Merge remote-tracking branch 'origin/master' into worktree-session-hover-menu-fix

# Conflicts:
#	apps/web/tests/workspace-management.e2e.ts

.agents/notes/implemented/bug-fix/2026-07-27-glob-sampling.i18n.yaml

@@ -0,0 +1,6 @@
+# Bilingual-pair consistency record (docs/i18n/README.md): the git blob hash of each
+# side as of the last confirmed-consistent state. Both languages carry equal authority;
+# after editing either side, bring the other along and re-record with:
+#   pnpm run verify-translation-pairing --write .agents/notes/implemented/bug-fix/2026-07-27-glob-sampling.md
+2026-07-27-glob-sampling.md: 67912cf290b127a96819d57f30387197af68323d
+2026-07-27-glob-sampling.zh.md: e339a5f87b483849df60feb726b061fe85074300

.agents/notes/implemented/bug-fix/2026-07-27-glob-sampling.md

@@ -0,0 +1,47 @@
+# Agent Note: Sample over-cap glob results across the tree
+
+Status: implemented
+
+English | [中文](2026-07-27-glob-sampling.zh.md)
+
+## Problem
+
+Asked what a workspace contained, an agent described one subfolder as if it were the whole project. The workspace held 22 top-level entries and 11,485 files. `glob {"pattern":"*"}` matched 10,030 paths, but all 100 inline paths sat under one recently unpacked subtree, so the model never saw the other 21 entries.
+
+Three individually valid behaviors composed into the false impression. A glob without `/` matches basenames at any depth, so `*` means every file in the tree rather than the shell's current-directory expansion. Ripgrep's `--sort=modified` is ascending, so an archive's restored old timestamps put that subtree first. The inline page then took the head of that order without saying that it represented only one concentrated slice.
+
+## Decision
+
+A result that fits within `globMaxResults` remains complete and byte-for-byte modification-time ordered. The required `sampleOverCapGlobResults` config has no fallback: `false` retains the modification-time head for an over-cap result, while `true` samples round-robin across the complete result's top-level entries. In sampling mode, every entry receives one slot before any receives a second, exhausted groups drop out, relative order remains stable within each group, and grouping is relative to the actual search root, including an explicit `path`.
+
+In sampling mode, the footer states that the page is a cross-entry sample rather than the modification-time head and reports how many top-level entries it reaches when that fact adds information. When more top-level entries exist than inline slots, it tells the model to narrow `path`. Head mode keeps the ordinary capped-result footer. When spill succeeds, both modes preserve the complete sorted list in the artifact.
+
+The prompt and schema state the configured over-cap ordering, that a pattern without `/` matches at any depth, and that glob returns files, never directory entries. The shipped CLI composition explicitly selects head mode; deployments that want representative capped pages select sampling mode. Directory orientation remains ordinary shell work in deployments that expose the model-facing bash tool: use `ls` for one directory, and glob for a named file-path pattern across the tree. `ctx.fs.listDir` remains an internal provider primitive used by skill discovery; this decision adds no model-facing `list` tool.
+
+## Alternatives considered
+
+**Keep the modification-time head as the only behavior.** Rejected after measuring the failure shape. Some deployments need the stable ordering, but a deployment that values workspace orientation can explicitly select representative data instead of asking the model to distrust the only paths it received.
+
+**Give the sampling choice a default.** Rejected. No product-wide evidence establishes either ordering as the implicit contract, so every composition selects one and misconfiguration fails at load.
+
+**Sample every result.** Rejected. A complete result loses nothing to truncation, so modification-time order remains useful for age-oriented questions. Sampling begins only when the head stops describing the whole.
+
+**Switch to newest-first order.** Rejected. It merely changes which concentrated subtree can dominate and removes the existing oldest-first contract without making a capped page representative.
+
+**Sample only past a skew threshold.** Rejected. No current evidence supports a deployment-wide threshold, and the model could not know which ordering contract applied. The existing cap is the explainable transition.
+
+**Balance recursively below the top level.** Deferred. First-segment balance fixes the observed failure; deeper balancing needs an unsupported depth-versus-breadth policy.
+
+**Add a model-facing `list` tool.** Rejected after implementation review. The default coding composition already exposes general bash and the model understands `ls`; a duplicate tool would add permanent schema/prompt tokens plus ordering, pagination, symlink, escaping, UI, and snapshot contracts without a distinct security or policy benefit. Thin deployments without a model-facing bash tool do not gain directory orientation from this change.
+
+**Reject `*` or silently anchor separator-free patterns.** Rejected. The same basename-at-any-depth behavior makes `*.ts` useful across a tree. Documenting the rule preserves working ripgrep semantics.
+
+## Consequences
+
+A sampling-mode over-cap page no longer answers age-order questions from its inline paths; its footer says so, and the spill artifact retains the complete sorted view. Sampling balances only the first segment beneath the search root, so a deeper hot subtree can still dominate within one top-level entry. Head mode retains the concentration risk as an explicit deployment trade-off.
+
+The tool surface does not grow. Every composition must set `sampleOverCapGlobResults`; changing it alters glob's prompt, schema description, and over-cap Native rendering. The canonical output keeps `root` so sampling mode can recover its grouping basis, while fitting results remain unchanged.
+
+## Testing
+
+Package tests pin the required config, both over-cap modes, their prompt and schema descriptions, concentrated and flat results, explicit roots, more groups than the JavaScript argument limit, exhausted groups, fewer slots than groups, and paths outside the workdir. The `fs-glob-sampling` ACP scenario explicitly enables sampling, boots a minimal real Loader/app/local-bash composition, and executes the real search plugin against a deterministic `rg` process fixture; its result spans four top-level entries instead of returning one subtree's head.

.agents/notes/implemented/bug-fix/2026-07-27-glob-sampling.zh.md

@@ -0,0 +1,47 @@
+# Agent Note: 跨目录树采样超出上限的 glob 结果
+
+Status: implemented
+
+[English](2026-07-27-glob-sampling.md) | 中文
+
+## 问题
+
+用户询问工作区包含什么内容时，一个 agent（智能体）把某个子文件夹描述成了整个项目。该工作区有 22 个顶层条目和 11,485 个文件。`glob {"pattern":"*"}` 匹配到 10,030 条路径，但内联显示的 100 条路径全部位于一棵近期解压的子树中，因此模型完全没有看到其余 21 个条目。
+
+三个单独看都合理的行为叠加后造成了错误印象。不含 `/` 的 glob 会匹配任意深度的文件名，因此 `*` 表示目录树中的每个文件，而不是 shell 对当前目录执行的展开。Ripgrep 的 `--sort=modified` 按升序排列，因此归档包还原出的旧时间戳会让该子树排在最前。随后，内联页面直接截取这一顺序的前部，却没有说明它只代表集中于一处的切片。
+
+## 决策
+
+未超过 `globMaxResults` 的结果仍保持完整，且按修改时间排序的内容逐字节不变。必填的 `sampleOverCapGlobResults` 配置没有回退值：`false` 会为超过上限的结果保留按修改时间排序的前部，`true` 则会在完整结果的顶层条目之间按轮转方式采样。采样模式下，每个条目都先获得一个位置，之后才有条目获得第二个位置；已经用尽的分组会退出轮转；各组内部的相对顺序保持稳定；分组以实际搜索根为基准，显式指定 `path` 时也如此。
+
+采样模式下，footer 会说明当前页面是跨条目的样本，而不是按修改时间排序的前部；当触达的顶层条目数能提供额外信息时，还会报告该数量。若顶层条目数量超过内联位置数，footer 会要求模型缩小 `path`。保留前部的模式沿用达到上限时的普通 footer。spill 成功时，两种模式都会在该产物中保留完整排序列表。
+
+提示词与 schema 会说明配置所指定的超限结果排序方式、不含 `/` 的模式会匹配任意深度，以及 glob 只返回文件而绝不返回目录条目。随产品交付的 CLI（命令行界面）组合显式选择保留前部的模式；希望达到上限的页面具有代表性的部署则选择采样模式。在向模型暴露 bash 工具的部署中，目录定位仍由普通 shell 操作完成：查看一个目录使用 `ls`，跨目录树按指定文件路径模式查找则使用 glob。skill（技能）发现流程仍将 `ctx.fs.listDir` 作为内部提供方原语使用；本决策不会新增面向模型的 `list` 工具。
+
+## 考虑过的替代方案
+
+**只保留按修改时间排序的前部。** 测量实际故障形态后否决。某些部署需要这种稳定排序；但重视工作区定位的部署可以显式选择具有代表性的数据，而不必要求模型怀疑自己拿到的唯一一批路径。
+
+**为采样选项提供默认值。** 否决。没有全产品范围的证据支持把任一排序作为隐式契约，因此每个组合都必须选择一种，配置错误则在加载时失败。
+
+**对所有结果采样。** 否决。完整结果没有因截断损失任何信息，因此按修改时间排序仍有助于回答关注新旧时间的问题。只有当截取前部已经无法描述整体时，才开始采样。
+
+**改为最新优先排序。** 否决。这只会改变哪一棵结果集中的子树可能占据主导；既取消了现有的最旧优先契约，也没有让受限页面更具代表性。
+
+**仅在偏斜超过阈值时采样。** 否决。目前没有证据支持适用于所有部署的统一阈值，模型也无法判断当前采用的是哪一种排序契约。现有上限是可以清楚解释的切换点。
+
+**在顶层以下递归平衡。** 暂缓。按第一路径段做平衡已经修复观测到的故障；更深层的平衡需要一套尚无依据的深度与广度取舍策略。
+
+**新增面向模型的 `list` 工具。** 实现评审后否决。默认编程组合已经提供通用 bash，模型也理解 `ls`；重复工具会永久增加 schema 与提示词所占的 token，并引入排序、分页、符号链接、转义、UI 与快照契约，却没有独立的安全或策略收益。不向模型提供 bash 工具的精简部署也不会因本次改动获得目录定位能力。
+
+**拒绝 `*`，或在不含分隔符的模式前静默加上根目录锚点。** 否决。同样的「在任意深度匹配文件名」行为使 `*.ts` 可以有效地跨目录树搜索。记录这条规则能够保留正常工作的 Ripgrep 语义。
+
+## 影响
+
+采样模式下，超过上限的 glob 页面无法再根据内联路径回答按时间判断新旧的问题；footer 会明确说明这一点，spill 产物仍保留完整的排序视图。采样只平衡搜索根下的第一路径段，因此某个顶层条目内部较深处、结果密集的子树仍可能占据主导。保留前部的模式则把集中风险作为显式部署取舍保留下来。
+
+工具接口不会扩大。每个组合都必须设置 `sampleOverCapGlobResults`；更改该值会改变 glob 的提示词、schema 描述以及超过上限时的 Native 渲染。规范输出保留 `root`，以便采样模式恢复其分组基准；未超过上限的结果保持不变。
+
+## 测试
+
+包测试锁定了必填配置、两种超过上限模式及其提示词和 schema 描述、结果集中与扁平两种情况、显式根目录、分组数超过 JavaScript 参数个数上限、分组耗尽、位置数少于分组数，以及工作目录以外的路径。`fs-glob-sampling` ACP（Agent Client Protocol）场景会显式启用采样，启动最小化的真实 Loader/app/local-bash 组合，并让真实搜索插件对接确定性的 `rg` 进程 fixture（测试前置数据）；其结果覆盖 4 个顶层条目，而不是只返回某棵子树的前部。

.agents/notes/implemented/bug-fix/2026-07-28-themed-scrollbars-and-reserved-gutter.i18n.yaml

@@ -2,5 +2,5 @@
 # side as of the last confirmed-consistent state. Both languages carry equal authority;
 # after editing either side, bring the other along and re-record with:
 #   pnpm run verify-translation-pairing --write .agents/notes/implemented/bug-fix/2026-07-28-themed-scrollbars-and-reserved-gutter.md
-2026-07-28-themed-scrollbars-and-reserved-gutter.md: 38228c868bb00210118e8110feb722fb81d0d56c
-2026-07-28-themed-scrollbars-and-reserved-gutter.zh.md: 9fafe1faa9303b5e2e23a1b3064904f71494026d
+2026-07-28-themed-scrollbars-and-reserved-gutter.md: b45f70b126d083916c756afb88a8b646a4e9bb85
+2026-07-28-themed-scrollbars-and-reserved-gutter.zh.md: 8afa36429ce7e6e061b014d63dcb20e5a642a84c

.agents/notes/implemented/bug-fix/2026-07-28-themed-scrollbars-and-reserved-gutter.md

@@ -20,13 +20,13 @@ 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. Eight surfaces rebind today: the command popup, the slash menu, the model-select panel, the settings panel, the shared `ui-primitives` menu card, the composer input card, the question composer card, and the todo panel. Most declare it on the elevated card rather than on the scrolling descendant, because the elevation is a property of the surface and custom properties inherit down 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. 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.
 
-The last four were missed in the first implementation and found in review, which is why the rebinding contract is now checked mechanically rather than by inspection: a sheet that scrolls somewhere and paints an elevated surface somewhere must rebind.
+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.
 
 The elevated set is resolved from the palette's own dark elevation ladder — the surface tokens whose dark value lands on `bg-layer-2` or `bg-layer-3`, which is the step the l1/l2 split encodes. Deriving it instead from the sheets that already rebind was the first attempt and is unsound: such a set can only confirm what someone already remembered, and a surface nobody has rebound yet — exactly the case the check exists for — defines itself as unelevated. `--dsw-specific-tip` proved it, resolving to the menu surface's rung while the todo panel scrolled on it unrebound and the derived check stayed green.
 
-Scope is by token family, not by geometry: only `--dsw-alias-bg-*` and `--dsw-specific-*` name a surface. `--dsw-alias-button-*`, `--dsw-alias-interactive-*`, and `--dsw-alias-markdown-*` reach the same rungs while naming a control or an inline span that no scroll container renders its bar against. Shape cannot make that call, since a floating button legitimately carries a radius, a shadow, and a fixed size. The check is per sheet rather than per rule because the card and the descendant that scrolls are separate rules, and CSS text does not express which contains which.
+Scope is by token family, not by geometry: only `--dsw-alias-bg-*` and `--dsw-specific-*` name a surface. `--dsw-alias-button-*`, `--dsw-alias-interactive-*`, and `--dsw-alias-markdown-*` reach the same rungs while naming a control or an inline span that no scroll container renders its bar against. Shape cannot make that call, since a floating button legitimately carries a radius, a shadow, and a fixed size. The check is per sheet rather than per rule because the card and the descendant that scrolls are separate rules. That approximation cannot detect a scrolling component embedded in an elevated card painted by another package's stylesheet, as `DirectoryBrowser` inside `Modal` demonstrated; cross-sheet composition remains a review and assembled-UI responsibility.
 
 The track and the corner stay transparent, so the thumb reads against whatever surface scrolls under it; only the thumb and its hover state carry a token color.
 

.agents/notes/implemented/bug-fix/2026-07-28-themed-scrollbars-and-reserved-gutter.zh.md

@@ -20,13 +20,13 @@ 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。目前有八处抬升表面做了重新绑定：命令浮层、斜杠菜单、模型选择面板、设置面板、`ui-primitives` 共用菜单卡片、输入条卡片、提问组件卡片与待办面板。多数把声明写在抬升卡片上而非滚动的后代元素上，因为抬升层级是这个表面的属性，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。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。
 
-后四处在最初的实现里被漏掉、由评审发现，因此重新绑定契约现在由机械检查把关，而不再依赖人工审阅：一张样式表只要在某处滚动、又在某处绘制抬升表面，就必须重新绑定。
+`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。
 
 抬升表面集合是从调色板自身的暗色抬升阶梯解析出来的——暗色取值落在 `bg-layer-2` 或 `bg-layer-3` 上的那些表面 token，而这一档正是 l1/l2 之分所编码的层级差。最初的做法是从已经做了重新绑定的样式表反向推导，那是不成立的：这样得到的集合只能确认别人已经记得的部分，而尚无人重新绑定的表面——恰恰就是这项检查存在的理由——会把自己定义成「非抬升」。`--dsw-specific-tip` 证明了这一点：它解析到与菜单表面相同的那一档，待办面板在它上面滚动却没有重新绑定，而推导式的检查依然是绿的。
 
-判定范围依据 token 家族而非几何形状：只有 `--dsw-alias-bg-*` 与 `--dsw-specific-*` 表述的是表面。`--dsw-alias-button-*`、`--dsw-alias-interactive-*` 与 `--dsw-alias-markdown-*` 会落到相同档位，但它们表述的是控件或行内片段，没有任何滚动容器会把滚动条画在它们之上。形状无法做这个判断，因为悬浮按钮本来就会带圆角、阴影和固定尺寸。这项检查以样式表为粒度而非以规则为粒度，因为卡片与真正滚动的后代元素是两条不同的规则，而 CSS 文本无法表达谁包含谁。
+判定范围依据 token 家族而非几何形状：只有 `--dsw-alias-bg-*` 与 `--dsw-specific-*` 表述的是表面。`--dsw-alias-button-*`、`--dsw-alias-interactive-*` 与 `--dsw-alias-markdown-*` 会落到相同档位，但它们表述的是控件或行内片段，没有任何滚动容器会把滚动条画在它们之上。形状无法做这个判断，因为悬浮按钮本来就会带圆角、阴影和固定尺寸。这项检查以样式表为粒度而非以规则为粒度，因为卡片与真正滚动的后代元素是两条不同的规则。这种近似检查无法检测嵌在由另一个包的样式表绘制的抬升卡片中的滚动组件，`Modal` 内的 `DirectoryBrowser` 就证明了这一点；跨样式表的组合仍需在评审和组装后 UI 层面把关。
 
 轨道与两条滚动条相交的角落保持透明，因此滑块是以其下滚动的任何表面为背景被看到；只有滑块及其 hover 状态带 token 颜色。
 

.agents/notes/implemented/bug-fix/2026-07-29-human-transcript-append-origin.i18n.yaml

@@ -0,0 +1,6 @@
+# Bilingual-pair consistency record (docs/i18n/README.md): the git blob hash of each
+# side as of the last confirmed-consistent state. Both languages carry equal authority;
+# after editing either side, bring the other along and re-record with:
+#   pnpm run verify-translation-pairing --write .agents/notes/implemented/bug-fix/2026-07-29-human-transcript-append-origin.md
+2026-07-29-human-transcript-append-origin.md: a296b93d538d9c28bd61ee8fd0530863b4bfd878
+2026-07-29-human-transcript-append-origin.zh.md: 96e0cd1038fe8904dfd4c1eceaae9b25339c5dca

.agents/notes/implemented/bug-fix/2026-07-29-human-transcript-append-origin.md

@@ -0,0 +1,53 @@
+# Agent Note: The human transcript projects append-origin events
+
+Status: implemented
+
+English | [中文](2026-07-29-human-transcript-append-origin.zh.md)
+
+## Problem
+
+The terminal and the host history gateway both treated the model-visible surface as the human transcript. A successful compaction replaces a surface range with one checkpoint node, so the moment that replacement landed the terminal dropped every message it shadowed — conversation the user had already read — and re-ran that destructive rebuild on any later replacement. The same confusion reached pagination: `maxMessages` counted every `user/message`, `assistant/message`, and `steering/message` in the window, so a model-only replacement copy consumed a page slot the human never filled, and the cut could land between a compaction's log-only provenance and the replacement that cites it.
+
+Nothing was lost from the log. `Session.events` still held every original message and full tool result; the surface only decides what the model is sent next. The defect was entirely in the projection.
+
+## Decision
+
+Model and human projections are separate, and the event's own marker decides which one an event belongs to. `dsh-session` exports the marker split `isAppendSurfaceEvent(event)` and `isReplacementSurfaceEvent(event)` over the two `SurfaceOp` variants, from the browser-safe `surface` module. Append-origin events are the durable source for a transcript; replacement copies stay model-only. Everything that must send exactly what the model sees — `deriveMessages`, token accounting, the compaction backends, tool pairing, injected-context liveness, cross-session reference projection — keeps reading `session.surface`.
+
+The terminal replays the transcript from append-origin surface events and keeps a shadowed step's tool cards paired through `transcriptToolCallIds`, which reads the append-origin `assistant/message` rather than surface membership. A landed compaction contributes one dim `… earlier context was compacted …` row at its own log position: the marker reports where the model stopped seeing that history instead of erasing it. The framed checkpoint payload never renders, and both paths classify a surface event by the same marker, so a compaction that arrives live and the same log replayed after resume produce the same transcript. Only replay re-derives `tool/call` pairing: a call event carries no marker of its own and inherits membership from the `assistant/message` that advertised it, which the live listener has necessarily just rendered.
+
+A checkpoint is recognized through the compaction seam's own contract — `isCompactCheckpointSource`, the backend-independent marker `CompactService` requires on the replacement user message — so the terminal depends on the declared vocabulary, not on the shape of the replacement. `dsh-session-reference` already consumes that predicate to project another session's log; this is the same question asked by a different reader. Other replacements are silent: a pruned `tool/result` and a regenerated `assistant/message` rewrite one node for the model and mark no boundary in the conversation.
+
+`session.history` counts only append-origin messages toward `maxMessages`. Each page remains one contiguous raw event range, so a compaction's `compact/summary` provenance stays on the page of the replacement that cites it.
+
+No persisted event, RPC envelope, compaction transaction, or model-visible surface changed, and no migration is required.
+
+## Deferred
+
+The browser client still builds its conversation from the model surface through `FoldAdapter`, so compaction still collapses web history to a single context row. The same predicate is the fix there, together with an append-order transcript projection and a marker component; that work is a separate change against `packages/client/runtime` and `packages/client/ui-conversation`.
+
+That work must handle a page whose checkpoint cites a `surfaceOp.start` outside the window: pagination no longer spends quota on the checkpoint, so it never cuts on the checkpoint's provenance group, and `FoldAdapter` pads absent events with a non-surface sentinel — so `SurfaceManager` rejects the range and `nodes()` falls back to `degradedSeqs()` with a logged error. The hole predates this change (counting could already run past a checkpoint into the range it shadows), but the old rule accidentally covered the case where the checkpoint was the oldest counted message and pulled the whole shadowed range onto its page. `degradedSeqs()` — every surface-eligible event in append order — is already close to the transcript projection A2 needs, which is the shape to build deliberately rather than reach as a degradation. Rendering compaction *progress* — a terminal indicator while a compaction runs — needs the bracket-first ordering that the queued manual `/compact` work introduces, and is likewise out of scope here. The marker also carries no scale: the checkpoint's `sourceEventSeqs` already hold the shadowed count, so a count or range would tell a reader how much each row folded. That belongs with progress, where the reader meets the other half of the same information. Whoever takes it should fold the terminal's two replacement branches — replay and the live listener, textually identical and 600 lines apart — into one `renderReplacement(event)` first, so the marker's content has a single home.
+
+## Alternatives considered
+
+**Recognize a checkpoint by shape (a replacement `user/message`).** Rejected: it reads a coincidence of today's producers instead of a declared contract, and any future producer that replaces a range with a user message would silently inherit the compaction marker. The seam already publishes `COMPACT_CHECKPOINT_SOURCE` precisely so consumers can recognize a checkpoint independently of the backend.
+
+**Keep rendering the checkpoint as an injected-context card.** Rejected: the framed checkpoint is an instruction envelope written for the model, not human conversation content. Showing it while hiding the history it replaced inverts what the reader needs.
+
+**Persist a second display transcript.** Rejected: the append-only log already contains the authoritative source material, so a parallel record buys nothing and adds migration and consistency work.
+
+**Derive the marker from the `compact/*` bracket instead of the checkpoint.** Rejected for the transcript: the bracket is a pair of time-point markers around an operation, while the transcript needs the position where the surface actually changed. The bracket is the right source for progress and duration, which this change does not render.
+
+**Classify events by re-folding the log, as `session-query` does for search (`current` / `shadowed` / `log-only`).** Rejected: a fold answers a whole-log question, while a projection asks a per-event one that the event's own marker already answers in constant time.
+
+## Consequences
+
+Compaction no longer erases terminal history; a session compacted several times shows one marker per landed compaction, in log order. Pagination pages can carry more raw events than before, because quota is spent only on messages a human or model actually produced.
+
+`rebuildTranscript` now materializes a component per append-origin event in the whole log, and it runs on mount, on a terminal color-scheme change, and on every reasoning toggle. Compaction used to bound that work for exactly the long sessions compaction serves, so the cost now grows with session length instead of with the surface. That is the trade the fix exists to make — preserved history is the point — but a windowing or reuse strategy belongs to whoever first measures a slow rebuild, not to a later profiler wondering why the work grew.
+
+`dsh-tui` gains a dependency on the `dsh-compact` seam for one pure predicate, mirroring `dsh-session-reference`'s existing use. The terminal still needs no compaction backend at runtime.
+
+Two behaviors changed with their tests. The surface-replacement terminal test previously pinned erasure ("hides shadowed tool calls") and now pins preservation plus exactly one marker, including a pruned result copy, a regenerated assistant message, and a foreign plugin's replacement all rendering nothing. The compaction snapshot scenario wrote a `workspace-context` source while claiming to pin compaction; it now writes a real checkpoint source, and its three fixtures are re-recorded to show the preserved prompt, the full tool card, and the marker.
+
+The live/replay equivalence above is fixture-pinned, not only asserted here: `surface-replayed-compaction` mounts with the replacement already stored and records byte-identical to the live path's `surface-after-compaction-wide`. Changing either path breaks that equality, which is the point — the resume projection is what regressed for users, and the two fixtures must move together.

.agents/notes/implemented/bug-fix/2026-07-29-human-transcript-append-origin.zh.md

@@ -0,0 +1,53 @@
+# Agent Note: 人类可读记录投影追加来源的事件
+
+Status: implemented
+
+[English](2026-07-29-human-transcript-append-origin.md) | 中文
+
+## Problem
+
+终端与宿主历史网关都把模型可见的 surface 当作人类可读记录（transcript）。一次成功的压缩（compaction）会用一个检查点节点替换一段 surface 范围，因此该替换一落地，终端就丢弃了它所遮蔽的每条消息——那些是用户已经读过的对话——并在此后任何替换到来时重新执行这次破坏性重建。同样的混淆也波及分页：`maxMessages` 统计窗口内的每个 `user/message`、`assistant/message` 和 `steering/message`，于是仅供模型使用的替换副本占用了一个人类从未填充的页面额度，而切分点还可能落在压缩的仅日志溯源信息与引用它的替换之间。
+
+日志本身没有丢失任何内容。`Session.events` 仍保存着每条原始消息和完整的工具结果；surface 只决定接下来发送给模型的内容。缺陷完全在投影层。
+
+## Decision
+
+模型投影与人类投影是分开的，而事件属于哪一种由事件自身的标记决定。`dsh-session` 在浏览器安全的 `surface` 模块中导出按两种 `SurfaceOp` 变体划分的谓词 `isAppendSurfaceEvent(event)` 与 `isReplacementSurfaceEvent(event)`。追加来源的事件是记录的持久来源，替换副本仅供模型使用。凡是必须准确发送模型所见内容的部分——`deriveMessages`、token 记账、压缩后端、工具配对、注入上下文的存活判断、跨会话引用投影——都继续读取 `session.surface`。
+
+终端从追加来源的 surface 事件回放记录，并通过 `transcriptToolCallIds` 让被遮蔽步骤的工具卡片保持配对：该函数读取追加来源的 `assistant/message`，而不是 surface 成员关系。已落地的压缩会在其自身日志位置贡献一行暗色 `… earlier context was compacted …`：这行标记报告模型从何处起不再看到那段历史，而不是把它抹掉。带框的检查点载荷从不渲染，且两条路径都按同一个标记对 surface 事件分类，因此实时到达的压缩与恢复后回放同一份日志会产生相同的记录。只有回放会重新推导 `tool/call` 的配对关系：调用事件自身不携带标记，其归属继承自公布它的 `assistant/message`，而实时监听器必然刚刚渲染过后者。
+
+检查点通过压缩接缝自身的契约来识别——`isCompactCheckpointSource`，即 `CompactService` 要求替换用户消息携带的、与后端无关的标记——因此终端依赖的是已声明的词汇，而不是替换的形态。`dsh-session-reference` 已经在用该谓词投影另一个会话的日志；这里只是另一个读者提出同样的问题。其他替换保持静默：被裁剪的 `tool/result` 与重新生成的 `assistant/message` 只是为模型重写一个节点，并不在对话中标出边界。
+
+`session.history` 只把追加来源的消息计入 `maxMessages`。每一页仍是一段连续的原始事件区间，因此压缩的 `compact/summary` 溯源信息会与引用它的替换留在同一页。
+
+持久事件、RPC 信封、压缩事务与模型可见的 surface 都没有变化，也不需要迁移。
+
+## Deferred
+
+浏览器客户端仍通过 `FoldAdapter` 从模型 surface 构建会话，因此压缩在 Web 端仍会把历史折叠成一行上下文。那里的修复用的是同一个谓词，另需按追加顺序的记录投影与一个标记组件；该工作是针对 `packages/client/runtime` 与 `packages/client/ui-conversation` 的独立变更。
+
+该工作必须处理这样一页：其检查点引用的 `surfaceOp.start` 落在窗口之外。分页不再为检查点消耗额度，因此永远不会按检查点的溯源分组切分；而 `FoldAdapter` 会用一个非 surface 的哨兵事件填补缺失事件——于是 `SurfaceManager` 拒绝该范围，`nodes()` 退化为 `degradedSeqs()` 并记录一条错误。这个缺口早于本次变更（此前计数就可能越过检查点进入它所遮蔽的范围），但旧规则恰好覆盖了这样一种情形：检查点是最旧的被计数消息，其溯源分组把整段被遮蔽的范围一起拉到该页。`degradedSeqs()`——按追加顺序的每个 surface 可入事件——已经很接近 A2 所需的记录投影，因此那正是应当刻意构建的形态，而不是作为退化路径被动落到的结果。渲染压缩*进度*——压缩运行期间的终端指示——需要排队式手动 `/compact` 工作引入的“先开括号”顺序，同样不在本次范围内。标记同样不携带规模信息：检查点的 `sourceEventSeqs` 已经包含被遮蔽的数量，因此一个计数或区间可以告诉读者每一行折叠了多少内容。这件事属于进度那一侧，读者正是在那里遇到同一份信息的另一半。接手者应当先把终端里两处替换分支——回放与实时监听器，文本完全相同却相隔 600 行——合并为一个 `renderReplacement(event)`，让标记的内容只有一个归处。
+
+## Alternatives considered
+
+**按形态识别检查点（一个替换型 `user/message`）。** 被否决：那读取的是当前生产者的巧合而非已声明的契约，而未来任何用用户消息替换一段范围的生产者都会静默地继承压缩标记。接缝已经发布 `COMPACT_CHECKPOINT_SOURCE`，正是为了让消费方与后端无关地识别检查点。
+
+**继续把检查点渲染为注入上下文卡片。** 被否决：带框的检查点是为模型撰写的指令信封，不是人类对话内容。展示它却隐藏它替换掉的历史，正好颠倒了读者的需要。
+
+**持久化第二份展示用记录。** 被否决：仅追加的日志已经包含权威源材料，平行记录换不来任何东西，反而增加迁移与一致性工作。
+
+**用 `compact/*` 括号而不是检查点来推导标记。** 就记录而言被否决：括号是围绕一次操作的一对时间点标记，而记录需要的是 surface 真正发生变化的位置。括号适合作为进度与耗时的来源，而本次变更并不渲染这些。
+
+**像 `session-query` 为搜索所做的那样重新折叠日志来分类事件（`current`／`shadowed`／`log-only`）。** 被否决：折叠回答的是整份日志的问题，而投影问的是逐事件的问题，事件自身的标记已能以常数时间给出答案。
+
+## Consequences
+
+压缩不再抹掉终端历史；被压缩多次的会话会按日志顺序显示每次落地压缩对应的一行标记。分页的每一页可以携带比以前更多的原始事件，因为额度只花在人类或模型真正产生的消息上。
+
+`rebuildTranscript` 现在会为整份日志中的每个追加来源事件物化一个组件，并在挂载时、终端配色方案变化时以及每次切换 reasoning 时运行。压缩此前正好为压缩所服务的那些长会话限制了这项工作量，因此这份开销现在随会话长度增长，而不再随 surface 增长。这正是本次修复要做的取舍——保留历史才是目的——但窗口化或复用策略属于第一个真正测到重建变慢的人，而不属于日后某个疑惑工作量为何增长的性能分析者。
+
+`dsh-tui` 为一个纯谓词新增了对 `dsh-compact` 接缝的依赖，与 `dsh-session-reference` 现有用法一致。终端在运行时仍然不需要任何压缩后端。
+
+两项行为随其测试一起改变。表层替换的终端测试此前钉住的是抹除（“隐藏被遮蔽的工具调用”），现在钉住的是保留加恰好一行标记，其中被裁剪的结果副本、重新生成的 assistant 消息以及来自其他插件的替换都不渲染任何内容。压缩快照场景此前声称钉住压缩，却写入了 `workspace-context` 来源；现在它写入真实的检查点来源，并重新录制三份 fixture，以显示被保留的提示、完整的工具卡片和那行标记。
+
+上文的实时／回放等价性由 fixture 钉住，而不只是在此断言：`surface-replayed-compaction` 在挂载时替换已经存在，其录制结果与实时路径的 `surface-after-compaction-wide` 逐字节一致。改动任一路径都会破坏这项相等——这正是要点：回放投影才是当初对用户造成回归的部分，两份 fixture 必须一起变动。

.agents/notes/implemented/bug-fix/2026-07-30-approval-panel-command-cap.i18n.yaml

@@ -0,0 +1,6 @@
+# Bilingual-pair consistency record (docs/i18n/README.md): the git blob hash of each
+# side as of the last confirmed-consistent state. Both languages carry equal authority;
+# after editing either side, bring the other along and re-record with:
+#   pnpm run verify-translation-pairing --write .agents/notes/implemented/bug-fix/2026-07-30-approval-panel-command-cap.md
+2026-07-30-approval-panel-command-cap.md: 941f7eda187f263f2d8af6aa643d493c92a3669b
+2026-07-30-approval-panel-command-cap.zh.md: 939a700934f6467947028d988da9a694169e203e

.agents/notes/implemented/bug-fix/2026-07-30-approval-panel-command-cap.md

@@ -0,0 +1,52 @@
+# Agent Note: The approval takeover shares the composer's text cap
+
+Status: implemented
+
+English | [中文](2026-07-30-approval-panel-command-cap.zh.md)
+
+## Problem
+
+The approval panel is a composer takeover: while a sandbox escalation waits, it replaces the InputBar in the composer seat with the model's justification, the paired command, and a refuse/allow row. Both texts are unbounded model output, and the card had no height cap. A long command — the realistic shape, since escalation happens on the command the sandbox just denied, and a denied command is often a long inline write — grew the card until the action row left the viewport. The user could read the request and not answer it: the buttons existed, off screen, in a sticky footer that had already used the whole column.
+
+The InputBar the panel replaces has always been capped (14 lines, then the textarea scrolls), so the takeover was also the one composer state that could grow without limit — the seat's height jumped on election and jumped back on answer.
+
+## Decision
+
+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 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.
+
+The panel's card rebinds `--dsh-scrollbar-thumb{,-hover}` to the l2 pair, as every scrolling surface on an elevated background must ([scrollbar contract](../../../../packages/client/ui-theme/src/styles/scrollbar.css)).
+
+## Alternatives considered
+
+**Cap the whole card instead of the text region.** One declaration, no restructuring, and it reads as the literal "same max height as the input box". Rejected because the card holds the strip and the action row: at 336px total the justification and command would get ~250px, less room than the draft they replace, and the numbers would only agree by coincidence of the strip's height. Capping the text region makes both seats top out at the same text height, which is the property that keeps the footer from jumping.
+
+**Cap against the viewport like the question composer (`min(60vh, 520px)`).** The sibling takeover already does this, so it is the local precedent. Rejected because the designer's request was parity with the InputBar, and the two takeovers are not the same shape: the question composer's scroll content is a list of options the user must compare, which wants as much viewport as it can get, while the approval panel's is one command the user skims before deciding. A viewport-relative cap would also make the seat's height jump on election again, in the other direction.
+
+**Ellipsize or truncate the command.** No scroll region, no cap, and the buttons stay put. Rejected because the command is the thing being approved: hiding its tail asks the user to consent to text they cannot read. Truncation is also unrecoverable here — the panel is the whole approval UI, so there is no "show more" surface to fall back to.
+
+**Leave the action row inside the scroll region and cap the region.** Fewer moving parts than pinning the row. Rejected because it reproduces the defect inside the card: the buttons scroll out of the region, and the user has to discover a scrollbar to reach them.
+
+## Consequences
+
+- 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 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.
+
+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.
+
+Reproducing the off-screen buttons needs a card taller than the scrollport, not merely a tall card. The composer seat is `position: sticky; bottom: 0`, so while the card still fits it stays pinned to the viewport bottom and the buttons remain visible — at 900x1000 the uncapped card ate the whole transcript yet kept its action row on screen. Only once the card outgrows the scrollport does sticky stop being able to hold the bottom edge, and the row goes under.
+
+The geometry block and the golden are replay-only, so record mode reaches the fixture write instead of aborting on layout.
+
+The scenario keeps exactly one golden — the waiting panel — and asserts the answered state on the world instead (the decided outcome, the file the escalated command wrote, `DONE`, the panel gone, the composer re-enabled). An answered-transcript golden was recorded first and failed on Linux CI: the denied first attempt renders the OS's own refusal, and that text is platform-specific (`bash: notes.txt: Operation not permitted` on macOS against `bash: line 1: notes.txt: Read-only file system` on Linux). Any scenario whose transcript contains a sandbox-denied command inherits that, so the denial belongs in assertions, never in a golden.
+
+The panel ships as a client-module bundle: `pnpm run build:web` alone does not pick up a change to `ApprovalPanel.module.css` or a new `data-` hook in `ApprovalPanel.tsx` — the package build must run first, or the browser lane asserts against an older client than the tree.

.agents/notes/implemented/bug-fix/2026-07-30-approval-panel-command-cap.zh.md

@@ -0,0 +1,52 @@
+# Agent Note: 审批接管面板与输入框共用同一文本高度上限
+
+Status: implemented
+
+[English](2026-07-30-approval-panel-command-cap.md) | 中文
+
+## 问题
+
+审批面板是一次 composer 接管：当一次沙箱越权申请处于等待状态时，它在 composer 容器中取代 InputBar，展示模型给出的理由、与之配对的命令，以及一行拒绝／允许按钮。这两段文本都是长度不受限的模型输出，而卡片当时没有任何高度上限。命令一长——而这正是现实中的常见形态，因为越权申请针对的就是沙箱刚刚拒绝的那条命令，而被拒绝的命令往往是一次很长的内联写入——卡片就会一直变高，直到操作按钮行离开视口。用户能读到这次申请，却无法回应它：按钮存在，只是在屏幕之外，位于一个已经占满整列的吸底容器里。
+
+被它取代的 InputBar 一直是有上限的（14 行，之后由 textarea 自行滚动），因此这次接管也是 composer 唯一一个可以无限增高的状态——被选中时容器高度骤增，回应之后又骤降。
+
+## 决策
+
+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。
+
+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。
+
+面板卡片把 `--dsh-scrollbar-thumb{,-hover}` 重新绑定到 l2 那一对，这是每一个位于高层表面上的滚动区域都必须做的（[滚动条约定](../../../../packages/client/ui-theme/src/styles/scrollbar.css)）。
+
+## 曾考虑的替代方案
+
+**给整张卡片设上限，而不是给文本区域设。** 一条声明，不需要重构结构，而且它读起来就是字面意义上的"与输入框相同的最大高度"。之所以否决：卡片还装着状态条和操作按钮行——总高 336px 时，理由与命令只能分到约 250px，比它们所取代的草稿区更矮，而且两边数字能对上纯属状态条高度的巧合。给文本区域设上限，才能让两种状态在同一文本高度处收住，而这正是让底部不再跳动的那条性质。
+
+**像提问 composer 那样按视口设上限（`min(60vh, 520px)`）。** 同为接管面板的兄弟组件已经这么做了，因此这是本地既有先例。之所以否决：设计同学的要求是与 InputBar 对齐，而两个接管面板形态并不相同——提问 composer 的滚动内容是一组需要用户互相比较的选项，能占多少视口就该占多少；审批面板的滚动内容则是一条命令，用户在决定之前扫读即可。按视口设上限还会让容器高度在被选中时再次跳动，只是方向相反。
+
+**对命令做省略号或截断处理。** 不需要滚动区域，不需要上限，按钮也不会移位。之所以否决：命令正是被审批的对象，隐去它的尾部等于要求用户为自己读不到的文本背书。在这里截断还是不可恢复的——面板就是审批的全部界面，没有"展开更多"的落脚处。
+
+**把操作按钮行留在滚动区域内，只给该区域设上限。** 比把按钮行固定住少动几处。之所以否决：这会把缺陷搬进卡片内部——按钮滚出该区域，用户得先发现有滚动条才能碰到它们。
+
+## 后果
+
+- 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。
+- 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。
+
+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。
+
+要复现按钮跑到屏幕外，需要的是比滚动视口更高的卡片，而不只是一张很高的卡片。composer 容器为 `position: sticky; bottom: 0`，因此在卡片尚能容纳时它会一直吸附在视口底部，按钮仍然可见——在 900x1000 下，未设上限的卡片吃掉了整个会话流，却仍把操作按钮行留在屏幕内。只有当卡片长过滚动视口，sticky 才再也无法守住底边，按钮行随之沉入折叠之下。
+
+几何断言块与 golden 仅在回放模式下执行，这样录制模式才能走到写入 fixture 那一步，而不是在布局检查处中断。
+
+该场景只保留一份 golden —— 等待中的面板；回应之后的状态改为对世界作断言（决策结果、越权命令写出的那个文件、`DONE`、面板消失、输入框重新可用）。最初还录了一份"已回应会话流"的 golden，它在 Linux CI 上失败了：第一次被拒绝的尝试渲染的是操作系统自己的拒绝文本，而这段文本因平台而异（macOS 为 `bash: notes.txt: Operation not permitted`，Linux 为 `bash: line 1: notes.txt: Read-only file system`）。任何会话流中含有被沙箱拒绝命令的场景都会继承这一点，因此这类拒绝只能进断言，绝不能进 golden。
+
+该面板以客户端模组包的形式发布：单跑 `pnpm run build:web` 不会带上对 `ApprovalPanel.module.css` 的改动，也不会带上 `ApprovalPanel.tsx` 中新增的 `data-` 钩子——必须先执行包构建，否则浏览器测试通道会对着一个比工作树更旧的客户端做断言。

.agents/notes/implemented/bug-fix/2026-07-31-composer-glyph-layer-tracks-the-textarea.i18n.yaml

@@ -0,0 +1,6 @@
+# Bilingual-pair consistency record (docs/i18n/README.md): the git blob hash of each
+# side as of the last confirmed-consistent state. Both languages carry equal authority;
+# after editing either side, bring the other along and re-record with:
+#   pnpm run verify-translation-pairing --write .agents/notes/implemented/bug-fix/2026-07-31-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

.agents/notes/implemented/bug-fix/2026-07-31-composer-glyph-layer-tracks-the-textarea.md

@@ -0,0 +1,77 @@
+# 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.

.agents/notes/implemented/bug-fix/2026-07-31-composer-glyph-layer-tracks-the-textarea.zh.md

@@ -0,0 +1,77 @@
+# 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。