.variable(name, provider)` 注册 `{{name}}` 值。组装过程将它们解析到 waterfall 可见的变量映射中。渲染阶段拒绝以下情况：引用未知的自有属性、已注册的提供方返回 `undefined`、格式错误的完整引用、以及仍包含闭合 `}}` 的不平衡引用；孤立的未匹配 `{{` 保留为行文，替换后的值不会被重新扫描。注册阶段拒绝无效或重复的变量名，section 名称也必须唯一。段可设置 `interpolate: false` 来原样保留生成的文档；`tools:sdk` 使用此设置，因为工具描述和 schema 可能会介绍自身的 `{{…}}` 语法。
 
 `dsh-agent-loop` 注册两个内置变量，均为上下文 agent 的纯投影：`model`（= `options.model`）和 `cwd`（= `session.header.cwd`）。示例 persona 写 `powered by the {{model}} model`——模型名称只在 `model:` 配置键中声明一次。`{{cwd}}` 仅在 ACP 示例中演示：每个 ACP 会话携带客户端的 cwd，而配置预创建的 stdio agent 没有 cwd（在那里声称 `{{cwd}}` 的 persona 会导致该轮次失败——这是有意为之）。变量留在 loop 插件上（不同于下面的 section）：它们是本循环驱动的 agent 的运行时事实，替换循环自行提供自己的变量。
 
@@ -47,7 +47,7 @@ Status: implemented
 - **循环自行组合一行 identity 文本**：在必须保持精简的那个包（「用插件，不改循环」）中硬编码面向模型的行文，且在 section 流水线之外构成第二条组合路径。（identity 确实以代码字面量交付——但作为 `dsh-system-prompt` 注册的普通 section，其 `system-prompt/assemble` waterfall 仍是部署需要移除它时的逃生阀。）
 - **通过 `agent/request` waterfall 注入模型名称**：提示词文本会在两处组合，更早渲染的 persona 也可能与最终已路由 header 不一致。拥有延迟路由的请求插件还必须拥有该模型在提示词中更早出现的声明。
 - **在每个 persona 中手写模型名称**：与上方一行的 `model:` 键重复，配置修改后静默失实；正是本决策要治愈的病症。
-- **宽松插值（未知引用保留原样或替换为空）**：一个拼写错误 `{{modle}}`（或一个空洞）会被发送给模型，直到 transcript（文本记录）审查时才会被发现。
+- **宽松插值（未知引用保留原样或替换为空）**：一个拼写错误 `{{modle}}`（或一个空洞）会被发送给模型，直到 transcript（文本记录）审查时才会被发现。仅保留未知名称仍会替换工具文档中的已注册名称。
 - **在配置中为每个 subagent 实例编写措辞**：面向模型的行文回到每个部署 × 实例中，重蹈在 leaf YAML 中手写指导的漂移。**根据提供方名称选择措辞**：`providerName` 本身是配置，重命名提供方后会静默获得错误的措辞。
 - **在 `apply` 时解析提供方（加载顺序要求）**与**仅用 section 承载 subagent 措辞（在 assemble 时惰性解析）**：提供方生命周期事件的替代方案；两者均在[提供方生命周期事件 Agent Note](../../archived/architecture/2026-07-05-subagent-provider-lifecycle-events.md)中被否决。
 
@@ -60,7 +60,7 @@ Status: implemented
 
 - tui-agent 的提示词通过一条组装路径依次渲染 identity、带插值模型名的 persona，然后是 fs/shell/web 指导。
 - fork 和 fresh subagent 的描述反映提供方是否继承已完成的对话轮次；工具随提供方生命周期变化而出现、消失和重新措辞。
-- 未知、无值、格式错误或不平衡的变量引用会指明 section 名称并抛出异常；重复的 section、变量和工具注册同样抛出异常。
+- 在启用插值的段中，未知、无值、格式错误或不平衡的变量引用会指明 section 名称并抛出异常；重复的 section、变量和工具注册同样抛出异常。
 - 快照回放与提示词无关：它按轮次和步骤索引已记录的分片流，不比较发出的请求。
 
 ## 后果
@@ -69,4 +69,4 @@ Status: implemented
 - `{{model}}` 在组装时反映 `AgentOptions.model`。如果一个插件在 `agent/request` waterfall 中切换模型，提示词对该步骤的声明就会过时；如果一个插件在那里提供模型（options.model 未设置——循环文档中记载的回退路径），变量在渲染时无值，包含 `{{model}}` 的 persona 会在 waterfall 运行前失败。两者的补救方式相同，就是归属规则本身：拥有延迟绑定模型事实的插件在 `system-prompt/assemble` waterfall 上提前声明它（`assembly.variables['model'] = …`）——一个归属方，两处声明；一个循环测试端到端固定了 supply 路径。已接受。
 - 当一个已绑定的提供方不存在时（尚未激活、已卸载、HMR（热模块替换）重载中），subagent 工具不存在，该窗口内的模型请求中不会包含它。这是诚实的状态——替代方案是注册一个 description 或执行都不可信的工具。
 - 严格性意味着 persona 可能在渲染时导致轮次失败（例如在无 cwd 的会话上使用 `{{cwd}}`）。失败是受控的——该轮次以 `error` 结束，循环存活——且这是一个我们希望明确暴露的撰写错误。
-- 目前没有在提示词行文中转义字面 `{{name}}` 的语法；如果真实提示词确实需要，再行添加。
+- 插值文本仍不支持行内转义；字面文本段无需转义。PTC 单元测试覆盖两种模式和运行时语言，录制的 `ptc-turn` 场景在模型可见的提示词中保留工具模板示例。

.agents/notes/implemented/architecture/2026-07-10-single-file-executable-sdk-runtime-distribution.i18n.yaml

@@ -2,5 +2,5 @@
 # side as of the last confirmed-consistent state. Both languages carry equal authority;
 # after editing either side, bring the other along and re-record with:
 #   pnpm run verify-translation-pairing --write .agents/notes/implemented/architecture/2026-07-10-single-file-executable-sdk-runtime-distribution.md
-2026-07-10-single-file-executable-sdk-runtime-distribution.md: 7792e24a5869be6b7bae7787a6a481f3075ba740
-2026-07-10-single-file-executable-sdk-runtime-distribution.zh.md: 857bfaec80da962fac0403f2239e0a5e71954194
+2026-07-10-single-file-executable-sdk-runtime-distribution.md: 8c55c142137f0ab24869bd57ffed3835b7b0516a
+2026-07-10-single-file-executable-sdk-runtime-distribution.zh.md: 52ca2e13901d9469e2f9663936acb766a934e8f7

.agents/notes/implemented/architecture/2026-07-23-client-plugin-loading-model.i18n.yaml

@@ -2,5 +2,5 @@
 # side as of the last confirmed-consistent state. Both languages carry equal authority;
 # after editing either side, bring the other along and re-record with:
 #   pnpm run verify-translation-pairing --write .agents/notes/implemented/architecture/2026-07-23-client-plugin-loading-model.md
-2026-07-23-client-plugin-loading-model.md: 21bad78792c6b5aad48b51f454f6c08c0400ad72
-2026-07-23-client-plugin-loading-model.zh.md: 2758f28f3bd34131ece3bed74152fbfe0b36174e
+2026-07-23-client-plugin-loading-model.md: c5576b148c5991dd498d3aed605e3c2e3395774b
+2026-07-23-client-plugin-loading-model.zh.md: c7d6982c2680995bd4698ddbff052a7708f69997

.agents/notes/implemented/architecture/2026-07-23-client-plugin-loading-model.md

@@ -52,7 +52,7 @@ What happens between `dsh web` starting and the UI appearing? Three stages: the
 
 **Host side — compose the graph.**
 
-1. The composing app (`apps/cli`) ships the roster as ordinary rows in its `cordis.yml` config tree — client plugin packages are entry rows like every host plugin, including the always-mounted `client-hmr` row. A roster row that fails to import is caught by `assertEntriesLoaded`; a row whose fiber rejects is reported with its original stack by `assertEntriesActivated` ([host boot decision](2026-07-24-web-config-tree-boot-and-transport-layering.md)).
+1. The composing app (`apps/cli`) ships the roster as ordinary rows in its `cordis.yml` config tree — client plugin packages are entry rows like every host plugin, including the always-mounted `client-hmr` row. `auditStartupEntries` reports failed imports, activation errors with their original stacks, and pending dependencies; optional entries warn and required entries reject startup ([startup policy](2026-09-09-consumer-owned-startup-strictness.md)).
 2. The `dsh-client-modules` node half (the package is dual-face: its browser half is the module table) resolves each live Loader entry through the same `name` and owning-tree `baseUrl` inputs that imported its Host face, then reads the nearest owning package.json `dsh.client` declaration and composes `window.__DSH_BOOT__`: `{ rev, entries: [{ id, url, rev, inject?, immediately?, external? }], batches: [{ phase, url, rev, entries }] }`. The manifest package name is the browser module identity even when an overlay names a relative source or built entry file. Distinct active Loader sources resolving to one package name fail composition; after one source unloads, the surviving source supplies the row without a fiber restart. The row's three optional fields come from manifests, never hand-copied. Composition orders requested dynamic rows before their consumers, rejects synchronous request cycles, and assigns every row to exactly one initial batch. It refuses declared plugins without built `./client` bundles and groups their package/path rows under one required source-build instruction; malformed declaration fields also fail activation, and the Host audit reports either error from the FAILED fiber.
 3. Scanning is incremental per package — there is no full-rescan code path. Each cordis `internal/plugin` emission marks the fiber's entry name dirty (entry-less fibers drop O(1)); a microtask flush reconciles each dirty name against live loader entries, with package metadata (including the negative "not a client package" verdict) cached per entry name and owning-tree base URL for the process lifetime and bundle re-hashing reachable only through `rebuilt(id)`. The activation pass seeds the same dirty set from current entries and flushes synchronously, so first scan and steady state share one implementation. Initial rows receive an opaque process nonce plus sequence without hashing their artifacts; startup combo revisions hash the combined script inputs plus indexed map, and the rows plus batch descriptors hash into `graph.rev`. The graph types are single-sourced in the modules package's `./client` export — the webserver knows nothing about the graph, while modules registers the combo route and contributes structured index-injection rows.
 

.agents/notes/implemented/architecture/2026-07-23-client-plugin-loading-model.zh.md

@@ -52,7 +52,7 @@ Host 会快照每个已构建插件产物，并把每个调度阶段的有序 ro
 
 **host 侧——组合这张图。**
 
-1. 负责组合的 app（`apps/cli`）把名册作为普通行放进它的 `cordis.yml` 配置树——client 插件包与每个 host 插件一样是 entry 行，包括无条件挂载的 `client-hmr` 行。名册行 import 失败由 `assertEntriesLoaded` 捕获；fiber reject 的行则由 `assertEntriesActivated` 报告原始 stack（[host boot 决策](2026-07-24-web-config-tree-boot-and-transport-layering.zh.md)）。
+1. 负责组合的 app（`apps/cli`）把名册作为普通行放进它的 `cordis.yml` 配置树——client 插件包与每个 host 插件一样是 entry 行，包括无条件挂载的 `client-hmr` 行。`auditStartupEntries` 报告 import 失败、带原始 stack 的激活错误，以及待满足的依赖；optional entry 输出 warning，required entry 则使启动失败（[启动策略](2026-09-09-consumer-owned-startup-strictness.zh.md)）。
 2. `dsh-client-modules` 的 node 半（该包是双面的：浏览器半就是模块表）使用 Host face import 时相同的 `name` 与所属 tree `baseUrl` 解析每个 live Loader entry，再读取最近归属 package.json 的 `dsh.client` 声明并组合出 `window.__DSH_BOOT__`：`{ rev, entries: [{ id, url, rev, inject?, immediately?, external? }], batches: [{ phase, url, rev, entries }] }`。即使 overlay 指向相对的 source 或 built entry 文件，manifest 包名仍是浏览器模块身份。若不同的 active Loader source 解析到同一包名，组合会失败；一个来源卸载后，仍存活的来源无需重启 fiber 即可提供该 row。Row 的三个可选字段都来自 manifest，永不人肉抄写。组合会把被请求的动态图 row 排到消费者之前、拒绝同步请求环，并把每个 row 恰好分配给一个初始批次。它会拒绝没有已构建 `./client` bundle 的已声明插件，并把它们的 package/path 行归到一条源码构建要求下；畸形声明字段同样会让激活失败，Host 检查会从 FAILED fiber 报告这两类错误。
 3. 扫描是单包增量——不存在全量重扫代码路径。每次 cordis `internal/plugin` 发射把该 fiber 的 entry 名标脏（无 entry 的 fiber O(1) 丢弃）；微任务 flush 把每个脏名对账 live loader entries，包元数据（含「非 client 包」的否定结论）按 entry 名与所属 tree base URL 缓存至进程结束，bundle 重哈希只经 `rebuilt(id)` 可达。激活趟从当前 entries 灌同一脏集合并同步 flush，初扫与稳态共享一条实现。初始 row 使用不透明的进程 nonce 加序号，不对其产物求哈希；启动 combo revision 对合并脚本输入及 indexed map 求哈希，row 与批次描述再共同哈希进 `graph.rev`。图类型单源在 modules 包的 `./client` 出口——webserver 对图一无所知；modules 会注册 combo 路由并贡献结构化 index 注入行。
 

.agents/notes/implemented/architecture/2026-07-24-single-harness-home-resolver.i18n.yaml

@@ -2,5 +2,5 @@
 # side as of the last confirmed-consistent state. Both languages carry equal authority;
 # after editing either side, bring the other along and re-record with:
 #   pnpm run verify-translation-pairing --write .agents/notes/implemented/architecture/2026-07-24-single-harness-home-resolver.md
-2026-07-24-single-harness-home-resolver.md: 02578674b4547bae54b2edce99c6c4ef53748588
-2026-07-24-single-harness-home-resolver.zh.md: 5764be3f85901196cd2a2db77d1fc3838668163f
+2026-07-24-single-harness-home-resolver.md: 1191c52bfa222b807e9aa301b250d2efcc8b0953
+2026-07-24-single-harness-home-resolver.zh.md: cefa124b7cb46bab92157af34576f59d3053da9f

.agents/notes/implemented/architecture/2026-07-24-single-harness-home-resolver.md

@@ -23,6 +23,8 @@ explicit configured path  >  $DSH_HOME  >  ~/.dsh
 
 An empty or whitespace-only `$DSH_HOME` is treated as unset; otherwise `resolve('')` would silently place the home at the current working directory. The harness keeps all user data under one root; there is no XDG config/data/cache split. `dshHomePath(...segments)` joins deployment-owned children onto that root, and `dsh-app-boot` exposes it to Loader `!!js` config expressions before mounting entries, so shipped compositions derive `sessions` and `storages` without copying the resolver. `dshHomeDisplay()` names a resolved root symbolically for user-facing paths — `~/.dsh` for the default home, `$DSH_HOME` for any configured home — so the user-global `AGENTS.md` label never leaks an absolute machine path. It replaces agent-instructions's bespoke default-vs-`$DSH_HOME` check.
 
+`dshCachePath(...segments)` derives paths below the resolved home's `cache` directory. An initial `{ dshHome }` option preserves a provider's explicit home override. It resolves paths without creating directories; callers own directory creation. `attachment-local` uses this helper for regenerable request-image variants while retaining durable attachment objects in their versioned storage tree, so clearing the cache cannot remove Session attachments. Existing request-image cache entries are left in place and are not read or copied; a cache miss regenerates the variant from its durable attachment.
+
 `@deepseek-ai/dsh-home` is deleted. Home-owning providers and boot packages import `resolveDshHome` from `dsh-home-paths`; composition bundles contain only the resolved configuration rows.
 
 `dsh-telemetry` and its separate home policy are absent under the [SDK project toolchain removal](../../archived/simplification/2026-08-11-remove-sdk-project-toolchain.md), leaving this resolver as the sole home policy.

.agents/notes/implemented/architecture/2026-07-24-single-harness-home-resolver.zh.md

@@ -23,6 +23,8 @@ explicit configured path  >  $DSH_HOME  >  ~/.dsh
 
 空或仅含空白的 `$DSH_HOME` 被当作未设置处理；否则，`resolve('')` 会悄悄把 home 落在当前工作目录。harness 把所有用户数据都放在同一个根目录下；不存在 XDG 的 config/data/cache 拆分。`dshHomePath(...segments)` 将部署负责的子路径拼接到该根目录下，`dsh-app-boot` 在挂载条目前向 Loader `!!js` 配置表达式暴露它，因此出厂组合无需复制解析器即可派生 `sessions` 和 `storages`。`dshHomeDisplay()` 为面向用户的路径以符号形式命名已解析的根目录——默认 home 显示为 `~/.dsh`，任何已配置的 home 显示为 `$DSH_HOME`——这样用户全局的 `AGENTS.md` 标签就绝不会泄露机器上的绝对路径。它取代了 agent-instructions 中自定义的「默认值 vs `$DSH_HOME`」判断。
 
+`dshCachePath(...segments)` 在解析出的主目录下的 `cache` 目录中派生路径。首个 `{ dshHome }` 选项保留提供方显式配置的主目录覆盖值。它只解析路径，不创建目录；目录创建由调用方负责。`attachment-local` 将此函数用于可重新生成的请求图片版本，持久附件对象仍保留在其带版本的存储树中，因此清空缓存不会删除 Session 附件。已有请求图片缓存条目保留在原处，不再读取或复制；缓存未命中时从持久附件重新生成请求版本。
+
 `@deepseek-ai/dsh-home` 被删除。拥有 home 配置的提供方与 boot 包从 `dsh-home-paths` 导入 `resolveDshHome`；组合包只包含解析后的配置行。
 
 `dsh-telemetry` 及其独立 home 策略已随 [SDK 项目工具链移除](../../archived/simplification/2026-08-11-remove-sdk-project-toolchain.md)一并消失，因此该解析器是唯一的 home 策略。

.agents/notes/implemented/architecture/2026-07-24-web-config-tree-boot-and-transport-layering.i18n.yaml

@@ -2,5 +2,5 @@
 # side as of the last confirmed-consistent state. Both languages carry equal authority;
 # after editing either side, bring the other along and re-record with:
 #   pnpm run verify-translation-pairing --write .agents/notes/implemented/architecture/2026-07-24-web-config-tree-boot-and-transport-layering.md
-2026-07-24-web-config-tree-boot-and-transport-layering.md: e7c0781e1504bce12a8b0d2197bd47b37d3e873a
-2026-07-24-web-config-tree-boot-and-transport-layering.zh.md: 9a0affa397151a3994c5731bebda4b45a4139b6c
+2026-07-24-web-config-tree-boot-and-transport-layering.md: c8a1db3cd7bd715d8912b22a0b41a9ef660c1316
+2026-07-24-web-config-tree-boot-and-transport-layering.zh.md: 41fcf65b06829199c203faa806bd17a310b68bcf

.agents/notes/implemented/architecture/2026-07-24-web-config-tree-boot-and-transport-layering.md

@@ -12,7 +12,7 @@ English | [中文](2026-07-24-web-config-tree-boot-and-transport-layering.zh.md)
 
 ## Decision
 
-**Composition is one flat assembled tree.** `apps/cli/config/base.cordis.yml` plus `apps/cli/config/web.cordis.yml` holds every row — the host runtime (32 rows), the `api-gateway` row, the `webserver` row, and the `dsh.client` rows (the browser roster; the modules row is simultaneously a host row). No spine bundle: every plugin is one row and every config field is yml-editable. That stance later became repository-wide, with the rows both surfaces share factored into `apps/cli/config/base.cordis.yml` and each surface reduced to an overlay ([shared-base overlays](../../archived/simplification/2026-07-29-shared-base-config-overlays.md)). The `dsh-client-hmr` row is an ordinary always-on bundle row (originally appended in code by `--dev`; the flag is retired). Row order carries no load semantics; activation is service-availability driven. The shared audit rejects imports with no fiber, awaits only failed fibers to recover original activation errors, and reports services that leave a fiber `PENDING`; before throwing, it marks those exact rejection reasons through one process checkpoint so `installFailLoud` coalesces Loader's duplicate notification while unrelated unhandled rejections remain fatal. The Node app-boot artifact embeds `@cordisjs/plugin-include` while leaving `@cordisjs/plugin-loader` external, so the include's `EntryTree` and the host bind to one Loader peer instead of splitting a config tree across two Loader implementations.
+**Composition is one flat assembled tree.** `apps/cli/config/base.cordis.yml` plus `apps/cli/config/web.cordis.yml` holds every row — the host runtime (32 rows), the `api-gateway` row, the `webserver` row, and the `dsh.client` rows (the browser roster; the modules row is simultaneously a host row). No spine bundle: every plugin is one row and every config field is yml-editable. That stance later became repository-wide, with the rows both surfaces share factored into `apps/cli/config/base.cordis.yml` and each surface reduced to an overlay ([shared-base overlays](../../archived/simplification/2026-07-29-shared-base-config-overlays.md)). The `dsh-client-hmr` row is an ordinary always-on bundle row (originally appended in code by `--dev`; the flag is retired). Row order carries no load semantics; activation is service-availability driven. `auditStartupEntries` reports failed imports, reads failed fibers for original activation errors, and lists services that leave a fiber `PENDING`. The [startup policy](2026-09-09-consumer-owned-startup-strictness.md) makes optional failures warnings and required failures fatal. Reported rejection reasons stay marked through one process checkpoint so `installFailLoud` coalesces Loader's duplicate notification while unrelated unhandled rejections remain fatal. The Node app-boot artifact embeds `@cordisjs/plugin-include` while leaving `@cordisjs/plugin-loader` external, so the include's `EntryTree` and the host bind to one Loader peer instead of splitting a config tree across two Loader implementations.
 
 **Boot glue is a class pair.** `AppCLIEntry` (apps/cli) and `AppWebEntry` (the shell kernel) hold only what must exist independently of cordis: argv facts, the composed patch set, the parsed boot manifest, the module system instance, loading-page handles — everything else lives in plugins. `AppCLIEntry.run()` is three stages: layered env (ambient > cwd `.env` > `$DSH_HOME/.env`, closing the defect above) → patch composition → Loader include boot plus the activation audit. `AppWebEntry.run()` mirrors it browser-side: parse `window.__DSH_BOOT__` into a `BootManifest` (two views: npm-package rows for the module table, cordis-plugin rows for entry composition; malformed wire throws), build the module system, render the loading page, prefetch the `immediately` tier in parallel with Context/Loader setup, **await the prefetch before creating entries** (materialization is `tree.import`'s synchronous require, unprotected by fiber inject waiting; cross-package require edges such as i18n → runtime/client need every immediately-tier factory registered first — an empirically found 10–25% boot race otherwise), adopt the modules entry, create the graph rows, settle, sweep.
 

.agents/notes/implemented/architecture/2026-07-24-web-config-tree-boot-and-transport-layering.zh.md

@@ -12,7 +12,7 @@ Status: implemented
 
 ## 决策
 
-**组合结果是一棵平铺配置树。** `apps/cli/config/base.cordis.yml` 与 `apps/cli/config/web.cordis.yml` 共同持有全部行——host 运行时（32 行）、`api-gateway` 行、`webserver` 行、`dsh.client` 行（浏览器 roster；modules 行同时是 host 行）。不做主干 bundle：每插件一行、每个 config 字段 yml 可改。这一立场后来推广到全仓：两个 surface 共享的配置项被抽取进 `apps/cli/config/base.cordis.yml`，各 surface 则收敛为一份 overlay（[共享 base overlay](../../archived/simplification/2026-07-29-shared-base-config-overlays.md)）。`dsh-client-hmr` 行是普通的始终启用的 bundle 行（最初由 `--dev` 在代码中追加；该旗标已废除）。行序无装载语义；激活由服务可用性驱动。共享 audit 会拒绝没有 fiber 的 import、仅等待失败的 fiber 以恢复原始激活错误，并报告让 fiber 停在 `PENDING` 的服务；抛出错误前，审计会通过一个进程级检查点标记这些 rejection 的确切原因，从而让 `installFailLoud` 将 Loader 的重复通知合并为一次，而无关的未处理 rejection 仍然致命。Node app-boot 产物内嵌 `@cordisjs/plugin-include`，但将 `@cordisjs/plugin-loader` 保持为外部依赖，因此 include 的 `EntryTree` 与 host 会绑定到同一个 Loader peer，而不会让一棵配置树横跨两个 Loader 实现。
+**组合结果是一棵平铺配置树。** `apps/cli/config/base.cordis.yml` 与 `apps/cli/config/web.cordis.yml` 共同持有全部行——host 运行时（32 行）、`api-gateway` 行、`webserver` 行、`dsh.client` 行（浏览器 roster；modules 行同时是 host 行）。不做主干 bundle：每插件一行、每个 config 字段 yml 可改。这一立场后来推广到全仓：两个 surface 共享的配置项被抽取进 `apps/cli/config/base.cordis.yml`，各 surface 则收敛为一份 overlay（[共享 base overlay](../../archived/simplification/2026-07-29-shared-base-config-overlays.md)）。`dsh-client-hmr` 行是普通的始终启用的 bundle 行（最初由 `--dev` 在代码中追加；该旗标已废除）。行序无装载语义；激活由服务可用性驱动。`auditStartupEntries` 报告 import 失败、读取失败 fiber 的原始激活错误，并列出让 fiber 停在 `PENDING` 的服务。[启动策略](2026-09-09-consumer-owned-startup-strictness.zh.md)将 optional failure 作为 warning，将 required failure 视为致命错误。已报告的 rejection 原因会保持标记至一个进程级检查点，使 `installFailLoud` 合并 Loader 的重复通知，而无关的未处理 rejection 仍然致命。Node app-boot 产物内嵌 `@cordisjs/plugin-include`，但将 `@cordisjs/plugin-loader` 保持为外部依赖，因此 include 的 `EntryTree` 与 host 会绑定到同一个 Loader peer，而不会让一棵配置树横跨两个 Loader 实现。
 
 **boot 胶水由两个类组成。** `AppCLIEntry`（apps/cli）与 `AppWebEntry`（壳内核）只持有那些必须独立于 cordis、提前存在的东西：argv 事实、合成的 patch 集、解析出的 boot manifest（元数据清单）、模块系统实例、loading 页句柄——其余一律进插件。`AppCLIEntry.run()` 三段：分层 env（ambient > cwd `.env` > `$DSH_HOME/.env`，顺手关掉上述缺陷）→ patch 合成 → Loader include boot 加 activation audit。`AppWebEntry.run()` 在浏览器侧镜像它：把 `window.__DSH_BOOT__` 解析成 `BootManifest`（双视角：npm 包行给模块表、cordis 插件行给 entry 组合；畸形 wire 大声抛）、建模块系统、渲染 loading 页、immediately 层预取与 Context/Loader 准备并行、**create entry 之前等预取齐**（物化是 `tree.import` 的同步 require，不受 fiber inject 等待保护；i18n → runtime/client 这类跨包 require 边要求 immediately 层工厂全部注册完——否则有实测 10–25% 的 boot 竞态）、收编 modules entry、逐一创建图行、settle、sweep。
 

.agents/notes/implemented/architecture/2026-08-03-per-session-agent-presets.i18n.yaml

@@ -2,5 +2,5 @@
 # side as of the last confirmed-consistent state. Both languages carry equal authority;
 # after editing either side, bring the other along and re-record with:
 #   pnpm run verify-translation-pairing --write .agents/notes/implemented/architecture/2026-08-03-per-session-agent-presets.md
-2026-08-03-per-session-agent-presets.md: 8af48979b49f08c8e3ac945f648acbb615757a98
-2026-08-03-per-session-agent-presets.zh.md: 2889487e989d093c162848c3c972d46fece3726d
+2026-08-03-per-session-agent-presets.md: c2c9f670df480662804e086fbf150c773d8f4fc8
+2026-08-03-per-session-agent-presets.zh.md: 9a4760be9e4a9d1d2de0a7f8d2d3bce755c04480

.agents/notes/implemented/architecture/2026-08-03-per-session-agent-presets.md

@@ -33,7 +33,7 @@ The `agent-presets` user-settings namespace carries `modeSelectionEnabled` and `
 
 **The effective default is read per resolution, never snapshotted.** A cached value would need a `watch` subscription and a reload path to stay honest, and the resolved scope already re-reads a hot-reloaded document. The Host setting itself applies when an unnamed session is resolved afterwards. An explicit Web Settings action additionally routes its accepted effective default through the existing blank-session selection path only when the captured session id is still current and blank; it never recomposes a running session or rewrites that session's history. The session log enforces the same invariant from the other side — the header records the id a session was CREATED with and an `agent-preset/selected` event records any later blank-session switch, so a reader resolves the pair (`resolveSessionPreset`) and never the header alone: a resume rebuilds the composition its history was produced under rather than the deployment default at resume time, a cold transcript's presenters resolve in that composition's layer, and the gateway rejects an attempt to adopt a live session under a preset other than the one it currently runs. A snapshot would make the two disagree at exactly the moment the setting changes.
 
-**A directly-plugged subtree is invisible to the boot audit.** It never links itself to an `Entry`, so it is absent from `ctx.loader.entries()` and `assertEntriesActivated` cannot see it. The mount audits its own rows instead, reading the tree through an `Include` subclass that publishes it.
+**A directly-plugged subtree is invisible to the boot audit.** It never links itself to an `Entry`, so it is absent from `ctx.loader.entries()` and `auditStartupEntries` cannot see it. The mount reads its own rows through an `Include` subclass that publishes the tree and requires every enabled row to activate, independently of the [application startup policy](2026-09-09-consumer-owned-startup-strictness.md).
 
 **A preset can only name a group because the app registers one.** Sharing a realm across rows is a `cordis:group` row, and a preset living outside this workspace — the authored ones under the Harness home, which is the point — cannot resolve `@cordisjs/plugin-group` by name: Node's upward `node_modules` walk never reaches the harness from there. `boot()` therefore registers `cordis:group` beside `cordis:include` as a loader builtin, so both load through the ambient module pipeline rather than through the included tree's own specifier resolution. Without it the `isolate` vocabulary above is expressible one row at a time only, and a provider could never be grouped with its consumers.
 

.agents/notes/implemented/architecture/2026-08-03-per-session-agent-presets.zh.md

@@ -33,7 +33,7 @@ Status: implemented
 
 **有效默认值在每次解析时读取，绝不保存快照。** 缓存下来就需要一个 `watch` 订阅和一条重载路径才能保持诚实，而解析后的 scope 本来就会重读热重载过的文档。Host 设置本身会在此后解析未指名会话时生效。Web Settings 中的明确操作还会把已接受的有效默认值送入既有的空白会话选择链路，但只在操作前捕获的会话 id 仍是当前空白会话时对齐；它绝不会重新组装运行中的会话，也不会改写该会话的历史。session 日志从另一侧执行同一条不变量——header 记录会话**创建时**的 id，此后空白期的任何切换由 `agent-preset/selected` 事件记录，因此读取方解析的是两者之和（`resolveSessionPreset`）、绝不单看 header：恢复重建的是其历史所产出的那份组装而不是恢复时的部署默认值，冷读记录的 presenter 在那份组装的层里解析，网关也会拒绝把一个活着的会话收编到它当前运行的 preset 以外的 preset 之下。快照会让两者恰好在设置改变的那一刻各说各话。
 
-**直接挂载的子树对启动审计不可见。** 它不会把自己关联到 `Entry`，因此不在 `ctx.loader.entries()` 中，`assertEntriesActivated` 也看不到它。改由挂载过程自行校验各行，通过一个会公开自身 tree 的 `Include` 子类读取。
+**直接挂载的子树对启动审计不可见。** 它不会把自己关联到 `Entry`，因此不在 `ctx.loader.entries()` 中，`auditStartupEntries` 也看不到它。挂载过程通过一个会公开自身 tree 的 `Include` 子类读取各行，并要求每个启用行都激活，不受[应用启动策略](2026-09-09-consumer-owned-startup-strictness.zh.md)影响。
 
 **preset 能写出 group，是因为 app 注册了它。** 跨行共享 realm 就是一个 `cordis:group` 行，而住在本工作区之外的 preset——也就是 Harness home 下由人或 agent 创作的那些，正是这套设计的目的——无法按名字解析 `@cordisjs/plugin-group`：Node 向上查找 `node_modules` 的路径从那里永远走不到 harness。因此 `boot()` 把 `cordis:group` 与 `cordis:include` 并排注册为 loader builtin，两者都经由环境模块管线加载，而不依赖被包含树自身的说明符解析。没有它，上文那套 `isolate` 词汇就只能一行一行地表达，提供方也永远无法与它的消费方归入同一组。
 

.agents/notes/implemented/architecture/2026-08-25-electron-desktop-packaging-and-updates.i18n.yaml

@@ -2,5 +2,5 @@
 # side as of the last confirmed-consistent state. Both languages carry equal authority;
 # after editing either side, bring the other along and re-record with:
 #   pnpm run verify-translation-pairing --write .agents/notes/implemented/architecture/2026-08-25-electron-desktop-packaging-and-updates.md
-2026-08-25-electron-desktop-packaging-and-updates.md: 6d22777c8913911dbb1d89c09de9e891636873c8
-2026-08-25-electron-desktop-packaging-and-updates.zh.md: 5108492ba6f029847735f570aa77c2cb505f981c
+2026-08-25-electron-desktop-packaging-and-updates.md: 4a4dfe8910905b3c35fbdfdcaedd34a556b532f3
+2026-08-25-electron-desktop-packaging-and-updates.zh.md: 69877127578ae4d61643653403b736dbb5e7b1e6

.agents/notes/implemented/architecture/2026-08-28-subprocess-native-containment.i18n.yaml

@@ -2,5 +2,5 @@
 # side as of the last confirmed-consistent state. Both languages carry equal authority;
 # after editing either side, bring the other along and re-record with:
 #   pnpm run verify-translation-pairing --write .agents/notes/implemented/architecture/2026-08-28-subprocess-native-containment.md
-2026-08-28-subprocess-native-containment.md: 8e1e12a8536f56c9ccb515cec4c07c730c8b9b84
-2026-08-28-subprocess-native-containment.zh.md: 83bc20a29d937bca9530ca107712e4a7b39d8d4d
+2026-08-28-subprocess-native-containment.md: 7523f9d66e7a302f6ce9c77d93671a5b1303a5aa
+2026-08-28-subprocess-native-containment.zh.md: 252a8e9fd058cf37a1a7a70f09394a6811d58580

.agents/notes/implemented/architecture/2026-08-28-subprocess-native-containment.md

@@ -22,7 +22,7 @@ The first eligible Linux ordinary or PTY call in one runtime deeply checks the e
 
 The parent creates one 0700 directory with a complete 0600 `launch-request.json` containing the final target cwd and environment. The private `DSH_SUBPROCESS_RUNNER` value locates that request while the runner starts from the provider cwd and a bootstrap-safe environment. `systemd-run --user --scope --quiet --collect --expand-environment=no` registers its process in the scope, then the one-shot bootstrap removes and validates the request, changes to the target cwd, restores the complete target environment, resolves a bare executable with the target PATH rules, clears `FD_CLOEXEC` on fd 0 through fd 2, and calls libc `execve()` with the original argv. The bootstrap becomes the target in place and preserves its inherited stdio; it does not remain as a supervisor.
 
-Request consumption or a manager observation of a loaded unit establishes scope ownership. Unit absence before either fact remains unresolved while the direct launcher is running. If that launcher exits while the request remains unconsumed, the direct result rejects with startup failure unless its observed signal matches a termination requested while the launcher was running. A matching signal preserves the actual exit outcome for both ordinary and PTY launches; a recorded startup error always takes precedence. Range observation independently resolves an empty range when the launcher has exited and the unit is absent. The parent checks this unresolved interval every 50 milliseconds; after establishment, state queries back off exponentially to the existing 5-second systemctl bound. Each query reads both `LoadState` and `ActiveState`: loaded `inactive` or `failed`, or an established unit becoming `not-found`/`inactive` or otherwise collected away, proves the range empty. `active`, `activating`, `reloading`, and `deactivating` remain nonterminal. Unknown or malformed combinations and unreadable manager results reject `waitForExit()` instead of claiming quiescence. `terminate()` wakes a sleeping observer for an immediate recheck, and settlement cancels the losing backoff sleep. A strict sibling `startup-error.json` carries only request/bootstrap or target pre-exec failure, and the parent removes this spawn's private paths at observable lifecycle completion.
+Request consumption or a manager observation of a loaded unit establishes scope ownership. Unit absence before either fact remains unresolved while the direct launcher is running. If that launcher exits while the request remains unconsumed, the direct result rejects with startup failure unless its observed signal matches a termination requested while the launcher was running. A matching signal preserves the actual exit outcome for both ordinary and PTY launches; a recorded startup error always takes precedence. Range observation independently resolves an empty range when the launcher has exited and the unit is absent. The parent checks this unresolved interval every 50 milliseconds; after establishment, state queries back off exponentially to the existing 5-second systemctl bound. Each query reads `LoadState`, `ActiveState`, and `TasksCurrent`: loaded `inactive` or `failed`, or an established unit becoming `not-found`/`inactive` or otherwise collected away, proves the range empty. `active`, `activating`, `reloading`, and `deactivating` remain nonterminal, except that once termination was requested and the launcher has exited, a still-active unit reporting no processes is an empty range: the manager ends a scope only on the populated-to-empty transition, so a payload killed before it entered the cgroup never triggers one. That conclusion stops the leftover unit so transient units cannot accumulate, while an unreported or `[not set]` process count stays unknown and keeps waiting. Unknown or malformed combinations and unreadable manager results reject `waitForExit()` instead of claiming quiescence. `terminate()` wakes a sleeping observer for an immediate recheck, and settlement cancels the losing backoff sleep. A strict sibling `startup-error.json` carries only request/bootstrap or target pre-exec failure, and the parent removes this spawn's private paths at observable lifecycle completion.
 
 The ordinary target result still comes from the same child process. The PTY path uses the same request and bootstrap without a resident runner, so the `node-pty` PID, process group, session leader, controlling terminal, foreground `inputWaiting`, `/dev/tty`, readiness, and direct terminal outcome retain their existing meanings while scope membership covers `setsid` and reparented descendants.
 
@@ -54,7 +54,7 @@ This note owns the current native-containment mechanism. It partially updates th
 
 ## Verification
 
-- Provider and Linux protocol suites pin synchronous NUL rejection before launch side effects, strict request/error decoding, target cwd and complete environment restoration, private-variable collision, symlink-sensitive PATH traversal with preserved argv, close-on-exec removal for inherited stdio, pre-exec error ownership, failed-deep-probe retry plus successful-deep-probe caching with per-call manager checks, the three scope-establishment states including requested versus unexpected exits with an unconsumed request, `LoadState`/`ActiveState` parsing, `reloading`, terminate wake-up with losing-delay cancellation, bounded established-scope backoff, and exactly-once PTY managed-owner cleanup.
+- Provider and Linux protocol suites pin synchronous NUL rejection before launch side effects, strict request/error decoding, target cwd and complete environment restoration, private-variable collision, symlink-sensitive PATH traversal with preserved argv, close-on-exec removal for inherited stdio, pre-exec error ownership, failed-deep-probe retry plus successful-deep-probe caching with per-call manager checks, the three scope-establishment states including requested versus unexpected exits with an unconsumed request, `LoadState`/`ActiveState`/`TasksCurrent` parsing, releasing a leftover scope left active with no processes beside the live-client, unterminated, and unset-count cases, `reloading`, terminate wake-up with losing-delay cancellation, bounded established-scope backoff, and exactly-once PTY managed-owner cleanup.
 - Windows protocol and Win32 suites pin exactly two result branches, numeric-only target exits, ordinary-error start cancellation with raw parent-local reasons, the reduced `name`/`message`/`code`/`syscall`/`path` error record, the fixed `2`/`3`/`267` to `ENOENT`, `740` to `EACCES`, `5` to `EPERM`, `193` to `EFTYPE`, and remaining-code to `UNKNOWN` mapping, start delivery after runner spawn, empty-range settlement after pre-spawn failure, explicit ordinally sorted target environment blocks with `=C:` preservation and double-NUL termination, `uv_get_osfhandle()` carrier mapping and unsigned invalid-sentinel rejection, the null-device ignored-stdin carrier and piped non-ignored stdin, result-send and IPC-disconnect failures, direct-result latching before stdio settlement, active-process quiescence, and unique handle cleanup.
 - A keyless [`bash-startup-timeout`](../../../../snapshots/session/bash-startup-timeout/snapshot.yml) Session snapshot pins the model-facing timeout result. A Linux user-systemd fixture holds the launch request unconsumed at an input barrier and verifies cancellation plus range settlement.
 - Real Linux user-systemd tests run one ordinary and one `node-pty` `setsid`/reparent scenario through the production entry. They prove scope signalling and collection, bare executable lookup, escaped-descendant termination, range settlement, and unchanged PTY PID, session, controlling-terminal, foreground-input, `/dev/tty`, readiness, and startup-failure semantics.

.agents/notes/implemented/architecture/2026-08-28-subprocess-native-containment.zh.md

@@ -22,7 +22,7 @@ detached POSIX 进程组、Windows direct-parent 遍历与 PTY 后代扫描只
 
 parent 创建一个 0700 目录，其中的完整 0600 `launch-request.json` 保存最终 target cwd 与环境。私有 `DSH_SUBPROCESS_RUNNER` 值负责定位该 request，runner 则从 provider cwd 与 bootstrap-safe 环境启动。`systemd-run --user --scope --quiet --collect --expand-environment=no` 先把自身进程注册到 scope，再由 one-shot bootstrap 删除并校验 request、切换到 target cwd、恢复完整 target 环境、按 target PATH 规则解析裸可执行文件、清除 fd 0 至 fd 2 的 `FD_CLOEXEC`，并使用原始 argv 调用 libc `execve()`。bootstrap 会原地成为 target 并保留继承的 stdio，不作为常驻 supervisor。
 
-request 被消费或 manager 已观察到 loaded unit 都能建立 scope ownership。在这两项事实出现前，只要 direct launcher 仍在运行，unit absence 就保持未决。如果 launcher 退出时 request 仍未消费，direct result 会以 startup failure reject，除非实际观察到的信号匹配 launcher 仍在运行时请求的终止信号。普通进程与 PTY 进程遇到匹配信号时都会保留实际退出结果；已记录的 startup error 始终优先。launcher 已退出且 unit 不存在时，range observation 会独立结算 empty-range wait。parent 每 50 毫秒检查一次这段未决区间；建立后，状态查询按指数增长间隔退避，最多达到既有的 5 秒 systemctl 上限。每次查询同时读取 `LoadState` 与 `ActiveState`：loaded `inactive` 或 `failed`，以及已经建立的 unit 变为 `not-found`/`inactive` 或被 collect 卸载，都能证明 range 为空。`active`、`activating`、`reloading` 与 `deactivating` 仍是非终态。未知或 malformed 组合以及不可读的 manager 结果会使 `waitForExit()` reject，而不是宣称完全停稳。`terminate()` 会唤醒正在休眠的 observer 立即复查，结算时会取消未胜出的退避 sleep。严格的同目录 `startup-error.json` 只承载 request／bootstrap 或 target pre-exec failure，parent 会在可观察生命周期完成时移除本次 spawn 的私有路径。
+request 被消费或 manager 已观察到 loaded unit 都能建立 scope ownership。在这两项事实出现前，只要 direct launcher 仍在运行，unit absence 就保持未决。如果 launcher 退出时 request 仍未消费，direct result 会以 startup failure reject，除非实际观察到的信号匹配 launcher 仍在运行时请求的终止信号。普通进程与 PTY 进程遇到匹配信号时都会保留实际退出结果；已记录的 startup error 始终优先。launcher 已退出且 unit 不存在时，range observation 会独立结算 empty-range wait。parent 每 50 毫秒检查一次这段未决区间；建立后，状态查询按指数增长间隔退避，最多达到既有的 5 秒 systemctl 上限。每次查询读取 `LoadState`、`ActiveState` 与 `TasksCurrent`：loaded `inactive` 或 `failed`，以及已经建立的 unit 变为 `not-found`/`inactive` 或被 collect 卸载，都能证明 range 为空。`active`、`activating`、`reloading` 与 `deactivating` 仍是非终态；例外是：一旦请求过终止且 launcher 已经退出，报告没有任何进程却仍 active 的 unit 就是空 range——manager 只在观测到 populated→empty 转变时才结束 scope，因此进入 cgroup 前就被杀死的 payload 永远不会触发该转变。该判定会 stop 掉这个遗留 unit，使 transient unit 不会累积；未上报或为 `[not set]` 的进程数仍视为未知并继续等待。未知或 malformed 组合以及不可读的 manager 结果会使 `waitForExit()` reject，而不是宣称完全停稳。`terminate()` 会唤醒正在休眠的 observer 立即复查，结算时会取消未胜出的退避 sleep。严格的同目录 `startup-error.json` 只承载 request／bootstrap 或 target pre-exec failure，parent 会在可观察生命周期完成时移除本次 spawn 的私有路径。
 
 普通 target result 仍来自同一个 child process。PTY 路径复用同一 request 与 bootstrap，但不增加常驻 runner，因此 `node-pty` PID、进程组、session leader、控制终端、前台 `inputWaiting`、`/dev/tty`、readiness 与 direct terminal outcome 保留既有含义，同时 scope membership 覆盖 `setsid` 与 reparent 后代。
 
@@ -54,7 +54,7 @@ selector 是 per-spawn locator 或 sentinel，不是凭据或持久格式。Linu
 
 ## Verification
 
-- provider 与 Linux 协议测试套件固定同步 NUL 拒绝发生在启动副作用之前、严格 request／error 解码、target cwd 与完整环境恢复、私有变量碰撞、保留 argv 且对 symlink 敏感的 PATH 遍历、为继承 stdio 清除 close-on-exec、pre-exec error ownership、失败深度 probe 重试与成功深度 probe 缓存及逐调用 manager 检查、三种 scope 建立状态（包括 request 未消费时的请求终止与意外退出）、`LoadState`／`ActiveState` 解析、`reloading`、带未胜出 delay 取消的 terminate wake-up、建立后有上限的退避，以及 PTY managed-owner 恰好一次 cleanup。
+- provider 与 Linux 协议测试套件固定同步 NUL 拒绝发生在启动副作用之前、严格 request／error 解码、target cwd 与完整环境恢复、私有变量碰撞、保留 argv 且对 symlink 敏感的 PATH 遍历、为继承 stdio 清除 close-on-exec、pre-exec error ownership、失败深度 probe 重试与成功深度 probe 缓存及逐调用 manager 检查、三种 scope 建立状态（包括 request 未消费时的请求终止与意外退出）、`LoadState`／`ActiveState`／`TasksCurrent` 解析、释放被留在 active 且没有任何进程的遗留 scope（连同 client 仍存活、未请求终止与进程数未上报三种情形）、`reloading`、带未胜出 delay 取消的 terminate wake-up、建立后有上限的退避，以及 PTY managed-owner 恰好一次 cleanup。
 - Windows 协议与 Win32 测试套件固定恰好两个 result 分支、只含数字的 target exit、使用普通 error 的 start cancellation 与 parent 原样保留的本地 reason、缩减到 `name`／`message`／`code`／`syscall`／`path` 的 error record、固定的 `2`／`3`／`267` 到 `ENOENT`、`740` 到 `EACCES`、`5` 到 `EPERM`、`193` 到 `EFTYPE` 及其余 code 到 `UNKNOWN` 的映射、runner spawn 后才发送 start、spawn 前 failure 的 empty-range settlement、按序数显式排序的 target 环境块及 `=C:` 保留和双 NUL 结尾、`uv_get_osfhandle()` carrier 映射与 unsigned invalid sentinel 拒绝、null-device ignored-stdin carrier 与非 ignore stdin pipe、result-send 与 IPC-disconnect failure、stdio settlement 前的 direct-result 锁存、active-process 完全停稳，以及唯一 handle cleanup。
 - 无需密钥的 [`bash-startup-timeout`](../../../../snapshots/session/bash-startup-timeout/snapshot.yml) Session 快照固定模型可见的超时结果。Linux user-systemd fixture 通过输入屏障保持启动请求未消费，并验证取消与 range settlement。
 - 真实 Linux user-systemd 测试会分别通过生产入口运行一条普通命令与一条 `node-pty` `setsid`／reparent 场景。它们证明 scope signalling 与 collection、裸可执行文件查找、逃逸后代终止、range settlement，以及不变的 PTY PID、session、控制终端、前台输入、`/dev/tty`、readiness 与 startup-failure 语义。

.agents/notes/implemented/architecture/2026-09-08-desktop-bundled-runtime-and-external-plugins.i18n.yaml

@@ -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/2026-09-08-desktop-bundled-runtime-and-external-plugins.md
+2026-09-08-desktop-bundled-runtime-and-external-plugins.md: 581a4ad31121bfda651cb99e550487a0a712e943
+2026-09-08-desktop-bundled-runtime-and-external-plugins.zh.md: fbbb25a4a3446d52ac56dc35bb176c0b706d7cb4

.agents/notes/implemented/architecture/2026-09-08-desktop-bundled-runtime-and-external-plugins.md

@@ -0,0 +1,61 @@
+# Agent Note: Bundle the Desktop runtime and retain external plugins
+
+Status: implemented
+
+English | [中文](2026-09-08-desktop-bundled-runtime-and-external-plugins.zh.md)
+
+Profile mutation and recovery follow the [in-place profile decision](2026-09-09-desktop-in-place-profile.md).
+
+## Problem
+
+Installing the core dependency graph during Desktop initialization repeats work already done by the release builder. An offline store eliminates downloads but retains extraction, package-manager startup, and installation costs. Users need the application to start with its production packages present while retaining ordinary npm plugin installation and plugin state across application upgrades.
+
+Separate package directories can load duplicate Cordis or service modules. Retaining plugin files also does not prove compatibility with a new host API or Node runtime.
+
+## Decision
+
+[Runtime preparation](../../../../apps/desktop/scripts/prepare-dsh.ts) materializes the production graph once at build time and ships it through `extraResources/dsh`. The Electron shell stays in ASAR. A bundled upstream Node process runs the private Desktop Host from resources and loads enabled plugins from `$DSH_HOME/profiles/desktop`.
+
+Desktop has not been released. This is its first installation format; there are no readers or migrations for the unpublished seed-based profile. This note supersedes core seed installation and single-project dependency ownership in the [Desktop packaging decision](2026-08-25-electron-desktop-packaging-and-updates.md). That note continues to own release identity, signing, portless transport, process ownership, and Electron-only plugin authorization. No existing note is fully superseded or archived.
+
+## Package ownership
+
+The resource descriptor records the exact release, Node version, platform, architecture, shared package versions, and final file hashes. The runtime tree contains ordinary files and directories, without links back to pnpm’s build store. Native Mach-O files are signed before hashing; the application signer preserves their bytes and checks the inventory after signing. An explicit `dsh/node_modules` resource mapping bypasses electron-builder’s root `node_modules` exclusion, and the copied tree is verified before any signing or notarization.
+
+The [Desktop file policy](../../../../apps/desktop/scripts/runtime-file-policy.ts) applies after production npm installation and before native signing or descriptor generation. npm publication lists serve library consumers and can include declarations, maps, tests, and native build inputs; they do not identify the files needed by the Desktop process. The Desktop copy omits declarations and recognized source maps because Host execution uses JavaScript and generated Typert artifacts, clears inherited `NODE_OPTIONS`, and does not enable source mapping. Reviewed plugin lifecycle builds cover native dependencies, not arbitrary TypeScript compilation. Published npm packages and external plugin directories retain their own files. Source debugger navigation is a development-package capability.
+
+Package-specific exclusions remove Domino tests, fs-ext compilation outputs, Koffi's Windows import library, and non-target node-pty prebuilds and debug symbols. The policy retains native executable dependencies, node-pty's ConPTY source distribution, licenses, and unrecognized assets; broad `src`, `test`, `.ts`, or `.map` exclusions could remove executable code or runtime data. Copy tests preserve sentinel assets and seal the filtered inventory; the bundled-Node [payload smoke](../../../../apps/desktop/tests/fixtures/runtime-payload-smoke.mjs) verifies PTY output, native file seeking, FFI, image conversion, and HTML parsing. Runtime preparation still verifies every retained byte and boots the complete Host with an external plugin.
+
+Every first-party package in the dsh and private Host production closures is shared. The profile contains directory symlinks to those resource packages, or junctions on Windows. Links resolve to real host package directories under normal Node resolution. Host and plugin imports therefore share the same module instance for each resolved export. Distinct ESM and CommonJS conditional exports remain distinct entry points; a link cannot merge a package’s dual implementations.
+
+External plugins declare shared host packages as peers. Ordinary dependencies remain plugin-owned and may differ from the versions used by dsh. Validation rejects incompatible enabled peers, nested or aliased copies of shared packages, private package links, and dependency resolution through CLI or other ancestor directories. A third-party package requiring host-wide instance identity must be explicitly added to the runtime’s shared inventory; matching version numbers alone are insufficient.
+
+The profile manifest records exact installed plugin dependencies separately from its enabled bundle list. Disabling a plugin preserves its package, lockfile entry, and user configuration. The shared links are Desktop-owned derived state, recorded separately from pnpm; package-manager operations run without those links, then Desktop recreates and validates them.
+
+## Transactions and upgrades
+
+First launch creates profile metadata and host links without running pnpm, preserving unrelated files. Compatible release changes or application relocation refresh links and validate enabled peers in place. Node version, platform, or architecture changes reinstall the locked plugin graph and run approved native builds.
+
+Native canonical paths identify shared package directories. Windows launchers can vary path casing without moving the application; string equality would trigger unnecessary profile preparation. Profile cleanup explicitly unlinks every nested directory link before removing real directories. A Windows fixture under Electron 44 reproduces recursive `fs.rmSync` deleting files through a nested junction, while bundled upstream Node 24.17 preserves them. Cleanup qualification therefore includes the real Electron runtime; Node-only tests do not establish target preservation.
+
+Dependency mutations install with scripts disabled, validate the plugin graph and host links, run the reviewed pending lifecycle builds, and validate again. This permits approved native dependencies to resolve host peers while preventing accidental duplicate host packages from reaching startup. The `allowBuilds` policy remains explicit; unsupported build-requiring dependencies fail the transaction.
+
+Desktop stops the Host before package mutations and waits for pnpm exit before restarting it. The [in-place decision](2026-09-09-desktop-in-place-profile.md) owns partial failures and persistent retry state. Recorded host links identify owned directories independently of package-operation completion.
+
+The [immediate-window decision](2026-09-09-desktop-immediate-window-and-direct-start.md) owns direct Host startup and recovery in the main window. Users can update, remove, disable, or re-enable plugins and retry startup. Incompatible plugins are not silently deleted or automatically downgraded. Each backend launch requires the current runtime identity.
+
+## Alternatives considered
+
+Full runtime verification belongs to packaging. Startup reads the descriptor, checks shared package records and required Host entries, and uses the recorded runtime identity for profile reuse. The [release-validation decision](2026-09-09-desktop-build-release-validation.md) assigns release and target compatibility checks to packaging. It neither enumerates nor hashes installed runtime files, including on first launch or after an upgrade. Reading every file before backend loading adds startup I/O proportional to the distribution size. Installed content changes therefore are not detected by a startup checksum comparison; unusable modules fail when loaded. Build-time verification still rejects changed, missing, extra, or linked files against the recorded inventory.
+
+- **Install the bundled offline seed at startup.** This preserves an ordinary pnpm installation procedure but repeats core extraction and installation on every affected machine. Materialized resources remove that work at the cost of more application files and release-builder responsibility.
+- **Link all host dependencies into plugins.** This unnecessarily couples ordinary plugin dependencies to the host. Only the explicit shared inventory is linked; private packages retain independent versions.
+- **Use hardlinks.** They cannot represent directories, may not cross volumes, share writable bytes, and retain old inodes after application replacement. Directory symlinks and Windows junctions express the intended package target.
+- **Use `NODE_PATH` or preserve symlink paths.** These do not provide uniform ESM resolution or shared module identity. Normal package lookup through explicit links is directly testable.
+- **Keep core packages in ASAR.** The backend uses upstream Node rather than Electron’s patched filesystem. Ordinary `extraResources` also preserves native loading and subprocess paths.
+
+## Consequences
+
+Core package installation is absent from first launch and compatible upgrades. Metadata checks and backend loading still cost startup time; no release latency or download-size improvement is claimed without measurement. Plugin preservation is conditional on host API and native runtime compatibility, with a visible recovery path when that condition fails.
+
+The [Desktop README](../../../../apps/desktop/README.md) owns operational guidance. Focused tests cover real pnpm installation and approved builds, shared ESM instance identity, private dependency versions, relocation, disabled plugins, native rebuild selection, activation failures, and transaction locking. Signed installed-artifact upgrades, macOS notarization, Windows junction/native behavior, release size and startup benchmarks, and real-model GUI recordings remain release-environment qualification requirements; unit fixtures do not substitute for them.

.agents/notes/implemented/architecture/2026-09-08-desktop-bundled-runtime-and-external-plugins.zh.md

@@ -0,0 +1,61 @@
+# Agent Note: 内置 Desktop 运行时并保留外部插件
+
+Status: implemented
+
+[English](2026-09-08-desktop-bundled-runtime-and-external-plugins.md) | 中文
+
+profile 修改与恢复遵循[直接修改 profile 决策](2026-09-09-desktop-in-place-profile.zh.md)。
+
+## 问题
+
+Desktop 初始化时安装核心依赖图，会重复发布构建器已经完成的工作。离线 store 消除了下载，但仍有解压、包管理器启动和安装成本。用户需要应用在生产依赖已就绪时启动，同时保留普通 npm 插件安装能力，以及跨应用升级的插件状态。
+
+分离的包目录可能加载重复的 Cordis 或服务模块。仅保留插件文件也不能证明它与新的宿主 API 或 Node 运行时兼容。
+
+## 决策
+
+[运行时准备](../../../../apps/desktop/scripts/prepare-dsh.ts)在构建时物化一次生产依赖图，并通过 `extraResources/dsh` 分发。Electron 壳保留在 ASAR 中。内置上游 Node 进程从资源启动私有 Desktop Host，并从 `$DSH_HOME/profiles/desktop` 加载已启用插件。
+
+Desktop 尚未发布。这是它的第一种安装格式；不提供未发布 seed profile 的读取器或迁移。本记录取代 [Desktop 打包决策](2026-08-25-electron-desktop-packaging-and-updates.zh.md)中的核心 seed 安装和单项目依赖归属部分。该记录继续负责发布身份、签名、无端口传输、进程归属和仅限 Electron 的插件授权。没有现有记录被完全取代或归档。
+
+## 包归属
+
+资源描述文件记录精确发布版本、Node 版本、平台、架构、共享包版本和最终文件哈希。运行时树包含普通文件和目录，不包含指回 pnpm 构建 store 的链接。原生 Mach-O 文件先签名再哈希；应用签名器保留其字节，并在签名后检查清单。明确的 `dsh/node_modules` 资源映射绕过 electron-builder 对根 `node_modules` 的排除，并在任何签名或公证前验证复制后的依赖树。
+
+[桌面文件规则](../../../../apps/desktop/scripts/runtime-file-policy.ts)在生产 npm 依赖安装之后、原生签名或描述文件生成之前执行。npm 发布列表服务于库的使用者，可以包含声明、map、测试和原生构建输入，不能直接表示桌面进程需要哪些文件。桌面副本排除声明和已识别的 source map，因为 Host 执行 JavaScript 和生成的 Typert 产物，清除继承的 `NODE_OPTIONS`，且不开启源码映射。经过审核的插件生命周期构建面向原生依赖，不执行任意 TypeScript 编译。已发布的 npm 包和外部插件目录保留各自的文件。源码调试导航由开发包提供。
+
+包专用排除项包括 Domino 测试、fs-ext 编译产物、Koffi 的 Windows 导入库，以及非目标平台的 node-pty 预构建文件和调试符号。规则保留原生可执行依赖、node-pty 的 ConPTY 源分发内容、许可证和未知资源；宽泛排除 `src`、`test`、`.ts` 或 `.map` 可能移除可执行代码或运行时数据。复制测试保留哨兵资源并封存过滤后的清单；内置 Node 的[产物 smoke](../../../../apps/desktop/tests/fixtures/runtime-payload-smoke.mjs)验证 PTY 输出、原生文件定位、FFI、图像转换和 HTML 解析。运行时准备仍会验证每个保留字节，并携带外部插件启动完整 Host。
+
+dsh 与私有 Host 生产闭包中的每个第一方包都共享。profile 包含指向这些资源包的目录软链接，在 Windows 上使用 junction。正常 Node 解析会把链接解析到实际宿主包目录。因此，宿主与插件对每个已解析导出的导入共享同一模块实例。不同的 ESM 与 CommonJS 条件导出仍是不同入口；链接不能合并包的两套实现。
+
+外部插件把共享宿主包声明为 peer。普通依赖由插件拥有，可以不同于 dsh 使用的版本。验证拒绝已启用插件的不兼容 peer、共享包的嵌套或别名副本、私有包链接，以及通过 CLI 或其他祖先目录解析依赖。如果第三方包需要宿主范围的实例身份，必须明确加入运行时共享清单；版本号相同并不足够。
+
+profile manifest 分别记录精确的已安装插件依赖和已启用 bundle 列表。停用插件会保留其包、锁文件条目和用户配置。共享链接是 Desktop 拥有的派生状态，独立于 pnpm 记录；包管理器操作不携带这些链接，随后 Desktop 重建并验证它们。
+
+## 事务与升级
+
+首次启动创建 profile 元数据和宿主链接，不运行 pnpm，并保留无关文件。兼容的发布变化或应用移动会直接刷新链接并验证已启用的 peer。Node 版本、平台或架构变化时，会重新安装锁定的插件依赖图并运行获准的原生构建。
+
+共享包目录使用原生规范路径识别。Windows 启动器可能改变路径大小写而不移动应用；字符串相等判断会触发不必要的 profile 准备。profile 清理在移除真实目录前，显式解除每一个嵌套目录链接。Windows 夹具在 Electron 44 下复现了递归 `fs.rmSync` 沿嵌套 junction 删除目标文件，而内置上游 Node 24.17 会保留它们。因此清理验收包含真实 Electron 运行时；仅在 Node 下测试不能证明目标文件会保留。
+
+依赖修改先禁用脚本安装，验证插件依赖图和宿主链接，运行经过审查的待执行生命周期构建，再次验证。这允许已批准的原生依赖解析宿主 peer，同时阻止意外的重复宿主包进入启动过程。`allowBuilds` 策略保持明确；不受支持且需要构建的依赖会使事务失败。
+
+Desktop 在包修改前停止 Host，并等待 pnpm 退出后再重启它。[直接修改决策](2026-09-09-desktop-in-place-profile.zh.md)规定部分失败和持久重试状态的处理方式。记录的宿主链接用于识别自有目录，与包操作是否完成相互独立。
+
+[立即显示窗口决策](2026-09-09-desktop-immediate-window-and-direct-start.zh.md)规定实际 Host 启动和主窗口恢复。用户可以更新、删除、禁用或重新启用插件并重试启动。不兼容插件不会被静默删除或自动降级。每次后端启动都要求当前运行时标识。
+
+## 考虑过的替代方案
+
+完整运行时验证属于打包流程。启动读取描述文件，检查共享包记录和必要的 Host 入口，并使用记录的运行时身份复用 profile。[发布验证决策](2026-09-09-desktop-build-release-validation.zh.md)把发布与目标兼容性检查交给打包流程。首次启动和升级后启动都不枚举已安装运行时文件或计算其哈希。在后端加载前读取每个文件，会增加与分发体积成正比的启动 I/O。因此，启动不会通过校验和比较检测已安装内容的变化；不可用模块在加载时失败。构建时验证仍按记录的清单拒绝内容变化、缺失、多余或链接文件。
+
+- **启动时安装内置离线 seed。** 这保留普通 pnpm 安装流程，但会在每台受影响机器上重复核心解压与安装。物化资源消除了这部分工作，代价是更多应用文件和发布构建器责任。
+- **把所有宿主依赖链接给插件。** 这会让普通插件依赖与宿主产生不必要的耦合。只链接明确的共享清单；私有包保留独立版本。
+- **使用硬链接。** 它不能表示目录，可能无法跨卷，共享可写字节，并在应用替换后保留旧 inode。目录软链接和 Windows junction 能表达预期的包目标。
+- **使用 `NODE_PATH` 或保留软链接路径。** 它们不能提供统一的 ESM 解析或共享模块身份。通过明确链接进行正常包查找可以直接测试。
+- **把核心包留在 ASAR。** 后端使用上游 Node，而不是 Electron 修改过的文件系统。普通 `extraResources` 也能保留原生加载和子进程路径。
+
+## 影响
+
+首次启动和兼容升级不安装核心包。元数据检查和后端加载仍需要启动时间；没有测量前，不声称发布启动延迟或下载体积改善。插件保留以宿主 API 和原生运行时兼容为条件，条件不满足时提供可见的恢复入口。
+
+[Desktop README](../../../../apps/desktop/README.zh.md)负责操作说明。定向测试覆盖真实 pnpm 安装与已批准构建、共享 ESM 实例身份、私有依赖版本、应用移动、停用插件、原生重建选择、激活失败和事务锁。签名安装产物升级、macOS 公证、Windows junction 与原生行为、发布体积与启动基准，以及真实模型 GUI 录制仍是发布环境验收要求；单元夹具不能替代这些验证。

.agents/notes/implemented/architecture/2026-09-09-consumer-owned-startup-strictness.i18n.yaml

@@ -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/2026-09-09-consumer-owned-startup-strictness.md
+2026-09-09-consumer-owned-startup-strictness.md: e009e66ead25ef0a5e6001d33663e32bc04d19d2
+2026-09-09-consumer-owned-startup-strictness.zh.md: 58c364056f5b0dc41e018cd5488be983662401d4

.agents/notes/implemented/architecture/2026-09-09-consumer-owned-startup-strictness.md

@@ -0,0 +1,39 @@
+# Agent Note: Consumer-owned startup strictness
+
+Status: implemented
+
+English | [中文](2026-09-09-consumer-owned-startup-strictness.zh.md)
+
+## Problem
+
+Best-effort Loader reconciliation preserves usable plugins, but applications still need a minimum set of capabilities. An HTTP application without its listening server is not running, while an unavailable tool can be omitted without making the remaining application unusable. Cordis cannot infer this distinction from plugin implementation or dependency state.
+
+## Decision
+
+DSH owns startup strictness outside vendored Cordis. App-boot audits the settled initial tree against one global list of stable entry ids. A listed entry that is present, enabled, and not active rejects startup and disposes the application. A listed id that is absent or disabled has no effect. The bootstrap Include is required by entry identity because a missing or invalid root configuration prevents application assembly. Other inactive entries produce one warning and leave successful siblings running.
+
+The required ids are `agent-loop`, `webserver`, `modules`, `connection`, `headless-runner`, `acp`, and `sdk-jsonrpc-server`. They represent shared Agent execution, application endpoints, and Web bootstrap/transport. Web needs its client module registry and authenticated connection even when the HTTP server can listen without them. Providers already required through injection need no separate entry: their absence leaves a listed consumer pending or failed.
+
+The audit treats a throwing `disabled` expression as an entry failure, not a disabled entry, because evaluation never established whether to skip it. The same optional/required policy applies to that failure.
+
+The audit runs only during initial application boot. Later config HMR remains best effort and keeps the failed candidate visible for repair.
+
+This policy governs [Web host boot](2026-07-24-web-config-tree-boot-and-transport-layering.md), including its [client plugin roster](2026-07-23-client-plugin-loading-model.md). [Per-session presets](2026-08-03-per-session-agent-presets.md) own a separate strict subtree audit.
+
+## Alternatives considered
+
+- **Add transactional and best-effort modes to vendored Loader.** Rejected because strictness belongs to the application or resource owner, while a Loader group contains unrelated plugins. A mode would also expand the vendor patch and leave callers to select a policy at every group.
+- **Declare required entries in each profile.** Rejected because the same application endpoints would be duplicated across profile data and custom profiles. A global list treats missing ids as irrelevant while keeping stable shipped ids authoritative.
+- **Make every startup failure optional.** Rejected because a process that cannot expose its selected application endpoint must report launch failure.
+
+## Consequences
+
+Stable required entry ids are part of application assembly. Renaming one requires updating the list and its tests. Optional plugin failures remain visible in Loader state and stderr without tearing down active siblings. Required failures use the same detailed import, activation, or pending-service diagnostic before app-boot disposes the root.
+
+## Testing
+
+App-boot unit tests cover absent and disabled required ids, optional import failure, config evaluation failure, synchronous and asynchronous `apply()` failure, pending dependencies, and required failure teardown. The built Web-profile acceptance serves the full UI with optional failures and exits nonzero without readiness when the required HTTP port is occupied or `modules` or `connection` cannot activate.
+
+The [Web process matrix](../../../../apps/cli/tests/profiles/web/tests/web-failure-matrix.expected.e2e.ts) independently exercises optional and required failures at startup and after native patch-file edits. Authenticated HTTP requests and plugin lifecycle files distinguish a usable application from a surviving process. These keyless process checks complement the [controlled-delivery unit tests](../testing/2026-09-09-user-patch-hmr-test-delivery.md): unit tests isolate reconciliation failures, while the process tests also require the shipped launcher, native watcher, and bounded shutdown to work together.
+
+The matrix enables Chokidar's `awaitWriteFinish` to acknowledge stable file contents before each reload; otherwise its short change-event suppression window can discard the next test edit. Native events remain required, and assertions wait for observed activation or failure rather than a fixed settling sleep. This is explicit test configuration, not evidence for the default watcher timing.

.agents/notes/implemented/architecture/2026-09-09-consumer-owned-startup-strictness.zh.md

@@ -0,0 +1,39 @@
+# Agent Note：由 consumer 持有启动严格语义
+
+Status: implemented
+
+[English](2026-09-09-consumer-owned-startup-strictness.md) | 中文
+
+## 问题
+
+Best-effort Loader reconcile 会保留可用 plugin，但应用仍需一组最小 capability。HTTP 应用没有 listening server 就不算运行，而一个 tool 不可用时可以仅省略该 tool，剩余应用仍然可用。Cordis 无法从 plugin 实现或依赖状态推断这一区别。
+
+## 决策
+
+DSH 在 vendored Cordis 之外持有启动严格语义。App-boot 用一份全局稳定 entry id list 审计已结算的初始 tree。List 中存在、启用且未 active 的 entry 会使启动 reject，并拆卸应用。List 中缺失或禁用的 id 不产生影响。Bootstrap Include 按 entry 身份被视为 required，因为根配置缺失或无效会阻止应用组装。其他 inactive entry 输出一次 warning，并让成功 sibling 继续运行。
+
+Required id 为 `agent-loop`、`webserver`、`modules`、`connection`、`headless-runner`、`acp` 和 `sdk-jsonrpc-server`。它们分别代表共享 Agent 执行、应用 endpoint，以及 Web 启动与传输。即使 HTTP server 不依赖它们也能监听，Web 仍需要客户端模块注册表和经过认证的连接。通过注入已成为必需项的 provider 不需要单列：它们缺失时，已列出的消费方会保持 pending 或失败。
+
+审计将 `disabled` 表达式抛出的异常视为 entry 失败，而不是 entry 已禁用，因为求值未能确定是否跳过它。该失败遵循相同的 optional/required 策略。
+
+该审计只在应用首次启动时运行。之后的 config HMR 仍采用 best effort，并保留 failed candidate 供后续修复。
+
+该策略适用于 [Web host 启动](2026-07-24-web-config-tree-boot-and-transport-layering.zh.md)，包括其 [client 插件名册](2026-07-23-client-plugin-loading-model.zh.md)。[按会话的 preset](2026-08-03-per-session-agent-presets.zh.md)持有独立的严格子树审计。
+
+## 考虑过的替代方案
+
+- **给 vendored Loader 增加 transactional 与 best-effort mode。** 拒绝，因为严格语义属于应用或资源 owner，而一个 Loader group 包含互不相关的 plugin。Mode 还会扩大 vendor patch，并要求 caller 为每个 group 选择 policy。
+- **在每个 profile 中声明 required entry。** 拒绝，因为相同应用 endpoint 会在 profile data 与 custom profile 中重复。全局 list 会忽略缺失 id，同时让稳定的随附 id 保持权威。
+- **把所有启动失败都视为 optional。** 拒绝，因为无法暴露所选应用 endpoint 的进程必须报告启动失败。
+
+## 后果
+
+稳定的 required entry id 是应用 assembly 的一部分。重命名时必须同步更新 list 与测试。Optional plugin failure 会保留在 Loader state 和 stderr 中，但不会拆卸 active sibling。Required failure 使用相同的详细 import、activation 或 pending-service 诊断，然后由 app-boot 拆卸 root。
+
+## 测试
+
+App-boot 单元测试覆盖缺失和禁用的 required id、optional import failure、config evaluation failure、同步和异步 `apply()` failure、pending dependency，以及 required failure teardown。构建后的 Web-profile acceptance 会在 optional failure 存在时继续提供完整 UI，并在 required HTTP port 被占用或 `modules`、`connection` 无法激活时以非零码退出，且不报告就绪。
+
+[Web 进程矩阵](../../../../apps/cli/tests/profiles/web/tests/web-failure-matrix.expected.e2e.ts)分别验证启动时和原生补丁文件修改后的 optional 与 required 失败。经过认证的 HTTP 请求和插件生命周期文件区分可用应用与仅存活的进程。这些无需密钥的进程检查与[受控事件投递单元测试](../testing/2026-09-09-user-patch-hmr-test-delivery.zh.md)互补：单元测试隔离配置协调失败，进程测试还要求随附启动器、原生监听器和有界关闭流程协同工作。
+
+矩阵启用 Chokidar 的 `awaitWriteFinish`，在每次重载前确认文件内容已稳定；否则它的短暂 change 事件抑制窗口可能丢弃下一次测试编辑。测试仍然依赖原生事件，并等待观察到激活或失败，而不是固定时长的休眠。这是显式测试配置，不能证明默认监听器的时序行为。

.agents/notes/implemented/architecture/2026-09-09-deprecate-synchronous-session-event-reads.i18n.yaml

@@ -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/2026-09-09-deprecate-synchronous-session-event-reads.md
+2026-09-09-deprecate-synchronous-session-event-reads.md: a2a86b0d6500269531ca1088738aaa612076bfb4
+2026-09-09-deprecate-synchronous-session-event-reads.zh.md: 38468708a5323b6ebf94377f3cc8f2c86dc95dde

.agents/notes/implemented/architecture/2026-09-09-deprecate-synchronous-session-event-reads.md

@@ -0,0 +1,51 @@
+# Agent Note: Deprecate synchronous reads of arbitrary Session events
+
+Status: implemented
+
+English | [中文](2026-09-09-deprecate-synchronous-session-event-reads.zh.md)
+
+## Problem
+
+Synchronous access to arbitrary event positions makes consumers depend on the complete Session event sequence being immediately available in memory. The storage direction is to stop retaining that complete sequence in memory. Once historical events require storage I/O, the runtime cannot preserve the same synchronous read guarantee without retaining the history or blocking on storage.
+
+New callers increase that dependency even when they read only one old event. Repeated history scans after resume also make ordinary domain logic depend on historical storage instead of the state it actually needs.
+
+## Decision
+
+All operations that synchronously read arbitrary positions or ranges of Session event history are deprecated, including `Session.eventAt()`, `Session.snapshotEvents()`, and `Session.ownEvents()`. Existing logic may remain unmigrated for now, but new calls are prohibited. New aliases or wrappers that expose the same synchronous historical access are prohibited as well.
+
+The three methods carry this rule in `@deprecated` JSDoc. This is an API-use decision; the current Session implementation still retains the complete event sequence in memory.
+
+Repository test files, including `scripts/**/*.spec.{ts,tsx}`, may call these three readers to inspect emitted events and exercise Session history behavior. The test-file lint override allows `snapshotEvents`, `eventAt`, and `ownEvents`; all other deprecated names remain errors. This allowance also covers unrelated declarations with the same three names under the current linter. It does not apply to production source or non-test repository scripts.
+
+### State needed after resume
+
+Design durable event fields and Session projections together so each domain can reconstruct the state its consumers need. Restore that state during resume, then maintain it incrementally from newly committed events. After resume, ordinary logic reads the projection or processes the delivered current event instead of looking back through historical events. Reading already-maintained projection state synchronously does not require arbitrary access to the event log.
+
+Historical content presented on demand uses explicit asynchronous pagination and progressive loading, with each read limited to the requested window. Loading the complete sequence behind a synchronous helper preserves the dependency this decision removes.
+
+### Operations that require complete history
+
+Fork and a small number of operations may genuinely need a complete historical sequence or inherited prefix. Their need for those records remains valid and requires an explicit storage read. It does not require the whole sequence to remain resident or grant an exception for new calls to deprecated synchronous readers. Each such consumer must establish why its result requires the complete sequence rather than projected state or a limited historical window.
+
+## Alternatives considered
+
+**Keep synchronous single-event reads while deprecating only full snapshots.** A single requested event may also be absent from memory. Restricting the result size does not remove the storage dependency, and helpers such as `ownEvents()` retain the same assumption for a suffix.
+
+**Keep the complete sequence resident to preserve the read APIs.** This lets consumer convenience dictate Session memory retention and prevents the intended storage design. Projections preserve required state, while explicit historical reads preserve access to the records themselves.
+
+**Require every existing caller to migrate immediately.** Existing logic may defer migration under this decision. Preventing new dependencies bounds the remaining work without making every existing consumer part of the same change.
+
+**Disable deprecation lint throughout test files or exempt production readers by name.** Tests only need the three reader names; other deprecated APIs must remain errors. Production code retains the prohibition on new synchronous reads.
+
+**Add a separate test-only reader API.** The three existing readers already expose the observations these tests need. Wrapping them adds an API and production-import checks without changing those observations.
+
+## Consequences
+
+New domain behavior must make its event data and projected state sufficient for resumed execution. User-requested history may still load progressively, and genuine full-history operations still have a storage path to design. This decision does not claim that resume or fork already avoids loading the complete log.
+
+Calls outside test files carry line-scoped `typescript/no-deprecated` waivers that identify deferred migration or delegation between deprecated readers. Remove a waiver when its deprecated call is removed; copying a waiver to a new production call violates this policy. The executable lint check accepts test reads and existing waived reads, rejects unwaived production reads, and rejects unrelated deprecated APIs in tests. Documentation checks verify the source-equivalent API declarations and bilingual records.
+
+## Related decisions
+
+The [session immutability decision](2026-06-11-dev-invariants-over-deep-readonly.md) continues to own event ownership and freezing; this decision supersedes its recommendation to use synchronous historical readers. The [required projection reader decision](2026-08-19-session-projection-mandatory-seam.md) continues to own missing-state failures and typed state reads. The [recallable compaction proposal](../../proposed/feature/2026-07-06-recallable-compaction.md) retains its recall design while replacing its proposed synchronous history access with paged reads.

.agents/notes/implemented/architecture/2026-09-09-deprecate-synchronous-session-event-reads.zh.md

@@ -0,0 +1,51 @@
+# Agent Note: 弃用对会话任意位置事件的同步读取
+
+Status: implemented
+
+[English](2026-09-09-deprecate-synchronous-session-event-reads.md) | 中文
+
+## 问题
+
+同步访问任意事件位置，使消费方依赖完整会话事件序列在内存中随时可用。存储的演进方向是不再将完整序列保存在内存中。历史事件一旦需要存储 I/O，运行时若不继续保留历史或阻塞等待存储，就无法维持相同的同步读取保证。
+
+即使只读取一个旧事件，新增调用也会加深这种依赖。恢复后反复扫描历史，还会使普通领域逻辑依赖历史存储，而不是它真正需要的状态。
+
+## 决策
+
+弃用所有同步读取会话事件历史任意位置或区间的操作，包括 `Session.eventAt()`、`Session.snapshotEvents()` 和 `Session.ownEvents()`。现有逻辑可以暂不迁移，但禁止新增调用。同样禁止新增暴露相同同步历史访问能力的别名或包装层。
+
+这三个方法通过 `@deprecated` JSDoc 声明该规则。这是 API 使用决策；当前 Session 实现仍在内存中保留完整事件序列。
+
+仓库测试文件（包括 `scripts/**/*.spec.{ts,tsx}`）可以调用这三个读取方法，以检查已发出的事件并验证 Session 历史行为。测试文件 lint override 允许 `snapshotEvents`、`eventAt` 和 `ownEvents`，其他弃用名称仍报错。在当前 linter 下，此豁免也覆盖使用这三个名称的其他声明。它不适用于生产源码或非测试仓库脚本。
+
+### 恢复后需要的状态
+
+结合设计持久事件字段与会话投影，使每个领域都能重建其消费方所需的状态。在恢复期间还原这些状态，随后通过新提交的事件增量维护。恢复后，普通逻辑读取投影或处理当前收到的事件，而不是回头查找历史事件。同步读取已经维护好的投影状态，不需要任意访问事件日志。
+
+按需展示的历史内容采用显式异步分页和渐进式加载，每次读取限制在请求的窗口内。在同步辅助方法背后加载完整序列，仍保留了本决策要消除的依赖。
+
+### 确实需要完整历史的操作
+
+fork 等少数操作可能确实需要完整历史序列或继承前缀。它们对这些记录的需求仍然成立，需要通过显式存储读取来满足。这不要求完整序列常驻内存，也不构成新增已弃用同步读取调用的例外。每个此类消费方都必须说明，为什么其结果需要完整序列，而不能只使用投影状态或局部历史窗口。
+
+## 曾考虑的替代方案
+
+**只弃用完整快照，保留同步单事件读取。** 单个被请求的事件也可能不在内存中。限制返回结果大小并不能消除存储依赖，`ownEvents()` 等辅助方法对后缀序列也保留着相同假设。
+
+**为保留读取 API 而让完整序列常驻内存。** 这会让消费方的便利性决定会话的内存保留方式，阻碍预期的存储设计。投影保留所需状态，显式历史读取则保留对记录本身的访问能力。
+
+**要求立即迁移所有现有调用方。** 本决策允许现有逻辑推迟迁移。禁止新增依赖可以限制剩余工作规模，无需将每个现有消费方都纳入同一次修改。
+
+**在整个测试文件中禁用弃用 lint，或按名称豁免生产代码中的读取方法。** 测试只需要这三个读取方法名；其他弃用 API 必须继续报错。生产代码仍禁止新增同步读取。
+
+**添加单独的测试专用读取 API。** 三个现有读取方法已经提供这些测试所需的观察结果。包装它们会增加一个 API 及生产代码导入检查，却不会改变这些观察结果。
+
+## 后果
+
+新增领域行为必须使其事件数据与投影状态足以支持恢复后的执行。用户请求的历史仍可渐进式加载，确实需要完整历史的操作仍需设计存储读取路径。本决策不表示恢复或 fork 已经能够避免加载完整日志。
+
+测试文件之外的调用带有逐行的 `typescript/no-deprecated` 豁免，说明暂缓迁移或已弃用读取方法之间的委托。删除已弃用调用时应一并删除其豁免；将豁免复制到新增生产调用违反本策略。实际执行 lint 的检查允许测试读取和已有豁免调用，拒绝未豁免的生产读取，并拒绝测试中其他已弃用 API。文档检查校验与源码一致的 API 声明及双语记录。
+
+## 相关决策
+
+[会话不可变性决策](2026-06-11-dev-invariants-over-deep-readonly.zh.md)继续负责事件所有权与冻结；本决策取代其中对同步历史读取方法的使用建议。[投影读取必需性决策](2026-08-19-session-projection-mandatory-seam.zh.md)继续负责缺失状态时的失败规则与类型化状态读取。[可回溯压缩提案](../../proposed/feature/2026-07-06-recallable-compaction.zh.md)保留其回溯设计，并将其中拟议的同步历史访问替换为分页读取。

.agents/notes/implemented/architecture/2026-09-09-desktop-build-release-validation.i18n.yaml

@@ -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/2026-09-09-desktop-build-release-validation.md
+2026-09-09-desktop-build-release-validation.md: 23aeeaa98d2647f8b9f54e94ea2aedaa82b19e55
+2026-09-09-desktop-build-release-validation.zh.md: 266bca5bbd593907c9a8fadd43050113ede94100

.agents/notes/implemented/architecture/2026-09-09-desktop-build-release-validation.md

@@ -0,0 +1,23 @@
+# Agent Note: Verify Desktop release compatibility during packaging
+
+Status: implemented
+
+English | [中文](2026-09-09-desktop-build-release-validation.zh.md)
+
+## Problem
+
+The shell and runtime descriptor ship together. Comparing their release facts on every launch repeats packaging checks without proving that installed executable bytes match the descriptor.
+
+## Decision
+
+The packaging verifier owns descriptor schema, shell version, platform, architecture, declared Host protocol version, and Node/pnpm semver validation. Startup reads the fields needed for profile preparation and retains shared-package record and Host-entry checks. The actual Host ready message still validates its protocol version.
+
+This partially supersedes startup release compatibility checks in the [bundled-runtime decision](2026-09-08-desktop-bundled-runtime-and-external-plugins.md). That note retains package ownership and distribution rationale.
+
+## Alternatives considered
+
+Repeating descriptor comparisons can diagnose a mixed installation earlier, but cannot establish executable integrity. Reintroducing them requires a concrete installation failure that packaging validation and actual Host diagnostics cannot adequately explain.
+
+## Consequences
+
+Startup does not reject a descriptor solely because its declared release schema, target, or Host protocol differs, or its Node/pnpm version strings are not semver. Packaging still rejects these cases and shell-version mismatch. Tests distinguish startup reads from packaging verification.

.agents/notes/implemented/architecture/2026-09-09-desktop-build-release-validation.zh.md

@@ -0,0 +1,23 @@
+# Agent Note: 在打包时验证 Desktop 发布兼容性
+
+Status: implemented
+
+[English](2026-09-09-desktop-build-release-validation.md) | 中文
+
+## 问题
+
+壳与运行时描述文件一起发布。每次启动比较其中的发布信息会重复打包检查，却不能证明已安装的可执行文件字节与描述文件一致。
+
+## 决策
+
+打包验证器负责描述文件 schema、shell 版本、平台、架构、声明的 Host 协议版本，以及 Node/pnpm semver 验证。启动读取准备 profile 所需的字段，并保留共享包记录和 Host 入口检查。实际 Host ready 消息仍然验证其协议版本。
+
+本决策部分取代[内置运行时决策](2026-09-08-desktop-bundled-runtime-and-external-plugins.zh.md)中的启动发布兼容性检查。该记录继续负责包归属与分发的理由。
+
+## 考虑过的替代方案
+
+重复比较描述文件能更早诊断混装，但不能证明可执行文件完整性。重新引入这些比较需要具体的安装故障，且打包验证和实际 Host 诊断无法充分解释该故障。
+
+## 后果
+
+启动不会仅因描述文件声明的发布 schema、目标或 Host 协议不同，或 Node/pnpm 版本字符串不是 semver 而拒绝运行。打包仍拒绝这些情况和 shell 版本不匹配。测试区分启动读取与打包验证。

.agents/notes/implemented/architecture/2026-09-09-desktop-immediate-window-and-direct-start.i18n.yaml

@@ -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/2026-09-09-desktop-immediate-window-and-direct-start.md
+2026-09-09-desktop-immediate-window-and-direct-start.md: 1c2209c044495de1fb9d3410e5ac777759e8fa6b
+2026-09-09-desktop-immediate-window-and-direct-start.zh.md: 49b08e4eeea2703c9bcf868d7b290d56eba79542

.agents/notes/implemented/architecture/2026-09-09-desktop-immediate-window-and-direct-start.md

@@ -0,0 +1,33 @@
+# Agent Note: Show the Desktop window before starting the Host
+
+Status: implemented
+
+English | [中文](2026-09-09-desktop-immediate-window-and-direct-start.zh.md)
+
+Profile mutation and recovery follow the [in-place profile decision](2026-09-09-desktop-in-place-profile.md).
+
+## Problem
+
+Waiting for backend readiness leaves users without a window during profile preparation and module loading. A complete staged health-check process repeats backend startup before the application starts its serving process, while plugin startup can still fail in the serving process.
+
+## Decision
+
+Electron creates the main window with a local loading page before profile reconciliation or Host startup. The page depends only on packaged shell assets and receives starting, ready, or error state through the owned preload. Readiness loads the product UI in that window; startup failures display diagnostics and available recovery actions. Closing during loading cancels further startup work and waits for the pending child to exit.
+
+The main window owns recovery because the failed Host cannot supply its own controls. Error pages retain diagnostics, restart, and reinstallation guidance. Disabling plugins and resetting Desktop are available only in a packaged application with loaded runtime metadata and available resources. Reset removes all profile contents except its held lock, without a backup; shared product data and the Harness-home environment file remain intact. The profile directory remains in place so another transaction cannot acquire a replacement lock during cleanup. Self-contained recovery controls use intercepted form navigation when preload is unavailable. A crashed renderer invalidates the navigation cache so the startup page loads again.
+
+Desktop starts the actual Host after preparing the profile in place, without booting a separate health-check backend. Package mutations retain dependency validation, approved lifecycle builds, runtime identity checks, and locking. Failures retain partial changes for explicit repair; there is no automatic profile rollback.
+
+This partially supersedes staged backend probes and waiting to create the main window in the [packaging decision](2026-08-25-electron-desktop-packaging-and-updates.md) and [bundled-runtime decision](2026-09-08-desktop-bundled-runtime-and-external-plugins.md). Those notes retain release, signing, transport, resource ownership, and dependency-transaction rationale. Full runtime file verification remains a packaging operation.
+
+## Alternatives considered
+
+**Keep a complete staged health check.** It can reject some startup failures before activation, but executes plugin initialization twice and cannot guarantee that the serving process will start. The actual Host result provides the diagnostic needed for explicit recovery.
+
+**Keep the main window hidden until readiness.** This avoids presenting a loading page but gives users no visible progress or interaction while the backend loads. A shell-owned page can remain available when Host startup fails.
+
+## Consequences
+
+Users can see startup progress and recover from failures before the product UI is available. A responsive window does not imply that the backend is ready, and startup latency still requires installed-artifact measurement. Profile changes remain in place after activation fails.
+
+Verification covers a delayed Host with a visible loading page, one serving startup for a fresh profile, failure and retry in the same window, plugin management during recovery, and closing while a child is starting. Installed GUI evidence complements lifecycle and transaction tests.

.agents/notes/implemented/architecture/2026-09-09-desktop-immediate-window-and-direct-start.zh.md

@@ -0,0 +1,33 @@
+# Agent Note: Show the Desktop window before starting the Host
+
+Status: implemented
+
+[English](2026-09-09-desktop-immediate-window-and-direct-start.md) | 中文
+
+profile 修改与恢复遵循[直接修改 profile 决策](2026-09-09-desktop-in-place-profile.zh.md)。
+
+## 问题
+
+等待后端就绪会让用户在准备 profile 和加载模块期间看不到窗口。完整的 staging 健康检查进程会在应用启动服务进程前重复启动后端，而插件在实际服务进程中仍然可能启动失败。
+
+## 决策
+
+Electron 在 profile 校准或 Host 启动前创建带本地加载页的主窗口。该页面仅依赖已打包的壳资源，并通过自有 preload 接收 starting、ready 或 error 状态。就绪后在同一窗口加载产品 UI；启动失败时显示诊断和可用恢复操作。加载期间关闭窗口会取消后续启动工作，并等待正在启动的子进程退出。
+
+主窗口提供恢复操作，因为失败的 Host 无法提供自身控件。错误页保留诊断、重启和重装指导。只有已打包应用加载了运行时元数据且资源可用时，才提供禁用插件和重置 Desktop。重置会删除 profile 中除所持锁文件外的所有内容，不保留备份；共享产品数据和 Harness-home 环境文件保持完整。profile 目录保持原位，避免清理期间另一事务获取替代锁。preload 不可用时，独立恢复控件使用被拦截的表单导航。渲染进程崩溃会使导航缓存失效，以重新加载启动页。
+
+Desktop 直接准备 profile 后启动实际 Host，不另行启动健康检查后端。包修改保留依赖验证、获准生命周期构建、运行时标识检查和锁。失败后保留部分修改，供显式修复；不自动回滚 profile。
+
+本决策部分取代[打包决策](2026-08-25-electron-desktop-packaging-and-updates.zh.md)和[内置运行时决策](2026-09-08-desktop-bundled-runtime-and-external-plugins.zh.md)中的 staging 后端探针与延迟创建主窗口。这两份记录仍保留发布、签名、传输、资源归属与依赖事务的理由。完整运行时文件验证仍属于打包操作。
+
+## 考虑过的替代方案
+
+**保留完整 staging 健康检查。** 它可以在激活前拒绝部分启动失败，但会执行两次插件初始化，也不能保证服务进程能够启动。实际 Host 结果提供显式恢复所需的诊断。
+
+**在就绪前隐藏主窗口。** 这避免显示加载页，但后端加载期间用户看不到进度，也无法交互。壳拥有的页面可以在 Host 启动失败时继续使用。
+
+## 后果
+
+用户可以在产品 UI 可用前看到启动进度并从失败中恢复。窗口能够响应不代表后端已经就绪，启动延迟仍需通过已安装产物测量。激活失败后，profile 修改保留在原位。
+
+验证覆盖 Host 延迟时可见的加载页、新 profile 只启动一次服务进程、同一窗口中的失败与重试、恢复期间的插件管理，以及子进程正在启动时关闭应用。安装后 GUI 证据补充生命周期与事务测试。

.agents/notes/implemented/architecture/2026-09-09-desktop-in-place-profile.i18n.yaml

@@ -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/2026-09-09-desktop-in-place-profile.md
+2026-09-09-desktop-in-place-profile.md: 9a221bb334983a18366baeec00a179646dc76385
+2026-09-09-desktop-in-place-profile.zh.md: 53a744ae9ea284c32c3807f67ad365d270217287

.agents/notes/implemented/architecture/2026-09-09-desktop-in-place-profile.md

@@ -0,0 +1,27 @@
+# Agent Note: Modify the Desktop profile in place
+
+Status: implemented
+
+English | [中文](2026-09-09-desktop-in-place-profile.zh.md)
+
+## Problem
+
+Staging preserves an old plugin installation but adds profile copying, directory moves, a recovery journal, and rollback state. Local plugin changes accept explicit repair after failure instead of this complexity.
+
+## Decision
+
+Desktop stops the Host and modifies the current profile directly. Shared host links are detached for package changes and restored when the operation settles. Package locking, dependency validation, and approved native builds remain. Compatible upgrades refresh links without copying plugin files.
+
+Package or Host failures retain partial changes for repair and retry. There is no staging profile, activation journal, directory-swap recovery, or automatic rollback. Existing scratch directories are not interpreted or deleted.
+
+This supersedes staging and rollback in [2026-08-25-electron-desktop-packaging-and-updates](2026-08-25-electron-desktop-packaging-and-updates.md), [2026-09-08-desktop-bundled-runtime-and-external-plugins](2026-09-08-desktop-bundled-runtime-and-external-plugins.md), [2026-09-09-desktop-immediate-window-and-direct-start](2026-09-09-desktop-immediate-window-and-direct-start.md). Other release, module-identity, and window-lifecycle decisions remain active.
+
+A persistent `desktop-packages-pending` marker precedes package writes or native-runtime rebuilding and is removed only after installation, approved builds, and validation succeed. A later launch with that marker reinstalls the locked graph and retries pending builds even when recorded runtime metadata already matches. Ordinary unchanged startups reuse the profile without scanning the plugin dependency graph; package mutations and runtime reconciliation retain validation.
+
+## Alternatives considered
+
+Staging protects the previous installation at the cost of copying and crash recovery. Versioned directories still need preparation, selection, and cleanup. Direct writes give up automatic recovery; reintroduction requires an unattended-recovery product requirement that justifies these costs.
+
+## Consequences
+
+Tests cover offline initialization, in-place upgrades, failure before writes, retained changes after Host failure, partial pnpm failure, restored host links, and exclusive package ownership. Signed application and GUI acceptance remain release-environment checks.

.agents/notes/implemented/architecture/2026-09-09-desktop-in-place-profile.zh.md

@@ -0,0 +1,27 @@
+# Agent Note: 直接修改 Desktop profile
+
+Status: implemented
+
+[English](2026-09-09-desktop-in-place-profile.md) | 中文
+
+## 问题
+
+staging 能保留旧插件安装，但增加 profile 复制、目录移动、恢复日志和回滚状态。本地插件变更接受失败后显式修复，以避免这些复杂度。
+
+## 决策
+
+Desktop 停止 Host 后直接修改当前 profile。修改包前解除宿主共享链接，操作结束后恢复链接。保留包锁、依赖验证和已批准的原生构建。兼容升级只刷新链接，不复制插件文件。
+
+包操作或 Host 失败会保留部分修改，供修复和重试。不使用 staging profile、激活日志、目录切换恢复或自动回滚。已有临时目录不会被解释或删除。
+
+本决策取代以下记录中的 staging 和回滚：[2026-08-25-electron-desktop-packaging-and-updates](2026-08-25-electron-desktop-packaging-and-updates.zh.md), [2026-09-08-desktop-bundled-runtime-and-external-plugins](2026-09-08-desktop-bundled-runtime-and-external-plugins.zh.md), [2026-09-09-desktop-immediate-window-and-direct-start](2026-09-09-desktop-immediate-window-and-direct-start.zh.md)。其他发布、模块实例和窗口生命周期决策继续有效。
+
+持久的 `desktop-packages-pending` 标记先于包写入或原生运行时重建，仅在安装、获准构建和验证成功后删除。后续启动发现该标记时，会重新安装锁定的依赖图并重试待执行构建，即使记录的运行时元数据已经匹配。普通未变化的启动复用 profile，不扫描插件依赖图；包修改和运行时校准保留验证。
+
+## 考虑过的替代方案
+
+staging 以复制和崩溃恢复为代价保护旧安装。版本化目录仍需要准备、选择和清理。直接写入放弃自动恢复；只有无人值守恢复的产品要求能证明这些成本合理时，才重新引入。
+
+## 后果
+
+测试覆盖离线初始化、原地升级、写入前失败、Host 失败后保留修改、pnpm 部分失败、宿主链接恢复及独占包操作。签名应用和 GUI 验收仍由发布环境负责。

.agents/notes/implemented/architecture/2026-09-10-command-identities-and-composer-file-action.i18n.yaml

@@ -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/2026-09-10-command-identities-and-composer-file-action.md
+2026-09-10-command-identities-and-composer-file-action.md: 696c3486c5e007416da3fd5e7be1d4b41cb5f582
+2026-09-10-command-identities-and-composer-file-action.zh.md: 2ce803bad5d39a6d4f40b665fb4ef7456e987912

.agents/notes/implemented/architecture/2026-09-10-command-identities-and-composer-file-action.md

@@ -0,0 +1,33 @@
+# Agent Note: Command identities and composer-owned File action
+
+Status: implemented
+
+English | [中文](2026-09-10-command-identities-and-composer-file-action.zh.md)
+
+## Problem
+
+Matching a command's English description to a client dictionary makes punctuation changes affect localization and inserted command tokens. A same-name override can also copy that description without implementing the first-party command. The File menu entry needs the composer's live attachment policy, including mount, lock, and submission state; a separate command-plugin check cannot determine that state.
+
+## Decision
+
+The command registry preserves an optional branded `CommandDefinitionId` as `definitionId` on the effective definition and descriptor. First-party producers choose their package name as the stable identity. Scoped shadowing selects the complete descriptor and never inherits the shadowed definition's identity. The identifier is discovery metadata, not an authorization claim, and does not enter command lifecycle events.
+
+The client command directory resolves input through its private `resolution.ts`. Exact registered names take priority; Chinese and English aliases select only the corresponding first-party definition in the effective Session catalog. Menu claims use the current locale's spelling; typed claims retain the supplied spelling; submissions use the resolved registered name. `presentation.ts` owns only sections, labels, descriptions, and icons. Resolution helpers and section constants are not exported from the plugin entrypoint.
+
+Conversation registers the File action through the injected command service and owns its localized label. The mounted input binds its file-dialog opener and one live availability query. Both menu filtering and invocation use that query, so lock, unmount, subagent, and submission state apply consistently. The binding and dispatch remain package-internal callbacks; no cross-plugin pick-files event is needed. The assembly uses a narrow structural action-registration face because command UI consumes Conversation's input types; a reverse compiler-project dependency would form a cycle. Its registration test checks against the command plugin's contribution type.
+
+This note supersedes only identity matching, input-resolution placement, and File-action ownership in the [composer menu decision](../feature/2026-09-08-composer-menu-sections-and-localized-rows.md). That note retains the menu, scrolling, claim-retention, and composition-timing decisions.
+
+## Alternatives considered
+
+**Match names and English descriptions.** Copy is editable and can be duplicated by unrelated definitions. A stable identity separates presentation selection from copy without moving localized text onto the Host.
+
+**Match names alone.** An agent-scoped override may deliberately provide a different command under the same name. It must not inherit first-party presentation or aliases unless it explicitly carries that identity.
+
+**Keep the File action in the command plugin.** The input owns attachment acceptance and the DOM lifetime. Maintaining a second availability condition splits one policy between owners.
+
+**Export helpers for reuse.** No production consumer needs the section constants or resolution helper as a plugin API. Tests import internal modules directly.
+
+## Consequences
+
+First-party producers and client identity mappings must agree on stable identifiers; third-party definitions may omit them. Display copy can evolve independently. Registry tests cover descriptor preservation and shadowing; client tests cover description edits in both locales, alias resolution, and exact-name priority. Composer tests cover live availability, opener replacement, unmount, and action-registration disposal. The existing Session-driven menu and command scenarios continue to own assembled browser output.

.agents/notes/implemented/architecture/2026-09-10-command-identities-and-composer-file-action.zh.md

@@ -0,0 +1,33 @@
+# Agent Note: 命令标识与输入框拥有的文件动作
+
+Status: implemented
+
+[English](2026-09-10-command-identities-and-composer-file-action.md) | 中文
+
+## 问题
+
+用命令的英文说明匹配客户端词典，会让标点修改影响本地化和填入的命令 token。同名覆盖也可能复制这段说明，却没有实现内置命令的行为。「文件」菜单项需要输入框实时的附件策略，包括挂载、锁定和提交状态，命令插件中的另一套检查无法确定这些状态。
+
+## 决策
+
+命令注册表在有效定义和描述符上保留可选的品牌类型 `CommandDefinitionId` 字段 `definitionId`。内置命令的提供方用包名作为稳定标识。作用域覆盖选择完整描述符，不继承被遮蔽定义的标识。该标识是发现元数据，不代表授权，也不进入命令生命周期事件。
+
+客户端命令目录通过内部的 `resolution.ts` 解析输入。精确注册名优先，中英文别名只选择会话有效目录中对应的内置定义。菜单认领使用当前语言的写法，手输认领保留原写法，提交使用解析到的注册名。`presentation.ts` 只负责分节、标题、说明和图标。解析辅助函数与小节常量不从插件入口导出。
+
+Conversation 通过注入的命令服务注册「文件」动作，并负责其本地化标题。已挂载输入框绑定文件选择器回调和一个实时可用性查询。菜单过滤与实际调用使用同一个查询，统一处理锁定、卸载、subagent 和提交状态。绑定与调用留在包内回调中，不需要跨插件的文件选择事件。组装层使用窄的结构化动作注册接口，因为命令 UI 已消费 Conversation 的输入类型，反向增加编译项目依赖会形成循环。注册测试使用命令插件的贡献项类型进行校验。
+
+本记录只接管[输入框菜单决策](../feature/2026-09-08-composer-menu-sections-and-localized-rows.zh.md)中的标识匹配、输入解析归属和文件动作归属。原记录继续负责菜单、滚动、认领保留和组合输入时序的决策。
+
+## 考虑过的替代方案
+
+**按名称和英文说明匹配。** 文案可以修改，也可以被无关定义复制。稳定标识将展示选择与文案分开，无需把本地化文本移到宿主。
+
+**只按名称匹配。** agent（智能体）作用域覆盖可能刻意在同名下提供另一种命令。除非它显式携带对应标识，否则不应继承内置展示或别名。
+
+**把文件动作留在命令插件。** 输入框负责附件接收和 DOM 生命周期。维护另一套可用性条件会把同一策略分给两个模块。
+
+**导出辅助函数以供复用。** 没有生产调用方需要把小节常量或解析辅助函数作为插件 API。测试直接导入内部模块。
+
+## 后果
+
+内置命令提供方与客户端映射必须使用一致的稳定标识，第三方定义可以省略标识。显示文案可以独立调整。注册表测试覆盖描述符保留和作用域覆盖，客户端测试覆盖两种语言下修改说明、别名解析和精确名称优先。输入框测试覆盖实时可用性、回调替换、卸载和动作注册的 dispose（资源释放）。现有会话驱动的菜单和命令场景继续负责组装后的浏览器输出。

.agents/notes/implemented/bug-fix/2026-07-31-fail-loud-releases-the-terminal.i18n.yaml

@@ -2,5 +2,5 @@
 # side as of the last confirmed-consistent state. Both languages carry equal authority;
 # after editing either side, bring the other along and re-record with:
 #   pnpm run verify-translation-pairing --write .agents/notes/implemented/bug-fix/2026-07-31-fail-loud-releases-the-terminal.md
-2026-07-31-fail-loud-releases-the-terminal.md: e5196121a850997b5eff045a26dd7c638776196c
-2026-07-31-fail-loud-releases-the-terminal.zh.md: 5cec9fe7d4ac75df1b5c88f2ea4cb0a665730719
+2026-07-31-fail-loud-releases-the-terminal.md: a6b8ccc5a0ad885a349b32707face7bfce9d7bed
+2026-07-31-fail-loud-releases-the-terminal.zh.md: 728fca424d2622af6222d51a3240eafd0e117986

.agents/notes/implemented/bug-fix/2026-07-31-fail-loud-releases-the-terminal.md

@@ -15,7 +15,7 @@ $ 1;2;4cecho hello
 zsh: command not found: 4cecho
 ```
 
-The Loader mounts entries concurrently, so entry failure order is not startup order. `ui-tui` activates and calls pi-tui's `ProcessTerminal.start()`, which puts stdin in raw mode, enables bracketed paste, and writes the Kitty keyboard-protocol probe — a sequence ending in a Device Attributes query (`ESC [ c`). A sibling entry (here `llm-pi-ai`) then rejects on its own config. At the time, that rejection surfaced as an unhandled rejection, and `installFailLoud` wrote one stderr line and called `process.exit(1)` immediately. (The transactional Loader now settles config-tree failures through `boot()`, which disposes the partial context itself; the release hook remains the guard for rejections `boot()` cannot see — a plugin's detached async work rejecting during or after mounting.)
+The Loader mounts entries concurrently, so entry failure order is not startup order. `ui-tui` activates and calls pi-tui's `ProcessTerminal.start()`, which puts stdin in raw mode, enables bracketed paste, and writes the Kitty keyboard-protocol probe — a sequence ending in a Device Attributes query (`ESC [ c`). A sibling entry (here `llm-pi-ai`) then rejects on its own config. At the time, that rejection surfaced as an unhandled rejection, and `installFailLoud` wrote one stderr line and called `process.exit(1)` immediately. (The [Loader activation audit](../simplification/2026-09-09-nontransactional-loader.md) reports config-tree failures through `boot()`, which disposes the partial context itself; the release hook remains the guard for rejections `boot()` cannot see — a plugin's detached async work rejecting during or after mounting.)
 
 Nothing disposed the tree, so `ProcessTerminal.stop()` never ran: raw mode, bracketed paste, and the keyboard protocol stayed set on the shell that outlived the process. The terminal's answer to the Device Attributes query (`1;2;4c`) arrived after exit and was read by the shell as typed input — the literal text above.
 

.agents/notes/implemented/bug-fix/2026-07-31-fail-loud-releases-the-terminal.zh.md

@@ -17,7 +17,7 @@ zsh: command not found: 4cecho
 
 Loader 并发挂载各个条目，因此条目失败的顺序并不等于启动顺序。`ui-tui` 会先激活并调用 pi-tui 的 `ProcessTerminal.start()`，它把 stdin 置为 raw 模式、启用 bracketed paste，并写出 Kitty 键盘协议探测序列——该序列以一个 Device Attributes 查询（`ESC [ c`）结尾。随后某个同级条目（这里是 `llm-pi-ai`）因自身配置而 rejection。
 
-在当时，该 rejection 以未处理 rejection 的形式浮现，而 `installFailLoud` 只写一行 stderr 就立即调用 `process.exit(1)`。（事务化 Loader 现在让配置树失败经 `boot()` 结算，由它自行 dispose（资源释放）部分构建的上下文；release 钩子仍然守护 `boot()` 看不到的 rejection——插件游离的异步工作在挂载期间或挂载之后失败。）没有任何环节 dispose 这棵树，因此 `ProcessTerminal.stop()` 从未执行：raw 模式、bracketed paste 和键盘协议都残留在比进程活得更久的 shell 上。终端对 Device Attributes 查询的回应（`1;2;4c`）在进程退出之后才到达，被 shell 当作用户输入读入——也就是上面那段字面文本。
+在当时，该 rejection 以未处理 rejection 的形式浮现，而 `installFailLoud` 只写一行 stderr 就立即调用 `process.exit(1)`。（[Loader 激活检查](../simplification/2026-09-09-nontransactional-loader.zh.md) 让配置树失败经 `boot()` 报告，由它自行 dispose（资源释放）部分构建的上下文；release 钩子仍然守护 `boot()` 看不到的 rejection——插件游离的异步工作在挂载期间或挂载之后失败。）没有任何环节 dispose 这棵树，因此 `ProcessTerminal.stop()` 从未执行：raw 模式、bracketed paste 和键盘协议都残留在比进程活得更久的 shell 上。终端对 Device Attributes 查询的回应（`1;2;4c`）在进程退出之后才到达，被 shell 当作用户输入读入——也就是上面那段字面文本。
 
 `/exit` 路径从不受影响，因为它会 dispose 整棵树，从而进入 TUI 自身的 `shutdown()`：先 `drainInput()`（吸收尚未返回的响应），再 `ui.stop()`。缺陷在于**启动失败**没有通往这同一套拆卸流程的路径。
 

.agents/notes/implemented/bug-fix/2026-09-10-built-bundle-css-exemption.i18n.yaml

@@ -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/2026-09-10-built-bundle-css-exemption.md
+2026-09-10-built-bundle-css-exemption.md: e5a55689696f885b8ea9411a5b3a7598295fff35
+2026-09-10-built-bundle-css-exemption.zh.md: 1c1aab35df72497946b02bf2a7f67a2271c804e5

.agents/notes/implemented/bug-fix/2026-09-10-built-bundle-css-exemption.md

@@ -0,0 +1,27 @@
+# Agent Note: Built-bundle exemption for a failing stylesheet
+
+Status: implemented
+
+English | [中文](2026-09-10-built-bundle-css-exemption.zh.md)
+
+## Problem
+
+The [Node import sweep](../../../../packages/experimental/webworker-runtime/tests/compile/transform-corpus-check.ts) exempts the Dockkit bundle because Node cannot load its stylesheets, and admitted one exact stylesheet path as the evidence: `packages/client/ui-dockkit/lib/components/dockkit.module.css`. The built bundle imports the workspace package `@deepseek-ai/dsh-client-ui-primitives` before its own stylesheet, and the `tsx` launcher resolves that specifier through tsconfig `paths` into the dependency's `src` tree, so the sweep reports `ERR_UNKNOWN_FILE_EXTENSION` for `packages/client/ui-primitives/src/StateDot.module.css`. The pinned path cannot match on a tree with client build output, and the Windows complete-gate inventory reported the exempt bundle as an unexpected baseline failure.
+
+## Decision
+
+The Dockkit exemption admits Node's unknown-`.css`-extension refusal for any stylesheet. Another extension, another error code, and an unrelated error message stay findings, as does an exempt bundle that imports cleanly.
+
+## Alternatives considered
+
+**Admit the dependency's source stylesheet alongside the pinned one.** That file is what the sweep reports, but the bundle's import order and the launcher's path mapping select it. Pinning it would certify those two details instead of the `.css` exemption.
+
+**Classify every CSS exemption by the same rule.** The Dockkit pin is the recorded evidence this change corrects; the other two stylesheet exemptions were never classified, and tightening them would change what they admit beyond the reported defect.
+
+**Drop the classification and admit any failure.** A bundle that stopped importing for an unrelated reason would then hide inside the exemption total.
+
+## Consequences
+
+The sweep reports the Dockkit bundle when it stops importing for any reason other than Node's unknown-`.css`-extension refusal, and the entry no longer asserts which stylesheet fails. [Scoped resolve/load hooks](../../../../packages/experimental/webworker-runtime/tests/compile/transform-corpus.spec.ts) exercise an admitted Dockkit stylesheet, the dependency's source stylesheet, another extension, an arbitrary message, another error code, and a stale exemption without modifying shared build artifacts.
+
+The [CI observation decision](../testing/2026-09-08-ci-completion-observations.md) keeps the fixture completion and isolation decisions it owns; its built-client classification paragraph keeps the sweep summary and links here for the admitted evidence.

.agents/notes/implemented/bug-fix/2026-09-10-built-bundle-css-exemption.zh.md

@@ -0,0 +1,27 @@
+# Agent Note: 构建产物豁免以失败的样式表为准
+
+Status: implemented
+
+[English](2026-09-10-built-bundle-css-exemption.md) | 中文
+
+## 问题
+
+[Node import sweep](../../../../packages/experimental/webworker-runtime/tests/compile/transform-corpus-check.ts)因 Node 无法加载 Dockkit bundle 的样式表而豁免该 bundle，并曾以一个精确的样式表路径作为接受证据：`packages/client/ui-dockkit/lib/components/dockkit.module.css`。该已构建 bundle 在导入自身样式表之前先导入 workspace 包 `@deepseek-ai/dsh-client-ui-primitives`，`tsx` 启动器又通过 tsconfig `paths` 把这个说明符解析进依赖的 `src` 树，因此 sweep 报告的是 `packages/client/ui-primitives/src/StateDot.module.css` 的 `ERR_UNKNOWN_FILE_EXTENSION`。固定路径在带有 client 构建输出的树上无法匹配，Windows 完整门禁的清单于是把这个豁免 bundle 报告为意外的基线失败。
+
+## 决策
+
+Dockkit 豁免接受 Node 对任意样式表因未知 `.css` 扩展名而拒绝加载。其他扩展名、其他错误码和无关的错误消息仍计为发现，能顺利完成导入的豁免 bundle 同样计为发现。
+
+## 考虑过的替代方案
+
+**在固定路径之外同时接受依赖的源样式表。** sweep 报告的正是该文件，但选中它的是 bundle 的导入顺序和启动器的路径映射。固定该路径认定的将是这两处细节，而非 `.css` 豁免。
+
+**用同一规则分类每个 CSS 豁免。** Dockkit 固定路径是本次修改所纠正的已记录证据；另外两个样式表豁免从未被分类，收窄它们会在所报告缺陷之外改变其接受的内容。
+
+**放弃分类，接受任何失败。** 届时因无关原因停止导入的 bundle 会隐藏在豁免总数之内。
+
+## 后果
+
+只要 Dockkit bundle 因 Node 未知 `.css` 扩展名拒绝之外的任何原因停止导入，sweep 就会报告它，该条目也不再断言失败的是哪个样式表。[限定范围的 resolve/load hook](../../../../packages/experimental/webworker-runtime/tests/compile/transform-corpus.spec.ts)覆盖被接受的 Dockkit 样式表、依赖的源样式表、其他扩展名、任意消息、其他错误码和陈旧豁免，不修改共享构建产物。
+
+[CI 观察决策](../testing/2026-09-08-ci-completion-observations.zh.md)保留其拥有的 fixture 完成与隔离决策；其已构建 Client 导入分类段落保留 sweep 摘要，并就被接受的证据链接到本文。

.agents/notes/implemented/bug-fix/2026-09-10-deepseek-image-token-calculator-v41.i18n.yaml

@@ -2,5 +2,5 @@
 # side as of the last confirmed-consistent state. Both languages carry equal authority;
 # after editing either side, bring the other along and re-record with:
 #   pnpm run verify-translation-pairing --write .agents/notes/implemented/bug-fix/2026-09-10-deepseek-image-token-calculator-v41.md
-2026-09-10-deepseek-image-token-calculator-v41.md: d8042b04c1ed7824720485aa3ccf584f913d0726
-2026-09-10-deepseek-image-token-calculator-v41.zh.md: 6be386da1c66c469329d03e7b86c8c0e2c22ce3a
+2026-09-10-deepseek-image-token-calculator-v41.md: 668738221ebe3e634be7e4fe9d07c9627e0ceb4a
+2026-09-10-deepseek-image-token-calculator-v41.zh.md: a483282c6da667c2b8e32accb091e065419d44fc

.agents/notes/implemented/bug-fix/2026-09-10-deepseek-image-token-calculator-v41.md

@@ -12,7 +12,7 @@ English | [中文](2026-09-10-deepseek-image-token-calculator-v41.zh.md)
 
 `image-tokens.ts` is rewritten as a verbatim port of the `v41` configuration. The constants are a 14px patch, 3:1 per-axis downsampling, a 544×544 total-pixel floor, and a 1024-token cap. The grid formula is `rows × (cols + 1) + 2` with no odd-row extra row, no parity correction, and no even-row trimming in the solver. There is no alignment pad, so the estimate is exact rather than a worst-case upper bound, and there is no aspect-ratio clamp, so extreme aspect ratios reach the cap through the solver's one-row and one-column branches. The over-budget path is a single closed-form solve followed by the published assertion; the decrementing retry loop existed only for the odd-row layout. The provider's fixpoint iteration over the projected dimensions is unchanged.
 
-The test vectors are re-pinned from the published calculator. The request-pricing tests, package README, and this note carry the new numbers; the pixel budget the harness applies before pricing (`DEFAULT_REQUEST_IMAGE_PIXEL_BUDGET`, 640,000 total pixels) and the catalog model ids are unchanged.
+The test vectors are re-pinned from the published calculator. The request-pricing tests, package README, and this note carry the new numbers. This change left the 640,000 total-pixel projection the harness applied before pricing and the catalog model ids as they were; the [successor](2026-09-10-deepseek-v41-request-image-projection.md) later replaced that projection with the same grid, so omitting `imagePixelBudget` now projects onto the token grid while a positive integer or `low` keeps a total-pixel budget.
 
 ## Alternatives considered
 
@@ -24,4 +24,4 @@ The test vectors are re-pinned from the published calculator. The request-pricin
 
 ## Consequences
 
-An 800×800 request image costs 422 tokens instead of 349, while a 640×480 image costs 206 instead of 209 and a low-budget 512×512 image costs 184 instead of 201. Compaction pressure changes with the retained image dimensions. The 640,000-pixel budget does not imply a 422-token ceiling: an 8192×1 image stays within that pixel budget and costs 1024 tokens. The estimate no longer carries a three-token conservative margin; provider usage remains the authoritative anchor once a request completes. Sessions replayed through `llm-replay` use their fixture's `imageRequestTokens` and are unaffected.
+An 800×800 request image costs 422 tokens instead of 349, while a 640×480 image costs 206 instead of 209 and a low-budget 512×512 image costs 184 instead of 201. Compaction pressure changes with the retained image dimensions. The request projection later moved onto the same grid ([successor](2026-09-10-deepseek-v41-request-image-projection.md)), so the 640,000-pixel budget that made a square request image cost 422 tokens no longer applies. The estimate no longer carries a three-token conservative margin; provider usage remains the authoritative anchor once a request completes. Sessions replayed through `llm-replay` use their fixture's `imageRequestTokens` and are unaffected.

.agents/notes/implemented/bug-fix/2026-09-10-deepseek-image-token-calculator-v41.zh.md

@@ -12,7 +12,7 @@ Status: implemented
 
 `image-tokens.ts` 重写为 `v41` 配置的逐句移植。常量为 14px patch、每轴 3:1 降采样、544×544 总像素下限、1024 token 上限。网格公式为 `rows × (cols + 1) + 2`，没有奇数行额外行、没有奇偶校正、求解器也不再把行数截成偶数。没有对齐 pad，所以估算值是精确值而非最坏情况上界；没有宽高比钳制，所以极端长宽比会经求解器的单行和单列分支到达上限。超预算路径是一次闭式求解加上公开的断言；逐步递减的重试循环只服务于奇数行布局。提供方对投影尺寸的定点迭代保持不变。
 
-测试向量按公开计算器重新固定。request-pricing 测试、包 README 和本 note 使用新数字；harness 在定价前应用的像素预算（`DEFAULT_REQUEST_IMAGE_PIXEL_BUDGET`，640,000 总像素）和 catalog 模型 id 不变。
+测试向量按公开计算器重新固定。request-pricing 测试、包 README 和本 note 使用新数字。本次改动保留了 harness 在定价前应用的 640,000 总像素投影和 catalog 模型 id；[后续决策](2026-09-10-deepseek-v41-request-image-projection.zh.md)把该投影换成了同一套网格，现在省略 `imagePixelBudget` 走 token 网格，正整数或 `low` 仍走总像素预算。
 
 ## 备选方案
 
@@ -24,4 +24,4 @@ Status: implemented
 
 ## 后果
 
-800×800 请求图片的计价从 349 变为 422 token，640×480 图片从 209 变为 206，低预算下的 512×512 图片从 201 变为 184。压缩压力随保留图片的尺寸变化。640,000 像素预算不意味着 422 token 上限：8192×1 图片在该像素预算内，仍计 1024 token。估算值不再带 3 token 的保守余量；请求完成后，提供方 usage 仍是权威锚点。经 `llm-replay` 回放的会话使用各自 fixture 的 `imageRequestTokens`，不受影响。
+800×800 请求图片的计价从 349 变为 422 token，640×480 图片从 209 变为 206，低预算下的 512×512 图片从 201 变为 184。压缩压力随保留图片的尺寸变化。请求投影后来改为同一套网格（[后续决策](2026-09-10-deepseek-v41-request-image-projection.zh.md)），让正方形请求图片计 422 token 的 640,000 像素预算已不再适用。估算值不再带 3 token 的保守余量；请求完成后，提供方 usage 仍是权威锚点。经 `llm-replay` 回放的会话使用各自 fixture 的 `imageRequestTokens`，不受影响。

.agents/notes/implemented/bug-fix/2026-09-10-deepseek-v41-request-image-projection.i18n.yaml

@@ -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/2026-09-10-deepseek-v41-request-image-projection.md
+2026-09-10-deepseek-v41-request-image-projection.md: 2bf7ac3f7f09ff7aea1067030f4474306d3f9a8e
+2026-09-10-deepseek-v41-request-image-projection.zh.md: 4f0a8cfd3a71d54a54bf74d6819235ed7d0749f8

.agents/notes/implemented/bug-fix/2026-09-10-deepseek-v41-request-image-projection.md

@@ -0,0 +1,31 @@
+# Agent Note: DeepSeek request images on the published token grid
+
+Status: implemented
+
+English | [中文](2026-09-10-deepseek-v41-request-image-projection.zh.md)
+
+## Problem
+
+The harness projected every DeepSeek request image under a 640,000 total-pixel budget, a value chosen for the retired V4 vision model and kept unchanged when the [token estimator moved to the `v41` calculator](2026-09-10-deepseek-image-token-calculator-v41.md). The current Flash model retains far more: it pads each edge to whole 14px patches, groups 3×3 patches into one token cell, and keeps the largest aspect-preserving grid whose token count `rows × (columns + 1) + 2` fits 1024. A square image keeps 1302×1302 pixels, a 16:9 image keeps a 1708×966 grid, and extreme aspect ratios keep up to about 1.8 million pixels. The 640,000-pixel projection therefore sent a 2000×2000 screenshot as 800×800, roughly 38% of the pixels the model would have used, and the estimator priced that reduced version at 422 tokens instead of the 994 the model charges for the full grid. The Vision guide's "about 1300×1300 total pixels" describes only the square case; the exact rule is the token grid.
+
+Two smaller gaps sat beside it. The request version had no per-side cap: the provider rejects any image over 4096 pixels per side once a request carries 15 or more images, while normalization admits an 8192-pixel long edge, so a many-image session could fail on one thin image. The 1 MiB encoded-byte target was sized for 640,000-pixel outputs and would push a 1302×1302 photograph down the JPEG quality ladder.
+
+## Decision
+
+The route chooses each request image's dimensions; the attachment provider only resizes and encodes to them. `ImageRequestPolicy` in `dsh-attachment` becomes `ImageRequestTarget`: a width, a height, and the byte target for one attachment. `readImageRequest` resizes by the source long edge alone without enlargement, so the encoder derives the short edge as the route predicted, and keys its cache by the attachment id, target dimensions, byte target, encoder settings, and the new `request-image-v6` transform version, so no earlier cache entry or upload mapping is reused. `dsh-attachment` keeps two provider-neutral geometry exports: `requestImageDimensions` for a total-pixel budget and `longEdgeDimensions` for an exact long edge with a rounded short edge.
+
+`llm-deepseek` owns the provider rule. `image-tokens.ts` keeps the verbatim `v41` solver and adds `deepSeekRequestImageDimensions`: the source itself when its patch-padded grid fits the cap, otherwise the source aspect ratio at the solved grid's long edge, so a 3840×2160 source is sent as 1708×961 and the provider pads it to its 1708×966 grid. `resolveRequestImageTarget` applies that solver when `imagePixelBudget` is omitted, `requestImageDimensions` for a positive integer or the 512×512 `low` preset, then a 4096-pixel per-side cap on every request image so the image count never changes a target, and the route's 2 MiB byte target. Pricing prices `deepSeekImageTokens` of the same target, so the estimator and the sent image come from one solver. The pi-ai route derives its targets from its unchanged 2048×2048 pixel budget. Small images are never enlarged because the provider scales up below 544×544 pixels itself.
+
+## Alternatives considered
+
+**Raise the pixel budget to 1302×1302.** A total-pixel budget is right only for squares: a 16:9 source would be sent at 1.69 million pixels when the grid keeps 1.65 million, and a 4:1 source when it keeps 1.59 million, while extreme ratios would lose detail the grid keeps. One rule that reproduces the provider removes the guesswork.
+
+**A `token-grid` projection kind on the attachment policy, with the solver in `dsh-attachment`.** This was built first: the policy became a closed union of `pixel-budget` and `token-grid`, the solver moved into `request-projection.ts`, and `deepSeekImageTokens` imported it back. It put one provider's layout formula and patch constants into the provider-neutral package under a generic-looking name, needed an `unscaled` flag so the store could tell "send the source" from "send the solved size", and would grow a new union member for every provider rule. Handing the store a finished target keeps the provider rule beside the provider's pricing and leaves the store with no projection vocabulary at all.
+
+**Send the solver's exact grid dimensions with a fill resize.** The solved grid edges are whole patches and differ from the source aspect ratio by under one patch. Filling that box would distort the image slightly even though the provider does the same on its side; preserving the source aspect ratio can change how many token cells the rounded short edge covers. A 1224×1429 source is sent as 1187×1386: the published calculator gives 959 tokens for the source and 992 for the sent dimensions. Request generation and pricing share the target dimensions, and pricing applies the published calculator to that target.
+
+**Keep the 1 MiB target.** The target is not a cap: an output over it is still sent at the smallest ladder quality. At 1302×1302 a JPEG photograph at quality 85 lands between 400 KB and 1.2 MB, so 2 MiB keeps most images at the top quality and lets more PNG screenshots pass through losslessly, while the inline base64 fallback still holds about seven such images under its 20 MiB bound.
+
+## Consequences
+
+A square source now reaches the model at up to 1302×1302 pixels and 994 tokens instead of 800×800 and 422, so image-heavy sessions reach compaction pressure sooner and the estimator applies the published token rules to the sent target dimensions. Every existing request-image cache entry and DeepSeek Files API mapping is regenerated on the next request. Thin images keep their full grid until the per-side cap applies: an 8192×78 source costs 396 tokens under the grid but is sent as 4096×39. `llm-replay` does not project images, so keyless snapshots cannot record the sent dimensions; the `llm-deepseek` adapter tests pin the resolved targets and the projected handle text against a mock server, and the local store tests resize real images to targets.

.agents/notes/implemented/bug-fix/2026-09-10-deepseek-v41-request-image-projection.zh.md

@@ -0,0 +1,31 @@
+# Agent Note: DeepSeek 请求图片按官方 token 网格投影
+
+Status: implemented
+
+[English](2026-09-10-deepseek-v41-request-image-projection.md) | 中文
+
+## 问题
+
+harness 此前把每张 DeepSeek 请求图片投影到 640,000 总像素预算内。这个值是为已下线的 V4 视觉模型选的，[token 预估器改用 `v41` 计算器](2026-09-10-deepseek-image-token-calculator-v41.zh.md)时没有改动它。当前 Flash 模型保留的远多于此：它把每条边补齐到整数个 14 px patch，把 3×3 个 patch 归为一个 token 格，再保留 token 数 `rows × (columns + 1) + 2` 不超过 1024 的最大等比网格。正方形图片保留 1302×1302 像素，16:9 图片保留 1708×966 的网格，极端宽高比最多保留约 180 万像素。因此 640,000 像素投影把一张 2000×2000 的截图缩成 800×800 发出，只有模型本可使用像素的约 38%，预估器为这个缩小版计 422 token，而模型对完整网格收 994 token。图像理解指南里的「约 1300×1300 总像素」只描述正方形的情况，确切规则是 token 网格。
+
+旁边还有两个较小的缺口。请求版本没有单边上限：请求包含 15 张及以上图片时，提供方拒绝任何单边超过 4096 像素的图片，而规范化允许 8192 像素长边，多图会话可能因一张细长图失败。1 MiB 编码字节目标是按 640,000 像素输出定的，会把 1302×1302 的照片压到 JPEG 质量阶梯的低档。
+
+## 决策
+
+请求图片的尺寸由路由决定，附件提供方只负责缩放和编码。`dsh-attachment` 里的 `ImageRequestPolicy` 改为 `ImageRequestTarget`，即一张附件的目标宽、高和字节目标。`readImageRequest` 只按源图长边缩放且不放大，编码器按路由预测的方式推出短边；缓存按附件 id、目标尺寸、字节目标、编码参数和新的 `request-image-v6` 变换版本取键，因此之前的缓存条目和上传映射都不会被复用。`dsh-attachment` 保留两个提供方无关的几何导出：按总像素预算的 `requestImageDimensions`，以及长边精确、短边四舍五入的 `longEdgeDimensions`。
+
+提供方规则归 `llm-deepseek`。`image-tokens.ts` 保留逐字移植的 `v41` 求解器，并新增 `deepSeekRequestImageDimensions`：补齐 patch 后的网格在上限内就发源图本身，否则按源图宽高比取求解网格的长边，于是 3840×2160 的源图以 1708×961 发送，提供方再把它补齐到 1708×966 的网格。`resolveRequestImageTarget` 在省略 `imagePixelBudget` 时用这个求解器，正整数或 512×512 的 `low` 预设用 `requestImageDimensions`，然后对每张请求图片加 4096 像素单边上限，使图片数量不会改变目标，最后带上路由的 2 MiB 字节目标。计价对同一个目标算 `deepSeekImageTokens`，预估器和发出的图片来自同一个求解器。pi-ai 路由从它不变的 2048×2048 像素预算推导目标。小图不放大，因为提供方自己会放大 544×544 像素以下的图片。
+
+## 备选方案
+
+**把像素预算提高到 1302×1302。** 总像素预算只对正方形正确：16:9 的源图会按 169 万像素发送而网格只保留 165 万，4:1 的源图网格只保留 159 万，极端比例又会丢掉网格本会保留的细节。一条复现提供方的规则消除了猜测。
+
+**在附件策略上加 `token-grid` 投影种类，求解器放进 `dsh-attachment`。** 最初就是这样做的：策略变成 `pixel-budget` 和 `token-grid` 的封闭联合，求解器搬进 `request-projection.ts`，`deepSeekImageTokens` 再从那里引回来。这把一家提供方的布局公式和 patch 常量放进了提供方无关的包，还起了个看似通用的名字；存储层需要一个 `unscaled` 标志来区分「发源图」和「发求解尺寸」；以后每多一家提供方规则，联合就要多长一个分支。把算好的目标交给存储层，提供方规则和它的计价放在一起，存储层不需要任何投影词汇。
+
+**用填充缩放发送求解器的精确网格尺寸。** 求解出的网格边长是整数个 patch，与源图宽高比相差不到一个 patch。填充到这个框会轻微变形，尽管提供方那侧也会这样做；保持源图宽高比可能改变取整后短边覆盖的 token 格数。1224×1429 的源图以 1187×1386 发送，官方计算器对源图计 959 token，对发送尺寸计 992 token。请求生成和定价共享目标尺寸，定价按目标尺寸应用官方计算规则。
+
+**保留 1 MiB 目标。** 目标不是上限：超过它的输出仍会以阶梯最小质量发送。1302×1302 的 JPEG 照片在质量 85 时约 400 KB 到 1.2 MB，2 MiB 让多数图片停在最高质量，也让更多 PNG 截图无损直发，而内联 base64 回退在 20 MiB 上界内仍能容纳约七张这样的图片。
+
+## 后果
+
+正方形源图现在最多以 1302×1302 像素、994 token 到达模型，而不是 800×800 和 422，因此图片密集的会话更早触及 compaction 压力，预估器按发送目标尺寸应用官方 token 计算规则。所有已有的请求图片缓存条目和 DeepSeek Files API 映射在下次请求时重新生成。细长图在单边上限生效前保留完整网格：8192×78 的源图在网格下计 396 token，但以 4096×39 发送。`llm-replay` 不投影图片，keyless 快照记录不到发送尺寸；`llm-deepseek` 适配器测试对着 mock 服务器固定了解析出的目标和投影后的句柄文本，本地存储测试把真实图片缩放到目标尺寸。

.agents/notes/implemented/feature/2026-09-08-composer-menu-sections-and-localized-rows.i18n.yaml

@@ -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/feature/2026-09-08-composer-menu-sections-and-localized-rows.md
+2026-09-08-composer-menu-sections-and-localized-rows.md: 3cccb3c038ae4c358860f09533e966e9121b987e
+2026-09-08-composer-menu-sections-and-localized-rows.zh.md: 31c879b17ac6788bfade50a3245903c3e2e56224

.agents/notes/implemented/feature/2026-09-08-composer-menu-sections-and-localized-rows.md

@@ -0,0 +1,43 @@
+# Agent Note: Composer menu sections, localized rows, and the File row
+
+Status: implemented
+
+English | [中文](2026-09-08-composer-menu-sections-and-localized-rows.zh.md)
+
+## Problem
+
+The composer's `+` button and a typed `/` listed every command in Host registration order as `name description`, all lowercase, with no glyphs and no grouping, beside a separate paperclip button for files. Under Chinese the rows stayed English because Host descriptors carry English text only, and a user who knew a command by its Chinese title could neither find it by that title nor see what to type. Issue #3567 and the design doc for it ask for two sections in usage order, a glyph and a left-aligned title per row with the description right-aligned, capitalized English titles, Chinese titles and descriptions that stay searchable in both languages and show the English command name, a Chinese fill for the Plan and Goal claims, and a File entry inside the menu.
+
+## Decision
+
+`ui-commands` owns the menu's presentation in `src/client/presentation.ts`: an Add section (`file`, `goal`, `plan`, `feedback`) and a Commands section (`compact`, `permission`, `model`, `export`), each in usage order, with rows outside both lists closing Commands in catalog order. The section lists are keyed by name; an empty query returns the sectioned rows, and a typed query returns the flat ranking of every visible row so the best match is always first.
+
+The six built-in Host commands get their localized title, description, glyph, and claim token from the client. Stable identity selection and input resolution follow the [command identity decision](../architecture/2026-09-10-command-identities-and-composer-file-action.md). Contributions carry `label()`, `description()`, and `icon`, read on every candidate pass, so the `/model` row localizes without re-registration.
+
+A menu pick fills the locale's claim token: under Chinese, picking Plan fills `/计划 ` and the submission executes `/plan `. A typed token keeps its typed spelling as the claim, because the composer reads the arguments after the token it holds. The effective Session catalog resolves Chinese and English aliases through the same input path in every locale.
+
+The File row is an `action` contribution: a bare invocation consumes the trigger token and runs a client callback without submitting a message. Conversation owns that registration, its live availability, and the hidden file input. The menu replaces the separate paperclip button; the `+` button's accessible name and tooltip read "Add files or run commands".
+
+`ui-input-trigger` renders the new row anatomy: `InputTriggerCandidate.label` is the title and a second search key of the shared `rankByName`, the name renders as a trailing alias when the label differs from it, `icon` accepts an icon component beside the reference glyph tokens, and the description is right-aligned. `ui-primitives` gains the Plan glyph from the design doc, a static ring for Compact, and the permission shield contour.
+
+The menu uses a 400 px border-box height cap, which fits both headings and the eight built-in rows before the viewport clamp reduces it. A real overflow keeps a 10 px draggable WebKit rail around a 4 px visible thumb, insets the track from the rounded ends, and shows a bottom fade until the viewport reaches the final row; Firefox keeps its standard thin scrollbar.
+
+## Alternatives considered
+
+**Release a claim when its separator is deleted.** The complete command name still identifies the selected command. Keeping the claim until the name changes preserves its highlight through argument replacement and avoids relying on an IME-generated space to run ordinary keydown adjudication. The shared input machine applies this rule to every command token, including failure recovery; neither the command name nor the locale selects a separate implementation.
+
+**Restore placeholders directly on native composition end.** Browsers can deliver that event before Lexical reconciles the final text. The shared editor binding keeps command hints and ordinary placeholders hidden while either native or editor composition remains active, and reevaluates visibility after an editor commit, including a cancellation that changes no text. Keyboard submit guards retain their separate post-composition window.
+
+**Localize descriptions on the Host.** The Host has no locale and its catalog is shared by every client; localized product copy belongs to the client.
+
+**Keep sections under a typed query.** Ranking inside sections put a prefix hit in Commands below weaker matches in Add (typing `e` listed Feedback and File above Export); a flat ranking keeps the best match first, and the headings only carry information while the list is complete.
+
+**Keep the paperclip beside the menu.** The design doc and the issue's acceptance criteria integrate the entry into the menu and forbid the old icon from showing twice; the hidden file input and the drop gate stay where they were.
+
+**A contribution-owned copy for the built-in Host rows.** Host packages have no client half to register from, so the copy has to live on the client; one table keyed by name in the package that already owns the `/` source keeps the design decision in one place.
+
+**A `token` alias only under the active locale.** Resolving through every dictionary costs nothing and lets a draft persisted under one locale submit under another.
+
+## Consequences
+
+The menu and localized presentation remain client-owned, while the Host owns effective command definitions. Adding a localized first-party command requires its identity mapping, dictionary entries, icon, and menu position. Session-driven browser scenarios and owner-local ARIA expectations cover the menu. Unrelated same-name definitions keep their own copy and no first-party glyph. The command identity decision owns this distinction and the File action's lifecycle.

.agents/notes/implemented/feature/2026-09-08-composer-menu-sections-and-localized-rows.zh.md

@@ -0,0 +1,43 @@
+# Agent Note: Composer 菜单分节、本地化行与「文件」行
+
+Status: implemented
+
+[English](2026-09-08-composer-menu-sections-and-localized-rows.md) | 中文
+
+## 问题
+
+composer 的 `+` 按钮与键入的 `/` 按宿主注册顺序把每个命令列成 `name description`，全部小写，没有图标也不分组，旁边还有一个单独的回形针按钮用于添加文件。中文界面下各行仍是英文，因为宿主描述符只携带英文文案；只知道命令中文名的用户既搜不到它，也看不出该输入什么。Issue #3567 及其设计稿要求：按使用频次分成两个小节，每行带图标、标题左对齐、说明右对齐，英文标题首字母大写，中文标题与说明在两种语言下都能搜索并同时显示英文命令名，计划与目标的声明用中文填入，以及把「文件」入口放进菜单。
+
+## 决策
+
+`ui-commands` 在 `src/client/presentation.ts` 里拥有菜单的展示：「添加」小节（`file`、`goal`、`plan`、`feedback`）与「指令」小节（`compact`、`permission`、`model`、`export`），各按使用频次排列，不在两个清单里的行按目录顺序排在「指令」末尾。小节清单按名字索引；空查询返回分节的行，输入查询后返回所有可见行的平铺排序，最佳匹配始终在第一位。
+
+六个内置宿主命令从客户端取得本地化标题、说明、图标和认领 token。稳定标识选择与输入解析遵循[命令标识决策](../architecture/2026-09-10-command-identities-and-composer-file-action.zh.md)。贡献项携带 `label()`、`description()` 和 `icon`，每次生成候选项时读取，因此 `/model` 行无需重新注册即可本地化。
+
+菜单选中填入当前语言的认领 token：中文下选中「计划」填入 `/计划 `，提交时执行 `/plan `。手输 token 保留原写法作为认领，因为输入框从它持有的 token 之后读取参数。会话有效目录在所有界面语言下通过同一输入路径解析中英文别名。
+
+「文件」行是一个 `action` 贡献项：裸调用消费触发 token 后运行客户端回调，不提交消息。Conversation 负责该注册、实时可用性和隐藏文件输入框。菜单取代单独的回形针按钮，`+` 按钮的无障碍名称与提示为「添加文件或调用指令」。
+
+`ui-input-trigger` 渲染新的行结构：`InputTriggerCandidate.label` 是标题，也是共享 `rankByName` 的第二个搜索键；label 与名字不同时名字渲染为尾随别名；`icon` 在引用图标 token 之外接受图标组件；说明右对齐。`ui-primitives` 新增设计稿给出的计划图标、压缩用的静态环形，以及权限盾形轮廓。
+
+菜单采用 400 px 的 border-box 高度上限，在视口限制缩小高度前可容纳两个小节标题与八个内置入口。内容确实溢出时，WebKit 系浏览器用 10 px 的可拖动区域承载 4 px 的可见滑块，轨道避开圆角两端，视口抵达最后一行前显示底部渐隐；Firefox 保留标准细滚动条。
+
+## 考虑过的替代方案
+
+**删除分隔空格时释放命令认领。** 完整命令名仍能标识已选命令。保留认领直到命令名改变，可以在替换参数时保持高亮，也无需依赖输入法生成的空格触发普通按键裁决。共享输入状态机对所有命令 token 使用这个规则，提交失败后的恢复也相同；命令名和界面语言都不选择另一套实现。
+
+**在原生组合输入结束事件中直接恢复占位文字。** 浏览器可能先发送该事件，Lexical 随后才完成最终文字更新。共享编辑器绑定在原生输入法或编辑器仍处于组合输入时隐藏命令提示和普通占位文字，并在编辑器提交更新后重新判断显隐，包括没有文字变化的取消操作。键盘提交保护保留独立的组合输入结束后保护时段。
+
+**在宿主侧本地化说明。** 宿主没有语言设置，目录由所有客户端共享，本地化产品文案属于客户端。
+
+**输入查询后保留小节。** 在小节内排序会把「指令」里的前缀命中排在「添加」里较弱的匹配之下（输入 `e` 时反馈与文件排在导出之上）；平铺排序让最佳匹配始终靠前，小节标题只在列表完整时才有信息量。
+
+**在菜单旁保留回形针。** 设计稿与 issue 的验收条件都要求把入口整合进菜单且旧图标不再重复显示；隐藏的文件输入框与拖放门槛保持原位。
+
+**由贡献项拥有内置宿主行的文案。** 宿主包没有客户端一半可供注册，文案只能放在客户端；在已经拥有 `/` source 的包里用一张按名字索引的表，能把设计决策集中在一处。
+
+**只在当前语言下解析 `token` 别名。** 遍历所有词典解析没有成本，还能让在一种语言下持久化的草稿在另一种语言下提交。
+
+## 后果
+
+菜单与本地化展示由客户端负责，宿主负责有效命令定义。新增本地化内置命令需要对应的标识映射、词典条目、图标和菜单位置。会话驱动的浏览器场景与归属模块的 ARIA 预期覆盖菜单。无关的同名定义保留自己的文案，不获得内置图标。命令标识决策负责这一划分和文件动作的生命周期。

.agents/notes/implemented/feature/2026-09-08-feedback-dialog-and-categories.i18n.yaml

@@ -2,5 +2,5 @@
 # side as of the last confirmed-consistent state. Both languages carry equal authority;
 # after editing either side, bring the other along and re-record with:
 #   pnpm run verify-translation-pairing --write .agents/notes/implemented/feature/2026-09-08-feedback-dialog-and-categories.md
-2026-09-08-feedback-dialog-and-categories.md: 02a41f08541cca85dad7cd3c2c984b5e7cd8fd6d
-2026-09-08-feedback-dialog-and-categories.zh.md: 40d83360e52bff569120820b78959a7cc3f4db4f
+2026-09-08-feedback-dialog-and-categories.md: ee656946a9e5d2921a02a1b642041e5965661f31
+2026-09-08-feedback-dialog-and-categories.zh.md: ef2b2d4670aad12398125925dd1e1a7e7c10a5f0

.agents/notes/implemented/feature/2026-09-08-feedback-dialog-and-categories.md

@@ -12,9 +12,9 @@ The Web client had two disconnected feedback paths with no visible outcome. `/fe
 
 `command-feedback` owns the category taxonomy as the `FeedbackCategory` union and the `FEEDBACK_CATEGORIES` tuple in its client-safe `./types` export, and `feedback/record` becomes `{ text?, category? }`: blank text is recorded as absent, and an entry with neither member still records, because the log delivery that the feedback authorizes is the content. The same package publishes the `sessionFeedback.record` Remote through `TypertRemoteService`, resolving the live Session by id and calling the existing `recordFeedback` producer, so the dialog records the same event as the command without command bookkeeping. `message-feedback` adds the optional `category` to `MessageFeedbackItem` and `MessageFeedbackPutRequest`, validates stored values against the tuple, and counts a category change as a material edit.
 
-`ui-message-feedback` becomes the Web feedback surface. A per-session `FeedbackSurface` owns the message-feedback controller, a `FeedbackDialogController` for the draft, the submission, and the toast sequence, and the routing between them: a message target puts a negative judgment with the dialog's category and note through the message controller, the Session target records through `ctx.remote.sessionFeedback`. A `FeedbackDialog` entry of `conversation.input.overlay` renders the Modal and Toast primitives from the dialog store. A decoration on the Host's `feedback` command opens the dialog for the Session from a menu pick or a bare Enter while `/feedback <text>` still reaches the Host; it uses the `action` kind this PR adds to `CommandUiSpec`, a bare invocation that consumes the trigger token and runs a client callback without submitting anything. Dislike opens the same dialog for the message. Like calls `toggle`, which now reports the rating it committed, so the row acknowledges a recorded Like and stays silent on a retraction. The note popover, `clearNote`, and `clear` are removed: the dialog is the only note editor, a rating switch stores the bare judgment, and clicking a recorded rating retracts it.
+`ui-message-feedback` becomes the Web feedback surface. A per-session `FeedbackSurface` owns the message-feedback controller, a `FeedbackDialogController` for the draft, the submission, and the toast sequence, and the routing between them: a message target puts its selected judgment with the dialog's category and note through the message controller, while the Session target records through `ctx.remote.sessionFeedback`. A `FeedbackDialog` entry of `conversation.input.overlay` renders the Modal and Toast primitives from the dialog store. A decoration on the Host's `feedback` command opens the dialog for the Session from a menu pick or a bare Enter while `/feedback <text>` still reaches the Host; it uses the `action` kind in `CommandUiSpec`, which consumes the trigger token and runs a client callback without submitting anything. The later [symmetric message feedback submission](2026-09-10-symmetric-message-feedback-submission.md) decision owns the rating entry rule: either unrecorded rating opens the dialog, while clicking the recorded rating retracts it. The note popover, `clearNote`, and `clear` remain absent because the dialog is the only note editor.
 
-The dialog is the shared Modal card at the design's width; the design's checkbox for including the conversation log is not built, because the log travels with every feedback event and is not optional. An oversized description still fails on submit with `note-too-large`; the dialog stays open with the code.
+The dialog is the shared Modal card at the design's width; the design's checkbox for including the conversation log is not built, because the log travels with every feedback event and is not optional. An oversized description still fails on submit with `note-too-large`; the dialog stays open with its draft and a warning toast presents the localized failure.
 
 ## Alternatives considered
 
@@ -24,9 +24,9 @@ The dialog is the shared Modal card at the design's width; the design's checkbox
 
 **Keep the note popover beside the dialog.** Two editors for one note with different reachability would leave the row two-line at some widths, the defect the popover was introduced to avoid, and the design shows only the thumbs.
 
-**A Toast per message control.** The composer overlay already mounts once per Session, and the dialog owns the toast sequence, so one owner serves the Like path and the dialog path alike.
+**A Toast per message control.** The composer overlay already mounts once per Session, and the dialog owns the toast sequence, so one owner serves both message-rating paths and the Session dialog.
 
-**A dialog kind in `CommandUiSpec`.** An action that consumes the token and runs a client callback is all the dialog needs; PR #3745 introduces the same `action` kind for its File row, so whichever lands second keeps one definition.
+**A dialog kind in `CommandUiSpec`.** An action that consumes the token and runs a client callback is all the dialog needs; the File row uses the same `action` kind, so one definition serves both entries.
 
 ## Consequences
 

.agents/notes/implemented/feature/2026-09-08-feedback-dialog-and-categories.zh.md

@@ -12,9 +12,9 @@ Web 客户端有两条互不相连的反馈路径，且都没有可见结果。`
 
 `command-feedback` 在其客户端可用的 `./types` 导出中以 `FeedbackCategory` 联合类型与 `FEEDBACK_CATEGORIES` 元组拥有分类表，`feedback/record` 变为 `{ text?, category? }`：空白文本记为缺省，两个成员都没有的条目仍会记录，因为反馈所授权的日志投递本身就是内容。同一个包通过 `TypertRemoteService` 发布 `sessionFeedback.record` Remote，按 id 找到 live Session 后调用已有的 `recordFeedback` 生产方，因此弹窗记录的是与命令相同的事件，只是没有命令簿记。`message-feedback` 给 `MessageFeedbackItem` 与 `MessageFeedbackPutRequest` 加上可选 `category`，按元组校验已存值，并把分类变化算作实质编辑。
 
-`ui-message-feedback` 成为 Web 反馈界面。每个 Session 一个 `FeedbackSurface`，拥有消息反馈控制器、负责草稿、提交与 toast 序号的 `FeedbackDialogController`，以及两者之间的路由：消息目标经消息控制器 put 一条带弹窗分类与备注的差评，Session 目标经 `ctx.remote.sessionFeedback` 记录。`conversation.input.overlay` 的 `FeedbackDialog` 条目从弹窗 store 渲染 Modal 与 Toast 基元。宿主 `feedback` 命令上的装饰让菜单选中或不带参数的回车为 Session 打开弹窗，而 `/feedback <text>` 仍到达宿主；它使用本 PR 给 `CommandUiSpec` 新增的 `action` 种类：裸调用消费触发 token 后运行一个客户端回调，不提交任何内容。点踩为消息打开同一个弹窗。点赞调用 `toggle`，它现在会报告自己提交的评分，因此该行只对记录成功的点赞做确认，撤回时保持沉默。备注浮层、`clearNote` 与 `clear` 被移除：弹窗是唯一的备注编辑器，切换评分只存判断本身，再次点击已记录的评分即撤回。
+`ui-message-feedback` 成为 Web 反馈界面。每个 Session 一个 `FeedbackSurface`，拥有消息反馈控制器、负责草稿、提交与 toast 序号的 `FeedbackDialogController`，以及两者之间的路由：消息目标经消息控制器 put 一条带弹窗分类与备注的所选评分，Session 目标经 `ctx.remote.sessionFeedback` 记录。`conversation.input.overlay` 的 `FeedbackDialog` 条目从弹窗 store 渲染 Modal 与 Toast 基元。宿主 `feedback` 命令上的装饰让菜单选中或不带参数的回车为 Session 打开弹窗，而 `/feedback <text>` 仍到达宿主；它使用 `CommandUiSpec` 中的 `action` 种类：裸调用消费触发 token 后运行一个客户端回调，不提交任何内容。后续的[对称消息反馈提交](2026-09-10-symmetric-message-feedback-submission.zh.md)决策拥有评分入口规则：任一未记录的评分都会打开弹窗，再次点击已记录的评分则撤回。备注浮层、`clearNote` 与 `clear` 继续保持移除，因为弹窗是唯一的备注编辑器。
 
-弹窗是共用的 Modal 卡片，宽度按设计稿；设计稿里「包括当前对话的日志」复选框不做，因为日志随每个反馈事件一起投递，不是可选项。超长描述仍在提交时以 `note-too-large` 失败；弹窗带着失败码保持打开。
+弹窗是共用的 Modal 卡片，宽度按设计稿；设计稿里「包括当前对话的日志」复选框不做，因为日志随每个反馈事件一起投递，不是可选项。超长描述仍在提交时以 `note-too-large` 失败；弹窗保留草稿并通过警告 toast 展示本地化错误。
 
 ## 考虑过的替代方案
 
@@ -24,9 +24,9 @@ Web 客户端有两条互不相连的反馈路径，且都没有可见结果。`
 
 **在弹窗之外保留备注浮层。** 同一条备注有两个可达性不同的编辑器，会让该行在某些宽度下变成两行，正是当初引入浮层要避免的缺陷，而且设计稿只有两个拇指。
 
-**每个消息控件各自一个 Toast。** 输入框浮层已经按 Session 挂载一次，弹窗又拥有 toast 序号，因此一个持有者同时服务点赞路径与弹窗路径。
+**每个消息控件各自一个 Toast。** 输入框浮层已经按 Session 挂载一次，弹窗又拥有 toast 序号，因此一个持有者同时服务两种消息评分路径与 Session 弹窗。
 
-**在 `CommandUiSpec` 里新增 dialog 种类。** 一个消费 token 后运行客户端回调的 action 已经够用；PR #3745 为它的「文件」行引入了同一个 `action` 种类，后合并的一方保留一份定义即可。
+**在 `CommandUiSpec` 里新增 dialog 种类。** 一个消费 token 后运行客户端回调的 action 已经够用；「文件」行使用同一个 `action` 种类，一份定义即可服务两个条目。
 
 ## 后果
 

.agents/notes/implemented/feature/2026-09-08-shared-file-type-icons.i18n.yaml

@@ -2,5 +2,5 @@
 # side as of the last confirmed-consistent state. Both languages carry equal authority;
 # after editing either side, bring the other along and re-record with:
 #   pnpm run verify-translation-pairing --write .agents/notes/implemented/feature/2026-09-08-shared-file-type-icons.md
-2026-09-08-shared-file-type-icons.md: f8c87cb3259f757483a6658d79306d3f02d90488
-2026-09-08-shared-file-type-icons.zh.md: ea115cd63b3144ded2c99f818bb9364eea8adfef
+2026-09-08-shared-file-type-icons.md: 1e3fe7267961a8525bbcdf946eacaa933bcf3e4f
+2026-09-08-shared-file-type-icons.zh.md: fd1f38ebb33f80c73d17cc48f6f9640c309f97f2

.agents/notes/implemented/feature/2026-09-08-shared-file-type-icons.md

@@ -14,11 +14,11 @@ Client feature plugins can share React components only through `@deepseek-ai/dsh
 
 `FileTypeIcon` accepts a path, shared `IconProps`, the explicit `kind`, and an optional project-file snapshot. Traditional file types render the supplied 28px document and folder contours as inline SVG. Excel, Markdown, PDF, PPT, and Word foreground marks scale to 122% around their visual center; the remaining marked traditional glyphs use 112%, while the file body and folded corner retain their source geometry and the generic file has no invented center mark. The sheet is a solid category color, the foreground mark and ordinary folded corner are white, and the generic file has a darker grey corner. CSS assigns the supplied category palette through static design tokens: DeepSeek blue for code/HTML/Markdown, the lighter DeepSeek blue for Word, green for Excel, two amber steps for folder/PPT, red for PDF, and neutral grey for unknown files. Image and video share the supplied violet through a component-local variable because the design platform has no matching violet token. A caller may override a traditional sheet through `--dsh-file-type-icon-color`.
 
-Recognized code and configuration files render the corresponding 20px square artwork scaled to the requested icon size. These technology marks retain their embedded multicolor fills and are the explicit exception to the ordinary current-color icon rule. The map selects React before TypeScript/JavaScript, Angular filename suffixes before their base extension, Docker/Node/Git/Make/CMake by filename rules, and Flutter only when the optional project snapshot contains a `pubspec.yaml` whose text includes `flutter:`. Markdown and SVG remain owned by the traditional Markdown and image categories. CSV and TSV use the code glyph in file cards, rows, and preview titles; their clickable links also use code. Both `.env` and names ending in `.env` use the environment glyph. Every traditional and technology SVG is `aria-hidden`, and the card, row, or button that owns the file identity supplies the accessible name.
+Recognized code and configuration files render the corresponding 20px square artwork scaled to the requested icon size. The embedded static table contains exactly the 48 established `CodeFileType` entries; archive-only artwork does not add a category, and an adjacent manifest records the responsible design owner and source digests. Tests reject scripts, event attributes, external references, and duplicate ids in that table. `CodeFileIcon` replaces local SVG ids with a per-component prefix before insertion so repeated gradients and clip paths remain independent. These technology marks retain their embedded multicolor fills and are the explicit exception to the ordinary current-color icon rule. The map selects React before TypeScript/JavaScript, Angular filename suffixes before their base extension, Docker/Node/Git/Make/CMake by filename rules, and Flutter only when the optional project snapshot contains a `pubspec.yaml` whose text includes `flutter:`. Markdown and SVG remain owned by the traditional Markdown and image categories. CSV and TSV use the code glyph in file cards, rows, and preview titles; their clickable links also use code. Both `.env` and names ending in `.env` use the environment glyph. Every traditional and technology SVG is `aria-hidden`, and the card, row, or button that owns the file identity supplies the accessible name.
 
 `LinkIcon` delegates extension classification to `classifyFileType` and folds the detailed result into its existing link vocabulary: code and HTML use `code`, images use `image`, PDF/Word/Excel/PPT use `document`, and Markdown/video/unknown files use `other`. Extensionless names remain `other` in link contexts, so the 14px clickable-link appearance defined by the [clickable-link decision](2026-09-04-web-clickable-link-styles.md) does not change.
 
-Attachment upload cards, sent-message file cards, queued-file rows, and workspace file rows render `FileTypeIcon`. The Files tab title renders its explicit `folder` kind at 16px. Explicit delivery cards also use `FileTypeIcon` at 28px and `fileExtension` for their fallback metadata. The two metadata rows use `fileExtension` rather than local parsers; a leading-dot basename such as `.env` therefore displays `ENV`, while an absent or trailing suffix displays no extension label. Image content continues to render as a preview rather than a file-type glyph, and produced-file links and Markdown file mentions continue to use `LinkIcon` because they are link surfaces.
+Attachment upload cards, sent-message file cards, queued-file rows, and workspace file rows render `FileTypeIcon`. The Files tab title renders its explicit `folder` kind at 16px. Explicit delivery cards use `FileTypeIcon` at 20px and `fileExtension` for their fallback metadata. The two metadata rows use `fileExtension` rather than local parsers; a leading-dot basename such as `.env` therefore displays `ENV`, while an absent or trailing suffix displays no extension label. Image content continues to render as a preview rather than a file-type glyph, and produced-file links and Markdown file mentions continue to use `LinkIcon` because they are link surfaces.
 
 ## Alternatives considered
 
@@ -32,12 +32,13 @@ Attachment upload cards, sent-message file cards, queued-file rows, and workspac
 
 ## Testing
 
-The `ui-primitives` specs cover case-insensitive suffixes, both path separators, leading-dot files, missing and trailing suffixes, common named files, unknown fallback, all detailed code mappings, rule priority, Flutter context, every supplied technology SVG, instance-safe gradient ids, `aria-hidden`, sizing/class forwarding, distinct artwork, the 112% and 122% foreground-mark transforms, and the traditional solid-sheet/contrast-mark layers without literal SVG colors. A stylesheet spec pins every traditional category-to-color mapping, the caller override, and the local violet value. The existing `LinkIcon` classification table pins its coarse output, including `Makefile` remaining `other`. Attachment, chat, queue, and sidebar component suites exercise the migrated render paths; their accessibility output does not change because the glyphs remain decorative.
+The `ui-primitives` specs cover case-insensitive suffixes, both path separators, leading-dot files, missing and trailing suffixes, common named files, unknown fallback, all detailed code mappings, rule priority, Flutter context, the exact 48-key artwork set, rejected archive-only categories, static-markup safety, instance-safe SVG ids and references, `aria-hidden`, sizing/class forwarding, distinct artwork, the 112% and 122% foreground-mark transforms, and the traditional solid-sheet/contrast-mark layers without literal SVG colors. A stylesheet spec pins every traditional category-to-color mapping, the caller override, and the local violet value. The existing `LinkIcon` classification table pins its coarse output, including `Makefile` remaining `other`. Attachment, chat, queue, and sidebar component suites exercise the migrated render paths; their accessibility output does not change because the glyphs remain decorative.
 
 ## Consequences
 
 - Client packages use one filename parser and one detailed file-type table instead of importing or recreating feature-local logic.
 - A new suffix joins the detailed table only when an existing glyph truthfully represents it. If its link category differs from the current adapter, the change must also decide whether the 14px link appearance changes.
 - Code and configuration artwork preserves its embedded palette and does not accept the traditional `--dsh-file-type-icon-color` override.
+- The fixed 48-entry artwork table adds about 35 kB uncompressed and 17 kB gzip to the shared browser bundle; adding categories must justify that static baseline cost.
 - The primitive owns no copy but does own the default file-type palette. Consumers continue to own accessible labels and surrounding text, and may replace the category color through `--dsh-file-type-icon-color`.
 - The detailed category names describe presentation, not MIME validation. A suffix is a display hint and does not establish file contents or trust.

.agents/notes/implemented/feature/2026-09-08-shared-file-type-icons.zh.md

@@ -14,11 +14,11 @@ Status: implemented
 
 `FileTypeIcon` 接受路径、共享 `IconProps`、显式 `kind`和可选的项目文件快照。传统文件类型把所提供的 28px 文档与文件夹轮廓渲染为 inline SVG。Excel、Markdown、PDF、PPT、Word 的前景标记围绕自身视觉中心缩放至 122%，其余带标记的传统图形使用 112%；文件底板与折角保持源图几何，通用文件不凭空增加中心标记。底板使用实色分类颜色，前景标记与普通折角使用白色，通用文件使用较深的灰色折角。CSS 通过静态设计 token 分配所提供的分类调色板：code/HTML/Markdown 使用 DeepSeek 蓝，Word 使用较浅的 DeepSeek 蓝，Excel 使用绿色，folder/PPT 使用两档琥珀色，PDF 使用红色，未知文件使用中性灰。image 与 video 通过组件本地变量共用所提供的紫色，因为设计平台没有匹配的紫色 token。调用方可通过 `--dsh-file-type-icon-color` 覆盖传统底板颜色。
 
-已识别的代码与配置文件把对应的 20px 方形图稿缩放到请求的图标尺寸。这些技术标记保留自身内嵌的多色填充，是普通 current-color 图标规则的明确例外。映射让 React 优先于 TypeScript/JavaScript、Angular 文件名后缀优先于基础扩展名，并按文件名识别 Docker/Node/Git/Make/CMake；只有可选项目快照包含内容带 `flutter:` 的 `pubspec.yaml` 时才选择 Flutter。Markdown 与 SVG 仍由传统 Markdown 和图片类别拥有。CSV 和 TSV 在文件卡片、文件行及预览标题中使用 code 图标，其可点击链接也使用 code。`.env` 和以 `.env` 结尾的文件名均使用环境配置图标。所有传统与技术 SVG 都是 `aria-hidden` 的，拥有文件身份的卡片、行或按钮提供无障碍名称。
+已识别的代码与配置文件把对应的 20px 方形图稿缩放到请求的图标尺寸。内嵌静态表只包含现有 48 个 `CodeFileType` 条目；资源包中额外的图稿不会新增类别，相邻 manifest 记录负责的设计归属方与来源摘要。测试会拒绝该表中的脚本、事件属性、外部引用与重复 id。`CodeFileIcon` 在插入前为本地 SVG id 加上组件实例前缀，使重复渐变与裁剪路径互不干扰。这些技术标记保留自身内嵌的多色填充，是普通 current-color 图标规则的明确例外。映射让 React 优先于 TypeScript/JavaScript、Angular 文件名后缀优先于基础扩展名，并按文件名识别 Docker/Node/Git/Make/CMake；只有可选项目快照包含内容带 `flutter:` 的 `pubspec.yaml` 时才选择 Flutter。Markdown 与 SVG 仍由传统 Markdown 和图片类别拥有。CSV 和 TSV 在文件卡片、文件行及预览标题中使用 code 图标，其可点击链接也使用 code。`.env` 和以 `.env` 结尾的文件名均使用环境配置图标。所有传统与技术 SVG 都是 `aria-hidden` 的，拥有文件身份的卡片、行或按钮提供无障碍名称。
 
 `LinkIcon` 委托 `classifyFileType` 做扩展名分类，再把精细结果折叠进原有链接词汇：code 与 HTML 使用 `code`，图片使用 `image`，PDF/Word/Excel/PPT 使用 `document`，Markdown、video 与未知文件使用 `other`。无扩展名文件在链接语境中仍是 `other`，因此[可点击链接决策](2026-09-04-web-clickable-link-styles.zh.md)定义的 14px 外观不变。
 
-附件上传卡片、已发送消息文件卡片、排队文件行和工作区文件行渲染 `FileTypeIcon`。Files 标签页标题使用显式的 `folder` 类别，尺寸为 16px。显式交付卡片也使用 28px 的 `FileTypeIcon`，并通过 `fileExtension` 提供默认元数据。两处元数据行使用 `fileExtension`，不再保留本地解析器；`.env` 这样的前导点 basename 会显示 `ENV`，无后缀或末尾点号则不显示扩展名 label。图片内容继续渲染为预览而不是文件类型图形，产物文件链接与 Markdown 文件提及继续使用 `LinkIcon`，因为它们属于链接表面。
+附件上传卡片、已发送消息文件卡片、排队文件行和工作区文件行渲染 `FileTypeIcon`。Files 标签页标题使用显式的 `folder` 类别，尺寸为 16px。显式交付卡片使用 20px 的 `FileTypeIcon`，并通过 `fileExtension` 提供默认元数据。两处元数据行使用 `fileExtension`，不再保留本地解析器；`.env` 这样的前导点 basename 会显示 `ENV`，无后缀或末尾点号则不显示扩展名 label。图片内容继续渲染为预览而不是文件类型图形，产物文件链接与 Markdown 文件提及继续使用 `LinkIcon`，因为它们属于链接表面。
 
 ## 备选方案
 
@@ -32,12 +32,13 @@ Status: implemented
 
 ## 测试
 
-`ui-primitives` 测试覆盖不区分大小写的后缀、两种路径分隔符、前导点文件、无后缀与末尾点号、常见具名文件、未知回退、全部细分代码映射、规则优先级、Flutter 上下文、每一份技术 SVG、实例安全的渐变 id、`aria-hidden`、尺寸/class 转发、不同图稿、112% 与 122% 前景标记变换，以及不含 SVG 字面颜色的传统实色底板/对比标记层。样式表测试钉住每一项传统类别到颜色的映射、调用方覆盖变量与本地紫色值。既有 `LinkIcon` 分类表钉住它的粗粒度输出，包括 `Makefile` 仍为 `other`。附件、聊天、队列和侧边栏组件测试覆盖迁移后的渲染路径；由于图形仍是装饰性的，其无障碍输出不变。
+`ui-primitives` 测试覆盖不区分大小写的后缀、两种路径分隔符、前导点文件、无后缀与末尾点号、常见具名文件、未知回退、全部细分代码映射、规则优先级、Flutter 上下文、精确的 48 键图稿集、被拒绝的资源包额外类别、静态 markup 安全性、实例安全的 SVG id 与引用、`aria-hidden`、尺寸/class 转发、不同图稿、112% 与 122% 前景标记变换，以及不含 SVG 字面颜色的传统实色底板/对比标记层。样式表测试钉住每一项传统类别到颜色的映射、调用方覆盖变量与本地紫色值。既有 `LinkIcon` 分类表钉住它的粗粒度输出，包括 `Makefile` 仍为 `other`。附件、聊天、队列和侧边栏组件测试覆盖迁移后的渲染路径；由于图形仍是装饰性的，其无障碍输出不变。
 
 ## 后果
 
 - 客户端包使用一个文件名解析器与一份精细文件类型表，不再 import 或重新实现功能包本地逻辑。
 - 新后缀只在已有图形能准确表达它时加入精细表。若它的链接类别与当前适配不同，这次改动还必须决定是否改变 14px 链接外观。
 - 代码与配置图稿保留内嵌调色板，不接受传统图形的 `--dsh-file-type-icon-color` 覆盖。
+- 固定的 48 项图稿表为共享浏览器 bundle 增加约 35 kB 未压缩体积和 17 kB gzip 体积；新增类别必须证明这份静态基线成本是必要的。
 - primitive 不拥有文案，但拥有默认文件类型调色板。消费方继续拥有无障碍 label 与周围文字，并可通过 `--dsh-file-type-icon-color` 替换分类颜色。
 - 精细类别名称描述展示意图，不是 MIME 校验。后缀只是展示提示，不能证明文件内容或可信度。

.agents/notes/implemented/feature/2026-09-10-connection-indicator-refinements.i18n.yaml

@@ -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/feature/2026-09-10-connection-indicator-refinements.md
+2026-09-10-connection-indicator-refinements.md: 3b6b4c7fd0962edbb87be17c2fa45eded46ee86b
+2026-09-10-connection-indicator-refinements.zh.md: 0bd5dc3e0b5b69cf2d9491e2d3ce5711c96898f5

.agents/notes/implemented/feature/2026-09-10-connection-indicator-refinements.md

@@ -0,0 +1,29 @@
+# Agent Note: Connection indicator state and interaction refinements
+
+Status: implemented
+
+English | [中文](2026-09-10-connection-indicator-refinements.zh.md)
+
+## Problem
+
+The sidebar connection pill hid its affordance behind a hover swap: outage and retry-attempt states replaced their label with **Reconnect now** on hover or focus, so every state had to reserve the widest supplied label to keep the control from resizing. A retry that resolved in under a second flickered the connecting pill in and out, and state changes and unmounts jumped with no transition.
+
+## Decision
+
+**The disconnected pill shows its action statically.** [ConnectionIndicator.tsx](../../../../packages/client/ui-primitives/src/ConnectionIndicator.tsx) renders a permanent retry glyph (`IconRefreshOutline14`) beside the outage copy (`连接异常，刷新重试` / `Disconnected`; the Chinese copy also names the retry action); clicking the pill still reconnects immediately. The hover label swap and the hidden widest-label size-reservation spans are gone, so the pill sizes to its current label. The connecting state shows a rotating-arc spinner instead of the exclamation glyph. Appearance and removal fade over 150ms — swaps between visible states replace content in place: `EXIT_MS` delays unmount to match the stylesheet's `.leaving` transition, and `prefers-reduced-motion` disables every animation and transition. Chrome settles at 28px height, 8px horizontal padding, 4px icon gap, 13px radius, and a 1px border of the label color at 20% alpha.
+
+**The shell owns attempt pacing.** [SettingsRoot.tsx](../../../../packages/client/ui-settings-general/src/client/SettingsRoot.tsx) keeps the connecting pill visible for at least `CONNECTING_MIN_VISIBLE_MS` (800ms) so sub-second retries do not flicker; every attempt, manual or automatic, reads the one label `重新连接中` (`connection.connecting`). The two-second recovery confirmation (`RECOVERY_CONFIRMATION_MS`) starts when the recovered pill becomes visible, so a hold that delays its appearance never shortens the confirmation. Both timings are built-in presentation constants of their owners, not configuration.
+
+## Alternatives considered
+
+**Animating width changes.** A FLIP-style measured pixel transition (remember the old width, pin it, transition to the new measurement) needs a layout effect and imperative style writes; the fade-only change reads calm enough without them.
+
+**Swapping to the retry glyph only on hover.** Showing the retry glyph permanently states the affordance without requiring any pointer interaction, matching the static label; a hover cross-fade adds interaction-dependent state and conveys nothing extra.
+
+**Scaling on enter/exit.** A 0.98 scale beside the opacity fades reads as jitter at 12px text, so only opacity animates.
+
+**Naming manual and automatic attempts differently.** `ConnectionController.emitState` deduplicates repeated `connecting` states across backoff attempts, so the shell cannot observe attempt boundaries: a shell-held manual-retry flag either flips the label mid-hold or sticks across later automatic attempts. Distinguishing the copy correctly requires the connection layer to expose the attempt origin, which this change does not need — both attempt kinds read the same label.
+
+## Consequences
+
+`ConnectionIndicator`'s `reconnectLabel` prop and its size-reservation spans are removed from the pre-stable API; the sole consumer (`ui-settings-general`) is updated in the same change. `settings-root.client.spec.tsx` pins the 800ms hold, the single attempt label held steady through the hold, and the visibility-based confirmation window; `atoms.client.spec.tsx` pins the exit-duration unmount; `lifecycle-chrome.e2e.ts` and its ARIA golden replay the recovery flow in a real browser. Both packages' READMEs restate the interaction.

.agents/notes/implemented/feature/2026-09-10-connection-indicator-refinements.zh.md

@@ -0,0 +1,29 @@
+# Agent Note: 连接指示器状态与交互细化
+
+Status: implemented
+
+[English](2026-09-10-connection-indicator-refinements.md) | 中文
+
+## Problem
+
+侧边栏连接药丸把操作提示藏在悬停切换里：断连与重试状态在悬停或聚焦时把文案替换为**立即重连**，因此每个状态都要为最宽的 label 预留空间以避免控件变形。一次不到一秒就恢复的重试会让连接中药丸闪现闪没，状态切换和消失也没有任何过渡、十分突兀。
+
+## Decision
+
+**断连药丸静态地展示其动作。** [ConnectionIndicator.tsx](../../../../packages/client/ui-primitives/src/ConnectionIndicator.tsx) 在断连文案旁常驻渲染重试图形（`IconRefreshOutline14`），文案为 `连接异常，刷新重试` / `Disconnected`（中文文案同时点明重试动作）；点击药丸仍会立即重连。悬停换文案和隐藏的最宽 label 占位 span 全部移除，药丸宽度随当前 label 自适应。连接中状态改用旋转圆弧 spinner 取代感叹号图形。出现与移除以 150ms 淡入淡出——可见状态之间的切换则原地替换内容：`EXIT_MS` 延迟卸载以匹配样式表的 `.leaving` 过渡，`prefers-reduced-motion` 会禁用全部动画与过渡。外观定为高 28px、水平内边距 8px、图标间距 4px、圆角 13px，以及 label 颜色 20% 透明度的 1px 边框。
+
+**外壳拥有尝试节奏。** [SettingsRoot.tsx](../../../../packages/client/ui-settings-general/src/client/SettingsRoot.tsx) 让连接中药丸至少可见 `CONNECTING_MIN_VISIBLE_MS`（800ms），亚秒级重试不再闪动；无论手动还是自动，每次尝试都显示同一个文案`重新连接中`（`connection.connecting`）。2 秒恢复确认（`RECOVERY_CONFIRMATION_MS`）从恢复药丸实际可见时起算，驻留推迟其出现也不会缩短确认时长。两个时长是各自持有方的内置展示常量，不是配置。
+
+## Alternatives considered
+
+**给宽度变化加动画。** FLIP 式的像素测量过渡（记住旧宽度、钉住、过渡到新测量值）需要一个 layout effect 和命令式样式写入；纯淡入淡出已足够平静，无需这些。
+
+**仅在悬停时切换为重试图形。** 常驻显示重试图形无需任何指针交互就说明了操作，与静态文案一致；悬停交叉渐变引入依赖交互的状态却不传达更多信息。
+
+**进出场缩放。** 0.98 的缩放叠加在透明度淡入淡出上，在 12px 文字上读起来像抖动，因此只保留透明度。
+
+**手动与自动尝试使用不同命名。** `ConnectionController.emitState` 会去重退避尝试之间重复的 `connecting` 状态，外壳观察不到尝试边界：外壳自持的手动重试标志要么在驻留中途翻转文案，要么在后续自动尝试中一直滞留。要正确区分文案需要连接层暴露尝试来源，而本次变更并不需要——两种尝试显示同一文案。
+
+## Consequences
+
+`ConnectionIndicator` 的 `reconnectLabel` prop 及其占位 span 从 pre-stable API 中移除；唯一消费者（`ui-settings-general`）在同一变更中更新。`settings-root.client.spec.tsx` 固定 800ms 驻留、驻留期间保持不变的单一尝试文案，以及按可见时刻起算的确认窗口；`atoms.client.spec.tsx` 固定退出时长后的卸载；`lifecycle-chrome.e2e.ts` 及其 ARIA golden 在真实浏览器中回放恢复流程。两个包的 README 重述了该交互。

.agents/notes/implemented/feature/2026-09-10-symmetric-message-feedback-submission.i18n.yaml

@@ -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/feature/2026-09-10-symmetric-message-feedback-submission.md
+2026-09-10-symmetric-message-feedback-submission.md: 51c754dbd3ecc566e15ec1b7a276fc2a3cf7c551
+2026-09-10-symmetric-message-feedback-submission.zh.md: 0ba5d1a1c53f08b76c5bbc32b84687c3dcd38ec2

.agents/notes/implemented/feature/2026-09-10-symmetric-message-feedback-submission.md

@@ -0,0 +1,25 @@
+# Agent Note: Symmetric message feedback submission
+
+Status: implemented
+
+English | [中文](2026-09-10-symmetric-message-feedback-submission.zh.md)
+
+## Problem
+
+The assistant-message rating controls used different commit points. Like recorded a positive rating immediately, while Dislike opened the feedback dialog and recorded only after Submit. The asymmetry made an accidental Like durable before confirmation and prevented positive feedback from carrying the same optional category and description as negative feedback.
+
+## Decision
+
+The [feedback dialog and categories](2026-09-08-feedback-dialog-and-categories.md) decision owns the shared form and durable taxonomy. Both unrecorded ratings open the shared feedback dialog and record only after Submit. A message `FeedbackDialogTarget` carries the selected `positive` or `negative` rating, and `FeedbackSurface` passes that rating with the dialog entry to `MessageFeedbackController.rate`. The action row reads the committed item before either action: clicking its current rating calls the injected `retract` operation, while clicking an absent or opposite rating opens the dialog. `retract` rechecks the committed rating inside the controller's serialized mutation queue and becomes a no-op after a concurrent change, so it cannot turn stale UI intent into a bare rating put. The dialog remains optional-input: submitting without a category or description records the selected rating and raises the acknowledgement toast; dismissing it records nothing.
+
+## Alternatives considered
+
+**Keep Like as an immediate action.** This preserves one fewer click for positive feedback, but keeps two submission models beside each other and prevents positive reports from carrying context.
+
+**Require the dialog to retract a recorded rating.** Retraction has no category or description to collect, and an extra confirmation would make the existing undo action less direct.
+
+**Create separate positive and negative forms.** The fields, validation, failures, and acknowledgement are identical; carrying the rating in the existing target keeps one draft and submission lifecycle.
+
+## Consequences
+
+Neither rating creates a `feedback/message-put` event until the user submits the dialog. Positive and negative records can both include a category and note, while clicking the active rating continues to create `feedback/message-delete` without opening the dialog. Unit coverage pins the shared action path and target routing, and the keyless Web scenarios submit both ratings through the real dialog before checking durable events and telemetry release.

.agents/notes/implemented/feature/2026-09-10-symmetric-message-feedback-submission.zh.md

@@ -0,0 +1,25 @@
+# Agent Note: 对称消息反馈提交
+
+Status: implemented
+
+[English](2026-09-10-symmetric-message-feedback-submission.md) | 中文
+
+## 问题
+
+助手消息的两个评分控件使用不同的提交时点。点赞会立即记录正面评分，点踩则先打开反馈弹窗，仅在用户提交后记录。这种不对称会让误触点赞在确认前就持久化，也让正面反馈无法携带与负面反馈相同的可选分类和描述。
+
+## 决策
+
+[反馈弹窗与分类](2026-09-08-feedback-dialog-and-categories.zh.md)决策负责共用表单和持久化分类表。两种未记录的评分都打开共用反馈弹窗，并且只在提交后记录。消息 `FeedbackDialogTarget` 携带所选的 `positive` 或 `negative` 评分，`FeedbackSurface` 将该评分与弹窗条目一起传给 `MessageFeedbackController.rate`。动作行会在任一操作前读取已提交条目：点击当前评分会调用注入的 `retract` 操作，点击未记录或相反评分则打开弹窗。`retract` 会在控制器的串行变更队列内重新检查已提交评分，并在并发变更后变为无操作，因此不会把陈旧的 UI 意图转成裸评分 put。弹窗的输入仍可全部留空：不选分类也不填描述时提交会记录所选评分并弹出确认 toast；关闭弹窗不会记录任何内容。
+
+## 考虑过的替代方案
+
+**保留点赞立即提交。** 这能让正面反馈少一次点击，但会让并列的两个评分继续使用不同提交模式，也无法让正面反馈携带上下文。
+
+**撤回已记录评分也必须经过弹窗。** 撤回没有需要收集的分类或描述，多一次确认会让现有撤销操作变得不够直接。
+
+**分别创建正面和负面表单。** 两者的字段、校验、失败处理与确认完全相同；在现有目标中携带评分即可共用一套草稿与提交生命周期。
+
+## 后果
+
+用户提交弹窗前，两种评分都不会创建 `feedback/message-put` 事件。正面与负面记录都可以包含分类和备注，而点击当前评分仍会直接创建 `feedback/message-delete`，无需打开弹窗。单元测试固定共用动作路径与目标路由，无密钥 Web 场景则通过真实弹窗提交两种评分，再检查持久事件与遥测投递。

.agents/notes/implemented/process/2026-09-09-parallel-macos-notarization.i18n.yaml

@@ -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/process/2026-09-09-parallel-macos-notarization.md
+2026-09-09-parallel-macos-notarization.md: 152da4661207bf130389e600c16398335fd9fe70
+2026-09-09-parallel-macos-notarization.zh.md: d5375e39a6348b49ffe5bb93c2c056151d5abe75

.agents/notes/implemented/process/2026-09-09-parallel-macos-notarization.md

@@ -0,0 +1,37 @@
+# Agent Note: Parallel macOS notarization from isolated App copies
+
+Status: implemented
+
+English | [中文](2026-09-09-parallel-macos-notarization.zh.md)
+
+## Problem
+
+The Desktop release distributes a DMG for installation and a ZIP for updates. Waiting for App notarization before creating the DMG serializes two Apple submissions. A proxy improves upload throughput but does not overlap the independent service waits. Stapling modifies the App, so concurrent notarization and packaging cannot safely share that writable directory.
+
+## Decision
+
+The fixed-target installer command signs and verifies one App, then creates two independent copies with `ditto`. The App lane notarizes and staples its copy, verifies its signature, ticket, and Gatekeeper acceptance, and asks electron-builder to create the ZIP and updater metadata. The DMG lane immediately packages its copy, signs the image, and uses the existing artifact-completion hook to notarize, staple, and verify the image. Each electron-builder process receives the actual `.app` path through `--prepackaged`, an isolated output directory, and `--publish never`.
+
+The ZIP contains an individually stapled App. The DMG contains the signed App without an individually stapled ticket; its outer ticket covers the nested code, following Apple's [container guidance](https://developer.apple.com/documentation/xcode/packaging-mac-software-for-distribution). Apple describes [ticket ingestion when Gatekeeper checks the outer container](https://developer.apple.com/forums/thread/125512). Independent extraction of the unstapled App relies on an online or cached ticket; the ZIP supplies an embedded ticket. The directory-only command continues to notarize and staple its App.
+
+Both lanes settle before error propagation or temporary-directory cleanup. Only two successful lanes allow promotion of the DMG, ZIP, ZIP blockmap, and channel metadata. The stapled App replaces the signed directory build, and the caller writes the release completion record last. An error leaves that record absent, so the existing upload validation rejects the incomplete release. Separate output directories also prevent concurrent writes to electron-builder diagnostics and channel metadata.
+
+This refines the notarization ordering in the [Desktop packaging decision](../architecture/2026-08-25-electron-desktop-packaging-and-updates.md); that note remains the owner of release identity, signatures, update ownership, and publishing requirements.
+
+## Alternatives considered
+
+**Share one App between both lanes.** A DMG reader can overlap with `stapler` writes, producing a nondeterministic bundle. Independent copies keep the submitted and distributed bytes stable within each lane.
+
+**Only notarize the DMG.** ZIP updates are distributed independently and need a stapled App. Retaining both submissions keeps that independent qualification explicit.
+
+**Keep serial notarization and only use a proxy.** The same 214.84 MiB DMG uploads in 203.83 seconds directly and 38.59 seconds through the tested system proxy, but neither route removes the serial dependency between Apple submissions. Proxy configuration remains a build-host concern; the packaging script does not change host network settings.
+
+**Run both targets in one electron-builder call before App notarization finishes.** The ZIP must read the stapled copy. Separate prepackaged invocations preserve electron-builder's own archive, blockmap, and metadata implementation without changing its target scheduling or patching the dependency.
+
+## Consequences
+
+On 2026-09-09, full arm64 packaging on the same Mac through the same system proxy took 490.78 seconds with parallel notarization versus 751.33 seconds serially, a 34.7% reduction. The artifact lanes began 8 milliseconds apart and completed in 307.36 seconds for App/ZIP and 268.54 seconds for DMG. Apple accepted both submissions; their uploads completed about one second apart. Each configuration has one full-build sample, so cache and Apple queue variation prevent attributing the entire difference to concurrency.
+
+Two temporary App copies and separate artifact directories increase peak disk usage. Two uploads can contend for network bandwidth, and Apple can queue either submission independently; phase timings describe observed behavior rather than a CI latency budget. A failed lane waits for the other lane to finish before cleanup, which can delay failure reporting but avoids deleting files still owned by a child process.
+
+The [orchestration tests](../../../../apps/desktop/tests/package-macos.spec.ts) use barriers to prove overlap, ticket isolation, both-error collection, and refusal to promote incomplete artifacts. A controlled serial regression fails the overlap assertion. Real signed macOS packaging, extracted ZIP verification, DMG integrity and nested signature checks, and final upload-plan validation qualify the platform tools; cross-version installed updates and offline installation on a clean Mac remain release qualification work.

.agents/notes/implemented/process/2026-09-09-parallel-macos-notarization.zh.md

@@ -0,0 +1,37 @@
+# Agent Note: 基于隔离 App 副本的并行 macOS 公证
+
+Status: implemented
+
+[English](2026-09-09-parallel-macos-notarization.md) | 中文
+
+## 问题
+
+Desktop 发布同时提供用于安装的 DMG 和用于更新的 ZIP。等待 App 公证完成后才创建 DMG，会让两次 Apple 提交串行执行。代理可以提高上传吞吐量，但不能让两个独立的服务等待过程重叠。钉票会修改 App，因此并发公证和打包不能安全地共享同一个可写目录。
+
+## 决策
+
+固定目标安装包命令先签名并验证一个 App，再通过 `ditto` 创建两个独立副本。App 路线公证其副本并钉票，验证签名、票据与 Gatekeeper 接受状态，再由 electron-builder 生成 ZIP 和更新元数据。DMG 路线立即封装其副本、签署映像，再通过现有 artifact-completion hook 公证映像、钉票并验证。每个 electron-builder 进程都通过 `--prepackaged` 接收真正的 `.app` 路径、独立的输出目录和 `--publish never`。
+
+ZIP 包含已单独钉票的 App。DMG 包含已签名但未单独附加票据的 App；根据 Apple 的[容器说明](https://developer.apple.com/documentation/xcode/packaging-mac-software-for-distribution)，外层票据覆盖内嵌代码。Apple 还说明了 [Gatekeeper 检查外层容器时接收票据的行为](https://developer.apple.com/forums/thread/125512)。单独提取未钉票 App 依赖在线或缓存票据；ZIP 则提供内嵌票据。仅生成目录的命令仍会公证 App 并钉票。
+
+错误传播和临时目录清理前必须等待两路均结束。只有两路都成功，才允许移入 DMG、ZIP、ZIP blockmap 和频道元数据。已钉票的 App 替换签名目录构建，调用方最后写入发布完成记录。发生错误时该记录保持缺失，现有上传校验因而会拒绝不完整发布。独立输出目录还避免了 electron-builder 诊断文件与频道元数据的并发写入。
+
+本决策细化了 [Desktop 打包决策](../architecture/2026-08-25-electron-desktop-packaging-and-updates.zh.md)中的公证顺序；原决策继续负责发布身份、签名、更新归属与发布要求。
+
+## 考虑过的替代方案
+
+**两路共享一个 App。** DMG 读取可能与 `stapler` 写入重叠，使 bundle 内容不确定。独立副本使每条路线中提交与分发的字节保持稳定。
+
+**只公证 DMG。** ZIP 更新独立分发，需要已钉票的 App。保留两次提交可以明确维持这项独立验收。
+
+**保留串行公证，仅使用代理。** 同一个 214.84 MiB DMG 直连上传耗时 203.83 秒，通过所测系统代理上传耗时 38.59 秒，但两种网络路径都不能消除 Apple 提交间的串行依赖。代理配置仍由构建主机负责；打包脚本不修改主机网络设置。
+
+**App 公证结束前，在同一次 electron-builder 调用中运行两个目标。** ZIP 必须读取已钉票副本。独立的 prepackaged 调用可以保留 electron-builder 自己的归档、blockmap 和元数据实现，无需修改目标调度或给依赖打补丁。
+
+## 影响
+
+2026-09-09，在同一台 Mac、同一系统代理下，arm64 完整打包在并行公证时耗时 490.78 秒，串行时耗时 751.33 秒，缩短 34.7%。两条产物路线相隔 8 毫秒启动，App/ZIP 耗时 307.36 秒，DMG 耗时 268.54 秒。Apple 接受了两次提交，两者上传完成时间仅相差约一秒。每种配置只有一次完整构建样本，缓存与 Apple 队列变化使我们不能将全部差值归因于并发。
+
+两个临时 App 副本与独立产物目录增加了磁盘峰值占用。两次上传可能争用网络带宽，Apple 也可能分别排队处理；阶段计时记录实际行为，不构成 CI 延迟预算。一路失败后会等待另一路结束再清理，这可能延迟错误报告，但能避免删除仍由子进程持有的文件。
+
+[编排测试](../../../../apps/desktop/tests/package-macos.spec.ts)通过同步屏障验证重叠执行、票据隔离、收集两路错误，以及拒绝移入不完整产物。受控的串行回归会使重叠断言失败。真实签名 macOS 打包、ZIP 解压后验证、DMG 完整性与内嵌签名检查、最终上传计划验证用于验收平台工具；跨版本已安装应用更新和干净 Mac 上的离线安装仍属于发布验收工作。

.agents/notes/implemented/simplification/2026-09-09-nontransactional-loader.i18n.yaml

@@ -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/simplification/2026-09-09-nontransactional-loader.md
+2026-09-09-nontransactional-loader.md: 7ea4db0abf0fb65e443126dbca9fd822e47ce2d0
+2026-09-09-nontransactional-loader.zh.md: 5293efb58998f68a25e1143a2d043da32ef06bfb

.agents/notes/implemented/simplification/2026-09-09-nontransactional-loader.md

@@ -0,0 +1,35 @@
+# Agent Note: Keep Loader mutations non-transactional
+
+Status: implemented
+
+English | [中文](2026-09-09-nontransactional-loader.zh.md)
+
+## Problem
+
+Transactional config reload preserves an old plugin generation after a failed edit, but requires Loader to own candidate imports, lifecycle settlement, rollback, option identity, and Include serialization. These changes make the vendored implementation substantially different from its pinned sources. Application startup and profile patch watching also depend on that settlement implicitly.
+
+## Decision
+
+Revert the five commits in [#932](https://github.com/deepseek-harness/deepseek-harness/pull/932), resolving package moves and retaining independent later behavior. The reported merge commit belongs to the larger #936 dependency chain; reverting its first-parent diff would remove unrelated repository-plugin support. The [vendor ledger](../../../../vendor/README.md#local-modifications) records every retained source change against the unchanged pins.
+
+Loader changes entry options eagerly. EntryGroup starts siblings concurrently and logs application failures; EntryTree waits for outstanding work without rejecting failed fibers. Neither restores a previous plugin or configuration. Include retains parse validation and patch reapplication, but plugin failures can leave a partially applied tree.
+
+Application consumers own their completion checks. The CLI waits for its fallback HMR service before installing live patch watchers. The directory chooser checks the entries it mounts. The chooser and browser package runner capture the first fiber-disposal result before removing the entry, then await it before reporting teardown complete. Preset mounting waits for its subtree and reports import, activation, and missing-service failures. [App boot](../../../../packages/boot/app-boot/README.md) owns exact patch-file watching, activation audits, and partial-context cleanup. Web startup audits activation before printing a URL or opening the browser. These adaptations preserve existing consumer behavior after the reverse patch.
+
+Fiber, Entry, and isolate keep their upstream update return behavior. App boot observes discarded restart promises through the existing `internal/update` waterfall and waits for fibers before auditing a patch reload. The detached import-completion observer handles both fiber outcomes; the fiber still retains its failure for an explicit audit. Durable Include writes drain before and after child removal so a later teardown write cannot erase an earlier terminal write failure.
+
+Two #932-specific vendor changes remain: awaited initial-file creation and forced rereading in Include, and Schemastery conditional exports. Restoring the pre-#932 debounced write/read sequence reproduces `ENOENT` in the missing-file initialization test. Keeping these two lines preserves the existing `initial` option without an application-side file writer or a second YAML serializer. Removing Schemastery exports reproduces `ERR_REQUIRE_ESM_RACE_CONDITION` while the Web preset suite boots: Node falls back to the CJS entry during concurrent ESM imports. The HMR injection decorators, conditional patch cloning, and update return values use the pre-#932 behavior. Explicit `workspace:^` dependencies make the #932 workspace-link switch and dedicated lockfile check unnecessary.
+
+## Alternatives considered
+
+**Keep transactional Loader updates.** They provide automatic recovery from a rejected plugin candidate, but retain the vendored lifecycle machinery being removed. Parse failures can be contained without plugin rollback.
+
+**Restore every vendored file verbatim.** This would also remove lazy injected config evaluation, conditional disabled entries, lifecycle disposal fixes, durable writes, and module-loader compatibility. Those changes have independent consumers and remain recorded in the vendor ledger.
+
+**Move generic rollback into app boot.** This would retain the same candidate-generation and restoration obligations under another owner. Applications instead report failures and allow a later valid edit to recover.
+
+## Consequences
+
+A plugin activation failure can leave the new options and a failed fiber in place. Callers that require active plugins must audit after settlement; awaiting `Loader.create()` alone does not establish activation. Automatic plugin rollback requires a separate future decision with evidence that its recovery benefit warrants the additional lifecycle implementation.
+
+[Live-patch tests](../testing/2026-09-09-user-patch-hmr-test-delivery.md) retain controlled event delivery and native watcher coverage, while asserting failure reporting without rollback. The [terminal-release policy](../bug-fix/2026-07-31-fail-loud-releases-the-terminal.md) remains applicable to fatal errors and partial boot teardown. Web preset composition and a real CLI webhook-created model Session provide the application-level verification beyond hand-mounted plugins.

.agents/notes/implemented/simplification/2026-09-09-nontransactional-loader.zh.md

@@ -0,0 +1,35 @@
+# Agent Note: 保持 Loader 更改非事务化
+
+Status: implemented
+
+[English](2026-09-09-nontransactional-loader.md) | 中文
+
+## 问题
+
+事务化配置重载会在编辑失败后保留旧插件代次，但要求 Loader 负责候选导入、生命周期结算、回滚、选项对象身份和 Include 串行化。这些更改使 vendor 实现与固定来源产生显著差异。应用启动和 profile patch 监视也隐式依赖该结算行为。
+
+## 决策
+
+撤销 [#932](https://github.com/deepseek-harness/deepseek-harness/pull/932) 中的五个提交，解决包移动冲突并保留后续独立行为。记录的合并提交属于更大的 #936 依赖链；撤销其第一父提交差异还会删除无关的仓库插件支持。[Vendor 修改记录](../../../../vendor/README.md#local-modifications) 按不变的固定来源记录每项保留的源码更改。
+
+Loader 立即更改条目选项。EntryGroup 并发启动同级条目并记录应用失败；EntryTree 等待未完成的工作，但不因失败的 fiber 而拒绝。两者均不恢复旧插件或配置。Include 保留解析校验和 patch 重应用，但插件失败可能留下部分应用的配置树。
+
+应用消费者负责完成检查。CLI 在安装实时 patch 监视器前等待其回退 HMR 服务。目录选择器检查其挂载的条目。选择器和浏览器包运行器在移除条目前取得首次 fiber 释放的结果，并等待该结果后才报告拆卸完成。预设挂载等待其子树，并报告导入、激活和缺少服务的失败。[应用启动](../../../../packages/boot/app-boot/README.zh.md) 负责精确 patch 文件监视、激活检查和部分上下文清理。Web 启动在打印 URL 或打开浏览器前检查激活状态。这些适配维持应用反向补丁后已有的消费者行为。
+
+Fiber、Entry 和 isolate 保持上游的更新返回行为。App boot 通过现有的 `internal/update` waterfall 观察被丢弃的重启 promise，并在检查 patch 重载前等待 fiber。游离的导入完成观察器处理 fiber 的两种结果；fiber 仍保留失败信息供显式检查。Include 的持久写入在删除子条目前后均排空，防止后续拆卸写入掩盖更早的终止性写入失败。
+
+保留两项 #932 专属 vendor 改动：Include 中等待初始文件创建并强制重新读取，以及 Schemastery 条件导出。恢复 #932 前的防抖写入和读取顺序，会在缺失文件初始化测试中复现 `ENOENT`。保留这两行可以维持已有 `initial` 选项，而无需在应用侧增加文件写入器或另一份 YAML 序列化逻辑。移除 Schemastery exports 后，Web preset 测试在启动时复现 `ERR_REQUIRE_ESM_RACE_CONDITION`：并发 ESM 导入使 Node 回退到 CJS 入口。HMR 注入装饰器、条件 patch 克隆及更新返回值采用 #932 前的行为。显式 `workspace:^` 依赖使 #932 的 workspace 链接开关与专用锁文件检查不再必要。
+
+## 考虑过的替代方案
+
+**保留事务化 Loader 更新。** 它们能从被拒绝的插件候选自动恢复，但会保留本次删除的 vendor 生命周期机制。解析失败可以独立于插件回滚进行处理。
+
+**逐字恢复所有 vendor 文件。** 这还会删除延迟注入配置求值、条件禁用条目、生命周期释放修复、持久写入和模块加载器兼容性。这些更改具有独立消费者，并继续记录在 vendor 修改记录中。
+
+**将通用回滚移入应用启动。** 这会在另一归属下保留相同的候选代次与恢复义务。应用改为报告失败，并允许后续有效编辑恢复。
+
+## 后果
+
+插件激活失败可能保留新选项和失败的 fiber。要求插件处于激活状态的调用者必须在结算后检查；仅等待 `Loader.create()` 不能证明激活。自动插件回滚需要后续独立决策，并证明其恢复收益值得额外的生命周期实现。
+
+[实时 patch 测试](../testing/2026-09-09-user-patch-hmr-test-delivery.zh.md) 保留受控事件投递和原生监视覆盖，同时断言不回滚时的失败报告。[终端释放策略](../bug-fix/2026-07-31-fail-loud-releases-the-terminal.zh.md) 仍适用于致命错误和部分启动拆卸。Web preset 组合与真实 CLI webhook 创建的模型 Session 提供手动挂载插件之外的应用级验证。

.agents/notes/implemented/testing/2026-09-06-client-assembly-test-line.i18n.yaml

@@ -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/testing/2026-09-06-client-assembly-test-line.md
+2026-09-06-client-assembly-test-line.md: 0d7ed86e4f0c8898441737c8d6a4c631586268a2
+2026-09-06-client-assembly-test-line.zh.md: 8b4381dbfa24819e1b1f086b7c318d388b681ef5

.agents/notes/implemented/testing/2026-09-06-client-assembly-test-line.md

@@ -0,0 +1,90 @@
+# Agent Note: Whole-client test tier over an endpoint-named Remote mock
+
+Status: implemented
+
+English | [中文](2026-09-06-client-assembly-test-line.zh.md)
+
+## Problem
+
+Browser feature specs each hand-build their bench: a bare Cordis context, stand-ins for `locale`, `connection`, and `remote`, and slot declarations the real declarer would have made. Their assertions therefore describe the bench, not the product: a plugin that adds a settings section, a declarer that reloads through the Loader, or a Connection that reconnects is invisible to them, and every bench repeats the same forty lines with small drift.
+
+API client specs drive their objects through a programmable fake of the Remote surface. The fake re-implements Gateway semantics it should only exercise: follow-stream opening snapshots derived from a history list, page cuts, stream pumps with delivery promises, and an envelope layer. Each of those is a second implementation of a contract the product already owns, and it lets tests describe behavior the generated client cannot produce, such as a unary call that rejects.
+
+No focused source test boots the client the way production does — `bootClient` creating one Loader entry per manifest row, then `mountClient` — so composition faults between plugins are hidden by hand-built benches.
+
+## Decision
+
+A whole-client tier lives in `@deepseek-ai/dsh-client-test-runtime` under the deep import `src/assembly/`, and a new test-support package `@deepseek-ai/dsh-remote-mock` answers Remote traffic by endpoint name. Both are described by their READMEs ([client-runtime](../../../../packages/test-support/client-runtime/README.md), [remote-mock](../../../../packages/test-support/remote-mock/README.md)); this note records the decisions behind them.
+
+**The roster is read from the bundles, never copied.** `bundleRoster(bundles)` parses each bundle's `dsh.bundle.patch` with the include plugin's own YAML dialect (`entryListSchema`, which carries `!!js`) and composes the layers with its `applyEntryPatches`, then keeps every enabled row whose package declares `dsh.client.platform === 'web'`, carrying that declaration's `inject` and `immediately`. `webApp` is the `web` profile's roster (`dsh-base`, then `dsh-web-app`), computed at import. A spec names what it tests and derives the rest: `webApp.closure([row])` keeps a row plus its transitive `inject` cone; `pick` and `without` exist for deliberate cuts. The test runtime stays a Client-face package: it imports no Host module, uses no dynamic import for this, and its client-face `types` adds `node` beside `client-build-environment` so the reader can use `node:fs`. `closure` treats the shell's static platform modules (`PLATFORM_MODULES`) as satisfied without a row.
+
+**The production boot path runs unchanged.** `TestClient.start(plan, mock)` installs the mock as the Connection carrier (`__DSH_TRANSPORT__ = { rpc: mock.rpc }`), loads each roster row's `/client` module, registers its factory through the production module facade's `pendingQueue`, boots through `bootClient`, optionally mounts through `mountClient`, and waits for `connected`. `reload(name)` rebuilds a Loader entry through client-hmr's exported `tearDownEntryFiber`; `unload(name)` removes it; `dispose()` tears down and then fails the test on any endpoint that had no rule. jsdom lacks `EventSource` and `ResizeObserver`; `start` installs inert stand-ins only where the global is absent. Boots and entry rebuilds run one at a time per worker, since the `connection` plugin reads the transport global at apply, and each installs the acting client's transport first; the transport and shims are held by reference count, the first client installing them and the last dispose restoring them, so overlapping clients in one test each connect to their own mock, also after a `reload` of the `connection` row.
+
+**`remote.<ns>` is a contract-free proxy, not the generated client.** The `@deepseek-ai/dsh-api-remotes` row is dropped because its generated clients exist only in built `lib/`. For every `remote.<ns>` a roster row injects, plus every namespace the mock has a rule for, the tier provides one Proxy: `ctx.remote.<ns>.<method>(...args)` calls the endpoint `<ns>/<method>` over the roster's own Connection with the positional args, as a stream when the mock registered a stream script for it and as a unary call otherwise. Cordis resolves `ctx.remote.<ns>` to the service `remote.<ns>`, so the Gateway client itself is untouched. A unary answer returns unchanged; a unary rejection folds the way the generated client folds a carrier throw, through the Gateway client's exported `carrierFailure` and `cancelledFailure`, so product code that fires a Remote call without awaiting sees no rejection. Stream items and failures pass through as the stream yields them.
+
+**Native mocks own response configuration and call assertions.** Tests use `mock.remote.<namespace>.<method>` with `mockResolvedValue`, `mockResolvedValueOnce`, `mockReturnValueOnce`, or `mockImplementation`; the generated API supplies the signatures. Each mock instance owns its native response queue. Reusable tables register only default values or positional handlers, and each endpoint retains only its latest default. Stateful callbacks and deferred promises belong to individual tests. `ok` builds the success envelope. Streams require explicit declarations and can receive scripts over the opening args and a handle (`push`, `end`, `fail`); a declaration without a script produces a stream miss, while undeclared endpoints default to unary calls. Values are not validated. `mock.streams` controls scripted streams and exposes readiness/drain waits; `mock.log` records carrier calls (`pending`, `answered`, `failed`), scripted-stream state, first-argument `requests(endpoint?)`, and unmatched requests. `RemoteMock.create()` answers `$events` with a ready frame so the client can connect.
+
+**Vitest owns each test's mock and client lifetime.** `createClientTest(plan, options)` adds native `mock`, `remote`, and `start` fixtures. The mock is fresh and carries the default responses; `remote` is its namespace proxy, and explicit `start()` leaves startup responses configurable and shares one startup promise within the test. Teardown waits for startup, disposes the successful client even after an assertion failure, checks missing responses, and rejects later starts. Callers await startup failures. Independently owned clients still use `TestClient.start`. Scenario data configures native mocks directly; returning a mutation response and updating subsequent describe responses remain separate actions.
+
+**`remoteDefaultResponses` holds default responses for the boot-time Remote endpoints.** The table lists exactly the endpoints the `web` roster calls while booting and rendering with no sessions, no workspaces, and default settings, each row commented with its caller. A spec layers its own `RemoteTable` on top; a new boot-time call fails the spec at `dispose()`.
+
+`mock.remote` uses native `@vitest/spy.fn` functions shared by direct callers and Connection dispatch. `MockedRemote` applies Vitest's deep mock type transformation to the entire generated namespace map; an empty map weakens only this proxy to `any`. Production `Context` and Remote declarations remain strict, with no namespace-specific type copies or compiler flags. The [proxy typing guidance](../../../../packages/test-support/remote-mock/README.md#remote-proxy) requires build-backed local type checking even when unbuilt tests pass.
+
+## Product exports added for the tier
+
+- `client/connection`: `ClientTransportHooks.rpc?` publishes the already decoded carrier the `?fixture` path used internally; `fetch` becomes optional.
+- `client/hmr`: `tearDownEntryFiber(entry)` is the registry-first fiber teardown `reload` already performed.
+- `client/modules`: `parseDshClient` and `exactPackageSpecifier` are exported from the client face and shared by the Host and roster reader. Test factories use the existing registration queue. The roster-to-boot-graph synthesis has only test consumers and lives in the tier.
+- `client/web`: `bootClient` and `mountClient` are extracted from `AppWebEntry.run()`, which now calls them.
+- `api/gateway`: `carrierFailure` and `cancelledFailure` are exported so a stand-in for the generated client folds identically.
+
+## Alternatives considered
+
+**Running the generated `/remote` clients from built `lib/`.** Rejected: it makes source-plane specs depend on a build artifact, and the proxies need only the unary-or-stream declaration the mock already holds.
+
+**A generated static roster module with a drift gate.** Rejected after review: it is a copy of bundle data inside the test package, and every subset written against it is a hand-list that misses rows. Reading the bundles at import through the include plugin's own schema and patch application removes the copy, the generator, and the gate.
+
+**A Host compile face for the test runtime, a dynamic import of a Host module, or a vitest `globalSetup` handing rosters through `provide`/`inject`.** Rejected: a Client test runtime must not import Host code, dynamic imports hide the dependency, and a config-level channel hides the roster's source. The composition functions the launcher uses are face-neutral, so none of these is needed.
+
+**A hand-written test-side YAML and patch parser.** Rejected: `entryListSchema` and `applyEntryPatches` are the launcher's own and carry no Host Context merge; the tier writes only file reading, package.json location, and the web-row filter.
+
+**A mock module standing in for the Gateway client.** Rejected: the mock must not interfere with Gateway internals; installing it on the Connection carrier keeps retry, folding, and stream semantics real.
+
+**A second typed Gateway implementation with an `Api` generic, envelope and error classes, and a fixtures directory.** Rejected: it duplicates Gateway declarations and encoding. The mock derives method types from the generated namespace map and declares only unary or stream behavior at runtime.
+
+**Proxies passing unary rejections through unchanged.** Rejected: product code never awaits a Remote rejection because the generated client folds carrier throws, so an unmatched endpoint produced unhandled rejections; folding through the exported helpers restores the client's face.
+
+**Keeping CallContext and wrapping native spies in an adapter.** Rejected: it makes each test unwrap a synthetic call and retains counter/state machinery with no business-spec consumer. Positional handlers use the existing test ecosystem directly.
+
+**A separate `once` / `sequence` DSL and fallback rule stack.** Rejected: native per-instance queues already express the deferred responses and temporary failures used by consumers. Immutable table declarations with a cursor per registration would allow shared one-shot tables, but no current shared table requires them. The tier gives up newest-registration-first fallback and table-level repeat-last declarations; tests use native queue order and a persistent default instead. Response values and stateful handlers remain borrowed, not cloned.
+
+**A separate pre-materialized-module option or `staticModules`.** Rejected: the existing pending registration queue accepts the same factories before Loader startup. `staticModules` bypasses factory materialization and does not share graph prefetch/invalidation behavior; the queue removes the extra option without losing that behavior. Assembly-only helpers remain in their leaf modules rather than the recommended entry's exports.
+
+**A compiler-wide fallback flag or private augmentation package.** Rejected: declaration merging affects every file in a TypeScript Program that reaches the import; `private: true` only prevents publication. Separate test compiler graphs and additional policy checks add configuration maintenance without narrowing the fallback to its actual helper consumers. A local conditional type limits weakened inference to those consumers.
+
+**Per-namespace helpers with selected method lists and separate spy aliases.** Rejected: they repeat operation names and controls already supplied by native mocks. The generic proxy derives every method from the production namespace map, while fixtures own the returned data rather than a second implementation of domain writes or publication.
+
+## Consequences
+
+Specs boot real plugins: the whole `web` roster costs about five seconds cold and well under a second warm, and a three-row cone about twenty milliseconds per test. Assertions read product facts — the real section list, the real declarer, a Loader rebuild, a second `$events` generation on reconnect — and change when the product changes.
+
+The proxies skip the generated client's zod validation, wire-name mapping, and scoped-identity injection; mock rules read positional `args`, and the generated clients stay covered by the built-artifact e2e lanes. `remoteDefaultResponses` must gain a row when a plugin adds a boot-time call, and fails loud until it does. The two bundle names of the `web` profile are repeated once in `WEB_PROFILE_BUNDLES`, mirroring the launcher's `PROFILE_TEMPLATES.web`, and no check links the two: the client test program cannot import `@deepseek-ai/dsh-app-boot`, whose Host `Context` merges collide with the Client ones, and the test runtime takes no Host dependency even for tests. A template change therefore has to be carried to that constant by hand.
+
+The shared functions keep production and test callers on one implementation. `AppWebEntry.run()` mounts the Loader after the immediate-tier prefetch settles; application entry creation remains after prefetch, so serializing Loader setup with prefetch does not advance application activation.
+
+Native stream overrides may return their own iterable. The caller then owns consumption and cancellation; these iterables bypass scripted-stream logs and controls. Registered scripts retain managed queues and cancellation. This distinction preserves native mock behavior without adding another iterator wrapper or changing pull timing.
+
+## Deferred
+
+Product facts the tier surfaced and leaves as they are:
+
+- No `declare module` augmentation declares `Context.connection`; consumers read `ctx.get('connection') as ConnectionHandle`, and `TestClient.connection` is the typed entry the tier offers.
+- `TestClient.start` has no page-URL option, so a spec that needs the `connection` plugin to classify the page as off-loopback reconfigures the jsdom instance vitest exposes on `globalThis.jsdom`, a private detail of the jsdom environment provider.
+- `ISessions` exposes no queue observation point, so queue frames reach a `Session` through `handleControlFrame` directly rather than over the `session/control` stream.
+- A durable event pushed twice with the same seq is dropped at the tail of `RemoteJournalStream` as a replay and never reaches `SessionQueueMirror.acceptDurable`.
+- The vendored Loader rejects `create()` when a module import fails, so the import-failed branch of `assertEntriesActive` is unreachable through `create()`.
+- The session-controller client casts `ctx.remote as unknown as SessionRemotes`; in the client test program the cast is redundant, since the generated `/remote` merges are visible there.
+
+## Testing
+
+`packages/test-support/remote-mock/tests/` covers rules, streams, the log, and the carrier face; the `assembly-` specs under `packages/test-support/client-runtime/tests/` cover the roster reader on the real bundles and on a scratch installation, module loading, the proxies including their fold, and `TestClient` under jsdom and plain Node. Seven converted specs use the tier. In `packages/client/ui-settings-general/tests/`, the shell and apply specs boot the whole `web` roster; the apply spec reads its Chinese copy from the Host settings document the mock answers and reconfigures the jsdom page URL for the off-loopback branch. In `packages/api/session-controller/tests/`, the Session, queue-store, and pending-submission specs drive their objects over the roster's real Connection through the `remote.<ns>` proxies, with the Gateway client's own `$stream` retry loop, over the gateway's dependency cone, and the client-apply spec boots the plugin's dependency cone, delivering Remote events as emit frames on `$events`. In `packages/api/workspace-controller/tests/`, the transport spec boots the plugin's cone for apply cases and the gateway cone for hand-built stream and controller cases, since a rostered plugin would share the follow endpoint. Each package keeps a `tests/remote/` module with its default responses and frame builders. The fixture tests include an expected assertion failure and independently observe completed client cleanup; settings reload tests observe replaced registration identities, and write tests assert every mutation argument. Teardown-failure tests execute the real tree disposer before reporting the injected failure and observe the `$events` stream's cancelled state.

.agents/notes/implemented/testing/2026-09-06-client-assembly-test-line.zh.md

@@ -0,0 +1,90 @@
+# Agent Note: 基于端点具名 Remote mock 的整机客户端测试档
+
+Status: implemented
+
+[English](2026-09-06-client-assembly-test-line.md) | 中文
+
+## 问题
+
+浏览器功能 spec 各自手拼测试台：一个裸 Cordis context、`locale`、`connection`、`remote` 的替身，以及本该由真声明者做出的 slot 声明。它们的断言因此描述的是测试台而不是产品：一个插件新增了设置 section、一个声明者经 Loader 重载、一个 Connection 重连，对它们都不可见，而每个测试台都重复着同样的四十行并各有细小漂移。
+
+API 客户端 spec 用一个可编程的 Remote 面假件驱动对象。这个假件重新实现了它本该只是调用的 Gateway 语义：从历史列表推导 follow 流的开场快照、切页、带投递 promise 的流泵，以及一层信封。每一样都是产品已有契约的第二份实现，它还让测试描述出生成客户端做不出来的行为，比如一次会 reject 的一元调用。
+
+没有聚焦源码测试按生产方式起客户端——`bootClient` 按 manifest 每行建一个 Loader entry，再 `mountClient`——插件之间的组合故障会被手工测试台遮住。
+
+## 决定
+
+整机档放在 `@deepseek-ai/dsh-client-test-runtime` 的深 import `src/assembly/` 下，新的 test-support 包 `@deepseek-ai/dsh-remote-mock` 按端点名应答 Remote 流量。两者的用法由各自 README 描述（[client-runtime](../../../../packages/test-support/client-runtime/README.zh.md)、[remote-mock](../../../../packages/test-support/remote-mock/README.zh.md)）；本文记录它们背后的决定。
+
+**roster 从 bundle 现读，绝不拷贝。** `bundleRoster(bundles)` 用 include 插件自己的 YAML 方言（带 `!!js` 的 `entryListSchema`）解析每个 bundle 的 `dsh.bundle.patch`，用它的 `applyEntryPatches` 合成各层，再保留每个未禁用且其包声明 `dsh.client.platform === 'web'` 的行，带上该声明的 `inject` 与 `immediately`。`webApp` 是 `web` profile 的 roster（先 `dsh-base`、再 `dsh-web-app`），import 时算出。spec 点名它要测的东西，其余推导：`webApp.closure([row])` 保留一行及其传递 `inject` 锥；`pick` 与 `without` 留给刻意裁剪。测试运行时仍是 Client 面的包：它不 import 任何 Host 模块，此处不用动态 import，其 client 面的 `types` 在 `client-build-environment` 之外加了 `node`，好让读取器使用 `node:fs`。`closure` 把 shell 静态种入的平台模块（`PLATFORM_MODULES`）视为无需行即已满足。
+
+**生产启动路径原样运行。** `TestClient.start(plan, mock)` 把 mock 装成 Connection 载体（`__DSH_TRANSPORT__ = { rpc: mock.rpc }`），加载每个 roster 行的 `/client` 模块，经生产模块 facade 的 `pendingQueue` 登记其工厂，经 `bootClient` 启动，可选经 `mountClient` 挂载，然后等 `connected`。`reload(name)` 经 client-hmr 导出的 `tearDownEntryFiber` 重建一个 Loader entry；`unload(name)` 移除它；`dispose()` 拆掉一切，然后对任何没有规则的端点让测试失败。jsdom 没有 `EventSource` 与 `ResizeObserver`；`start` 只在全局缺失处装惰性替身。每个 worker 内启动与 entry 重建逐个进行，因为 `connection` 插件在 apply 时读传输全局，每次都先装上当事客户端的传输；传输与替身按引用计数持有，第一个客户端安装、最后一次 dispose 恢复，因此同一测试里重叠的客户端各连各的 mock，`reload` 了 `connection` 行之后也是。
+
+**`remote.<ns>` 是无契约代理，不是生成客户端。** `@deepseek-ai/dsh-api-remotes` 行被去掉，因为它生成的客户端只存在于构建后的 `lib/`。对 roster 行注入的每个 `remote.<ns>`，加上 mock 有规则的每个命名空间，本档各提供一个 Proxy：`ctx.remote.<ns>.<method>(...args)` 经 roster 自己的 Connection 用位置参数调用端点 `<ns>/<method>`，mock 为它登记了流脚本就走流、否则走一元。Cordis 把 `ctx.remote.<ns>` 解析到服务 `remote.<ns>`，所以 Gateway 客户端本身不动。一元应答原样返回；一元拒绝按生成客户端折叠载体抛错的方式折叠，经 Gateway 客户端导出的 `carrierFailure` 与 `cancelledFailure`，因此不等待就发出 Remote 调用的产品代码看不到任何 reject。流的项与失败按流吐出的样子直传。
+
+**原生 mock 负责响应配置和调用断言。** 测试通过 `mock.remote.<namespace>.<method>` 使用 `mockResolvedValue`、`mockResolvedValueOnce`、`mockReturnValueOnce` 或 `mockImplementation`，签名由生成的 API 提供。每个 mock 实例独立持有原生响应队列。可复用的表只登记默认值或位置参数 handler，每个端点仅保留最新默认响应。有状态回调和 deferred promise 归各测试所有。`ok` 构造成功信封。流需要显式声明，可提供接收打开参数与句柄（`push`、`end`、`fail`）的脚本；无脚本的声明产生流漏配，未声明端点默认走一元。值不校验。`mock.streams` 控制脚本流并提供打开／排空等待；`mock.log` 记录载体调用（`pending`、`answered`、`failed`）、脚本流状态、首参数 `requests(endpoint?)` 和未匹配请求。`RemoteMock.create()` 为 `$events` 应答 ready 帧，让客户端可以连接。
+
+**Vitest 拥有每条测试的 mock 和客户端生命周期。** `createClientTest(plan, options)` 增加原生 `mock`、`remote` 与 `start` fixture。mock 每次新建并携带默认响应；`remote` 是它的命名空间 Proxy，显式 `start()` 留出配置启动期应答的时机，同一测试共用一个启动 Promise。收尾等待启动，即使断言失败也销毁成功创建的客户端、检查漏配，并拒绝后续启动。启动错误由调用方 await 观察。分别拥有多个客户端时仍用 `TestClient.start`。场景数据直接配置原生 mock；返回 mutation 应答与更新后续 describe 应答仍是两个独立操作。
+
+**`remoteDefaultResponses` 是启动期 Remote 端点的默认响应。** 这张表恰好列出 `web` roster 在没有 session、没有 workspace、默认设置下启动并渲染时会打的端点，每行注明调用方。spec 在其上叠加自己的 `RemoteTable`；新的启动期调用会在 `dispose()` 时让 spec 失败。
+
+`mock.remote` 使用直接调用方与 Connection 分发共用的原生 `@vitest/spy.fn` 函数。`MockedRemote` 对完整生成的命名空间映射应用 Vitest 深层 mock 类型转换；映射为空时仅这个 Proxy 弱化为 `any`。生产 `Context` 与 Remote 声明保持严格，不需要命名空间专属类型副本或编译器 Flag。[Proxy 类型指引](../../../../packages/test-support/remote-mock/README.zh.md#remote-proxy)要求即使无构建测试通过，本地也必须执行构建后的类型检查。
+
+## 为本档新增的产品导出
+
+- `client/connection`：`ClientTransportHooks.rpc?` 公开 `?fixture` 路径内部已在用的已解码载体；`fetch` 变为可选。
+- `client/hmr`：`tearDownEntryFiber(entry)` 就是 `reload` 本来执行的 registry 先行的 fiber 拆除。
+- `client/modules`：`parseDshClient` 与 `exactPackageSpecifier` 从 client 面导出，由 Host 和 roster 读取器共用。测试工厂使用已有注册队列。roster 行到 boot graph 的合成只有测试消费者，放在本档里。
+- `client/web`：`bootClient` 与 `mountClient` 从 `AppWebEntry.run()` 抽出，后者现在调用它们。
+- `api/gateway`：导出 `carrierFailure` 与 `cancelledFailure`，让生成客户端的替身折叠得一模一样。
+
+## 考虑过的替代方案
+
+**从构建后的 `lib/` 运行生成的 `/remote` 客户端。** 否决：它让源码面的 spec 依赖构建产物，而代理只需要 mock 已持有的一元或流声明。
+
+**带漂移门禁的生成静态 roster 模块。** 评审后否决：它是测试包内的一份 bundle 数据拷贝，基于它写的每个子集都是会漏行的手列清单。用 include 插件自己的 schema 与补丁应用在 import 时读 bundle，把拷贝、生成器和门禁一起去掉。
+
+**给测试运行时加 Host 编译面、动态 import 一个 Host 模块，或用 vitest `globalSetup` 经 `provide`/`inject` 传 roster。** 否决：Client 测试运行时不得 import Host 代码，动态 import 藏起依赖，配置层通道藏起 roster 的来源。启动器用的合成函数本就面中立，这些都不需要。
+
+**自写一套测试侧的 YAML 与补丁解析器。** 否决：`entryListSchema` 与 `applyEntryPatches` 就是启动器自己的，不带 Host Context 合并；本档只写读文件、定位 package.json 和 web 行过滤。
+
+**用一个 mock 模块替代 Gateway 客户端。** 否决：mock 不得干涉 Gateway 内部；装在 Connection 载体上让重试、折叠与流语义都保持真实。
+
+**第二套带类型的 Gateway 实现，包含 `Api` 泛型、信封与错误类以及 fixtures 目录。** 否决：它重复 Gateway 声明与编解码。mock 从生成的命名空间映射派生方法类型，运行时只声明一元或流行为。
+
+**代理把一元拒绝原样直传。** 否决：产品代码从不等待 Remote 的 reject，因为生成客户端会折叠载体抛错，于是没匹配的端点造成未处理的拒绝；经导出的辅助函数折叠恢复了客户端的面。
+
+**保留 CallContext，再用适配器包装原生 spy。** 否决：每个测试都要解开合成调用对象，还保留没有业务 spec 消费者的计数／状态机制。位置参数 handler 直接使用现有测试生态。
+
+**独立的 `once` / `sequence` DSL 与回退规则栈。** 否决：按实例持有的原生队列已经能表达消费方使用的延迟响应和临时失败。不可变表声明配合每次登记的游标虽能共享一次性响应表，但当前没有共享表需要它。本档放弃最新登记优先的回退和表级末项重复声明，测试改用原生队列顺序与持续默认响应。响应值和有状态 handler 仍按引用借用，不做克隆。
+
+**独立的预加载模块选项或 `staticModules`。** 否决：现有待注册队列能在 Loader 启动前接收同样的工厂。`staticModules` 绕过工厂物化，也不共享图的预取／失效行为；队列在保留这些行为的同时省掉额外选项。仅装配内部使用的 helper 保留在叶模块，不从推荐入口再导出。
+
+**编译器级全局降级 Flag 或私有类型增补包。** 否决：声明合并会影响同一 TypeScript Program 中能够到达该导入的所有文件；`private: true` 只阻止发布。拆分测试编译图或增加策略检查会增加配置维护成本，却不能把降级限制在真正使用它的 Helper 中。局部条件类型只弱化这些消费方的类型推断。
+
+**为每个命名空间编写方法清单和独立 spy 别名的 Helper。** 否决：它们重复操作名称和原生 mock 已经提供的控制功能。通用 Proxy 从生产命名空间映射派生每个方法，fixture 拥有返回数据，而不再实现另一份领域写入或发布机制。
+
+## 后果
+
+spec 起的是真插件：整个 `web` roster 冷启动约五秒、热启动远低于一秒，三行的锥每例约二十毫秒。断言读的是产品事实——真实的 section 清单、真实的声明者、一次 Loader 重建、重连时的第二代 `$events`——并随产品变化而变化。
+
+代理跳过了生成客户端的 zod 校验、wire 字段名映射与 scoped 身份注入；mock 规则读位置 `args`，生成客户端仍由构建产物 e2e 车道覆盖。插件新增启动期调用时 `remoteDefaultResponses` 必须加一行，加之前会响亮失败。`web` profile 的两个 bundle 名在 `WEB_PROFILE_BUNDLES` 里重复了一次，对应启动器的 `PROFILE_TEMPLATES.web`，且两者之间没有机检联系：客户端测试程序不能 import `@deepseek-ai/dsh-app-boot`（其 Host `Context` 合并与 Client 的冲突），而测试运行时连测试也不引入 Host 依赖。模板变更因此要靠人工带到这个常量。
+
+共享函数让生产和测试调用方使用同一份实现。`AppWebEntry.run()` 在立即层预取落定后挂载 Loader；应用 entry 仍在预取之后创建，因此让 Loader 安装与预取串行不会提前应用激活。
+
+原生流覆盖可以返回自有 iterable，此时调用方负责消费与取消，这些 iterable 不参与脚本流日志或控制。已登记的脚本仍使用受控队列与取消机制。这一区分保留原生 mock 行为，无需再添加 iterator 包装器或改变拉取时机。
+
+## 遗留事项
+
+本档暴露出来、原样保留的产品事实：
+
+- 没有任何 `declare module` 增强声明 `Context.connection`；消费方一律 `ctx.get('connection') as ConnectionHandle`，`TestClient.connection` 是本档提供的带类型入口。
+- `TestClient.start` 没有页面 URL 选项，需要 `connection` 插件把页面判为 off-loopback 的 spec 只能重配 vitest 挂在 `globalThis.jsdom` 上的 jsdom 实例，这是 jsdom 环境提供者的私有细节。
+- `ISessions` 没有 queue 观察点，queue 帧只能直接经 `handleControlFrame` 到达 `Session`，而不是走 `session/control` 流。
+- 同 seq 的 durable 事件二次推送在 `RemoteJournalStream` 尾部被当作重放丢弃，到不了 `SessionQueueMirror.acceptDurable`。
+- vendored Loader 在模块 import 失败时让 `create()` reject，因此 `assertEntriesActive` 的 import 失败分支经 `create()` 不可达。
+- session-controller 客户端把 `ctx.remote` cast 成 `SessionRemotes`；在客户端测试程序里这个 cast 是多余的，因为生成的 `/remote` 合并在那里可见。
+
+## 测试
+
+`packages/test-support/remote-mock/tests/` 覆盖规则、流、日志与载体面；`packages/test-support/client-runtime/tests/` 下的 `assembly-` 系列 spec 覆盖在真 bundle 与临时安装上的 roster 读取器、模块加载、含折叠的代理，以及 jsdom 与纯 Node 下的 `TestClient`。七条改造后的 spec 使用本档。`packages/client/ui-settings-general/tests/` 下，shell 与 apply 两条起整个 `web` roster；apply 从 mock 应答的 Host settings 文档读它的中文文案，并为 off-loopback 分支重配 jsdom 页面 URL。`packages/api/session-controller/tests/` 下，Session、queue-store、pending-submission 三条在 gateway 依赖锥上经 `remote.<ns>` 代理走 roster 的真 Connection 驱动对象（`$stream` 的重试循环仍是 Gateway 客户端自己的），client-apply 起插件的依赖锥，把 Remote 事件作为 `$events` 上的 emit 帧投递。`packages/api/workspace-controller/tests/` 下，transport 的 apply 用例起插件锥、手工构造流与 controller 的用例起 gateway 锥，因为进了 roster 的插件会共用 follow 端点。每个包在 `tests/remote/` 保有自己的默认响应与帧构造。fixture 测试包含预期的断言失败，并独立观察客户端清理完成；settings 重载测试观察注册身份被替换，写入测试断言全部 mutation 参数。teardown 失败测试先执行真实树清理，再报告注入的失败，并观察 `$events` 流的取消状态。

.agents/notes/implemented/testing/2026-09-06-frontend-performance-budgets.i18n.yaml

@@ -2,5 +2,5 @@
 # side as of the last confirmed-consistent state. Both languages carry equal authority;
 # after editing either side, bring the other along and re-record with:
 #   pnpm run verify-translation-pairing --write .agents/notes/implemented/testing/2026-09-06-frontend-performance-budgets.md
-2026-09-06-frontend-performance-budgets.md: 4c1a99e10c9b2d7ba7ecbd38f2249ac84e9d330b
-2026-09-06-frontend-performance-budgets.zh.md: f563c55e6440de85d71110cbfb2533c7f99218ce
+2026-09-06-frontend-performance-budgets.md: 4d69dda8a04d4e9807c18c8f1b978ca7fc87a883
+2026-09-06-frontend-performance-budgets.zh.md: c9a174df55a2c175232eb2f4d8468b054e532b9c

.agents/notes/implemented/testing/2026-09-06-frontend-performance-budgets.md

@@ -16,7 +16,7 @@ The existing serial benchmark inventory includes two frontend owners: [active re
 
 The browser input contains 240 closed turns, 40 tool results, and 20 code fences, plus mixed-language prose and reasoning. Historical Assistant records carry matching compact streams built through the production accumulator with 12-character reasoning/text deltas and 8-character tool-argument deltas; empty streams would omit stored and transferred payload costs. Nine older-page actions exhaust this input from its observed 25-turn initial window; the readiness probe follows mounted turn growth rather than duplicating the pagination algorithm. Each sample uses a fresh scaffold and browser. Setup, seeding, browser launch, initial shell load, and sidebar expansion are excluded from open timing. Open ends at transcript availability and an editable composer; page and navigation timings end at their target DOM state. Two animation frames include a rendering opportunity, not hardware presentation or a guarantee that every offscreen node painted.
 
-The continuation sends 120 text deltas at 16 ms replay pacing. The input witness is installed before Enter submission from the focused composer; typing retains focus without a mouse-refocus action or a separate pre-input animation-frame wait. First/final marker lookups stay inside the latest Assistant step and retain visible-state waits. Diagnostics capture reply markers and focus immediately after the first-visible wait, plus browser-clock timestamps and focus at the first actual input event. Replay never waits for input; starvation can still fail overlap. The synchronous input witness reads that same bounded reply. Whole-history text and accessibility queries add observer CPU and garbage collection to the measured interval, so reducing that observer work is benchmark repair, not product optimization. It records Enter-to-first-visible-reply, trusted draft typing whose first actual input event observes the first reply but no completion marker, complete reply wall time through settled persistence and the new rendered turn-tail, and Chromium main-thread task duration. The complete wall budget adds the fixed 1984 ms scripted pacing to a scaled overhead allowance; input and completion have their own enforced budgets. Post-GC browser heap and DOM counts remain diagnostics because one endpoint does not prove a leak.
+The continuation sends 120 text deltas at 16 ms replay pacing. The input witness is installed before Enter submission from the focused composer; typing retains focus without a mouse-refocus action or a separate pre-input animation-frame wait. First/final marker lookups sample visible text inside the latest Assistant step on animation frames, avoiding selector retry backoff. Hidden text and markers in older steps cannot satisfy the observer. The first observation captures reply markers and focus in the browser; diagnostics are retrieved after typing to avoid an additional pre-input round trip. Diagnostics also capture browser-clock timestamps and focus at the first actual input event. Replay never waits for input; starvation can still fail overlap. The synchronous input witness reads that same bounded reply. Whole-history text and accessibility queries add observer CPU and garbage collection to the measured interval, so reducing that observer work is benchmark repair, not product optimization. It records Enter-to-first-visible-reply, trusted draft typing whose first actual input event observes the first reply but no completion marker, complete reply wall time through settled persistence and the new rendered turn-tail, and Chromium main-thread task duration. The complete wall budget adds the fixed 1984 ms scripted pacing to a scaled overhead allowance; input and completion have their own enforced budgets. Post-GC browser heap and DOM counts remain diagnostics because one endpoint does not prove a leak.
 
 Reconnect uses three fresh compiled plain-Node children. Each creates a 100,000-delta reasoning prefix with distinct timestamps and two compact records before timing `ClientAssistantStream.replace()`. GC precedes the baseline and follows replacement while the result remains reachable; replacement time excludes both collections. The report consumes the result after collection and checks that the next dense live frame remains accepted. This measures reconstruction, not transport, rendering, or an entire reconnect workflow.
 

.agents/notes/implemented/testing/2026-09-06-frontend-performance-budgets.zh.md

@@ -16,7 +16,7 @@ Node 对话折叠很快，并不能证明浏览器能绘制长对话或在流式
 
 浏览器输入包含 240 个已关闭轮次、40 个工具结果和 20 个代码块，以及混合语言正文和推理。历史 Assistant 记录携带匹配的紧凑 stream，通过生产 accumulator 按 12 字符推理/文本 delta 和 8 字符工具参数 delta 构建；空 stream 会遗漏存储与传输负载成本。从观察到的初始 25 轮窗口开始，九次更早分页操作读完该输入；就绪探针跟踪已挂载轮次增长，不复制分页算法。每个样本使用全新 scaffold 和浏览器。环境准备、数据播种、浏览器启动、初始 shell 加载及侧栏展开不计入打开时间。打开测量在对话可用且输入框可编辑时结束；分页与导航测量在目标 DOM 状态出现时结束。两次动画帧包含一次渲染机会，不代表硬件显示或保证每个屏幕外节点都已绘制。
 
-续接以 16 ms 重放间隔发送 120 个文本 delta。输入观察器在从已聚焦输入框按 Enter 提交前安装；键入保留焦点，不执行鼠标重新聚焦，也不单独等待输入前动画帧。首段/最终标记查找限制在最新 Assistant step，并保留可见状态等待。诊断在首段可见等待后立即记录回复标记和焦点，并记录首个实际输入事件的浏览器时钟时间戳与焦点。重放从不等待输入；响应阻塞仍可能导致重叠失败。同步输入证据读取同一个受限回复。全历史文本与无障碍查询会向测量区间加入观察器 CPU 和垃圾回收成本，因此减少此类观察工作属于基准修正，而非产品优化。它记录 Enter 提交到首段可见回复的时间、首个实际输入事件观察到首段回复且完成标记尚未出现时的真实草稿键入、直到持久化结算并渲染新 turn-tail 的完整回复壁钟时间，以及 Chromium 主线程任务时间。完整壁钟预算在缩放后的额外开销额度上加固定的 1984 ms 脚本节奏；输入和完成均有独立执行的预算。强制 GC 后的浏览器 heap 和 DOM 数量仍仅供诊断，因为单个终点不能证明泄漏。
+续接以 16 ms 重放间隔发送 120 个文本 delta。输入观察器在从已聚焦输入框按 Enter 提交前安装；键入保留焦点，不执行鼠标重新聚焦，也不单独等待输入前动画帧。首段/最终标记查找在动画帧上采样最新 Assistant step 内的可见文本，避免选择器重试退避。隐藏文本和较早步骤中的标记均不能满足观察条件。首次观察在浏览器内记录回复标记与焦点；诊断在键入后取回，避免增加输入前的通信往返。诊断还记录首个实际输入事件的浏览器时钟时间戳与焦点。重放从不等待输入；响应阻塞仍可能导致重叠失败。同步输入证据读取同一个受限回复。全历史文本与无障碍查询会向测量区间加入观察器 CPU 和垃圾回收成本，因此减少此类观察工作属于基准修正，而非产品优化。它记录 Enter 提交到首段可见回复的时间、首个实际输入事件观察到首段回复且完成标记尚未出现时的真实草稿键入、直到持久化结算并渲染新 turn-tail 的完整回复壁钟时间，以及 Chromium 主线程任务时间。完整壁钟预算在缩放后的额外开销额度上加固定的 1984 ms 脚本节奏；输入和完成均有独立执行的预算。强制 GC 后的浏览器 heap 和 DOM 数量仍仅供诊断，因为单个终点不能证明泄漏。
 
 重连使用三个全新编译后的纯 Node 子进程。各进程在计时 `ClientAssistantStream.replace()` 前创建包含不同时间戳、两条紧凑记录和 100,000 个 delta 的推理前缀。在基线前执行 GC，并在结果仍可达时于替换后再次 GC；替换时间不含两次回收。报告在回收后消费结果，并检查下一个稠密序号的实时 frame 仍被接受。这测量重建，不测量传输、渲染或完整重连工作流。
 

.agents/notes/implemented/testing/2026-09-08-ci-completion-observations.i18n.yaml

@@ -2,5 +2,5 @@
 # side as of the last confirmed-consistent state. Both languages carry equal authority;
 # after editing either side, bring the other along and re-record with:
 #   pnpm run verify-translation-pairing --write .agents/notes/implemented/testing/2026-09-08-ci-completion-observations.md
-2026-09-08-ci-completion-observations.md: 8af4685a5e2c51c1edf4a41b47088916223e7a6c
-2026-09-08-ci-completion-observations.zh.md: e3147bb7ac39877b46820862e9af4645803a37a2
+2026-09-08-ci-completion-observations.md: e3de87a70419778407b5eb230cec64ce088dd0c0
+2026-09-08-ci-completion-observations.zh.md: 4e95d6fb9dc9b1ecac2e3d0dfa262b819de16261

.agents/notes/implemented/testing/2026-09-08-ci-completion-observations.md

@@ -30,7 +30,7 @@ The [LSP backpressure test](../../../../packages/lsp/lsp-stdio/tests/instance.sp
 
 ### Built-client import classification
 
-The [Node import sweep](../../../../packages/experimental/webworker-runtime/tests/compile/transform-corpus-check.ts) admits the Dockkit bundle only when Node reports `ERR_UNKNOWN_FILE_EXTENSION` for its exact `dockkit.module.css` path. Other errors and unexpectedly successful exempt imports fail. Scoped resolve/load hooks exercise expected CSS failure, arbitrary failure, another stylesheet, another error code, and stale exemption without modifying shared build artifacts.
+The [Node import sweep](../../../../packages/experimental/webworker-runtime/tests/compile/transform-corpus-check.ts) admits the Dockkit bundle only when Node reports `ERR_UNKNOWN_FILE_EXTENSION` for a `.css` file, and fails every other error and every unexpectedly successful exempt import; the [stylesheet exemption decision](../bug-fix/2026-09-10-built-bundle-css-exemption.md) owns which stylesheets that covers. Scoped resolve/load hooks exercise expected CSS failure, another stylesheet, another extension, arbitrary failure, another error code, and stale exemption without modifying shared build artifacts.
 
 ## Alternatives considered
 

.agents/notes/implemented/testing/2026-09-08-ci-completion-observations.zh.md

@@ -30,7 +30,7 @@ Status: implemented
 
 ### 已构建 Client 的导入分类
 
-[Node import sweep](../../../../packages/experimental/webworker-runtime/tests/compile/transform-corpus-check.ts)只有在 Node 针对准确的 `dockkit.module.css` 路径报告 `ERR_UNKNOWN_FILE_EXTENSION` 时才接受 Dockkit bundle。其他错误以及意外成功的豁免导入都会失败。限定范围的 resolve/load hook 覆盖预期 CSS 失败、任意失败、其他 stylesheet、其他错误码和过期豁免，不修改共享构建产物。
+[Node import sweep](../../../../packages/experimental/webworker-runtime/tests/compile/transform-corpus-check.ts)只有在 Node 针对某个 `.css` 文件报告 `ERR_UNKNOWN_FILE_EXTENSION` 时才接受 Dockkit bundle，其他每个错误以及每个意外成功的豁免导入都会失败；该豁免覆盖哪些样式表由[样式表豁免决策](../bug-fix/2026-09-10-built-bundle-css-exemption.zh.md)拥有。限定范围的 resolve/load hook 覆盖预期 CSS 失败、其他样式表、其他扩展名、任意失败、其他错误码和陈旧豁免，不修改共享构建产物。
 
 ## 考虑过的替代方案
 

.agents/notes/implemented/testing/2026-09-08-ci-readiness-and-completion.i18n.yaml

@@ -2,5 +2,5 @@
 # side as of the last confirmed-consistent state. Both languages carry equal authority;
 # after editing either side, bring the other along and re-record with:
 #   pnpm run verify-translation-pairing --write .agents/notes/implemented/testing/2026-09-08-ci-readiness-and-completion.md
-2026-09-08-ci-readiness-and-completion.md: 2249df5467189975aca2d73ec56c6cf82ec7b62f
-2026-09-08-ci-readiness-and-completion.zh.md: ccf8e7c16903b66e39bc48e99460e5c5179ab51b
+2026-09-08-ci-readiness-and-completion.md: c72048ef90b987e7d27992b1754736b2ce895093
+2026-09-08-ci-readiness-and-completion.zh.md: 1188d66249b3092416435da2e8496a657ad00677

.agents/notes/implemented/testing/2026-09-08-ci-readiness-and-completion.md

@@ -14,6 +14,8 @@ The [ACP coverage run](https://github.com/deepseek-harness/deepseek-harness/acti
 
 A [worker-runtime coverage failure](https://github.com/deepseek-harness/deepseek-harness/actions/runs/34248221544/job/102135631932) exhausts the slow-binding fixture's one-second compute allowance. Concurrent native Windows reproductions exceed that allowance before calling the binding. Worker initialization contributes measured active time; the delayed binding contributes idle time.
 
+The [Windows coverage run](https://github.com/deepseek-harness/deepseek-harness/actions/runs/34324325375/job/102377982193) reports an SDK subprocess exit beyond a fixture's 200 ms confirmation window and an Inspector Worker startup beyond its ten-second default. Neither the protocol-error routing case nor the Cordis tree projection case measures those latency guarantees.
+
 ## Decision
 
 The [webhook browser test](../../../../apps/web/tests/github-ready-review.e2e.ts) observes the model request caused by delivery before checking Session registration. The [feedback test](../../../../apps/web/tests/feedback-command.e2e.ts) waits for the empty composer and enabled attachment control before comparing ARIA output. Matching consecutive snapshots cannot prove that the command RPC has settled: its event stream can publish the acknowledgement first.
@@ -28,6 +30,8 @@ The [subagent teardown decision](2026-09-07-subagent-teardown-test-budgets.md) o
 
 The [worker-runtime binding test](../../../../packages/code-runtime/code-runtime-worker-thread/tests/runtime.spec.ts) allows five seconds of compute for source-worker initialization and delays the binding for 6.5 seconds. Charging that idle delay would still exceed the entire compute allowance. The case retains its 15-second test limit and 30-second wall ceiling, registers Context and reply-timer cleanup, and leaves the hot-loop, decoy-dispatch, wall-ceiling, and abort controls at their existing limits. Production budgets remain unchanged.
 
+The [SDK subagent protocol-error test](../../../../packages/subagent/subagent-dsh-sdk/tests/subagent-dsh-sdk.spec.ts) uses the provider's normal shutdown and exit grace periods and registers disposal before its assertions. The [Inspector tree tests](../../../../packages/experimental/inspector/tests/cordis-tree.host.spec.ts) pass the active test budget to Worker startup and register cleanup while startup is still pending. A cancelled test cannot receive a late-ready handle; cleanup awaits initialization and closes a successfully started Worker. Failed initialization already terminates the Worker before rejecting. A controlled late-start test verifies cancellation and closure through the real Worker's HTTP endpoint. Production defaults remain unchanged.
+
 ## Alternatives considered
 
 **Larger independent waits.** Rejected where a completion promise already exists. A separate polling deadline continues to compete with the execution lane's budget.

.agents/notes/implemented/testing/2026-09-08-ci-readiness-and-completion.zh.md

@@ -14,6 +14,8 @@ Status: implemented
 
 一次 [worker runtime coverage 失败](https://github.com/deepseek-harness/deepseek-harness/actions/runs/34248221544/job/102135631932)耗尽了慢 binding 夹具的一秒计算额度。原生 Windows 并发复现在调用 binding 前已超过该额度。Worker 初始化会累计所测的活跃时间；延迟的 binding 累计空闲时间。
 
+[Windows 覆盖率运行](https://github.com/deepseek-harness/deepseek-harness/actions/runs/34324325375/job/102377982193)报告了 SDK 子进程退出超过测试设置的 200 毫秒确认期限，以及 Inspector Worker 启动超过十秒默认期限。协议错误转发用例和 Cordis 树投影用例都不衡量这些延迟保证。
+
 ## 决策
 
 [Webhook 浏览器测试](../../../../apps/web/tests/github-ready-review.e2e.ts)观察投递触发的模型请求后再检查 Session 注册。[反馈测试](../../../../apps/web/tests/feedback-command.e2e.ts)在比较 ARIA 输出前等待输入框清空且附件按钮启用。连续两次快照相同不能证明命令 RPC 已完成：事件流可能先发布确认消息。
@@ -28,6 +30,8 @@ Status: implemented
 
 [Worker runtime binding 测试](../../../../packages/code-runtime/code-runtime-worker-thread/tests/runtime.spec.ts)为源码 worker 初始化保留五秒计算额度，并将 binding 延迟设为 6.5 秒。若将该空闲延迟计费，仍会超过整个计算额度。用例保留 15 秒测试期限与 30 秒墙钟上限，登记 Context 和回复定时器的清理，并保持热循环、诱饵 dispatch、墙钟上限及取消控制用例的原有限制。生产预算不变。
 
+[SDK 子 Agent 协议错误测试](../../../../packages/subagent/subagent-dsh-sdk/tests/subagent-dsh-sdk.spec.ts)使用提供方正常的关闭和退出等待时间，并在断言前登记清理。[Inspector 树测试](../../../../packages/experimental/inspector/tests/cordis-tree.host.spec.ts)将当前测试预算传给 Worker 启动，并在启动尚未完成时登记清理。取消后的测试不会收到随后才就绪的实例；清理等待初始化完成，并关闭成功启动的 Worker。初始化失败时，启动操作会在拒绝前终止 Worker。受控的延迟启动测试通过真实 Worker 的 HTTP 端点验证取消和关闭。生产默认值不变。
+
 ## 考虑过的替代方案
 
 **增大独立等待时限。** 已有完成 Promise 时不采用。独立轮询期限仍会与执行通道的预算竞争。

.agents/notes/implemented/testing/2026-09-09-user-patch-hmr-test-delivery.i18n.yaml

@@ -2,5 +2,5 @@
 # side as of the last confirmed-consistent state. Both languages carry equal authority;
 # after editing either side, bring the other along and re-record with:
 #   pnpm run verify-translation-pairing --write .agents/notes/implemented/testing/2026-09-09-user-patch-hmr-test-delivery.md
-2026-09-09-user-patch-hmr-test-delivery.md: 427cf938d38eaf5c351fc0334bfd7df766d5663c
-2026-09-09-user-patch-hmr-test-delivery.zh.md: c2a3a14c9f322e748dbfd131a98138b13c5d47a6
+2026-09-09-user-patch-hmr-test-delivery.md: fadd191650411f07c3b8b1f35c94e73869fbc8e2
+2026-09-09-user-patch-hmr-test-delivery.zh.md: b108cf25766ff108d87c429e9f208c2d4cf7afd2

.agents/notes/implemented/testing/2026-09-09-user-patch-hmr-test-delivery.md

@@ -1,4 +1,4 @@
-# Agent Note: User-patch transactions control filesystem event delivery
+# Agent Note: User-patch tests control filesystem event delivery
 
 Status: implemented
 
@@ -6,22 +6,22 @@ English | [中文](2026-09-09-user-patch-hmr-test-delivery.zh.md)
 
 ## Problem
 
-The [macOS Sandbox run](https://github.com/deepseek-harness/deepseek-harness/actions/runs/34238200206/job/102101292119) times out while waiting for the first user-patch addition. Concurrent local reproductions show no filesystem notification reaching HMR. A polling variant also misses a subsequent edit while HMR has no pending refresh. These failures prevent the transaction assertions from exercising the parser, activation, and rollback behavior they own.
+The [macOS Sandbox run](https://github.com/deepseek-harness/deepseek-harness/actions/runs/34238200206/job/102101292119) times out while waiting for the first user-patch addition. Concurrent local reproductions show no filesystem notification reaching HMR. A polling variant also misses a subsequent edit while HMR has no pending refresh. These failures prevent the refresh assertions from exercising the parser, activation, and recovery behavior they own.
 
 ## Decision
 
-The [user-patch transaction test](../../../../packages/boot/app-boot/tests/user-patches.spec.ts) writes real patch files and delivers their add, change, and unlink events through a Chokidar watcher without native watch handles. HMR registration, refresh serialization, Include recomposition, plugin activation, failure broadcasting, rollback, and recovery remain real. The fixture restores its watcher factory and disposes the Context even when setup fails before the local cleanup block.
+The [user-patch test](../../../../packages/boot/app-boot/tests/user-patches.spec.ts) writes real patch files and delivers their add, change, and unlink events through a Chokidar watcher without native watch handles. App-boot watcher registration, refresh serialization, Include recomposition, plugin activation, failure reporting, and recovery remain real; plugin rollback is absent under the [Loader policy](../simplification/2026-09-09-nontransactional-loader.md). The fixture restores its watcher factory and disposes the Context even when setup fails before the local cleanup block.
 
-The separate [HMR config tests](../../../../packages/boot/app-boot/tests/hmr-config.spec.ts) own native notification delivery, including add/change/unlink, initially absent parents, and filesystem aliases. The transaction test does not establish operating-system delivery guarantees.
+The separate [watcher tests](../../../../packages/boot/app-boot/tests/watch-config.spec.ts) own native notification delivery, including add/change/unlink, initially absent parents, and filesystem aliases. The refresh test does not establish operating-system delivery guarantees.
 
 ## Alternatives considered
 
-**Native notifications for every transaction assertion.** Rejected because it repeats the native delivery dependency across each parser and activation state transition. A missing event obscures which downstream behavior is broken.
+**Native notifications for every refresh assertion.** Rejected because it repeats the native delivery dependency across each parser and activation state transition. A missing event obscures which downstream behavior is broken.
 
 **Polling and fixed settling delays.** Rejected because neither acknowledges delivery of the next edit. Chokidar readiness does not expose completion of Node's asynchronous initial polling baseline; a local polling reproduction still misses changes. Increasing the test deadline cannot recover an event that was never emitted.
 
-**Mock HMR registration or Include.** Rejected because the test must retain transactional recomposition and last-good-state assertions after activation and parse failures.
+**Mock HMR registration or Include.** Rejected because the test must retain real recomposition, report activation failures, and preserve the running configuration after parse failures.
 
 ## Consequences
 
-The transaction sequence retains every semantic assertion and removes fixed change-throttle sleeps. Independent concurrent processes exercise isolation, and a forced setup failure verifies watcher closure and factory restoration before the next case. Native watcher failures remain visible in their owning tests and require their own diagnosis.
+The refresh sequence retains its semantic assertions and removes fixed change-throttle sleeps. Independent concurrent processes exercise isolation, and a forced setup failure verifies watcher closure and factory restoration before the next case. Native watcher failures remain visible in their owning tests and require their own diagnosis.

.agents/notes/implemented/testing/2026-09-09-user-patch-hmr-test-delivery.zh.md

@@ -1,4 +1,4 @@
-# Agent Note: 用户 patch 事务控制文件系统事件投递
+# Agent Note: 用户 patch 测试控制文件系统事件投递
 
 Status: implemented
 
@@ -6,22 +6,22 @@ Status: implemented
 
 ## 问题
 
-[macOS Sandbox 运行](https://github.com/deepseek-harness/deepseek-harness/actions/runs/34238200206/job/102101292119) 在等待首次用户 patch 新增时超时。本地并发复现表明，没有文件系统通知到达 HMR。轮询变体也会遗漏后续修改，此时 HMR 没有待执行的刷新。这些失败阻止事务断言执行其负责验证的解析、激活与回滚行为。
+[macOS Sandbox 运行](https://github.com/deepseek-harness/deepseek-harness/actions/runs/34238200206/job/102101292119) 在等待首次用户 patch 新增时超时。本地并发复现表明，没有文件系统通知到达 HMR。轮询变体也会遗漏后续修改，此时 HMR 没有待执行的刷新。这些失败阻止刷新断言执行其负责验证的解析、激活与回滚行为。
 
 ## 决策
 
-[用户 patch 事务测试](../../../../packages/boot/app-boot/tests/user-patches.spec.ts) 写入真实 patch 文件，并通过不持有原生监听句柄的 Chokidar watcher 投递 add、change 和 unlink 事件。HMR 注册、刷新串行化、Include 重组、插件激活、失败广播、回滚与恢复仍使用真实实现。即使初始化在进入局部清理块前失败，夹具也会恢复 watcher 工厂并销毁 Context。
+[用户 patch 测试](../../../../packages/boot/app-boot/tests/user-patches.spec.ts) 写入真实 patch 文件，并通过不持有原生监听句柄的 Chokidar watcher 投递 add、change 和 unlink 事件。应用启动监视器注册、刷新串行化、Include 重组、插件激活、失败报告与恢复仍使用真实实现；[Loader 策略](../simplification/2026-09-09-nontransactional-loader.zh.md) 不提供插件回滚。即使初始化在进入局部清理块前失败，夹具也会恢复 watcher 工厂并销毁 Context。
 
-独立的 [HMR 配置测试](../../../../packages/boot/app-boot/tests/hmr-config.spec.ts) 负责原生通知投递，包括 add/change/unlink、初始不存在的父目录和文件系统别名。事务测试不验证操作系统的投递保证。
+独立的 [监视器测试](../../../../packages/boot/app-boot/tests/watch-config.spec.ts) 负责原生通知投递，包括 add/change/unlink、初始不存在的父目录和文件系统别名。刷新测试不验证操作系统的投递保证。
 
 ## 考虑过的替代方案
 
-**每个事务断言都使用原生通知。** 不采用，因为这会让每次解析器与激活状态转换都重复依赖原生投递。事件缺失会掩盖下游究竟哪个行为出现问题。
+**每个刷新断言都使用原生通知。** 不采用，因为这会让每次解析器与激活状态转换都重复依赖原生投递。事件缺失会掩盖下游究竟哪个行为出现问题。
 
 **轮询与固定等待。** 不采用，因为两者都不能确认下一次修改已经投递。Chokidar 就绪状态不暴露 Node 异步初始轮询基线的完成时刻；本地轮询复现仍会遗漏修改。延长测试期限无法恢复从未发出的事件。
 
-**Mock HMR 注册或 Include。** 不采用，因为测试必须保留事务重组，以及激活和解析失败后的最后有效状态断言。
+**Mock HMR 注册或 Include。** 不采用，因为测试必须保留真实重组、报告激活失败，并在解析失败后保留运行中的配置。
 
 ## 影响
 
-事务序列保留所有语义断言，并移除固定的 change 节流等待。独立并发进程验证隔离性，强制初始化失败则验证 watcher 在下一用例前关闭、工厂在下一用例前恢复。原生 watcher 失败仍在其所属测试中可见，需要单独诊断。
+刷新序列保留其语义断言，并移除固定的 change 节流等待。独立并发进程验证隔离性，强制初始化失败则验证 watcher 在下一用例前关闭、工厂在下一用例前恢复。原生 watcher 失败仍在其所属测试中可见，需要单独诊断。

.agents/notes/implemented/testing/2026-09-10-hosted-image-test-assumptions.i18n.yaml

@@ -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/testing/2026-09-10-hosted-image-test-assumptions.md
+2026-09-10-hosted-image-test-assumptions.md: 5cf3b14503162d67d0fbb27743228cb1a8d7ffe2
+2026-09-10-hosted-image-test-assumptions.zh.md: db82566c11b50b9203d30f7be6aec1378512f94b

.agents/notes/implemented/testing/2026-09-10-hosted-image-test-assumptions.md

@@ -0,0 +1,41 @@
+# Agent Note: Hosted-image assumptions in the coverage suite
+
+Status: implemented
+
+English | [中文](2026-09-10-hosted-image-test-assumptions.zh.md)
+
+## Problem
+
+The [failover leg](../process/2026-09-09-blacksmith-failover-leg.md) runs this suite on pools this repository does not own — Blacksmith's ephemeral images, and the in-house `vm-backup` and `dsh-win-ci` standbys. On the hosted image the coverage lanes failed on host properties their cases never named: whether the host offered a usable user-systemd scope decided which containment a mocked PTY exit raced; the wall-clock grace a managed scope needed before it could take a `SIGKILL` was below what a loaded image provides; a starved reader coalesced writes the illegal-UTF-8 residual cases assumed arrived as separate chunks; and a Windows Server image refuses `CoCreateInstance(CLSID_FileOpenDialog)` outright.
+
+## Decision
+
+Every case names the host property it depends on, so the same revision reports the same verdict on the in-house pool and on a hosted image.
+
+Terminal cases that drive a mocked PTY exit pin the containment they need (`internals = { platform: 'darwin' }` in `packages/subprocess/subprocess-local/tests/local.spec.ts`); under the host's native scope the mocked exit races the scope bootstrap and fails as `terminal scope exited before its bootstrap consumed the launch request`. Mocking the `linux-scope.ts` probes to reach the same path was removed: the platform pin skips both probes, so the mock could not change the selected path.
+
+`disposal contains a spawn-failure rejection that races teardown` asserts the settlement contract instead of one winner of the race: a bootstrap that published its pre-exec failure rejects with that failure, and a teardown that stopped the bootstrap first settles as the requested `SIGTERM`. Only the Linux scope records the stopped arm, because the win32 job owner turns a cancelled start into a rejection and the fallback launcher rejects the missing directory.
+
+`plugin-config dispose graces reach the real ACP run` configures 5000ms dispose graces. At 150ms the hosted image escalated while the scope could not take the signal — `systemctl` failed the kill (`Failed to send signal SIGKILL to auxiliary processes: Invalid argument`) and the teardown reported a failure the configuration never asked for. Its mock refuses stdin EOF and `SIGTERM` by design, so the case waits out both graces (~10s) and carries a 30s case budget, above the 5000ms default the local unit entry grants.
+
+Both illegal-UTF-8 residual cases in `packages/experimental/code-runtime-python/tests/runtime.spec.ts` pace their writes with `time.sleep(0.001)`: `os.sched_yield()` lets a loaded reader coalesce the writes into one chunk, and the coalesced chunk is what the wrapped `Buffer.concat` measures (the hosted image measured 2563 against the 2048 bound with a correct implementation). Their payloads stay above that bound — 3200 bytes for the `0xFF` case and 1100 `ED A0 80` sequences, 3300 raw bytes, for the CESU-8 case, past the 3072-byte budget a raw-byte undercount reaches — so the undercount still flushes above 2048. Each carries a 20s case budget for the paced writes plus the interpreter start.
+
+The stray-output sealing test in `packages/experimental/code-runtime-python/tests/stray-fragments.spec.ts` keeps a real Python child but splits its stdout reads into single-byte events. OS pipe coalescing cannot guarantee the 1024 fragments needed to seal a block: [run 34465259316](https://github.com/deepseek-harness/deepseek-harness/actions/runs/34465259316) passed all assertions but missed that branch. The controlled reads exercise repeated sealing and the final newline merge; exact output and bounded copy volume detect dropped bytes and repeated prefix copies.
+
+The Linux coverage lane grants `DSH_COVERAGE_TEST_TIMEOUT_MS: '90000'`, matching the Windows coverage lane, because the disposal cases in `subprocess-local` and `bash-sandbox` exceed the 5000ms default when the lane's partitions, workers, and sibling gates share one host.
+
+The Windows folder-dialog smoke probes `CoCreateInstance(CLSID_FileOpenDialog)` through PowerShell instead of gating on `process.platform`. An image that answers `CLASS_E_CLASSNOTAVAILABLE` (0x80040111) runs the clean-rejection case and skips the real-dialog case, so `win32-dialog.ts` keeps its file coverage without a host that can open a dialog. Every exception from that activation reads as refusal, so a host failing the probe for another reason only loses the real-dialog case; a probe that cannot run at all keeps the win32 assumption.
+
+## Alternatives considered
+
+**Excluding the coverage lanes from the hosted leg.** Rejected: the leg exists to run the same suite on another pool, and the failures named real host dependencies rather than a suite the pool cannot support.
+
+**Raising only the lane's per-test budget.** Rejected: a wider budget does not change the cases whose cost or behaviour is deterministic — a trapped ACP child still waits out both graces, and a coalesced reader still inflates the measured peak.
+
+**Keeping the `linux-scope.ts` probe mocks.** Rejected as inert: the platform pin selects the fallback path before either probe is called, so the mock changed no execution path.
+
+**Cutting the illegal-UTF-8 payloads to keep the cases fast.** Rejected: below the 2048 bound the assertion can no longer fail for the undercount it names, which leaves the regression unguarded.
+
+## Consequences
+
+The suite's verdict no longer depends on which pool served the lane, at the cost of fixtures pinned to one containment choice: the `linux-scope` and win32-job paths keep their own dedicated cases instead of being reached through these ones. The ACP dispose case costs about 10s of wall clock per run and each residual case about 3.5s, all deterministic rather than host-paced. Hosted-image evidence: [run 34449848541](https://github.com/deepseek-harness/deepseek-harness/actions/runs/34449848541) failed on these cases, [run 34457655892](https://github.com/deepseek-harness/deepseek-harness/actions/runs/34457655892) is green with this diff plus the 90000ms lane budget, and `windows node 24 / coverage` is green on five consecutive hosted runs, where the probe reports the refusal (`clsid-probe=refused`).

.agents/notes/implemented/testing/2026-09-10-hosted-image-test-assumptions.zh.md

@@ -0,0 +1,41 @@
+# Agent Note: coverage 套件里的托管镜像假设
+
+Status: implemented
+
+[English](2026-09-10-hosted-image-test-assumptions.md) | 中文
+
+## 问题
+
+[故障切换支路](../process/2026-09-09-blacksmith-failover-leg.zh.md)会把这套测试跑在本仓库不拥有的池上——Blacksmith 的临时镜像，以及自有的 `vm-backup` 与 `dsh-win-ci` 备用池。在托管镜像上，coverage 各通道的失败来自用例从未点明的宿主属性：宿主是否提供可用的用户级 systemd scope，决定了被 mock 的 PTY 退出会与哪种 containment 竞争；托管 scope 接受 `SIGKILL` 之前所需的墙钟宽限，低于负载镜像实际提供的量；读端被抢占时会把非法 UTF-8 残余用例假定为独立分块的写入合并成一个分块；以及 Windows Server 镜像直接拒绝 `CoCreateInstance(CLSID_FileOpenDialog)`。
+
+## 决策
+
+每个用例都点明它依赖的宿主属性，因此同一份修订在自有池与托管镜像上给出同样的结论。
+
+驱动被 mock 的 PTY 退出的终端用例钉死自己需要的 containment（`packages/subprocess/subprocess-local/tests/local.spec.ts` 中的 `internals = { platform: 'darwin' }`）；在宿主的原生 scope 下，被 mock 的退出会与 scope 的 bootstrap 竞争，并以 `terminal scope exited before its bootstrap consumed the launch request` 失败。为走到同一路径而 mock `linux-scope.ts` 的探针已被删除：平台钉死会让两个探针都不被调用，因此该 mock 无法改变选中的路径。
+
+`disposal contains a spawn-failure rejection that races teardown` 断言结算契约，而不是这场竞争的某一方获胜：已经发布其 pre-exec 失败的 bootstrap 以该失败 reject，先停住 bootstrap 的 teardown 则以被请求的 `SIGTERM` 结算。只有 Linux scope 会记录停止这一支，因为 win32 job owner 会把被取消的启动转成 rejection，fallback 启动器则因目录缺失而 reject。
+
+`plugin-config dispose graces reach the real ACP run` 配置 5000ms 的 dispose 宽限。在 150ms 时，托管镜像在 scope 还无法接受信号时就升级了信号——`systemctl` 的 kill 失败（`Failed to send signal SIGKILL to auxiliary processes: Invalid argument`），teardown 上报了一个配置从未要求的失败。该用例的 mock 按设计既拒绝 stdin EOF 也拒绝 `SIGTERM`，因此用例会等满两个宽限（约 10s），并自带 30s 的用例预算，高于本地单测入口授予的 5000ms 默认值。
+
+`packages/experimental/code-runtime-python/tests/runtime.spec.ts` 的两个非法 UTF-8 残余用例都用 `time.sleep(0.001)` 控制写入节奏：`os.sched_yield()` 会让被抢占的读端把多次写入合并成一个分块，而被包裹的 `Buffer.concat` 测量的正是该分块（在正确实现下，托管镜像测得 2563，超过了 2048 的界）。两个用例的载荷都保持在该界之上——`0xFF` 用例 3200 字节，CESU-8 用例 1100 个 `ED A0 80` 序列（3300 原始字节，超过按原始字节计费会触及的 3072 字节预算）——因此少计仍然会在 2048 之上触发 flush。两者各自带有 20s 的用例预算，容纳带节奏的写入与解释器启动。
+
+`packages/experimental/code-runtime-python/tests/stray-fragments.spec.ts` 的原生输出分块封存测试保留真实 Python 子进程，但把 stdout 读取拆成单字节事件。操作系统的管道合并无法保证达到封存一块所需的 1024 个片段：[run 34465259316](https://github.com/deepseek-harness/deepseek-harness/actions/runs/34465259316) 的全部断言通过，却未覆盖该分支。可控读取覆盖反复封存和末尾换行合并；精确输出与复制总量上限检测字节丢失和前缀反复复制。
+
+Linux coverage 通道授予 `DSH_COVERAGE_TEST_TIMEOUT_MS: '90000'`，与 Windows coverage 通道一致，因为当该通道的分区、worker 与同级门禁共用一个宿主时，`subprocess-local` 与 `bash-sandbox` 的处置用例会超过 5000ms 默认值。
+
+Windows 文件夹对话框冒烟测试改为通过 PowerShell 探测 `CoCreateInstance(CLSID_FileOpenDialog)`，而不再按 `process.platform` 分流。回答 `CLASS_E_CLASSNOTAVAILABLE`（0x80040111）的镜像会跑干净的拒绝用例并跳过真实对话框用例，因此 `win32-dialog.ts` 在没有可开对话框的宿主上仍保有文件覆盖率。该激活过程抛出的任何异常都按拒绝解读，因此因其它原因探测失败的宿主只会失去真实对话框用例；完全无法运行的探测则保留 win32 假设。
+
+## 备选方案
+
+**把 coverage 通道排除出托管支路。** 否决：该支路的意义就是把同一套测试跑在另一个池上，而这些失败点出的是真实的宿主依赖，不是一个该池无法支撑的套件。
+
+**只抬高通道的每用例预算。** 否决：更宽的预算改变不了那些成本或行为确定的用例——被 trap 的 ACP 子进程仍然会等满两个宽限，被合并的读端仍然会抬高测得的峰值。
+
+**保留 `linux-scope.ts` 的探针 mock。** 因无效而否决：平台钉死会在任一探针被调用前就选中 fallback 路径，因此该 mock 没有改变任何执行路径。
+
+**削减非法 UTF-8 的载荷以让用例更快。** 否决：低于 2048 的界之后，断言再也无法为它所点名的少计而失败，等于让该回归失去守护。
+
+## 后果
+
+套件的结论不再取决于哪个池服务了这条通道，代价是被钉在某一 containment 选择上的 fixture：`linux-scope` 与 win32-job 两条路径仍由各自的专用用例覆盖，而不是经由这些用例抵达。ACP 处置用例每次运行约 10s 墙钟，每个残余用例约 3.5s，且都是确定成本而非随宿主浮动。托管镜像证据：[run 34449848541](https://github.com/deepseek-harness/deepseek-harness/actions/runs/34449848541) 在这些用例上失败，[run 34457655892](https://github.com/deepseek-harness/deepseek-harness/actions/runs/34457655892) 在本改动加 90000ms 通道预算下转绿，`windows node 24 / coverage` 在连续五次托管运行中为绿，其中探针报告拒绝（`clsid-probe=refused`）。

.agents/notes/proposed/feature/2026-07-06-recallable-compaction.i18n.yaml

@@ -2,5 +2,5 @@
 # side as of the last confirmed-consistent state. Both languages carry equal authority;
 # after editing either side, bring the other along and re-record with:
 #   pnpm run verify-translation-pairing --write .agents/notes/proposed/feature/2026-07-06-recallable-compaction.md
-2026-07-06-recallable-compaction.md: f0cf1b0602ad7dd5ae719fdd1e3b569b0bf5b448
-2026-07-06-recallable-compaction.zh.md: 8e8793f28be848e6231e86a3eaba099e68d9947e
+2026-07-06-recallable-compaction.md: 07b89116b72a1c1aac2d79ee5aa41ff6e5d02a4c
+2026-07-06-recallable-compaction.zh.md: 7347cb08c94665a6e5bb88655ec08d04921d08eb