Merge origin/master at 52ec34f796

2026-08-04 18:01:22 +08:00
parent 8e4a1b3396 7f3ea406e1
commit 8a24a68333
196 changed files with 3768 additions and 1267 deletions
@@ -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: 02347f2964942b89ec1f0a6ec483f4c2b2f9e68c
-2026-07-23-client-plugin-loading-model.zh.md: ea927d35860fbbba567c47cea0ee3a45133ce0f4
+2026-07-23-client-plugin-loading-model.md: 2dc0c68e5f20bd790c2362f92c16dece171babf5
+2026-07-23-client-plugin-loading-model.zh.md: ce7850e37b9ae2735565a35ce3de28f4f290ed04
@@ -14,7 +14,9 @@ The browser client runs the same cordis plugin mechanism, so it needs the same s

 Conventional frontend engineering digests all dependencies at build time: one bundle, externals resolved by the bundler, nothing left to manage at runtime. Runtime module management on top of that is the unusual requirement here. The client therefore splits into two layers: the upper layer is cordis plugin loading through the same vendored Loader, and the lower layer is module-granular dependency management — `dsh-client-modules`.

-The lower layer supplies four capabilities: externals (the platform list), remote arrival (bundle fetch plus lazy factory registration), versioning (content-hash revs), and hot update (invalidate/prefetch).
+The lower layer supplies four capabilities: externals (the platform list), remote arrival (same-origin external classic scripts plus lazy factory registration), versioning (content-hash revs), and hot update (invalidate/prefetch).
+
+Plugin bundles are built independently outside Vite's module graph. Feeding response text into an inline script leaves the browser with a dynamic source execution: no standard source-map chain connects the network resource, generated bundle, and TypeScript/TSX source, so performance profiles and stacks stop at generated `client.js`; the module system must also buffer the complete source and split one arrival responsibility across fetch and execute transport seams.

 On top of that, client and host plugins register and load consistently: a package declares `dshClient` once, the host scans the declaration into the boot graph, and the same Loader semantics govern entries on both sides.

@@ -46,10 +48,18 @@ Four edge rules govern imports across the two kinds. None of them depends on any

 The browser mirrors the host's division of labor. `dsh-client-modules` (`ClientModuleSystem`) takes the module-system seat that Node's internal ESM loader holds host-side; the same vendored `@cordisjs/plugin-loader` keeps the governance seat on both sides. The line between them in one sentence: **the module system owns module identity and bytes — how code arrives, registers, and becomes an export surface; the Loader owns plugin lifecycle — when a plugin mounts, what it waits for, and how it is torn down.**

-`ClientModuleSystem` is a lazy CJS table. Executing a bundle only **registers** its factory — the bundle calls `window.__ModuleLoader__.load({ id, factory })` and nothing else happens. Every module body side effect, CSS injection included, lives inside the factory closure and runs at materialization: the first `require`/import of that id, memoized after that. A factory that requires a registered-but-unmaterialized sibling materializes it recursively, so no sort order exists anywhere. When asked to import an id, the table resolves through a fixed branch order: seed word → memoized record → static registration (shell-own modules, e.g. app-shell) → registered factory → graph-row fetch + execute → loud throw. That final throw is the runtime mirror of the build-time purity gate. The system also keeps per-module bookkeeping — owned `<style data-plugin>` tag ids, observed require edges — and exposes the two verbs HMR needs: `prefetch(id)` (fetch + execute, registration only; concurrent calls share one in-flight task) and `invalidate(id)` (drop factory, record, and consumed text so the next arrival refetches).
+`ClientModuleSystem` is a lazy CJS table. Executing a bundle only **registers** its factory — the bundle calls `window.__ModuleLoader__.load({ id, factory })` and nothing else happens. Every module body side effect, CSS injection included, lives inside the factory closure and runs at materialization: the first `require`/import of that id, memoized after that. A factory that requires a registered-but-unmaterialized sibling materializes it recursively, so no sort order exists anywhere. When asked to import an id, the table resolves through a fixed branch order: seed word → memoized record → static registration (shell-own modules, e.g. app-shell) → registered factory → graph-row external classic-script load → loud throw. That final throw is the runtime mirror of the build-time purity gate. The system also keeps per-module bookkeeping — owned `<style data-plugin>` tag ids, observed require edges — and exposes the two verbs HMR needs: `prefetch(id)` (load the script and register its factory; concurrent calls share one in-flight task) and `invalidate(id)` (drop the factory and record so the next arrival reloads it).

 The vendored Loader consumes the module system through its `internal` seam — the only call site is `tree.import` — and owns everything entry-shaped: entry creation, fiber activation through cordis service waiting (PENDING until injected services exist, cascading when a service is provided), update/refresh, teardown. The governance code is byte-identical to the host side, per vendor policy. Browserization is compile-time mapping in the shell's vite config: a `node:module` stub alias plus `process.*` defines make `ModuleLoader.fromInternal()` return undefined — exactly the empty slot the shell fills. The module system mounts as `ctx.modules`.

+### External-script arrival and source maps
+
+Each graph row's `url` goes to a same-origin external classic `<script src>` with `async` set. The browser owns the network request and script execution; the node is removed as soon as `load` or `error` settles so HMR cannot accumulate dead nodes. Successful settlement also requires the graph row's factory id to exist in the module table, or arrival fails; registration still does not run the factory, so the side-effect boundary remains first materialization.
+
+The shared tsdown preset emits `client.js.map` for every plugin and rewrites first-party source paths into the browser-resolvable repository shape `/packages/<group>/<package>/src/...`. Other workspace sources inlined into a bundle likewise resolve to their `packages/` owner, while dependency paths remain unchanged; `sourcesContent` carries the source, so the host only serves the map at `/plugins/<id>/client.js.map` and exposes no source route. The Vite shell also emits source maps, letting both shell code and out-of-graph plugins map stacks and performance profiles back to TypeScript/TSX.
+
+`rev` remains the script URL's query parameter and content-consistency anchor, and the bundle and map are both served with `no-cache`. An external script's `error` event exposes neither response status nor body, so failure diagnostics name only the URL; the same-origin host and build-stamped handoff id form the identity boundary, while the post-`load` factory-presence check rejects an artifact that did not register the expected id.
+
 ### The loading flow, end to end

 What happens between `dsh web` starting and the UI appearing? Three stages: the host composes and serves a graph, the shell prefetches, then cordis orchestrates.
@@ -58,11 +68,11 @@ What happens between `dsh web` starting and the UI appearing? Three stages: the

 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, and `--dev` appends the `client-hmr` row in code (`AppCLIEntry`) before the host activation audit so the same check covers it. 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)).
 2. The `dsh-client-modules` node half (the package is dual-face: its browser half is the module table) scans loader entries' package.json `dshClient` declarations and composes `window.__DSH_BOOT__`: `{ rev, entries: [{ id, url, rev, inject?, immediately? }] }`. The `inject` edges and the `immediately` mark come from manifests, never hand-copied. 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 name forever 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. Each bundle's content hash is its `rev` (cache busting + HMR diff anchor), the row set hashes into `graph.rev`, and every row is fetch-served: `/plugins/<id>/client.js?rev=…`. The graph types are single-sourced in the modules package's `./impl` export — the webserver knows nothing about the graph (it is a plain route-registration plugin; modules registers the bundle route and taps the index render itself).
+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 name forever 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. Each bundle's content hash is its `rev` (cache busting + HMR diff anchor), the row set hashes into `graph.rev`, and every row is served as a script resource at `/plugins/<id>/client.js?rev=…`, with its source map at the same path plus `.map`. The graph types are single-sourced in the modules package's `./client` export — the webserver knows nothing about the graph (it is a plain route-registration plugin; modules registers the bundle route and taps the index render itself).

 Why is the roster yml rows and not a scan? Because which plugins compose into a deployment is a composition decision, not a package property — a dshClient package existing in the repo does not mean this deployment mounts it, so discovery-by-scan cannot make that call; the node half scans only what the tree actually mounted.

-**Phase one — the module face.** The shell builds the module system over the graph, then prefetches every `immediately` row in parallel. Prefetch is fetch + execute, which registers factories only. A single row's prefetch failure is swallowed here: phase two's import retries the fetch and owns the loud failure, so one bad row cannot mask the others. `immediately` is a prefetch mark — not a barrier, not an identity. The package declares it, the registry carries it into the row. The infrastructure plugins (connection, runtime, ui-theme, i18n, plus hmr) declare it; UI plugins simply arrive on demand.
+**Phase one — the module face.** The shell builds the module system over the graph, then prefetches every `immediately` row in parallel. Prefetch loads the external script and registers its factory only. A single row's prefetch failure is swallowed here: phase two's import retries the load and owns the loud failure, so one bad row cannot mask the others. `immediately` is a prefetch mark — not a barrier, not an identity. The package declares it, the registry carries it into the row. The infrastructure plugins (connection, runtime, ui-theme, i18n, plus hmr) declare it; UI plugins simply arrive on demand.

 **Phase two — the plugin face.**

@@ -81,7 +91,7 @@ How does a rebuilt bundle become a reload signal? The hmr node half observes it
 On the browser side, the driver reloads one plugin per frame, serialized:

 1. `invalidate` — drop the stale factory and record. A live factory would make the next step a no-op.
-2. `prefetch` — fetch + execute + register the fresh factory, while the old fiber still serves.
+2. `prefetch` — load the external script and register the fresh factory, while the old fiber still serves.
 3. `registry.delete` — before touching the fiber. A bare fiber dispose trips the vendored Loader's self-dispose branch, which would disable the entry permanently.
 4. Drain the old fiber's disposers.
 5. Remove owned `<style data-plugin>` tags.
@@ -112,9 +122,9 @@ The support boundary, stated honestly. Reload is coarse by design: fresh fiber,

 ## Consequences

-One governance implementation runs on both sides of the wire; the browser-specific surface is one module system plus one reload plugin. Plugin packages have one shape, so the purity gate covers them all. Dependency edges and the boot tier live with their owners — the manifests — while the composing app holds only the roster and the `--dev` switch. The drift classes stay structurally closed: share-list hand-sync, load-order coupling, cross-plugin imports, roster/tier double bookkeeping.
+One governance implementation runs on both sides of the wire; the browser-specific surface is one module system plus one reload plugin. Plugin packages have one shape, so the purity gate covers them all. Dependency edges and the boot tier live with their owners — the manifests — while the composing app holds only the roster and the `--dev` switch. The drift classes stay structurally closed: share-list hand-sync, load-order coupling, cross-plugin imports, roster/tier double bookkeeping. Browser-native script loading preserves the standard mapping among plugin network resources, generated bundles, and TypeScript/TSX sources, while the module system keeps only one replaceable `loadBundle` seam.

-Costs accepted: the vendored Loader carries idle machinery in the browser (EntryTree persistence is a no-op, groups/isolation unused); every plugin edit in dev pays a bundle rebuild plus fiber remount; graph `inject` rows are informational — activation truth is service-level — so a mismatch surfaces at the settled sweep, not at graph validation; and the three not-yet-promoted libraries keep their static-import export surface until their DI conversions land.
+Costs accepted: the vendored Loader carries idle machinery in the browser (EntryTree persistence is a no-op, groups/isolation unused); every plugin edit in dev pays a bundle rebuild plus fiber remount; graph `inject` rows are informational — activation truth is service-level — so a mismatch surfaces at the settled sweep, not at graph validation; the three not-yet-promoted libraries keep their static-import export surface until their DI conversions land; every bundle gains a source-map artifact; and external-script failures provide only coarse URL diagnostics instead of the HTTP status available to an explicit fetch.

 Roster endgame (landed 2026-07-25 with the config-tree boot move): the roster lives in `apps/cli/config/web.cordis.yml`, `mountWebPlugins` and the `CLIENT_PACKAGES` constant are gone, and recomposing a deployment means swapping the yml/overlay. The graph composer moved from a webserver-side registry into the `dsh-client-modules` node half (the package upgraded to dual-face per this note's promotion rule — its consumer now reaches it through cordis DI), and the transport split landed alongside: the webserver became a plain route-registration plugin, `/api/*` binding moved to the connection node half over the upgraded `api-gateway` plugin (`dsh-host-apiproxy` providing `ctx.apiProxy`), and the dev bundle watch + SSE channel moved to the hmr node half.

@@ -129,4 +139,5 @@ Roster endgame (landed 2026-07-25 with the config-tree boot move): the roster li
 | Import maps | Ruled out earlier; the DI require table is the terminal mechanism |
 | Full ctx-ification now (react and libraries via services, no module table) | The module-axis extreme; parked — the upgrade law walks there one package at a time instead |
 | Eager instantiation with a frozen table | Requires arrival-time ordering; lazy CJS registration makes recursive `require` self-ordering and matches the naive-puller phase split |
+| Fetch response text, then inject an inline `<script>` | Makes the module system buffer the complete source and maintain separate fetch/execute seams; dynamic source execution also breaks the browser-native association among the network resource, source map, and profile |
 | Builder-push rebuild channel (`POST /plugins/rebuilt` from the orchestrator's `onSuccess`) | Couples reload to one blessed builder process and a second wire protocol; the webserver already holds every bundle path, and stat polling covers the torn-write race (re-hash on every stat change) that once justified pushing |
@@ -14,7 +14,9 @@ host 侧，cordis 插件装载站在 Node 的模块机制之上——require cac

 常规前端工程在构建期消化全部依赖：单一 bundle，external 由打包器解决，运行时无物可管。在此之上再做运行时模块管理，正是这里的特殊需求。client 因此拆成两层：上层是经同一份 vendored Loader 的 cordis 插件装载，下层是模块粒度的依赖管理——`dsh-client-modules`。

-下层供给四项能力：external（平台清单）、远程到达（bundle 拉取加惰性工厂登记）、版本化（内容哈希 rev）、热更新（invalidate/prefetch）。
+下层供给四项能力：external（平台清单）、远程到达（同源外部 classic script 加惰性工厂登记）、版本化（内容哈希 rev）、热更新（invalidate/prefetch）。
+
+插件 bundle 独立构建在 Vite 模块图之外。若把响应文本塞进内联 script，浏览器只能看到一次动态源码执行：网络资源、生成 bundle、TypeScript/TSX 源码之间没有标准 sourcemap 链，性能 profile 与 stack 只能落到生成后的 `client.js`；模块系统还要持有整份源码文本，并把同一项到达职责拆成 fetch 与 execute 两道传输 seam。

 在此之上，client 与 host 插件以一致的方式注册与装载：包声明一次 `dshClient`，host 把声明扫描进 boot 图，同一套 Loader 语义在两侧治理 entry。

@@ -46,10 +48,18 @@ manifest 拥有包的装载契约：它的 `inject` 依赖边，加可选的 `im

 浏览器复刻 host 侧的分工。`dsh-client-modules`（`ClientModuleSystem`）坐上 host 侧由 Node 内部 ESM loader 占据的模块系统席位；同一份 vendored `@cordisjs/plugin-loader` 在两侧都坐治理席。二者的分界线一句话说尽：**模块系统拥有模块身份与字节——代码怎么到达、怎么登记、怎么变成导出面；Loader 拥有插件生命周期——插件何时挂载、等待什么、如何拆除。**

-`ClientModuleSystem` 是一张 lazy CJS 表。执行 bundle 只**登记**其工厂——bundle 调用 `window.__ModuleLoader__.load({ id, factory })`，此外什么都不发生。模块体的一切副作用（包括 CSS 注入）都住在工厂闭包里，在物化时运行：物化即该 id 的首次 `require`/import，此后记忆化。工厂若 require 一个已登记未物化的同伴，就递归物化它，因此任何地方都不存在排序。被要求 import 一个 id 时，表按固定分支顺序解析：种子词条 → 记忆化的记录 → 静态登记（壳自有模块，如 app-shell）→ 已登记的工厂 → 图行 fetch + 执行 → 大声抛错。最后这一抛是构建期纯度门禁在运行期的镜像。系统还保管逐模块的簿记——名下 `<style data-plugin>` 标签 id、观测到的 require 边——并暴露 HMR（热模块替换）需要的两个动词：`prefetch(id)`（fetch + 执行、只登记；并发调用共享同一在途任务）与 `invalidate(id)`（丢弃工厂、记录与已消费文本，下次到达即重新拉取）。
+`ClientModuleSystem` 是一张 lazy CJS 表。执行 bundle 只**登记**其工厂——bundle 调用 `window.__ModuleLoader__.load({ id, factory })`，此外什么都不发生。模块体的一切副作用（包括 CSS 注入）都住在工厂闭包里，在物化时运行：物化即该 id 的首次 `require`/import，此后记忆化。工厂若 require 一个已登记未物化的同伴，就递归物化它，因此任何地方都不存在排序。被要求 import 一个 id 时，表按固定分支顺序解析：种子词条 → 记忆化的记录 → 静态登记（壳自有模块，如 app-shell）→ 已登记的工厂 → 图行外部 classic script 加载 → 大声抛错。最后这一抛是构建期纯度门禁在运行期的镜像。系统还保管逐模块的簿记——名下 `<style data-plugin>` 标签 id、观测到的 require 边——并暴露 HMR（热模块替换）需要的两个动词：`prefetch(id)`（加载脚本、只登记工厂；并发调用共享同一在途任务）与 `invalidate(id)`（丢弃工厂与记录，下次到达即重新加载）。

 vendored Loader 经其 `internal` seam 消费模块系统——唯一调用点是 `tree.import`——并拥有一切 entry 形状的事务：entry 创建、fiber 经 cordis 服务等待的激活（注入的服务未就位即保持 PENDING，服务 provide 时级联激活）、update/refresh、拆除。治理代码按 vendor 政策与 host 侧逐字节相同。浏览器化是壳 vite 配置里的编译期映射：一个 `node:module` stub 别名加若干 `process.*` define，使 `ModuleLoader.fromInternal()` 返回 undefined——这正是留给壳来填的空槽。模块系统挂载为 `ctx.modules`。

+### 外部脚本到达与源码映射
+
+每个图行的 `url` 交给一个带 `async` 的同源外部 classic `<script src>`。浏览器拥有网络请求与脚本执行；`load` 或 `error` 结算后节点立即移除，避免 HMR 累积失效节点。成功结算还要求图行对应的工厂 id 已出现在模块表中，否则到达失败；登记仍不运行工厂，副作用边界继续落在首次物化。
+
+共享 tsdown 预设为每个插件产出 `client.js.map`，并把第一方源码路径重写成浏览器可识别的仓库形状 `/packages/<group>/<package>/src/...`。内联进 bundle 的其他 workspace 源码同样回到其 `packages/` 归属，依赖包路径保持原样；`sourcesContent` 承载源码，因此 host 只需在 `/plugins/<id>/client.js.map` 供给 map，无需开放源码路由。Vite 壳也产出 sourcemap，使壳代码与图外插件都能从 stack 和性能 profile 回到 TypeScript/TSX。
+
+`rev` 继续作为脚本 URL 的查询参数和内容一致性锚点，bundle 与 map 都以 `no-cache` 供给。外部脚本的 `error` 事件不给响应状态与正文，因此失败诊断只报告 URL；同源 host 供给与构建期写入的 handoff id 是身份边界，`load` 后的工厂存在性检查负责拒绝未登记预期 id 的产物。
+
 ### 装载流程，端到端

 从 `dsh web` 启动到 UI 出现之间发生了什么？三个阶段：host 组合并供给一张图，壳预取，然后 cordis 编排。
@@ -58,11 +68,11 @@ vendored Loader 经其 `internal` seam 消费模块系统——唯一调用点

 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，`--dev` 由代码（`AppCLIEntry`）在 host 激活检查之前追加 `client-hmr` 行，使同一项检查覆盖它。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.md)）。
 2. `dsh-client-modules` 的 node 半（该包是双面的：浏览器半就是模块表）扫描 loader entry 的 package.json `dshClient` 声明，组合出 `window.__DSH_BOOT__`：`{ rev, entries: [{ id, url, rev, inject?, immediately? }] }`。`inject` 边与 `immediately` 标记都来自 manifest，永不人肉抄写。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）按名永久缓存，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。每个 bundle 的内容哈希是其 `rev`（缓存失效 + HMR diff 锚点），行集合哈希进 `graph.rev`，每一行都经 fetch 供给：`/plugins/<id>/client.js?rev=…`。图类型单源在 modules 包的 `./client` 出口——webserver 对图一无所知（它是朴素路由注册插件；bundle 路由和 index 渲染 tap 都由 modules 自己注册）。
+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）按名永久缓存，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。每个 bundle 的内容哈希是其 `rev`（缓存失效 + HMR diff 锚点），行集合哈希进 `graph.rev`，每一行都作为脚本资源供给：`/plugins/<id>/client.js?rev=…`，对应 sourcemap 位于同一路径加 `.map`。图类型单源在 modules 包的 `./client` 出口——webserver 对图一无所知（它是朴素路由注册插件；bundle 路由和 index 渲染 tap 都由 modules 自己注册）。

 为什么名册是 yml 行而不是扫描？因为哪些插件组合进一次部署是组合决策，不是包属性——一个 dshClient 包存在于仓库里，不代表这次部署要挂载它，扫描发现无从替人做这个决定；node 半只扫描配置树实际挂载了的东西。

-**第一层——模块面。**壳在图之上建起模块系统，然后并行预取每个 `immediately` 行。预取即 fetch + 执行，只登记工厂。单行预取失败在这里被吞下：第二层 import 时会重试 fetch 并拥有那次大声失败，因此一个坏行藏不住其他行。`immediately` 是预取标记——不是屏障，不是身份。包声明它，注册表把它带进图行。基础设施插件（connection、runtime、ui-theme、i18n，外加 hmr）声明它；UI 插件则径直按需到达。
+**第一层——模块面。**壳在图之上建起模块系统，然后并行预取每个 `immediately` 行。预取即加载外部脚本，只登记工厂。单行预取失败在这里被吞下：第二层 import 时会重试加载并拥有那次大声失败，因此一个坏行藏不住其他行。`immediately` 是预取标记——不是屏障，不是身份。包声明它，注册表把它带进图行。基础设施插件（connection、runtime、ui-theme、i18n，外加 hmr）声明它；UI 插件则径直按需到达。

 **第二层——插件面。**

@@ -81,7 +91,7 @@ vendored Loader 经其 `internal` seam 消费模块系统——唯一调用点
 浏览器侧，驱动插件每帧重载一个插件，串行执行：

 1. `invalidate`——丢弃陈旧的工厂与记录。工厂还活着会让下一步变成 no-op。
-2. `prefetch`——fetch + 执行 + 登记新工厂，旧 fiber 此刻仍在服役。
+2. `prefetch`——加载外部脚本并登记新工厂，旧 fiber 此刻仍在服役。
 3. `registry.delete`——先于任何 fiber 操作。裸做 fiber dispose 会触发 vendored Loader 的自 dispose 分支，把 entry 永久停用。
 4. 排空旧 fiber 的各 disposer。
 5. 移除名下的 `<style data-plugin>` 标签。
@@ -112,9 +122,9 @@ vendored Loader 经其 `internal` seam 消费模块系统——唯一调用点

 ## Consequences

-wire 两侧跑着同一份治理实现；浏览器特有的表面只是一套模块系统加一个重载插件。插件包只有一种形态，纯度门禁因此覆盖全部插件。依赖边与启动档位都与其所有者——manifest——同住，负责组合的 app 只握名册与 `--dev` 开关。各漂移缺陷类被结构性关死：共享清单人肉同步、装载顺序耦合、跨插件 import、名册/档位双重记账。
+wire 两侧跑着同一份治理实现；浏览器特有的表面只是一套模块系统加一个重载插件。插件包只有一种形态，纯度门禁因此覆盖全部插件。依赖边与启动档位都与其所有者——manifest——同住，负责组合的 app 只握名册与 `--dev` 开关。各漂移缺陷类被结构性关死：共享清单人肉同步、装载顺序耦合、跨插件 import、名册/档位双重记账。浏览器原生脚本装载使插件网络资源、生成 bundle 与 TypeScript/TSX 源码保持标准映射，模块系统也只保留一道可替换的 `loadBundle` seam。

-接受的代价：vendored Loader 在浏览器里背着闲置机件（EntryTree 持久化是 no-op，分组/隔离未用）；开发期每次修改插件都要付一次 bundle 重建加 fiber 重挂；图中 `inject` 行仅是信息性说明——激活的真相在服务层——因此不匹配会在 settled 扫描时浮出，而不是在图校验时被拦下；三个尚未升格的库在各自的 DI 转换落地之前保持静态 import 的导出面。
+接受的代价：vendored Loader 在浏览器里背着闲置机件（EntryTree 持久化是 no-op，分组/隔离未用）；开发期每次修改插件都要付一次 bundle 重建加 fiber 重挂；图中 `inject` 行仅是信息性说明——激活的真相在服务层——因此不匹配会在 settled 扫描时浮出，而不是在图校验时被拦下；三个尚未升格的库在各自的 DI 转换落地之前保持静态 import 的导出面；每个 bundle 多出一份 sourcemap 产物，外部脚本失败也只能给出粗粒度的 URL 诊断，不能像显式 fetch 那样报告 HTTP 状态。

 名册的终局（2026-07-25 随配置树 boot 迁移落地）：名册住 `apps/cli/config/base.cordis.yml` 与 `apps/cli/config/web.cordis.yml`，`mountWebPlugins` 与 `CLIENT_PACKAGES` 常量已消失，重组一次部署等于换 yml/overlay。图的组合器从 webserver 侧的注册表迁进 `dsh-client-modules` 的 node 半（该包按本 note 的升级法则升格为双面——其消费方现经 cordis DI 到达），传输拆分同轮落地：webserver 变为朴素路由注册插件，`/api/*` 绑定迁到 connection 的 node 半、走升格后的 `api-gateway` 插件（`dsh-host-apiproxy` 提供 `ctx.apiProxy`），dev 的 bundle 监视与 SSE 通道迁到 hmr 的 node 半。

@@ -129,4 +139,5 @@ wire 两侧跑着同一份治理实现；浏览器特有的表面只是一套模
 | import map | 早已排除；DI require 表是终局机制 |
 | 现在就彻底 ctx 化（react 与库全走服务，不设模块表） | 模块轴上的极端形态；搁置——升级法则改为一次一包走向它 |
 | 冻结表 + 到达即实例化 | 要求按到达时刻排序；lazy CJS 登记让递归 `require` 自行定序，且与朴素拉取器的分层相合 |
+| fetch 响应文本后注入内联 `<script>` | 模块系统必须缓冲整份源码并维护 fetch/execute 两道 seam；动态源码执行也切断浏览器网络资源、sourcemap 与 profile 的原生关联 |
 | 构建器推送重建通道（编排器在 `onSuccess` 里 POST `/plugins/rebuilt`） | 把重载耦合到一个钦定的构建器进程和第二套 wire 协议；webserver 本就握有每个 bundle 路径，stat 轮询（每次 stat 变化即重哈希）已兜住当年为推送辩护的撕裂写竞态 |
@@ -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/2026-07-28-themed-scrollbars-and-reserved-gutter.md
-2026-07-28-themed-scrollbars-and-reserved-gutter.md: b45f70b126d083916c756afb88a8b646a4e9bb85
-2026-07-28-themed-scrollbars-and-reserved-gutter.zh.md: 8afa36429ce7e6e061b014d63dcb20e5a642a84c
+2026-07-28-themed-scrollbars-and-reserved-gutter.md: ba3d9d3c94cf9a775c43c14e292187b3a43935fe
+2026-07-28-themed-scrollbars-and-reserved-gutter.zh.md: 51d668f584ee254183d5d1cf709140df2556e8d0
@@ -20,7 +20,7 @@ The rules sit on `body`, not `html`. `design-platform.css` declares the `--dsw-a

 The two renderings are mutually exclusive, and the exclusion is enforced rather than assumed. A non-`auto` `scrollbar-width` or `scrollbar-color` makes Chromium and Safari discard every `::-webkit-scrollbar*` rule for that element, `::-webkit-scrollbar-thumb:hover` included. Declaring both unconditionally therefore leaves the hover token rendering nowhere at all: the engines that implement the hover pseudo-element are exactly the ones the standard properties silence, and Firefox has no hover pseudo-element to fall back on. The standard properties consequently sit inside `@supports not selector(::-webkit-scrollbar)`, which is true only where the pseudo-element is unimplemented, so Firefox takes the standard path and WebKit-based engines take the pseudo-element path. The WebKit rules are not gated in turn: an engine without those pseudo-elements drops them as unknown selectors, so a gate would only restate what selector matching already does. An engine too old for the `selector()` function makes the condition invalid, which evaluates false and selects the pseudo-element path — the correct side for the pre-16.4 Safari that is the realistic case for that reading.

-Both paths read one indirection pair, `--dsh-scrollbar-thumb` and `--dsh-scrollbar-thumb-hover`, bound on `body` to the l1 (base-surface) tokens. **This is the rebinding contract, and it is the part the CSS alone does not state**: an elevated surface sets `--dsh-scrollbar-thumb: var(--dsw-alias-scrollbar-bg-l2)` and `--dsh-scrollbar-thumb-hover: var(--dsw-alias-scrollbar-hover-l2)` on its own container, and that one rebind reaches the standard properties and the WebKit pseudo-elements together. The pair is rebound as a pair; rebinding the resting thumb alone leaves the hover state on the base-surface token. The mechanically discoverable subset is owned by `packages/client/ui-theme/tests/scrollbar-styles.spec.ts`: any sheet that both scrolls and paints an elevated surface must rebind, so this note no longer maintains a complete surface inventory. Most declare the pair on the elevated card rather than on the scrolling descendant, because elevation belongs to the surface and custom properties inherit to whichever child actually scrolls.
+Both paths read one indirection pair, `--dsh-scrollbar-thumb` and `--dsh-scrollbar-thumb-hover`, bound on `body` to the l1 (base-surface) tokens. **This is the rebinding contract, and it is the part the CSS alone does not state**: an elevated surface sets `--dsh-scrollbar-thumb: var(--dsw-alias-scrollbar-bg-l2)` and `--dsh-scrollbar-thumb-hover: var(--dsw-alias-scrollbar-hover-l2)` on its own container, and that one rebind reaches the standard properties and the WebKit pseudo-elements together. The pair is rebound as a pair; rebinding the resting thumb alone leaves the hover state on the base-surface token. `transparent` is the pair's other legal target, added when the sidebar's bars [started following the pointer](../feature/2026-08-04-pointer-revealed-sidebar-scrollbars.md); the gate below admits those two and nothing else. The mechanically discoverable subset is owned by `packages/client/ui-theme/tests/scrollbar-styles.spec.ts`: any sheet that both scrolls and paints an elevated surface must rebind, so this note no longer maintains a complete surface inventory. Most declare the pair on the elevated card rather than on the scrolling descendant, because elevation belongs to the surface and custom properties inherit to whichever child actually scrolls.

 Four surfaces — `Menu`, `InputBar`, `QuestionComposer`, and `TodoPanel` — were missed in the first implementation and found in review, which is why the per-sheet rebinding contract is checked mechanically rather than by inspection.

@@ -54,7 +54,7 @@ The gutter and the sheet's `::-webkit-scrollbar` width are jointly necessary aga

 ## Consequences

- Every scroll container in the client draws the themed thumb: `rgb(229, 229, 229)` on a light base surface, `rgb(60, 60, 61)` on a dark one, and `rgb(84, 85, 87)` for a dark elevated surface that rebinds to the l2 pair.
+- Every scroll container in the client draws the themed thumb: `rgb(229, 229, 229)` on a light base surface, `rgb(60, 60, 61)` on a dark one, and `rgb(84, 85, 87)` for a dark elevated surface that rebinds to the l2 pair. The sidebar's regions draw theirs only under the pointer, through the same indirection.
 - The two renderings are separately specified, so a change to the thumb's geometry or hover behavior has to be made twice — once in `scrollbar-width`/`scrollbar-color`, once in the pseudo-elements. Routing both through the indirection pair confines that duplication to the properties Firefox and WebKit do not share.
 - The hover tokens (`--dsw-alias-scrollbar-hover-l1`/`-l2`) render only on the pseudo-element path. Firefox states one thumb color through `scrollbar-color` and derives its own hover treatment, so a design change to the hover colors is visible in Chromium and Safari and not in Firefox. This is a limit of `scrollbar-color`, not of the sheet.
 - `body *` matches every element, for two properties whose effect the user agent already limits to elements that actually scroll. The cost is a broad selector; the alternative was a rebinding contract that does not work.
@@ -20,7 +20,7 @@ Status: implemented

 两种渲染互斥，而这种互斥是被强制的，不是假定的。`scrollbar-width` 或 `scrollbar-color` 只要取非 `auto` 值，Chromium 与 Safari 就会丢弃该元素上的全部 `::-webkit-scrollbar*` 规则，`::-webkit-scrollbar-thumb:hover` 也在其中。因此无条件地同时声明会让 hover token 在任何地方都得不到渲染：实现了 hover 伪元素的引擎，恰恰就是被标准属性静音的那些，而 Firefox 没有 hover 伪元素可作退路。于是标准属性写在 `@supports not selector(::-webkit-scrollbar)` 之内，该条件只在伪元素未被实现处为真，因此 Firefox 走标准属性路径，WebKit 系引擎走伪元素路径。WebKit 规则不再反向加门禁：不实现这些伪元素的引擎会把它们当作未知选择器丢弃，因此加门禁只是重述选择器匹配本身已经做的事。对于旧到不支持 `selector()` 函数的引擎，该条件无效，从而求值为假并选中伪元素路径——对于这条判断下现实存在的 16.4 之前的 Safari，这正是正确的一侧。

-两条路径都读取同一组间接变量 `--dsh-scrollbar-thumb` 与 `--dsh-scrollbar-thumb-hover`，它们在 `body` 上绑定到 l1（基础表面）token。**这就是重新绑定契约，也是单看 CSS 无法得知的部分**：抬升表面在自己的容器上设置 `--dsh-scrollbar-thumb: var(--dsw-alias-scrollbar-bg-l2)` 与 `--dsh-scrollbar-thumb-hover: var(--dsw-alias-scrollbar-hover-l2)`，这一次重新绑定同时作用于标准属性和 WebKit 伪元素。这组变量必须成对重新绑定；只改静止态滑块会让 hover 状态仍留在基础表面的 token 上。可由机械检查发现的子集归 `packages/client/ui-theme/tests/scrollbar-styles.spec.ts` 所有：任何既滚动又绘制抬升表面的样式表都必须重新绑定，因此本 note 不再维护完整的表面清单。多数把这组变量声明在抬升卡片上而非滚动的后代元素上，因为抬升层级属于这个表面，而自定义属性会继承到真正滚动的那个子元素。
+两条路径都读取同一组间接变量 `--dsh-scrollbar-thumb` 与 `--dsh-scrollbar-thumb-hover`，它们在 `body` 上绑定到 l1（基础表面）token。**这就是重新绑定契约，也是单看 CSS 无法得知的部分**：抬升表面在自己的容器上设置 `--dsh-scrollbar-thumb: var(--dsw-alias-scrollbar-bg-l2)` 与 `--dsh-scrollbar-thumb-hover: var(--dsw-alias-scrollbar-hover-l2)`，这一次重新绑定同时作用于标准属性和 WebKit 伪元素。这组变量必须成对重新绑定；只改静止态滑块会让 hover 状态仍留在基础表面的 token 上。这组变量另一个合法的目标是 `transparent`，它随侧边栏滚动条[改为跟随指针](../feature/2026-08-04-pointer-revealed-sidebar-scrollbars.md)一并引入；下文的门禁只接受这两种目标。可由机械检查发现的子集归 `packages/client/ui-theme/tests/scrollbar-styles.spec.ts` 所有：任何既滚动又绘制抬升表面的样式表都必须重新绑定，因此本 note 不再维护完整的表面清单。多数把这组变量声明在抬升卡片上而非滚动的后代元素上，因为抬升层级属于这个表面，而自定义属性会继承到真正滚动的那个子元素。

 `Menu`、`InputBar`、`QuestionComposer` 与 `TodoPanel` 这四个表面在最初的实现里被漏掉、由评审发现，因此逐样式表的重新绑定契约由机械检查而非人工审阅把关。

@@ -54,7 +54,7 @@ Status: implemented

 ## 后果

- 客户端的每个滚动容器都绘制带主题的滑块：亮色基础表面为 `rgb(229, 229, 229)`，暗色基础表面为 `rgb(60, 60, 61)`，重新绑定到 l2 的暗色抬升表面为 `rgb(84, 85, 87)`。
+- 客户端的每个滚动容器都绘制带主题的滑块：亮色基础表面为 `rgb(229, 229, 229)`，暗色基础表面为 `rgb(60, 60, 61)`，重新绑定到 l2 的暗色抬升表面为 `rgb(84, 85, 87)`。侧边栏内的滚动区域经由同一组间接变量，只在指针到达时才绘制滑块。
 - 两种渲染分别指定，因此改动滑块的几何或 hover 行为需要改两处：一处在 `scrollbar-width`／`scrollbar-color`，一处在伪元素。让两者都经由这组间接变量，把这份重复限制在 Firefox 与 WebKit 不共用的那些属性上。
 - hover token（`--dsw-alias-scrollbar-hover-l1`／`-l2`）只在伪元素路径上渲染。Firefox 通过 `scrollbar-color` 只表述一个滑块颜色，其 hover 表现由引擎自行推导，因此对 hover 颜色的设计改动在 Chromium 与 Safari 上可见，在 Firefox 上不可见。这是 `scrollbar-color` 本身的限制，不是这张样式表的限制。
 - `body *` 匹配所有元素，涉及的两个属性其效果本就被浏览器限制在实际会滚动的元素上。代价是一个覆盖面很宽的选择器；另一种选择是一个不生效的重新绑定契约。
@@ -2,5 +2,5 @@
 # side as of the last confirmed-consistent state. Both languages carry equal authority;
 # after editing either side, bring the other along and re-record with:
 #   pnpm run verify-translation-pairing --write .agents/notes/implemented/bug-fix/2026-07-30-approval-panel-command-cap.md
-2026-07-30-approval-panel-command-cap.md: 941f7eda187f263f2d8af6aa643d493c92a3669b
-2026-07-30-approval-panel-command-cap.zh.md: 939a700934f6467947028d988da9a694169e203e
+2026-07-30-approval-panel-command-cap.md: 1e4b75c106b71605a004ef35301445da76ec8cf0
+2026-07-30-approval-panel-command-cap.zh.md: aa9f708b78c8ace735e1c96c0ff9ab0a7a4d526a
@@ -14,7 +14,7 @@ The InputBar the panel replaces has always been capped (14 lines, then the texta

 The panel's justification and command move into one scroll region (`data-approval-scroll`) capped at the same height as the composer's draft area; the amber strip and the action row sit outside it, so both buttons are in the card at every content length.

-The cap is one value with two consumers, declared as `--dsh-composer-text-max-height: 336px` on `ConversationRoot`'s `.composerSeat` — the composer chain's only shared ancestor, since the fallback InputBar and an elected takeover render as siblings. `InputBar`'s mirror and the panel's scroll region both read it, so the seat cannot cap its two states differently: what the designer asked for ("unify it with the input box's max height") is now a fact of the stylesheet rather than a number repeated in two files. The region is `box-sizing: border-box` so the cap is its outer height, the same box the composer's draft area occupies.
+The cap is one value with two consumers, declared as `--dsh-composer-text-max-height: 336px` on `ConversationRoot`'s `.composerSeat` — the composer chain's only shared ancestor, since the fallback InputBar and an elected takeover render as siblings. `InputBar`'s draft scrollport and the panel's scroll region both read it, so the seat cannot cap its two states differently: what the designer asked for ("unify it with the input box's max height") is now a fact of the stylesheet rather than a number repeated in two files. The region is `box-sizing: border-box` so the cap is its outer height, the same box the composer's draft area occupies.

 The region is a tab stop (`tabIndex={0}`, named `role="group"`). Unlike the question composer's scroll body, whose option rows are focusable and pull the container along, this one holds nothing but text: without its own tab stop a keyboard-only user could reach the buttons and never the command's tail, and approve what they could not finish reading.

@@ -34,12 +34,12 @@ The panel's card rebinds `--dsh-scrollbar-thumb{,-hover}` to the l2 pair, as eve

 - A long command scrolls inside the card and the refuse/allow buttons stay on screen. Measured on the built client at 900x1000 and 900x700: the region reports `scrollHeight` past `clientHeight`, and both buttons stay inside the card and inside the viewport.
 - Electing the takeover no longer changes how tall the composer seat can get, so the transcript above it does not reflow by hundreds of pixels when an approval arrives or resolves.
- The InputBar's 14-line cap now resolves through a custom property inherited from `.composerSeat`. Rendering the bar outside that seat would drop the declaration (an unresolved `var()` with no fallback), so a future composer host has to carry the property — which is why it is declared on the shared seat rather than the app root.
+- The InputBar's 14-line cap now resolves through a custom property inherited from `.composerSeat`, on the box that scrolls its draft ([one scrollport for both text layers](2026-07-31-composer-text-layers-share-one-scrollport.md) moved the declaration off the auto-grow mirror). Rendering the bar outside that seat would drop the declaration (an unresolved `var()` with no fallback), so a future composer host has to carry the property — which is why it is declared on the shared seat rather than the app root.
 - The scenario's recorded command is a 200-token blob, far longer than a round trip needs. That cost is deliberate: the cap is unfalsifiable without content that passes it, and the model compresses any regular payload (the first recording turned "alpha 400 times" into `printf 'alpha %.0s' {1..400}`, a one-line command that proves nothing).

 ## Verification

-`apps/web/tests/approval-composer.e2e.ts` drives the real composition: a read-only session, a denied write, the model's escalation retry, and the answer clicked through the panel. The geometry assertion runs on the live panel at two viewport heights and is guarded against holding vacuously — the region must actually be scrolling, and the measured cap must equal the composer's own, which the test reads off the live textarea before sending rather than hardcoding the px value.
+`apps/web/tests/approval-composer.e2e.ts` drives the real composition: a read-only session, a denied write, the model's escalation retry, and the answer clicked through the panel. The geometry assertion runs on the live panel at two viewport heights and is guarded against holding vacuously — the region must actually be scrolling, and the measured cap must equal the composer's own, which the test reads off the live draft scrollport before sending rather than hardcoding the px value.

 Confirmed both directions against the built client. With the cap reverted, the region reports `scrolls: false` and grows to the command's full height (1798px for the recorded blob at 900x1000, against 336px capped); at 900x700 the card is 680px tall against a 700px viewport and the action row's bottom lands at y=749 — below the fold, the designer's report exactly. With the cap restored the scenario passes in replay.

@@ -14,7 +14,7 @@ Status: implemented

 面板的理由与命令移入同一个滚动区域（`data-approval-scroll`），其高度上限与 composer 的草稿区完全相同；琥珀色状态条与操作按钮行位于该区域之外，因此无论内容多长，两个按钮都留在卡片内。

-这个上限是一个值、两个消费者，以 `--dsh-composer-text-max-height: 336px` 声明在 `ConversationRoot` 的 `.composerSeat` 上——它是 composer 链唯一的共同祖先，因为兜底的 InputBar 与被选中的接管面板是兄弟节点。`InputBar` 的 mirror 与面板的滚动区域都读取它，于是同一个容器不可能给它的两种状态设出不同上限：设计同学要求的"可以跟输入框最大高度统一"，如今是样式表中的一个事实，而不是抄在两个文件里的一个数字。该区域取 `box-sizing: border-box`，因此上限指的是它的外框高度，与 composer 草稿区占据的是同一个盒子。
+这个上限是一个值、两个消费者，以 `--dsh-composer-text-max-height: 336px` 声明在 `ConversationRoot` 的 `.composerSeat` 上——它是 composer 链唯一的共同祖先，因为兜底的 InputBar 与被选中的接管面板是兄弟节点。`InputBar` 的草稿滚动容器与面板的滚动区域都读取它，于是同一个容器不可能给它的两种状态设出不同上限：设计同学要求的"可以跟输入框最大高度统一"，如今是样式表中的一个事实，而不是抄在两个文件里的一个数字。该区域取 `box-sizing: border-box`，因此上限指的是它的外框高度，与 composer 草稿区占据的是同一个盒子。

 该区域自身是一个 Tab 停靠点（`tabIndex={0}`，带名称的 `role="group"`）。提问 composer 的滚动体不需要这样做——它的选项行本身可聚焦，会把容器一起带过去；而这里除文本之外别无内容：没有自己的停靠点，仅用键盘的用户能走到按钮却走不到命令尾部，于是可能批准了自己没读完的东西。

@@ -34,12 +34,12 @@ Status: implemented

 - 长命令在卡片内滚动，拒绝／允许按钮留在屏幕内。在构建产物客户端上于 900x1000 与 900x700 实测：该区域报告的 `scrollHeight` 超过 `clientHeight`，两个按钮都留在卡片内、也都留在视口内。
 - 选中接管面板不再改变 composer 容器能达到的高度，因此审批到来或解决时，上方的会话流不会有数百像素的重排。
- InputBar 的 14 行上限现在通过一个自 `.composerSeat` 继承而来的自定义属性解析。把输入栏渲染到该容器之外会丢掉这条声明（一个没有兜底值的未解析 `var()`），因此未来的 composer 宿主必须带上这个属性——这也正是它声明在共享容器上、而不是应用根节点上的原因。
+- InputBar 的 14 行上限现在通过一个自 `.composerSeat` 继承而来的自定义属性解析，且落在真正滚动草稿的那个盒子上（[两层文本共用同一个滚动容器](2026-07-31-composer-text-layers-share-one-scrollport.md)把该声明从自增高镜像层移了出去）。把输入栏渲染到该容器之外会丢掉这条声明（一个没有兜底值的未解析 `var()`），因此未来的 composer 宿主必须带上这个属性——这也正是它声明在共享容器上、而不是应用根节点上的原因。
 - 该场景录制的命令是一段 200 个 token 的字符块，远超一次往返所需。这个代价是有意付出的：没有能越过上限的内容，这个上限无法被证伪，而模型会把任何规整的载荷压缩掉（第一次录制时，模型把"alpha 重复 400 次"写成了 `printf 'alpha %.0s' {1..400}`，一条什么也证明不了的单行命令）。

 ## 验证

-`apps/web/tests/approval-composer.e2e.ts` 驱动的是真实组合：一个只读会话、一次被拒绝的写入、模型的越权重试，以及在面板上点击完成的回应。几何断言在两个视口高度上针对活动面板执行，并有守卫防止它空洞地成立——该区域必须确实处在滚动状态，且实测上限必须等于 composer 自身的上限，后者由测试在发送之前从活动 textarea 上读出，而不是把该像素值写死。
+`apps/web/tests/approval-composer.e2e.ts` 驱动的是真实组合：一个只读会话、一次被拒绝的写入、模型的越权重试，以及在面板上点击完成的回应。几何断言在两个视口高度上针对活动面板执行，并有守卫防止它空洞地成立——该区域必须确实处在滚动状态，且实测上限必须等于 composer 自身的上限，后者由测试在发送之前从活动的草稿滚动容器上读出，而不是把该像素值写死。

 在构建产物客户端上双向确认过。撤销上限后，该区域报告 `scrolls: false`，并长到命令的完整高度（900x1000 下，录制的字符块为 1798px，而设上限后为 336px）；在 900x700 下卡片高 680px、视口高 700px，操作按钮行底边落在 y=749——正在折叠之下，与设计同学的反馈完全一致。恢复上限后，该场景在回放模式下通过。

@@ -1,77 +0,0 @@
-# Agent Note: The composer's glyph layer tracks the textarea's scroll offset
-
-Status: implemented
-
-English | [中文](2026-07-31-composer-glyph-layer-tracks-the-textarea.zh.md)
-
-## Problem
-
-A composer draft longer than the 14-line cap could not be scrolled. The caret moved and the selection moved, but the words stayed frozen at line 1 — no wheel gesture, drag, or arrow key brought the end of a long draft on screen, so the bottom of anything past ~14 lines was unreachable and unreadable while writing it.
-
-The cap itself was working. The composer paints its text in two stacked layers ([InputBar](../../../../packages/client/ui-conversation/src/client/skeleton/InputBar.tsx)): the `<textarea>` owns the value, the selection, and the caret but renders its own glyphs `color: transparent`, and every visible character is painted by the `[data-input-backdrop]` div beneath it, which also carries the claim-token highlight, the chips, and the ghost hint. That split is what makes chips and highlights possible at all — a textarea cannot style a range of its own text.
-
-The two layers were coupled in geometry but not in scroll. The backdrop is `position: absolute; inset: 0; overflow: hidden`: it is clipped, not scrolled, and nothing in the browser links its offset to the textarea's. Below the cap that is invisible, because both layers rest at offset 0 and the mirror div sizes the box to the draft. At the cap the textarea starts scrolling and the backdrop does not follow, so the layer the user actually reads never moves.
-
-The defect is therefore exactly as old as the cap, and it hid behind the resting state: a short draft, the state every screenshot and every existing fixture captured, renders identically with and without the coupling.
-
-## Decision
-
-`InputBar` mirrors the textarea's `scrollTop` onto the backdrop from one `scroll` listener, registered beside the existing wheel-chaining listener in the same effect (the textarea is never unmounted — the inert state renders the same element disabled).
-
-One listener is the whole coupling, because every way the box moves ends in a `scroll` event on the textarea. A gesture scrolls it; an edit scrolls the caret into view; a draft that shrinks past the current offset clamps it. The clamp case is the one that looks like it needs separate handling and does not: the two layers share an extent, so they clamp to the same maximum, and the textarea's clamp fires the `scroll` that mirrors it.
-
-That shared extent is not free, and mirroring an offset is only correct while it holds. Two things break it, both discovered in review, both failing in the same direction — a backdrop shorter than the textarea, so the assignment clamps and the glyphs sit below the caret. A textarea reserves a line box for the caret after a final newline; `white-space: pre-wrap` collapses a text node's trailing newline and generates none. A draft ending in a newline therefore made the backdrop exactly one line shorter than the textarea — measured 628 against 652 — so the assignment clamped and the glyphs sat a line behind the caret at the very bottom. The backdrop now carries the same trailing-line sentinel the mirror div already did: its content is the decoration walk plus one `'\n'`, which the same collapse absorbs when the draft does not end in a newline and which supplies the missing line box when it does. Measured across plain, trailing-newline, soft-wrapping, unbreakable-run, and interior-blank-line drafts, the two extents now agree in every case.
-
-The second premise is wrap width, and it is asserted rather than fixed. Only `.input` scrolls, so only `.input` can lose content width to a scrollbar that consumes layout space, and a narrower `.input` wraps a long draft onto more lines — worth 2 to 5 lines for an 8px difference, measured on a standalone harness, while at equal widths a textarea and a div agree exactly. Measured on the running app across the three engines Playwright ships, the widths agree on two and not on the third:
-
-| engine | `.input` / `.backdrop` / `.mirror` wrap width | extents |
-|---|---|---|
-| chromium | 776 / 776 / 776 | equal |
-| firefox | 776 / 776 / 776 | equal |
-| WebKit | **768** / 776 / 776 | equal for the drafts measured |
-
-WebKit's textarea loses 8px to its scrollbar while the clipped layers keep theirs. That gap predates this change and is not closed here; the mirror is unaffected on the drafts measured because the extents still agree, but a draft whose wrapping is sensitive at exactly that width would make `.input` taller and clamp the mirrored offset. The scenario asserts the equality on the lane's engine, so a regression into that state fails loudly rather than silently.
-
-`scrollbar-gutter: stable` on the shared metrics block was tried and removed. WebKit applies it to `overflow-y: auto` but not to `overflow: hidden`, so it left `.input` at 768 against 776 — exactly the gap it was meant to close — while costing chromium 8px of text width unconditionally. Closing this needs one geometry every engine agrees on, not that property.
-
-The mirror is one-directional: the textarea is the authority because it owns the caret, and the caret is what the browser scrolls to.
-
-## Alternatives considered
-
-**Give the backdrop `overflow: auto` and let it scroll itself.** It would then have a scroll offset of its own to keep in step, which is the same problem plus a second scrollbar painted over the input. The backdrop is a projection of the textarea, not an independently navigable surface.
-
-**Drop the backdrop and style the textarea's own text.** This removes the layer split and the whole class of desync with it. Rejected because it is not implementable: a textarea renders one uniform text run, so the claim-token highlight, the chips, and the ghost hint — the reasons the backdrop exists — have no way to be expressed. Losing them to fix scrolling trades a bounded defect for a feature deletion.
-
-**Render the draft in a `contenteditable` div instead of a textarea.** One element, one scroll offset, styleable ranges. Rejected as far out of proportion to the defect: `contenteditable` would put IME composition, undo/redo, selection semantics, and paste normalization back on us, all of which the textarea plus the input machine currently handle, and the machine already owns an undo log that assumes a textarea's value semantics.
-
-**Scroll the backdrop from the existing wheel handler instead of a `scroll` listener.** The handler already runs on every wheel over the textarea, so it looks like the natural place. Rejected because it covers only one of the ways the box scrolls: typing at the end, `End`, arrow keys, drag-selection past the edge, and scrollbar drags all move the textarea without a wheel event. Listening to `scroll` is listening to the thing itself rather than to one of its causes.
-
-**Reserve the scrollbar gutter on all three layers with `scrollbar-gutter: stable`.** Adopted, then reverted on measurement. The reasoning was that whatever a platform's scrollbar costs, three layers reserving it stay equal — and `overflow: hidden` is a scroll container, so the spec says the clipped layers honour it. Chromium agrees (8px reserved on each, widths 768/768/768). WebKit does not: it reserves for `overflow-y: auto` and not for `overflow: hidden`, leaving 768 against 776 — the same gap, unclosed — so the property bought nothing on the one engine where the divergence is observable while costing every chromium user 8px of text column. Reverted in favour of asserting the premise and recording the WebKit gap.
-
-**Suppress the textarea's scrollbar instead of reserving a gutter on the other layers.** `scrollbar-width: none` on `.input` would equalize the widths without narrowing the text column. Rejected because the composer deliberately shows a thumb once the draft passes the cap — `.card` binds the l2 scrollbar tokens for exactly that — and removing it takes away the only affordance that says a long draft continues below.
-
-**Translate the backdrop with `transform: translateY(-scrollTop)` instead of scrolling it.** A transform is not clamped by content height, so it would paper over any extent divergence — including the trailing-newline one — without matching the layers. Rejected because the divergence is the actual defect: unequal extents also mean the two layers disagree about where the last line sits, and hiding that behind an unclamped transform would leave a mismatch that resurfaces the moment anything measures the backdrop. Fixing the extent keeps one truth about the draft's height.
-
-**Add a second mirror in a layout effect keyed on the committed draft.** This shipped in the first version of the change, on the theory that an edit reflows both layers without necessarily moving the textarea, and that a shrinking draft clamps each layer independently. Both premises are false, and it was removed after mutation-testing each hook alone against the built client: with only the layout effect disabled the browser scenario stays green, while disabling only the `scroll` listener fails it. Typing scrolls the caret into view, which is an ordinary `scroll`; a shrinking draft clamps both layers to the same maximum because their extents are equal, and the textarea's clamp fires `scroll` too. The specific hazard the effect was imagined to cover — React replacing the backdrop's children when the decoration set changes shape, resetting its offset — does not occur: measured in chromium, replacing every child of an `overflow: hidden` box preserves `scrollTop` (300 stays 300), and the only replacement that zeroes it is one that shrinks the content below the offset, which is the clamp case already covered.
-
-**Sync in the `onChange` handler.** Rejected for the same reason plus one of its own: it fires before React commits the new draft to the backdrop, so it would mirror against the previous layout.
-
-## Consequences
-
- A draft past the cap scrolls its glyphs. Measured in the browser scenario: after a wheel gesture over a 40-line draft the last line sits inside the visible box and the first has scrolled out above it; before, the last line stayed a full draft-height below the box while the textarea's own offset had moved.
- The coupling is one-directional and cheap — one assignment of one number, no measurement, no layout read beyond `scrollTop` — so it adds nothing to the typing path's cost.
- Chips, claim-token highlights, and text-ref marks stay aligned with their glyphs while scrolled, because they are positioned inside the backdrop and move with it. Nothing about the decoration walk changes.
- The composer's two-layer design keeps this hazard: any future layer added beside the backdrop needs the same mirroring, and any change to how a layer reserves its last line box breaks the extent equality the mirror depends on. The e2e scenario asserts both — the relation the user cares about (which line is on screen) and the extent equality underneath it — so a future divergence fails on the invariant rather than on a screenshot.
- Extent equality is asserted, not assumed. It is the premise that turns "mirror the offset" from correct into subtly wrong, and it failed for the trailing-newline shape before the sentinel.
- Wrap-width equality is the other premise, and it does NOT hold universally: WebKit lays `.input` out 8px narrower than the glyph layers. That predates this change and is left open, with the measurement recorded above and an assertion on the lane's engine. A draft whose wrapping turns on those 8px would clamp the mirror on WebKit.
- The composer's layout is unchanged. An earlier revision narrowed the text column by 8px on every platform to chase the wrap-width premise; measurement showed it did not buy the guarantee, so the metrics are the same as before this change.
-
-## Testing
-
-The unit spec in [input-bar.spec.tsx](../../../../packages/client/ui-conversation/tests/input-bar.spec.tsx) proves the mirroring path runs: it stubs both offsets, because jsdom reports `scrollHeight === clientHeight` for every element and never scrolls one, and asserts the backdrop follows the textarea to a new offset and back to the top. Reverting the `ref` makes it fail.
-
-The user-visible fact needs a real engine, so [composer-draft-scroll.e2e.ts](../../../../apps/web/tests/composer-draft-scroll.e2e.ts) measures it in chromium against the built client: a 40-line draft in a fresh workspace's blank composer, zero model calls, with a DOM Range over the backdrop's own text reporting where the first and last lines sit relative to the visible box. A vacuity guard asserts the draft actually overflows the capped box first. A separate case drives the trailing-newline shape and asserts the two extents are equal before asserting the glyphs reach the end; each layer's maximum is observed by asking for an impossible offset and reading back the clamp, not computed from `scrollHeight`. A third asserts the gutter premise: equal wrap widths, and a reserved band greater than zero on each layer. The band is what keeps that assertion from being vacuous — the widths would also match with no reservation at all on this engine's overlay scrollbar, and it is the reservation, not the match, that carries the guarantee to a platform whose scrollbar takes real width.
-
-Confirmed both directions against the built client. With the mirroring reverted and the packages rebuilt, the wheel case fails on the layer offsets, the typing case fails with it, and the golden diff reads `last draft line is on screen: false` while `textarea moved: true` — the reported symptom stated as a fixture. The resting-state case passes in both builds, which is the point: it is the state that hid the defect.
-
-Note that the composer ships inside a client-module bundle, so `pnpm run build:web` alone does not pick up a change to `InputBar.tsx` — the package build must run for the browser lane to see it, and a scenario run against a stale `lib/` asserts against an older client than the tree.
@@ -1,77 +0,0 @@
-# Agent Note: composer 的字形层跟随 textarea 的滚动偏移
-
-Status: implemented
-
-[English](2026-07-31-composer-glyph-layer-tracks-the-textarea.md) | 中文
-
-## 问题
-
-草稿一旦超过 14 行的高度上限，就无法再滚动。光标会动，选区会动，但文字始终冻结在第 1 行——无论滚轮、拖拽还是方向键，都无法把长草稿的末尾带到可见范围内，因此约 14 行之后的内容在书写过程中既够不着也读不到。
-
-高度上限本身是正常工作的。composer 的文本由两层叠放绘制（见 [InputBar](../../../../packages/client/ui-conversation/src/client/skeleton/InputBar.tsx)）：`<textarea>` 持有取值、选区与光标，但它自己的字形以 `color: transparent` 渲染；用户看到的每一个字符都由其下的 `[data-input-backdrop]` 层绘制，该层同时承载 claim token 高亮、chip 与提示影子文本。这一拆分正是 chip 与高亮得以存在的前提——textarea 无法为自身文本的某个区间单独设置样式。
-
-两层在几何上是耦合的，在滚动上却不是。backdrop 为 `position: absolute; inset: 0; overflow: hidden`：它只做裁剪，不做滚动，浏览器也不会把它的偏移与 textarea 关联起来。未达上限时这一点不可见，因为两层都停在偏移 0，且镜像层会把盒子撑到草稿的高度。一旦触及上限，textarea 开始滚动而 backdrop 不跟随，于是用户真正在读的那一层从不移动。
-
-因此该缺陷与高度上限同龄，并且藏在静止状态背后：短草稿——也就是所有截图与既有 fixture（测试前置数据）所捕获的那个状态——在有无该耦合时渲染完全一致。
-
-## 决策
-
-`InputBar` 通过一个 `scroll` 监听把 textarea 的 `scrollTop` 镜像到 backdrop 上，该监听与既有的滚轮接力监听注册在同一个 effect 中（textarea 从不卸载——失效状态渲染的是同一个元素的 disabled 形态）。
-
-一个监听即构成完整耦合，因为这个盒子移动的每一种方式最终都会在 textarea 上产生 `scroll` 事件：手势使它滚动；编辑会把光标滚入可见范围；草稿缩短到当前偏移之下时它会被钳位。看似需要单独处理、实则不需要的正是钳位这一种：两层共享同一滚动范围，因此它们会钳位到同一个最大值，而 textarea 的钳位本身就会触发那次完成镜像的 `scroll`。
-
-这个「共享的滚动范围」并非白得，而镜像偏移只有在它成立时才是正确的。有两件事会破坏它，都是在审查中被发现的，且失效方向相同——backdrop 比 textarea 矮，于是赋值被钳制、字形落到光标之下。textarea 会在末尾换行之后为光标保留一个行盒，而 `white-space: pre-wrap` 会折叠文本节点的尾随换行、不生成任何行盒。因此以换行结尾的草稿会让 backdrop 恰好比 textarea 少一行——实测为 628 对 652——于是该赋值被钳制，滚到最底部时字形比光标落后一行。现在 backdrop 也带上了镜像层早已具备的同一枚尾行哨兵：其内容为装饰扫描的结果再加一个 `'\n'`；草稿不以换行结尾时它被同一次折叠吸收，以换行结尾时它补上缺失的那个行盒。对纯文本、尾随换行、软折行、不可断长串以及中间空行五类草稿实测，两侧范围在每种情形下均相等。
-
-第二个前提是折行宽度，它是被断言的，而不是被修复的。只有 `.input` 会滚动，因此也只有 `.input` 会把内容宽度让给一条占布局宽度的滚动条；`.input` 一旦更窄，长草稿就会折出更多行——在独立环境实测，8px 的宽度差值 2 到 5 行，而宽度相等时 textarea 与 div 完全一致。在运行中的应用上、对 Playwright 自带的三个引擎实测，两个相等、一个不等：
-
-| 引擎 | `.input` / `.backdrop` / `.mirror` 折行宽度 | 滚动范围 |
-|---|---|---|
-| chromium | 776 / 776 / 776 | 相等 |
-| firefox | 776 / 776 / 776 | 相等 |
-| WebKit | **768** / 776 / 776 | 所测草稿下相等 |
-
-WebKit 的 textarea 把 8px 让给了自己的滚动条，而两个被裁剪的图层没有。该差距先于本次改动存在，本 PR 未予关闭；在所测草稿下滚动范围仍然相等，因此镜像不受影响，但一份恰好在该宽度上折行敏感的草稿会让 `.input` 更高、从而钳制镜像偏移。场景在测试通道所用引擎上断言了这项相等性，因此一旦回退到那种状态会显式失败，而不是悄然发生。
-
-共享度量块上的 `scrollbar-gutter: stable` 曾被采用又被移除：WebKit 对 `overflow-y: auto` 应用它、对 `overflow: hidden` 不应用，于是 `.input` 仍是 768 对 776——正是它本想关闭的那个差距——同时又让 chromium 无条件损失 8px 文本宽度。要关闭它，需要一套所有引擎都认同的几何，而不是这个属性。
-
-该镜像是单向的：textarea 是权威方，因为它持有光标，而浏览器滚动的目标正是光标。
-
-## 曾考虑的替代方案
-
-**给 backdrop 加 `overflow: auto`，让它自行滚动。** 那样它就有了一个属于自己的滚动偏移需要同步，问题原样保留，还额外多出一条画在输入框上的滚动条。backdrop 是 textarea 的投影，而不是一个可独立导航的界面。
-
-**去掉 backdrop，直接为 textarea 自身文本设置样式。** 这会消除分层，连同整类失步问题一并消除。之所以否决，是因为它根本无法实现：textarea 只渲染一段统一的文本流，因此 claim token 高亮、chip 与提示影子文本——backdrop 存在的全部理由——都无从表达。为修滚动而放弃它们，是拿一个有界的缺陷去换一次功能删除。
-
-**改用 `contenteditable` div 承载草稿，不再用 textarea。** 一个元素、一个滚动偏移、区间可设样式。之所以否决，是它与该缺陷的体量严重不相称：`contenteditable` 会把 IME 组词、撤销／重做、选区语义与粘贴规范化重新压回我们身上，而这些目前都由 textarea 加输入状态机处理，且状态机已持有一份以 textarea 取值语义为前提的撤销日志。
-
-**在既有的滚轮处理函数里滚动 backdrop，而不是新增 `scroll` 监听。** 该处理函数本就在 textarea 上的每次滚轮时运行，看似是自然的落点。之所以否决，是它只覆盖了盒子滚动的其中一种成因：在末尾输入、`End`、方向键、拖选越过边缘、拖动滚动条，都会在没有滚轮事件的情况下移动 textarea。监听 `scroll` 是在监听事情本身，而不是它的某一个成因。
-
-**用 `scrollbar-gutter: stable` 让三层一起预留滚动条 gutter。** 曾经采用，实测后回退。当初的推理是：无论平台滚动条占多少宽度，三层都预留同样多即可保持相等；而且 `overflow: hidden` 也是滚动容器，按规范应当遵守该声明。chromium 确实如此（三层各预留 8px，宽度 768/768/768）。WebKit 不然：它对 `overflow-y: auto` 预留、对 `overflow: hidden` 不预留，结果仍是 768 对 776——差距原样保留——于是该属性在唯一能观测到这一偏差的引擎上一无所获，却让每一位 chromium 用户损失 8px 文本列。改为断言该前提并记录 WebKit 的差距。
-
-**改为抑制 textarea 的滚动条，而不是给另外两层预留 gutter。** 在 `.input` 上写 `scrollbar-width: none` 同样能让宽度相等，且不必收窄文本列。之所以否决：草稿超过上限后 composer 是有意显示滚动条滑块的——`.card` 正是为此绑定了 l2 滚动条 token——去掉它就等于拿走了「下面还有内容」这一唯一提示。
-
-**改用 `transform: translateY(-scrollTop)` 平移 backdrop，而不是滚动它。** transform 不受内容高度钳制，因此它能把任何范围偏差——包括尾随换行这一种——一并掩盖，却并不让两层真正对齐。之所以否决，是因为这个偏差本身就是真正的缺陷：范围不等同时意味着两层对末行位置的判断不一致，把它藏在一个不受钳制的 transform 之后，只会让这一失配在任何人去测量 backdrop 的那一刻重新浮现。修正范围本身，才能让草稿高度只有一个事实来源。
-
-**再加一个以已提交草稿为 key 的 layout effect 作为第二道镜像。** 该改动的第一版确实带着它，理由是：一次编辑会让两层重排却不一定让 textarea 移动，且草稿变短时两层各自独立地被钳位。这两个前提都不成立，因此在针对构建产物客户端逐个变异测试每个 hook 之后将其移除：仅禁用 layout effect 时浏览器场景全绿，而仅禁用 `scroll` 监听则会失败。输入会把光标滚入可见范围，那就是一次普通的 `scroll`；草稿变短时两层因范围相等而钳位到同一个最大值，且 textarea 的钳位同样会触发 `scroll`。该 effect 本想覆盖的那个具体隐患——React 在装饰集合形状变化时替换 backdrop 的全部子节点，从而重置其偏移——并不会发生：在 chromium 中实测，替换一个 `overflow: hidden` 盒子的全部子节点会保留 `scrollTop`（300 仍为 300），唯一会将其归零的替换是把内容缩短到偏移之下，而那正是已被覆盖的钳位情形。
-
-**在 `onChange` 处理函数里同步。** 除上述同样的理由外还有其自身的问题：它在 React 把新草稿提交到 backdrop 之前触发，因而会按上一次的布局做镜像。
-
-## 后果
-
- 超过上限的草稿会滚动其字形。浏览器场景实测：在 40 行草稿上做一次滚轮手势后，最后一行位于可见盒子之内，第一行已滚出上方；此前最后一行仍停在盒子下方整整一个草稿高度处，而 textarea 自身的偏移已经移动了。
- 该耦合是单向且廉价的——一次对一个数字的赋值，没有测量，除 `scrollTop` 外没有额外的布局读取——因此不会给输入路径增加开销。
- chip、claim token 高亮与文本引用标记在滚动时始终与其字形对齐，因为它们定位在 backdrop 内部并随之移动。装饰扫描本身没有任何改动。
- composer 的双层设计保留了这一隐患：日后在 backdrop 旁新增的任何一层都需要同样的镜像；而任何改变某一层如何保留其末行行盒的改动，都会破坏镜像所依赖的范围相等性。e2e 场景对两者都做了断言——用户真正关心的关系（哪一行在屏幕上），以及其下的范围相等性——因此日后一旦出现偏差，失败会落在不变量上，而不是落在某张截图上。
- 范围相等性是被断言的，而非被假定的。它正是那个能把「镜像偏移」从正确变为微妙错误的前提，并且在加入哨兵之前，它在尾随换行这一形态上确实不成立。
- 折行宽度相等是另一个前提，而它并非普遍成立：WebKit 把 `.input` 排得比字形层窄 8px。该问题先于本次改动存在，此处保持开放，上文记录了实测数值，并在测试通道所用引擎上加了断言。一份折行恰好取决于这 8px 的草稿会在 WebKit 上钳制镜像。
- composer 的布局没有变化。此前有一版为追求折行宽度前提而在所有平台把文本列收窄了 8px；实测表明它并不能带来该保证，因此度量与改动前保持一致。
-
-## 验证
-
-[input-bar.spec.tsx](../../../../packages/client/ui-conversation/tests/input-bar.spec.tsx) 中的单元用例证明镜像路径确实执行：它对两侧偏移都做了桩替换——因为 jsdom 对任何元素都报告 `scrollHeight === clientHeight` 且从不滚动任何元素——并断言 backdrop 既跟随 textarea 到新的偏移，也跟随它回到顶部。撤掉那个 `ref` 会让它失败。
-
-用户可见的事实需要真实引擎，因此 [composer-draft-scroll.e2e.ts](../../../../apps/web/tests/composer-draft-scroll.e2e.ts) 在 chromium 中针对构建产物客户端测量它：在全新工作区空白会话的 composer 中放入 40 行草稿，零模型调用，用一个跨越 backdrop 自身文本的 DOM Range 报告首行与末行相对于可见盒子的位置。一个防空转守卫会先断言草稿确实溢出了设有上限的盒子。另有一个独立用例驱动尾随换行这一形态，先断言两侧范围相等，再断言字形确实抵达末尾；每一层的最大值都通过请求一个不可能的偏移再读回其钳位结果来观测，而非由 `scrollHeight` 计算得出。第三个用例断言 gutter 前提：折行宽度相等，且每层预留的带宽大于零。正是这条「带宽」使该断言不至于空转——在本引擎的 overlay 滚动条下，即使完全不预留，两侧宽度也会相等；把保证传递到滚动条真正占宽的平台上的，是那次预留，而不是这次相等。
-
-已双向确认。撤掉镜像并重新构建各包后，滚轮用例在两层偏移上失败，输入用例随之失败，golden 差异读作 `last draft line is on screen: false` 而 `textarea moved: true`——即以 fixture（测试前置数据）形式陈述的原始现象。静止状态用例在两种构建下都通过，这正是要点所在：它就是掩盖了该缺陷的那个状态。
-
-注意 composer 随客户端模块 bundle 一同发布，因此仅运行 `pnpm run build:web` 不会纳入对 `InputBar.tsx` 的改动——必须运行包构建，浏览器测试通道才能看到它；针对陈旧 `lib/` 运行的场景，断言的是比当前工作树更旧的客户端。
@@ -0,0 +1,6 @@
+# Bilingual-pair consistency record (docs/i18n/README.md): the git blob hash of each
+# side as of the last confirmed-consistent state. Both languages carry equal authority;
+# after editing either side, bring the other along and re-record with:
+#   pnpm run verify-translation-pairing --write .agents/notes/implemented/bug-fix/2026-07-31-composer-text-layers-share-one-scrollport.md
+2026-07-31-composer-text-layers-share-one-scrollport.md: ba11384409714d6a64a964d63197705acf39d213
+2026-07-31-composer-text-layers-share-one-scrollport.zh.md: 9a4a2bcc9f947d385d1ff9fc6367b9a3eab17733
@@ -0,0 +1,82 @@
+# Agent Note: The composer's two text layers share one scrollport
+
+Status: implemented
+
+English | [中文](2026-07-31-composer-text-layers-share-one-scrollport.zh.md)
+
+## Problem
+
+The composer paints its text in two stacked layers ([InputBar](../../../../packages/client/ui-conversation/src/client/skeleton/InputBar.tsx)): the `<textarea>` owns the value, the selection, and the caret but renders its own glyphs `color: transparent`, and every visible character is painted by the `[data-input-backdrop]` div beneath it, which also carries the claim-token highlight, the chips, and the ghost hint. That split is what makes chips and highlights possible at all — a textarea cannot style a range of its own text. The draft box is capped at 14 lines, so past the cap something has to scroll.
+
+Two layers with two scroll offsets fail in two stages, and this change is the second one.
+
+The first was static. The backdrop is `position: absolute; inset: 0; overflow: hidden` — clipped, not scrolled — and nothing in the browser links its offset to the textarea's, so past the cap the caret moved and the words stayed frozen at line 1. Below the cap that is invisible, because both layers rest at 0; the defect was exactly as old as the cap and hid behind the resting state that every screenshot captured. It was fixed by mirroring the textarea's `scrollTop` onto the backdrop from a `scroll` listener, and that made the two layers agree **at rest**.
+
+The second is what a user then reported: swiping quickly from the top of a long draft throws the caret up out of its own text, as though it carried momentum, and it settles back a moment later. The mirror is the cause. A wheel gesture scrolls the textarea on the compositor, off the main thread; the `scroll` event that drives the assignment is dispatched afterwards, so for the frames in between the caret sits at the new offset and every glyph sits at the old one. Measured on a standalone harness of the same geometry, moving the offset 200px and reading the caret's distance to its own glyphs before the task ends: chromium 203px, firefox 202px, WebKit 203px of separation, settling to the fixed line-box constant (3/2/3) a frame or two later. Every engine, every gesture, proportional to how fast the user scrolls.
+
+No listener can close that gap, because the gap is the definition of a listener: it runs after the thing it reacts to. Anything that keeps two boxes equal in JavaScript is a frame behind a compositor that moves one of them without asking.
+
+## Decision
+
+One scrolling box, holding both layers.
+
+`[data-input-scroll]` is the composer's only scrollport and carries the 14-line cap. Inside it, the auto-grow stack is as tall as the **whole** draft — the hidden mirror div is in normal flow and no longer capped, so it sizes the stack to the full text — and the backdrop and textarea ride that height absolutely. The textarea is `overflow: hidden` with no scrollable overflow of its own; it can no longer hold an offset at all.
+
+The browser then applies one offset to both layers, in the same frame, on the same compositor. The caret is bound to its glyphs by construction rather than by upkeep: there is no code to run, no event to wait for, and no state that can be one frame stale. The wheel-chaining handler stays, retargeted from the textarea to the scrollport, and remains the only listener on the box.
+
+Two things the previous mechanism needed are gone with it:
+
+**The backdrop's trailing-line sentinel.** It existed to keep the two boxes' scroll extents equal — a textarea reserves a line box for the caret after a final newline while `white-space: pre-wrap` collapses a text node's trailing newline, so a draft ending in a newline made the backdrop one line shorter and clamped the mirrored offset a line above the caret. With one scrollport the backdrop's own extent decides nothing: the mirror div sizes the stack for both layers, both start at the same top, and a layer whose content ends earlier simply paints nothing on the last line. The shape is worth keeping in mind rather than the mechanism: it is the one that measured 628 against 652 when the two boxes had to agree on a height.
+
+**The wrap-width premise.** All three layers now resolve their width inside the scrollport, so a scrollbar that consumes layout space costs them the same width by construction. This closes the divergence the superseded note recorded as open and unfixable by any property: WebKit reserved gutter space for the `overflow-y: auto` textarea and not for the `overflow: hidden` layers beside it, laying the textarea out 768 against 776 — worth 2 to 5 wrapped lines on a long draft, i.e. glyphs under the wrong caret on the one engine where it was observable. Measured on the harness after the change, all three layers report one width on all three engines Playwright ships.
+
+**Edits the composer performs itself now ask for the reveal.** Paste and cut suppress the native edit — the machine owns the draft and the undo log — and restore the caret with `setSelectionRange`, which reveals nothing: measured in chromium and WebKit, pasting a long block leaves the view where it was while the caret sits at the end of what was pasted. That defect predates this change (Firefox happened to reveal it, in the old geometry only) and is fixed here because one scrollport is what finally makes the reveal ours to perform. The two restores share one helper that measures the caret against the hidden mirror — same draft, same metrics, same wrap width, so a Range collapsed at the caret's index reports where the caret is without a caret API — and scrolls the minimum that brings it inside, which is what the browser does for typing.
+
+One shape needs a rule of its own, because the engines disagree about it: a caret straight after a newline sits on a line with nothing on it to measure, which is where a trailing-newline draft ends. chromium returns **no client rects at all** for the collapsed position — an all-zero box, which would send the reveal the wrong way — firefox reports the line above, and WebKit the right one. The helper therefore measures the newline the caret just left, whose box is the line it came from, and steps one line down; all three then land on the same offset (649 of 652, with the caret's line at 315 inside the 336px box). A non-collapsed Range over that newline returns a real rectangle on all three engines even when the draft ends in consecutive newlines, so each trailing blank line composes with the same rule.
+
+Revealing the caret is the one thing that now depends on the browser rather than on us: with no offset of its own, the textarea's scroll-into-view has to walk up to the scrollport. It does, on every engine measured — typing at the draft's end brings the scrollport to the caret (625, 626 and 628 of a 628px maximum in chromium, firefox and WebKit), walking the caret back up with `ArrowUp` scrolls back to it, and typing after scrolling away returns to it.
+
+## Alternatives considered
+
+**Mirror `scrollTop` onto the backdrop from a `scroll` listener.** The superseded decision, and correct at rest: it is what made a long draft scrollable at all. Rejected now because it cannot be correct in motion — it is a main-thread reaction to a compositor-thread fact — and because it needed two premises to stay true that the single scrollport does not need at all (equal extents, equal wrap widths), each of which had already failed once.
+
+**Translate the backdrop with `transform: translateY(-scrollTop)` instead of assigning an offset.** Same lag: still main-thread, still driven by the same event. It additionally papers over extent divergence rather than making the layers agree, so a mismatch resurfaces the moment anything measures the backdrop.
+
+**Drive the backdrop from a scroll-driven animation (`animation-timeline: scroll()`).** This would run the coupling on the compositor and genuinely eliminate the lag while keeping two boxes. Rejected on support: Safari does not implement it and Firefox has only recently, so the composer would keep the reported defect on the engines that lack it, and the fallback path would be the mechanism being replaced.
+
+**Scroll both layers from JavaScript, with the textarea `overflow: hidden` and a wheel handler assigning both offsets in one task.** No divergence during wheel gestures, since nothing scrolls without us. Rejected because it replaces native scrolling — momentum, trackpad rubber-banding, scrollbar dragging, keyboard scrolling — with a hand-written approximation, and the caret-reveal path (the browser setting the textarea's own offset) still lands asynchronously.
+
+**Keep the cap on the mirror and just wrap today's structure in a scroller.** The layers would stay window-sized, not draft-sized: `inset: 0` on an absolutely positioned child resolves against the scrollport's padding box, not its scrollable overflow area, so both layers would scroll away from the content that is supposed to be underneath them. The stack has to be the full draft height for the arrangement to mean anything.
+
+**Give the backdrop `overflow: auto` and let it scroll itself.** It would then have an offset of its own to keep in step, which is the same problem plus a second scrollbar painted over the input. The backdrop is a projection of the textarea, not an independently navigable surface.
+
+**Drop the backdrop and style the textarea's own text.** This removes the layer split and the whole class of desync with it. Rejected because it is not implementable: a textarea renders one uniform text run, so the claim-token highlight, the chips, and the ghost hint — the reasons the backdrop exists — have no way to be expressed. Losing them to fix scrolling trades a bounded defect for a feature deletion.
+
+**Render the draft in a `contenteditable` div.** One element, one offset, styleable ranges. Rejected as far out of proportion: `contenteditable` would put IME composition, undo/redo, selection semantics, and paste normalization back on us, all of which the textarea plus the input machine currently handle, and the machine already owns an undo log that assumes a textarea's value semantics.
+
+**`scrollbar-gutter: stable` on all three layers.** Tried and reverted while the textarea was the scroller: WebKit applied it to `overflow-y: auto` and not to `overflow: hidden`, leaving the same 8px gap unclosed while costing every chromium user 8px of text column. Moot now — the layers share a containing block, so there is nothing to reserve.
+
+**`scrollbar-width: none` on the scrolling layer.** Would equalize widths by hiding the thumb. Rejected: the composer deliberately shows one once the draft passes the cap — `.card` binds the l2 scrollbar tokens for exactly that — and it is the only affordance saying a long draft continues below.
+
+## Consequences
+
+- The caret cannot leave its glyphs. The browser scrolls one box, so the separation between where the textarea puts a line and where the backdrop paints it is a fixed line-box constant at every offset, mid-gesture included. The scenario measures exactly that number.
+- The scrollbar moved from the textarea to the scrollport — the same visual place, one box out. The `.card` l2 token binding still inherits down to it.
+- Chips, claim-token highlights, and text-ref marks stay aligned with their glyphs while scrolled, because they are positioned inside the backdrop and move with it. The decoration walk is unchanged apart from the dropped sentinel.
+- On Firefox and WebKit, clicking into a composer whose draft overflows the cap now also scrolls the conversation transcript to its bottom: the caret's scroll-into-view walks past the composer's scrollport up to the transcript scrollport, which a textarea shorter than its box never made it do. Chromium does not. Measured, with `overscroll-behavior: contain` and `contain: paint` both tried and neither stopping the walk — there is no CSS that ends scroll-into-view chaining. Accepted: it scrolls toward the bottom, where the composer already sits, and the alternative is a caret visibly detached from its text on every engine.
+- Paging and drag-selection are unchanged, both measured old against new. `PageDown`/`PageUp` never moved a textarea's caret in the first place — chromium scrolls a page and leaves `selectionStart` where it was, in both geometries; only the box that scrolls differs. Drag-selecting past the bottom edge still auto-scrolls, and to the same place (chromium 628/628, firefox 625/620, WebKit 170/170 — WebKit's slower autoscroll is equally slow before and after).
+- The composer's own `focus()` calls pass `preventScroll` — the unlock/session-switch effect and the focus-keeping mousedown on the toolbar buttons — so a focus nobody gestured for cannot move the transcript through the taller textarea's reveal chain. Suppressing that walk hands the caret back to us on the one path where it matters: the composer DOM is reused across sessions, so switching to a longer draft keeps the previous offset while the value swap puts the caret at the new draft's end. Measured on all three engines, that leaves the caret 940px below a box sitting at 0; the effect therefore reveals it in its own scrollport, landing at 625 of 628 — what the old geometry reached at 628 through the browser. The mousedown path needs no reveal: the caret has not moved, and the next keystroke gets the browser's native one. A separate reveal-only effect handles a non-empty draft that arrives after render: `ConversationSession` seeds a persisted draft in its own mount effect, which runs AFTER this component's, so the first reveal would otherwise measure an empty mirror and never run again for the draft that then appeared. The second effect never focuses, so ordinary empty/non-empty transitions such as send-clear or failed-send restore cannot take focus from another control.
+- Undo and redo can change the draft without a caret restore or reveal: the machine replays a previous draft and the DOM selection stays where the browser clamps it. That predates this change and is unchanged by it — named here because the two restores that DO reveal make the omission look deliberate, and the helper is sitting right there if it is ever reported.
+- Any layer added beside the backdrop belongs INSIDE the scrollport and must be as tall as the draft, or it reintroduces exactly this defect. This is the composer's standing hazard: the two-layer split is load-bearing for chips and highlights, so the coupling has to be structural, not maintained.
+
+## Testing
+
+The unit spec in [input-bar.spec.tsx](../../../../packages/client/ui-conversation/tests/input-bar.spec.tsx) asserts what jsdom can see: that one scrolling box contains both the textarea and the backdrop, that the backdrop's text is now the draft and nothing else, and that a late persisted draft reveals its caret without taking focus from another control. jsdom reports `scrollHeight === clientHeight` for every element and never scrolls one, so the geometry belongs to the browser scenario; the wheel-chaining cases stub the scrollport's metrics rather than the textarea's.
+
+[composer-draft-scroll.e2e.ts](../../../../apps/web/tests/composer-draft-scroll.e2e.ts) measures the rest in chromium against the built client: a 40-line draft in a fresh workspace's blank composer, zero model calls. Every metric is read in the caret's own coordinate frame — where the textarea places line n, offset included — against a DOM Range over the backdrop's text for the same line, because that difference is what a user sees. The decisive case changes the offset and re-reads that difference **before the task ends**, which is before any `scroll` listener could have run: 0 with one scrollport, and the full delta with a mirror. A vacuity guard asserts the draft overflows the capped box first, and separate cases cover the cap, one wrap width across all three layers, a wheel gesture, a trailing-newline draft, and the caret-reveal path that the textarea's own scrolling used to handle — typing after scrolling away must bring the scrollport back to the caret.
+
+A separate case covers the paste path end to end: a short draft, the caret at its end, and one `paste` event carrying real clipboard data — the same event a Cmd-V delivers, through the same handler — then the offset and the last pasted line. It waits on the offset rather than on the draft overflowing, because the restore lands one frame after the machine commits the draft; a build without the reveal fails that wait.
+
+The two-geometry comparison behind the decision was measured on a standalone harness before implementing, since the old and new arrangements cannot both exist in the app at once: the same-task separation is 203/202/203px old against 3/2/3px new (chromium/firefox/WebKit), the wrap widths 768-against-776 old on WebKit against 1264/1264/1264 new on all three, and the textarea's own scrollable overflow 0 in the new geometry, which is what makes a second offset impossible rather than merely equal.
+
+Note that the composer ships inside a client-module bundle, so `pnpm run build:web` alone does not pick up a change to `InputBar.tsx` — the package build must run for the browser lane to see it, and a scenario run against a stale `lib/` asserts against an older client than the tree.
@@ -0,0 +1,82 @@
+# Agent Note: composer 的两层文本共用同一个滚动容器
+
+Status: implemented
+
+[English](2026-07-31-composer-text-layers-share-one-scrollport.md) | 中文
+
+## 问题
+
+composer 的文本由两层叠放绘制（见 [InputBar](../../../../packages/client/ui-conversation/src/client/skeleton/InputBar.tsx)）：`<textarea>` 持有取值、选区与光标，但它自己的字形以 `color: transparent` 渲染；用户看到的每一个字符都由其下的 `[data-input-backdrop]` 层绘制，该层同时承载 claim token 高亮、chip 与提示影子文本。这一拆分正是 chip 与高亮得以存在的前提——textarea 无法为自身文本的某个区间单独设置样式。草稿框的高度上限为 14 行，因此超过上限之后总得有东西滚动。
+
+两层各自持有一个滚动偏移，会分两个阶段失效，本次改动是其中的第二个。
+
+第一个阶段是静态的。backdrop 为 `position: absolute; inset: 0; overflow: hidden`——只裁剪、不滚动——浏览器也不会把它的偏移与 textarea 关联起来，于是超过上限之后光标在动而文字冻结在第 1 行。未达上限时这一点不可见，因为两层都停在 0；该缺陷与高度上限同龄，藏在所有截图所捕获的那个静止状态背后。当时的修复是用一个 `scroll` 监听把 textarea 的 `scrollTop` 镜像到 backdrop 上，这让两层在**静止时**保持一致。
+
+第二个阶段正是随后被用户报告的现象：从长草稿的顶部快速滑到底部时，光标像带着惯性一样从自己的文字里往上飞出去，过一会儿又落回原位。原因就是那个镜像。滚轮手势在合成线程上滚动 textarea，不经过主线程；而驱动赋值的 `scroll` 事件在其后才派发，因此中间这些帧里光标位于新偏移、每一个字形却仍位于旧偏移。在同一套几何的独立环境上实测：把偏移改变 200px 并在本任务结束之前读取光标与其字形之间的距离，chromium 分离 203px、firefox 202px、WebKit 203px，一两帧之后才收敛到固定的行盒常量（3/2/3）。每个引擎、每次手势都如此，且滑动越快分离越远。
+
+任何监听都无法消除这个间隙，因为这个间隙正是监听的定义：它在自己所响应的那件事之后才运行。凡是用 JavaScript 维持两个盒子相等的做法，都会落后于那个不经询问就搬动其中一个盒子的合成器一帧。
+
+## 决策
+
+只保留一个滚动盒，让它同时装下两层。
+
+`[data-input-scroll]` 是 composer 唯一的滚动容器，14 行的高度上限落在它身上。容器内部的自增高栈与**整份**草稿等高——隐藏的镜像层处于常规流中且不再设上限，因此由它把栈撑到完整文本高度——backdrop 与 textarea 以绝对定位骑在这个高度上。textarea 为 `overflow: hidden`，自身没有可滚动溢出，也就再无法持有任何偏移。
+
+于是浏览器在同一帧、同一个合成器上，把同一个偏移施加给两层。光标与字形的绑定来自结构本身，而不是来自持续维护：没有代码要跑，没有事件要等，也没有任何状态可能落后一帧。滚轮接力处理器保留，只是从 textarea 改挂到滚动容器上，并且仍是这个盒子上唯一的监听。
+
+上一版机制所需要的两样东西随它一起消失：
+
+**backdrop 的尾行哨兵。** 它的存在只是为了让两个盒子的滚动范围相等——textarea 会在末尾换行之后为光标保留一个行盒，而 `white-space: pre-wrap` 会折叠文本节点的尾随换行，因此以换行结尾的草稿会让 backdrop 少一行，把镜像偏移钳制在光标上方一行。改为单一滚动容器后，backdrop 自身的范围不再决定任何事：镜像层为两层统一定高，两层顶端对齐，内容更早结束的那一层只是在最后一行什么都不画。值得记住的是这类草稿形状而不是那套机制：正是它在「两个盒子必须就高度达成一致」的时代量出了 628 对 652。
+
+**折行宽度这一前提。** 现在三层都在滚动容器内部解析自身宽度，因此一条占布局宽度的滚动条对它们的代价由结构保证相等。这也就关闭了被取代的那篇笔记记录为「悬置且没有任何属性能修」的分歧：WebKit 会为 `overflow-y: auto` 的 textarea 预留槽位，却不为它旁边 `overflow: hidden` 的层预留，把 textarea 排成 768 对 776——在长草稿上值 2 到 5 个折行，也就是在唯一能观察到它的那个引擎上把字形放到了错误的光标之下。改动后在独立环境实测，Playwright 自带的三个引擎上三层宽度均一致。
+
+**由 composer 自己完成的编辑，现在会主动请求回视。** 粘贴与剪切都会抑制原生编辑——草稿与撤销日志归状态机所有——再用 `setSelectionRange` 恢复光标，而这不会带来任何回视：在 chromium 与 WebKit 上实测，粘贴一大段之后视图停在原处，光标却落在所粘内容的末尾。该缺陷早于本次改动（Firefox 只在旧几何下恰好会回视），在此修复，是因为单一滚动容器才终于让「回视」成为我们能自己做的事。两处恢复共用一个 helper：它以隐藏的镜像层为标尺——同一份草稿、同一套度量、同一折行宽度，因此在光标索引处折叠一个 Range 就能报出光标位置，无需任何 caret API——并且只滚动到刚好把该行带进可见范围为止，与浏览器为输入所做的一致。
+
+有一种形状需要单独的规则，因为引擎之间在这里并不一致：紧跟在换行之后的光标，落在一条没有任何内容可供度量的行上——以换行结尾的草稿正是终止于此。chromium 对这个折叠位置**根本不返回任何 client rect**（一个全零盒子，会把回视带向反方向），firefox 报的是上一行，WebKit 报的才是对的那一行。因此该 helper 改为度量光标刚离开的那个换行——它的盒子就是光标来的那一行——再往下走一行；三者随即落在同一个偏移上（649/652，光标所在行位于 336px 盒内的 315）。即使草稿以连续换行结尾，覆盖该换行的非折叠 Range 在三个引擎上都会返回真实矩形，因此每个尾随空行都能沿用同一条规则定位。
+
+现在唯一依赖浏览器而非依赖我们自己的，是把光标滚入可见范围：textarea 没有了自己的偏移，它的 scroll-into-view 必须向上走到滚动容器。实测的每个引擎都会这么做——在草稿末尾输入会把滚动容器带到光标处（chromium、firefox、WebKit 分别为 625、626、628，最大值 628），用 `ArrowUp` 把光标一路走回去会滚回去，滚离光标后再输入也会回到光标。
+
+## 备选方案
+
+**用 `scroll` 监听把 `scrollTop` 镜像到 backdrop 上。** 被取代的那个决策，静止时是正确的：正是它让长草稿第一次可以滚动。现在被否决，是因为它在运动中不可能正确——它是主线程对合成线程事实的响应——并且它需要两个前提持续成立（范围相等、折行宽度相等），而单一滚动容器根本不需要这两个前提，其中每一个都已经失效过一次。
+
+**用 `transform: translateY(-scrollTop)` 平移 backdrop，而不是赋一个偏移。** 延迟相同：仍在主线程，仍由同一个事件驱动。它还会把范围分歧盖住而不是让两层真正一致，于是只要有什么东西去测量 backdrop，错配就会重新浮现。
+
+**用滚动驱动动画（`animation-timeline: scroll()`）驱动 backdrop。** 这会把耦合放到合成器上运行，在保留两个盒子的前提下确实能消除延迟。因支持度被否决：Safari 尚未实现、Firefox 也是近期才有，因此在缺少它的引擎上 composer 会保留被报告的缺陷，而其回退路径正是这次要替换掉的机制。
+
+**两层都由 JavaScript 驱动滚动：textarea 设 `overflow: hidden`，滚轮处理器在同一个任务里给两个偏移赋值。** 滚轮手势期间不会分离，因为没有我们就没有东西会滚动。被否决，是因为它用手写近似替换了原生滚动——惯性、触控板回弹、拖拽滚动条、键盘滚动——而且光标回视路径（浏览器设置 textarea 自己的偏移）仍然是异步落地的。
+
+**把上限留在镜像层上，只在今天的结构外面套一个滚动容器。** 那样两层仍是「窗口大小」而非「草稿大小」：绝对定位子元素的 `inset: 0` 是相对滚动容器的 padding box 解析的，而不是相对其可滚动溢出区域，于是两层会从本该垫在它们下面的内容上滚开。栈必须与整份草稿等高，这套排布才有意义。
+
+**给 backdrop 加 `overflow: auto`，让它自己滚动。** 那样它就有了一个自己的偏移需要保持同步，即同一个问题再加一条画在输入框上的滚动条。backdrop 是 textarea 的投影，不是一个可独立导航的界面。
+
+**去掉 backdrop，直接给 textarea 自己的文本上样式。** 这会消除分层，也一并消除这一整类失步。被否决是因为它不可实现：textarea 只渲染一段统一的文本，claim token 高亮、chip 与提示影子文本——backdrop 存在的理由——无从表达。为修滚动而失去它们，是拿一个有界缺陷换一次功能删除。
+
+**用 `contenteditable` div 渲染草稿。** 一个元素、一个偏移、区间可上样式。因代价与缺陷严重不成比例被否决：`contenteditable` 会把输入法组合、撤销/重做、选区语义与粘贴规范化重新压回我们身上，而这些目前都由 textarea 加输入状态机处理，且状态机已经持有一份假定 textarea 取值语义的撤销日志。
+
+**给三层都加 `scrollbar-gutter: stable`。** 在 textarea 还是滚动者时试过并已回退：WebKit 只对 `overflow-y: auto` 生效、不对 `overflow: hidden` 生效，那 8px 的差距原样留着，却让每个 chromium 用户无条件损失 8px 文本列宽。现在已无意义——三层共享同一包含块，没有什么需要预留。
+
+**给滚动的那一层加 `scrollbar-width: none`。** 靠隐藏滑块来抹平宽度。被否决：草稿超过上限时 composer 是有意显示滑块的——`.card` 绑定 l2 滚动条 token 正是为此——而它是唯一提示「长草稿在下面还有」的可供性。
+
+## 影响
+
+- 光标不可能离开自己的字形。浏览器滚动的是同一个盒子，因此「textarea 把某一行放在哪」与「backdrop 把这一行画在哪」之间的距离在任何偏移下都是一个固定的行盒常量，手势进行中也不例外。场景测试度量的正是这个数。
+- 滚动条从 textarea 移到了滚动容器上——视觉位置相同，只是外移了一层。`.card` 的 l2 token 绑定仍会继承下去。
+- chip、claim token 高亮与文本引用标记在滚动时仍与其字形对齐，因为它们定位在 backdrop 内部、随之移动。除去掉哨兵之外，装饰扫描没有变化。
+- 在 Firefox 与 WebKit 上，点进一个草稿超过上限的 composer 现在还会把会话记录滚动到底部：光标的 scroll-into-view 会越过 composer 的滚动容器一路走到会话记录的滚动容器，而一个比自身盒子矮的 textarea 从不会引发这一步。chromium 不会。已实测，并试过 `overscroll-behavior: contain` 与 `contain: paint`，两者都拦不住这次上行——没有任何 CSS 能终止 scroll-into-view 的接力。接受：它滚向底部，而 composer 本来就在底部，而其替代方案是在每个引擎上都出现光标与文字明显分离。
+- 翻页与拖拽选区的行为未变，二者均做了新旧对照实测。`PageDown`/`PageUp` 本来就不会移动 textarea 的插入点——chromium 是滚动一页并保持 `selectionStart` 不变，新旧几何皆然，区别只在于滚的是哪个盒子。拖拽选区越过下边缘仍会自动滚动，且落点一致（chromium 628/628、firefox 625/620、WebKit 170/170——WebKit 自动滚动较慢，但改动前后一样慢）。
+- composer 自己发起的 `focus()` 全部加了 `preventScroll`——解锁/切会话的 effect，以及工具栏按钮上那个保持焦点的 mousedown——因此一次没有任何手势要求的聚焦，不会再通过更高的 textarea 的回视链把 transcript 挪走。抑制这条链之后，光标就回到了我们手上，而这在一条路径上确实要紧：composer 的 DOM 跨会话复用，因此切到更长的草稿时旧偏移会留着，而换值会把光标放到新草稿的末尾。三引擎实测，这会让光标落在停在 0 的盒子下方 940px 处；于是该 effect 会在自己的滚动容器里把它带回来，落点 625/628——正是旧几何靠浏览器达到的 628。mousedown 那条不需要回视：光标没动过，而下一次敲键会拿到浏览器原生的回视。另一个只负责回视的 effect 会处理渲染后才到达的非空草稿：`ConversationSession` 在自己的 mount effect 中注入持久化草稿，而该 effect 在本组件的 effect 之后运行，否则第一次回视会量到空镜像，且不会再为随后出现的草稿重跑。第二个 effect 从不聚焦，因此发送后清空或发送失败后恢复这类普通的空/非空转换不会从其他控件夺走焦点。
+- 撤销/重做可以在不恢复光标、也不回视的情况下改动草稿：状态机重放上一版草稿，DOM 选区停在浏览器钳位后的位置。这早于本次改动且未被改动——在此点名，是因为另外两处恢复都会回视，会让这处遗漏看起来像有意为之；真被报告时 helper 就在旁边。
+- 任何新增在 backdrop 旁边的层都属于滚动容器**内部**，并且必须与草稿等高，否则就会重新引入这一缺陷。这是 composer 长期存在的风险点：两层拆分对 chip 与高亮是承重的，因此耦合必须来自结构，而不是靠维护。
+
+## 测试
+
+[input-bar.spec.tsx](../../../../packages/client/ui-conversation/tests/input-bar.spec.tsx) 中的单元用例断言 jsdom 能看见的部分：同一个滚动盒同时包含 textarea 与 backdrop，backdrop 的文本现在就是草稿本身、不多不少，且渲染后才到达的持久化草稿会回视其光标，同时不从其他控件夺走焦点。jsdom 对任何元素都报告 `scrollHeight === clientHeight` 且从不滚动，因此几何属于浏览器场景；滚轮接力用例改为桩接滚动容器的度量，而非 textarea 的。
+
+[composer-draft-scroll.e2e.ts](../../../../apps/web/tests/composer-draft-scroll.e2e.ts) 在 chromium 中针对构建产物度量其余部分：全新工作区空会话的 composer 中一份 40 行草稿，零模型调用。每个度量都在光标自己的坐标系里读取——即 textarea 把第 n 行放在哪，含其自身偏移——再与 backdrop 同一行文本上的 DOM Range 相比，因为这个差值正是用户看到的东西。决定性的用例改变偏移，并**在本任务结束之前**重新读取该差值，也就是在任何 `scroll` 监听可能运行之前：单一滚动容器下为 0，镜像方案下则是整个增量。空洞性保护先断言草稿确实超过了带上限的盒子；其余用例分别覆盖高度上限、三层同一折行宽度、滚轮手势、以换行结尾的草稿，以及过去由 textarea 自身滚动承担的光标回视路径——滚离光标后输入，必须把滚动容器带回光标处。
+
+另有一个用例端到端覆盖粘贴路径：短草稿、光标停在末尾，然后派发一个携带真实剪贴板数据的 `paste` 事件——与 Cmd-V 送达的是同一个事件，走同一个处理器——再检查偏移与所粘内容的最后一行。它等待的是偏移而不是「草稿是否溢出」，因为恢复发生在状态机提交草稿之后的下一帧；没有这次回视的构建会卡在这个等待上失败。
+
+支撑该决策的两套几何对比是在实现之前于独立环境度量的，因为新旧排布无法在应用里同时存在：同任务分离度旧为 203/202/203px、新为 3/2/3px（chromium/firefox/WebKit），折行宽度旧在 WebKit 上为 768 对 776、新在三个引擎上均为 1264/1264/1264，而新几何下 textarea 自身的可滚动溢出为 0——正是这一点让第二个偏移不可能存在，而不只是碰巧相等。
+
+注意 composer 打包在 client-module bundle 内，因此只跑 `pnpm run build:web` 并不会带上 `InputBar.tsx` 的改动——必须先跑包构建，浏览器泳道才看得到；对着过期 `lib/` 跑场景，断言的是比当前代码树更旧的客户端。
@@ -1,6 +1,6 @@
 # Bilingual-pair consistency record (docs/i18n/README.md): the git blob hash of each
 # side as of the last confirmed-consistent state. Both languages carry equal authority;
 # after editing either side, bring the other along and re-record with:
-#   pnpm run verify-translation-pairing --write .agents/notes/implemented/bug-fix/2026-07-31-composer-glyph-layer-tracks-the-textarea.md
-2026-07-31-composer-glyph-layer-tracks-the-textarea.md: d60a100be98683b5f7a7c88edf7585d275134730
-2026-07-31-composer-glyph-layer-tracks-the-textarea.zh.md: eab3f9e3fe3bddb426836113d08f1839329119d5
+#   pnpm run verify-translation-pairing --write .agents/notes/implemented/bug-fix/2026-08-03-cli-signal-shutdown-escalation.md
+2026-08-03-cli-signal-shutdown-escalation.md: 2746b5784baad0f3b14258280cd56a621db07c15
+2026-08-03-cli-signal-shutdown-escalation.zh.md: 0bda83327d4cc8fe2edb61f8145a89138610901e
@@ -0,0 +1,52 @@
+# Agent Note: Bounded, escalating signal shutdown for Web and headless
+
+Status: implemented
+
+English | [中文](2026-08-03-cli-signal-shutdown-escalation.zh.md)
+
+## Problem
+
+The default telemetry mount added SIGINT/SIGTERM handlers to `dsh web` and `dsh -p` so process exit could drain the Cordis tree instead of dropping queued telemetry. Each handler used a one-way boolean latch and exited only after `ctx.fiber.dispose()` settled. Headless normal completion also awaited that disposal without a bound.
+
+A user then reproduced `dsh -p` hanging immediately after the observation URL and ignoring repeated `Ctrl+C`; `DSH_TELEMETRY_DISABLED=1` removed the hang, while a standalone Node handler in the same Linux sandbox received SIGINT. This isolated the pending disposer to telemetry rather than terminal signal forwarding. OTel's `BatchLogRecordProcessor.shutdown()` awaits `exporter.forceFlush()` before the `exportTimeoutMillis`-bounded completion promise, and the OTLP exporter's `forceFlush()` waits directly on its in-flight HTTP Promise. A proxy/sandbox connection that never obtains a socket can therefore leave provider shutdown pending despite both configured SDK timeouts.
+
+The latch then turned that telemetry defect into an unkillable CLI: normal completion was already awaiting the single-shot root disposal; the first SIGINT joined the same pending disposal and set the signal latch; later SIGINTs returned at the latch, so the process had no remaining escape. A signal received before normal completion had the same unbounded wait. Web used the same latch shape.
+
+Telemetry's own timeouts cannot prove that the whole plugin tree settles. Any current or future disposer can wedge, and the process boundary must preserve both a graceful first attempt and a user-controlled way out.
+
+## Decision
+
+The fix has two ownership layers. The OTel backend adds `shutdownTimeoutMillis` (default and shipped value: three seconds) around the SDK provider's complete shutdown Promise. Crossing it rejects into the telemetry coordinator's existing contained-failure path, allowing the Cordis tree to finish disposal; pending records may be lost because OTel exposes no cancellation for the transport Promise.
+
+Web and headless share `createProcessShutdown`, one process-level controller around root disposal:
+
+- Normal shutdown calls coalesce onto one disposal and retain the first requested exit code; they never escalate one another.
+- The first signal starts the same graceful disposal and a referenced five-second exit backstop. Disposal success or failure exits once; neither can cancel the process exit.
+- A signal received while shutdown is pending forces immediate exit with that signal path's code. This includes the first `Ctrl+C` after headless normal completion has already entered disposal, and a second signal after a signal initiated the drain.
+- The five-second bound is a process-safety invariant, not a deployment tunable. It is long enough for the telemetry deployment's ordinary drain ceiling while still bounding any wedged disposer at the launcher boundary.
+
+Headless preserves exit 0 for a completed turn, exit 1 for another turn-end reason or API business error, 130 for SIGINT, and 143 for SIGTERM. Web preserves its existing SIGTERM exit 0 and SIGINT exit 130 behavior.
+
+This supersedes the [telemetry deployment Note's](../feature/2026-07-31-web-telemetry-default-mount.md) assumption that SDK exporter/processor timeouts bound complete provider shutdown, and its earlier decision to defer a process-level backstop. The backend owns its export loss/latency policy and closes the known SDK `forceFlush()` gap; the launcher owns the outer guarantee that no plugin can trap the process indefinitely.
+
+## Alternatives considered
+
+**Bound only the telemetry backend's `shutdown()`.** Insufficient because it protects the known OTel wait but cannot protect the launcher from another plugin's disposer.
+
+**Restore Node's default immediate signal exit.** Rejected because a healthy first signal should still flush telemetry and release other resources. Immediate exit is the explicit escalation path, not the default.
+
+**Add only the five-second timeout.** Rejected because a user pressing `Ctrl+C` again is asking to stop waiting now. Swallowing that intent for the rest of the grace period recreates the reported behavior at a shorter duration.
+
+## Consequences
+
+A healthy exit still disposes the complete Cordis tree. The known telemetry wait releases after at most three seconds; any other wedged exit lasts at most five seconds without further input, and a repeated signal ends it immediately. Forced or deadline-bounded exit can interrupt telemetry export or remaining cleanup, which is intentional only after the graceful contract has failed or the user has explicitly escalated.
+
+The controller is launcher infrastructure rather than a Cordis plugin: it makes no claim that disposal completed, and it does not weaken the lifecycle rule that ordinary disposers must reach quiescence.
+
+## Testing
+
+`apps/cli/tests/process-shutdown.spec.ts` pins resolved and rejected disposal, the five-second backstop, normal-call coalescing, a signal interrupting normal disposal, and second-signal escalation.
+
+`apps/cli/tests/headless-shutdown.e2e.ts` boots the real shipped Web/headless Loader tree in a PTY with a test-only plugin whose disposer announces entry and never settles. The test sends SIGINT after the observation URL, waits for proof that disposal started, sends SIGINT again, and requires exit 130. The source/artifact launch resolver keeps the same regression on both execution planes. This PTY case covers the user-visible process state; no model-output snapshot changes.
+
+`packages/telemetry/session-telemetry-otel/tests/otel.spec.ts` holds a real OTLP request open after timer export begins and pins that Cordis disposal returns at `shutdownTimeoutMillis`, despite the SDK's `forceFlush()` remaining pending. The collector is then released so the still-observed provider Promise settles cleanly.
@@ -0,0 +1,52 @@
+# Agent Note：Web 与 headless 的有界信号关闭和重复信号强制退出
+
+状态：已实现
+
+[English](2026-08-03-cli-signal-shutdown-escalation.md) | 中文
+
+## 问题
+
+默认挂载遥测后，`dsh web` 与 `dsh -p` 新增了 SIGINT/SIGTERM 处理器，使进程退出时可以排空 Cordis 插件树，而不是丢弃排队中的遥测数据。每个处理器都使用单向布尔闩锁（latch），并且只有在 `ctx.fiber.dispose()` 结算后才退出。headless 正常完成时同样会无界等待整棵树执行 dispose（资源释放）。
+
+随后有用户复现，`dsh -p` 在打印观察 URL 后立即卡死，重复按 `Ctrl+C` 也没有反应；设置 `DSH_TELEMETRY_DISABLED=1` 后不再卡死，而同一 Linux 沙箱中的独立 Node 信号处理器能够收到 SIGINT。这将待结算的 disposer 定位到遥测，而非终端信号转发。OTel 的 `BatchLogRecordProcessor.shutdown()` 会先等待 `exporter.forceFlush()`，再进入受 `exportTimeoutMillis` 限制的完成 promise；OTLP 导出器的 `forceFlush()` 则直接等待正在进行的 HTTP Promise。因此，代理／沙箱连接始终无法取得 socket 时，即使已经配置两项 SDK 超时，也会让提供方关闭一直待结算。
+
+闩锁随后把这个遥测缺陷变成无法终止的 CLI（命令行界面）：正常完成流程已经在等待单次根级 dispose；第一次 SIGINT 会加入同一个待结算的 dispose，并设置信号闩锁；后续 SIGINT 在闩锁处直接返回，因此进程再无退出途径。正常完成之前收到信号时，同样会陷入无界等待。Web 使用的闩锁结构与此相同。
+
+遥测自身的超时无法证明整棵插件树都能结算。任何当前或未来的 disposer 都可能卡死；进程边界既要保留第一次优雅关闭的机会，也必须给用户留下强制退出的途径。
+
+## 决策
+
+修复分为两层归属。OTel 后端围绕 SDK 提供方的完整关闭 Promise 增加 `shutdownTimeoutMillis`（默认值和交付值均为 3 秒）。超过该截止时间时会 reject，并进入遥测协调器现有的失败隔离路径，使 Cordis 插件树能够完成 dispose；由于 OTel 未公开取消传输 Promise 的能力，待处理记录可能丢失。
+
+Web 与 headless 共用 `createProcessShutdown`，它是围绕根级 dispose 建立的进程级控制器：
+
+- 多次正常关闭调用会汇合到同一次 dispose，并保留首次请求的退出码；这些调用不会相互触发强制退出。
+- 第一个信号会启动同一次优雅 dispose，并设置一个带引用的 5 秒退出兜底。dispose 无论成功或失败都会触发且仅触发一次退出；任何一种结果都无法取消进程退出。
+- 关闭待结算期间收到信号时，会立即按该信号路径的退出码强制退出。这既包括 headless 正常完成已经进入 dispose 后收到的第一次 `Ctrl+C`，也包括由信号启动排空后收到的第二个信号。
+- 5 秒上限是进程安全不变式，而不是部署调节项。它足以覆盖遥测部署的常规排空时限，同时仍在启动器边界为任何卡死的 disposer 设置等待上限。
+
+headless 对完成的轮次仍以 0 退出，对其他轮次结束原因或 API 业务错误仍以 1 退出，对 SIGINT 以 130 退出，对 SIGTERM 以 143 退出。Web 保留现有行为：SIGTERM 以 0 退出，SIGINT 以 130 退出。
+
+这项决策取代了[遥测部署 Agent Note](../feature/2026-07-31-web-telemetry-default-mount.md) 中 SDK 导出器／处理器超时能够限制提供方完整关闭流程的假设，也取代了其中暂缓进程级退出兜底的决定。后端负责导出数据丢失与延迟策略，并封住已知的 SDK `forceFlush()` 缺口；启动器负责最外层保证，确保任何插件都无法无限期困住进程。
+
+## 考虑过的替代方案
+
+**只限制遥测后端的 `shutdown()`。** 仍不充分：它能保护已知的 OTel 等待，但无法保护启动器免受其他插件 disposer 的影响。
+
+**恢复 Node 默认的信号即时退出。** 不予采纳：收到第一个信号时，健康流程仍应刷新遥测数据并释放其他资源。即时退出是显式的强制退出路径，而非默认行为。
+
+**只增加 5 秒超时。** 不予采纳：用户再次按下 `Ctrl+C`，就是要求立即停止等待。若在剩余宽限期内继续吞掉这一意图，只是缩短了报告中故障的持续时间，并未解决问题。
+
+## 后果
+
+健康的退出流程仍会对整棵 Cordis 插件树执行 dispose。已知的遥测等待最多会在 3 秒后解除；其他退出流程卡死时，如无进一步输入，最多等待 5 秒，再次收到信号则立即结束进程。强制退出或受截止时间限制的退出可能中断遥测导出或尚未完成的清理工作；只有优雅关闭契约已经失败，或用户明确要求强制退出时，才会有意接受这一结果。
+
+该控制器属于启动器基础设施，而不是 Cordis 插件：它不会声称 dispose 已经完成，也不会削弱普通 disposer 必须达到完全停稳状态的生命周期规则。
+
+## 测试
+
+`apps/cli/tests/process-shutdown.spec.ts` 固定了 dispose 成功与失败、5 秒退出兜底、正常调用汇合、信号中断正常 dispose，以及第二次信号强制退出的行为。
+
+`apps/cli/tests/headless-shutdown.e2e.ts` 在 PTY 中启动真实交付的 Web/headless Loader 插件树，并挂载一个仅用于测试的插件；该插件的 disposer 会声明已经进入清理流程，但永不结算。测试在观察地址出现后发送 SIGINT，等待 dispose 已启动的证据，再次发送 SIGINT，并要求进程以 130 退出。源码／产物启动解析器使两个执行平面都覆盖同一项回归。该 PTY 用例覆盖用户可见的进程状态；模型输出快照没有变化。
+
+`packages/telemetry/session-telemetry-otel/tests/otel.spec.ts` 在定时器导出开始后保持一条真实 OTLP 请求打开，并固定以下行为：即使 SDK 的 `forceFlush()` 仍待结算，Cordis dispose 也会在 `shutdownTimeoutMillis` 到期时返回。随后测试释放 collector，使仍受观察的提供方 Promise 干净结算。
@@ -2,5 +2,5 @@
 # side as of the last confirmed-consistent state. Both languages carry equal authority;
 # after editing either side, bring the other along and re-record with:
 #   pnpm run verify-translation-pairing --write .agents/notes/implemented/feature/2026-07-28-web-terminal-card.md
-2026-07-28-web-terminal-card.md: 14896b1d88e5cfd2e4c58830c7a1bca1e54ed823
-2026-07-28-web-terminal-card.zh.md: 16c9004f8f80b720b25b76ba5c04f308b0fccbaf
+2026-07-28-web-terminal-card.md: 0e5f3e2157ebfc4e71aead26c15b6ee91958a5d5
+2026-07-28-web-terminal-card.zh.md: 1285d3fbb46ebd32ff163feac632cd487e8a04f1
@@ -6,7 +6,7 @@ English | [中文](2026-07-28-web-terminal-card.zh.md)

 ## Problem

-The bash tool declares `card: 'terminal'` for both its call and its result ([render-intent union](../architecture/2026-07-02-tool-render-intent-union.md)): the call view carries the command, an optional model-authored description, and the working directory; the result view carries the output, exit code, and terminating signal. That view already reaches the browser — host, connection, and runtime deliver it onto `ConversationSnapshot` as `callView`/`resultView` — and the TUI already renders it as a `$`-prompt card with an exit line and a head/tail height cap.
+The bash tool declares `card: 'terminal'` for both its call and its result ([render-intent union](../architecture/2026-07-02-tool-render-intent-union.md)): the call view carries the command, an optional model-authored description, and the working directory; the result view carries the output, exit code, and terminating signal. That view already reaches the browser — host, connection, and runtime deliver it onto `ConversationSnapshot` as `callView`/`resultView` — and the former TUI rendered it as a `$`-prompt card with an exit line and a head/tail height cap.

 The Web client ignored it. `packages/client/ui-conversation/src/client/contract/tool-call-model.ts` derived every row from raw tool args, and `skeleton/DetailsPanel.tsx` flattened every tool's content blocks into one `<pre>` with `white-space: pre-wrap; word-break: break-word`. Two defects followed from soft-wrapping and from having no height bound: multi-column output (`ls`, a table, box drawing) folded into a paragraph and lost the column alignment that is the whole point of that output, and a long single-column listing stretched the details panel to the length of the listing.

@@ -19,7 +19,7 @@ The component's contract:
 - **Prompt lines, one per command line.** Each line of the command gets its own row: label, then that line verbatim. A `command` carrying two shell commands on two lines therefore reads as the two commands it is, instead of collapsing into one ellipsized row. The label is the cwd's last path segment, or `~` when the cwd equals the `home` prop — a browser has no `$HOME`, so the caller supplies the absolute home directory and the collapse simply does not apply without it. A view with no cwd renders a plain `$`. A trailing newline is a terminator, not an empty final command. Only the FIRST row carries the label: the view knows one working directory — where the call started — and a later line may run somewhere else entirely, since a `cd` in the command is enough to move it. Repeating the label down the rows would state a directory per line that nothing here knows, which is the same reason the run-state dot appears once. Later rows keep a bare `$` so they still read as prompts.
 - **One run-state dot for the call, on the first row.** `StateDot` in three of its four states: the chase while running, red for the exit status that also renders the pill, green for a clean settle — the same indicator a tool row's leading icon uses, so a row and its own card cannot disagree about one command. The dot exists because the first question a reader has about a shell command is whether it is still running, and without it that had to be inferred from the absence of output — which a settled command producing no output also looks like. It sits out of flow in a gutter the card reserves as its OWN left padding, so it neither indents its command nor depends on the command's text metrics to line up. The reservation is padding rather than margin because every render site rewrites `margin` wholesale to set its own indent, which silently cancelled a margin-based gutter and let a container clip the dot. Exactly one dot, whatever the line count: the exit status the view carries is the whole call's, and bash reports no per-command status, so a dot per line would assert of a line that succeeded inside a failing call that the line itself failed. The single visually hidden text label carries the same scope, since `StateDot` is `aria-hidden` and one label per row would read to assistive technology as several distinct outcomes.
 - **No soft wrapping.** Output lines are `white-space: pre` inside a horizontally scrolling box. Column alignment survives; a long line scrolls instead of folding.
- **Height cap with an expand control.** Output longer than `DEFAULT_TERMINAL_MAX_LINES` (16) lines shows `ceil(max/2)` head lines plus the remaining tail lines, with a button in between that reports the hidden count and expands. The count is of parsed lines after the trailing output terminator is dropped, so an N-line output ending in a newline is N lines. The split arithmetic is the same as the TUI transcript's collapsed tool card (`packages/ui/tui/src/components/transcript.ts`), so one command's head and tail slices agree between the two front ends.
+- **Height cap with an expand control.** Output longer than `DEFAULT_TERMINAL_MAX_LINES` (16) lines shows `ceil(max/2)` head lines plus the remaining tail lines, with a button in between that reports the hidden count and expands. The count is of parsed lines after the trailing output terminator is dropped, so an N-line output ending in a newline is N lines. The split arithmetic preserves the former TUI transcript's collapsed-card behavior, so the established head/tail selection stays stable.
 - **ANSI color.** `anser` splits the SGR runs; `ui-primitives/src/ansi.ts` resolves each run into an inline style rendered as React spans. A foreground-only run maps the basic 16 colors onto `--dsw-*` theme tokens so authored color stays legible under both themes; a run that paints its own background keeps anser's literal rgb for both so its intended contrast survives, as do 256-palette, truecolor, and the two basic colors this design system has no token for. Sequences that carry no color (OSC strings, non-CSI escapes, inert C0 controls) are stripped before parsing so they never reach the DOM as literal characters. Cursor movements resolve before that strip, into a per-line column buffer rather than by string surgery, because carriage return and backspace only MOVE the cursor — neither erases anything, so what a reader sees is whatever each column last had written to it. `100%` then a carriage return and `OK` shows `OK0%`, since the redraw is shorter than the frame beneath it; a trailing `abc` plus a backspace still shows `abc`, since nothing overwrote the `c`; `abc` plus two backspaces and `XY` shows `aXY`. Each of these was checked against a real terminal, because the earlier truncate-and-delete approximations looked right and were not. SGR state is stamped per column as a terminal stores it per cell, so a partial overwrite keeps each surviving character's own color: red `bad`, three backspaces, then `ok` shows `okd` with the `d` still red. A CSI sequence occupies no column and changes only the state later writes are stamped with, which is also why a carriage return does not reset color, and why SGR state threads from one line to the next rather than closing at each newline. Erase-in-line is part of the same replay, because `\r\x1b[K` is the single idiom every spinner and progress bar writes — modelling the `\r` alone left the previous frame's tail standing, which is text the terminal never showed. Only `m` accumulates into a cell's style; a cursor or erase sequence must not, or the state string grows per redraw and emits boundaries anser has to discard. SGR is held per cell as a NORMALIZED record (foreground, background, attribute set), not as the sequence history: accumulating raw sequences made every state boundary re-emit the whole chain, so output that switches color without a full reset emitted O(n^2) characters — 3200 such cells produced 25 MB and a `RangeError` well under bash's own output cap. The record also lets the attribute closers every chalk-based tool writes (`39`, `49`, `22`, `24`, …) actually close their attribute, and each boundary emits one canonical sequence for the state it opens. A run also has to CLOSE: the replay converges to the state the scan ended in, not the last written cell's, because a reset after the final write changes no cell yet ends the run — without that a line finishing in `\x1b[0m` leaked its color onto every later line. The cursor advances by terminal columns, so a tab reaches the next 8-column stop, a wide character takes two (its spacer blanking rather than closing the gap once the lead cell is overwritten), and a combining mark takes none. Width follows emoji PRESENTATION rather than the U+2600-U+27BF block: `\u2713`, the check every progress line writes, is one column, so treating the block as wide misaligned exactly the output this card exists for. Writing over either half of a wide pair blanks the other, since a terminal cannot leave one cell of a two-cell glyph standing: `a\tb` then a redraw of `XY` shows `XY      b`, since a two-character redraw cannot reach column 8.
 - **Exit status and copy.** A non-zero exit code or a signal renders a status pill, matching the exit-status distinction the bash tool's own renderer draws; a clean exit renders none, and settled empty output renders a dimmed placeholder — judged on the parsed lines the card renders, not on the raw text, since output that is only escapes or control bytes survives a `trim()` yet parses to nothing visible and would otherwise draw blank rows plus a copy control for invisible bytes. The copy control copies the raw output text, not the rendered tree, so the prompt line and the pill stay out of the clipboard.

@@ -6,7 +6,7 @@ Status: implemented

 ## Problem

-bash 工具的调用与结果都声明 `card: 'terminal'`（[渲染意图联合类型](../architecture/2026-07-02-tool-render-intent-union.md)）：调用视图携带命令、一段可选的模型撰写描述以及工作目录，结果视图携带输出、退出码与终止信号。该视图早已抵达浏览器——host、connection 与 runtime 把它投递到 `ConversationSnapshot` 的 `callView`/`resultView` 上——TUI 也早已把它渲染为带 `$` 提示符的卡片，附退出行与首尾高度上限。
+bash 工具的调用与结果都声明 `card: 'terminal'`（[渲染意图联合类型](../architecture/2026-07-02-tool-render-intent-union.md)）：调用视图携带命令、一段可选的模型撰写描述以及工作目录，结果视图携带输出、退出码与终止信号。该视图早已抵达浏览器——host、connection 与 runtime 把它投递到 `ConversationSnapshot` 的 `callView`/`resultView` 上——原 TUI 曾把它渲染为带 `$` 提示符的卡片，附退出行与首尾高度上限。

 Web client 却对它视而不见。`packages/client/ui-conversation/src/client/contract/tool-call-model.ts` 仅从原始工具参数推导每一行，`skeleton/DetailsPanel.tsx` 则把所有工具的内容块压平进一个 `<pre>`，样式为 `white-space: pre-wrap; word-break: break-word`。软换行加上没有高度约束，带来两个缺陷：多列输出（`ls`、表格、制表符绘图）被折成一段文字，丢掉了这类输出赖以存在的列对齐；而单列的长列表会把详情面板拉长到与列表等长。

@@ -19,7 +19,7 @@ Web client 却对它视而不见。`packages/client/ui-conversation/src/client/c
 - **提示符行，每条命令行一行。** 命令的每一行各占一行：标签，其后原样跟随该行。因此一个在两行上承载两条 shell 命令的 `command` 就读作它本身的两条命令，而不是被压成一行并省略号截断。标签取 cwd 的最后一段路径，当 cwd 等于 `home` prop 时取 `~`——浏览器没有 `$HOME`，因此由调用方提供绝对家目录，不提供时该折叠不生效。视图不带 cwd 时渲染一个纯 `$`。末尾换行是终止符，不是一条空的末命令。只有**第一行**携带该标签：视图只知道一个工作目录——调用开始处的那个——而后面的行完全可能在别处运行，命令里一个 `cd` 就足以改变它。把标签在各行重复，等于陈述一个此处无人知晓的逐行目录，这与运行状态点只出现一次是同一个理由。其余行保留一个裸 `$`，因此它们仍读作提示符。
 - **整次调用一枚运行状态点，位于第一行。** 它是 `StateDot` 四种状态中的三种：运行期间为追逐动画，与渲染状态徽章相同的退出状态为红色，干净落定为绿色——与工具行行首图标使用同一个指示器，因此一行与其自身的卡片不可能对同一条命令产生分歧。该状态点存在的理由是：读者对一条 shell 命令的第一个问题就是它是否仍在运行；没有它时，这一点只能从「没有输出」推断，而一条落定后无输出的命令看起来也一样。它以脱离文档流的方式落在卡片以**自身左内边距**预留的落区里，因此既不会缩进其命令，也不依赖命令自身的文本度量来与之对齐。该预留用 padding 而非 margin，是因为每个渲染点都会整条重写 `margin` 来设定自己的缩进——那会静默取消基于 margin 的落区，并让容器把状态点裁掉。无论有多少行，都只有一枚：视图携带的退出状态属于整次调用，而 bash 不报告逐条命令的状态，因此每行一枚状态点就等于在断言——一条在失败调用中其实成功了的命令行自身失败了。那一处视觉隐藏的文本标签具有相同的作用域，因为 `StateDot` 是 `aria-hidden`，而每行一个标签会被辅助技术读成好几个各自独立的结果。
 - **不软换行。** 输出行使用 `white-space: pre`，置于横向滚动的容器内。列对齐得以保留；长行滚动，而非折行。
- **高度上限与展开控件。** 输出超过 `DEFAULT_TERMINAL_MAX_LINES`（16）行时，显示 `ceil(max/2)` 行首部加余下的尾部行数，中间是一个按钮，报告被隐藏的行数并可展开。计数针对的是剥除输出末尾终止符之后解析出的行，因此以换行结尾的 N 行输出就是 N 行。切分算法与 TUI transcript 折叠态工具卡片（`packages/ui/tui/src/components/transcript.ts`）完全一致，因此同一条命令的首尾切片在两个前端之间吻合。
+- **高度上限与展开控件。** 输出超过 `DEFAULT_TERMINAL_MAX_LINES`（16）行时，显示 `ceil(max/2)` 行首部加余下的尾部行数，中间是一个按钮，报告被隐藏的行数并可展开。计数针对的是剥除输出末尾终止符之后解析出的行，因此以换行结尾的 N 行输出就是 N 行。切分算法保留原 TUI transcript 折叠卡片的行为，因此既有的首尾选择保持稳定。
 - **ANSI 颜色。** `anser` 切分 SGR 分段；`ui-primitives/src/ansi.ts` 把每段解析为内联样式，渲染成 React span。只设前景色的分段把基本 16 色映射到 `--dsw-*` 主题 token，使作者指定的颜色在两种主题下都可读；自行绘制背景的分段则前后景都保留 anser 给出的字面 rgb，以保住它意图中的对比度，256 色板、truecolor 以及本设计系统没有对应 token 的两种基本色同样如此。不承载颜色的转义序列（OSC 串、非 CSI 转义、无显示意义的 C0 控制符）在解析前被剥除，因此绝不会以字面字符抵达 DOM。光标移动在该剥除之前先行结算，且落在逐行的列缓冲里而不是靠字符串手术，因为回车与退格**只移动**光标——两者都不擦除任何东西，所以读者看到的就是每一列最后被写入的内容。`100%` 后接回车再接 `OK` 显示为 `OK0%`，因为这次重绘比它下面的帧更短；末尾 `abc` 加一个退格仍显示 `abc`，因为没有任何东西覆盖过那个 `c`；`abc` 加两个退格再接 `XY` 显示 `aXY`。这些用例都对照真实终端核实过，因为先前「截断加删除」的近似看起来是对的，实际并不对。SGR 状态按列打戳，与终端按单元格存储颜色的方式一致，因此部分覆盖会保留每个存活字符自身的颜色：红色 `bad`、三个退格、再写 `ok`，显示为 `okd` 且那个 `d` 仍是红的。CSI 序列不占列，只改变后续写入被打上的状态——这也正是回车不会重置颜色的原因，以及 SGR 状态会从一行延续到下一行、而不是在每个换行处关闭的原因。行内擦除属于同一次重放，因为 `\r\x1b[K` 是每个 spinner 与进度条都会写的同一个惯用法——只建模 `\r` 会让上一帧的尾巴留在原处，那是终端从未显示过的文本。只有 `m` 会累加进单元格样式；光标或擦除序列不能累加，否则状态串会随每次重绘线性增长，并发出 anser 只能丢弃的边界。SGR 按单元格以**归一化记录**保存（前景、背景、属性集合），而不是序列历史：累积原始序列会让每个状态边界重新发射整条链，因此不做完整 reset 的换色输出会发射 O(n^2) 个字符——3200 个这样的单元格产生 25 MB 并最终 `RangeError`，远低于 bash 自身的输出上限。该记录也让所有 chalk 系工具写出的属性闭合码（`39`、`49`、`22`、`24` 等）真正闭合其属性，且每个边界只为它开启的状态发射一条规范序列。一个分段也必须**收束**：重放收敛到扫描结束时的状态，而不是最后一个被写入单元格的状态——因为最后一次写入之后的 reset 不改变任何单元格，却结束了该分段；没有这一步，以 `\x1b[0m` 结尾的行会把颜色泄漏到其后所有行。光标按终端列推进，因此制表符前进到下一个 8 列制表位、宽字符占两列（其续列在首列被覆盖后变为空白而非合拢），组合标记不占列。宽度依据 emoji **presentation** 而非 U+2600–U+27BF 整个区块：`\u2713`——每条进度行都会写的对勾——只占一列，把该区块整体当作双宽恰好会错位这张卡片赖以存在的那类输出。写入宽字符对的任一半都会把另一半清成空白，因为终端无法让一个双格字形只留下一格：`a\tb` 之后用 `XY` 重绘显示为 `XY      b`，因为两个字符的重绘到不了第 8 列。
 - **退出状态与复制。** 非零退出码或信号渲染一枚状态徽章，与 bash 工具自身渲染器所作的退出状态区分一致；干净退出不渲染徽章，落定后的空输出渲染一处变暗的占位文字——该判定读的是卡片实际渲染的解析行，而非原始文本，因为只含转义或控制字节的输出能通过 `trim()` 却解析不出任何可见内容，否则就会画出一片空行外加一个把不可见字节写进剪贴板的复制控件。复制控件复制的是原始输出文本而非渲染后的树，因此提示符行与徽章不会进入剪贴板。

@@ -2,5 +2,5 @@
 # side as of the last confirmed-consistent state. Both languages carry equal authority;
 # after editing either side, bring the other along and re-record with:
 #   pnpm run verify-translation-pairing --write .agents/notes/implemented/feature/2026-07-30-search-render-card.md
-2026-07-30-search-render-card.md: 36f772d7198ef30d6c243549cbaa9f16c780268b
-2026-07-30-search-render-card.zh.md: 7d7ba352f19f3fb83cb6b7d049980776dc2c277e
+2026-07-30-search-render-card.md: 29544cb703f3ab048f4e7702935887ecf05179bf
+2026-07-30-search-render-card.zh.md: e0ae21924e82152ba629ab628476113ad9afe3d8
@@ -18,7 +18,7 @@ The discriminant is `shape`, not `kind`, deliberately: the same presentation mod

 One view with two shapes rather than two cards, because both tools are the same visual object — a search result — and a web consumer switches on one `card` value, then on `shape` for the row layout. The discriminated `shape` keeps each variant's fields non-optional (a matches view always has `files`, a paths view always has `paths`) instead of a single interface where every shape-specific field is optional.

-The view carries **no** result text. An earlier revision attached the model-facing `result.content` to the view; that was a no-op for every consumer (the TUI already falls back to `result.content`, and web fallbacks read the raw `tool/result` content), and it serialized the whole search text a second time into the persisted view. The view is the structured shape only; a UI without a search card falls back to the raw `tool/result` content.
+The view carries **no** result text. An earlier revision attached the model-facing `result.content` to the view; that was a no-op because consumer fallbacks already read the raw `tool/result` content, and it serialized the whole search text a second time into the persisted view. The view is the structured shape only; a UI without a search card falls back to the raw result content.

 The card tag is result-time only. A search call stays a `GenericCallView` (`kind: 'search'`): the pending state has no matches or paths to show, so there is nothing a `SearchCallView` would carry that the generic title does not. This is the asymmetry with the terminal card, whose call view carries the command, cwd, and description that exist before execution; a search's structured content exists only after `execute`.

@@ -30,7 +30,7 @@ The card tag is result-time only. A search call stays a `GenericCallView` (`kind

 The `SearchMeta` member shapes are object-literal `type` aliases, not the `SearchFileMatches`/`SearchLineMatch` interfaces the view exposes, because only a type alias is assignable to the `JsonValue` index signature `presentationMeta` returns; the two are structurally identical, so the projected value still reads back as a `SearchResultView`.

-The TUI (`packages/ui/tui/src/components/transcript.ts`) needs no dedicated arm: its result-view switch handles `terminal` and `diff` explicitly, and a `search` view falls through to the same dim generic body, reading the model-facing text from `this.result?.content`. Because the search view carries no `content` of its own and grep/glob returned a generic card before this PR, the TUI output stays byte-identical to the pre-search-card fallback. The web frontend that renders the structured `files`/`paths` shape is a separate later PR; this PR is the backend contract and its two producers.
+A consumer without a dedicated `search` arm falls back to the same generic body and reads the model-facing text from the raw result. Because the search view carries no `content` of its own and grep/glob returned a generic card before this PR, that fallback stays byte-identical to the pre-search-card path. The frontend that renders the structured `files`/`paths` shape is independent of this backend contract and its two producers.

 ## Alternatives considered

@@ -48,7 +48,7 @@ The TUI (`packages/ui/tui/src/components/transcript.ts`) needs no dedicated arm:

 `grep` and `glob` now compute `presentationMeta` on every non-nested successful call, a bounded projection over the already-retained matches or paths — the same retention outcome the render consumes, so there is no second retention pass and no doubled search text on the wire. The serialized meta is bounded by `searchMetaMaxBytes`, so a broad search no longer persists an unbounded structured copy into the session log.

-A UI without a search card renders the raw `tool/result` content, so no consumer regresses, and the TUI stays byte-identical. The web consumer that renders the structured shape reads `truncated`/`total` and the per-file groups; because the view carries only the retained, byte-bounded page, a UI wanting the complete result follows the spill locator in the model-facing text, exactly as the model does.
+A UI without a search card renders the raw `tool/result` content, so no consumer regresses. A consumer that renders the structured shape reads `truncated`/`total` and the per-file groups; because the view carries only the retained, byte-bounded page, a UI wanting the complete result follows the spill locator in the model-facing text, exactly as the model does.

 ## Testing

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

 用一个带两种形状的视图而非两张卡片，因为两个工具是同一个视觉对象 —— 一个搜索结果 —— web 消费方先在一个 `card` 值上分支，再在 `shape` 上分支决定行布局。判别式 `shape` 让每个变体的字段保持非可选（matches 视图总有 `files`，paths 视图总有 `paths`），而不是一个所有形状相关字段都可选的单一接口。

-该视图**不**携带结果文本。早期版本曾把面向模型的 `result.content` 附到视图上；那对每个消费方都是 no-op（TUI 本就回退到 `result.content`，web 回退读原始 `tool/result` 内容），却把整段搜索文本又序列化进持久化视图一遍。视图只承载结构化形状；无 search 卡片的 UI 回退到原始 `tool/result` 内容。
+该视图**不**携带结果文本。早期版本曾把面向模型的 `result.content` 附到视图上；但消费方的回退路径本就读取原始 `tool/result` 内容，因此这不会产生效果，却会把整段搜索文本又序列化进持久化视图一遍。视图只承载结构化形状；无 search 卡片的 UI 回退到原始结果内容。

 卡片标签只在结果时存在。搜索调用保持为 `GenericCallView`（`kind: 'search'`）：pending 状态没有匹配或路径可展示，所以 `SearchCallView` 能携带的东西不会比 generic 标题更多。这是与 terminal 卡片的不对称之处 —— terminal 的调用视图携带执行前就存在的命令、cwd、description；搜索的结构化内容只在 `execute` 之后才存在。

@@ -30,7 +30,7 @@ Status: implemented

 `SearchMeta` 的成员形状是对象字面量 `type` 别名，而非视图暴露的 `SearchFileMatches`/`SearchLineMatch` 接口，因为只有 type 别名可赋给 `presentationMeta` 返回的 `JsonValue` 索引签名；两者结构等价，所以投影值仍读回为 `SearchResultView`。

-TUI（`packages/ui/tui/src/components/transcript.ts`）不需要专门分支：它的结果视图 switch 显式处理 `terminal` 与 `diff`，`search` 视图落入同一个变暗的 generic body，从 `this.result?.content` 读取面向模型的文本。因为搜索视图不带自己的 `content`，而本 PR 之前 grep/glob 返回的是 generic 卡片，所以 TUI 输出与无 search 卡片的回退逐字节一致。渲染结构化 `files`/`paths` 形状的 web 前端是另一个后续 PR；本 PR 是后端契约及其两个生产者。
+没有专用 `search` 分支的消费方会回退到同一个 generic body，并从原始结果中读取面向模型的文本。因为搜索视图不带自己的 `content`，而本 PR 之前 grep/glob 返回的是 generic 卡片，所以该回退与引入 search 卡片之前的路径逐字节一致。渲染结构化 `files`/`paths` 形状的前端独立于这个后端契约及其两个生产者。

 ## 考虑过的备选

@@ -48,7 +48,7 @@ TUI（`packages/ui/tui/src/components/transcript.ts`）不需要专门分支：

 `grep` 与 `glob` 现在在每次非嵌套的成功调用上计算 `presentationMeta`，这是对已保留匹配或路径的一次有界投影 —— 与 render 消费的是同一份保留产出，所以没有第二次保留计算，线上也没有翻倍的搜索文本。序列化 meta 受 `searchMetaMaxBytes` 约束，所以宽泛搜索不再把无界的结构化副本持久化进会话日志。

-无 search 卡片的 UI 渲染原始 `tool/result` 内容，所以没有消费方退化，TUI 也逐字节一致。渲染结构化形状的 web 消费方读 `truncated`/`total` 与按文件分组；因为视图只携带保留的、字节有界的页，想要完整结果的 UI 跟随面向模型文本里的 spill 定位符，与模型的做法完全一致。
+无 search 卡片的 UI 渲染原始 `tool/result` 内容，所以没有消费方退化。渲染结构化形状的消费方读 `truncated`/`total` 与按文件分组；因为视图只携带保留的、字节有界的页，想要完整结果的 UI 跟随面向模型文本里的 spill 定位符，与模型的做法完全一致。

 ## 测试

@@ -2,5 +2,5 @@
 # side as of the last confirmed-consistent state. Both languages carry equal authority;
 # after editing either side, bring the other along and re-record with:
 #   pnpm run verify-translation-pairing --write .agents/notes/implemented/feature/2026-07-30-web-read-card.md
-2026-07-30-web-read-card.md: 1fb3d61a113d26f6daf023fc791f3638055b5be0
-2026-07-30-web-read-card.zh.md: 946bcca95bcc9bb50beb1ef22e77e4a21728b538
+2026-07-30-web-read-card.md: 26d1634b6be86666980e65f842de51c868c26efd
+2026-07-30-web-read-card.zh.md: 749177e93e8e3b2e34aed46d1fe99226395a6686
@@ -16,7 +16,7 @@ Add a fourth `card` tag, `read`, to the [render-intent union](../architecture/20

 The read tool projects the structured window through `output.presentationMeta`, the same persisted channel write/edit use for their applied-diff hunks ([canonical tool output contract](../architecture/2026-07-20-canonical-tool-output-contract.md)). `presentationMeta` runs once for a top-level surface call, returns `{ path, offset, lines, totalLines, lang? }` as JSON the session validates and stores on the result's `meta`, and `presentResult` narrows that meta back into the `ReadResultView` on both live and replay paths. `offset` (the 1-based first line the window requested) rides along because a byte cap below the first selected line yields an empty `lines` array with a positive `totalLines`; without the persisted `offset` a replayed card of such a window could not report where it starts or where a continuation resumes, and the last-line and re-parse fallbacks are both lossy. Without this channel the line array and total would be unreachable: the raw output object is not on the wire, and re-parsing the `N: text` text is lossy and fragile against the truncation footer.

-`presentResult` returns `undefined` — the generic fallback — whenever the meta is absent or malformed (`readMetaFromMeta` narrows it defensively, so a replay of an older logged result never throws), whenever the result is an error, and whenever the single text block is not the read envelope. A pre-card logged result — a valid read envelope with no persisted `meta`, recorded before this card existed — takes that same `undefined` path deliberately: the client falls back to the raw `result.content`, so it shows the enveloped `<path>/<type>/<content>` text rather than the envelope-stripped generic card the old presenter returned. This is the accepted degradation under the [pre-release stance](../../../../AGENTS.md#pre-release-stance-foundation-over-blast-radius): reject the old on-disk format rather than add an envelope-stripping compatibility branch, since this PR re-records every published fixture and the session format promises no backward compatibility. On the success path `presentResult` carries `content` (the envelope-stripped text) alongside the structured fields, so a UI without the read capability, including the current TUI, renders the file text through the generic/default card arm exactly as before. The TUI's `renderBody` switch (`packages/ui/tui/src/components/transcript.ts`) is not `assertNever`-exhaustive: `terminal` and `diff` have arms and everything else falls through to the generic arm, which reads `view.content`. That default arm alone was not enough: `render()` sets `genericContent` — and with it the dim-Markdown `dimBody` treatment — on a separate gate that was `card === 'generic'` only, so a `read` card would have kept the text but lost its dim styling. The gate now admits `card: 'read'` too, taking `content` down the same dim-Markdown path, so a read renders in the TUI exactly as it did before the read card existed. Beyond that one gate the TUI needs no read-specific code.
+`presentResult` returns `undefined` — the generic fallback — whenever the meta is absent or malformed (`readMetaFromMeta` narrows it defensively, so a replay of an older logged result never throws), whenever the result is an error, and whenever the single text block is not the read envelope. A pre-card logged result — a valid read envelope with no persisted `meta`, recorded before this card existed — takes that same `undefined` path deliberately: the client falls back to the raw `result.content`, so it shows the enveloped `<path>/<type>/<content>` text rather than the envelope-stripped generic card the old presenter returned. This is the accepted degradation under the [pre-release stance](../../../../AGENTS.md#pre-release-stance-foundation-over-blast-radius): reject the old on-disk format rather than add an envelope-stripping compatibility branch, since this PR re-records every published fixture and the session format promises no backward compatibility. On the success path `presentResult` carries `content` (the envelope-stripped text) alongside the structured fields, so a UI without the read capability renders the file text through its generic/default card arm. The former TUI established the need for this fallback: its non-exhaustive result switch read `view.content`, while a separate dim-Markdown gate also had to admit `card: 'read'`. That frontend has since been removed, but the content fallback remains part of the view contract for any consumer without a structured read card.

 ### Language hint derivation

@@ -34,13 +34,13 @@ The read tool projects the structured window through `output.presentationMeta`,

 ## Consequences

-`ToolResultView` has a fourth member. Every consumer that switches on `card` keeps compiling: the TUI and the current Web client route an unknown card to their generic path, and the read card carries `content` so that path shows the file text. The Web frontend that renders the line-numbered, syntax-highlighted view from `lines`/`lang`/`totalLines` is a separate follow-up PR; this PR is the backend that makes the data reachable. Until that lands, a read renders exactly as it did before (the generic text card) everywhere.
+`ToolResultView` has a fourth member. A consumer may render the structured `lines`/`lang`/`totalLines` shape or route an unsupported card to its generic path; the read card carries `content` so the latter still shows the file text. This producer change is the backend that makes the structured data reachable without requiring every consumer to implement the richer view at once.

 The read tool now computes `presentationMeta` for every top-level read, a small per-call projection (a `lines.map` and one `langFromPath` call) on data already in hand. The meta is persisted with the session log, so a read result is slightly larger on disk — the line array it already rendered as text, now also structured.

 ## Testing

-`packages/fs/tool-fs/tests/read-render.spec.ts` unit-tests `langFromPath` (known extensions case-insensitively, extension read after the last segment and last dot, and the `undefined` cases: dotfile, extensionless, trailing dot, unknown) and `readMetaFromMeta` (a well-formed narrow with and without `lang`, and every rejection: non-object, array, missing or wrong-typed `path`/`totalLines`/`lines`, a malformed line entry, a non-string `lang`, and — because the function narrows the opaque persisted `meta` boundary — the semantically invalid paths a well-typed replayed JSON can still carry: an `offset` that is not a 1-based integer, a first line `number` below `offset`, a line `number` that is not a 1-based integer (`0`, `1.5`, `NaN`, `Infinity`), a `totalLines` that is not a non-negative integer (`-1`, `1.5`, `NaN`), and lines whose numbers duplicate, decrease, or exceed `totalLines`; it also narrows an empty window at a positive `offset` (a byte cap below the first selected line). `packages/fs/tool-fs/tests/tools.spec.ts` pins the tool wiring: `execute` attaches the structured window (with and without a `lang` hint) as `meta`, `presentResult` narrows it into a `card: 'read'` view carrying the envelope-stripped `content`, and the decline paths (error result, non-single-text content, malformed envelope with valid meta, and valid envelope with absent or malformed meta) all fall back to `undefined`. Both changed source files hold per-file 100% coverage. This PR carries the snapshot evidence for the persisted meta and the extended union, not for a new rendered view: the re-recorded ACP session fixtures (`fs-read`, `fs-read-window`, `fs-edit`, `fs-policy-reject`, `fs-write-overwrite`, `parallel-tool-calls`, `workspace-context`, `workspace-edit`) pin the persisted read `meta` (with `{{cwd}}`-tokenized paths), and `cordis-inspect-jsdoc` pins the four-member `ToolResultView` union. The keyless snapshot and assembled-application transcript for the rendered read card belong to the follow-up Web PR that consumes the view, since this PR adds no new product-user-visible rendering — the TUI routes the read card through its existing generic dim-Markdown fallback (`transcript.ts` treats `card: 'read'` like `card: 'generic'`), so its output is unchanged. The `apps/cli` `parallel-file-reads` terminal golden (`apps/cli/tests/snapshots/parallel-file-reads/terminal.expected.txt`) pins exactly that: a real replay executes the read tool, renders it through the new `card: 'read'` gate, and the golden's dim-Markdown rows are byte-for-byte what a generic read produced before this card existed.
+`packages/fs/tool-fs/tests/read-render.spec.ts` unit-tests `langFromPath` (known extensions case-insensitively, extension read after the last segment and last dot, and the `undefined` cases: dotfile, extensionless, trailing dot, unknown) and `readMetaFromMeta` (a well-formed narrow with and without `lang`, and every rejection: non-object, array, missing or wrong-typed `path`/`totalLines`/`lines`, a malformed line entry, a non-string `lang`, and — because the function narrows the opaque persisted `meta` boundary — the semantically invalid paths a well-typed replayed JSON can still carry: an `offset` that is not a 1-based integer, a first line `number` below `offset`, a line `number` that is not a 1-based integer (`0`, `1.5`, `NaN`, `Infinity`), a `totalLines` that is not a non-negative integer (`-1`, `1.5`, `NaN`), and lines whose numbers duplicate, decrease, or exceed `totalLines`; it also narrows an empty window at a positive `offset` (a byte cap below the first selected line). `packages/fs/tool-fs/tests/tools.spec.ts` pins the tool wiring: `execute` attaches the structured window (with and without a `lang` hint) as `meta`, `presentResult` narrows it into a `card: 'read'` view carrying the envelope-stripped `content`, and the decline paths (error result, non-single-text content, malformed envelope with valid meta, and valid envelope with absent or malformed meta) all fall back to `undefined`. Both changed source files hold per-file 100% coverage. This PR carries the snapshot evidence for the persisted meta and the extended union, not for a new rendered view: the re-recorded ACP session fixtures (`fs-read`, `fs-read-window`, `fs-edit`, `fs-policy-reject`, `fs-write-overwrite`, `parallel-tool-calls`, `workspace-context`, `workspace-edit`) pin the persisted read `meta` (with `{{cwd}}`-tokenized paths), and `cordis-inspect-jsdoc` pins the four-member `ToolResultView` union. The then-current terminal snapshot also pinned that a consumer's generic dim-Markdown fallback stayed byte-identical; the structured card's own assembled-application transcript belonged to its consuming frontend change.

 ## Related

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

 read 工具通过 `output.presentationMeta` 投影结构化窗口，这与 write/edit 用来投影其应用 diff hunk 的持久化通道相同（[规范化工具输出契约](../architecture/2026-07-20-canonical-tool-output-contract.md)）。`presentationMeta` 对一次顶层 surface 调用运行一次，返回 `{ path, offset, lines, totalLines, lang? }` 作为会话校验并存储在结果 `meta` 上的 JSON，`presentResult` 在 live 和回放路径上都把该 meta 收窄回 `ReadResultView`。`offset`（窗口请求的 1-based 起始行）一并携带，是因为当字节上限低于首个选中行时，窗口会返回空的 `lines` 数组而 `totalLines` 为正；没有持久化的 `offset`，这类窗口的回放 card 就无法报告它从哪行开始、或续读应从哪行继续，而末行推断与文本重解析两种兜底都有损。没有这个通道，行数组和总数就无法触及：原始输出对象不在线上，而重新解析 `N: text` 文本既有损又对截断脚注脆弱。

-`presentResult` 在以下情况返回 `undefined`——即 generic 回退：meta 缺失或畸形（`readMetaFromMeta` 防御性收窄它，因此回放旧的已记录结果永不抛错）、结果是错误、以及单个文本块不是 read 信封。本 card 出现之前记录的结果——信封合法但无持久化 `meta`——有意走同一条 `undefined` 路径：客户端回退到原始 `result.content`，因此显示带 `<path>/<type>/<content>` 信封的原文，而非旧展示器返回的剥信封 generic card。这是 [pre-release 立场](../../../../AGENTS.md#pre-release-stance-foundation-over-blast-radius)下接受的降级：拒绝旧的磁盘格式，而非加一个剥信封的兼容分支——本 PR 已重录全部已发布 fixtures，且 session 格式不承诺向后兼容。在成功路径上，`presentResult` 在结构化字段之外携带 `content`（剥信封后的文本），因此不具备 read 能力的 UI（包括当前的 TUI）通过 generic/default card 分支渲染文件文本，与之前完全一致。TUI 的 `renderBody` switch（`packages/ui/tui/src/components/transcript.ts`）不是 `assertNever` 穷尽的：`terminal` 和 `diff` 有分支，其余都落入 generic 分支，该分支读取 `view.content`。仅有该默认分支还不够：`render()` 在一个独立门控上设置 `genericContent`（连同 dim-Markdown 的 `dimBody` 处理），该门控原先只判 `card === 'generic'`，因此 `read` card 虽保留文本却会丢失 dim 样式。现在该门控也接纳 `card: 'read'`，让 `content` 走同一条 dim-Markdown 路径，因此 read 在 TUI 中的渲染与 read card 出现之前完全一致。除这一处门控外，TUI 无需 read 专属代码。
+`presentResult` 在以下情况返回 `undefined`——即 generic 回退：meta 缺失或畸形（`readMetaFromMeta` 防御性收窄它，因此回放旧的已记录结果永不抛错）、结果是错误、以及单个文本块不是 read 信封。本 card 出现之前记录的结果——信封合法但无持久化 `meta`——有意走同一条 `undefined` 路径：客户端回退到原始 `result.content`，因此显示带 `<path>/<type>/<content>` 信封的原文，而非旧展示器返回的剥信封 generic card。这是 [pre-release 立场](../../../../AGENTS.md#pre-release-stance-foundation-over-blast-radius)下接受的降级：拒绝旧的磁盘格式，而非加一个剥信封的兼容分支——本 PR 已重录全部已发布 fixtures，且 session 格式不承诺向后兼容。在成功路径上，`presentResult` 在结构化字段之外携带 `content`（剥信封后的文本），因此不具备 read 能力的 UI 会通过自己的 generic/default card 分支渲染文件文本。原 TUI 证明了这条回退的必要性：它的非穷尽结果 switch 读取 `view.content`，而另一道 dim-Markdown 门控也必须接纳 `card: 'read'`。该前端随后被移除，但对任何没有结构化 read 卡片的消费方而言，content 回退仍是视图契约的一部分。

 ### 语言提示推导

@@ -34,13 +34,13 @@ read 工具通过 `output.presentationMeta` 投影结构化窗口，这与 write

 ## Consequences

-`ToolResultView` 多了第四个成员。每个在 `card` 上 switch 的消费者都继续编译：TUI 和当前 Web 客户端把未知 card 路由到其 generic 路径，而 read card 携带 `content` 使该路径显示文件文本。从 `lines`/`lang`/`totalLines` 渲染带行号、语法高亮视图的 Web 前端是单独的后续 PR；本 PR 是让数据可触及的后端。在它落地前，read 在各处的渲染与之前完全一致（generic 文本 card）。
+`ToolResultView` 多了第四个成员。消费方可以渲染结构化的 `lines`/`lang`/`totalLines` 形状，也可以将不支持的 card 路由到 generic 路径；read card 携带 `content`，所以后者仍会显示文件文本。本次生产者变更是让结构化数据可触及的后端，无需每个消费方同时实现更丰富的视图。

 read 工具现在为每次顶层 read 计算 `presentationMeta`，这是对已在手数据的一次小投影（一次 `lines.map` 和一次 `langFromPath` 调用）。meta 随会话日志持久化，因此 read 结果在磁盘上略大——它已渲染为文本的行数组，现在也以结构化形式存在。

 ## Testing

-`packages/fs/tool-fs/tests/read-render.spec.ts` 单测 `langFromPath`（已知扩展名的大小写不敏感、扩展名在最后一段与最后一个点之后读取、以及 `undefined` 各情况：dotfile、无扩展名、结尾的点、未知）与 `readMetaFromMeta`（含与不含 `lang` 的良构收窄，以及每种拒绝：非对象、数组、缺失或类型错误的 `path`/`totalLines`/`lines`、畸形行项、非字符串 `lang`，以及——因为该函数收窄持久化的 opaque `meta` 边界——良构类型的回放 JSON 仍可能携带的语义无效路径：不是 1-based 整数的 `offset`、小于 `offset` 的首行 `number`、不是 1-based 整数的行 `number`（`0`、`1.5`、`NaN`、`Infinity`）、不是非负整数的 `totalLines`（`-1`、`1.5`、`NaN`）、以及行号重复、递减或超过 `totalLines` 的情况；并且收窄正 `offset` 处的空窗口（字节上限低于首个选中行））。`packages/fs/tool-fs/tests/tools.spec.ts` 固定工具接线：`execute` 把结构化窗口（含与不含 `lang` 提示）作为 `meta` 附上、`presentResult` 把它收窄为携带剥信封 `content` 的 `card: 'read'` 视图、以及各拒绝路径（错误结果、非单文本内容、meta 有效但信封畸形、信封有效但 meta 缺失或畸形）都回退到 `undefined`。两个改动的源文件保持逐文件 100% 覆盖率。本 PR 携带的是持久化 meta 与扩展后联合类型的快照证据，而非新渲染视图的证据：重录的 ACP session fixtures（`fs-read`、`fs-read-window`、`fs-edit`、`fs-policy-reject`、`fs-write-overwrite`、`parallel-tool-calls`、`workspace-context`、`workspace-edit`）钉住持久化的读取 `meta`（含 `{{cwd}}` 令牌化路径），`cordis-inspect-jsdoc` 钉住四成员的 `ToolResultView` 联合类型。已渲染读取 card 的 keyless 快照与组装应用 transcript 属于消费该视图的后续 Web PR，因为本 PR 不新增任何面向产品用户可见的渲染——TUI 通过其现有的通用 dim-Markdown 回退路由读取 card（`transcript.ts` 把 `card: 'read'` 当作 `card: 'generic'` 处理），因此其输出保持不变。`apps/cli` 的 `parallel-file-reads` 终端 golden（`apps/cli/tests/snapshots/parallel-file-reads/terminal.expected.txt`）正钉住这一点：一次真实回放执行 read 工具、经新的 `card: 'read'` 门渲染，golden 的 dim-Markdown 行与本 card 出现前 generic read 所产出的逐字节一致。
+`packages/fs/tool-fs/tests/read-render.spec.ts` 单测 `langFromPath`（已知扩展名的大小写不敏感、扩展名在最后一段与最后一个点之后读取、以及 `undefined` 各情况：dotfile、无扩展名、结尾的点、未知）与 `readMetaFromMeta`（含与不含 `lang` 的良构收窄，以及每种拒绝：非对象、数组、缺失或类型错误的 `path`/`totalLines`/`lines`、畸形行项、非字符串 `lang`，以及——因为该函数收窄持久化的 opaque `meta` 边界——良构类型的回放 JSON 仍可能携带的语义无效路径：不是 1-based 整数的 `offset`、小于 `offset` 的首行 `number`、不是 1-based 整数的行 `number`（`0`、`1.5`、`NaN`、`Infinity`）、不是非负整数的 `totalLines`（`-1`、`1.5`、`NaN`）、以及行号重复、递减或超过 `totalLines` 的情况；并且收窄正 `offset` 处的空窗口（字节上限低于首个选中行））。`packages/fs/tool-fs/tests/tools.spec.ts` 固定工具接线：`execute` 把结构化窗口（含与不含 `lang` 提示）作为 `meta` 附上、`presentResult` 把它收窄为携带剥信封 `content` 的 `card: 'read'` 视图、以及各拒绝路径（错误结果、非单文本内容、meta 有效但信封畸形、信封有效但 meta 缺失或畸形）都回退到 `undefined`。两个改动的源文件保持逐文件 100% 覆盖率。本 PR 携带的是持久化 meta 与扩展后联合类型的快照证据，而非新渲染视图的证据：重录的 ACP session fixtures（`fs-read`、`fs-read-window`、`fs-edit`、`fs-policy-reject`、`fs-write-overwrite`、`parallel-tool-calls`、`workspace-context`、`workspace-edit`）钉住持久化的读取 `meta`（含 `{{cwd}}` 令牌化路径），`cordis-inspect-jsdoc` 钉住四成员的 `ToolResultView` 联合类型。当时的终端快照还钉住了消费方的 generic dim-Markdown 回退保持逐字节一致；结构化卡片自身的组装应用 transcript 则属于消费它的前端变更。

 ## Related

@@ -2,5 +2,5 @@
 # side as of the last confirmed-consistent state. Both languages carry equal authority;
 # after editing either side, bring the other along and re-record with:
 #   pnpm run verify-translation-pairing --write .agents/notes/implemented/feature/2026-07-30-web-result-card.md
-2026-07-30-web-result-card.md: deec27832aba2d5d868889f7306cbaef4f0b90b4
-2026-07-30-web-result-card.zh.md: 037e029332fbb665d90860d7e11c2fd117e6eb45
+2026-07-30-web-result-card.md: 838b13a7f3240c753e5cd1af6909389055c6352d
+2026-07-30-web-result-card.zh.md: 6e5170fcad82d709b050fb05e8efcfe955f20391
@@ -16,13 +16,13 @@ One tag with a `kind` discriminant, not two tags. Both calls are web retrieval a

 `presentationMeta` carries what render text cannot. The structured result object a tool returns from `execute` does NOT reach a client over the wire — only the model-facing `render` text and, when declared, the `output.presentationMeta` JSON projected onto the `tool/result` event's `meta` do. For `web_search` the meta is the ONLY faithful route to `{url, title?, snippet?, publishedAt?}`: the render collapses those fields into one lossy free-text line, so a consumer cannot reparse them. For `web_fetch` the meta is a smaller but real gain: `url`/`statusCode` are recoverable from the deterministic `Fetched <url> (HTTP <n>)` header line, but `truncated` is the effective truncation — provider cap, pre-conversion source cut, or the deployment's `fetchMaxOutputChars` output cap — which a client cannot recompute because it does not know that cap. The fetch card and the model-facing text derive `truncated` from one shared `renderFetchOutput(result, maxOutputChars)` helper, so the card never disagrees with the footer the model saw. This mirrors the write/edit diff template (`packages/fs/tool-fs/src/diff.ts`): a `*MetaFromValue` projector feeds `output.presentationMeta`, and a `*MetaFromResult` narrower reads `result.meta` back with a defensive fallback to the generic card. `web_fetch`'s body is already markdown in the result content, so it is not duplicated into meta.

-Neither result view carries a `content` copy. A UI that does not render the structured `web` card falls back to the raw `tool/result` content. The TUI does exactly this: it renders no structured web body, and its transcript renderer routes a `web` view's fallback content through the same dim Markdown path as a generic card's content (`packages/ui/tui/src/components/transcript.ts`, where both `render` and `renderBody` narrow the `generic` arm to `view.content` and give a `web` view the same `this.result?.content` fallback). Copying the result content into the view would duplicate up to `fetchMaxOutputChars` characters on the same delivered frame for no gain (the same rejection the meta section applies to the fetch body), so the views omit it and the fallback path renders the identical text. Each view sets its result-state `title` from the call args (`args.query` / `args.url`) so a window-truncated replay that dropped the call head still has a title, the way write/edit reset title at result time.
+Neither result view carries a `content` copy. A UI that does not render the structured `web` card falls back to the raw `tool/result` content, the same input a generic card consumes. Copying that content into the view would duplicate up to `fetchMaxOutputChars` characters on the same delivered frame for no gain (the same rejection the meta section applies to the fetch body), so the views omit it and the fallback path renders the identical text. Each view sets its result-state `title` from the call args (`args.query` / `args.url`) so a window-truncated replay that dropped the call head still has a title, the way write/edit reset title at result time.

 `presentResult` returns `undefined` (the generic card) on an error result and on absent or malformed `meta`, because presentation runs on replay of arbitrary logged results (possibly from an older schema) and must never throw. The narrowers validate every field defensively; an empty source list is valid meta, not malformed.

 ## Consequences

-The web frontend consumer is a separate later PR: this PR adds the contract arm and makes the two tools emit it, with no client-side rendering. The one observable change is that the `web_search`/`web_fetch` `tool/result` events now persist a `data.meta` payload (the `web-fetch` keyless snapshot is refreshed accordingly); the model-facing render text and the TUI presentation are unchanged (the TUI falls back to the same result content). The assembled-application transcript snapshot that exercises a `web` card belongs to the consumer PR that renders it, delivered there. Any existing `ToolResultView` consumer that switches exhaustively must add a `web` arm; the TUI does not switch exhaustively and needs none. `apiproxy`'s session schema already accepts any `card` string (`packages/host/apiproxy/src/api/sessions.schema.ts`), so the new view crosses the wire without a schema change.
+The frontend consumer was a separate later PR: this producer change adds the contract arm and makes the two tools emit it, with no client-side rendering. Its one observable change is that the `web_search`/`web_fetch` `tool/result` events persist a `data.meta` payload (the `web-fetch` keyless snapshot was refreshed accordingly); model-facing render text and generic fallback content stay unchanged. The assembled-application transcript snapshot that exercises a `web` card belongs to the consumer change that renders it. Any `ToolResultView` consumer that switches exhaustively must add a `web` arm; a non-exhaustive consumer may use the raw-result fallback. `apiproxy`'s session schema already accepts any `card` string (`packages/host/apiproxy/src/api/sessions.schema.ts`), so the new view crosses the wire without a schema change.

 A future web tool that wants this card declares `presentResult` returning a `card: 'web'` view with its own `kind`; adding a third `kind` is a union edit plus the frontend's branch, not a new card tag.

@@ -16,13 +16,13 @@ Status: implemented

 `presentationMeta` 携带 render 文本无法携带的东西。工具从 `execute` 返回的结构化结果对象**不会**经由 wire 抵达客户端——只有面向模型的 `render` 文本，以及（声明时）投影到 `tool/result` 事件 `meta` 上的 `output.presentationMeta` JSON 会。对 `web_search`，meta 是得到 `{url, title?, snippet?, publishedAt?}` 的**唯一**忠实途径：render 把这些字段压进一行有损的自由文本，消费者无法重新解析。对 `web_fetch`，meta 是更小但真实的收益：`url`/`statusCode` 可从确定格式的 `Fetched <url> (HTTP <n>)` header 行还原，但 `truncated` 是有效截断——provider cap、转换前源截断，或部署的 `fetchMaxOutputChars` 输出上限——客户端无法重算，因为它不知道那个上限。抓取卡片与面向模型的文本都从同一个 `renderFetchOutput(result, maxOutputChars)` helper 派生 `truncated`，因此卡片绝不会与模型看到的脚注分叉。这照搬 write/edit 的 diff 模板（`packages/fs/tool-fs/src/diff.ts`）：一个 `*MetaFromValue` 投影器喂给 `output.presentationMeta`，一个 `*MetaFromResult` 收窄器读回 `result.meta`，并在失败时防御性回退到 generic 卡片。`web_fetch` 的正文已是结果内容中的 markdown，因此不重复写入 meta。

-两个结果视图都不携带 `content` 副本。不渲染结构化 `web` 卡片的 UI 回退到原始 `tool/result` 内容。TUI 正是如此：它不渲染结构化的 web 正文，其 transcript 渲染器把 `web` 视图的回退内容与 generic 卡片的内容路由进同一条 dim Markdown 路径（`packages/ui/tui/src/components/transcript.ts` 中 `render` 与 `renderBody` 都把 `generic` 分支收窄为 `view.content`，并给 `web` 视图相同的 `this.result?.content` 回退）。把结果内容复制进视图会在同一投递帧上重复最多 `fetchMaxOutputChars` 个字符却毫无收益（与 meta 一节对抓取正文的否决同理），因此视图省略它，回退路径渲染完全相同的文本。每个视图从调用参数设置其结果期 `title`（`args.query`／`args.url`），因此丢掉了调用头的窗口截断重放仍有标题，与 write/edit 在结果期重设 title 的做法一致。
+两个结果视图都不携带 `content` 副本。不渲染结构化 `web` 卡片的 UI 回退到原始 `tool/result` 内容，这也是 generic 卡片消费的输入。把结果内容复制进视图会在同一投递帧上重复最多 `fetchMaxOutputChars` 个字符却毫无收益（与 meta 一节对抓取正文的否决同理），因此视图省略它，回退路径渲染完全相同的文本。每个视图从调用参数设置其结果期 `title`（`args.query`／`args.url`），因此丢掉了调用头的窗口截断重放仍有标题，与 write/edit 在结果期重设 title 的做法一致。

 `presentResult` 在错误结果、以及 `meta` 缺失或畸形时返回 `undefined`（即 generic 卡片），因为 presentation 会在对任意已记录结果（可能来自旧 schema）的重放中运行，绝不能抛错。收窄器防御性地校验每个字段；空来源列表是有效 meta，而非畸形。

 ## Consequences

-web 前端消费者是一个独立的后续 PR：本 PR 新增契约分支并让两个工具发出它，不含客户端渲染。唯一可观察的变化是 `web_search`/`web_fetch` 的 `tool/result` 事件现在持久化一个 `data.meta` 载荷（`web-fetch` keyless 快照随之刷新）；面向模型的 render 文本与 TUI 呈现不变（TUI 回退到相同的结果内容）。渲染 `web` 卡片的组装应用 transcript 快照属于渲染它的消费者 PR，在那里交付。任何做穷尽 switch 的现有 `ToolResultView` 消费者都必须新增一个 `web` 分支；TUI 并不穷尽 switch，无需新增。`apiproxy` 的会话 schema 已接受任意 `card` 字符串（`packages/host/apiproxy/src/api/sessions.schema.ts`），因此新视图无需 schema 变更即可跨 wire。
+前端消费方由后续独立 PR 交付：本次生产者变更新增契约分支并让两个工具发出它，不含客户端渲染。其唯一可观察的变化是 `web_search`/`web_fetch` 的 `tool/result` 事件持久化一个 `data.meta` 载荷（`web-fetch` keyless 快照当时随之刷新）；面向模型的 render 文本与 generic 回退内容保持不变。渲染 `web` 卡片的组装应用 transcript 快照属于渲染它的消费方变更。任何做穷尽 switch 的 `ToolResultView` 消费方都必须新增一个 `web` 分支；非穷尽消费方可以使用原始结果回退。`apiproxy` 的会话 schema 已接受任意 `card` 字符串（`packages/host/apiproxy/src/api/sessions.schema.ts`），因此新视图无需 schema 变更即可跨 wire。

 未来想用此卡片的 web 工具，声明一个返回带自有 `kind` 的 `card: 'web'` 视图的 `presentResult`；新增第三个 `kind` 是一次联合类型编辑加前端的分岔，而非一个新的 card 标签。

@@ -2,5 +2,5 @@
 # side as of the last confirmed-consistent state. Both languages carry equal authority;
 # after editing either side, bring the other along and re-record with:
 #   pnpm run verify-translation-pairing --write .agents/notes/implemented/feature/2026-07-31-web-telemetry-default-mount.md
-2026-07-31-web-telemetry-default-mount.md: e9ec7d0cda37db44e753c9aee572763b7e24ada6
-2026-07-31-web-telemetry-default-mount.zh.md: 68b411d0668772ce81d7f323c2d286714a223ca4
+2026-07-31-web-telemetry-default-mount.md: c1525a44196991d5059969ea70af34501b199323
+2026-07-31-web-telemetry-default-mount.zh.md: f203ea1943387beda445bd81ac78fa2cc0471d45
@@ -10,15 +10,15 @@ The telemetry seam and OTel backend ([revival Note](2026-07-23-session-telemetry

 ## Decision

-The shared `dsh` core (`apps/cli/config/base.cordis.yml`) mounts the `telemetry-otel` row by default with a baked-in production endpoint, so every surface — TUI, web, and headless — reports; this is the **internal-testing deployment stance** — reporting is on when an endpoint exists, and users opt out through the environment. Each surface's exit path drains the queue: web/headless dispose on SIGINT/SIGTERM (headless gained those handlers in this change), and the TUI's normal exit runs `disposeRootAndExit` (root dispose, 5s bounded — above the ~1s drain ceiling configured here) while its `/resume` handoff disposes the root before `execve`.
+The shared `dsh` base (`apps/cli/config/base.cordis.yml`) mounts the `telemetry-otel` row by default with a baked-in production endpoint, so Web and headless report; the raw-config command also mounts it before applying its required deployment overlay. This is the **internal-testing deployment stance** — reporting is on when an endpoint exists, and users opt out through the environment. Web and headless use the [bounded, escalating process-shutdown controller](../bug-fix/2026-08-03-cli-signal-shutdown-escalation.md) on SIGINT/SIGTERM, giving the backend's three-second shutdown deadline time to drain before the five-second launcher bound.

 | Ruling | Value | Rationale |
 |---|---|---|
-| Mount surface | base.cordis.yml (TUI + web + headless) | One deployment stance for every surface; per-surface divergence would need a reason, and none exists |
+| Mount surface | base.cordis.yml (raw config + Web + headless) | One deployment stance for every tree that loads the shared base; the raw overlay decides whether that deployment creates sessions |
 | Endpoint | `DSH_TELEMETRY_OTLP_URL`, default `https://harness-telemetry.deepseeksvc.com/v1/logs` | Internal collector; the env override serves local/dev runs |
 | Opt-out switch | any non-empty `DSH_TELEMETRY_DISABLED` (including `0`/`false`) disables | A privacy switch prefers off-by-mistake over on-by-mistake; a row can only be disabled at AppCLIEntry's patch layer (config has no disable semantic, and the switch must precede the load-time `exporter.url` validation) |
 | Cadence | `processor.scheduledDelayMillis: 10000` (10s/batch) | Streaming while the session runs, never exit-time-only; a crash loses at most the last unexported interval |
-| Exit-drain bound | `exporter.timeoutMillis: 1000` + `maxExportBatchSize: 2048` (== maxQueueSize) + `exportTimeoutMillis: 1500` | Dispose must release within ~1s against an unreachable collector: timeoutMillis doubles as the per-attempt socket timeout and the retry deadline (1s effectively disables the SDK's 5-try backoff), and aligning batch size with the queue cap makes the drain a single batch; SDK defaults can stall 40s+ |
+| Exit-drain bound | `exporter.timeoutMillis: 1000` + `maxExportBatchSize: 2048` (== maxQueueSize) + `exportTimeoutMillis: 1500` + `shutdownTimeoutMillis: 3000` | Ordinary unreachable-collector failure releases in ~1s: timeoutMillis is the per-attempt socket timeout and retry deadline, while one queue-sized batch avoids sequential drain multiplication. The DSH-owned 3s outer bound covers the SDK's preceding unbounded `forceFlush()` wait when the transport Promise never obtains a socket. |
 | Compression | `compression: gzip` | Event bodies carry full content; cross-datacenter bandwidth |
 | CI isolation | top-level `env: DSH_TELEMETRY_DISABLED: '1'` in all 8 GitHub workflows | Every CI channel that boots the web composition (e2e/snapshot/built smokes) must not stream test sessions to the production endpoint |

@@ -30,7 +30,7 @@ The keyless integration test `apps/cli/tests/telemetry-web.e2e.ts` pins the depl

 **A config field instead of an env patch for the switch.** Infeasible: cordis rows have no config-level disable semantic, and `exporter.url` validation fails loud at plugin construction, so the switch must take effect before the Loader — AppCLIEntry's patch layer is the only seat.

-**A `Promise.race` timeout backstop around exit.** Deferred: the parameter set already bounds the worst-case drain to ~1.5-3s (typically <100ms), measured SIGINT-to-exit 110ms-1.1s; the unbounded drip-feed-response risk stays under observation, and on real evidence the race lands inside the backend's `shutdown()` (never the coordinator — that would decide loss semantics for every backend).
+**A `Promise.race` timeout backstop around exit.** Originally deferred because the SDK parameters appeared to bound the backend's drain to ~1.5-3s (typically <100ms), with measured SIGINT-to-exit of 110ms-1.1s. A Linux sandbox reproduction later proved that `BatchLogRecordProcessor.shutdown()` can wait forever in `exporter.forceFlush()` before reaching its `exportTimeoutMillis`-bounded completion Promise. The [CLI shutdown fix](../bug-fix/2026-08-03-cli-signal-shutdown-escalation.md) therefore adds both a three-second backend bound for that specific gap and a five-second process-level bound plus repeated-signal escape for the whole plugin tree.

 ## Consequences

@@ -10,15 +10,15 @@ Status: implemented

 ## Decision

-`dsh` 共享核心（`apps/cli/config/base.cordis.yml`）默认挂载 `telemetry-otel` 行，内置生产 endpoint，因此所有 surface——TUI、web、headless——都上报；这是**内部测试期的部署立场**——有 endpoint 就报，用户可经环境变量退出。各 surface 的退出路径都会排空队列：web/headless 在 SIGINT/SIGTERM 上 dispose（headless 的信号处理是本次补上的），TUI 的正常退出走 `disposeRootAndExit`（根 dispose，5s 兜底——高于此处配置的 ~1s drain 上界），其 `/resume` 移交也在 `execve` 前 dispose 根。
+`dsh` 共享 base（`apps/cli/config/base.cordis.yml`）默认挂载 `telemetry-otel` 行，内置生产 endpoint，因此 Web 与 headless 都会上报；原始配置命令也会先挂载该行，再应用其必需的部署 overlay。这是**内部测试期的部署立场**——有 endpoint 就上报，用户可通过环境变量退出。Web 与 headless 在 SIGINT/SIGTERM 时使用[有界、可升级的进程关闭控制器](../bug-fix/2026-08-03-cli-signal-shutdown-escalation.md)，在启动器 5 秒上限到期前，先给后端 3 秒关闭截止时间完成排空。

 | 决策项 | 取值 | 理由 |
 |---|---|---|
-| 挂载面 | base.cordis.yml（TUI + web + headless） | 所有 surface 一个部署立场；按 surface 分化需要理由，而当前没有 |
+| 挂载面 | base.cordis.yml（原始配置 + Web + headless） | 所有加载共享 base 的配置树采用同一个部署立场；原始配置 overlay 决定该部署是否创建会话 |
 | endpoint | `DSH_TELEMETRY_OTLP_URL`，缺省 `https://harness-telemetry.deepseeksvc.com/v1/logs` | 内部 collector；env 覆盖供本地/联调 |
 | 退出开关 | `DSH_TELEMETRY_DISABLED` 非空（含 `0`/`false`）即关 | 隐私向开关取「宁关勿误开」；行级 disable 只能在 AppCLIEntry 的 patch 层做（config 无 disable 语义，且必须先于 `exporter.url` 的加载期校验生效） |
 | 上报节奏 | `processor.scheduledDelayMillis: 10000`（10s/批） | 流式回流，非退出才报；崩溃至多丢最后一个未导出间隔 |
-| 退出 drain 上界 | `exporter.timeoutMillis: 1000` + `maxExportBatchSize: 2048（== maxQueueSize）` + `exportTimeoutMillis: 1500` | collector 不可达时 dispose 必须 ~1s 内放行：timeoutMillis 同时是单次 socket 超时与重试 deadline（1s 等效关掉 SDK 5 次 backoff），批大小对齐队列上限使 drain 恒为单批；默认参数下最坏可卡 40s+ |
+| 退出 drain 上界 | `exporter.timeoutMillis: 1000` + `maxExportBatchSize: 2048（== maxQueueSize）` + `exportTimeoutMillis: 1500` + `shutdownTimeoutMillis: 3000` | collector 不可达的常规故障会在约 1s 内放行：timeoutMillis 是单次 socket 超时与重试 deadline，使用与队列等大的单批可避免依次排空导致耗时倍增。由 DSH 管理的 3s 外层上限覆盖 SDK 先执行的无界 `forceFlush()` 等待，即传输 Promise 始终无法取得 socket 的情况。 |
 | 压缩 | `compression: gzip` | 事件 body 含全文，跨机房带宽 |
 | CI 隔离 | 全部 8 个 GitHub workflow 顶层 `env: DSH_TELEMETRY_DISABLED: '1'` | CI 启动 web 组合的所有通道（e2e/snapshot/built smoke）不得向生产 endpoint 泄测试会话 |

@@ -30,7 +30,7 @@ Status: implemented

 **开关做成 config 字段而非 env patch。** 不可行：cordis 行没有 config 层的 disable 语义，且 `exporter.url` 校验在插件构造期 fail-loud，开关必须在 Loader 之前生效——AppCLIEntry patch 层是唯一落点。

-**退出时 `Promise.race` 兜底超时。** 暂缓：参数组合已把最坏 drain 压到 ~1.5-3s（典型 <100ms），实测 SIGINT→退出 110ms-1.1s；drip-feed 慢滴响应的无界等待风险留观，出现实证再在 backend `shutdown()` 内加 race（不放 coordinator——那会替所有 backend 决定丢失语义）。
+**退出时 `Promise.race` 兜底超时。** 最初暂缓，是因为 SDK 参数看似已经将后端排空耗时限制在约 1.5-3s（通常 <100ms），实测 SIGINT 到退出耗时 110ms-1.1s。后来在 Linux 沙箱中复现并证明，`BatchLogRecordProcessor.shutdown()` 可能在 `exporter.forceFlush()` 中永久等待，无法进入受 `exportTimeoutMillis` 限制的完成 Promise。因此，[CLI 关闭修复](../bug-fix/2026-08-03-cli-signal-shutdown-escalation.md) 既为这一特定缺口增加 3 秒后端上限，也为整棵插件树增加 5 秒进程级上限和重复信号退出途径。

 ## Consequences

@@ -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/2026-08-03-web-turn-run-time.md
+2026-08-03-web-turn-run-time.md: b6e79b34f45ebe46d9ce752b6333cdfce6fc4dd6
+2026-08-03-web-turn-run-time.zh.md: 73d9be4d2119278f8a78c6858bae353a4ef5d62f
@@ -0,0 +1,27 @@
+# Agent Note: Web turn run time and hover-revealed time chrome
+
+Status: implemented
+
+English | [中文](2026-08-03-web-turn-run-time.zh.md)
+
+## Problem
+
+The Web chat shows when a message arrived but not how long the agent worked on it. Long turns give no live progress signal beyond the static activity label, and after the turn settles the wall time is not recoverable from the UI. Meanwhile the always-visible clock row adds visual noise to every message.
+
+## Decision
+
+Turn wall time uses the existing logged `turn/start` and `turn/end` timestamps, with no new session events. The client Session folds each in-window pair into `turnTimings`; the actions-owning assistant footer renders `endTime - startTime` as a localized `Ran for {duration}` label after the turn ends. The running `TurnStatus` clock uses the latest timing without an end, so reload preserves elapsed time, steering does not reset it, and a retry starts from its own logged boundary. Both readings use the same localized formatter and whole-second floor. The clock appears only after 15 seconds and is hidden from the live region so screen readers announce the activity status without replaying every tick.
+
+Time chrome (clock and run time) is hover-revealed: message containers opt in with a `data-time-hover-root` attribute, and `MessageIconActions.module.css` fades the time label in on container `:hover`/`:focus-within`. The rule is scoped to `@media (hover: hover)`, so touch devices keep the always-visible label; opacity (not display) keeps the layout stable. Copy/branch icons stay always visible.
+
+## Alternatives considered
+
+**Deriving timing from message nodes.** The nearest user or steering timestamp is available in the rendered transcript, but it mismeasures retry turns and lets mid-turn steering reset the live clock. Existing turn boundary events provide the authoritative timestamps without changing the log format.
+
+**Anchoring the live clock to component mount.** Simpler, but a mid-turn reload would restart the clock at zero and disagree with the eventual footer label. Mount time remains only the fallback when `turn/start` is outside the loaded window.
+
+**Hiding the whole actions row until hover.** Copy and branch are affordances worth discovering, and row-level show/hide risks layout shift. Only the passive time text is hover-gated.
+
+## Consequences
+
+Turn duration is visible live and after settlement without new session events, and both readings share exact log boundaries and formatting. The settled duration includes activity after the last assistant text up to `turn/end`; the label is absent when `turn/start` is outside the loaded window. Time chrome no longer competes with message content at rest, and the ticking clock remains visual rather than repeatedly announced.
@@ -0,0 +1,27 @@
+# Agent Note: Web 轮次运行时长与悬停显示的时间附属元素
+
+Status: implemented
+
+[English](2026-08-03-web-turn-run-time.md) | 中文
+
+## 问题
+
+Web 聊天界面会显示消息的到达时间，却不显示 agent（智能体）处理这条消息花了多久。长轮次除静态活动标签外没有任何实时进度信号，轮次结束后也无法从 UI 中还原实际耗时。与此同时，始终可见的时钟行给每条消息都增加了视觉噪音。
+
+## 决策
+
+轮次实际耗时（wall time）采用日志中已有的 `turn/start` 和 `turn/end` 时间戳，不新增任何会话事件。客户端 Session 将加载窗口内的每对边界归并到 `turnTimings` 中；轮次结束后，承载操作图标的 assistant 页脚把 `endTime - startTime` 渲染为本地化的 `Ran for {duration}` 标签。运行中的 `TurnStatus` 时钟采用最新一条没有结束时间的计时记录，因此重新加载会保留已用时长，steering（中途引导）不会重置计时，重试也从自身的日志边界开始。两处读数共用同一个本地化格式化器，并向下取整到整秒。该时钟在 15 秒后才出现，并从实时区域中隐藏，因此屏幕阅读器会播报活动状态而不会重复播报每次时钟跳动。
+
+时钟与运行时长这类时间附属元素（time chrome）在悬停时才显示：消息容器通过 `data-time-hover-root` 属性显式启用该行为，`MessageIconActions.module.css` 在容器处于 `:hover`/`:focus-within` 时以淡入方式显示时间标签。该规则限定在 `@media (hover: hover)` 之内，触屏设备因此保持标签始终可见；显隐通过 opacity（而非 display）实现，布局保持稳定。复制与分支图标始终可见。
+
+## 考虑过的替代方案
+
+**从消息节点推导计时。** 渲染后的 transcript（文本记录）中可以取得最近的用户或 steering 时间戳，但这会错误计算重试轮次，并让轮次中途的 steering 重置实时时钟。已有的轮次边界事件无需改变日志格式即可提供权威时间戳。
+
+**将实时时钟锚定到组件挂载时刻。** 更简单，但轮次进行中重新加载会让时钟从零重新计时，并与最终的页脚标签不一致。仅当 `turn/start` 位于已加载窗口之外时，才回退到挂载时刻。
+
+**将整个操作行隐藏至悬停时才显示。** 复制与分支是值得让用户发现的操作入口，而整行级别的显隐切换有布局偏移的风险。只有被动的时间文本由悬停控制显隐。
+
+## 后果
+
+轮次时长在运行中和结束后都可见，且不需要新的会话事件；两处读数共用精确的日志边界和格式化方式。结束后的时长包括最后一条 assistant 文本之后、直至 `turn/end` 的活动；若 `turn/start` 位于已加载窗口之外，则不显示标签。未交互时，时间附属元素不再与消息内容争夺注意力，持续跳动的时钟也只保留视觉呈现，不会被重复播报。
@@ -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/2026-08-04-pointer-revealed-sidebar-scrollbars.md
+2026-08-04-pointer-revealed-sidebar-scrollbars.md: eb86f710e44cd8f1fb2cef41db99504af42413bb
+2026-08-04-pointer-revealed-sidebar-scrollbars.zh.md: 598c10dc2373eeb821d05f349a662156b194c71b
@@ -0,0 +1,63 @@
+# Agent Note: The sidebar's scrollbars follow the pointer
+
+Status: implemented
+
+English | [中文](2026-08-04-pointer-revealed-sidebar-scrollbars.zh.md)
+
+## Problem
+
+The sidebar's session list overflows after a handful of sessions, and from that point its scrollbar is drawn permanently — in a column that is at rest most of the time, next to rows whose own chrome only appears on hover. It is the one piece of always-on furniture in the sidebar, and nothing about it is actionable until someone reaches for it. The product ask (2026-08-04) is to draw it only while the pointer is over the sidebar, with a short tail so it does not blink out on the way past.
+
+## Decision
+
+`SidebarRoot` tracks the pointer over the whole column and carries a `quietBars` class whenever it is outside. The rule that class selects rebinds ui-theme's indirection pair — `--dsh-scrollbar-thumb` and `--dsh-scrollbar-thumb-hover` — to `transparent`, so every scroll region nested under the column draws no thumb. The session list is the only one today; a future one inherits the behavior rather than opting into it.
+
+The tail is `SCROLLBAR_LINGER_MS = 2000`: leaving arms a timer, entering cancels a pending one, and only the timer firing puts the class back. A pointer that crosses the column's edge and returns — travelling around a portalled menu, or overshooting on the way to a row — never sees the thumb blink.
+
+Entering is the column's own `pointerenter`; leaving is decided against the column's box, from a `pointermove` listener that exists only while the bars are drawn. DOM containment cannot decide the leave, because ui-settings renders its full-viewport settings panel as a fixed-position *descendant* of this column: a pointer moved onto that panel — or onto the conversation after it closes — never fires `pointerleave` here, and the bars would stay drawn over a column nobody is pointing at. The element's own leave is kept for the one case geometry cannot see, a pointer that leaves the window and emits no further moves.
+
+The pointer surface is the column, not the list. A pointer heading for the bar crosses the logo row, the New Session capsule, and the search field first, so revealing on the list alone would surface the bar only once the pointer was already among the rows.
+
+`transparent` is what makes the reveal free of layout. `scrollbar-gutter: stable` on the list exists so rows never move ([the gutter note](../bug-fix/2026-07-28-themed-scrollbars-and-reserved-gutter.md)); rebinding a colour leaves that reservation in force, so the thumb appears in space the list was already holding for it.
+
+The indirection pair rather than a rule on the list, because that pair is ui-theme's documented rebinding seam: one declaration reaches both rendering paths (the WebKit pseudo-elements and Firefox's `scrollbar-color`), and custom properties inherit, which is what makes the column — rather than each scroll region in it — the thing that owns the state.
+
+That widens the rebinding contract, so its gate states the new shape rather than accepting it by silence: `ui-theme/tests/scrollbar-styles.spec.ts` admits exactly two rebind targets, the l2 pair or `transparent`, and judges the *rule* rather than each declaration — a mixed rule (`thumb: transparent` beside an l2 hover) would repaint the bar the moment the pointer reached it while passing a per-declaration check. The elevation half compares whole values against the pair's canonical spelling, which is also what rejects a crossed pair and a token wrapped in a literal expression; an l1 rebind and a bare colour were already out.
+
+Hiding no longer counts as elevating: only an l2 rebind exempts a sheet from "every sheet that scrolls on an elevated surface rebinds". A sheet that hides its bars and also scrolls on an elevated surface still owes the l2 pair for whatever draws a thumb there.
+
+## Alternatives considered
+
+**CSS `:hover` on the column, with no JavaScript state.** The whole mechanism in one rule, and it cannot express the tail: the bar would vanish on the frame the pointer crossed the edge, which is exactly when a pointer is travelling to the conversation or around a portalled menu. The ask names the tail, and a hover-only version reads as flicker.
+
+**Keep it in CSS and get the delay from a transition,** by registering `--dsh-scrollbar-thumb` through `@property` so the custom property becomes animatable and a `transition-delay` could hold the colour. Rejected on cost and on reach: the registration is global to every surface that reads the pair, for one column's timing, and the WebKit scrollbar pseudo-elements this palette actually renders through do not reliably transition — the delay would be specified where it cannot be observed.
+
+**Hide the bar itself** — `scrollbar-width: none`, or `display: none` on `::-webkit-scrollbar`. Rejected because it takes the reserved band with it: the bar would reappear by re-taking 8px and shift every row sideways under the pointer that revealed it, which is the regression the gutter reservation was added to fix.
+
+**Draw an overlay thumb in the app** and hide the native bar entirely, which is what a fully custom fade would need. It buys arbitrary styling and costs hit-testing, drag, wheel, momentum, and both palettes' hover states — a large owned surface for a cosmetic gain, in a client whose scrollbars are already themed through tokens.
+
+**Scope the reveal to the scrolling list rather than the column.** Fewer elements involved, and it puts the reveal at the wrong boundary: the pointer reaches the rows last, so the bar would appear after the user is already reading them, and every other scroll region added to the sidebar later would have to opt in by hand.
+
+**Reveal on scroll events too,** so a keyboard- or touch-driven scroll shows the bar. Rejected as drawing an affordance the input that triggered it cannot use; the rows themselves already show that the list moved.
+
+## Consequences
+
+- A list scrolled by keyboard or by a touch drag shows no thumb once the linger passes, since neither leaves a pointer over the column. The e2e pins this rather than only describing it.
+- Dragging the thumb itself out of the column does not hide it mid-drag: the scrollbar takes the pointer capture, so the page receives no `pointermove` while the button is held. Measured in Chromium — the bar stays drawn and keeps scrolling with the pointer 900px to its right, past the linger window.
+- The column starts quiet on a cold load and stays so until the pointer first moves over it. A pointer already parked there when the page loads fires nothing until it moves, which is the browser's rule rather than this shell's.
+- An elevated surface nested in the column that rebinds the pair to l2 for its own elevation overrides the quiet state and keeps its bar drawn. Nothing in the sidebar does this today.
+- The shell's DOM now carries a state class, so ui-sidebar's shell snapshots pin `quietBars` and a regression in the default state is a snapshot diff rather than something someone has to notice in a screenshot.
+
+## Testing
+
+`packages/client/ui-sidebar/tests/pointer-scrollbars.spec.tsx` drives the class through the transitions with fake timers: revealed on entry, still revealed one millisecond before the linger closes, quiet one millisecond after, and cancelled by a return within the window. Two more cover the geometric leave: a `pointermove` landing outside the column's box hides the bars without any DOM leave (the settings-panel shape), and one landing back inside cancels a pending hide. It also unmounts mid-linger and asserts no timer survives — a pending hide firing into a dead component is the failure this shape is prone to. The events are `pointerover`/`pointerout` carrying a `relatedTarget`, because React synthesizes enter and leave from those and ignores the raw ones.
+
+`packages/client/ui-sidebar/tests/scrollbar-quiet-styles.spec.ts` reads the sheet: the rule states both halves of the pair — rebinding the resting thumb alone would leave the hover colour painting the moment the pointer reached the bar — and states no `scrollbar-gutter`, which belongs to the scrolling region.
+
+`apps/web/tests/sidebar-scrollbar.e2e.ts` is where the two halves meet a real engine. It parks the pointer over the list before every colour reading, since a scenario that never moves the mouse would measure the quiet state throughout and read as vacuous green. Its own test then moves the pointer away, asserts the thumb is still drawn on the leave itself, polls until it resolves to `rgba(0, 0, 0, 0)`, re-measures the geometry there to prove the reservation held while the bar was hidden, and scrolls the list programmatically — what a keyboard or a touch drag does — to pin that a pointerless scroll draws nothing. The committed golden records the thumb at both pointer positions in both palettes.
+
+The e2e's control is a mutation, and it needs the plugin's own bundle: dropping `quietBars` from the shell, rebuilding `@deepseek-ai/dsh-client-ui-sidebar` and only then `build:web`, turns that test red on the thumb resolving to `rgb(229, 229, 229)` where it expects `rgba(0, 0, 0, 0)`. Rerunning `build:web` alone exercises a stale bundle and passes with the change removed, which is the trap [the gutter note](../bug-fix/2026-07-28-themed-scrollbars-and-reserved-gutter.md) documented.
+
+The widened gate has its own controls, each a one-declaration mutation of a real sheet: crossing `transparent` with an l2 hover, and wrapping an l2 token in `color-mix(…)`, each turn the pair assertion red.
+
+The recording that demonstrates this behavior has to be headed. Headless Chromium reserves the band (`offsetWidth - clientWidth` is 8) but paints no thumb into a captured frame — measured by counting thumb-coloured pixels in the band across the reveal, which stays at noise level in headless and jumps from 46 to 1466 in a headed run.
@@ -0,0 +1,63 @@
+# Agent Note：侧边栏的滚动条跟随指针
+
+Status: implemented
+
+[English](2026-08-04-pointer-revealed-sidebar-scrollbars.md) | 中文
+
+## Problem
+
+侧边栏的会话列表只要有十来个会话就会溢出，从那一刻起它的滚动条就一直画在那里——所处的这一列大部分时间都是静止的，而列表行自己的操作按钮只在悬停时才出现。它是侧边栏里唯一始终常驻的构件，而在有人真的伸手去拖它之前，它不提供任何可操作性。产品诉求（2026-08-04）是只在指针位于侧边栏内时才绘制它，并留一小段拖尾，避免指针路过时它一闪而灭。
+
+## Decision
+
+`SidebarRoot` 跟踪整列上的指针，只要指针不在列内就给根元素挂上 `quietBars` 类。该类选中的规则把 ui-theme 的那组间接变量——`--dsh-scrollbar-thumb` 与 `--dsh-scrollbar-thumb-hover`——重新绑定为 `transparent`，于是嵌套在这一列下的每个滚动区域都不绘制滑块。今天这样的区域只有会话列表；将来新增的区域会直接继承这一行为，而不需要逐个接入。
+
+拖尾是 `SCROLLBAR_LINGER_MS = 2000`：离开会启动一个定时器，进入会取消尚未触发的定时器，只有定时器真正触发才会把类加回去。指针越过列边界又折返时——绕过一个 portal 菜单，或是奔向某一行时冲过了头——不会看到滑块闪动。
+
+进入用的是列自身的 `pointerenter`；离开则按列的盒子判定，由一个只在滚动条可见期间存在的 `pointermove` 监听完成。DOM 包含关系无法判定离开：ui-settings 把整屏的设置面板渲染为这一列的 fixed 定位**后代**，指针移到该面板上——或在面板关闭后移到对话区——都不会在这里触发 `pointerleave`，滚动条就会继续画在一个没人指向的列上。元素自身的 leave 仍然保留，用于几何判定看不到的那一种情况：指针移出窗口后不再产生任何移动事件。
+
+承载指针的是整列，而不是列表。奔向滚动条的指针会先经过 logo 行、New Session 胶囊和搜索框，所以只在列表上显示，会让滚动条等到指针已经落在行中间时才出现。
+
+`transparent` 正是让这次显示不触发任何布局的原因。列表上的 `scrollbar-gutter: stable` 存在的意义就是让行永不移动（见[空槽 Agent Note（agent 决策记录）](../bug-fix/2026-07-28-themed-scrollbars-and-reserved-gutter.md)）；重新绑定的只是颜色，那份预留始终有效，所以滑块出现在列表本就为它留出的空间里。
+
+选择这组间接变量而不是给列表加规则，是因为这组变量正是 ui-theme 写明的重新绑定接缝：一次声明同时作用于两条渲染路径（WebKit 伪元素与 Firefox 的 `scrollbar-color`），而自定义属性会继承——这正是让整列、而不是列内每个滚动区域，成为该状态所有者的原因。
+
+这拓宽了重新绑定契约，因此它的门禁把新的形态明写出来，而不是默许通过：`ui-theme/tests/scrollbar-styles.spec.ts` 只接受两种重新绑定目标，即 l2 那一组或 `transparent`，并且判定的是**整条规则**而不是逐条声明——混合规则（`thumb: transparent` 与 l2 的 hover 并列）会在指针一碰到滚动条时重新上色，却能通过逐条检查。抬升那一半按整个值与这组变量的规范写法比对，这同时也拒绝了交叉绑定和被包在字面表达式里的 token；绑回 l1 与裸颜色本来就在门外。
+
+隐藏不再算作抬升：只有 l2 重绑才能让一张样式表免于「任何既滚动又绘制抬升表面的样式表都必须重新绑定」。既隐藏滚动条又在抬升表面上滚动的样式表，仍然欠着那里真正绘制滑块时所需的 l2。
+
+## Alternatives considered
+
+**只用列上的 CSS `:hover`，不引入 JavaScript 状态。** 整套机制只需一条规则，但它表达不出拖尾：指针越过边界的那一帧滑块就会消失，而那恰好是指针正奔向对话区或绕行 portal 菜单的时刻。诉求本身点名了拖尾，只有 hover 的版本读起来就是闪烁。
+
+**留在 CSS 里、用过渡拿到这段延迟**，即通过 `@property` 注册 `--dsh-scrollbar-thumb` 让该自定义属性可动画，再用 `transition-delay` 把颜色按住。因代价与作用范围被否决：这项注册对每个读取这组变量的表面都是全局的，却只为一列的时序服务；而且这套调色板实际渲染所走的 WebKit 滚动条伪元素并不可靠地支持过渡——延迟会被声明在观察不到它的地方。
+
+**直接把滚动条藏掉**——`scrollbar-width: none`，或对 `::-webkit-scrollbar` 用 `display: none`。被否决，因为这会连带取消那段预留：滚动条重新出现时要重新占走 8px，会把每一行都在显示它的那个指针底下横向推移，而这正是当初加入空槽预留所修掉的回归。
+
+**在应用内自绘一个覆盖式滑块**，并彻底隐藏原生滚动条，这是完全自定义淡入淡出所需要的做法。它换来任意样式，代价是命中测试、拖拽、滚轮、惯性以及两套调色板下的 hover 状态——在一个滚动条已由 token 统一主题化的客户端里，为观感付出的是一大片自持表面。
+
+**把显示范围收敛到滚动的列表而不是整列。** 涉及的元素更少，却把显示的边界放错了位置：指针最后才到达行，滚动条会等到用户已经在读这些行时才出现；而且日后加入侧边栏的其他滚动区域都得手工接入。
+
+**滚动事件也触发显示**，让键盘或触摸驱动的滚动同样显示滚动条。被否决，因为那是在为触发它的输入方式画一个它用不上的可供性；行本身已经说明列表移动过了。
+
+## Consequences
+
+- 用键盘或触摸拖动滚动的列表，在拖尾结束后不显示滑块，因为这两种方式都不会把指针留在列上。e2e 会钉住这一点，而不只是把它写下来。
+- 拖动滑块本身移出列不会在拖动中途把它隐藏：滚动条会接管指针捕获，按住按键期间页面收不到 `pointermove`。已在 Chromium 实测——指针拖到列右侧 900px 处、超过拖尾窗口后，滚动条依然绘制并继续滚动。
+- 冷启动时该列处于静默状态，直到指针第一次移到它上面为止。页面加载时就停在那里的指针在移动之前不会触发任何事件，这是浏览器的规则，而非这个外壳的。
+- 嵌套在列内、为自身抬升层级把这组变量重新绑定到 l2 的抬升表面，会覆盖静默状态并继续绘制自己的滚动条。今天侧边栏内没有这样的表面。
+- 外壳的 DOM 现在带有一个状态类，因此 ui-sidebar 的外壳快照会钉住 `quietBars`，默认状态出现回归时表现为快照 diff，而不是需要有人从截图里看出来的东西。
+
+## Testing
+
+`packages/client/ui-sidebar/tests/pointer-scrollbars.spec.tsx` 用假定时器把这个类走过各次跃迁：进入时显示，拖尾结束前 1 毫秒仍然显示，结束后 1 毫秒转为静默，以及窗口内折返会取消隐藏。另有两条覆盖几何判定的离开：落在列盒子之外的 `pointermove` 会在没有任何 DOM leave 的情况下隐藏滚动条（即设置面板那种形态），落回盒子之内的则取消待触发的隐藏。它还在拖尾进行中卸载组件并断言没有定时器存活——待触发的隐藏落到已销毁的组件上，正是这种写法容易犯的错。事件用的是带 `relatedTarget` 的 `pointerover`／`pointerout`，因为 React 由它们合成 enter 与 leave，而会忽略原生的那两个事件。
+
+`packages/client/ui-sidebar/tests/scrollbar-quiet-styles.spec.ts` 直接读样式表：该规则必须写出这组变量的两半——只重新绑定静止态滑块，会让指针一碰到滚动条就露出 hover 颜色——并且不得出现 `scrollbar-gutter`，那属于滚动区域自己。
+
+`apps/web/tests/sidebar-scrollbar.e2e.ts` 是两半在真实引擎里汇合的地方。它在每次读取颜色前先把指针停在列表上，因为一个从不移动鼠标的场景全程测到的都是静默状态，会变成空洞的绿。随后它自己的用例把指针移开，断言在 leave 当下滑块仍在绘制，轮询直到它解析为 `rgba(0, 0, 0, 0)`，在该状态下重新测量几何以证明滚动条隐藏期间那份预留依然生效，并以编程方式滚动列表——键盘或触摸拖动所做的事——来钉住无指针滚动不绘制任何滑块。提交的 golden 记录了两套调色板下、两个指针位置上的滑块颜色。
+
+这条 e2e 的对照是一次 mutation，而它需要插件自己的产物：把 `quietBars` 从外壳中去掉，先重新构建 `@deepseek-ai/dsh-client-ui-sidebar`、之后再跑 `build:web`，该用例会因为滑块解析为 `rgb(229, 229, 229)`、而期望 `rgba(0, 0, 0, 0)` 而变红。只重跑 `build:web` 用的是过期产物，即使改动已被删除也照样通过，这正是[空槽 Agent Note（agent 决策记录）](../bug-fix/2026-07-28-themed-scrollbars-and-reserved-gutter.md)记录过的陷阱。
+
+拓宽后的门禁也有自己的对照，每个都是对真实样式表的一处声明改动：把 `transparent` 与 l2 的 hover 混用，以及把 l2 token 包进 `color-mix(…)`，都会让这条成对断言变红。
+
+演示这一行为的录制必须用有头浏览器。无头 Chromium 会预留那条带（`offsetWidth - clientWidth` 为 8），却不会把滑块画进捕获帧——通过统计带内滑块色像素在显示前后的变化实测：无头一直停在噪声水平，有头则从 46 跳到 1466。
@@ -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/2026-08-04-web-composer-shared-width-axis.md
+2026-08-04-web-composer-shared-width-axis.md: 96cddda25bf79e9f2df7a0298039c27f316befca
+2026-08-04-web-composer-shared-width-axis.zh.md: 9a9f5a513bbce8f3e97698d58ab48b0abdefb142
@@ -0,0 +1,31 @@
+# Agent Note: Web composer shared width axis and control-row polish
+
+Status: implemented
+
+English | [中文](2026-08-04-web-composer-shared-width-axis.zh.md)
+
+## Problem
+
+The web conversation column sized each surface independently: the transcript column, the input card, the todo/goal/queue dock cards, and the ask-question/approval/plan-review takeover cards each carried their own hardcoded max-width (736/752/776/800px variants) and their own side paddings. The surfaces drifted a few pixels apart at full width and diverged further on narrow viewports, where some panels kept clearance from the screen edge and others went flush. Separately, the composer's control row had no adaptive behavior — on a narrow card the permission trigger's label squeezed the row — and the overlay menus anchored to the card could render wider than the card itself, painting past its right edge.
+
+## Decision
+
+One content width variable owns the whole column. `--dsh-chat-content-width` (748px) is declared on ConversationRoot's `.root` — the transcript and the composer seat are sibling subtrees, so the declaration must sit on their common ancestor for CSS custom-property inheritance to reach both. Every other geometry derives from it: the input card caps at `content + 32px` (`--dsh-composer-card-max-width`), the dock cards subtract four dock insets (4 × 8px) from the card and land back on the content width, and the takeover cards use the content width directly. The narrow-viewport invariant is expressed structurally, not numerically: content-width surfaces pad `calc(var(--dsh-composer-side-clearance) + 16px)` per side while the input card clears the bare clearance (16px), so "input card = content + 32px" holds at every viewport width, not just at the cap.
+
+The control row inside the card is a `container-type: inline-size` container, and the permission trigger drops its text label (keeping glyph + chevron) under a 460px container query. The query is anonymous on purpose: CSS modules hash `container-name` per module, so a name declared in InputBar's sheet can never match a query written in PermissionSelect's sheet — the two hashed names silently differ and the query never fires. Only triggers that carry a mode glyph collapse (`:has(.triggerIcon)`); a host-configured mode without one keeps its text as its sole identifier.
+
+Overlay menus anchored to the card (slash menu, command popupSelect) clamp to the anchor's width (`max-width: min(<design cap>, 100%)`), truncating long rows with ellipses instead of overflowing the card. Tooltip bubbles keep a 12px viewport-edge safety margin in the clamp (ui-primitives Tooltip).
+
+## Alternatives considered
+
+**Keep per-surface widths and align the numbers by hand.** Rejected: the drift this change removes was exactly the residue of hand-aligned constants; any future width change would need five coordinated edits with nothing enforcing the relation.
+
+**Declare the variables on `.composerStack`.** Rejected after trying it: the takeover panels are siblings of the stack in the composer seat and the transcript is a different subtree entirely, so the variables never reached them; the common ancestor (`.root`) is the only correct home.
+
+**A named container query for the label collapse.** Rejected by measurement: CSS modules scope `container-name` per module, so the cross-module name never matched and the query was dead. The anonymous query resolves against the nearest ancestor container, which is unambiguous here (the row is the only container).
+
+**JS ResizeObserver for the label collapse.** Rejected: a container query is declarative, needs no listener lifecycle, and the 460px threshold is a design choice either way.
+
+## Consequences
+
+Changing the column width is now a one-line edit with the ratio relations preserved by construction, which the 736 → 748 retune during review already exercised. The cost is indirection: the widths of five surfaces are no longer readable off their own stylesheets and require following the variable chain to ConversationRoot. The container-query collapse adds the constraint that InputBar's row stays a size container; removing that declaration silently disables the permission trigger's adaptive behavior. The anonymous query also means any future second container between the row and the trigger would capture it — if that happens, the query must move or the intermediate container must be avoided.
@@ -0,0 +1,31 @@
+# Agent Note: Web 输入区共享宽度轴与控制行打磨
+
+Status: implemented
+
+[English](2026-08-04-web-composer-shared-width-axis.md) | 中文
+
+## Problem
+
+Web 会话列的各个界面各自独立设定尺寸：转录列、输入卡片、todo/goal/queue 停靠卡片、ask-question/approval/plan-review 接管卡片各自硬编码 max-width（736/752/776/800px 等变体）与各自的侧边内边距。这些界面在全宽下彼此漂移几个像素，在窄视口下偏差更大——有的面板保留了到屏幕边缘的间隙，有的却贴边。另外，输入卡片的控制行没有自适应行为——窄卡片下权限触发器的文字标签会挤压整行；锚定在卡片上的浮层菜单也可能渲染得比卡片更宽，越过其右边缘。
+
+## Decision
+
+一个内容宽度变量拥有整列。`--dsh-chat-content-width`（748px）声明在 ConversationRoot 的 `.root` 上——转录与 composer 座位是兄弟子树，声明必须放在共同祖先上，CSS 自定义属性才能通过继承同时到达两者。其他几何全部由它推导：输入卡片上限为 `content + 32px`（`--dsh-composer-card-max-width`），停靠卡片从卡片宽度中减去四个停靠 inset（4 × 8px）正好回到内容宽度，接管卡片直接使用内容宽度。窄视口不变式以结构而非数值表达：内容宽度的界面每侧 pad `calc(var(--dsh-composer-side-clearance) + 16px)`，而输入卡片只留裸 clearance（16px），因此"输入卡片 = 内容 + 32px"在任意视口宽度下都成立，而不只是在上限处。
+
+卡片内的控制行是一个 `container-type: inline-size` 容器，权限触发器在 460px 容器查询下收起文字标签（保留图标 + 下拉箭头）。查询刻意匿名：CSS modules 按模块哈希 `container-name`，InputBar 样式表里声明的名字永远无法匹配 PermissionSelect 样式表里写的查询——两个哈希后的名字悄然不同，查询永不触发。只有带模式图标的触发器才收起（`:has(.triggerIcon)`）；没有图标的宿主自定义模式保留文字作为其唯一标识。
+
+锚定在卡片上的浮层菜单（slash 菜单、command popupSelect）钳制到锚点宽度（`max-width: min(<设计上限>, 100%)`），过长的行以省略号截断而不是溢出卡片。Tooltip 气泡在钳制中保留 12px 的视口边缘安全距离（ui-primitives Tooltip）。
+
+## Alternatives considered
+
+**保留各界面独立宽度，手工对齐数值。** 否决：本次改动消除的漂移正是手工对齐常量的残留；未来任何宽度调整都需要五处协同编辑，且没有任何机制强制这组关系。
+
+**把变量声明在 `.composerStack` 上。** 尝试后否决：接管面板在 composer 座位中是 stack 的兄弟节点，转录更是完全不同的子树，变量根本到不了它们；共同祖先（`.root`）是唯一正确的家。
+
+**用命名容器查询实现标签收起。** 经实测否决：CSS modules 按模块作用域化 `container-name`，跨模块名字永不匹配，查询是死的。匿名查询解析到最近的祖先容器，在这里没有歧义（该行是唯一的容器）。
+
+**用 JS ResizeObserver 实现标签收起。** 否决：容器查询是声明式的，无需监听器生命周期，而 460px 阈值无论哪种方案都是设计选择。
+
+## Consequences
+
+修改列宽现在是一行编辑，比例关系由构造保证——评审期间 736 → 748 的重调已经验证了这一点。代价是间接性：五个界面的宽度不再能从各自的样式表直接读出，需要沿变量链追到 ConversationRoot。容器查询收起增加了一个约束：InputBar 的行必须保持为尺寸容器；删掉那条声明会静默禁用权限触发器的自适应行为。匿名查询也意味着未来若在行与触发器之间出现第二个容器，它会截获该查询——届时查询必须迁移，或避免中间容器。
@@ -2,5 +2,5 @@
 # side as of the last confirmed-consistent state. Both languages carry equal authority;
 # after editing either side, bring the other along and re-record with:
 #   pnpm run verify-translation-pairing --write apps/cli/README.md
-README.md: 360ab26ec3dfecf2841a012fda8947d6a84fdfec
-README.zh.md: 3ffa5d7726a798b784c67fdb8c4154fddbdea7a4
+README.md: 6fdca68eed11dffe46bf2fbde9a7899359690dca
+README.zh.md: d8d7122729df1dd8aaed8207ddfb0a0470778b01
@@ -49,6 +49,8 @@ The production Web runner needs built package and frontend artifacts (`pnpm run

 `dsh -p "task"` uses the same base and Web composition with the startup personal config, starts its Web host on an OS-assigned port, runs one fresh persisted session, prints the final answer, and exits. It accepts neither `--config` nor raw config-dump flags.

+Web and headless process shutdown gives the plugin tree up to five seconds to dispose. The first `SIGINT`/`SIGTERM` starts that graceful drain; a second signal forces immediate exit. If headless normal completion is already stuck in disposal, the first `Ctrl+C` is the escalation and exits immediately instead of being swallowed.
+
 Both modes treat the invoking directory as the default workspace root, load applicable `AGENTS.md` or `CLAUDE.md` instructions with a 65,536-byte render budget, and use an in-memory SQLite session content index. Web watches valid personal config edits; headless reads the file once at startup. The [app-boot personal-config contract](../../packages/ui/app-boot/README.md#personal-config) owns layer precedence, credential storage, live-update failure behavior, and `$DSH_HOME` resolution.

 New sessions default to the `workspace-write` permission preset. Bash and filesystem mutations are restricted to the session workspace and platform temporary roots; reads, network access, and process visibility are not confined. `DSH_PERMISSION_MODE` changes the process fallback. Stored General-settings permissions affect later Web sessions, not an already-open one.
@@ -49,6 +49,8 @@ dsh web --dump-config

 `dsh -p "task"` 使用相同的 base 与 Web 组合及启动时个人配置，在由操作系统分配的端口上启动 Web 宿主，运行一个全新的持久会话，打印最终答案后退出。它不接受 `--config` 或原始配置输出标志。

+Web 与 headless 的进程关闭流程最多给插件树 5 秒执行 dispose。第一次 `SIGINT`/`SIGTERM` 会启动这次优雅排空；第二次信号会立即强制退出。如果 headless 的正常完成流程已经卡在 dispose 中，第一次 `Ctrl+C` 就会触发强制退出：进程立即结束，该信号不再被吞掉。
+
 两种模式都以调用目录作为默认 workspace 根目录，加载适用的 `AGENTS.md` 或 `CLAUDE.md` 指令，渲染预算为 65,536 字节，并使用内存 SQLite 会话内容索引。Web 会持续应用有效的个人配置编辑；headless 只在启动时读取该文件一次。层次优先级、凭据存储、实时更新失败行为与 `$DSH_HOME` 解析均由 [app-boot 个人配置契约](../../packages/ui/app-boot/README.md#personal-config) 统一定义。

 新会话默认使用 `workspace-write` 权限 preset。Bash 和文件系统写操作受限于会话 workspace 与平台临时根目录；读取、网络访问与进程可见性不受限制。`DSH_PERMISSION_MODE` 会改变进程回退值。已存储的常规设置权限会影响之后的 Web 会话，不会更改已打开的会话。
@@ -111,17 +111,19 @@
 # process out (the launchers patch the row disabled; config cannot disable
 # a row). Exports carry the harness home's anonymous user id ($DSH_HOME/.userid,
 # random UUID; delete the file to reset the identity) as the Resource's
-# user.id. The exporter/processor values bound the shutdown drain to ~1s
-# against an unreachable collector: exporter.timeoutMillis is both the
-# per-attempt socket timeout and the retry deadline (1s effectively
-# disables the SDK's 5-try backoff), maxExportBatchSize == maxQueueSize
-# (both explicit) makes the drain a single batch, and exportTimeoutMillis
-# is the processor's own cap on that one export cycle — the second bound
-# when the exporter's clock alone does not fire. Every CLI exit path drains it
-# by disposing the root on SIGINT/SIGTERM.
+# user.id. The exporter/processor values normally bound the shutdown drain
+# to ~1s against an unreachable collector: exporter.timeoutMillis is both
+# the per-attempt socket timeout and the retry deadline (1s effectively
+# disables the SDK's 5-try backoff), while maxExportBatchSize == maxQueueSize
+# (both explicit) makes the drain a single batch. The SDK awaits
+# exporter.forceFlush() outside exportTimeoutMillis, so the backend's 3s
+# shutdownTimeoutMillis is the load-bearing outer bound when a transport
+# promise never settles. Every CLI exit path drains it by disposing the root
+# on SIGINT/SIGTERM.
 - id: telemetry-otel
  name: '@deepseek-ai/dsh-session-telemetry-otel'
  config:
+    shutdownTimeoutMillis: 3000
    exporter:
      url: !!js process.env.DSH_TELEMETRY_OTLP_URL ?? 'https://harness-telemetry.deepseeksvc.com/v1/logs'
      compression: gzip
@@ -14,6 +14,7 @@ import type { MuxFrame } from '@deepseek-ai/dsh-host-apiproxy/api'
 import type { RpcRequest, RpcResponse } from '@deepseek-ai/dsh-host-apiproxy/api/rpc'
 import type { SessionId } from '@deepseek-ai/dsh-session'
 import { AppCLIEntry } from './app-cli-entry.ts'
+import { createProcessShutdown } from './process-shutdown.ts'

 /** Outcome of one headless turn: aggregated final text plus the turn-end reason kind. */
 interface TurnOutcome {
@@ -21,12 +22,12 @@ interface TurnOutcome {
  reason: string
 }

-/** Unwrap an RpcResponse or fail loud: business errors print and exit 1 (dispose first). */
-async function unwrap<T>(response: RpcResponse<T>, dispose: () => Promise<void>): Promise<T> {
+/** Unwrap an RpcResponse or fail loud: business errors print and exit 1 (shutdown first). */
+async function unwrap<T>(response: RpcResponse<T>, shutdown: () => Promise<void>): Promise<T> {
  if (response.result.ok) return response.result.value
  const { code, message } = response.result.error
  process.stderr.write(`dsh: ${code}: ${message}\n`)
-  await dispose()
+  await shutdown()
  process.exit(1)
 }

@@ -82,23 +83,16 @@ export async function runHeadless(task: string): Promise<void> {
    port: 0,
  })
  const { ctx, port } = await entry.run()
-  const dispose = async (): Promise<void> => { await ctx.fiber.dispose() }
-  // Signal exits must still dispose the tree: the composition mounts
-  // exit-drained plugins (telemetry's queued tail and shutdown marker would
-  // otherwise be lost), and Node's default signal exit skips disposal.
-  let signalled = false
-  const disposeAndExit = (code: number): void => {
-    if (signalled) return
-    signalled = true
-    void dispose().finally(() => { process.exit(code) })
-  }
-  process.on('SIGTERM', () => { disposeAndExit(143) })
-  process.on('SIGINT', () => { disposeAndExit(130) })
+  // Normal completion and signals share one bounded drain. A signal received
+  // during that drain escalates immediately instead of becoming a no-op.
+  const shutdown = createProcessShutdown(async () => { await ctx.fiber.dispose() })
+  process.on('SIGTERM', () => { shutdown.interrupt(143) })
+  process.on('SIGINT', () => { shutdown.interrupt(130) })
  // The headless session is web-observable while it runs (same composition).
  process.stderr.write(`dsh: observing at http://127.0.0.1:${String(port)}\n`)
  const api = new InProcessApiClient(toFetchHandler(ctx.apiProxy))

-  const created = await unwrap(await api.sessions.create({}), dispose)
+  const created = await unwrap(await api.sessions.create({}), () => shutdown.shutdown(1))

  // Open the stream before prompting so no frame is lost — kept in this order
  // even though in-process delivery has no race, so the code survives a move
@@ -111,11 +105,10 @@ export async function runHeadless(task: string): Promise<void> {
    sessionId: created.sessionId,
    mode: 'queue',
    content: [{ type: 'text', text: task }],
-  }), dispose)
+  }), () => shutdown.shutdown(1))

  const outcome = await done
  process.stdout.write(outcome.text + '\n')
  abort.abort()
-  await dispose()
-  process.exit(outcome.reason === 'completed' ? 0 : 1)
+  await shutdown.shutdown(outcome.reason === 'completed' ? 0 : 1)
 }
@@ -0,0 +1,58 @@
+/** Bounded, escalating process shutdown for the long-lived CLI surfaces. */
+
+/** Maximum grace allowed for the application tree to dispose before process exit. */
+export const PROCESS_SHUTDOWN_TIMEOUT_MS = 5_000
+
+/** Process-exit controller shared by normal completion and Unix signal handlers. */
+export interface ProcessShutdown {
+  /** Start or join graceful disposal before exiting with `code`. */
+  shutdown(code: number): Promise<void>
+  /** Start graceful disposal, or force exit when a shutdown is already running. */
+  interrupt(code: number): void
+}
+
+/**
+ * Create one process-exit controller around an application disposer.
+ * @param dispose - Whole-application teardown that resolves at quiescence.
+ * @param exit - Process exit boundary, replaceable by tests.
+ * @param timeoutMs - Grace before forced exit, replaceable by tests.
+ * @returns A controller whose normal calls coalesce and whose repeated signal call escalates.
+ */
+export function createProcessShutdown(
+  dispose: () => Promise<void>,
+  exit: (code: number) => void = (code) => { process.exit(code) },
+  timeoutMs = PROCESS_SHUTDOWN_TIMEOUT_MS,
+): ProcessShutdown {
+  let pending: Promise<void> | undefined
+  let timeout: ReturnType<typeof setTimeout> | undefined
+  let exited = false
+
+  const exitOnce = (code: number): void => {
+    if (exited) return
+    exited = true
+    /* v8 ignore else -- shutdown() arms the timer before any asynchronous exit path can run. */
+    if (timeout !== undefined) clearTimeout(timeout)
+    exit(code)
+  }
+
+  const shutdown = (code: number): Promise<void> => {
+    if (pending !== undefined) return pending
+    timeout = setTimeout(() => { exitOnce(code) }, timeoutMs)
+    pending = Promise.resolve().then(dispose).then(
+      () => { exitOnce(code) },
+      () => { exitOnce(code) },
+    )
+    return pending
+  }
+
+  return {
+    shutdown,
+    interrupt(code) {
+      if (pending !== undefined) {
+        exitOnce(code)
+        return
+      }
+      void shutdown(code)
+    },
+  }
+}
@@ -13,6 +13,7 @@ import type {} from '@deepseek-ai/dsh-host-webserver'
 import type {} from '@deepseek-ai/dsh-system-prompt'
 import type {} from '@deepseek-ai/dsh-tool-bash'
 import { AppCLIEntry } from './app-cli-entry.ts'
+import { createProcessShutdown } from './process-shutdown.ts'

 // The shipped base plus the Web application's overlay.
 const BASE_CONFIG = fileURLToPath(new URL('../config/base.cordis.yml', import.meta.url))
@@ -118,17 +119,12 @@ export async function runWeb(
  const { ctx, port: boundPort } = await entry.run()
  const resolvedLocalWebUrl = localWebUrl(ctx)

-  let exiting = false
-  const shutdown = (code: number): void => {
-    if (exiting) return
-    exiting = true
-    void Promise.resolve(ctx.fiber.dispose()).finally(() => { process.exit(code) })
-  }
+  const shutdown = createProcessShutdown(async () => { await ctx.fiber.dispose() })

  // Install shutdown handling before publishing readiness: supervisors may
  // send a signal as soon as they observe the URL line.
-  process.on('SIGTERM', () => { shutdown(0) })
-  process.on('SIGINT', () => { shutdown(130) })
+  process.on('SIGTERM', () => { shutdown.interrupt(0) })
+  process.on('SIGINT', () => { shutdown.interrupt(130) })

  // The entry's boot-time snapshot, not a fresh sample: the printed LAN URL
  // must name an address the /api trust fence was configured with.
@@ -0,0 +1,18 @@
+/** Test-only Cordis plugin whose disposer announces entry and never settles. */
+
+import { existsSync } from 'node:fs'
+
+/**
+ * Register a disposer that keeps process shutdown pending until it is forced.
+ * @param {import('cordis').Context} ctx - loader-mounted test plugin context.
+ */
+export function apply(ctx) {
+  const keepAlive = setInterval(() => {}, 60_000)
+  ctx.effect(() => async () => {
+    clearInterval(keepAlive)
+    const armFile = process.env.DSH_TEST_SHUTDOWN_ARM_FILE
+    if (armFile === undefined || !existsSync(armFile)) return
+    process.stderr.write('dsh-test: never-dispose started\n')
+    await new Promise(() => {})
+  })
+}
@@ -0,0 +1,121 @@
+import { mkdir, mkdtemp, rm, writeFile } from 'node:fs/promises'
+import { tmpdir } from 'node:os'
+import { join } from 'node:path'
+import { fileURLToPath, pathToFileURL } from 'node:url'
+import { execa } from 'execa'
+import { describe, expect, it } from 'vitest'
+import { LOADER_SMOKE_TEST_TIMEOUT_MS, resolveExampleLaunch } from '@deepseek-ai/dsh-loader-smoke'
+
+const dshBinScript = fileURLToPath(new URL('../src/bin.ts', import.meta.url))
+const tsconfigPath = fileURLToPath(new URL('../../../tsconfig.json', import.meta.url))
+const neverDisposePlugin = pathToFileURL(
+  fileURLToPath(new URL('./fixtures/never-dispose.mjs', import.meta.url)),
+).href
+
+const POSIX_HEADLESS_PTY_DRIVER = String.raw`
+import errno, json, os, pty, select, signal, sys, time
+node, launch_args_json, launch_env_json, cwd, timeout_seconds = sys.argv[1:]
+env = os.environ.copy()
+env.update(json.loads(launch_env_json))
+pid, fd = pty.fork()
+if pid == 0:
+    os.chdir(cwd)
+    os.execvpe(node, [node, *json.loads(launch_args_json)], env)
+
+markers = [b"dsh: observing at ", b"dsh-test: never-dispose started"]
+output = bytearray()
+marker_index = 0
+deadline = time.monotonic() + float(timeout_seconds)
+status = None
+while time.monotonic() < deadline:
+    ready, _, _ = select.select([fd], [], [], 0.05)
+    if ready:
+        try:
+            chunk = os.read(fd, 65536)
+        except OSError as error:
+            if error.errno != errno.EIO:
+                raise
+            chunk = b""
+        if chunk:
+            output.extend(chunk)
+    while marker_index < len(markers) and markers[marker_index] in output:
+        if marker_index == 0:
+            open(os.path.join(cwd, "shutdown-armed"), "w").close()
+        os.write(fd, b"\x03")
+        marker_index += 1
+    waited, candidate = os.waitpid(pid, os.WNOHANG)
+    if waited == pid:
+        status = candidate
+        break
+
+if status is None:
+    os.kill(pid, signal.SIGKILL)
+    _, status = os.waitpid(pid, 0)
+sys.stdout.buffer.write(output)
+if marker_index != len(markers):
+    sys.stderr.write(f"completed {marker_index}/{len(markers)} PTY actions before timeout\n")
+    sys.exit(124)
+actual_exit = os.waitstatus_to_exitcode(status)
+if actual_exit != 130:
+    sys.stderr.write(f"expected exit 130, got {actual_exit}\n")
+    sys.exit(125)
+`
+
+async function runHeadlessPtySmoke(): Promise<string> {
+  const cwd = await mkdtemp(join(tmpdir(), 'dsh-headless-shutdown-'))
+  try {
+    const home = join(cwd, '.dsh')
+    await mkdir(home, { recursive: true })
+    await writeFile(join(home, 'config.yaml'), [
+      '- insert:',
+      '    - id: never-dispose',
+      `      name: '${neverDisposePlugin}'`,
+      '',
+    ].join('\n'))
+    const launch = resolveExampleLaunch({
+      srcBin: dshBinScript,
+      configArgs: ['-p', 'never complete'],
+      tsconfigPath,
+      env: {
+        DSH_HOME: home,
+        DSH_AGENTS_HOME: join(cwd, '.agents'),
+        DEEPSEEK_API_KEY: 'keyless-shutdown-no-call',
+        DSH_TELEMETRY_DISABLED: '1',
+        DSH_TEST_SHUTDOWN_ARM_FILE: join(cwd, 'shutdown-armed'),
+      },
+    })
+    const timeoutMs = 15_000
+    const result = await execa('python3', [
+      '-c',
+      POSIX_HEADLESS_PTY_DRIVER,
+      launch.command,
+      JSON.stringify(launch.args),
+      JSON.stringify(launch.env),
+      cwd,
+      String(timeoutMs / 1_000),
+    ], {
+      stdin: 'ignore',
+      timeout: timeoutMs + 5_000,
+      killSignal: 'SIGKILL',
+      reject: false,
+      stripFinalNewline: false,
+    })
+    if (result.timedOut) {
+      throw new Error(`dsh headless PTY driver did not exit. stdout:\n${result.stdout}\nstderr:\n${result.stderr}`)
+    }
+    if (result.failed) {
+      throw new Error(`dsh headless PTY driver exited ${String(result.exitCode)}. stdout:\n${result.stdout}\nstderr:\n${result.stderr}`)
+    }
+    return result.stdout
+  } finally {
+    await rm(cwd, { recursive: true, force: true })
+  }
+}
+
+describe.skipIf(process.platform === 'win32')('headless process shutdown (real Loader tree in a PTY)', () => {
+  it('lets a second Ctrl+C force exit while the first signal is draining', async () => {
+    const output = await runHeadlessPtySmoke()
+    expect(output).toContain('dsh: observing at ')
+    expect(output).toContain('dsh-test: never-dispose started')
+  }, LOADER_SMOKE_TEST_TIMEOUT_MS)
+})
@@ -0,0 +1,131 @@
+import { afterEach, describe, expect, it, vi } from 'vitest'
+import {
+  createProcessShutdown,
+  PROCESS_SHUTDOWN_TIMEOUT_MS,
+} from '../src/process-shutdown.ts'
+
+function deferred(): { promise: Promise<void>; resolve: () => void; reject: (error: Error) => void } {
+  let resolve!: () => void
+  let reject!: (error: Error) => void
+  const promise = new Promise<void>((accept, fail) => {
+    resolve = accept
+    reject = fail
+  })
+  return { promise, resolve, reject }
+}
+
+afterEach(() => {
+  vi.useRealTimers()
+  vi.restoreAllMocks()
+})
+
+describe('process shutdown', () => {
+  it('exits once after graceful disposal resolves or rejects', async () => {
+    const resolvedExit = vi.fn()
+    const resolved = createProcessShutdown(() => Promise.resolve(), resolvedExit)
+    await resolved.shutdown(0)
+    expect(resolvedExit).toHaveBeenCalledOnce()
+    expect(resolvedExit).toHaveBeenCalledWith(0)
+
+    const rejectedExit = vi.fn()
+    const rejected = createProcessShutdown(() => Promise.reject(new Error('dispose failed')), rejectedExit)
+    await rejected.shutdown(1)
+    expect(rejectedExit).toHaveBeenCalledOnce()
+    expect(rejectedExit).toHaveBeenCalledWith(1)
+  })
+
+  it('uses process.exit as the default process boundary', async () => {
+    const exit = vi.spyOn(process, 'exit').mockImplementation(_code => undefined as never)
+    const shutdown = createProcessShutdown(() => Promise.resolve())
+
+    await shutdown.shutdown(7)
+
+    expect(exit).toHaveBeenCalledOnce()
+    expect(exit).toHaveBeenCalledWith(7)
+  })
+
+  it('forces exit when graceful disposal reaches its bound', async () => {
+    vi.useFakeTimers()
+    const disposal = deferred()
+    const exit = vi.fn()
+    const shutdown = createProcessShutdown(() => disposal.promise, exit)
+    const pending = shutdown.shutdown(0)
+
+    await vi.advanceTimersByTimeAsync(PROCESS_SHUTDOWN_TIMEOUT_MS - 1)
+    expect(exit).not.toHaveBeenCalled()
+    await vi.advanceTimersByTimeAsync(1)
+    expect(exit).toHaveBeenCalledOnce()
+    expect(exit).toHaveBeenCalledWith(0)
+
+    disposal.resolve()
+    await pending
+    expect(exit).toHaveBeenCalledOnce()
+  })
+
+  it('honors a caller-supplied grace period', async () => {
+    vi.useFakeTimers()
+    const disposal = deferred()
+    const exit = vi.fn()
+    const shutdown = createProcessShutdown(() => disposal.promise, exit, 25)
+    const pending = shutdown.shutdown(0)
+
+    await vi.advanceTimersByTimeAsync(24)
+    expect(exit).not.toHaveBeenCalled()
+    await vi.advanceTimersByTimeAsync(1)
+    expect(exit).toHaveBeenCalledOnce()
+
+    disposal.resolve()
+    await pending
+  })
+
+  it('lets Ctrl+C force a normal shutdown already stuck in disposal', async () => {
+    const disposal = deferred()
+    const exit = vi.fn()
+    const shutdown = createProcessShutdown(() => disposal.promise, exit)
+    const pending = shutdown.shutdown(0)
+
+    shutdown.interrupt(130)
+    expect(exit).toHaveBeenCalledOnce()
+    expect(exit).toHaveBeenCalledWith(130)
+
+    disposal.resolve()
+    await pending
+    expect(exit).toHaveBeenCalledOnce()
+  })
+
+  it('drains on the first signal and forces on the second signal', async () => {
+    const disposal = deferred()
+    const dispose = vi.fn(() => disposal.promise)
+    const exit = vi.fn()
+    const shutdown = createProcessShutdown(dispose, exit)
+
+    shutdown.interrupt(143)
+    await Promise.resolve()
+    expect(dispose).toHaveBeenCalledOnce()
+    expect(exit).not.toHaveBeenCalled()
+
+    shutdown.interrupt(130)
+    expect(exit).toHaveBeenCalledOnce()
+    expect(exit).toHaveBeenCalledWith(130)
+
+    disposal.resolve()
+    await shutdown.shutdown(0)
+    expect(exit).toHaveBeenCalledOnce()
+  })
+
+  it('coalesces normal shutdown calls without treating them as escalation', async () => {
+    const disposal = deferred()
+    const exit = vi.fn()
+    const shutdown = createProcessShutdown(() => disposal.promise, exit)
+
+    const first = shutdown.shutdown(0)
+    const second = shutdown.shutdown(1)
+    expect(second).toBe(first)
+    expect(exit).not.toHaveBeenCalled()
+
+    disposal.resolve()
+    await first
+    expect(exit).toHaveBeenCalledOnce()
+    expect(exit).toHaveBeenCalledWith(0)
+  })
+})
@@ -77,12 +77,14 @@ describe('web e2e: approval takeover keeps its actions reachable', () => {
    const input = page.locator('textarea').first()
    await input.waitFor({ timeout: 10_000 })

-    // The composer's own text cap, measured on the live textarea before the
-    // takeover replaces it. The panel's scroll region must stop at the same
-    // height (the designer's requirement: one cap for the composer seat), and
-    // measuring it here keeps the assertion free of the px value itself.
+    // The composer's own text cap, measured on the live draft scrollport before
+    // the takeover replaces it — the box that carries the cap, while the
+    // textarea inside it is as tall as the whole draft. The panel's scroll
+    // region must stop at the same height (the designer's requirement: one cap
+    // for the composer seat), and measuring it here keeps the assertion free of
+    // the px value itself.
    await input.fill(CAP_PROBE)
-    const composerCap = await input.evaluate(el => el.clientHeight)
+    const composerCap = await input.evaluate(el => el.closest('[data-input-scroll]')?.clientHeight ?? 0)
    expect(composerCap).toBeGreaterThan(0)
    await input.fill('')

@@ -1,7 +1,7 @@
 // @vitest-environment jsdom
 // The built-bundle boot smoke: the ONE assembled-jsdom test that loads the
 // real `packages/client/*/lib/client.js` artifacts through AppWebEntry's
-// ModuleLoader path (fetchBundle/executeBundle) and proves the boot graph
+// ModuleLoader path (loadBundle) and proves the boot graph
 // assembles — staged activation across the immediately tier and the inject
 // layers, per-plugin CSS injection, and a rendered journey reaching chat
 // content from the keyless FixtureApiClient transport.
@@ -91,11 +91,11 @@ it('boots the built plugin graph and renders a fixture session end to end', asyn
  win.__DSH_BOOT__ = { rev: 'fx', entries: PLUGINS.map(({ dir: _dir, ...plugin }) => plugin) }
  act(() => {
    const entry = new AppWebEntry(root, {
-      fetchBundle: (url) => {
+      loadBundle: async (url) => {
        const code = bundles.get(url)
-        return code === undefined ? Promise.reject(new Error(`missing built bundle ${url}`)) : Promise.resolve(code)
+        if (code === undefined) throw new Error(`missing built bundle ${url}`)
+        ;(0, eval)(code)
      },
-      executeBundle: (code) => { (0, eval)(code) },
    })
    void entry.run()
    unmount = () => { entry.dispose() }
@@ -608,6 +608,9 @@ describe('web e2e: long Chat scroll contract', () => {
      await world.page.getByRole('tab', { name: 'Trajectory', exact: true }).click()
      await world.page.getByLabel('Trajectory timeline').waitFor({ timeout: 30_000 })
      await world.page.setViewportSize({ width: 700, height: 900 })
+      // The narrow breakpoint auto-collapses the sidebar. Re-open it because
+      // this scenario switches sessions while pinning the narrow Chat scroll owner.
+      await world.page.getByRole('button', { name: 'Open sidebar', exact: true }).click()
      await world.page.getByRole('tab', { name: 'Chat', exact: true }).click()
      await nextPaint(world.page)
      await expectSameFlowTop(world.page, sessionAnchor)
@@ -1,36 +1,31 @@
 // Web e2e scenario: a composer draft longer than the 14-line cap scrolls its
-// GLYPHS, not just its caret.
+// GLYPHS AND ITS CARET AS ONE.
 //
 // The composer paints its text in two stacked layers (see
 // packages/client/ui-conversation/src/client/skeleton/InputBar.module.css): the
 // `<textarea>` carries the value, the selection and the caret but renders its
 // own glyphs `color: transparent`, and every visible character is painted by the
 // `[data-input-backdrop]` div underneath it, which also carries the claim-token
-// highlight, the chips and the ghost hint. The backdrop is `position: absolute;
-// inset: 0; overflow: hidden` — it is CLIPPED, not scrolled, and nothing in the
-// browser links its scroll offset to the textarea's.
+// highlight, the chips and the ghost hint.
 //
-// So past the cap the textarea scrolled and the words did not: the caret walked
-// off the bottom of a block of text frozen at line 1, and no gesture — wheel,
-// drag, arrow key — moved it. `InputBar` now mirrors the offset onto the
-// backdrop on every textarea `scroll`, which is the one event every way of
-// moving the box ends in.
+// Two layers can only stay together by moving together. They now do: both sit
+// inside `[data-input-scroll]`, the composer's single scrolling box, and are as
+// tall as the whole draft — so one offset, applied by the browser, moves the
+// caret and the words in the same frame. Scrolling the textarea and assigning
+// its offset to the backdrop looks equivalent and is not: a wheel gesture is
+// composited off the main thread, so the assignment lands frames late and the
+// caret visibly flies ahead of the text it belongs to.
 //
-// Mirroring an offset is only correct while both layers can reach it, so the
-// geometry underneath is asserted here alongside the visible outcome: the
-// backdrop's trailing-line sentinel (a textarea reserves a line box for the
-// caret after a final newline; `pre-wrap` collapses one), and one wrap width
-// across all three layers (only the textarea scrolls, so only it can lose
-// width to a scrollbar that consumes layout space). Either breaks the extent
-// equality, and an unreachable offset clamps the glyphs below the caret.
+// That failure is what the same-task measurement below pins. Every metric here
+// is read through the caret's own coordinate frame — where the textarea puts
+// line n — against where the backdrop paints line n, because that difference is
+// the defect a user sees, and it is the one number a mirror between two boxes
+// cannot hold at zero.
 //
-// Only a real engine can show this. Scrolling is layout: jsdom reports
+// Only a real engine can show any of this. Scrolling is layout: jsdom reports
 // `scrollHeight === clientHeight` for every element and never scrolls one, so
-// the unit spec in packages/client/ui-conversation/tests/input-bar.spec.tsx has
-// to stub both offsets and can only prove the mirroring code path runs. What is
-// asserted here instead is the user-visible fact that path exists for — after
-// scrolling to the end of a long draft, the LAST line is the one on screen —
-// measured with a DOM Range over the backdrop's own text.
+// the unit spec in packages/client/ui-conversation/tests/input-bar.spec.tsx can
+// only assert that one scrollport contains both layers.
 //
 // Zero model calls: a fresh workspace's blank session already carries a live
 // composer, and the scenario only types into it. A stray stream would fail loud
@@ -49,10 +44,10 @@ import { connectFreshWorkspace, newEnglishPage, saveFailureShot } from './suppor
 const SNAPSHOT_DIR = fileURLToPath(new URL('./snapshots/composer-draft-scroll', import.meta.url))
 /**
 * Committed golden of the composer's two-layer scroll geometry. The change
- * alters no DOM and no accessible name, so the aria goldens the other scenarios
- * commit are byte-identical with and without it; this records the relations
- * instead, which makes a shift in the cap or in the layer coupling a reviewable
- * diff rather than an assertion someone has to reconstruct.
+ * alters no accessible name, so the aria goldens the other scenarios commit are
+ * byte-identical with and without it; this records the relations instead, which
+ * makes a shift in the cap or in the layer coupling a reviewable diff rather
+ * than an assertion someone has to reconstruct.
 */
 const GEOMETRY_EXPECTED = join(SNAPSHOT_DIR, 'geometry.expected.md')
 const MODE = webSnapshotMode()
@@ -69,41 +64,54 @@ const DRAFT = Array.from({ length: DRAFT_LINES }, (_unused, index) => {
 }).join('\n')

 /**
- * A draft ending in a newline: the shape whose layer extents diverge without
- * the backdrop's trailing-line sentinel. A textarea reserves a line box for the
- * caret after a final newline; `white-space: pre-wrap` collapses a text node's
- * trailing newline and generates none, so the backdrop would come out exactly
- * one line shorter and the mirrored offset would clamp a line above the caret.
+ * A draft ending in a newline: the shape where the two layers reserve their
+ * final line box on different terms. A textarea keeps one for the caret after a
+ * final newline; `white-space: pre-wrap` collapses a text node's trailing
+ * newline and generates none. The hidden auto-grow mirror carries the newline
+ * and so decides the height for both, which is why the backdrop needs no
+ * padding of its own — but only a draft of this shape can show it.
 */
 const DRAFT_TRAILING_NEWLINE = `${DRAFT}\n`

-/** The composer's two text layers as the browser lays them out. */
+/** The composer's text layers as the browser lays them out. */
 interface ComposerMetrics {
  /** True when the draft is taller than the capped box — the situation under test. */
  overflows: boolean
-  /** Visible height of the textarea's content box: the cap in pixels. */
+  /** Visible height of the scrollport's content box: the cap in pixels. */
  clientHeight: number
  /** Whole lines that fit in the visible box, at the composer's own line-height. */
  visibleLines: number
-  /** The textarea's scroll offset, which the caret and the selection follow. */
-  inputScrollTop: number
-  /** The backdrop's scroll offset, which every visible glyph follows. */
-  backdropScrollTop: number
-  /** True when the two layers agree — the coupling this scenario exists for. */
-  layersAgree: boolean
+  /** The composer's one scroll offset, which the caret and the glyphs both follow. */
+  scrollTop: number
+  /** Furthest that offset can go. */
+  scrollMax: number
+  /**
+   * Scrollable overflow the textarea holds on its own — 0, or a second offset
+   * exists that nothing keeps equal to this one.
+   */
+  inputScrollable: number
+  /**
+   * Distance between where the caret sits for a draft line and where the
+   * backdrop paints that line, in pixels. A fixed value (the difference between
+   * a line box's top and its glyph box's) is alignment; a value that CHANGES
+   * with the scroll offset is the defect — the words trailing the caret.
+   */
+  caretGlyphGap: number
+  /**
+   * How much that gap moves when the offset changes inside a single task: 0
+   * here, because one box carries both layers. Assigning one box's offset to
+   * another cannot be 0 — a scroll event is dispatched after the task that
+   * moved the box, so between the two there is a frame with the caret at the
+   * new offset and the glyphs at the old one.
+   */
+  gapShiftOnScroll: number
  /**
   * Top of the LAST draft line relative to the visible box's top, in pixels: at
-   * most `clientHeight` when that line is on screen. This is the reported
-   * symptom as a number — with the layers uncoupled the backdrop stays at offset
-   * 0, so the last line sits a full draft-height below the box.
+   * most `clientHeight` when that line is on screen.
   */
  lastLineOffset: number
  /** Top of the FIRST draft line relative to the visible box's top: negative once it has scrolled out. */
  firstLineOffset: number
-  /** Furthest the textarea can scroll. */
-  inputMax: number
-  /** Furthest the backdrop can scroll — equal to `inputMax`, or the mirror clamps below the caret. */
-  backdropMax: number
  /** Content width the textarea wraps at. */
  inputWrapWidth: number
  /** Content width the backdrop wraps at — equal, or the layers break lines in different places. */
@@ -113,14 +121,16 @@ interface ComposerMetrics {
 }

 /**
- * Measure both composer layers in the page.
+ * Measure the composer's layers in the page, in the caret's coordinate frame.
 * @param page - the page under test.
- * @returns the two layers' offsets and where the draft's first and last lines sit.
+ * @returns the offset, the caret-to-glyph gap, and where the draft's first and last lines sit.
 */
 function measureComposer(page: Page): Promise<ComposerMetrics> {
  return page.evaluate(({ first, last }) => {
    const input = document.querySelector<HTMLTextAreaElement>('textarea:enabled')
    if (input === null) throw new Error('no live composer textarea in the DOM')
+    const scroll = input.closest<HTMLElement>('[data-input-scroll]')
+    if (scroll === null) throw new Error('the composer textarea is not inside a draft scrollport')
    const backdrop = input.parentElement?.querySelector<HTMLElement>('[data-input-backdrop]')
    if (backdrop === undefined || backdrop === null) throw new Error('no decoration backdrop beside the composer textarea')
    // The hidden auto-grow mirror: the textarea's next sibling, and the layer
@@ -128,47 +138,50 @@ function measureComposer(page: Page): Promise<ComposerMetrics> {
    // two that carry glyphs.
    const mirror = input.nextElementSibling
    if (!(mirror instanceof HTMLElement)) throw new Error('no auto-grow mirror after the composer textarea')
-    const box = input.getBoundingClientRect()
    // The draft carries no chips or claim token, so the decoration walk emits it
-    // as one text node — the backdrop's first, ahead of the trailing-line
-    // sentinel React renders as a second one. Both markers live in that first
-    // node, which is what the Range below needs.
+    // as a single text node, which is what the Range below needs.
    const text = backdrop.firstChild
    if (!(text instanceof Text)) throw new Error('backdrop does not open with a plain text node')
-    const offsetOf = (marker: string): number => {
+    const lineHeight = Number.parseFloat(getComputedStyle(input).lineHeight)
+    /** Where the backdrop paints the line holding `marker`, in viewport coordinates. */
+    const glyphTop = (marker: string): number => {
      const at = text.data.indexOf(marker)
      if (at < 0) throw new Error(`marker ${marker} missing from the backdrop text`)
      const range = document.createRange()
      range.setStart(text, at)
      range.setEnd(text, at + marker.length)
-      return range.getBoundingClientRect().top - box.top
+      return range.getBoundingClientRect().top
    }
-    const lineHeight = Number.parseFloat(getComputedStyle(input).lineHeight)
-    // Each layer's own maximum, probed by asking for an impossible offset and
-    // reading back what it clamped to, then restored. Reading scrollHeight -
-    // clientHeight instead would compute the maximum rather than observe it.
-    const restore = input.scrollTop
-    const restoreBackdrop = backdrop.scrollTop
-    input.scrollTop = 1e7
-    backdrop.scrollTop = 1e7
-    const inputMax = input.scrollTop
-    const backdropMax = backdrop.scrollTop
-    input.scrollTop = restore
-    backdrop.scrollTop = restoreBackdrop
+    const paddingTop = Number.parseFloat(getComputedStyle(input).paddingTop)
+    // Where the CARET sits on the draft's first line: the textarea lays its own
+    // (transparent) glyphs out from its border box, shifted by any offset it
+    // holds itself. Reading the caret's frame this way rather than the
+    // scrollport's is what makes the gap the user-visible quantity — it stays
+    // honest if the textarea ever starts scrolling on its own again.
+    const gap = (): number =>
+      Math.round(input.getBoundingClientRect().top + paddingTop - input.scrollTop - glyphTop(first))
+    // The same-task probe: move the offset and re-read the gap before the task
+    // ends, which is before any scroll event could have run a listener.
+    const before = gap()
+    const restore = scroll.scrollTop
+    scroll.scrollTop = restore === 0 ? 120 : 0
+    const gapShiftOnScroll = Math.abs(gap() - before)
+    scroll.scrollTop = restore
+    const box = scroll.getBoundingClientRect()
    return {
-      inputMax,
-      backdropMax,
      inputWrapWidth: input.clientWidth,
      backdropWrapWidth: backdrop.clientWidth,
      mirrorWrapWidth: mirror.clientWidth,
-      overflows: input.scrollHeight > input.clientHeight,
-      clientHeight: input.clientHeight,
-      visibleLines: Math.floor(input.clientHeight / lineHeight),
-      inputScrollTop: input.scrollTop,
-      backdropScrollTop: backdrop.scrollTop,
-      layersAgree: input.scrollTop === backdrop.scrollTop,
-      lastLineOffset: offsetOf(last),
-      firstLineOffset: offsetOf(first),
+      overflows: scroll.scrollHeight > scroll.clientHeight,
+      clientHeight: scroll.clientHeight,
+      visibleLines: Math.floor(scroll.clientHeight / lineHeight),
+      scrollTop: scroll.scrollTop,
+      scrollMax: scroll.scrollHeight - scroll.clientHeight,
+      inputScrollable: input.scrollHeight - input.clientHeight,
+      caretGlyphGap: before,
+      gapShiftOnScroll,
+      lastLineOffset: glyphTop(last) - box.top,
+      firstLineOffset: glyphTop(first) - box.top,
    }
  }, { first: FIRST_MARKER, last: LAST_MARKER })
 }
@@ -179,43 +192,56 @@ function measureComposer(page: Page): Promise<ComposerMetrics> {
 * Absolute glyph coordinates are deliberately absent: they depend on font
 * metrics and would make the fixture fail on a machine that measures text
 * differently — a golden that needs re-recording per platform documents the
- * platform, not the change. What is recorded is the cap, the layer agreement,
- * and which lines are on screen, each a comparison that survives any layout
- * keeping the coupling.
+ * platform, not the change. What is recorded is the cap, the caret-to-glyph
+ * relation, and which lines are on screen, each a comparison that survives any
+ * layout keeping the coupling.
 * @param top - metrics with the draft scrolled to its start.
 * @param bottom - metrics with the draft scrolled to its end.
 * @param trailingNewline - metrics with the trailing-newline draft scrolled to its end.
+ * @param pasted - metrics right after a long block was pasted at the draft's end.
 * @returns the golden body, without a trailing newline.
 */
-function renderGeometry(top: ComposerMetrics, bottom: ComposerMetrics, trailingNewline: ComposerMetrics): string {
+function renderGeometry(
+  top: ComposerMetrics, bottom: ComposerMetrics, trailingNewline: ComposerMetrics, pasted: ComposerMetrics,
+): string {
  return [
-    '# Composer draft scrolling (14-line cap, two text layers)',
+    '# Composer draft scrolling (14-line cap, two text layers, one scrollport)',
    '',
    '## At the start of the draft',
    '',
    `- draft overflows the capped box: ${String(top.overflows)}`,
    `- visible lines: ${String(top.visibleLines)}`,
-    `- both layers share one scroll extent: ${String(top.inputMax === top.backdropMax)}`,
+    `- the textarea holds no scroll offset of its own: ${String(top.inputScrollable === 0)}`,
    `- all three layers wrap at one width: ${String(
      top.inputWrapWidth === top.backdropWrapWidth && top.backdropWrapWidth === top.mirrorWrapWidth,
    )}`,
-    `- textarea scroll offset: ${String(top.inputScrollTop)}px`,
-    `- glyph layer tracks it: ${String(top.layersAgree)}`,
+    `- scroll offset: ${String(top.scrollTop)}px`,
+    `- caret and glyphs stay level when the offset changes: ${String(top.gapShiftOnScroll === 0)}`,
    `- first draft line is on screen: ${String(top.firstLineOffset >= 0 && top.firstLineOffset < top.clientHeight)}`,
    `- last draft line is on screen: ${String(top.lastLineOffset >= 0 && top.lastLineOffset < top.clientHeight)}`,
    '',
    '## Scrolled to the end of the draft',
    '',
-    `- textarea moved: ${String(bottom.inputScrollTop > 0)}`,
-    `- glyph layer tracks it: ${String(bottom.layersAgree)}`,
+    `- offset moved: ${String(bottom.scrollTop > 0)}`,
+    `- caret sits on its own glyphs: ${String(bottom.caretGlyphGap === top.caretGlyphGap)}`,
+    `- caret and glyphs stay level when the offset changes: ${String(bottom.gapShiftOnScroll === 0)}`,
    `- first draft line has scrolled out above: ${String(bottom.firstLineOffset < 0)}`,
    `- last draft line is on screen: ${String(bottom.lastLineOffset >= 0 && bottom.lastLineOffset < bottom.clientHeight)}`,
    '',
    '## Draft ending in a newline, scrolled to the end',
    '',
-    `- both layers share one scroll extent: ${String(trailingNewline.inputMax === trailingNewline.backdropMax)}`,
-    `- glyph layer tracks the caret: ${String(trailingNewline.layersAgree)}`,
-    `- last draft line is on screen: ${String(trailingNewline.lastLineOffset >= 0 && trailingNewline.lastLineOffset < trailingNewline.clientHeight)}`,
+    `- caret sits on its own glyphs: ${String(trailingNewline.caretGlyphGap === top.caretGlyphGap)}`,
+    `- the draft's own last line is on screen: ${String(
+      trailingNewline.lastLineOffset >= 0 && trailingNewline.lastLineOffset < trailingNewline.clientHeight,
+    )}`,
+    '',
+    '## Right after pasting a long block at the end',
+    '',
+    `- the composer scrolled to the caret it left: ${String(pasted.scrollTop > 0)}`,
+    `- caret and glyphs stay level when the offset changes: ${String(pasted.gapShiftOnScroll === 0)}`,
+    `- the pasted block's last line is on screen: ${String(
+      pasted.lastLineOffset >= 0 && pasted.lastLineOffset < pasted.clientHeight,
+    )}`,
  ].join('\n').trimEnd()
 }

@@ -251,17 +277,16 @@ describe('web e2e: composer draft scrolling', () => {
    // case below.
    await page.locator('textarea:enabled').first().hover()
    await page.mouse.wheel(0, -2000)
-    await expect.poll(async () => (await measureComposer(page)).inputScrollTop, { timeout: 10_000 }).toBe(0)
+    await expect.poll(async () => (await measureComposer(page)).scrollTop, { timeout: 10_000 }).toBe(0)
    const metrics = await measureComposer(page)
    // The cap is the composer seat's `--dsh-composer-text-max-height` (336px =
    // 14 x 24px lines). The count, not the pixels: it is the figma constant and
    // survives a device-pixel-ratio change.
    expect(metrics.visibleLines).toBe(14)
-    // Resting state: the draft's head is what a 40-line draft shows, and its
-    // tail is far below the box. Both layers sit at the origin, which is why the
-    // uncoupled build looks correct until something scrolls.
-    expect(metrics.inputScrollTop).toBe(0)
-    expect(metrics.layersAgree).toBe(true)
+    // One scrolling box: the textarea is as tall as the draft, so there is no
+    // second offset for the caret to hold while the glyphs hold another.
+    expect(metrics.inputScrollable).toBe(0)
+    expect(metrics.scrollTop).toBe(0)
    expect(metrics.firstLineOffset).toBeGreaterThanOrEqual(0)
    expect(metrics.firstLineOffset).toBeLessThan(metrics.clientHeight)
    expect(metrics.lastLineOffset).toBeGreaterThan(metrics.clientHeight)
@@ -270,19 +295,12 @@ describe('web e2e: composer draft scrolling', () => {

  it('lays out all three text layers at one wrap width', async () => {
    onTestFailed(() => saveFailureShot(page, 'web-e2e-composer-draft-scroll-wrap-width'))
-    // The premise under the mirror, asserted rather than assumed. Only .input
-    // scrolls, so only .input can lose content width to a scrollbar that
-    // consumes layout space; a narrower .input wraps a long draft onto more
-    // lines, ends up taller, and its larger maximum makes the mirrored offset
-    // clamp below the caret. Measured on a standalone harness, an 8px width
-    // difference is worth 2 to 5 lines on a wrap-sensitive draft.
-    //
-    // This holds on the lane's engine and is what a regression would break —
-    // it is NOT vacuous: measured on the same app, WebKit reports 768 against
-    // 776 here, which is the divergence the Agent Note records as a
-    // pre-existing, engine-specific limitation. The mirror is unaffected there
-    // today because the extents still agree; this assertion is what would
-    // notice if the lane's engine ever moved into the same state.
+    // A layer that breaks lines somewhere else puts the words under the wrong
+    // caret, and an 8px difference is worth 2 to 5 lines on a wrap-sensitive
+    // draft. The three now share a containing block — the scrollport — so a
+    // scrollbar that consumes layout space costs them the same width; before,
+    // only the textarea scrolled, and WebKit reserved gutter space for it alone
+    // (768 against 776) while chromium and firefox did not.
    const metrics = await measureComposer(page)
    expect(metrics.backdropWrapWidth).toBe(metrics.inputWrapWidth)
    // The mirror decides the box height, so it belongs in the same equality —
@@ -292,66 +310,112 @@ describe('web e2e: composer draft scrolling', () => {
    expect(tripwire.pageErrors).toEqual([])
  }, 60_000)

+  it('the glyphs cannot lag the caret: one task moves both', async () => {
+    onTestFailed(() => saveFailureShot(page, 'web-e2e-composer-draft-scroll-lag'))
+    // The reported symptom, isolated. A scroll offset changes and the caret's
+    // distance to its own glyphs is re-read before the task ends — before any
+    // `scroll` listener could have run. With the layers on one scrollport the
+    // browser moved both, so the distance is unchanged; with the glyph layer
+    // catching up in a listener it is off by the whole delta until a later
+    // frame, which is a caret flying away from its text mid-gesture.
+    const metrics = await measureComposer(page)
+    expect(metrics.gapShiftOnScroll).toBe(0)
+    expect(tripwire.pageErrors).toEqual([])
+  }, 60_000)
+
  it('a wheel gesture over a long draft moves the words, not only the caret', async () => {
    onTestFailed(() => saveFailureShot(page, 'web-e2e-composer-draft-scroll-wheel'))
    const input = page.locator('textarea:enabled').first()
    await input.hover()
-    // One delta past the whole draft: the textarea clamps at its own end, and
-    // the wheel-chaining handler leaves it native because the box is not yet at
-    // its edge when the gesture starts (the chaining itself is owned by the
-    // unit spec).
+    const resting = (await measureComposer(page)).caretGlyphGap
+    // One delta past the whole draft: the box clamps at its own end, and the
+    // wheel-chaining handler leaves it native because the box is not yet at its
+    // edge when the gesture starts (the chaining itself is owned by the unit spec).
    await page.mouse.wheel(0, 2000)
-    await expect.poll(async () => (await measureComposer(page)).inputScrollTop, { timeout: 10_000 })
+    await expect.poll(async () => (await measureComposer(page)).scrollTop, { timeout: 10_000 })
      .toBeGreaterThan(0)
    const metrics = await measureComposer(page)
-    // The coupling, stated directly.
-    expect(metrics.layersAgree).toBe(true)
+    // The caret is still on its own glyphs after the gesture.
+    expect(metrics.caretGlyphGap).toBe(resting)
    // The reported symptom, stated as what the user sees: the end of the draft
-    // is on screen and its beginning is not. On the uncoupled build the glyph
-    // layer stays at offset 0, so `lastLineOffset` is still a full draft below
-    // the box and `firstLineOffset` is still 0 — the text never moved.
+    // is on screen and its beginning is not.
    expect(metrics.lastLineOffset).toBeGreaterThanOrEqual(0)
    expect(metrics.lastLineOffset).toBeLessThan(metrics.clientHeight)
    expect(metrics.firstLineOffset).toBeLessThan(0)
    expect(tripwire.pageErrors).toEqual([])
  }, 60_000)

-  it('typing at the end of a scrolled draft keeps the layers together', async () => {
+  it('typing at the end of a scrolled draft brings the caret back into view', async () => {
    onTestFailed(() => saveFailureShot(page, 'web-e2e-composer-draft-scroll-edit'))
-    // The other way the box moves. Typing at the caret — parked at the draft's
-    // end by the wheel gesture — scrolls it into view, which is a `scroll` like
-    // any other; this pins that an edit is not a separate case needing its own
-    // mirror, which is why one listener is the whole implementation.
+    // The other way the box moves, and the one that depends on the browser: the
+    // textarea no longer scrolls, so revealing the caret after an edit is a
+    // scroll-into-view that has to walk up to the scrollport. Scroll away from
+    // the caret first, so the edit has somewhere to bring it back from.
    const input = page.locator('textarea:enabled').first()
    await input.press('End')
+    await input.hover()
+    await page.mouse.wheel(0, -2000)
+    await expect.poll(async () => (await measureComposer(page)).scrollTop, { timeout: 10_000 }).toBe(0)
    await input.pressSequentially(' tail')
    const metrics = await measureComposer(page)
-    expect(metrics.layersAgree).toBe(true)
+    expect(metrics.scrollTop).toBeGreaterThan(0)
    expect(metrics.lastLineOffset).toBeGreaterThanOrEqual(0)
    expect(metrics.lastLineOffset).toBeLessThan(metrics.clientHeight)
    expect(tripwire.pageErrors).toEqual([])
  }, 60_000)

+  it('pasting a long block scrolls to the caret it leaves at the end', async () => {
+    onTestFailed(() => saveFailureShot(page, 'web-e2e-composer-draft-scroll-paste'))
+    // The composer suppresses the native paste — the machine owns the draft and
+    // the undo log — and restores the caret programmatically, which reveals
+    // nothing on its own: measured in chromium and WebKit, the view stayed
+    // where it was while the caret sat at the end of the pasted block. The
+    // restore now scrolls it into view, and this is the case that proves it.
+    const input = page.locator('textarea:enabled').first()
+    await input.fill('one short line')
+    await input.press('End')
+    // A real `paste` event carrying real clipboard data, dispatched at the
+    // textarea: the same event a Cmd-V delivers, and it runs the same handler.
+    await input.evaluate((el, text) => {
+      const data = new DataTransfer()
+      data.setData('text/plain', text)
+      el.dispatchEvent(new ClipboardEvent('paste', { clipboardData: data, bubbles: true, cancelable: true }))
+      // Ending in a newline is the shape the engines disagree on: the caret
+      // lands on a line with nothing on it, where chromium reports no client
+      // rects at all for the collapsed position.
+    }, `\n${DRAFT}\n`)
+    await expect.poll(async () => (await measureComposer(page)).overflows, { timeout: 10_000 }).toBe(true)
+    // The restore lands one frame after the machine commits the draft, so the
+    // box overflows before it moves; waiting on the offset is waiting for the
+    // behavior itself, and its absence fails this poll.
+    await expect.poll(async () => (await measureComposer(page)).scrollTop, { timeout: 10_000 }).toBeGreaterThan(0)
+    const metrics = await measureComposer(page)
+    // The caret is at the end of what was pasted, so the draft's last line is
+    // what has to be on screen.
+    expect(metrics.lastLineOffset).toBeGreaterThanOrEqual(0)
+    expect(metrics.lastLineOffset).toBeLessThan(metrics.clientHeight)
+    expect(metrics.gapShiftOnScroll).toBe(0)
+    expect(tripwire.pageErrors).toEqual([])
+  }, 60_000)
+
  it('a draft ending in a newline scrolls to its true end, not a line above it', async () => {
    onTestFailed(() => saveFailureShot(page, 'web-e2e-composer-draft-scroll-trailing-newline'))
    // The layers reserve a final line box on different terms, so this shape is
-    // the one that separates equal extents from a mirror that clamps early.
+    // the one that separates a height every layer agrees on from a box measured
+    // one line short of the caret's own last position.
    const input = page.locator('textarea:enabled').first()
    await input.fill(DRAFT_TRAILING_NEWLINE)
    await expect.poll(async () => (await measureComposer(page)).overflows, { timeout: 10_000 }).toBe(true)
-    const extents = await measureComposer(page)
-    // The invariant the sentinel exists for. Without it the textarea measured
-    // 652 against the backdrop's 628 — one 24px line apart.
-    expect(extents.backdropMax).toBe(extents.inputMax)
    await input.hover()
    await page.mouse.wheel(0, 4000)
    await expect.poll(async () => {
      const m = await measureComposer(page)
-      return m.inputScrollTop === m.inputMax
+      return m.scrollTop === m.scrollMax
    }, { timeout: 10_000 }).toBe(true)
    const bottom = await measureComposer(page)
-    // At the very bottom the glyphs are level with the caret, not a line behind.
-    expect(bottom.layersAgree).toBe(true)
+    // At the very bottom the glyphs are level with the caret, and the draft's
+    // own last line — the one before the empty final line — is on screen.
+    expect(bottom.gapShiftOnScroll).toBe(0)
    expect(bottom.lastLineOffset).toBeGreaterThanOrEqual(0)
    expect(bottom.lastLineOffset).toBeLessThan(bottom.clientHeight)
    expect(tripwire.pageErrors).toEqual([])
@@ -365,11 +429,11 @@ describe('web e2e: composer draft scrolling', () => {
    await input.fill(DRAFT)
    await input.hover()
    await page.mouse.wheel(0, -2000)
-    await expect.poll(async () => (await measureComposer(page)).inputScrollTop, { timeout: 10_000 }).toBe(0)
+    await expect.poll(async () => (await measureComposer(page)).scrollTop, { timeout: 10_000 }).toBe(0)
    const top = await measureComposer(page)
    await input.hover()
    await page.mouse.wheel(0, 2000)
-    await expect.poll(async () => (await measureComposer(page)).inputScrollTop, { timeout: 10_000 })
+    await expect.poll(async () => (await measureComposer(page)).scrollTop, { timeout: 10_000 })
      .toBeGreaterThan(0)
    const bottom = await measureComposer(page)
    await input.fill(DRAFT_TRAILING_NEWLINE)
@@ -377,10 +441,25 @@ describe('web e2e: composer draft scrolling', () => {
    await page.mouse.wheel(0, 4000)
    await expect.poll(async () => {
      const m = await measureComposer(page)
-      return m.inputScrollTop === m.inputMax
+      return m.scrollTop === m.scrollMax
    }, { timeout: 10_000 }).toBe(true)
    const trailingNewline = await measureComposer(page)
-    await compareOrRefreshGolden(GEOMETRY_EXPECTED, renderGeometry(top, bottom, trailingNewline), MODE)
+    // The paste path, measured the way a user meets it: a short draft, the
+    // caret at its end, one long block pasted in.
+    await input.fill('one short line')
+    await input.press('End')
+    await input.evaluate((el, text) => {
+      const data = new DataTransfer()
+      data.setData('text/plain', text)
+      el.dispatchEvent(new ClipboardEvent('paste', { clipboardData: data, bubbles: true, cancelable: true }))
+      // The ordinary shape — not ending in a newline — so the collapsed branch
+      // of the reveal keeps a real engine under it; the case above owns the
+      // after-newline branch.
+    }, `\n${DRAFT}`)
+    await expect.poll(async () => (await measureComposer(page)).overflows, { timeout: 10_000 }).toBe(true)
+    await expect.poll(async () => (await measureComposer(page)).scrollTop, { timeout: 10_000 }).toBeGreaterThan(0)
+    const pasted = await measureComposer(page)
+    await compareOrRefreshGolden(GEOMETRY_EXPECTED, renderGeometry(top, bottom, trailingNewline, pasted), MODE)
    expect(tripwire.pageErrors).toEqual([])
  }, 60_000)

@@ -58,7 +58,7 @@ describe('web e2e: Models settings page configures a dormant provider', () => {
    await dialog.getByText('填入各提供方的 API 密钥即可使用其模型。').waitFor({ timeout: 10_000 })
    // The dormant pi-ai adapter contributes its whole installed catalog; no
    // provider is configured yet, so the page is one add button.
-    const add = dialog.getByRole('button', { name: '+ 添加提供方' })
+    const add = dialog.getByRole('button', { name: '添加提供方' })
    await add.waitFor({ timeout: 10_000 })
    // The button enables once the dormant catalog lands in the join.
    await expect.poll(async () => add.isEnabled(), { timeout: 10_000 }).toBe(true)
@@ -128,11 +128,11 @@ describe('assembled search card', () => {
    win.__DSH_BOOT__ = { rev: 'fx', entries: PLUGINS.map(({ dir: _dir, ...plugin }) => plugin) }
    act(() => {
      const entry = new AppWebEntry(root, {
-        fetchBundle: (url) => {
+        loadBundle: async (url) => {
          const code = bundles.get(url)
-          return code === undefined ? Promise.reject(new Error(`missing built bundle ${url}`)) : Promise.resolve(code)
+          if (code === undefined) throw new Error(`missing built bundle ${url}`)
+          ;(0, eval)(code)
        },
-        executeBundle: (code) => { (0, eval)(code) },
      })
      void entry.run()
      unmount = () => { entry.dispose() }
@@ -44,6 +44,12 @@
 // overlap, and `timeCoveredBy` measures it at 7. Each was mutation-checked with
 // the other assertions in its test silenced.
 //
+// The thumb is a pointer affordance (ui-sidebar rebinds the indirection pair
+// to `transparent` while the pointer is outside the column), so every
+// measurement below states which pointer position it was taken at: the
+// scenario parks the pointer over the sidebar before asserting a colour, and
+// the quiet state and its linger get their own test.
+//
 // Chromium also takes the `::-webkit-scrollbar*` path, not the standard
 // properties: scrollbar.css gates `scrollbar-width`/`scrollbar-color` behind
 // `@supports not selector(::-webkit-scrollbar)`, which is false here. The
@@ -106,6 +112,10 @@ interface ListMetrics {
  overflows: boolean
  /** Border-box width minus client width: the space the scrollbar takes out of the content area. */
  band: number
+  /** Distance from the scrollbar's right edge to the sidebar edge. */
+  scrollbarEdgeOffset: number
+  /** Distance from the first row background's right edge to the sidebar edge. */
+  rowEdgeInset: number
  /** Client-area right edge in viewport coordinates (`clientWidth` excludes the scrollbar band). */
  clientRight: number
  /** Border-box right edge in viewport coordinates. */
@@ -133,6 +143,8 @@ function measureList(page: Page): Promise<ListMetrics> {
    if (list === null) throw new Error('sidebar session list not in the DOM')
    const time = list.querySelector<HTMLElement>('[class*="time"]')
    if (time === null) throw new Error('no row relative-time element in the sidebar list')
+    const row = list.querySelector<HTMLElement>('[role="treeitem"]')
+    if (row === null) throw new Error('no row in the sidebar list')
    // Each indirection variable is resolved through its own throwaway probe
    // appended to the list: `var()` substitution then happens where the list
    // sits in the cascade, which is the claim, and `color` normalizes whatever
@@ -167,6 +179,9 @@ function measureList(page: Page): Promise<ListMetrics> {
    const style = getComputedStyle(list)
    const pseudoWidth = getComputedStyle(list, '::-webkit-scrollbar').width
    const barWidth = pseudoWidth === 'auto' ? 15 : Number.parseFloat(pseudoWidth)
+    const listRect = list.getBoundingClientRect()
+    const sidebarEdge = list.parentElement?.getBoundingClientRect().right
+    if (sidebarEdge === undefined) throw new Error('sidebar session list has no layout parent')
    return {
      gutter: style.scrollbarGutter,
      width: pseudoWidth,
@@ -177,9 +192,11 @@ function measureList(page: Page): Promise<ListMetrics> {
      token: resolve('--dsh-scrollbar-thumb'),
      hoverToken: resolve('--dsh-scrollbar-thumb-hover'),
      overflows: list.scrollHeight > list.clientHeight,
-      band: list.getBoundingClientRect().width - list.clientWidth,
-      clientRight: list.getBoundingClientRect().left + list.clientWidth,
-      borderRight: list.getBoundingClientRect().right,
+      band: listRect.width - list.clientWidth,
+      scrollbarEdgeOffset: sidebarEdge - listRect.right,
+      rowEdgeInset: sidebarEdge - row.getBoundingClientRect().right,
+      clientRight: listRect.left + list.clientWidth,
+      borderRight: listRect.right,
      timeRight: time.getBoundingClientRect().right,
      // The bar is drawn in the rightmost `barWidth` of the border box, whether
      // or not that space was reserved. Its width comes from the sheet where the
@@ -188,11 +205,60 @@ function measureList(page: Page): Promise<ListMetrics> {
      // absent. Taking the UA width as the fallback is what keeps the assertion
      // honest: assuming 0 there would report no occlusion precisely in the
      // state that has it.
-      timeCoveredBy: Math.max(0, time.getBoundingClientRect().right - (list.getBoundingClientRect().right - barWidth)),
+      timeCoveredBy: Math.max(0, time.getBoundingClientRect().right - (listRect.right - barWidth)),
    }
  })
 }

+/**
+ * Measure only overflow and row inset, which remain observable when every
+ * session is hidden under a collapsed workspace group.
+ * @param page - the page under test.
+ * @returns the list overflow state and first row's trailing inset.
+ */
+function measureRowInset(page: Page): Promise<Pick<ListMetrics, 'overflows' | 'rowEdgeInset'>> {
+  return page.evaluate(() => {
+    const list = document.querySelector<HTMLElement>('[role="tree"][aria-label="Sessions"]')
+    if (list === null) throw new Error('sidebar session list not in the DOM')
+    const row = list.querySelector<HTMLElement>('[role="treeitem"]')
+    if (row === null) throw new Error('no row in the sidebar list')
+    const sidebarEdge = list.parentElement?.getBoundingClientRect().right
+    if (sidebarEdge === undefined) throw new Error('sidebar session list has no layout parent')
+    return {
+      overflows: list.scrollHeight > list.clientHeight,
+      rowEdgeInset: sidebarEdge - row.getBoundingClientRect().right,
+    }
+  })
+}
+
+/** One palette's readings, taken at both pointer positions. */
+interface PaletteMetrics {
+  /** Everything measured with the pointer over the list, which is when a thumb exists. */
+  hovered: ListMetrics
+  /** `--dsh-scrollbar-thumb` with the pointer parked outside the column. */
+  quietThumb: string
+}
+
+/**
+ * Read one palette at both pointer positions, ending with the pointer back
+ * over the list so a caller measuring further leaves it revealed.
+ * @param page - the page under test.
+ * @returns the palette's quiet thumb and its hovered metrics.
+ */
+async function measurePalette(page: Page): Promise<PaletteMetrics> {
+  await pointAt(page, 'away')
+  // Poll rather than sleep the linger out: the wait is the column's, and a
+  // fixed sleep would either race it or pad every palette.
+  await expect.poll(async () => resolveThumb(page), { timeout: 10_000 }).toBe(NO_THUMB)
+  const quietThumb = await resolveThumb(page)
+  await pointAt(page, 'list')
+  // Poll the reveal too: the reading below is a colour, and taking it in the
+  // same tick as the pointer move would race React's flush and land a
+  // transparent thumb in the golden.
+  await expect.poll(async () => resolveThumb(page), { timeout: 10_000 }).not.toBe(NO_THUMB)
+  return { hovered: await measureList(page), quietThumb }
+}
+
 /**
 * Render the golden body: the resolved scrollbar style of the list in each
 * palette, plus the geometric relations the fix establishes.
@@ -208,20 +274,23 @@ function measureList(page: Page): Promise<ListMetrics> {
 * @param dark - metrics measured under the dark palette.
 * @returns the golden body, without a trailing newline.
 */
-function renderGeometry(light: ListMetrics, dark: ListMetrics): string {
-  const palette = (name: string, metrics: ListMetrics): string[] => [
+function renderGeometry(light: PaletteMetrics, dark: PaletteMetrics): string {
+  const palette = (name: string, { hovered: metrics, quietThumb }: PaletteMetrics): string[] => [
    `## ${name}`,
    '',
+    `- --dsh-scrollbar-thumb, pointer outside the sidebar: ${quietThumb}`,
    `- scrollbar-gutter: ${metrics.gutter}`,
    `- ::-webkit-scrollbar width: ${metrics.width}`,
    `- ::-webkit-scrollbar-track background: ${metrics.track}`,
    `- scrollbar-width: ${metrics.standardWidth}`,
    `- scrollbar-color: ${metrics.standardColor}`,
    `- ::-webkit-scrollbar-thumb:hover declarations: ${metrics.hoverRules.join(' | ')}`,
-    `- --dsh-scrollbar-thumb: ${metrics.token}`,
-    `- --dsh-scrollbar-thumb-hover: ${metrics.hoverToken}`,
+    `- --dsh-scrollbar-thumb, pointer over the list: ${metrics.token}`,
+    `- --dsh-scrollbar-thumb-hover, pointer over the list: ${metrics.hoverToken}`,
    `- list overflows: ${String(metrics.overflows)}`,
    `- reserved band: ${String(metrics.band)}px`,
+    `- scrollbar inset from the sidebar edge: ${String(metrics.scrollbarEdgeOffset)}px`,
+    `- row background inset from the sidebar edge: ${String(metrics.rowEdgeInset)}px`,
    `- relative time covered by the bar: ${String(metrics.timeCoveredBy)}px`,
    `- relative time ends inside the content area: ${String(metrics.timeRight <= metrics.clientRight)}`,
    `- content area ends before the border box: ${String(metrics.clientRight < metrics.borderRight)}`,
@@ -235,6 +304,48 @@ function renderGeometry(light: ListMetrics, dark: ListMetrics): string {
  ].join('\n').trimEnd()
 }

+/**
+ * Resolve `--dsh-scrollbar-thumb` as the list sees it, without the rest of the
+ * geometry. Own probe element for the same reason {@link measureList} uses
+ * one: `getComputedStyle` returns a live declaration.
+ * @param page - the page under test.
+ * @returns the resolved thumb colour, serialized as `rgb`/`rgba`.
+ */
+function resolveThumb(page: Page): Promise<string> {
+  return page.evaluate(() => {
+    const list = document.querySelector<HTMLElement>('[role="tree"][aria-label="Sessions"]')
+    if (list === null) throw new Error('sidebar session list not in the DOM')
+    const probe = document.createElement('span')
+    probe.style.color = 'var(--dsh-scrollbar-thumb)'
+    list.append(probe)
+    const value = getComputedStyle(probe).color
+    probe.remove()
+    return value
+  })
+}
+
+/** Fully transparent, which is how the quiet column spells "no thumb". */
+const NO_THUMB = 'rgba(0, 0, 0, 0)'
+
+/**
+ * Park the pointer over the session list or outside the sidebar entirely. The
+ * column reveals its scrollbars from real pointer movement, so a scenario that
+ * never moves the mouse measures the quiet state whatever it intended to.
+ * @param page - the page under test.
+ * @param where - `list` to point at the session list, `away` for the far side
+ * of the viewport (the conversation column).
+ */
+async function pointAt(page: Page, where: 'list' | 'away'): Promise<void> {
+  const box = await page.locator('[role="tree"][aria-label="Sessions"]').boundingBox()
+  if (box === null) throw new Error('sidebar session list has no layout box')
+  const viewport = page.viewportSize()
+  if (viewport === null) throw new Error('page has no viewport')
+  const target = where === 'list'
+    ? { x: box.x + box.width / 2, y: box.y + box.height / 2 }
+    : { x: viewport.width - 5, y: box.y + box.height / 2 }
+  await page.mouse.move(target.x, target.y)
+}
+
 /**
 * Reveal the seeded rows: every seeded session is unattached, so they all sit
 * in the collapsed Ungrouped bucket. Converges on expanded rather than
@@ -280,6 +391,10 @@ describe('web e2e: sidebar session list scrollbar (reserved gutter / themed thum
    await page.goto(scaffold.baseUrl, { waitUntil: 'load' })
    await page.waitForSelector('[class*="frame"]', { timeout: 30_000 })
    await expandSeededSessions(page)
+    // Every assertion about a thumb colour needs a drawn thumb, and the column
+    // only draws one under the pointer; the quiet state is asserted where it is
+    // the subject rather than left as an ambient condition of the whole file.
+    await pointAt(page, 'list')
  }, 180_000)

  afterAll(async () => {
@@ -299,6 +414,8 @@ describe('web e2e: sidebar session list scrollbar (reserved gutter / themed thum
    // drawn over it. Removing the declaration makes it exactly 0. The value
    // itself is not pinned — it tracks `scrollbar-width` and the platform.
    expect(metrics.band).toBeGreaterThan(0)
+    expect(metrics.scrollbarEdgeOffset).toBe(2)
+    expect(metrics.rowEdgeInset).toBe(12)
    // The reported symptom, stated directly: no part of the row's relative time
    // lies under the bar. Measures 7 on clean master — the `h` of `1h` is the
    // covered part. Unlike the client-edge comparison below it does not go
@@ -317,6 +434,48 @@ describe('web e2e: sidebar session list scrollbar (reserved gutter / themed thum
    expect(tripwire.pageErrors).toEqual([])
  }, 60_000)

+  it('draws no thumb until the pointer is over the column, and lingers on the way out', async () => {
+    onTestFailed(() => saveFailureShot(page, 'web-e2e-sidebar-scrollbar-pointer'))
+    const revealed = await resolveThumb(page)
+    expect(revealed).not.toBe(NO_THUMB)
+    await pointAt(page, 'away')
+    // The linger, measured as a state rather than a duration: the thumb is
+    // still drawn on the leave itself, and gone once the window has passed. A
+    // tighter timing assertion would pin the wall clock of a CI machine.
+    expect(await resolveThumb(page)).toBe(revealed)
+    await expect.poll(async () => resolveThumb(page), { timeout: 10_000 }).toBe(NO_THUMB)
+    // The reservation is unconditional, so nothing moved while the bar was
+    // hidden — this is what buys `transparent` over hiding the bar itself.
+    const quiet = await measureList(page)
+    expect(quiet.gutter).toBe('stable')
+    expect(quiet.band).toBeGreaterThan(0)
+    expect(quiet.timeCoveredBy).toBe(0)
+    // Scrolling without a pointer — what a keyboard or a touch drag does —
+    // leaves the column quiet. This is the change's one deliberate loss, and
+    // it is pinned here rather than only described, so making a scroll
+    // re-reveal the bar has to be a decision rather than a side effect.
+    await page.locator('[role="tree"][aria-label="Sessions"]').evaluate((el) => { el.scrollTop += 200 })
+    await page.waitForTimeout(500)
+    expect(await resolveThumb(page)).toBe(NO_THUMB)
+    await pointAt(page, 'list')
+    await expect.poll(async () => resolveThumb(page), { timeout: 10_000 }).toBe(revealed)
+    expect(tripwire.pageErrors).toEqual([])
+  }, 60_000)
+
+  it('keeps the row background inset when overflow disappears', async () => {
+    onTestFailed(() => saveFailureShot(page, 'web-e2e-sidebar-scrollbar-stable-inset'))
+    expect(await measureRowInset(page)).toEqual({ overflows: true, rowEdgeInset: 12 })
+    const bucket = page.getByText('Ungrouped', { exact: true }).locator('..').locator('..')
+    await bucket.click()
+    try {
+      await expect.poll(async () => (await measureRowInset(page)).overflows, { timeout: 10_000 }).toBe(false)
+      expect(await measureRowInset(page)).toEqual({ overflows: false, rowEdgeInset: 12 })
+    } finally {
+      await expandSeededSessions(page)
+    }
+    expect(tripwire.pageErrors).toEqual([])
+  }, 60_000)
+
  it('renders the themed thumb through the WebKit path in both palettes', async () => {
    onTestFailed(() => saveFailureShot(page, 'web-e2e-sidebar-scrollbar-theme'))
    const light = await measureList(page)
@@ -354,9 +513,9 @@ describe('web e2e: sidebar session list scrollbar (reserved gutter / themed thum

  it('matches the committed scrollbar geometry golden in both palettes', async () => {
    onTestFailed(() => saveFailureShot(page, 'web-e2e-sidebar-scrollbar-golden'))
-    const light = await measureList(page)
+    const light = await measurePalette(page)
    await page.evaluate(() => { document.body.setAttribute('data-ds-dark-theme', '') })
-    const dark = await measureList(page)
+    const dark = await measurePalette(page)
    await page.evaluate(() => { document.body.removeAttribute('data-ds-dark-theme') })
    await compareOrRefreshGolden(GEOMETRY_EXPECTED, renderGeometry(light, dark), MODE)
    expect(tripwire.pageErrors).toEqual([])
@@ -36,7 +36,7 @@
  - img
 - button "Branch into a new conversation":
  - img
- text: {{clock}}
+- text: {{clock}}Ran for {{duration}}
 - textbox "Message the agent"
 - button "Commands":
  - img
@@ -1,25 +1,31 @@
-# Composer draft scrolling (14-line cap, two text layers)
+# Composer draft scrolling (14-line cap, two text layers, one scrollport)

 ## At the start of the draft

 - draft overflows the capped box: true
 - visible lines: 14
- both layers share one scroll extent: true
+- the textarea holds no scroll offset of its own: true
 - all three layers wrap at one width: true
- textarea scroll offset: 0px
- glyph layer tracks it: true
+- scroll offset: 0px
+- caret and glyphs stay level when the offset changes: true
 - first draft line is on screen: true
 - last draft line is on screen: false

 ## Scrolled to the end of the draft

- textarea moved: true
- glyph layer tracks it: true
+- offset moved: true
+- caret sits on its own glyphs: true
+- caret and glyphs stay level when the offset changes: true
 - first draft line has scrolled out above: true
 - last draft line is on screen: true

 ## Draft ending in a newline, scrolled to the end

- both layers share one scroll extent: true
- glyph layer tracks the caret: true
- last draft line is on screen: true
+- caret sits on its own glyphs: true
+- the draft's own last line is on screen: true
+
+## Right after pasting a long block at the end
+
+- the composer scrolled to the caret it left: true
+- caret and glyphs stay level when the offset changes: true
+- the pasted block's last line is on screen: true
@@ -51,7 +51,7 @@
  - img
 - button "Branch into a new conversation":
  - img
- text: {{clock}}
+- text: {{clock}}Ran for {{duration}}
 - textbox "Message the agent"
 - button "Commands":
  - img
@@ -31,7 +31,7 @@
  - img
 - button "Branch into a new conversation":
  - img
- text: {{clock}}
+- text: {{clock}}Ran for {{duration}}
 - textbox "Message the agent"
 - button "Commands":
  - img
@@ -28,6 +28,7 @@
 - textbox "Describe what you want to build"
 - button "Commands":
  - img
+- tooltip "Commands"
 - 'button "Access mode, current: Workspace Write"': Workspace Write
 - button "Select model, current DeepSeek-V4-Flash":
  - text: DeepSeek-V4-Flash
@@ -23,7 +23,7 @@
  - img
 - button "Branch into a new conversation":
  - img
- text: {{clock}}
+- text: {{clock}}Ran for {{duration}}
 - textbox "Message the agent"
 - button "Commands":
  - img
@@ -20,7 +20,7 @@
  - img
 - button "Branch into a new conversation":
  - img
- text: {{clock}}
+- text: {{clock}}Ran for {{duration}}
 - textbox "Message the agent"
 - button "Commands":
  - img
@@ -25,7 +25,7 @@
  - img
 - button "Branch into a new conversation":
  - img
- text: {{clock}}
+- text: {{clock}}Ran for {{duration}}
 - textbox "Message the agent"
 - button "Commands":
  - img
@@ -19,7 +19,7 @@
  - img
 - button "Branch into a new conversation":
  - img
- text: {{clock}}
+- text: {{clock}}Ran for {{duration}}
 - textbox "Message the agent"
 - button "Commands":
  - img
@@ -20,7 +20,7 @@
  - img
 - button "Branch into a new conversation" [disabled]:
  - img
- text: Available only on the last message of a completed turn 7/25 {{clock}}
+- text: Available only on the last message of a completed turn 7/25 {{clock}}Ran for {{duration}}
 - button "Read a.txt":
  - img
  - img
@@ -46,7 +46,7 @@
  - img
 - button "Branch into a new conversation":
  - img
- text: 7/25 {{clock}}
+- text: 7/25 {{clock}}Ran for {{duration}}
 - textbox "Message the agent"
 - button "Commands":
  - img
@@ -17,4 +17,6 @@
      - text: minimax-cn
      - button "编辑"
      - button "删除"
-  - button "+ 添加提供方"
+  - button "添加提供方":
+    - img
+    - text: 添加提供方
@@ -36,7 +36,7 @@
  - img
 - button "Branch into a new conversation":
  - img
- text: {{clock}}
+- text: {{clock}}Ran for {{duration}}
 - textbox "Message the agent"
 - button "Commands":
  - img
@@ -31,7 +31,7 @@
  - img
 - button "Branch into a new conversation":
  - img
- text: {{clock}}
+- text: {{clock}}Ran for {{duration}}
 - textbox "Message the agent"
 - button "Commands":
  - img
@@ -30,6 +30,7 @@
    - textbox "Edit queued message": Edited queue item
    - button "Save queued message":
      - img
+    - tooltip "Save queued message"
    - button "Cancel editing":
      - img
 - textbox "Message the agent"
@@ -23,7 +23,7 @@
 - paragraph: partial
 - status: Deep diving...
 - region "To-dos":
-  - button "To-dos 1/2 tasks · 1 in progress"
+  - button "To-dos 1 completed · 1 in progress"
 - img
 - text: Ongoing Goal Keep the composer context panels aligned
 - button "Pause goal":
@@ -20,7 +20,7 @@
  - img
 - button "Branch into a new conversation":
  - img
- text: {{clock}} Edited queue item {{clock}}
+- text: {{clock}}Ran for {{duration}} Edited queue item {{clock}}
 - button "Copy":
  - img
 - button "Branch into a new conversation" [disabled]:
@@ -21,6 +21,7 @@
    - text: Edited queue item
    - button "Edit queued message":
      - img
+    - tooltip "Edit queued message"
    - button "Remove queued message":
      - img
    - button "Steer queued message":
@@ -33,7 +33,7 @@
  - img
 - button "Branch into a new conversation":
  - img
- text: 7/25 {{clock}}
+- text: 7/25 {{clock}}Ran for {{duration}}
 - button "Context compacted View compaction summary":
  - img
  - text: Context compacted View compaction summary
@@ -33,7 +33,7 @@
  - img
 - button "Branch into a new conversation":
  - img
- text: 7/25 {{clock}}
+- text: 7/25 {{clock}}Ran for {{duration}}
 - button "Context compacted View compaction summary":
  - img
  - text: Context compacted View compaction summary
@@ -2,32 +2,38 @@

 ## Light palette

+- --dsh-scrollbar-thumb, pointer outside the sidebar: rgba(0, 0, 0, 0)
 - scrollbar-gutter: stable
 - ::-webkit-scrollbar width: 8px
 - ::-webkit-scrollbar-track background: rgba(0, 0, 0, 0)
 - scrollbar-width: auto
 - scrollbar-color: auto
 - ::-webkit-scrollbar-thumb:hover declarations: var(--dsh-scrollbar-thumb-hover)
- --dsh-scrollbar-thumb: rgb(229, 229, 229)
- --dsh-scrollbar-thumb-hover: rgb(212, 212, 212)
+- --dsh-scrollbar-thumb, pointer over the list: rgb(229, 229, 229)
+- --dsh-scrollbar-thumb-hover, pointer over the list: rgb(212, 212, 212)
 - list overflows: true
 - reserved band: 8px
+- scrollbar inset from the sidebar edge: 2px
+- row background inset from the sidebar edge: 12px
 - relative time covered by the bar: 0px
 - relative time ends inside the content area: true
 - content area ends before the border box: true

 ## Dark palette

+- --dsh-scrollbar-thumb, pointer outside the sidebar: rgba(0, 0, 0, 0)
 - scrollbar-gutter: stable
 - ::-webkit-scrollbar width: 8px
 - ::-webkit-scrollbar-track background: rgba(0, 0, 0, 0)
 - scrollbar-width: auto
 - scrollbar-color: auto
 - ::-webkit-scrollbar-thumb:hover declarations: var(--dsh-scrollbar-thumb-hover)
- --dsh-scrollbar-thumb: rgb(60, 60, 61)
- --dsh-scrollbar-thumb-hover: rgb(84, 85, 87)
+- --dsh-scrollbar-thumb, pointer over the list: rgb(84, 85, 87)
+- --dsh-scrollbar-thumb-hover, pointer over the list: rgb(101, 103, 107)
 - list overflows: true
 - reserved band: 8px
+- scrollbar inset from the sidebar edge: 2px
+- row background inset from the sidebar edge: 12px
 - relative time covered by the bar: 0px
 - relative time ends inside the content area: true
 - content area ends before the border box: true
@@ -37,7 +37,7 @@
  - img
 - button "Branch into a new conversation":
  - img
- text: {{clock}}
+- text: {{clock}}Ran for {{duration}}
 - textbox "Message the agent"
 - button "Commands":
  - img
@@ -28,7 +28,7 @@
  - img
 - button "Branch into a new conversation":
  - img
- text: {{clock}} Now give the same explanation to a human reader. {{clock}}
+- text: {{clock}}Ran for {{duration}} Now give the same explanation to a human reader. {{clock}}
 - button "Copy":
  - img
 - button "Branch into a new conversation" [disabled]:
@@ -43,7 +43,7 @@
  - img
 - button "Branch into a new conversation":
  - img
- text: {{clock}}
+- text: {{clock}}Ran for {{duration}}
 - textbox "Message the agent"
 - button "Commands":
  - img
@@ -23,7 +23,7 @@
  - img
 - button "Branch into a new conversation":
  - img
- text: {{clock}}
+- text: {{clock}}Ran for {{duration}}
 - textbox "Message the agent"
 - button "Commands":
  - img
@@ -20,6 +20,9 @@ function rejectStandaloneServe(): Plugin {

 export default defineConfig({
  plugins: [rejectStandaloneServe(), react()],
+  build: {
+    sourcemap: true,
+  },
  resolve: {
    // Workspace packages resolve to SOURCE: package.json exports point at lib
    // for Node/type consumers, but the browser bundle must compile src directly
@@ -1219,10 +1219,9 @@ Requires: `sessions`

 ```ts config-catalog
 /**
- * Plugin configuration: two verbatim SDK option shapes plus nothing else.
- * `exporter.url` is the one field this package validates itself — required,
- * no default, must parse as an `http(s)` URL — because a missing endpoint
- * must fail at plugin load, not at first export.
+ * Plugin configuration: two verbatim SDK option shapes plus one DSH-owned
+ * shutdown bound. The package validates its endpoint and shutdown deadline
+ * because both must fail at plugin load rather than at first export or exit.
 */
 export interface Config {
  /**
@@ -1240,6 +1239,8 @@ export interface Config {
   * which this plugin fills); the SDK owns and documents these knobs.
   */
  processor?: Omit<BatchLogRecordProcessorOptions, 'exporter'>
+  /** Maximum time spent awaiting the SDK provider's complete shutdown path. */
+  shutdownTimeoutMillis?: number
 }
 ```

@@ -2,5 +2,5 @@
 # side as of the last confirmed-consistent state. Both languages carry equal authority;
 # after editing either side, bring the other along and re-record with:
 #   pnpm run verify-translation-pairing --write packages/client/README.md
-README.md: b111d67fa49e06227e324a33bd53417ad28c3a5b
-README.zh.md: b498008eb82f6ab357718f2af761f38e51140ef8
+README.md: 31d884c04a8b0233713b77b82b3d9cc7052003ca
+README.zh.md: 9c95db529306519136d6d68758889350e5dd65e4
@@ -11,7 +11,7 @@ The browser side of the dsh web GUI: shell kernel, module system, wire consumer,
 | `web-react/` | Shell-side React glue: `createSlotRenderer` + `SessionProvider` render seats | (renderer install) |
 | `connection/` | Wire consumer both ends: browser `ctx.connection` (shared api client + stream loop) and the node half mounting the `/api` route with its browser-trust fence | `ctx.connection` |
 | `runtime/` | Client cordis boot and React-free object services: slots, Sessions, Workspaces, per-session bindings | `ctx.slots` `ctx.sessions` `ctx.workspaces` |
-| `hmr/` | Dev-only hot reload for fetch-arrival client plugins (`--dev` graphs) | (dev entry) |
+| `hmr/` | Dev-only hot reload for script-loaded client plugins (`--dev` graphs) | (dev entry) |
 | `locale/` | Browser locale preference (`zh`/`en`) plus the ns×locale dictionary registry | `ctx.locale` |
 | `ui-slots/` | Slot registry pure core: SlotMap merging, single `register` API, the four-share props family | (types + core) |
 | `ui-theme/` | Theme preference over the `--dsw-*` token stylesheets (`light`/`dark`/`system`) | `ctx.theme` |
@@ -11,7 +11,7 @@ dsh web GUI 的浏览器侧：shell 内核、模块系统、协议消费层、
 | `web-react/` | shell 侧 React 胶水：`createSlotRenderer` + `SessionProvider` 渲染座位 | （渲染器安装） |
 | `connection/` | 协议两端的消费者：浏览器侧 `ctx.connection`（共享 api 客户端 + 流循环），node 半侧挂载带浏览器信任栅栏的 `/api` 路由 | `ctx.connection` |
 | `runtime/` | 客户端 cordis 启动与无 React 对象服务：slots、Session、Workspace、逐会话绑定 | `ctx.slots` `ctx.sessions` `ctx.workspaces` |
-| `hmr/` | 仅开发用的 fetch 到达型客户端插件热重载（`--dev` 图） | （开发条目） |
+| `hmr/` | 仅开发用的外部脚本加载型客户端插件热重载（`--dev` 图） | （开发条目） |
 | `locale/` | 浏览器语言偏好（`zh`／`en`）与 ns×locale 词典注册表 | `ctx.locale` |
 | `ui-slots/` | slot 注册表纯核心：SlotMap 合并、单一 `register` API、四份额 props 族 | （类型 + 核心） |
 | `ui-theme/` | 基于 `--dsw-*` token 样式表的主题偏好（`light`／`dark`／`system`） | `ctx.theme` |
@@ -2,5 +2,5 @@
 # side as of the last confirmed-consistent state. Both languages carry equal authority;
 # after editing either side, bring the other along and re-record with:
 #   pnpm run verify-translation-pairing --write packages/client/hmr/README.md
-README.md: 2b2f63c25cbf3a46babef78a4dfb52f859156887
-README.zh.md: 58fbad900d9ab86a9d28979f691f24de29e9b6f4
+README.md: f91a6c6f685c88a1ea19312985ad3e933222192a
+README.zh.md: 1d20a22d211c13d62089fb5618f40636ab7ae60a
@@ -2,9 +2,9 @@

 English | [中文](README.zh.md)

-Hot reload for fetch-arrival client plugins. A static-arrival entry composed only into `--dev` graphs (`dsh web --dev`); production graphs omit the row, so the shell-bundled code stays inert.
+Hot reload for script-loaded client plugins. A static-arrival entry composed only into `--dev` graphs (`dsh web --dev`); production graphs omit the row, so the shell-bundled code stays inert.

-The browser half subscribes to the system SSE channel (`GET /plugins/events`) and reloads one plugin per `rebuilt` frame, serialized through a queue (the bundle handoff slot is single). The sequence per frame — `prefetch` (fetch the new bundle before touching anything), `invalidate`, `registry.delete` (before the fiber: a bare fiber dispose trips the vendored Loader's self-dispose branch, which would mark the entry disabled), drain the old fiber, delete `entry.fiber`, remove owned `<style data-plugin>` tags, `entry.refresh()` re-imports and remounts, `fiber.await()` rethrows startup failures loud. Dependents reload through cordis itself: a fiber's activation epoch strings its service providers' uids, so replacing a provider's fiber cascades every dependent with zero client-side graph analysis. The node half detects rebuilds with one interval that stat-polls each graph bundle from a synchronous baseline, immediately re-hashes after adding a row, retains missing rows as dirty, and broadcasts only real rev changes; any tsdown watch process producing the bundle therefore triggers HMR with no builder→host channel.
+The browser half subscribes to the system SSE channel (`GET /plugins/events`) and reloads one plugin per `rebuilt` frame through a serialized queue. The sequence per frame — `invalidate`, `prefetch` (load and register the new bundle while the old fiber still serves), `registry.delete` (before the fiber: a bare fiber dispose trips the vendored Loader's self-dispose branch, which would mark the entry disabled), drain the old fiber, delete `entry.fiber`, remove owned `<style data-plugin>` tags, `entry.refresh()` re-imports and remounts, `fiber.await()` rethrows startup failures loud. Dependents reload through cordis itself: a fiber's activation epoch strings its service providers' uids, so replacing a provider's fiber cascades every dependent with zero client-side graph analysis. The node half detects rebuilds with one interval that stat-polls each graph bundle from a synchronous baseline, immediately re-hashes after adding a row, retains missing rows as dirty, and broadcasts only real rev changes; any tsdown watch process producing the bundle therefore triggers HMR with no builder→host channel.

 ## Model Experience

@@ -2,9 +2,9 @@

 [English](README.md) | 中文

-为通过 fetch 加载的客户端插件提供热重载。该静态加载配置项只组合进 `--dev` 图（`dsh web --dev`）；生产图省略该项，因此打包进 shell 的代码保持不活动。
+为通过外部脚本加载的客户端插件提供热重载。该静态加载配置项只组合进 `--dev` 图（`dsh web --dev`）；生产图省略该项，因此打包进 shell 的代码保持不活动。

-浏览器侧订阅系统 SSE（Server-Sent Events）通道（`GET /plugins/events`），每个 `rebuilt` 帧重载一个插件，并通过队列串行执行（组合包交接 slot 只能容纳一个）。每帧的顺序是：`prefetch`（在触碰任何内容前抓取新组合包）、`invalidate`、`registry.delete`（在 fiber dispose（资源释放）之前执行：仅 dispose fiber 会触发 vendored Loader 的 self-dispose 分支，把配置项标为禁用）、排空旧 fiber、删除 `entry.fiber`、移除自身拥有的 `<style data-plugin>` 标签、通过 `entry.refresh()` 重新导入并挂载、通过 `fiber.await()` 直接重新抛出启动失败。依赖方由 Cordis 自身重载：fiber 的激活 epoch 会串联其服务提供方的 uid，因此替换提供方 fiber 会级联所有依赖方，无需客户端图分析。node 侧使用一个 interval 检测重建：从同步基线开始 stat-poll 每个图组合包；新增一行后立即重新计算 hash；缺失行保持 dirty；只广播真实 rev 变更。因此，任何生成组合包的 tsdown watch 进程都能触发 HMR（热模块替换），无需 builder→host 通道。
+浏览器侧订阅系统 SSE（Server-Sent Events）通道（`GET /plugins/events`），每个 `rebuilt` 帧重载一个插件，并通过队列串行执行。每帧的顺序是：`invalidate`、`prefetch`（旧 fiber 仍在服务时加载并注册新组合包）、`registry.delete`（在 fiber dispose（资源释放）之前执行：仅 dispose fiber 会触发 vendored Loader 的 self-dispose 分支，把配置项标为禁用）、排空旧 fiber、删除 `entry.fiber`、移除自身拥有的 `<style data-plugin>` 标签、通过 `entry.refresh()` 重新导入并挂载、通过 `fiber.await()` 直接重新抛出启动失败。依赖方由 Cordis 自身重载：fiber 的激活 epoch 会串联其服务提供方的 uid，因此替换提供方 fiber 会级联所有依赖方，无需客户端图分析。node 侧使用一个 interval 检测重建：从同步基线开始 stat-poll 每个图组合包；新增一行后立即重新计算 hash；缺失行保持 dirty；只广播真实 rev 变更。因此，任何生成组合包的 tsdown watch 进程都能触发 HMR（热模块替换），无需 builder→host 通道。

 ## 模型体验

@@ -1,6 +1,6 @@
 {
  "name": "@deepseek-ai/dsh-client-hmr",
-  "description": "Dev-only hot-reload driver for fetch-arrival client entries: SSE rebuilt frames → prefetch/invalidate → fiber swap through the vendored Loader entry",
+  "description": "Dev-only hot-reload driver for script-loaded client entries: SSE rebuilt frames → invalidate/prefetch → fiber swap through the vendored Loader entry",
  "version": "0.0.1",
  "private": true,
  "type": "module",
@@ -2,7 +2,7 @@
 * client-hmr, browser half: hot-reload driver for client plugin entries.
 *
 * Listens on the host's system SSE channel (`GET /plugins/events`); on a
- * `rebuilt` frame it re-fetches the entry's bundle and swaps the cordis
+ * `rebuilt` frame it reloads the entry's bundle and swaps the cordis
 * fiber in place. Every graph entry is a plugin bundle under the web2 model
 * — `immediately` rows differ only in stage-one prefetch (a boot
 * optimization), so all rostered plugin packages share these reload semantics;
@@ -14,7 +14,7 @@
 * cascades into its UI dependents with no HMR-side bookkeeping.
 *
 * Reload order (lazy CJS table): invalidate (drop the stale factory and
- * materialized record) → prefetch (fetch + execute + register the fresh
+ * materialized record) → prefetch (load and register the fresh
 * factory) → registry-first teardown → drain old fiber unload → remove
 * owned `<style data-plugin>` tags → `entry.refresh()` materializes the new
 * factory. Invalidate MUST precede prefetch: a live factory makes prefetch
@@ -110,7 +110,7 @@ export function apply(ctx: Context): void {
    }
    // Invalidate first (drop stale factory + record — a live factory makes
    // prefetch a no-op and re-registration a loud duplicate), then run the
-    // async half while the old fiber still serves: fetch + execute registers
+    // async half while the old fiber still serves: script loading registers
    // the fresh factory with zero side effects (lazy CJS — module bodies run
    // at materialization, not execution).
    modLoader.invalidate(id)
@@ -2,5 +2,5 @@
 # side as of the last confirmed-consistent state. Both languages carry equal authority;
 # after editing either side, bring the other along and re-record with:
 #   pnpm run verify-translation-pairing --write packages/client/modules/README.md
-README.md: 99565b349d782c58752ac3e73ce7c0be527f78a8
-README.zh.md: a8ed0a4949ccefce53933b4f2fb8f51f5291684f
+README.md: 7d661c806955d0fac021dd6620994aab83c0f773
+README.zh.md: a1da42a552dbe8770fcb78bf458c01a0057ce8dd
@@ -6,9 +6,9 @@ Client module system: the browser peer of Node's internal ESM loader, built as a

 Lazy CJS model (web2): executing a plugin bundle only REGISTERS its factory (`window.__ModuleLoader__.load({id, factory})`); every module body side effect — CSS injection included — lives in the factory closure and runs at materialization (`factory(require)` → export surface, memoized in `loadCache`), not at script execution. A factory that requires another registered-but-unmaterialized module materializes it recursively, so load order needs no external sequencing; require cycles throw (factory-form CJS cannot deliver partial exports). `<id>/client` and the bare id name the same surface (a plugin bundle IS its package's client half).

-Resolution branch order (`import(specifier)`): platform seed word → shell instance; memoized record → surface; shell-own static registry (`registerStatic`, app-shell) → module; registered factory → materialize; graph row (`window.__DSH_BOOT__`) → fetch + execute + materialize; anything else throws — the runtime mirror of the build-time bundle purity gate. The synchronous `require` handed to factories walks the same order minus the fetch branch and records observed edges into the module record. `prefetch` is the stage-one arrival hook (fetch + execute, registration only; concurrent calls share one in-flight task); `invalidate` drops the factory and the materialized record so the next prefetch/import refetches (the HMR hook).
+Resolution branch order (`import(specifier)`): platform seed word → shell instance; memoized record → surface; shell-own static registry (`registerStatic`, app-shell) → module; registered factory → materialize; graph row (`window.__DSH_BOOT__`) → load its external classic script + materialize; anything else throws — the runtime mirror of the build-time bundle purity gate. The synchronous `require` handed to factories walks the same order minus the asynchronous load branch and records observed edges into the module record. `prefetch` is the stage-one arrival hook (script load and factory registration only; concurrent calls share one in-flight task); `invalidate` drops the factory and materialized record so the next prefetch/import reloads the script (the HMR hook).

-The Node half scans enabled Loader entries for web `dshClient` packages, resolves each `exports["./client"]`, hashes the built bundle into the boot graph, and serves it under `/plugins`. Source launch maps host imports to TypeScript source but still consumes this built client export; missing files share one build instruction followed by a package/path list, while unrelated filesystem errors remain separate failures.
+The Node half scans enabled Loader entries for web `dshClient` packages, resolves each `exports["./client"]`, hashes the built bundle into the boot graph, and serves it with its source map under `/plugins`. Source launch maps host imports to TypeScript source but still consumes this built client export; missing files share one build instruction followed by a package/path list, while unrelated filesystem errors remain separate failures.

 ## Model Experience

@@ -6,9 +6,9 @@

 惰性 CJS 模型（web2）：执行插件组合包只会注册其 factory（`window.__ModuleLoader__.load({id, factory})`）；每个模块主体的副作用（包括 CSS 注入）都位于 factory 闭包中，在物化时运行（`factory(require)` → 导出表层，并在 `loadCache` 中记忆化），不会在脚本执行时运行。如果 factory 依赖另一个已注册但尚未物化的模块，系统会递归物化它，因此加载顺序无需外部编排；require 循环会抛出异常（factory 形式的 CJS 无法提供部分导出）。`<id>/client` 与裸 id 指向同一表层（一个插件组合包就是其包的客户端侧）。

-解析分支顺序（`import(specifier)`）：平台种子词 → 外壳实例；记忆化记录 → 表层；外壳自身的静态注册表（`registerStatic`，app-shell）→ 模块；已注册 factory → 物化；模块图记录（`window.__DSH_BOOT__`）→ 抓取 + 执行 + 物化；其他情况一律抛出异常。这是构建时组合包纯度门禁的运行时镜像。交给 factory 的同步 `require` 采用相同顺序，但不含抓取分支，并把观察到的边记录到模块记录中。`prefetch` 是第一阶段加载钩子（抓取 + 执行，只注册；并发调用共享一个进行中的任务）；`invalidate` 会丢弃 factory 与物化记录，使下一次 prefetch/import 重新抓取；它是 HMR（热模块替换）钩子。
+解析分支顺序（`import(specifier)`）：平台种子词 → 外壳实例；记忆化记录 → 表层；外壳自身的静态注册表（`registerStatic`，app-shell）→ 模块；已注册 factory → 物化；模块图记录（`window.__DSH_BOOT__`）→ 加载外部 classic script + 物化；其他情况一律抛出异常。这是构建时组合包纯度门禁的运行时镜像。交给 factory 的同步 `require` 采用相同顺序，但不含异步加载分支，并把观察到的边记录到模块记录中。`prefetch` 是第一阶段到达钩子（只加载脚本并注册 factory；并发调用共享一个进行中的任务）；`invalidate` 会丢弃 factory 与物化记录，使下一次 prefetch/import 重新加载脚本；它是 HMR（热模块替换）钩子。

-Node 侧会扫描已启用的 Loader 配置项以发现 web `dshClient` 包，解析每个 `exports["./client"]`，把构建后的组合包哈希写入启动图，并通过 `/plugins` 提供该文件。源码启动会把宿主侧导入映射到 TypeScript 源码，但仍消费客户端导出的构建产物；缺失文件共享一条构建要求，随后以 package/path list 列出各项，而无关的文件系统错误仍是独立故障。
+Node 侧会扫描已启用的 Loader 配置项以发现 web `dshClient` 包，解析每个 `exports["./client"]`，把构建后的组合包哈希写入启动图，并通过 `/plugins` 提供该文件及其 sourcemap。源码启动会把宿主侧导入映射到 TypeScript 源码，但仍消费客户端导出的构建产物；缺失文件共享一条构建要求，随后以 package/path list 列出各项，而无关的文件系统错误仍是独立故障。

 ## 模型体验

@@ -17,11 +17,11 @@
 *
 * Resolution branch order (import): seed word → shell instance; memoized
 * record → surface; static registry (shell-own modules, e.g. app-shell) →
- * module; registered factory → materialize; graph row → fetch + execute +
- * materialize; anything else → throw (loud — the runtime mirror of the
+ * module; registered factory → materialize; graph row → load + materialize;
+ * anything else → throw (loud — the runtime mirror of the
 * build-time bundle purity gate). The synchronous `require` handed to
- * factories walks the same order minus the fetch branch: fetching is async,
- * so only already-executed bundles can be required — and cross-plugin value
+ * factories walks the same order minus the load branch: loading is async,
+ * so only already-registered bundles can be required — and cross-plugin value
 * imports are a build error anyway.
 *
 * This file is the browser-safe contract face (zero node imports): the
@@ -56,7 +56,7 @@ export interface WebBootEntry {
  rev: string
  /** Package-name dependency edges, informational (preflight display / HMR diffing). */
  inject?: string[]
-  /** Stage-one prefetch mark: fetch + execute (factory registration) during module-face boot. */
+  /** Stage-one prefetch mark: load the script for factory registration during module-face boot. */
  immediately?: boolean
 }

@@ -210,18 +210,17 @@ export interface ClientModuleLoader {
   */
  registerStatic(id: string, module: unknown): void
  /**
-   * Stage-one arrival: fetch the entry's bundle and execute it, registering
-   * its factory (no materialization — module side effects wait for import).
+   * Stage-one arrival: load the entry's script to register its factory (no
+   * materialization — module side effects wait for import).
   * No-op for static-registered ids and ids whose factory is already
   * registered; concurrent calls share one in-flight task. To force a fresh
-   * fetch (HMR), {@link invalidate} first.
+   * load (HMR), {@link invalidate} first.
   * @param id - graph entry name.
   */
  prefetch(id: string): Promise<void>
  /**
-   * Full reset of one module: drop its registered factory, its materialized
-   * record, and any consumed bundle text, so the next prefetch/import
-   * refetches and re-executes (the HMR invalidation hook).
+   * Full reset of one module: drop its registered factory and materialized
+   * record so the next prefetch/import reloads it (the HMR invalidation hook).
   * @param id - entry name to invalidate.
   */
  invalidate(id: string): void
@@ -233,11 +232,6 @@ export interface ClientModuleSystemOptions {
  modules: BootModuleRow[]
  /** Module-table seed: platform-singleton specifier → shell instance. */
  staticModules: Record<string, unknown>
-  /** Bundle fetch seam (parallelizable half). Defaults to same-origin fetch().text(). */
-  fetchBundle?: (url: string) => Promise<string>
-  /**
-   * Bundle execution seam (synchronously performs the load() registration).
-   * Defaults to a <script> element carrying the code.
-   */
-  executeBundle?: (code: string, url: string) => void
+  /** Bundle-load seam. Defaults to a same-origin classic `<script src>` element. */
+  loadBundle?: (url: string) => Promise<void>
 }
@@ -2,38 +2,28 @@
 * ClientModuleSystem — the implementation behind the {@link ClientModuleLoader}
 * seam. The conceptual contract (lazy CJS model, resolution branch order) is
 * documented on the public interfaces in `./manifest.ts`; this file owns the
- * state tables and the fetch/execute/materialize machinery.
+ * state tables and the load/materialize machinery.
 */
 import type {
  BootModuleRow, ClientModuleLoader, ClientModuleRecord,
  ClientModuleSystemOptions, ClientPluginHandoff, DshWindow,
 } from './manifest.ts'

-/** A registered-but-unmaterialized bundle: the factory plus its source URL (diagnostics). */
-interface RegisteredFactory {
-  factory: ClientPluginHandoff['factory']
-  url: string
-}
-
-/** Default bundle fetch seam: same-origin fetch().text(). */
-const defaultFetchBundle = async (url: string): Promise<string> => {
-  const res = await fetch(url)
-  if (!res.ok) throw new Error(`client-modules: bundle fetch ${url} answered ${String(res.status)}`)
-  return res.text()
-}
-
-/** Default bundle execution seam: a <script> element carrying the code. */
-const defaultExecuteBundle = (code: string, url: string): void => {
+/** Default bundle-load seam: same-origin external classic script. */
+const defaultLoadBundle = (url: string): Promise<void> => new Promise((resolve, reject) => {
  const el = document.createElement('script')
-  // Inline execution (not src) so the fetch half stays parallelizable; the
-  // sourceURL comment keeps devtools stack frames attributed to the bundle.
-  el.textContent = `${code}\n//# sourceURL=${url}`
-  document.head.appendChild(el)
-  // Execution is synchronous for inline scripts: the factory is registered by
-  // now, so the node (and its source text) has no further job. Removing it
-  // keeps repeated HMR rebuilds from accumulating dead script nodes.
-  el.remove()
-}
+  el.async = true
+  el.src = url
+  el.addEventListener('load', () => {
+    el.remove()
+    resolve()
+  }, { once: true })
+  el.addEventListener('error', () => {
+    el.remove()
+    reject(new Error(`client-modules: bundle script ${url} failed to load`))
+  }, { once: true })
+  document.head.append(el)
+})

 /**
 * A plugin bundle IS its package's client half: `<id>/client` (the exports
@@ -72,31 +62,21 @@ export class ClientModuleSystem implements ClientModuleLoader {

  private readonly seed: Map<string, unknown>
  private readonly statics = new Map<string, unknown>()
-  private readonly factories = new Map<string, RegisteredFactory>()
-  /** In-flight prefetch (fetch + execute) per id; concurrent callers share it. */
+  private readonly factories = new Map<string, ClientPluginHandoff['factory']>()
+  /** In-flight prefetch (script load) per id; concurrent callers share it. */
  private readonly pendingArrival = new Map<string, Promise<void>>()
  /** Materialization re-entrancy guard: factory-form CJS cannot deliver partial exports, so a cycle is fatal. */
  private readonly materializing = new Set<string>()
  private readonly graphRows = new Map<string, BootModuleRow>()
-  // Execution URL of the bundle currently being executed (bound into the
-  // factory registration so diagnostics can name the source).
-  private executingUrl = ''
-  // Graph id of the row currently being executed ('' outside arrive):
-  // the load sink cross-checks the handoff id against it so a mis-stamped
-  // bundle cannot register under another entry's identity.
-  private executingId = ''
-
-  private readonly fetchBundle: (url: string) => Promise<string>
-  private readonly executeBundle: (code: string, url: string) => void
+  private readonly loadBundle: (url: string) => Promise<void>

  /**
   * Build the module system over the parsed boot rows.
-   * @param options - module rows, module-table staticModules, fetch/execute seams.
+   * @param options - module rows, module-table staticModules, and bundle-load seam.
   */
  constructor(options: ClientModuleSystemOptions) {
    this.seed = new Map(Object.entries(options.staticModules))
-    this.fetchBundle = options.fetchBundle ?? defaultFetchBundle
-    this.executeBundle = options.executeBundle ?? defaultExecuteBundle
+    this.loadBundle = options.loadBundle ?? defaultLoadBundle

    for (const row of options.modules) {
      if (this.graphRows.has(row.id)) throw new Error(`client-modules: duplicate graph entry "${row.id}"`)
@@ -110,37 +90,22 @@ export class ClientModuleSystem implements ClientModuleLoader {
        // Registration is keyed by the handoff id; a duplicate means a bundle
        // executed twice without an invalidate — always a bug, always loud.
        if (this.factories.has(handoff.id)) throw new Error(`client-modules: duplicate factory registration for "${handoff.id}" (bundle executed twice without invalidate?)`)
-        // A fetched row's bundle must register the id its row names — a
-        // mis-stamped bundle registering under another entry's identity
-        // would let that entry silently materialize foreign exports.
-        if (this.executingId !== '' && handoff.id !== this.executingId) {
-          throw new Error(`client-modules: bundle ${this.executingUrl} registered "${handoff.id}" while arriving for "${this.executingId}" (mis-stamped bundle id)`)
-        }
-        this.factories.set(handoff.id, { factory: handoff.factory, url: this.executingUrl })
+        this.factories.set(handoff.id, handoff.factory)
      },
    }
  }

-  /** Fetch + execute one graph row so its factory is registered (idempotent per in-flight arrival). */
+  /** Load one graph row so its factory is registered (idempotent per in-flight arrival). */
  private arrive(row: BootModuleRow): Promise<void> {
    const { id, url } = row
    const pending = this.pendingArrival.get(id)
    if (pending !== undefined) return pending
    if (this.factories.has(id)) return Promise.resolve()
-    const task = (async (): Promise<void> => {
-      const code = await this.fetchBundle(url)
-      this.executingUrl = url
-      this.executingId = id
-      try {
-        this.executeBundle(code, url)
-      } finally {
-        this.executingUrl = ''
-        this.executingId = ''
-      }
+    const task = this.loadBundle(url).then(() => {
      if (!this.factories.has(id)) {
-        throw new Error(`client-modules: bundle ${url} executed without registering "${id}" via __ModuleLoader__.load`)
+        throw new Error(`client-modules: bundle ${url} loaded without registering "${id}" via __ModuleLoader__.load`)
      }
-    })().finally(() => { this.pendingArrival.delete(id) })
+    }).finally(() => { this.pendingArrival.delete(id) })
    this.pendingArrival.set(id, task)
    return task
  }
@@ -158,7 +123,7 @@ export class ClientModuleSystem implements ClientModuleLoader {
    this.materializing.add(id)
    try {
      const edges = new Set<string>()
-      const surface = registered.factory(this.makeRequire(edges))
+      const surface = registered(this.makeRequire(edges))
      const record: ClientModuleRecord = { id, surface, styles: claimStyles(id), edges }
      this.loadCache.set(id, record)
      return record
@@ -2,8 +2,8 @@
 * Node half of the client module system (dshClient dual-face package): scans
 * the host Loader's entries for `dshClient` packages, composes the
 * `window.__DSH_BOOT__` entry graph (wire single source: {@link WebBootEntry}
- * in `./client/manifest.ts`), serves `/plugins/<id>/client.js`, taps the
- * index render to inject the boot manifest, and provides the
+ * in `./client/manifest.ts`), serves `/plugins/<id>/client.js` and its source
+ * map, taps the index render to inject the boot manifest, and provides the
 * `clientModuleHost` service (the HMR node half's registration/notification
 * face).
 *
@@ -424,9 +424,15 @@ export class ClientModuleHostService extends Service {
    const pathname = decodeURIComponent(new URL(req.url ?? '/', 'http://x').pathname)
    // The id may contain a scope slash. Anything else under /plugins (including
    // /plugins/events when the HMR row is absent) is an unknown resource.
-    const path = pathname.startsWith('/plugins/') && pathname.endsWith('/client.js')
-      ? this.clientPath(pathname.slice('/plugins/'.length, -'/client.js'.length))
+    const prefix = '/plugins/'
+    const mapSuffix = '/client.js.map'
+    const bundleSuffix = '/client.js'
+    const isSourceMap = pathname.startsWith(prefix) && pathname.endsWith(mapSuffix)
+    const suffix = isSourceMap ? mapSuffix : bundleSuffix
+    const clientPath = pathname.startsWith(prefix) && pathname.endsWith(suffix)
+      ? this.clientPath(pathname.slice(prefix.length, -suffix.length))
      : undefined
+    const path = clientPath === undefined ? undefined : `${clientPath}${isSourceMap ? '.map' : ''}`
    if (path === undefined) {
      res.writeHead(404)
      res.end()
@@ -434,7 +440,10 @@ export class ClientModuleHostService extends Service {
    }
    try {
      const body = await readFile(path)
-      res.writeHead(200, { 'content-type': 'text/javascript; charset=utf-8', 'cache-control': 'no-cache' })
+      res.writeHead(200, {
+        'content-type': isSourceMap ? 'application/json; charset=utf-8' : 'text/javascript; charset=utf-8',
+        'cache-control': 'no-cache',
+      })
      res.end(body)
    } catch {
      // Registered but unreadable (bundle not built yet): loud 404 beats a silent SPA-fallback HTML page.
@@ -4,7 +4,7 @@
 * registers the factory), materialization on first import/require with
 * memoization and recursive self-sequencing, the resolution branch order,
 * shared in-flight arrival, invalidate-refetch (HMR), style claiming, the
- * default transport seams, and the loud failure modes (duplicate
+ * default transport seam, and the loud failure modes (duplicate
 * registration, cycles, table misses, double boot).
 */
 import { afterEach, describe, expect, it, vi } from 'vitest'
@@ -20,7 +20,6 @@ type Factory = ClientPluginHandoff['factory']
 afterEach(() => {
  vi.unstubAllGlobals()
  delete win.__ModuleLoader__
-  delete (document as unknown as Record<string, unknown>).__realmBridge
  for (const el of document.querySelectorAll('style, script')) el.remove()
 })

@@ -33,9 +32,9 @@ interface Bench {
 }

 /**
- * Loader over scripted bundles: fetch resolves to the row url (optionally
- * gated on a release callback); execute registers the scripted factory
- * through the window sink (`null` scripts a bundle that never calls load).
+ * Loader over scripted bundles: load records the row URL, optionally waits on
+ * a release callback, then registers the scripted factory through the window
+ * sink (`null` scripts a bundle that never calls load).
 */
 function bench(
  entries: BootModuleRow[],
@@ -47,15 +46,12 @@ function bench(
  const loader = new ClientModuleSystem({
    modules: entries,
    staticModules: opts.seed ?? {},
-    fetchBundle: (url) => {
+    loadBundle: async (url) => {
      fetched.push(url)
      if (opts.gated?.includes(url) === true) {
-        return new Promise((resolve) => { gates.set(url, () => { resolve(url) }) })
+        await new Promise<void>((resolve) => { gates.set(url, resolve) })
      }
-      return Promise.resolve(url)
-    },
-    executeBundle: (code) => {
-      const id = /\/plugins\/(.+)\/client\.js/.exec(code)?.[1]
+      const id = /\/plugins\/(.+)\/client\.js/.exec(url)?.[1]
      const factory = id === undefined ? undefined : bundles[id]
      if (factory == null || id === undefined) return
      win.__ModuleLoader__?.load({ id, factory })
@@ -65,7 +61,7 @@ function bench(
 }

 describe('lazy CJS arrival', () => {
-  it('prefetch fetches and executes but does not run the factory', async () => {
+  it('prefetch loads and registers but does not run the factory', async () => {
    const ran: string[] = []
    const b = bench([row('a')], { a: () => { ran.push('a'); return {} } })
    await b.loader.prefetch('a')
@@ -85,7 +81,7 @@ describe('lazy CJS arrival', () => {
    expect(b.loader.loadCache.get('a')?.id).toBe('a')
  })

-  it('import without prefetch fetches, executes, and materializes in one call', async () => {
+  it('import without prefetch loads, registers, and materializes in one call', async () => {
    const b = bench([row('a')], { a: () => ({ marker: 'direct' }) })
    const surface = await b.loader.import('a', '', {})
    expect((surface as { marker: string }).marker).toBe('direct')
@@ -228,7 +224,7 @@ describe('failure modes', () => {
 })

 describe('HMR reset', () => {
-  it('invalidate drops the factory and record so the module refetches and re-registers', async () => {
+  it('invalidate drops the factory and record so the module reloads and re-registers', async () => {
    let generation = 0
    const b = bench([row('a')], { a: () => ({ generation: ++generation }) })
    const first = await b.loader.import('a', '', {})
@@ -275,27 +271,35 @@ describe('style claiming', () => {
  })
 })

-describe('default transport seams', () => {
-  it('fetches same-origin and executes through an inline script tag', async () => {
-    // In a browser the loader's globalThis IS the page window; vitest's jsdom
-    // evaluates <script> in a separate realm that shares only the document,
-    // so the fixture bundle restores the sink from a document bridge before
-    // using the normal calling convention.
-    const code = 'window.__ModuleLoader__ = document.__realmBridge;\n'
-      + 'window.__ModuleLoader__.load({ id: "dee", factory: function () { return { marker: "via-script" } } })'
-    vi.stubGlobal('fetch', async () => ({ ok: true, text: async () => code }))
+describe('default transport seam', () => {
+  it('loads through an external classic script and removes the settled node', async () => {
+    const append = vi.spyOn(document.head, 'append').mockImplementation((...nodes) => {
+      const script = nodes[0]
+      if (!(script instanceof HTMLScriptElement)) throw new Error('expected script node')
+      expect(script.async).toBe(true)
+      expect(script.getAttribute('src')).toBe('/plugins/dee/client.js?rev=0')
+      queueMicrotask(() => {
+        win.__ModuleLoader__?.load({ id: 'dee', factory: () => ({ marker: 'via-script' }) })
+        script.dispatchEvent(new Event('load'))
+      })
+    })
    const loader: ClientModuleLoader = new ClientModuleSystem({ modules: [row('dee')], staticModules: {} })
-    ;(document as unknown as Record<string, unknown>).__realmBridge = win.__ModuleLoader__
    const surface = await loader.import('dee', '', {})
    expect((surface as { marker: string }).marker).toBe('via-script')
-    // The script node is removed right after its synchronous execution —
-    // repeated HMR rebuilds must not accumulate dead script nodes.
+    expect(append).toHaveBeenCalledOnce()
    expect([...document.querySelectorAll('script')]).toEqual([])
  })

-  it('a non-ok bundle response is loud with the status', async () => {
-    vi.stubGlobal('fetch', async () => ({ ok: false, status: 404 }))
+  it('a script load failure is loud and removes the node', async () => {
+    vi.spyOn(document.head, 'append').mockImplementation((...nodes) => {
+      const script = nodes[0]
+      if (!(script instanceof HTMLScriptElement)) throw new Error('expected script node')
+      queueMicrotask(() => { script.dispatchEvent(new Event('error')) })
+    })
    const loader = new ClientModuleSystem({ modules: [row('dee')], staticModules: {} })
-    await expect(loader.prefetch('dee')).rejects.toThrow('answered 404')
+    await expect(loader.prefetch('dee')).rejects.toThrow(
+      'bundle script /plugins/dee/client.js?rev=0 failed to load',
+    )
+    expect([...document.querySelectorAll('script')]).toEqual([])
  })
 })
@@ -1,12 +1,13 @@
 /** Node-half composition diagnostics for package metadata and built client bundles. */

 import { mkdirSync, mkdtempSync, realpathSync, rmSync, writeFileSync } from 'node:fs'
+import type { IncomingMessage, ServerResponse } from 'node:http'
 import { tmpdir } from 'node:os'
-import { join } from 'node:path'
+import { dirname, join } from 'node:path'
 import { pathToFileURL } from 'node:url'
 import { Context } from 'cordis'
 import { afterEach, describe, expect, it } from 'vitest'
-import type { HttpServerService } from '@deepseek-ai/dsh-host-webserver'
+import type { HttpServerService, WebRoute } from '@deepseek-ai/dsh-host-webserver'
 import { ClientModuleHostService } from '../src/index.ts'

 let root: string | undefined
@@ -33,8 +34,8 @@ function writePackage(packageName: string): string {
  return clientPath
 }

-/** Construct the node-half service over the enabled fixture entries. */
-function construct(packageNames: string[]): ClientModuleHostService {
+/** Construct the node-half service and capture its plugin-bundle route. */
+function constructWithRoute(packageNames: string[]): { service: ClientModuleHostService; route: WebRoute } {
  const ctx = new Context()
  ctx.baseUrl = pathToFileURL(root!).href + '/'
  ctx.provide('loader', {
@@ -44,13 +45,24 @@ function construct(packageNames: string[]): ClientModuleHostService {
      }
    },
  })
+  let route: WebRoute | undefined
  const httpServer: Pick<HttpServerService, 'port' | 'register' | 'tapIndex'> = {
    port: 0,
-    register: () => () => {},
+    register: (candidate) => {
+      if (candidate.path === '/plugins') route = candidate
+      return () => {}
+    },
    tapIndex: () => () => {},
  }
  ctx.provide('httpServer', httpServer as HttpServerService)
-  return new ClientModuleHostService(ctx)
+  const service = new ClientModuleHostService(ctx)
+  if (route === undefined) throw new Error('client bundle route was not registered')
+  return { service, route }
+}
+
+/** Construct the node-half service over the enabled fixture entries. */
+function construct(packageNames: string[]): ClientModuleHostService {
+  return constructWithRoute(packageNames).service
 }

 describe('client bundle activation', () => {
@@ -84,4 +96,40 @@ describe('client bundle activation', () => {
    expect(String(thrown)).toContain('EISDIR')
    expect(String(thrown)).not.toContain('pnpm run build')
  })
+
+  it('serves the source map beside a registered client bundle', async () => {
+    const packageName = '@fixture/source-map'
+    const clientPath = writePackage(packageName)
+    mkdirSync(dirname(clientPath), { recursive: true })
+    writeFileSync(clientPath, 'module.exports = {}\n')
+    const map = '{"version":3,"sources":["src/client/index.tsx"]}\n'
+    writeFileSync(`${clientPath}.map`, map)
+    const { route } = constructWithRoute([packageName])
+    let status = 0
+    let headers: Record<string, string> | undefined
+    let body = ''
+    const response = {
+      writeHead(nextStatus: number, nextHeaders?: Record<string, string>) {
+        status = nextStatus
+        headers = nextHeaders
+        return response
+      },
+      end(chunk?: Uint8Array) {
+        body = chunk === undefined ? '' : Buffer.from(chunk).toString('utf8')
+        return response
+      },
+    } as unknown as ServerResponse
+
+    await route.handler({
+      method: 'GET',
+      url: `/plugins/${packageName}/client.js.map`,
+    } as IncomingMessage, response)
+
+    expect(status).toBe(200)
+    expect(headers).toEqual({
+      'content-type': 'application/json; charset=utf-8',
+      'cache-control': 'no-cache',
+    })
+    expect(body).toBe(map)
+  })
 })
@@ -331,6 +331,8 @@ export interface ConversationSnapshot {
  sessionId: SessionId
  /** Human transcript plus retry notices and interrupted-turn terminal nodes in event order. */
  nodes: readonly ConversationNode[]
+  /** Exact in-window `turn/start` time and optional matching `turn/end` time. */
+  turnTimings: ReadonlyMap<number, { readonly startTime: number; readonly endTime?: number }>
  /** In-window completed turn number -> its `turn/end` event seq. */
  turnEnds: ReadonlyMap<number, number>
  partial: PartialAssistant | null
--- a/Show More
+++ b/Show More