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

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

.agents/notes/implemented/architecture/2026-06-21-bounded-llm-request-recovery.i18n.yaml

@@ -2,5 +2,5 @@
 # side as of the last confirmed-consistent state. Both languages carry equal authority;
 # after editing either side, bring the other along and re-record with:
 #   pnpm run verify-translation-pairing --write .agents/notes/implemented/architecture/2026-06-21-bounded-llm-request-recovery.md
-2026-06-21-bounded-llm-request-recovery.md: 83d47e3a7d91bbcd2ceaf7b11cf13316142eb3ed
-2026-06-21-bounded-llm-request-recovery.zh.md: 00dcbad3d1023ad33a22297bfe938b94bce839d4
+2026-06-21-bounded-llm-request-recovery.md: 5c76ed5d754ea40f41dff78cb56ee7fc139a32b1
+2026-06-21-bounded-llm-request-recovery.zh.md: 1fa56f3fe0405cab663c2843d423a78d910170dd

.agents/notes/implemented/architecture/2026-06-21-bounded-llm-request-recovery.md

@@ -60,11 +60,11 @@ For an eligible failure with budget remaining, the one-based transient retry cou
 
 The plugin owns a lifetime `AbortController` and tracks every active recovery callback, including delegated waterfall work and backoff. Effect cleanup first unregisters the listener, then aborts and awaits the active callbacks; abort wins over a late delegated retry decision, and a captured callback can neither retry nor enter the rest of its waterfall after disposal. This makes HMR disposal quiescent even though Cordis has already captured the listener.
 
-Before sleeping, `dsh-llm-retry` appends one non-surface `llm/retry` session event containing the turn, failed step, provider, policy mode, complete resolved-policy key, provider-policy retry number, mode-specific finite maximum when present, scheduled delay, and `LlmFailure`. The key sorts the code set and separates retry histories when a provider route is replaced by a behaviorally different same-mode policy. The plugin owns the `SessionEventMap` augmentation; `dsh-session` remains generic persistence and does not absorb the optional policy's vocabulary. The event says what was scheduled, not that the next request completed; cancellation during the delay is subsequently visible on `turn/end`. The event ships only with a production renderer and replay/snapshot coverage, because its purpose is operational state rather than trace collection.
+Before sleeping, `dsh-llm-retry` appends one non-surface `llm/retry` session event containing the turn, failed step, provider, policy mode, complete resolved-policy key, provider-policy retry number, mode-specific finite maximum when present, scheduled delay, and `LlmFailure`. The key sorts the code set and separates retry histories when a provider route is replaced by a behaviorally different same-mode policy. The plugin owns the `SessionEventMap` augmentation and exports the payload through its browser-safe `./types` subpath; `dsh-session` remains generic persistence and does not absorb the optional policy's vocabulary. The event says what was scheduled, not that the next request completed; cancellation during the delay is subsequently visible on `turn/end`. The event ships with production renderers and replay/snapshot coverage, because its purpose is operational state rather than trace collection.
 
 The listener calls `next()` for a non-transient code, an exhausted policy budget, or an over-cap provider delay. This preserves composition with context-overflow recovery and later policy plugins. For an owned failure it records and awaits the delay, then returns `{ kind: 'retry' }` without delegating. Turn cancellation and plugin disposal end the wait without returning a retry; the loop's cancellation/disposal checks remain authoritative.
 
-The agent-spine demo bundle loads the plugin so the shared stdio/TUI, one-shot CLI, and ACP example compositions use the same provider-routed policy. Library consumers retain explicit plugin composition: omitting the plugin leaves request failures terminal.
+The agent-spine demo bundle loads the plugin so the shared stdio/TUI, one-shot CLI, ACP, and headless example compositions use the same provider-routed policy. The shipped Web composition also loads it, so browser and command-line requests use the same provider defaults. Library consumers retain explicit plugin composition: omitting the plugin leaves request failures terminal.
 
 ### Make one layer own visible attempts
 
@@ -82,7 +82,7 @@ Boundary tests prove termination at both actual transports. The hand-written ada
 
 ### Keep attempts separate in the existing log
 
-A failed attempt may leave `assistant/chunk` events in its closed step, but it never appends `assistant/message` and never dispatches a tool. A retry closes the failed turn, opens the next numbered turn, reconstructs the request from the durable surface, and produces its own chunks. UIs may render live chunks while a step is open, then mark or clear that transient view when `llm/retry` identifies the failed step or `turn/end` records failure; message derivation continues to ignore the failed chunks.
+A failed attempt may leave `assistant/chunk` events in its closed step, but it never appends `assistant/message` and never dispatches a tool. A retry closes the failed turn, opens the next numbered turn, reconstructs the request from the durable surface, and produces its own chunks. UIs may render live chunks while a step is open, then mark or clear that transient view when `llm/retry` identifies the failed step or `turn/end` records failure. Web validates the complete retry payload contract, clears the failed partial at `llm/retry`, projects consecutive retry-turn events into one stable row updated to the latest attempt, and derives scheduled, started, or cancelled status from subsequent turn facts. Its countdown anchors the scheduled delay to browser receipt rather than the Host event clock, uses ceiling-rounded seconds with a one-second floor, animates only while unresolved, and keeps exact latest failure details collapsed behind the row. Retry nodes anchor their own trajectory turn even when the failed attempt has no assistant node. Message derivation continues to ignore the failed chunks, and Web applies the same projection during history rebuild so refreshing cannot resurrect discarded partials or duplicate retry rows.
 
 If recovery is exhausted, the final failure is stored once on `turn/end.reason` with the structured facts. If transient recovery continues, `llm/retry` is the durable home for that attempt's failure and delay. No standalone final-error event or response-id vocabulary is added.
 
@@ -116,7 +116,7 @@ If recovery is exhausted, the final failure is stored once on `turn/end.reason`
 - Pure unit tests cover transient-code selection, exponential backoff and jitter bounds, valid and over-cap `Retry-After`, exhausted budgets, deterministic timer/random seams, and abort during backoff.
 - Real agent-loop tests cover failure before chunks, partial chunks then failure, thrown and in-band failures, retry to success in a new turn, exhaustion to structured `turn/end.reason`, and composition with `dsh-compact-basic` context-overflow recovery.
 - The partial-chunk integration test proves failed chunks remain attributed to the failed step, no assistant message or tool side effect is committed for that step, and the successful retry has distinct provenance.
-- The plugin-owned `llm/retry` event is non-surface, survives JSONL and SQLite round trips, is ignored by message derivation, and drives TUI retraction plus scheduled-retry rendering. Keyless snapshots cover scheduling, cancellation, success, and exhaustion; ACP automation snapshots confirm that a discarded attempt stays off the wire while the recovered reply is emitted.
+- The plugin-owned `llm/retry` event is non-surface, survives JSONL and SQLite round trips, is ignored by message derivation, and drives TUI and Web retraction plus scheduled-retry rendering. Client tests cover complete wire validation, clock-independent countdown, cancellation versus completed retry labels, and trajectory attribution; keyless UI snapshots cover Web scheduling and success, a real Web composition test covers partial transport failure through recovery, and ACP automation snapshots confirm that a discarded attempt stays off the wire while the recovered reply is emitted.
 - Idle-watchdog tests prove the stable signal is rearmed only while `next()` is outstanding, disarmed during consumer think time and in `finally`, and classified separately from a total-call deadline and an earlier caller abort; adapter tests prove the signal stops the underlying request rather than merely detaching it.
 - Direct `ctx.llm.stream()` callers remain single-attempt and receive the same structured failure facts.
 

.agents/notes/implemented/architecture/2026-06-21-bounded-llm-request-recovery.zh.md

@@ -60,11 +60,11 @@ agent loop（智能体循环）会保留 `RequestError` 作为该精确的错误
 
 插件拥有一个全生命期 `AbortController`，并跟踪每个活跃的恢复回调，包括委托的 waterfall（瀑布式事件）工作与退避。effect 清理会先注销监听器，再中止并等待活跃回调；中止会胜过较晚到达的委托重试决策，被捕获的回调在插件释放后既不能重试，也不能进入其 waterfall 的剩余部分。尽管 Cordis 已捕获该监听器，此设计仍能使 HMR（热模块替换）释放达到完全停稳。
 
-休眠前，`dsh-llm-retry` 会追加一条不进入表层的 `llm/retry` 会话事件，其中包含轮次、失败步骤、提供方、策略 mode、完整的解析策略 key、提供方策略重试编号、该 mode 存在时的有限上限、计划延迟和 `LlmFailure`。该 key 会对 code 集排序，并在提供方路由被行为不同但 mode 相同的策略替换时分隔重试历史。该插件拥有 `SessionEventMap` 声明合并；`dsh-session` 继续负责通用持久化，不会吸收可选策略的词汇。事件记录已安排的内容，而不是下一个请求已完成；延迟期间取消随后会在 `turn/end` 中可见。因为该事件的目的是表示运行状态，而不是收集跟踪数据，所以它仅与生产渲染器及回放／快照覆盖一起交付。
+休眠前，`dsh-llm-retry` 会追加一条不进入表层的 `llm/retry` 会话事件，其中包含轮次、失败步骤、提供方、策略 mode、完整的解析策略 key、提供方策略重试编号、该 mode 存在时的有限上限、计划延迟和 `LlmFailure`。该 key 会对 code 集排序，并在提供方路由被行为不同但 mode 相同的策略替换时分隔重试历史。该插件拥有 `SessionEventMap` 声明合并，并通过其浏览器安全的 `./types` 子路径导出载荷；`dsh-session` 继续负责通用持久化，不会吸收可选策略的词汇。事件记录已安排的内容，而不是下一个请求已完成；延迟期间取消随后会在 `turn/end` 中可见。因为该事件的目的是表示运行状态，而不是收集跟踪数据，所以它会与生产渲染器及回放／快照覆盖一起交付。
 
 对非暂时性 code、耗尽的策略预算或超出上限的提供方延迟，监听器会调用 `next()`。这保留了与上下文溢出恢复及后续策略插件的组合能力。对自身处理的失败，它会记录并等待延迟，然后在不委托的情况下返回 `{ kind: 'retry' }`。轮次取消和插件释放会结束等待且不返回重试动作，此后仍以循环的取消／释放检查为准。
 
-agent-spine 演示组合包加载该插件，因此共享的 stdio/TUI、一次性 CLI（命令行界面）和 ACP（Agent Client Protocol）示例组合使用同一套按提供方路由的策略。库消费方仍需显式组合插件：省略该插件时，请求失败保持终态。
+agent-spine 演示组合包加载该插件，因此共享的 stdio/TUI、一次性 CLI（命令行界面）、ACP（Agent Client Protocol）和 headless 示例组合使用同一套按提供方路由的策略。随产品交付的 Web 组合也会加载该插件，因此浏览器请求与命令行请求使用相同的提供方默认值。库消费方仍需显式组合插件：省略该插件时，请求失败保持终态。
 
 ### 由单一层负责可见的尝试
 
@@ -82,7 +82,7 @@ agent-spine 演示组合包加载该插件，因此共享的 stdio/TUI、一次
 
 ### 在现有日志中分隔尝试
 
-一次失败尝试可以在已关闭的步骤中留下 `assistant/chunk` 事件，但绝不会追加 `assistant/message`，也不会分发工具。重试会关闭失败轮次，开启下一个编号轮次，从持久表层重建请求，并生成自己的分片。步骤仍处于打开状态时，UI 可以渲染实时分片；当 `llm/retry` 标识失败步骤，或 `turn/end` 记录失败时，UI 再标记或清除这份暂时视图。消息派生仍会忽略失败分片。
+一次失败尝试可以在已关闭的步骤中留下 `assistant/chunk` 事件，但绝不会追加 `assistant/message`，也不会分发工具。重试会关闭失败轮次，开启下一个编号轮次，从持久表层重建请求，并生成自己的分片。步骤仍处于打开状态时，UI 可以渲染实时分片；当 `llm/retry` 标识失败步骤，或 `turn/end` 记录失败时，UI 再标记或清除这份暂时视图。Web 会验证完整的重试载荷契约，在 `llm/retry` 到达时清除失败的部分输出，将连续重试轮次的事件投影为稳定的一行，并用最新一次尝试更新该行，再从后续轮次事实派生 scheduled、started 或 cancelled 状态。倒计时以浏览器收到事件的时刻为计划延迟的起点，而不是使用 Host 事件时钟；它按向上取整且不低于 1 秒的秒数显示，仅在重试尚未结束时显示动画，并把最近一次失败的准确详情折叠在该行之后。即使失败尝试没有 assistant 节点，重试节点也会锚定自身的轨迹轮次。消息派生仍会忽略失败分片；Web 在重建历史时也会应用同一投影，因此刷新页面不会让已丢弃的部分输出重新出现，也不会生成重复的重试行。
 
 如果恢复预算耗尽，最终失败会连同结构化事实在 `turn/end.reason` 中存储一次。如果暂时性恢复继续，`llm/retry` 就是该次尝试的失败与延迟的持久归属位置。本决策不增加独立的最终错误事件或响应 id 词汇。
 
@@ -116,7 +116,7 @@ agent-spine 演示组合包加载该插件，因此共享的 stdio/TUI、一次
 - 纯单元测试覆盖暂时性 code 选择、指数退避和抖动边界、有效及超出上限的 `Retry-After`、耗尽的预算、确定性定时器／随机数 seam，以及退避期间中止。
 - 真实 agent-loop 测试覆盖分片前失败、部分分片后失败、抛出及带内失败、在新轮次中重试至成功、耗尽后写入结构化 `turn/end.reason`，以及与 `dsh-compact-basic` 上下文溢出恢复的组合。
 - 部分分片集成测试证明：失败分片仍归属于失败步骤，该步骤不会提交 assistant 消息或工具副作用，成功的重试具有不同的来源信息。
-- 插件拥有的不进入表层的 `llm/retry` 事件可在 JSONL 和 SQLite 往返后保留，被消息派生忽略，并驱动 TUI 撤回和计划重试渲染。无密钥快照覆盖调度、取消、成功和耗尽；ACP 自动化快照确认，被丢弃的尝试不会通过协议发出，而恢复后的回复会正常发出。
+- 插件拥有的不进入表层的 `llm/retry` 事件可在 JSONL 和 SQLite 往返后保留，被消息派生忽略，并驱动 TUI 和 Web 撤回及计划重试渲染。客户端测试覆盖完整的 wire 验证、独立于时钟的倒计时、已取消与已完成重试标签的区别以及轨迹归属；无密钥 UI 快照覆盖 Web 的调度与成功，真实 Web 组合测试覆盖部分传输失败直至恢复，ACP 自动化快照确认，被丢弃的尝试不会通过协议发出，而恢复后的回复会正常发出。
 - 空闲看门狗测试证明：只有 `next()` 尚未完成时才会重新布防稳定信号；在消费方思考期间及 `finally` 中会解除布防；它与总调用 deadline 以及更早发生的调用方中止分开分类。适配器测试证明该信号会终止底层请求，而不只是与其脱离。
 - `ctx.llm.stream()` 的直接调用方仍只尝试一次，并收到相同的结构化失败事实。
 

.agents/notes/implemented/architecture/2026-07-19-gui-layering-and-rpc-protocol.i18n.yaml

@@ -2,5 +2,5 @@
 # side as of the last confirmed-consistent state. Both languages carry equal authority;
 # after editing either side, bring the other along and re-record with:
 #   pnpm run verify-translation-pairing --write .agents/notes/implemented/architecture/2026-07-19-gui-layering-and-rpc-protocol.md
-2026-07-19-gui-layering-and-rpc-protocol.md: b9718da4725316c64686adef24827e2984d8723d
-2026-07-19-gui-layering-and-rpc-protocol.zh.md: 2add148054e8f97c65600cd719fb4f8e0283f52d
+2026-07-19-gui-layering-and-rpc-protocol.md: b7081591cf7e5e3c586c74c5a71b4317376135cb
+2026-07-19-gui-layering-and-rpc-protocol.zh.md: 89557182ca7781f4fb59b8daf866aaca96cf20ee

.agents/notes/implemented/architecture/2026-07-19-gui-layering-and-rpc-protocol.md

@@ -204,7 +204,7 @@ The same domain tree as `ApiProxy`, but unary methods **take the business payloa
 | `callUnary` | mint → tap → POST full form → `serverResponseSchema` parse → **rpcId echo check** (mismatch throws) → tap → emit narrow form |
 | `readSse` | streaming fetch (not EventSource), `\n\n` framing, `data:` concatenation, ServerRequest full-form parse, tap, emit narrow `RpcRequest<frame>` |
 | `respond` | client-response passthrough (rpcId is an echo — never minted here); response body parsed by `rpcReceiptSchema` |
-| unary timeout | `AbortSignal.timeout` (default 30s, constructor-tunable); streams have no timeout (long-lived by nature) |
+| unary deadline | Ordinary unary calls use `AbortSignal.timeout` (default 30s, constructor-tunable); user-paced `host.pickDirectory` and `command.execute` omit that deadline but keep caller/connection cancellation; streams have no deadline |
 | `resolveBase` | browser = same-origin origin; no-location environment (Node) = the `http://dsh.internal` fake authority |
 
 ### The instance-level envelope observation aspect
@@ -234,7 +234,7 @@ All four quadrant full forms pass through `onEnvelope`; the base implementation
 
 ## Consequences
 
-Every client shape consumes one contract: adding a unary method is a five-step mechanical change radiating from a single signature, swapping a carrier touches only a `doFetch` subclass, and every wire message is zod-validated, observable through the envelope tap, and reconcilable by rpcId. The accepted costs: two groups of packages need explicit tsconfig paths entries, and the reserved seams (fork/inject/task.list/listModels/hostInstanceId) stay dormant until a real consumer arrives.
+Every client shape consumes one contract: adding a unary method is a five-step mechanical change radiating from a single signature, swapping a carrier touches only a `doFetch` subclass, and every wire message is zod-validated, observable through the envelope tap, and reconcilable by rpcId. Ordinary unary calls remain bounded, while `host.pickDirectory` and `command.execute` may stay pending until the operation finishes or caller/connection cancellation arrives; this accepts that a non-cooperative user-paced operation can hang its request rather than treating valid operation duration as transport failure. The other accepted costs: two groups of packages need explicit tsconfig paths entries, and the reserved seams (fork/inject/task.list/listModels/hostInstanceId) stay dormant until a real consumer arrives.
 
 ## Alternatives considered
 
@@ -252,3 +252,4 @@ Every client shape consumes one contract: adding a unary method is a five-step m
 | A DTO layer (a second wire-only structure set) | Core types reach the browser type-only at zero cost; a DTO is a permanent two-way synchronization tax |
 | Cursor resumption (implementing mux since) | Reconnect = rebuild (opencode-style) covers all v1 needs; the signature keeps the seat, implementation waits for a real consumer |
 | A createApiClient factory function (the original implementation) | Platform differences (transport/observation) are inheritance aspects, not parameters; the class family lets the fixture substitute at the protocol layer instead of wrapping a fake envelope |
+| Applying the 30-second transport deadline to `command.execute` | Command duration is operation work, not a transport-health budget; the deadline kills valid long-running handlers, while caller/connection cancellation already supplies the required stop path |

.agents/notes/implemented/architecture/2026-07-19-gui-layering-and-rpc-protocol.zh.md

@@ -202,7 +202,7 @@ export type ResponseValue<K> =
 | `callUnary` | mint → tap → POST 全形 → `serverResponseSchema` parse → **rpcId 回显校验**（不符即 throw）→ tap → 吐窄形 |
 | `readSse` | streaming fetch（非 EventSource）、`\n\n` 分帧、`data:` 拼接、ServerRequest 全形 parse、tap、吐窄形 `RpcRequest<帧>` |
 | `respond` | client-response 透传（rpcId 是回填，此处不 mint）；应答体 `rpcReceiptSchema` parse |
-| unary 超时 | `AbortSignal.timeout`（默认 30s，构造参数可调）；流不设超时（长连接本性） |
+| unary 时限 | 普通 unary 调用使用 `AbortSignal.timeout`（默认 30s，构造参数可调）；由用户掌控节奏的 `host.pickDirectory` 和 `command.execute` 不设该时限，但保留调用方／连接取消；流不设时限 |
 | `resolveBase` | 浏览器=同源 origin；无 location 环境（Node）=`http://dsh.internal` 假 authority |
 
 ### 实例级 envelope 观测切面
@@ -232,7 +232,7 @@ export type ResponseValue<K> =
 
 ## Consequences
 
-所有 client 形态消费同一契约：加一个 unary 方法是从单一签名辐射的五步机械改动，换载体只动一个 `doFetch` 子类，wire 上每条消息可 zod 校验、可经 envelope tap 观测、可按 rpcId 对账。接受的代价：两组包需要显式 tsconfig paths 条目；预留接缝（fork/inject/task.list/listModels/hostInstanceId）在真实消费者出现前保持休眠。
+所有 client 形态消费同一契约：加一个 unary 方法是从单一签名辐射的五步机械改动，换载体只动一个 `doFetch` 子类，wire 上每条消息可 zod 校验、可经 envelope tap 观测、可按 rpcId 对账。普通 unary 调用仍受时限约束，而 `host.pickDirectory` 与 `command.execute` 可保持挂起，直到操作完成或调用方／连接取消到来；若由用户掌控节奏的操作不自行结束，请求可能一直挂起，这是为避免把合理的操作时长视为传输失败而接受的代价。其余接受的代价：两组包需要显式 tsconfig paths 条目；预留接缝（fork/inject/task.list/listModels/hostInstanceId）在真实消费者出现前保持休眠。
 
 ## Alternatives considered
 
@@ -250,3 +250,4 @@ export type ResponseValue<K> =
 | DTO 层（wire 专用第二套结构） | core 类型 type-only 直达浏览器零成本；DTO 是永久的双向同步税 |
 | cursor 续传（mux since 实装） | 重连=重建（opencode 同款）覆盖 v1 全部需求；签名留座，实装等真实消费者 |
 | createApiClient 工厂函数（原实现） | 平台差异（传输/观测）是继承切面不是参数；类体系让 fixture 在协议层替换而不是包一层假信封 |
+| 对 `command.execute` 应用 30 秒传输时限 | 命令耗时属于操作本身，而非传输健康预算；该时限会终止本应继续运行的长时处理器，调用方／连接取消已提供所需的停止路径 |

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

@@ -2,5 +2,5 @@
 # side as of the last confirmed-consistent state. Both languages carry equal authority;
 # after editing either side, bring the other along and re-record with:
 #   pnpm run verify-translation-pairing --write .agents/notes/implemented/architecture/2026-07-23-client-plugin-loading-model.md
-2026-07-23-client-plugin-loading-model.md: b0873b7aa7bccd3d613f3113fa18207770952e5d
-2026-07-23-client-plugin-loading-model.zh.md: f3472dbfc5a78924e77337bf92ce5983c8492c4c
+2026-07-23-client-plugin-loading-model.md: 02347f2964942b89ec1f0a6ec483f4c2b2f9e68c
+2026-07-23-client-plugin-loading-model.zh.md: ea927d35860fbbba567c47cea0ee3a45133ce0f4

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

@@ -56,8 +56,8 @@ What happens between `dsh web` starting and the UI appearing? Three stages: the
 
 **Host side — compose the graph.**
 
-1. The composing app (`apps/cli`) ships the roster as ordinary rows in its `cordis.yml` config tree — client plugin packages are entry rows like every host plugin, and `--dev` appends the `client-hmr` row in code (`AppCLIEntry`) before the settle/sweep so the fail-loud triple covers it. A roster row that fails to import is caught by the boot's `assertEntriesLoaded`.
-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 a declared plugin without a built `./client` bundle, and any malformed declaration field — activation-time fail loud (a FAILED fiber the sweep reports).
+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).
 
 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.

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

@@ -56,8 +56,8 @@ vendored Loader 经其 `internal` seam 消费模块系统——唯一调用点
 
 **host 侧——组合这张图。**
 
-1. 负责组合的 app（`apps/cli`）把名册作为普通行放进它的 `cordis.yml` 配置树——client 插件包与每个 host 插件一样是 entry 行，`--dev` 由代码（`AppCLIEntry`）在 settle/sweep 之前追加 `client-hmr` 行，使 fail-loud 三件套一并覆盖它。名册行 import 失败由 boot 的 `assertEntriesLoaded` 捕获。
-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。它拒绝声明了插件却没有已构建 `./client` bundle 的包，也拒绝任何畸形的声明字段——激活期大声失败（FAILED fiber，由 sweep 上报）。
+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）。
 
 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。

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

@@ -2,5 +2,5 @@
 # side as of the last confirmed-consistent state. Both languages carry equal authority;
 # after editing either side, bring the other along and re-record with:
 #   pnpm run verify-translation-pairing --write .agents/notes/implemented/architecture/2026-07-24-web-config-tree-boot-and-transport-layering.md
-2026-07-24-web-config-tree-boot-and-transport-layering.md: a2080024d36d54162f4f4aa79896e51efd708f59
-2026-07-24-web-config-tree-boot-and-transport-layering.zh.md: 1c430bf2939f9f556f378bdeb78872a0091a592d
+2026-07-24-web-config-tree-boot-and-transport-layering.md: 88f94b1f58ae7a3451c7772f4a9ff7d6564254c0
+2026-07-24-web-config-tree-boot-and-transport-layering.zh.md: ea2a8f70a6c2d4207d4388a9303fbc6ce6e94238

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

@@ -12,9 +12,9 @@ English | [中文](2026-07-24-web-config-tree-boot-and-transport-layering.zh.md)
 
 ## Decision
 
-**Composition is one flat assembled tree.** `apps/cli/config/base.cordis.yml` plus `apps/cli/config/web.cordis.yml` holds every row — the host runtime (32 rows), the `api-gateway` row, the `webserver` row, and the `dshClient` rows (the browser roster; the modules row is simultaneously a host row). No spine bundle: every plugin is one row and every config field is yml-editable. That stance later became repository-wide, with the rows both surfaces share factored into `apps/cli/config/base.cordis.yml` and each surface reduced to an overlay ([shared-base overlays](../simplification/2026-07-29-shared-base-config-overlays.md)). `--dev` appends the `dsh-client-hmr` row in code before the settle sweep — prod and dev differ by exactly that row. Row order carries no load semantics; activation is service-availability driven, and the boot compensates with a fail-loud triple: `assertEntriesLoaded` (import failures), `installFailLoud` (late apply rejections), and an all-ACTIVE sweep (PENDING fibers — cordis inject waiting has no timeout).
+**Composition is one flat assembled tree.** `apps/cli/config/base.cordis.yml` plus `apps/cli/config/web.cordis.yml` holds every row — the host runtime (32 rows), the `api-gateway` row, the `webserver` row, and the `dshClient` rows (the browser roster; the modules row is simultaneously a host row). No spine bundle: every plugin is one row and every config field is yml-editable. That stance later became repository-wide, with the rows both surfaces share factored into `apps/cli/config/base.cordis.yml` and each surface reduced to an overlay ([shared-base overlays](../simplification/2026-07-29-shared-base-config-overlays.md)). `--dev` appends the `dsh-client-hmr` row in code before the settle audit — prod and dev differ by exactly that row. Row order carries no load semantics; activation is service-availability driven. The shared audit rejects imports with no fiber, awaits only failed fibers to recover original activation errors, and reports services that leave a fiber `PENDING`; before throwing, it marks those exact rejection reasons through one process checkpoint so `installFailLoud` coalesces Loader's duplicate notification while unrelated unhandled rejections remain fatal. The Node app-boot artifact embeds `@cordisjs/plugin-include` while leaving `@cordisjs/plugin-loader` external, so the include's `EntryTree` and the host bind to one Loader peer instead of splitting a config tree across two Loader implementations.
 
-**Boot glue is a class pair.** `AppCLIEntry` (apps/cli) and `AppWebEntry` (the shell kernel) hold only what must exist independently of cordis: argv facts, the composed patch set, the parsed boot manifest, the module system instance, loading-page handles — everything else lives in plugins. `AppCLIEntry.run()` is three stages: layered env (ambient > cwd `.env` > `$DSH_HOME/.env`, closing the defect above) → patch composition → Loader include boot plus the triple. `AppWebEntry.run()` mirrors it browser-side: parse `window.__DSH_BOOT__` into a `BootManifest` (two views: npm-package rows for the module table, cordis-plugin rows for entry composition; malformed wire throws), build the module system, render the loading page, prefetch the `immediately` tier in parallel with Context/Loader setup, **await the prefetch before creating entries** (materialization is `tree.import`'s synchronous require, unprotected by fiber inject waiting; cross-package require edges such as i18n → runtime/client need every immediately-tier factory registered first — an empirically found 10–25% boot race otherwise), adopt the modules entry, create the graph rows, settle, sweep.
+**Boot glue is a class pair.** `AppCLIEntry` (apps/cli) and `AppWebEntry` (the shell kernel) hold only what must exist independently of cordis: argv facts, the composed patch set, the parsed boot manifest, the module system instance, loading-page handles — everything else lives in plugins. `AppCLIEntry.run()` is three stages: layered env (ambient > cwd `.env` > `$DSH_HOME/.env`, closing the defect above) → patch composition → Loader include boot plus the activation audit. `AppWebEntry.run()` mirrors it browser-side: parse `window.__DSH_BOOT__` into a `BootManifest` (two views: npm-package rows for the module table, cordis-plugin rows for entry composition; malformed wire throws), build the module system, render the loading page, prefetch the `immediately` tier in parallel with Context/Loader setup, **await the prefetch before creating entries** (materialization is `tree.import`'s synchronous require, unprotected by fiber inject waiting; cross-package require edges such as i18n → runtime/client need every immediately-tier factory registered first — an empirically found 10–25% boot race otherwise), adopt the modules entry, create the graph rows, settle, sweep.
 
 **Config sources have one declaration place each.** yml static values are engineering defaults; the profile json (`./.dsh-tmp-profile/config.json`, read-only, never created, cwd-anchored until the `$DSH_HOME` migration) is user config mapped through a static `PROFILE_MAPPINGS` table onto target rows (`provider`/`model` → the `api-gateway` row, `persistenceRoot` → the jsonl row); CLI flags map onto the `webserver` row with a field set disjoint from the json's; env values enter through yml `!!js` expressions, never through the mapping table. Patches replace a row's config wholesale, so the entry class re-reads the yml row's static values (bypass parse) and merges overrides on top. An unmapped json key fails loud. The resolved frontend `distIndex` rides the same patch channel — an assembly fact, not user config.
 

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

@@ -12,9 +12,9 @@ Status: implemented
 
 ## 决策
 
-**组合结果是一棵平铺配置树。** `apps/cli/config/base.cordis.yml` 与 `apps/cli/config/web.cordis.yml` 共同持有全部行——host runtime（32 行）、`api-gateway` 行、`webserver` 行、`dshClient` 行（浏览器 roster；modules 行同时是 host 行）。不做 spine bundle：每插件一行、每个 config 字段 yml 可改。这一立场后来推广到全仓：两个 surface 共享的配置项被抽取进 `apps/cli/config/base.cordis.yml`，各 surface 则收敛为一份 overlay（[共享 base overlay](../simplification/2026-07-29-shared-base-config-overlays.md)）。`--dev` 在 settle sweep 之前由代码追加 `dsh-client-hmr` 行——prod 与 dev 的全部差异就是这一行。行序无装载语义；激活由服务可用性驱动，boot 以 fail-loud 三件套补偿：`assertEntriesLoaded`（import 失败）、`installFailLoud`（迟到的 apply 拒绝）、all-ACTIVE sweep（PENDING fiber——cordis inject 等待没有超时）。
+**组合结果是一棵平铺配置树。** `apps/cli/config/base.cordis.yml` 与 `apps/cli/config/web.cordis.yml` 共同持有全部行——host runtime（32 行）、`api-gateway` 行、`webserver` 行、`dshClient` 行（浏览器 roster；modules 行同时是 host 行）。不做 spine bundle：每插件一行、每个 config 字段 yml 可改。这一立场后来推广到全仓：两个 surface 共享的配置项被抽取进 `apps/cli/config/base.cordis.yml`，各 surface 则收敛为一份 overlay（[共享 base overlay](../simplification/2026-07-29-shared-base-config-overlays.md)）。`--dev` 在 settle audit 之前由代码追加 `dsh-client-hmr` 行——prod 与 dev 的全部差异就是这一行。行序无装载语义；激活由服务可用性驱动。共享 audit 会拒绝没有 fiber 的 import、仅等待失败的 fiber 以恢复原始激活错误，并报告让 fiber 停在 `PENDING` 的服务；抛出错误前，审计会通过一个进程级检查点标记这些 rejection 的确切原因，从而让 `installFailLoud` 将 Loader 的重复通知合并为一次，而无关的未处理 rejection 仍然致命。Node app-boot 产物内嵌 `@cordisjs/plugin-include`，但将 `@cordisjs/plugin-loader` 保持为外部依赖，因此 include 的 `EntryTree` 与 host 会绑定到同一个 Loader peer，而不会让一棵配置树横跨两个 Loader 实现。
 
-**boot 胶水是一对 class。** `AppCLIEntry`（apps/cli）与 `AppWebEntry`（壳内核）只持有独立于 cordis 必须提前存在的东西：argv 事实、合成的 patch 集、解析出的 boot manifest、模块系统实例、loading 页句柄——其余一律进插件。`AppCLIEntry.run()` 三段：分层 env（ambient > cwd `.env` > `$DSH_HOME/.env`，顺手关掉上述缺陷）→ patch 合成 → Loader include boot 加三件套。`AppWebEntry.run()` 在浏览器侧镜像它：把 `window.__DSH_BOOT__` 解析成 `BootManifest`（双视角：npm 包行给模块表、cordis 插件行给 entry 组合；畸形 wire 大声抛）、建模块系统、渲染 loading 页、immediately 层预取与 Context/Loader 准备并行、**create entry 之前等预取齐**（物化是 `tree.import` 的同步 require，不受 fiber inject 等待保护；i18n → runtime/client 这类跨包 require 边要求 immediately 层工厂全部注册完——否则有实测 10–25% 的 boot 竞态）、收编 modules entry、逐图行 create、settle、sweep。
+**boot 胶水是一对 class。** `AppCLIEntry`（apps/cli）与 `AppWebEntry`（壳内核）只持有独立于 cordis 必须提前存在的东西：argv 事实、合成的 patch 集、解析出的 boot manifest、模块系统实例、loading 页句柄——其余一律进插件。`AppCLIEntry.run()` 三段：分层 env（ambient > cwd `.env` > `$DSH_HOME/.env`，顺手关掉上述缺陷）→ patch 合成 → Loader include boot 加 activation audit。`AppWebEntry.run()` 在浏览器侧镜像它：把 `window.__DSH_BOOT__` 解析成 `BootManifest`（双视角：npm 包行给模块表、cordis 插件行给 entry 组合；畸形 wire 大声抛）、建模块系统、渲染 loading 页、immediately 层预取与 Context/Loader 准备并行、**create entry 之前等预取齐**（物化是 `tree.import` 的同步 require，不受 fiber inject 等待保护；i18n → runtime/client 这类跨包 require 边要求 immediately 层工厂全部注册完——否则有实测 10–25% 的 boot 竞态）、收编 modules entry、逐图行 create、settle、sweep。
 
 **每个配置源有唯一声明位置。** yml 静态值是工程默认；profile json（`./.dsh-tmp-profile/config.json`，只读、绝不创建、暂锚 cwd 直至 `$DSH_HOME` 迁移）是用户配置，经静态 `PROFILE_MAPPINGS` 表映射到目标行（`provider`/`model` → `api-gateway` 行，`persistenceRoot` → jsonl 行）；CLI flags 映射到 `webserver` 行、字段集与 json 不相交；env 值经 yml `!!js` 表达式进入，绝不进映射表。patch 整体替换行 config，故 entry 类旁路 parse 重读 yml 行静态值再叠加覆盖。未映射的 json 键 fail loud。解析出的前端 `distIndex` 走同一 patch 通道——装配事实，不是用户配置。
 

.agents/notes/implemented/architecture/2026-07-25-web-input-machine-and-slash-pipeline.i18n.yaml

@@ -2,5 +2,5 @@
 # side as of the last confirmed-consistent state. Both languages carry equal authority;
 # after editing either side, bring the other along and re-record with:
 #   pnpm run verify-translation-pairing --write .agents/notes/implemented/architecture/2026-07-25-web-input-machine-and-slash-pipeline.md
-2026-07-25-web-input-machine-and-slash-pipeline.md: 92bb91c3e892d928cedf18ec57c725a116b6ffc8
-2026-07-25-web-input-machine-and-slash-pipeline.zh.md: 5bee6df52f16d935aa4f4ccff8627a2d43d44c8c
+2026-07-25-web-input-machine-and-slash-pipeline.md: c3deadb34d3a633525dde701c92bcc98c05e5d6e
+2026-07-25-web-input-machine-and-slash-pipeline.zh.md: 7a6988423dcdffebb0a28735146439c8ade0a862

.agents/notes/implemented/architecture/2026-07-25-web-input-machine-and-slash-pipeline.md

@@ -63,7 +63,7 @@ Calls that stay un-evented (registry registration → explicit call → await):
 A trigger/menu/pick pipeline with zero knowledge of "commands":
 
 - The service holds only the source registry (`SlashSource{trigger: '/'|'@', name, order?, candidates, onPick, matchSpace?, matchEnter?}`; (trigger,name) unique; the optional `order` sorts the roster — lower first, default 0, ties keep registration order — and that sorted roster is both group order and polling order) and `sessionOf(sctx)`. Implementing a match hook IS the declaration of participation in space/enter adjudication; the pipeline polls in roster order, the first non-undefined answer wins, and no claimant means the default sink. matchSpace is synchronous (space fires mid-keystroke; hot cache only); matchEnter is asynchronous (it may await the source's own warmup, and a warmup failure rejects).
-- The controller holds the single authoritative hit (span included; retained for Space after the menu closes), the per-session menu store, the candidate-fetch generation, keyboard arbitration (combobox mode: focus stays in the textarea, ↑↓/Enter/Escape are intercepted and all pass the IME composition guard, with the single exception Shift+Enter unconditionally going first), and pick orchestration (outcome → self-dispatched bail events); a `dismiss()` verb backs MenuView's injected `onDismiss` (a pointer down outside both the menu and the surrounding composer card closes the menu; MenuView also localizes group titles through the `slash.menu` locale namespace and clamps its height to the viewport space above the composer via ui-primitives' `useAnchoredMaxHeight`); at each session scope's birth it runs `warm(projection)` once over the source roster — within that scope the projection holds only the stable sessionId, with no published/capability transitions; the scope disposer tears down the controller.
+- The controller holds the single authoritative hit (span included; retained for Space after the menu closes), the per-session menu store, the candidate-fetch generation, keyboard arbitration (combobox mode: focus stays in the textarea, ↑↓/Enter/Escape are intercepted and all pass the IME composition guard, with the single exception Shift+Enter unconditionally going first), and pick orchestration (outcome → self-dispatched bail events). `toggleSource(name, syntheticHit)` is the chrome-launch path: it seeds only that registered source over the caller's textarea selection and publishes `launcher = name` until close; ordinary typed tracking clears the launcher and restores the full trigger roster. Both paths render the same MenuView and execute the same `onPick` chain. A `dismiss()` verb backs MenuView's injected `onDismiss` (a pointer down outside both the menu and the surrounding composer card closes the menu; MenuView also localizes group titles through the `slash.menu` locale namespace and clamps its height to the viewport space above the composer via ui-primitives' `useAnchoredMaxHeight`); at each session scope's birth it runs `warm(projection)` once over the source roster — within that scope the projection holds only the stable sessionId, with no published/capability transitions; the scope disposer tears down the controller.
 - Trigger-detection word boundaries (`user@host` and URL `/` never trigger) and the guard tiers (plain: `/` everywhere + `@` inline / claimed: `/` suppressed, `@` live / frozen: none) are the frozen pure core.
 
 ### hub / facade: the resident shell and the strict-session input body
@@ -101,7 +101,7 @@ skill/@subagent references skip the placeholder + occurrence identity chain —
 - `conversation.input.dock` — the stacked strip above the input (QueueDock's read-only queue list lands here), ordered by `order`.
 - `conversation.composer.dock` — the stats band on the composer's top edge.
 - `conversation.input.left` / `conversation.input.right` — the tool-row left and right regions.
-- `conversation.input.plan` / `conversation.input.model` (single) — the tool row's two named control seats; the bar passes only `locked` (owner props), each stays empty until its owning plugin registers, no placeholder fallback.
+- `conversation.input.plan` / `conversation.input.model` (single) — the tool row's two named control seats; the bar passes only `locked` (owner props), each stays empty until its owning plugin registers, no placeholder fallback. The plan seat stays empty while inactive because the shared Command source owns entry; an effective plan target renders the warn-state `Plan ×` status button, whose only action is `/plan off`.
 - `conversation.hero.workspace` (root scope) — the Workspace picker shared by the no-session and blank Hero; a pick reuses or creates the target blank session through `connectWorkspace`, moving the draft where necessary before switching current.
 
 ### Testing discipline
@@ -122,6 +122,8 @@ The state machine's entire behavior is covered by pure-JS unit tests (event sequ
 | Space adjudication also claiming execute-kind commands | The misfire defense: after a space the whole line is an ordinary prompt; irreversible side effects keep explicit entry points only |
 | A generic tokenPattern decoration mechanism | Structured occurrence records replace pattern scanning |
 | A placeholder select resident in the tool row | Named seats stay empty until registration; a placeholder clashing with the real implementation is two sources of truth |
+| An always-visible Plan on/off toggle | The shared Command source already owns entry; a second entry point turns a status seat into redundant mode chrome |
+| A second plus-menu component/controller, or an Add/File group above Command | It would duplicate async candidates, keyboard highlight, focus retention, and pick state; the plus control is only a source-filtered launcher for the existing MenuView, and this scope has no file capability |
 | All references through U+FFFC chips (the pre-Decision-21 line) | Plain text + derived decoration carries zero identity state; the literal text IS the model projection, sparing undo/clipboard any special cases; the chip chain is kept for scenarios needing indivisible atomicity |
 
 ## Consequences

.agents/notes/implemented/architecture/2026-07-25-web-input-machine-and-slash-pipeline.zh.md

@@ -63,7 +63,7 @@ occurrence 表与 chip 三投影：
 对"命令"零知识的触发/菜单/pick 管线：
 
 - service 只有 source 注册表（`SlashSource{trigger: '/'|'@', name, order?, candidates, onPick, matchSpace?, matchEnter?}`；(trigger,name) 唯一；可选 `order` 对 roster 排序——越小越靠前、默认 0、同值保持注册序——排序后的 roster 同时是组序与轮询序）与 `sessionOf(sctx)`。实现 match 钩子即参与空格/回车裁决的声明；管线按 roster 序轮询，首个非 undefined 应答胜出，无人认领落 default sink。matchSpace 同步（空格在击键中触发，只许热缓存）；matchEnter 异步（可 await 源自身预热，预热失败即 reject）。
-- controller 持有唯一权威 hit（含 span；菜单关闭后为 Space 保留）、per-session menu store、候选 fetch generation、键盘仲裁（combobox 模式：焦点始终在 textarea，↑↓/Enter/Escape 拦截且全程过 IME composition 守卫，唯一例外 Shift+Enter 无条件先行）、pick 编排（outcome → 自派 bail 事件）；`dismiss()` 动词支撑 MenuView 注入的 `onDismiss`（指针落在菜单与所在 composer 卡片之外即关闭菜单；MenuView 还经 `slash.menu` locale 命名空间本地化组标题，并经 ui-primitives 的 `useAnchoredMaxHeight` 把高度收敛到 composer 上方的视口空间）；每个 session scope 出生时对 source roster 做一次 `warm(projection)`，projection 在该 scope 内只有稳定的 sessionId，无 published/能力跃迁；scope disposer 拆除 controller。
+- controller 持有唯一权威 hit（含 span；菜单关闭后为 Space 保留）、per-session menu store、候选 fetch generation、键盘仲裁（combobox 模式：焦点始终在 textarea，↑↓/Enter/Escape 拦截且全程过 IME composition 守卫，唯一例外 Shift+Enter 无条件先行），以及 pick 编排（outcome → 自派 bail 事件）。`toggleSource(name, syntheticHit)` 是 chrome launcher 路径：它基于调用方的 textarea selection，只 seed 对应的已注册 source，并发布 `launcher = name` 直至关闭；普通的键入式 tracking 会清除 launcher 并恢复完整的 trigger roster。两条路径渲染同一个 MenuView，并执行同一条 `onPick` 链。`dismiss()` 动词支撑 MenuView 注入的 `onDismiss`（指针落在菜单与所在 composer 卡片之外即关闭菜单；MenuView 还经 `slash.menu` locale 命名空间本地化组标题，并经 ui-primitives 的 `useAnchoredMaxHeight` 把高度收敛到 composer 上方的视口空间）；每个 session scope 出生时对 source roster 做一次 `warm(projection)`，projection 在该 scope 内只有稳定的 sessionId，无 published/能力跃迁；scope disposer 拆除 controller。
 - 触发检测词边界（`user@host`、URL `/` 永不触发）、守卫分档（plain：`/` 到处 + `@` 行内 / claimed：`/` 抑制、`@` 活 / frozen：全无）为冻结纯核。
 
 ### hub / facade：常驻外壳与严格 session 输入体
@@ -101,7 +101,7 @@ skill/@subagent 引用不走占位符 + occurrence 身份链——pick 直接把
 - `conversation.input.dock`——输入上方堆叠条（QueueDock 的队列只读列表落此），order 定序。
 - `conversation.composer.dock`——composer 上沿统计带。
 - `conversation.input.left` / `conversation.input.right`——工具行左右区。
-- `conversation.input.plan` / `conversation.input.model`（single）——工具行两具名控制位；bar 只传 `locked`（owner props），空到 owning 插件注册为止，无占位 fallback。
+- `conversation.input.plan` / `conversation.input.model`（single）——工具行两具名控制位；bar 只传 `locked`（owner props），空到 owning 插件注册为止，无占位 fallback。plan seat 未激活时保持为空，因为入口归共享 Command source 所有；有效 plan 目标会渲染 warn 状态的 `Plan ×` 状态按钮，其唯一动作是 `/plan off`。
 - `conversation.hero.workspace`（root scope）——无 session / blank Hero 共用的 Workspace picker；pick 经 `connectWorkspace` 复用或创建目标 blank session，必要时搬运 draft 后切 current。
 
 ### 测试纪律
@@ -122,6 +122,8 @@ skill/@subagent 引用不走占位符 + occurrence 身份链——pick 直接把
 | 空格裁决也认领即执行型命令 | 误触发防线：空格后整行是普通 prompt；不可逆副作用只留显式入口 |
 | 通用 tokenPattern 装饰机制 | 结构化 occurrence 记录取代模式扫描 |
 | 占位 select 常驻工具行 | 具名坑位空到注册为止；占位件与真实现冲突时是双真相源 |
+| 始终可见的 Plan 开／关切换 | 入口已归共享 Command source 所有；第二个入口会把状态 seat 变成冗余的 mode chrome |
+| 第二套加号菜单组件／controller，或在 Command 上方增加 Add/File 分组 | 这会重复异步候选、键盘高亮、焦点保留与 pick 状态；加号控件只是既有 MenuView 按 source 过滤的 launcher，且此 scope 没有文件能力 |
 | 引用一律走 U+FFFC chip（决策 21 前旧线） | 纯文本 + 派生装饰零身份状态；原文即模型投影，undo/剪贴板免特判；chip 链保留给需要不可分原子性的场景 |
 
 ## 后果

.agents/notes/implemented/architecture/2026-07-28-directory-picker-capability-seam.i18n.yaml

@@ -2,5 +2,5 @@
 # side as of the last confirmed-consistent state. Both languages carry equal authority;
 # after editing either side, bring the other along and re-record with:
 #   pnpm run verify-translation-pairing --write .agents/notes/implemented/architecture/2026-07-28-directory-picker-capability-seam.md
-2026-07-28-directory-picker-capability-seam.md: ad2aa904beddb2fe941883c3c1827702dbec9964
-2026-07-28-directory-picker-capability-seam.zh.md: 30e719ad9b4e8374496106b447e961a042c7d8b6
+2026-07-28-directory-picker-capability-seam.md: 495062f910785e1bb2f421dbb25c01c399d45567
+2026-07-28-directory-picker-capability-seam.zh.md: 62fc87212ab627ea8819dab55e3a769b4a5afc42

.agents/notes/implemented/architecture/2026-07-28-directory-picker-capability-seam.md

@@ -20,7 +20,7 @@ Placement and policy rulings folded into this decision:
 - **Dependency survey (hand-roll vs adopt).** Node's stdlib *is* the maintained cross-platform OS layer (`readdir(withFileTypes)`, `homedir`, path semantics); surveyed alternatives fail the dependency bar — file-manager packages (`node-file-manager`, `files-and-folders`, Syncfusion's provider) are whole HTTP apps (fit), drive-letter helpers (`drivelist` native addon, `windows-drive-letters` ~7y stale) fail health/proportionality. The browse backend is a thin adapter over stdlib.
 - **Hidden entries: return-and-flag.** The host stamps `hidden` (POSIX dot convention) and returns everything; the client filters. Display policy stays client-side, and the show-hidden toggle shipped as exactly that client-only change: a fixed-label footer toggle whose state lives in the pressed presentation (`aria-pressed` + check glyph), a dot-led path-draft prefix reveals the hidden entries it names, and the current selection is exempt from both the hidden and the prefix filter (it anchors the two-pane view). Windows' `FILE_ATTRIBUTE_HIDDEN` is not exposed by dirents — documented limitation until a native probe pays for itself.
 - **Path-editor cancel scope: the dialog card.** The browse client's path editor cancels on Escape and on focus leaving the card, both observed at a card-scope wrapper rather than the input — after Tab parks focus on a filtered row the input is off the event path, yet Escape must collapse the editor (not the dialog) and a later focus departure must still cancel. Non-cancel exemptions: window/tab focus loss, in-card focus moves, and pointer paths (rows and the toggle suppress focus steal on mousedown while editing). Separators for seeding and draft-tail filtering are inferred from `listing.home`; the wire-field alternative below records the deferred authoritative form. Combobox semantics between the editor and the list it filters (`aria-expanded`/`aria-controls`/active-descendant, result announcements) are likewise deferred — today they read to assistive tech as separate widgets.
-- **Navigation lands selection-anchored, progressively.** Away from the display root (the same collapse the crumb header renders, so crumbs and pane shape never disagree), the browse client's navigate commits the target level the moment it arrives — the editor closes and loading ends on that first settlement, so an Enter-submitted navigation is never withdrawn waiting on more — and a parent leg then upgrades the landing in place: the target's actual parent-level entry re-selected (platform case folding on Windows), its children on the right, so a crumb jump reads as stepping back one pane rather than collapsing to a single column. The parent leg runs under the landing's supersession scope and is aborted on the wire by any newer intent; a failed parent leg, or a truncated parent window lacking the target, leaves the committed single-pane landing — the upgrade must never orphan the selection it exists to anchor.
+- **Navigation lands selection-anchored, quiet, and bounded.** Away from the display root (the same collapse the crumb header renders, so crumbs and pane shape never disagree), the landing is two-pane: the target's actual parent-level entry re-selected (platform case folding on Windows), its children on the right, so a crumb jump reads as stepping back one pane rather than collapsing to a single column. Target and parent legs land as **one frame** when the parent leg settles within the 200ms wait bound — the stale view keeps rendering until then, so navigation swaps the panes without an intermediate single-pane flash — and past the bound the target commits alone at once (an Enter-submitted navigation is never held hostage by a stalled parent) with the late parent leg upgrading the landing in place. The parent leg runs under the landing's supersession scope and is aborted on the wire by any newer intent (Escape inside the landing window therefore withdraws the whole navigation); a failed parent leg, or a truncated parent window lacking the target, leaves the single-pane landing — the upgrade must never orphan the selection it exists to anchor. The loading indicator follows the same quiet rule: it floats over the content's bottom-right corner (never a layout-shifting row; the truncated/error rows own the bottom left and keep rendering through a scan) and only once a scan outlives a 300ms silence window, so a local listing swaps with nothing shown at all. Row picks are deliberately exempt from the one-frame rule: a pick's immediate pane split is its selected-state feedback (aria-current, crumbs following), while a navigation has nothing to acknowledge the click but the swap itself. Both timing constants are calibrated for local enumeration; a remote deployment (one RPC per level, commonly 100–400ms) would sit inside the silence window with no pressed state on the crumbs — revisit the window or add pressed feedback when a remote consumer lands.
 - **Symlinks: follow for enterability.** `stat` probes symlinks (broken/cyclic → skipped); crumbs keep the logical path the operator navigated, and `workspace.create` already canonicalizes via realpath at adoption.
 - **Listing levels are bounded, and streamed.** One `list` call returns at most `maxEntries` rows (config, default 1000 — GitHub's web-UI directory-listing bound). The level streams via `opendir` into a name-sorted window of `maxEntries + 1` candidates, so memory stays O(maxEntries) and enterability probing touches only windowed candidates; the wire `DirectoryListing` carries a required `truncated` flag so the client states incompleteness instead of silently missing tail entries. A windowed broken symlink is not backfilled from beyond the window — the eviction already marks the level truncated. Window insertion is binary with an O(1) full-window tail rejection (an oversized level must not pay a window scan per dirent), and `list(path, signal)` threads the carrier's request signal so a scan of a stalled network directory cannot outlive a disconnected caller — every await in the scan (open, each read, each symlink probe) races the signal, an aborted exit abandons rather than awaits the close (Node queues close behind in-flight reads), and abandoned settlements are swallowed so cleanup can never surface as an unhandled rejection. An unbounded level is a memory/responsiveness hole for large or adversarial directories.
 - **Whole-filesystem scope, no roots config.** `workspace.create` accepts arbitrary paths and the API serves bash-driving methods, so a browse root would be UX scoping, not a boundary; configurability without a consumer fails the evidence bar. Deferred until a deployment needs it.
@@ -38,7 +38,7 @@ Placement and policy rulings folded into this decision:
 
 ## Consequences
 
-- `cordis.yml` chooses the interaction; `apps/cli` mounts `-browse` (the shipped default — remote-capable picking out of the box), one row having swapped backend and UI together; `-native` remains the host-display alternative.
+- `cordis.yml` chooses the interaction; `apps/cli` mounts the [`-auto` chooser](../feature/2026-07-29-directory-picker-adaptive-default.md), which resolves the host's situation at boot and mounts `-native` or `-browse` itself, one row still swapping backend and UI together; composing a backend row directly pins the interaction.
 - The wire gains `host.listDirectory`/`host.createDirectory` and four error codes; the connection fixture serves a deterministic browse tree and a deterministic `pickDirectory` path for keyless assembled tests.
 - A future interaction (or an Electron provider of the `native` interaction) is one dual-face backend package — no gateway surgery, no ui-workspace edits.
 - `ApiProxyDefaults.pickDirectory` (test-only injection) is gone; tests provide a stub `ctx.directoryPicker` like any other service.

.agents/notes/implemented/architecture/2026-07-28-directory-picker-capability-seam.zh.md

@@ -20,7 +20,7 @@ web GUI 的"打开本地文件夹"流程被焊死在一种交互上：`host.pick
 - **依赖调研（手写 vs 引入）。** Node 标准库本身就是维护中的跨平台 OS 层（`readdir(withFileTypes)`、`homedir`、路径语义）；调研过的替代品都过不了依赖门槛——文件管理器包（`node-file-manager`、`files-and-folders`、Syncfusion 的 provider）是整套 HTTP 应用（契合度不过），盘符工具（原生插件 `drivelist`、约七年未更的 `windows-drive-letters`）健康度／比例失当。browse 后端是标准库上的薄适配。
 - **隐藏条目：返回并打标。** 宿主标注 `hidden`（POSIX 点前缀约定）并返回全部条目；客户端过滤。展示策略留在客户端，"显示隐藏"开关正是作为这一纯客户端改动落地：标签固定的 footer 开关，其状态由按下态呈现承载（`aria-pressed` + 勾选符号）；以点开头的路径草稿前缀会显出它所指名的隐藏条目；当前选中项则不受隐藏与前缀两种过滤影响（它锚定着双栏视图）。Windows 的 `FILE_ATTRIBUTE_HIDDEN` 不被 dirent 暴露——记为限制，直到原生探测值回其成本。
 - **路径编辑器的取消范围：对话框卡片。** browse 客户端的路径编辑器在按 Escape 与焦点离开卡片时取消，两者都在卡片范围的包装层而非输入框上监听——Tab 把焦点停到某个过滤命中的行之后，输入框已不在事件路径上，但 Escape 仍须收起编辑器（而非对话框），其后的焦点离开也仍须取消。不取消的豁免：窗口／标签页失焦、卡片内焦点移动，以及指针路径（编辑期间行与开关在 mousedown 时抑制焦点夺取）。预填与草稿末段过滤所用的分隔符从 `listing.home` 推断；下文的线上字段替代方案记录了被延期的权威形态。编辑器与其过滤的列表之间的 combobox 语义（`aria-expanded`／`aria-controls`／active-descendant、结果播报）同样被延期——目前二者在辅助技术看来是彼此独立的控件。
-- **导航以选中项为锚、渐进落地。** 在展示根之外（与 crumb 头部渲染的是同一塌缩，因此 crumb 与分栏形态永不相左），browse 客户端的导航在目标层级到达的那一刻即提交它——这次首个落定即关闭编辑器并结束加载，因此 Enter 提交的导航绝不会为等待更多内容而被撤回——随后父层级这一程就地升级这次落地：重新选中目标在父层级中的实际条目（Windows 上按平台惯例折叠大小写），右侧展示其子项，因此 crumb 跳转读作后退一栏，而不是塌缩成单列。父层级这一程在落地的 supersession 范围下运行，任何较新的意图都会在线上将其中止；父层级这一程失败，或被截断的父窗口缺少目标时，都保留已提交的单栏落地——升级的存在正是为了锚定选中项，绝不能反而让它悬空。
+- **导航以选中项为锚、安静且有界地落地。** 在展示根之外（与 crumb 头部渲染的是同一塌缩，因此 crumb 与分栏形态永不相左），落地即双栏：重新选中目标在父层级中的实际条目（Windows 上按平台惯例折叠大小写），右侧展示其子项，因此 crumb 跳转读作后退一栏，而不是塌缩成单列。父层级这一程在 200ms 等待上限内落定时，目标与父层级两程以**同一帧**落地——在此之前陈旧视图持续渲染，导航换栏时因此没有中间的单栏闪现——超出该上限则目标即刻单独提交（Enter 提交的导航绝不会被滞塞的父层级扣作人质），迟到的父层级这一程再就地升级这次落地。父层级这一程在落地的 supersession 范围下运行，任何较新的意图都会在线上将其中止（因此在落地窗口内按 Escape 即撤回整次导航）；父层级这一程失败，或被截断的父窗口缺少目标时，都保留单栏落地——升级的存在正是为了锚定选中项，绝不能反而让它悬空。加载指示器遵循同一安静规则：它浮于内容右下角（绝不是会挪动布局的一行；截断／错误行占据左下角，并在扫描期间持续渲染），且仅在扫描超出 300ms 静默窗口后才出现，因此本地列举切换时什么也不显示。行选取被刻意豁免于同一帧规则：选取后立即分栏本身就是其选中态反馈（aria-current、crumb 跟随），而导航除了换栏本身没有任何东西可确认这次点击。两个时序常量都按本地列举校准；远程部署（每层级一次 RPC，通常 100–400ms）会落在静默窗口之内、crumb 上却没有按下态——待远程消费方落地时，重新审视该窗口或补上按下反馈。
 - **符号链接：为可进入性而跟随。** 用 `stat` 探测符号链接（断链／循环→跳过）；面包屑保留操作者导航的逻辑路径，`workspace.create` 在接纳时本就做 realpath 规范化。
 - **列举层级有上限，且流式处理。** 单次 `list` 至多返回 `maxEntries` 行（配置项，默认 1000——GitHub 网页端目录列举的同一上限）。层级经 `opendir` 流入一个按名排序、容量 `maxEntries + 1` 的候选窗口，内存保持 O(maxEntries)，可进入性探测只触及窗口内候选；线上 `DirectoryListing` 携带必填的 `truncated` 标志，让客户端明示不完整而不是静默缺尾。窗口内的断链符号链接不从窗口外回填——发生过驱逐本身已把层级标记为截断。窗口插入为二分查找、满窗尾部单次比较即拒绝（超大层级不能为每个 dirent 付出一次全窗扫描），且 `list(path, signal)` 透传载体的请求信号，滞塞网络目录的扫描不会在调用方断连后继续存活——扫描中的每个 await（打开、每次读取、每次符号链接探测）都与信号赛跑，中止路径放弃而非等待 close（Node 会把 close 排在在飞读取之后），被放弃的 settlement 全部吞掉，清理不会以未处理拒绝的形式冒出。无上限的层级对超大或恶意构造的目录就是内存／响应性漏洞。
 - **全盘可浏览，不做 roots 配置。** `workspace.create` 接受任意路径且 API 本就提供驱动 bash 的方法，浏览根只会是 UX 范围而非边界；没有消费方的可配置性过不了证据门槛。等到有部署需要再做。
@@ -38,7 +38,7 @@ web GUI 的"打开本地文件夹"流程被焊死在一种交互上：`host.pick
 
 ## 后果
 
-- `cordis.yml` 决定交互形态；`apps/cli` 挂 `-browse`（随附默认——开箱即得可远程的选取），一行同时切换了后端与 UI；`-native` 仍是宿主屏幕方案。
+- `cordis.yml` 决定交互形态；`apps/cli` 挂 [`-auto` 选择器](../feature/2026-07-29-directory-picker-adaptive-default.md)，它在启动时判定宿主处境并自行挂载 `-native` 或 `-browse`，一行仍同时切换后端与 UI；直接组合某个后端行即固定交互。
 - 协议新增 `host.listDirectory`／`host.createDirectory` 与四个错误码；connection fixture 提供确定性浏览树与确定性 `pickDirectory` 路径供无密钥组装测试使用。
 - 未来的新交互（或提供 `native` 交互的 Electron 实现）只是一个双面后端包——无需网关手术，也不动 ui-workspace。
 - `ApiProxyDefaults.pickDirectory`（仅测试注入）删除；测试像提供其他服务一样提供 stub `ctx.directoryPicker`。

.agents/notes/implemented/architecture/2026-07-28-user-settings-seam.i18n.yaml

@@ -0,0 +1,6 @@
+# Bilingual-pair consistency record (docs/i18n/README.md): the git blob hash of each
+# side as of the last confirmed-consistent state. Both languages carry equal authority;
+# after editing either side, bring the other along and re-record with:
+#   pnpm run verify-translation-pairing --write .agents/notes/implemented/architecture/2026-07-28-user-settings-seam.md
+2026-07-28-user-settings-seam.md: bf93f95168b1b6d0dec5a9fc2c9aac5531f0564a
+2026-07-28-user-settings-seam.zh.md: 8cd4dfcbb2facdd590b2c24d453ad79e9badda4d

.agents/notes/implemented/architecture/2026-07-28-user-settings-seam.md

@@ -0,0 +1,35 @@
+# Agent Note: user-settings seam (`ctx.settings`) and the file provider
+
+Status: implemented
+
+English | [中文](2026-07-28-user-settings-seam.zh.md)
+
+> Scope: the `packages/settings/` capability family — the abstract seam, the file-backed provider, and the composition boundary between user settings and `cordis.yml`. The [web config-tree note](2026-07-24-web-config-tree-boot-and-transport-layering.md) recorded "the profile write path" as a deferral; this seam is that write path's owner. Consumer migrations (theme, locale, default model route) and the web `settings.*` RPC surface are follow-ups, not part of this note's shipped scope.
+
+## Problem
+
+User-editable configuration had no owner: `dsh web` read a cwd-anchored profile json through a static whitelist with no write path, the TUI read `$DSH_HOME/config.yaml` raw loader patches, and both froze at boot. A personal-settings page (web GUI) needs one cross-surface user layer with schema validation, a write path, and hot propagation — and peer products (Codex, Claude Code, Kimi, OpenCode, Pi) all converged on separating user preferences from extension composition. Loader-reactive config updates cannot carry this: `fiber.update` swaps entry config in place, so a plugin that read config at construction observes nothing and no callback tells it otherwise.
+
+## Decision
+
+**Two planes with a litmus test.** `cordis.yml` (+ Include patches) stays the composition plane: which plugins exist, wiring, deployment config, owned by the orchestrator and upgraded with the product. A settings namespace carries only the user-editable subset; the test is "should the personal config page edit it?" Values live in both planes without ambiguity because layering is the contract: schema defaults, then the registrant's composition `base` (its entry-config subset), then the user document section.
+
+**Three-package seam mirroring `session-persistence/`.** `dsh-settings` owns the abstract `Settings` service: namespace registry, layered resolution, schema validation, per-namespace deep-equal change detection, and the `settings/updated` commit event. Providers implement only `writable`/`load()`/`persist(ns, section)` and push externally observed documents through the protected `publish(doc)` — so hot-update semantics are identical across providers, and a network configuration-center backend (nacos-style, possibly read-only) is a sibling package away. `dsh-settings-local` is the file provider: YAML/JSON under `resolveSpec` (explicit defaulting to `<DSH_HOME>/settings.yaml`), chokidar watch, read-modify-write persists under a cross-process writer lock with atomic `0600` tmp+rename commits, leaf-level diff patching of the written namespace (comments survive untouched nodes), and content-equality self-write suppression ([write-path integrity note](2026-07-30-settings-write-path-integrity.md)).
+
+**Registrations are caller-fiber effects.** `register()` runs through the service proxy, so `this.ctx` is the registrant's context and the registration rides `ctx.effect`: disposing the registrant removes the namespace and its watchers (proven by the HMR disposal test), while the user's section keeps living in storage for the next owner.
+
+**Fail loud at rest, last-good in motion.** Boot-time and registration-time validation throw (invalid stored section fails the registering plugin; an existing-but-unparsable document fails provider load). Once live, a bad external edit warns and keeps the last good state per namespace — a hot reload must never take the process down. This asymmetry mirrors `Include.refresh()` and Kimi's safe runtime reload.
+
+**Consumers stay optional-by-construction.** A consumer registers inside `ctx.inject(['settings'], …)`; without a mounted provider it keeps resolving entry config alone, so every existing composition, demo, and snapshot works unchanged and migration is per-plugin.
+
+## Alternatives considered
+
+- **Include write-back as the user layer** (per-plugin config pages writing loader entry files, cordis-webui style): write-back would target per-composition files, binding user preferences to one `cordis.yml`; a per-user layer must survive template upgrades and serve TUI and web from one document.
+- **Loader-reactive `fiber.update` as the propagation channel**: constructor-time reads observe nothing; the seam's explicit `watch()` makes hot-update a consumer contract instead of framework magic.
+- **A domain-aware settings service** (getters per product area): the coupling objection from design review stands; the service stores, validates, and publishes — domain meaning stays with the registrant that owns the schema.
+- **Multi-layer precedence now** (system/managed/project tiers à la Codex/Claude Code): deferred until a real second layer exists; the resolve step is the single place layering would extend.
+- **A cross-process lockfile now** (Pi's proper-lockfile): initially deferred as "atomic replace plus watcher convergence until real contention shows up" — review showed convergence loses unobserved sibling namespaces, so the deferral is superseded by the [write-path integrity note](2026-07-30-settings-write-path-integrity.md)'s hand-rolled writer lock.
+
+## Consequences
+
+Deferred, in dependency order: the web `settings.raw`/`settings.describe`/`settings.update` RPC surface (which must redact `role('secret')` fields before exposure); first consumer migrations (`ui-theme`, locale, api-gateway default route) retiring `PROFILE_MAPPINGS` and the profile json; `${env:VAR}` value indirection for secrets; provider-side layering. The keyless snapshot obligation lands with the first model- or product-user-visible consumer, not with this infrastructure step.

.agents/notes/implemented/architecture/2026-07-28-user-settings-seam.zh.md

@@ -0,0 +1,35 @@
+# Agent Note：用户设置 seam（`ctx.settings`）与文件 provider
+
+Status: implemented
+
+[English](2026-07-28-user-settings-seam.md) | 中文
+
+> 范围：`packages/settings/` 能力族——抽象 seam、文件 provider，以及用户设置与 `cordis.yml` 的组合边界。[web config-tree note](2026-07-24-web-config-tree-boot-and-transport-layering.md) 曾把"profile 写路径"记为延后项；本 seam 就是该写路径的归属。消费者迁移（主题、语言、默认模型路由）与 web `settings.*` RPC 面是后续工作，不在本 note 已交付范围内。
+
+## 问题
+
+用户可编辑配置没有归属：`dsh web` 经静态白名单读 cwd 锚定的 profile json 且无写路径，TUI 读 `$DSH_HOME/config.yaml` 裸 loader patch，两者都在启动时冻结。个人设置页（web GUI）需要一个跨 surface 的用户层，带 schema 校验、写路径与热传导——同类产品（Codex、Claude Code、Kimi、OpenCode、Pi）也全部收敛于"用户偏好与扩展组合分离"。Loader 的 reactive 配置更新承载不了这件事：`fiber.update` 原地替换 entry config，构造期读过配置的插件毫无感知，也没有任何回调通知它。
+
+## 决策
+
+**两个面，一条判定。**`cordis.yml`（+ Include patches）仍是组合面：有哪些插件、接线、部署配置，归 orchestrator 所有并随产品升级。settings namespace 只承载用户可编辑子集；判定是"个人配置页应该能改它吗？"值可同时存在于两个面而不歧义，因为分层就是契约：schema 默认值，然后注册方的组合 `base`（其 entry 配置子集），最后用户文档分节。
+
+**镜像 `session-persistence/` 的三包 seam。**`dsh-settings` 拥有抽象 `Settings` 服务：namespace 注册表、分层解析、schema 校验、按 namespace 深相等变更检测，以及 `settings/updated` 提交事件。provider 只实现 `writable`/`load()`/`persist(ns, section)`，并通过受保护的 `publish(doc)` 推入外部观察到的文档——因此热更新语义对所有 provider 一致，网络配置中心后端（nacos 类，可能只读）只是一个平级包的距离。`dsh-settings-local` 是文件 provider：`resolveSpec` 显式默认到 `<DSH_HOME>/settings.yaml` 的 YAML/JSON、chokidar 监听、跨进程写锁下以 `0600` tmp+rename 原子提交的读-改-写 persist、对被写 namespace 的叶子级 diff 修补（未触碰节点的注释得以保留）、按内容相等抑制自写（[write-path integrity note](2026-07-30-settings-write-path-integrity.md)）。
+
+**注册是调用方 fiber 上的 effect。**`register()` 经服务代理调用，`this.ctx` 即注册方 context，注册挂在 `ctx.effect` 上：dispose 注册方即移除 namespace 及其观察者（HMR disposal 测试证明），而用户的分节继续留在存储中等待下一任 owner。
+
+**静止时响亮报错，运行中保留最后可用值。**启动期与注册期校验直接抛错（非法存量分节使注册插件加载失败；存在但不可解析的文档使 provider 加载失败）。运行中坏的外部编辑只告警并按 namespace 保留最后可用状态——热重载绝不拖垮进程。该不对称镜像 `Include.refresh()` 与 Kimi 的安全运行时重载。
+
+**消费者天然可选。**消费者在 `ctx.inject(['settings'], …)` 内注册；不挂 provider 时仍只按 entry 配置解析，因此所有既有组合、demo、snapshot 原样工作，迁移按插件渐进。
+
+## Alternatives considered
+
+- **以 Include 写回为用户层**（cordis-webui 式的按插件配置页写 loader entry 文件）：写回目标是按组合的文件，会把用户偏好绑死在某个 `cordis.yml` 上；用户层必须在模板升级中存活，并以同一文档服务 TUI 与 web。
+- **以 Loader reactive `fiber.update` 为传导通道**：构造期读取毫无感知；seam 的显式 `watch()` 把热更新变成消费者契约而非框架魔法。
+- **领域化的 settings 服务**（按产品域的 getter）：设计评审中的耦合反对成立；服务只做存储、校验、发布——领域含义留给拥有 schema 的注册方。
+- **现在就做多层优先级**（Codex/Claude Code 式 system/managed/project 层级）：延后到真实第二层出现；resolve 步骤是分层未来唯一的扩展点。
+- **现在就上跨进程锁**（Pi 的 proper-lockfile）：最初以"原子替换加 watcher 收敛，真实冲突出现再说"为由延后——评审发现收敛会丢失未观察到的同级 namespace，因此该延后已被 [write-path integrity note](2026-07-30-settings-write-path-integrity.md) 的手写写锁取代。
+
+## 后果
+
+按依赖顺序延后：web `settings.raw`/`settings.describe`/`settings.update` RPC 面（暴露前必须对 `role('secret')` 字段脱敏）；首批消费者迁移（`ui-theme`、语言、api-gateway 默认路由）并退役 `PROFILE_MAPPINGS` 与 profile json；面向密钥的 `${env:VAR}` 值间接引用；provider 侧分层。keyless snapshot 义务随第一个模型或产品用户可见的消费者落地，而非本基础设施步骤。

.agents/notes/implemented/architecture/2026-07-29-request-level-llm-config-credentials.i18n.yaml

@@ -0,0 +1,6 @@
+# Bilingual-pair consistency record (docs/i18n/README.md): the git blob hash of each
+# side as of the last confirmed-consistent state. Both languages carry equal authority;
+# after editing either side, bring the other along and re-record with:
+#   pnpm run verify-translation-pairing --write .agents/notes/implemented/architecture/2026-07-29-request-level-llm-config-credentials.md
+2026-07-29-request-level-llm-config-credentials.md: f12a2496a767decc3ce2b065f6be03009aec8992
+2026-07-29-request-level-llm-config-credentials.zh.md: 99fd90013a24746962ca02a5f4f18cdccd53f71a

.agents/notes/implemented/architecture/2026-07-29-request-level-llm-config-credentials.md

@@ -0,0 +1,29 @@
+# Agent Note: request-level LLM configuration and the credential seam
+
+Status: implemented
+
+English | [中文](2026-07-29-request-level-llm-config-credentials.zh.md)
+
+> Scope: the first production consumers of `ctx.settings` (the two LLM adapter plugins), the new `packages/credentials/` capability family, and the `packages/util/atomic-write` extraction. The follow-up wire surface (`settings.*`/`credentials.*` RPC, secret-role masking, the web settings form) is a separate PR and not part of this note's shipped scope.
+
+## Problem
+
+The [settings seam](2026-07-28-user-settings-seam.md) shipped without a production consumer, and the LLM adapters were the motivating one: both froze `apiKey`/`baseURL`/catalog into adapter instances at plugin load, so a changed key or endpoint needed a process restart, and a missing key failed plugin load — the worst possible first-run posture for a personal config page ("store a key, then restart"). Secrets were also headed the wrong way: the natural move (put `apiKey` in the settings document) would have forced masking, server-side backfill on `replace`, and dotfiles-sync warnings, a mitigation stack for a problem peer products simply do not have — Codex (`env_key` + auth.json), Reasonix (`api_key_env` + home `.env`), OpenCode/Pi (`auth.json`), Claude Code (`apiKeyHelper`) all keep secrets out of configuration files.
+
+## Decision
+
+**Per-request resolution, not fiber rebuilds.** The adapters take an options thunk (and a per-stream credential resolver) instead of frozen construction facts, resolving once per operation — the Pi pattern, with its tested semantics: two requests straddling a change see two configurations, one request resolves exactly once, and an in-flight stream keeps the facts it started with. This deletes the entire swap machinery a rebuild design needs (`DUPLICATE_ADAPTER` ordering, `NO_ADAPTER` windows, a deferred-activation state machine) and makes a missing key a *request-time* actionable failure (`MISSING_CREDENTIAL` naming every entry point) while the route stays registered and the catalog stays browsable. The one registration-captured fact — the retry policy the `ctx.llm` registry snapshots at `registerAdapter` (plus pi-ai's route *set*) — re-registers the same adapter instance in one synchronous section when it changes.
+
+**Secrets are references, values live behind `ctx.credentials`.** Configuration (both planes) carries `apiKeyEnv: DEEPSEEK_API_KEY`; the three-package credential seam resolves it per operation. `credentials-local` layers the live process environment (read-only, wins — a launch-time override is operator intent and must be *visibly* read-only, so shadowed writes reject instead of appearing to succeed) over `$DSH_HOME/.env` (writable, byte-preserving line edits, a quoting ladder dotenv reads back verbatim, wholesale snapshot replacement on reload so a deleted entry never lingers — the Claude Code additive-reapply lesson). Resolution order in the adapters is literal `apiKey` first (preserving the historical `config.apiKey ?? env` observable semantics), then the seam, then — only without a mounted seam — the raw environment variable.
+
+**Per-plugin namespaces, schema ≡ `Config`.** Each adapter registers its own namespace (`llm-deepseek`, `llm-pi-ai`) with its plugin `Config` schema and its `cordis.yml` entry as the composition `base` — a settings section is the same YAML shape as the entry config, and `resolveAdapterOptions`/`resolveProfiles` stay the one explicit resolve step for both. A live snapshot failing a beyond-schema bound keeps the last good facts (the seam's last-good philosophy extended one level up); the entry config itself still fails load. pi-ai's `providers` became a dict keyed by route so base and user layers merge per provider and the route set is structural; the array shape fails loud with migration directions, and an empty dict is the valid dormant posture — a composition ships the adapter bare and every route stays a user-plane decision.
+
+## Alternatives considered
+
+- **A bridge plugin (`dsh-llm-models`) owning one unified `models` dict** — with per-plugin namespaces there is nothing left to bridge, and the adapter-mapping rules it needed were pure invented indirection.
+- **Secrets in settings.yaml under `role('secret')` masking** — deleting the problem (references) beats mitigating it (mask + backfill + sync warnings); the coding-agent cohort is unanimous.
+- **Registry-level live retry policy** — making `providerRetryPolicy` re-read per call would silently change the `ctx.llm` capture contract every registration relies on; re-registering the route in place keeps that contract and stays observable.
+
+## Consequences
+
+Onboarding is restart-free end to end (pinned by the `missing-credential` headless snapshot and the credentials-rotation composition tests): boot keyless, browse the catalog, store the key, prompt again. The demos mount `settings-local` + `credentials-local` by default and inline no `!!js` key plumbing. `runLoaderSmoke` gained `expectedExitCode` so a designed failure surface can be pinned rather than masked. Deferred: the wire/UI surface must redact `role('secret')` fields before any RPC exposes `describe()`, settings-layer arrays still replace wholesale (the deepseek `models` list), and a settings section cannot remove a composition-provided pi-ai route (only override or extend). Review of this seam later reworked where the store lives and who may read it, made one request resolve one configuration generation, and made route replacement atomic ([credential boundaries note](2026-07-30-credential-boundaries-and-atomic-registration.md)).

.agents/notes/implemented/architecture/2026-07-29-request-level-llm-config-credentials.zh.md

@@ -0,0 +1,29 @@
+# Agent Note：请求级 LLM 配置与凭据 seam
+
+Status: implemented
+
+[English](2026-07-29-request-level-llm-config-credentials.md) | 中文
+
+> 范围：`ctx.settings` 的第一批生产消费方（两个 LLM 适配器插件）、新增的 `packages/credentials/` 能力族，以及 `packages/util/atomic-write` 的抽取。后续的 wire 面（`settings.*`/`credentials.*` RPC、secret 角色脱敏、web 设置表单）是单独的 PR，不在本 note 已交付范围内。
+
+## 问题
+
+[settings seam](2026-07-28-user-settings-seam.md) 落地时没有生产消费方，而 LLM 适配器正是当初驱动该 seam 的那个消费方：两个适配器都在插件加载时把 `apiKey`/`baseURL`/catalog 冻结进适配器实例，改密钥或端点就要重启进程，密钥缺失则直接使插件加载失败——对个人配置页而言，这是最糟糕的首次运行姿态（「先存密钥，再重启」）。机密的走向也不对：顺理成章的做法（把 `apiKey` 放进设置文档）会被迫引入脱敏、`replace` 时的服务端回填与 dotfiles 同步告警，为一个同类产品根本没有的问题堆起一整摞缓解措施——Codex（`env_key` + auth.json）、Reasonix（`api_key_env` + 家目录 `.env`）、OpenCode/Pi（`auth.json`）、Claude Code（`apiKeyHelper`）全都把机密挡在配置文件之外。
+
+## 决策
+
+**按请求解析，而非重建 fiber。**适配器改为接收一个 options thunk（外加按流调用的凭据解析器），不再持有冻结的构造期事实，每个操作解析一次——即 Pi 的模式，连同其经测试固定的语义：跨越一次变更的两个请求看到两份配置，一个请求恰好解析一次，进行中的流保持其起始事实。这删掉了重建式设计所需的整套切换机制（`DUPLICATE_ADAPTER` 顺序问题、`NO_ADAPTER` 窗口、延迟激活状态机），并把密钥缺失变成*请求时*可行动的失败（`MISSING_CREDENTIAL` 点名每个配置入口），同时路由保持注册、catalog 保持可浏览。唯一在注册期捕获的事实——`ctx.llm` 注册表在 `registerAdapter` 时快照的重试策略（外加 pi-ai 的路由*集合*）——在其变化时于一个同步区段内原地重新注册同一适配器实例。
+
+**机密是引用，值藏在 `ctx.credentials` 背后。**配置（两个面）携带 `apiKeyEnv: DEEPSEEK_API_KEY`；三包凭据 seam 按操作解析它。`credentials-local` 把活跃进程环境（只读、优先——启动时覆盖是操作者意图，必须*可见地*只读，因此被遮蔽的写入直接拒绝而不是表面成功）叠加在 `$DSH_HOME/.env` 之上（可写、保字节行级编辑、dotenv 能逐字读回的引号阶梯、重载时整体替换快照使删除的条目绝不滞留——来自 Claude Code 增量重放（additive reapply）的教训）。适配器内的解析顺序为：字面 `apiKey` 优先（保留历史 `config.apiKey ?? env` 的可观察语义），然后是 seam，最后——仅在未挂载 seam 时——原始环境变量。
+
+**按插件划分 namespace，schema ≡ `Config`。**每个适配器注册自己的 namespace（`llm-deepseek`、`llm-pi-ai`），schema 用其插件 `Config` schema，组合 `base` 用其 `cordis.yml` 条目——settings 分节与 entry 配置是同一种 YAML 形状，`resolveAdapterOptions`/`resolveProfiles` 对两者仍是唯一的显式 resolve 步骤。存活快照若违反 schema 之外的约束，则保留最后可用事实（seam 的最后可用值哲学向上延伸一层）；entry 配置本身仍会加载失败。pi-ai 的 `providers` 改为以路由为键的字典，base 层与用户层因此按提供方合并，路由集合也由结构直接表达；数组形状响亮失败并给出迁移指引，而空字典是合法的休眠姿态——组合可以裸挂该适配器，把每一条路由都留给用户面决定。
+
+## 曾考虑的替代方案
+
+- **由桥接插件（`dsh-llm-models`）持有统一的 `models` 字典**——有了按插件划分的 namespace，就没有什么可桥接的了；它所需的适配器映射规则纯属凭空发明的间接层。
+- **把机密放进 settings.yaml 并靠 `role('secret')` 脱敏**——删除问题本身（引用）胜过缓解问题（脱敏 + 回填 + 同步告警）；编码 agent 同类产品在这一点上口径一致。
+- **注册表级的实时重试策略**——让 `providerRetryPolicy` 每次调用都重读，会静默改变所有注册都依赖的 `ctx.llm` 捕获契约；原地重新注册路由既保住该契约，又保持可观察。
+
+## 后果
+
+上手流程端到端免重启（由 `missing-credential` headless 快照与凭据轮换组合测试固定）：无密钥启动、浏览 catalog、存入密钥、再次发起提示。demo 默认挂载 `settings-local` + `credentials-local`，不再内联任何 `!!js` 密钥接线。`runLoaderSmoke` 新增 `expectedExitCode`，使按设计出现的失败面可以被固定而非被掩盖。延后事项：wire/UI 面在任何 RPC 暴露 `describe()` 之前必须对 `role('secret')` 字段脱敏；settings 层的数组仍整体替换（deepseek 的 `models` 列表）；settings 分节无法移除组合提供的 pi-ai 路由（只能覆盖或扩展）。对该 seam 的评审随后改造了存储的所在位置与谁可以读取它，让一个请求解析出一个配置世代，并使路由替换成为原子操作（[credential boundaries note](2026-07-30-credential-boundaries-and-atomic-registration.md)）。

.agents/notes/implemented/architecture/2026-07-30-adapter-owned-max-token-defaults.i18n.yaml

@@ -0,0 +1,6 @@
+# Bilingual-pair consistency record (docs/i18n/README.md): the git blob hash of each
+# side as of the last confirmed-consistent state. Both languages carry equal authority;
+# after editing either side, bring the other along and re-record with:
+#   pnpm run verify-translation-pairing --write .agents/notes/implemented/architecture/2026-07-30-adapter-owned-max-token-defaults.md
+2026-07-30-adapter-owned-max-token-defaults.md: a522848fd4482f84859e587505b6a5e6f5c72d60
+2026-07-30-adapter-owned-max-token-defaults.zh.md: 8db6a06199fc1c4e73c86492d12dc86edafe8c7e

.agents/notes/implemented/architecture/2026-07-30-adapter-owned-max-token-defaults.md

@@ -0,0 +1,33 @@
+# Agent Note: Adapter-owned max-token defaults
+
+Status: implemented
+
+English | [中文](2026-07-30-adapter-owned-max-token-defaults.zh.md)
+
+## Problem
+
+An LLM adapter could serialize an explicit `GenerateOptions.maxTokens`, but its Cordis configuration could not establish a reconstructable conversation default. Applying a fallback only inside provider serialization would make the wire request differ from the durable `request/header`; putting every provider's default in Agent Loop would instead transfer deployment and model policy into the provider-neutral driver.
+
+## Decision
+
+`LlmResolvedModelInfo.defaultMaxTokens` carries an optional adapter-configured per-request output cap for one exact provider/model route. `LlmService` validates it as a positive safe integer and materializes it into `LlmCallConfig.maxTokens` only when the caller omitted a value. A prepared call identifies materialized `maxTokens` and `reasoningEffort` fields as adapter defaults; explicit request or Agent options remain unmarked and therefore win without clamping.
+
+The agent loop continues to prepare calls before logging `request/header`, so the effective config and its adapter-default provenance become durable request facts before dispatch. Before the next `agent/request` waterfall, the loop removes marked fields from the proposal; exact-model resolution then materializes the current route's defaults again. A provider/model switch therefore cannot mistake a previous adapter's default for an explicit override, while explicit conversation values persist. Direct `LlmService.stream()` calls resolve the same default at the final adapter boundary. The field is a request default rather than a hard model output limit; adapters that preserve provider-owned defaults omit it.
+
+The native DeepSeek adapter exposes `maxTokens` in Cordis config with a 256,000-token default and maps the effective value to `max_tokens`. Its default context capacity is 1,000,000 tokens: both built-in V4 entries publish that exact capacity, while configured entries without capacity and unlisted pass-through ids inherit the same adapter-wide fallback.
+
+## Alternatives considered
+
+**Apply the default only in DeepSeek serialization.** Rejected because the provider wire would contain a model-visible value absent from the durable request header.
+
+**Set `AgentOptions.maxTokens` in every shipped application.** Rejected because applications would duplicate adapter deployment policy, direct LLM calls would behave differently, and selecting another provider would retain a DeepSeek-specific cap.
+
+**Represent 256,000 as a hard per-model maximum.** Rejected because the configured value is the desired request budget, not evidence that every configured endpoint rejects larger outputs. Explicit callers remain authoritative.
+
+**Leave the provider default in control.** Rejected for the native DeepSeek deployment because the product requires a stable 256,000-token conversation budget across compatible endpoints.
+
+## Consequences
+
+DeepSeek conversations send `max_tokens: 256000` by default, and the same value plus its adapter provenance appear in the session request header. Deployments can change the adapter default through `llm-deepseek.config.maxTokens`; per-agent and per-request values override it. Changing the route rematerializes the new exact adapter's default instead of carrying DeepSeek's derived value forward. Other adapters retain their existing behavior until they intentionally publish `defaultMaxTokens`.
+
+The 256,000-token output budget reserves a large part of the one-million-token context on endpoints that pre-allocate requested output. Deployments whose gateway or model supports a smaller budget must lower `maxTokens`; the explicit configuration is preferable to an undocumented provider fallback.

.agents/notes/implemented/architecture/2026-07-30-adapter-owned-max-token-defaults.zh.md

@@ -0,0 +1,33 @@
+# Agent Note: 适配器持有的最大 token 默认值
+
+Status: implemented
+
+[English](2026-07-30-adapter-owned-max-token-defaults.md) | 中文
+
+## Problem
+
+LLM（大语言模型）适配器可以序列化显式的 `GenerateOptions.maxTokens`，但无法通过 Cordis 配置建立可重建的对话默认值。仅在提供方序列化中应用回退，会导致协议请求与持久 `request/header` 不一致；若将各提供方默认值都放进 agent loop（智能体循环），则会把部署与模型策略转移到提供方无关的驱动器中。
+
+## Decision
+
+`LlmResolvedModelInfo.defaultMaxTokens` 携带一条确切提供方／模型路由的可选单次请求输出上限，该值由适配器配置。`LlmService` 将其校验为正安全整数，并且仅在调用方省略值时才填入 `LlmCallConfig.maxTokens`。准备后的调用会将已填入的 `maxTokens` 和 `reasoningEffort` 字段标记为适配器默认值；显式请求值或 Agent 选项不带该标记，因此优先且不会被自动调整。
+
+agent loop 仍在记录 `request/header` 前准备调用，因此生效配置及其适配器默认值来源会在分派前成为持久请求事实。下一次 `agent/request` waterfall（瀑布式事件）前，agent loop 会从提议中移除带标记字段，随后精确模型解析会再次填入当前路由的默认值。因此，切换提供方／模型不会把前一个适配器的默认值误当成显式覆盖，而显式对话值则会保留。直接调用 `LlmService.stream()` 时，也会在最终适配器边界解析同一默认值。该字段是请求默认值，而非模型输出硬上限；保留提供方持有默认值的适配器会省略它。
+
+原生 DeepSeek 适配器在 Cordis 配置中公开 `maxTokens`，默认值为 256,000 token，并将生效值映射为 `max_tokens`。其默认上下文容量为 1,000,000 token：两个内置 V4 配置项均公布这一精确容量；不含容量的已配置项和未列出的原样传递 id 则继承同一个适配器级回退值。
+
+## Alternatives considered
+
+**仅在 DeepSeek 序列化中应用默认值。** 不予采纳，因为提供方协议会包含持久请求 header 中缺失的模型可见值。
+
+**在每个已发布应用中设置 `AgentOptions.maxTokens`。** 不予采纳，因为应用会重复适配器部署策略，直接 LLM 调用的行为将不同，而且选择另一个提供方后仍会保留 DeepSeek 专用上限。
+
+**将 256,000 表示为每模型硬上限。** 不予采纳，因为配置值是所需请求预算，无法证明每个已配置端点都会拒绝更大的输出。显式调用方仍具有最终决定权。
+
+**由提供方默认值控制。** 对原生 DeepSeek 部署不予采纳，因为产品要求各兼容端点都采用稳定的 256,000 token 对话预算。
+
+## Consequences
+
+DeepSeek 对话默认发送 `max_tokens: 256000`，会话请求 header 中也会出现相同的值及其适配器来源。部署可以通过 `llm-deepseek.config.maxTokens` 更改适配器默认值；每个 agent 和每次请求的值都会覆盖它。更改路由会重新填入新的精确适配器默认值，而不是继续沿用 DeepSeek 派生出的值。其他适配器会保留现有行为，直至主动公布 `defaultMaxTokens`。
+
+对于预分配请求输出的端点，256,000 token 的输出预算会占用 1,000,000 token 上下文中的很大部分。如果部署使用的 gateway 或模型仅支持较小预算，则必须调低 `maxTokens`；显式配置优于未记录的提供方回退值。

.agents/notes/implemented/architecture/2026-07-30-client-locale-full-rollout.i18n.yaml

@@ -0,0 +1,6 @@
+# Bilingual-pair consistency record (docs/i18n/README.md): the git blob hash of each
+# side as of the last confirmed-consistent state. Both languages carry equal authority;
+# after editing either side, bring the other along and re-record with:
+#   pnpm run verify-translation-pairing --write .agents/notes/implemented/architecture/2026-07-30-client-locale-full-rollout.md
+2026-07-30-client-locale-full-rollout.md: c080d9f240d4533ecd9694ceecfada8662c46425
+2026-07-30-client-locale-full-rollout.zh.md: 062d982e3d7ea62f3ca4c8fedb842e8336f0852c

.agents/notes/implemented/architecture/2026-07-30-client-locale-full-rollout.md

@@ -0,0 +1,45 @@
+# Agent Note: Full client copy rollout onto the typed locale seat, and the non-translation boundary
+
+Status: implemented
+
+English | [中文](2026-07-30-client-locale-full-rollout.zh.md)
+
+## Problem
+
+After the typed locale standard seat landed (`locale:` on register → framework-injected typed `t`), only four early adopters rode it; every other client package still shipped hardcoded, mixed-language literals. Migrating the rest required mechanisms and boundary decisions the early adopters never touched: how registration-time text (nav rows, view-tab labels) refreshes on a language switch; how the zero-cordis ui-primitives atoms receive copy; and which strings deliberately stay untranslated — an unrecorded boundary invites a future agent to "complete" the localization.
+
+## Decision
+
+**Registration-time text rides a label thunk.** A list registration's `label` accepts `SlotLabel = string | (() => string)`; owners projecting ledger rows resolve through `resolveSlotLabel` (never reading `options.label` raw) and make the read point follow the locale revision (outlets subscribe to the revision themselves; off-ledger projections such as the ui-settings nav fold the revision into their cache key and subscribe to both sources). Thunks evaluate per read, so a language switch causes zero ledger churn — no re-registration, versions stay put, and every `locale/change` re-registration wiring is deleted.
+
+**Component copy rides the standard `t` seat; deep children take `t` as a plain prop** typed `XxxProps['t']`. The dictionary canon is unchanged: `zh satisfies Record<string, string>` is the key source and `en satisfies Record<XxxKey, string>` locks bilingual balance.
+
+**Zero-cordis atoms (ui-primitives) take copy as props**: `labels` on `TerminalBlock`/`JsonTree`, `copyLabel`/`copiedLabel` on `CodeBlock`, `codeLabels` on `MarkdownText`, `truncatedLabel` on `JsonBlock`, `label` on `ConnectionBanner`, `closeLabel` on `Modal` — defaults are the previous hardcoded strings, so a consumer passing nothing renders byte-identical output. Localized plugins pass dictionary-driven labels from their own `t` seat; call sites passing object props memoize them on the `t` identity (`MarkdownText` caches its component table on the `codeLabels` identity).
+
+**The non-translation boundary (deliberate decisions, not debt):**
+
+- **Error/failure strings stay English**: client-authored fallbacks (`command failed`, plan-toggle failures), RpcError messages, and wire `error.message (code)` pass-throughs render verbatim.
+- **Design literals stay out of the dictionaries**: tool-row variant titles (Think/Bash/…), SYSTEM/USER-style kind badges, the Plan chip wordmark, the whole StatsLine — identical in both languages.
+- **ui-trajectory is deferred wholesale** (a developer inspection surface, terminology-dense, ruled separately).
+- **Boot copy stays hardcoded** (AppRoot renders before the locale service exists).
+
+**Derivation layers stay pure; localization happens at render.** ui-workspace's `relativeTime` returns structured `{unit, n}` composed with dictionary templates by the renderer; blank sessions and the Ungrouped bucket keep their stored titles, with the renderer substituting localized copy off the `blank` flag / absent `workspaceId`; **blank rows are excluded from search entirely** (a bilingual display title cannot match a single-language query stably). Dates use no Intl: format templates live in the dictionaries (message clock `clock.md`/`clock.ymd`, workspace hover `date.ymd`) and the formatters take `t` as a parameter, staying pure.
+
+**Test and e2e doctrine**: `makeTranslate(...dicts)` (dsh-client-test-runtime) mirrors the service lookup chain (first-dict-wins, key fallback, `{name}` interpolation); component specs stub the `t` seat with it, typed against real props seats. Web e2e uniformly opens through `newEnglishPage` (pins `dsh.locale=en` before boot) and the built-boot snapshot pins the same — goldens are immune to localization migrations; the settings language-switch scenario deliberately bypasses the helper to cover the zh default.
+
+The "apply layer subscribes to `locale/change` and re-registers for fresh labels" mechanism in the [settings/locale/theme layering note](../../proposed/architecture/2026-07-25-client-settings-locale-theme.md) is superseded by this decision (thunk + revision lifecycle).
+
+## Alternatives considered
+
+- **Keep labels as strings and re-register on switch** (the early adopters' original shape): boot already registers once per package, and `locale/change` listeners re-registering amplifies into a storm; ledger version churn also busts every version-keyed projection cache. Thunks move the refresh cost to read points that already follow the revision.
+- **A locale context/injection channel for ui-primitives**: breaks the zero-cordis boundary (atoms would depend on the runtime) and drags unlocalized consumers (ui-trajectory) along. Props let each consumer decide independently.
+- **Error strings in the dictionaries**: the error surface is a debugging surface — verbatim English is what gets searched and compared in reports; wire pass-throughs are untranslatable anyway, and half-translation manufactures mixed-language text.
+- **`toLocaleString()`/Intl for dates**: follows the browser/OS language, not the app locale, guaranteeing mixed text after a switch; the dictionary templates are tiny and isomorphic to the message clock.
+- **Blank rows matching search (against localized or stored titles)**: either choice yields "visible but unfindable" in one language; placeholder rows carry no information, so whole-row exclusion is the stable semantic.
+
+## Consequences
+
+- A language switch refreshes the whole UI instantly with zero re-registration; adopting a new package is three steps (dictionary + declare-merge + `locale: NS`), no hand-written glue.
+- Cost: list-label consumers must know `resolveSlotLabel` (a raw `options.label` read can now hold a function); the `SlotLabel` type catches most misuse statically.
+- ui-primitives' Chinese defaults still render Chinese under the English locale **until a consumer passes labels** — the unmigrated JsonTree consumer (ui-trajectory) showing its English defaults happens to match that package's all-English status quo.
+- Pinning e2e to English means the zh default is covered mainly by package-level component specs and the settings language-switch scenario; browser e2e no longer asserts zh copy.

.agents/notes/implemented/architecture/2026-07-30-client-locale-full-rollout.zh.md

@@ -0,0 +1,45 @@
+# Agent Note: client 文案全量接入 typed locale 席位与不翻译边界
+
+Status: implemented
+
+[English](2026-07-30-client-locale-full-rollout.md) | 中文
+
+## Problem
+
+typed locale 标准席位（`locale:` 注册声明 → 框架注入强类型 `t`）落地后，只有四个先行包接入；其余 client 包的文案仍是硬编码的中英混杂字面量。全量迁移需要几个先行包没有触及的机制与边界决定：注册期文本（导航行、视图 tab 的 label）在语言切换时如何刷新；zero-cordis 的 ui-primitives 原子组件如何拿到文案；哪些字符串**刻意不**本地化——没有记录的边界会诱使后来者"补完"翻译。
+
+## Decision
+
+**注册期文本走 label thunk。** ui-slots 的 list 注册项 `label` 接受 `SlotLabel = string | (() => string)`；owner 投影 ledger 行时必须经 `resolveSlotLabel` 解析（不裸读 `options.label`），并让读取点跟随 locale revision（outlet 自身订阅 revision；ledger 外的投影如 ui-settings 导航把 revision 并进缓存键、订阅双源）。thunk 每次读取时求值，语言切换零 ledger churn——没有重注册、version 不动，`locale/change` 重注册接线全部删除。
+
+**组件文案走标准 `t` 席位；深层子组件用 prop 下传**，类型写 `XxxProps['t']`。字典规范形态不变：`zh satisfies Record<string, string>` 为 key 源、`en satisfies Record<XxxKey, string>` 锁双语平衡。
+
+**zero-cordis 原子组件（ui-primitives）文案 props 化**：`TerminalBlock`/`JsonTree` 的 `labels`、`CodeBlock` 的 `copyLabel`/`copiedLabel`、`MarkdownText` 的 `codeLabels`、`JsonBlock` 的 `truncatedLabel`、`ConnectionBanner` 的 `label`、`Modal` 的 `closeLabel`——默认值即原硬编码字符串，不传 props 的消费者渲染逐字节不变。已本地化的插件从自己的 `t` 席位传字典驱动的 label；传对象 props 的调用点按 `t` 身份 memo（`MarkdownText` 的组件表按 `codeLabels` 身份缓存）。
+
+**不翻译边界（刻意决定，不是欠账）：**
+
+- **错误/失败类字符串一律英文**：client 自产的兜底串（`command failed`、plan 切换失败）、RpcError message、wire 透出的 `error.message (code)` 原样呈现。
+- **设计字面量不进字典**：tool 行 variant 标题（Think/Bash/…）、SYSTEM/USER 类 kind 徽标、Plan chip 字标、StatsLine 全部指标——中英界面显示一致。
+- **ui-trajectory 整包缓做**（开发者检查面，术语密集，单独裁决）。
+- **boot 文案保持硬编码**（AppRoot 渲染早于 locale 服务可用）。
+
+**派生层保持纯函数，本地化只在渲染层**：ui-workspace 的 `relativeTime` 返回结构化 `{unit, n}` 由渲染组合字典模板；blank 会话/未分组桶的存储标题不变，渲染按 `blank` 标志/`workspaceId` 缺席替换本地化文案；**搜索态 blank 行一律排除**（双语标题无法与单语查询稳定匹配）。日期不引 Intl：格式模板进字典（消息时钟 `clock.md`/`clock.ymd`，workspace hover `date.ymd`），格式化函数吃 `t` 参数保持纯。
+
+**测试与 e2e 口径**：`makeTranslate(...dicts)`（dsh-client-test-runtime）镜像服务查找链（首个命中字典胜出、key 兜底、`{name}` 插值），组件测试的 `t` 桩统一用它并以真实 props 席位定型。web e2e 统一 `newEnglishPage`（boot 前钉 `dsh.locale=en`），built-boot snapshot 同样钉 en——golden 对语言迁移免疫；settings 语言切换用例刻意绕开该 helper 覆盖 zh 默认态。
+
+[settings/locale/theme 分层 Note](../../proposed/architecture/2026-07-25-client-settings-locale-theme.md) 中"apply 层订阅 `locale/change` 重注册刷新 label"的机制已被本决定取代（thunk + revision 生命周期）。
+
+## Alternatives considered
+
+- **label 保持 string、语言切换时重注册**（先行包的旧形态）：boot 每包一次注册已很重，`locale/change` 监听者重注册会放大成风暴；ledger version 抖动还会击穿一切按 version 缓存的投影。thunk 把刷新成本移到读取点，读取点本来就跟随 revision。
+- **给 ui-primitives 造 locale context/注入通道**：破坏 zero-cordis 边界（原子组件从此依赖运行时），且强迫未本地化消费者（ui-trajectory）陪跑。props 化让每个消费者独立决定。
+- **错误串进字典**：错误面是排障面，英文原样最利于搜索与上报比对；且 wire 透出串本就不可译，半译反而制造混合语言。
+- **日期用 `toLocaleString()`/Intl**：跟随浏览器/OS 语言而非应用语言，切换后必然产生混合文本；字典模板量小且与消息时钟同构。
+- **blank 行参与搜索（匹配本地化标题或存储标题）**：任一选择都在某个语言下"看得见搜不到"；占位行本无信息量，整体排除语义最稳。
+
+## Consequences
+
+- 语言切换全 UI 即时刷新且零重注册；新包接入 = 字典 + declare-merge + `locale: NS` 三步，无手写胶水。
+- 代价：list label 的消费方必须知道 `resolveSlotLabel`（裸读 `options.label` 拿到函数）；类型上 `SlotLabel` 已挡住多数误用。
+- ui-primitives 的中文默认值在英文语言下依旧是中文，**直到消费点传 label**——未迁移包（ui-trajectory 的 JsonTree）显示英文默认恰好符合其整包英文现状。
+- e2e 英文钉死意味着 zh 默认态主要靠包级组件测试与 settings 语言切换用例覆盖，浏览器 e2e 不再验证 zh 文案。

.agents/notes/implemented/architecture/2026-07-30-command-row-copy-contract.i18n.yaml

@@ -0,0 +1,6 @@
+# Bilingual-pair consistency record (docs/i18n/README.md): the git blob hash of each
+# side as of the last confirmed-consistent state. Both languages carry equal authority;
+# after editing either side, bring the other along and re-record with:
+#   pnpm run verify-translation-pairing --write .agents/notes/implemented/architecture/2026-07-30-command-row-copy-contract.md
+2026-07-30-command-row-copy-contract.md: f6d5199389b3907780c501894e2861e6add85e77
+2026-07-30-command-row-copy-contract.zh.md: 4afaf31640c07e88765060681739f262f322769e

.agents/notes/implemented/architecture/2026-07-30-command-row-copy-contract.md

@@ -0,0 +1,35 @@
+# Agent Note: Command row copy is split between the row and the handler
+
+Status: implemented
+
+English | [中文](2026-07-30-command-row-copy-contract.zh.md)
+
+## Problem
+
+The web command row renders `title · summary` from one logged [command lifecycle pair](../../proposed/architecture/2026-07-27-session-projection-and-command-log.md): the title was the dispatched line rebuilt from `command/run` (`/permission workspace-write`) and the summary was `command/done`'s verbatim `text` (`Permission preset: workspace-write.`). Both halves were written without knowing about the other, so the row said the command name twice and its argument twice — the single worst case being the row a user gets for every Access-chip pick.
+
+## Decision
+
+The row's two halves have disjoint jobs, and each side is written to its own half alone.
+
+The row title is the bare command name — no `/`, no arguments. The `/` belongs to the composer's input grammar, not to a settled record, and the argument is not the row's to report: the summary already says what the command did. `GenericCommandCard` keeps the `命令` fallback for a cross-window node whose `command/run` page fell out of the client's window.
+
+A command handler's settlement `text` therefore never labels its value with the command's own name, because the surface that renders it has already said it. `/permission` returns `preset workspace-write`, bare `current preset workspace-write (available: …)`, and for a bad argument `unknown preset "bogus" (available: …)`. Read as a row this is `permission · preset workspace-write`; read as a standalone line — the TUI appends the same text as a notice — it still states which preset now applies.
+
+The rule bans the *label*, not the vocabulary. `Permission preset: workspace-write.` lost because `Permission preset:` is a caption for a value whose caption is already the title. A domain noun that happens to contain the command's name is not a caption and stays: `/plan` keeps `Plan mode off.` and `Plan mode on. Use /plan off to leave.` (`plan · Plan mode off.` names the mode, and the tail is an instruction, not an echo), and `/goal` keeps `Goal cleared.`. A handler that finds itself writing `<Command> <noun>:` in front of its own value is the case this rule catches.
+
+The log is unchanged: `command/run` keeps the structured `name`/`args` split, so a richer registered command row can still render arguments from the same node without a second data channel.
+
+## Alternatives considered
+
+**Keep the dispatched line as the title and only shorten the settlement text.** The argument would still appear on both sides of the separator (`permission workspace-write · preset workspace-write`), which is the repetition complained about.
+
+**Drop the settlement text from the collapsed row instead of the arguments.** It inverts the row's value: the outcome is what a durable record is for, and an error text would then have nowhere to land.
+
+**Have the row strip a leading command name from the settlement text.** Presentation would silently rewrite handler-authored text, and every handler that phrased its outcome differently would defeat the heuristic.
+
+**Ban the command's name from its settlement text outright, rewriting `/plan` and `/goal` to match.** The broader ban costs more than it buys: `Plan mode off.` and `Goal cleared.` are the clearest sentences those outcomes have, in the row and as standalone TUI notices both, and the shortenings that satisfy a name ban (`off.`, `cleared.`) read as fragments. Captions are the redundancy worth removing.
+
+## Consequences
+
+Every command row gets shorter, and the rule scales: a new command's author writes its outcome without knowing which surface renders it, and no surface has to de-duplicate. The cost is that the dispatched arguments leave the collapsed row — while a command is still executing the row shows only its name and `执行中…` — and that the no-caption rule is a convention the reviewer enforces, not a gate. The `/permission` texts are pinned by the permission package's command tests, and the assembled row copy by the [seeded-history](../../../../apps/web/tests/snapshots/seeded-history/command-row.expected.md) web golden, which reaches a real settled command row keylessly because `/permission` runs entirely on the host.

.agents/notes/implemented/architecture/2026-07-30-command-row-copy-contract.zh.md

@@ -0,0 +1,35 @@
+# Agent Note: Command row copy is split between the row and the handler
+
+Status: implemented
+
+[English](2026-07-30-command-row-copy-contract.md) | 中文
+
+## Problem
+
+Web 命令行由一对落库的[命令生命周期事件](../../proposed/architecture/2026-07-27-session-projection-and-command-log.md)渲染出 `标题 · 摘要`：标题是由 `command/run` 重建的分派命令行（`/permission workspace-write`），摘要是 `command/done` 的原样 `text`（`Permission preset: workspace-write.`）。两半各自成文、互不知情，于是一行里命令名出现两次、参数也出现两次——最糟的一例正是用户每次用 Access chip 切换权限时得到的那一行。
+
+## Decision
+
+命令行两半的职责互不重叠，各自只按自己那一半来写。
+
+行标题就是裸命令名——没有 `/`，也没有参数。`/` 属于编辑器的输入语法，不属于一条已落定的记录；参数也不该由这一行来报告：摘要已经说清了这条命令做了什么。对于 `command/run` 那一页已滑出客户端窗口的跨窗口节点，`GenericCommandCard` 仍保留 `命令` 兜底标题。
+
+因此，命令 handler 的落定 `text` 绝不用命令自身的名字给自己的值加标签——渲染它的界面已经说过一次了。`/permission` 返回 `preset workspace-write`，裸调用时返回 `current preset workspace-write (available: …)`，参数非法时返回 `unknown preset "bogus" (available: …)`。作为一行读是 `permission · preset workspace-write`；作为独立一句读——TUI 把同一段 text 作为通知追加——它依然说明了当下生效的是哪个预设。
+
+这条规则禁的是*标签*，不是用词。`Permission preset: workspace-write.` 之所以出局，是因为 `Permission preset:` 是给一个值加的题头，而这个题头正是标题本身。恰好含有命令名的领域名词不是题头，因此保留：`/plan` 仍返回 `Plan mode off.` 与 `Plan mode on. Use /plan off to leave.`（`plan · Plan mode off.` 说的是那个模式，句尾是一条指引，不是回声），`/goal` 仍返回 `Goal cleared.`。真正被这条规则拦下的，是 handler 在自己的值前面写出 `<命令名> <名词>：` 的那一类。
+
+日志本身未变：`command/run` 保留结构化的 `name`／`args` 拆分，因此更丰富的已注册命令行仍可从同一个节点渲染参数，无需第二条数据通道。
+
+## Alternatives considered
+
+**保留分派命令行作标题，只缩短落定文案。** 参数仍会出现在分隔点两侧（`permission workspace-write · preset workspace-write`），而这正是被指出的重复。
+
+**从折叠行中去掉落定文案，而不是去掉参数。** 这颠倒了这一行的价值：持久记录存在的意义就是结果，而错误文案将无处落脚。
+
+**由这一行从落定文案里剥掉开头的命令名。** 呈现层会悄悄改写 handler 写就的文案，而任何换一种措辞表达结果的 handler 都会让这套启发式失效。
+
+**彻底禁止命令名出现在自己的落定文案里，并把 `/plan`、`/goal` 一并改写。** 这种更宽的禁令代价大于收益：无论在行上还是作为独立的 TUI 通知，`Plan mode off.` 与 `Goal cleared.` 都是这些结果最清楚的句子，而满足"禁名字"所需的缩写（`off.`、`cleared.`）读起来只是残句。值得去掉的冗余是题头。
+
+## Consequences
+
+每一条命令行都变短了，而且这条规则可扩展：新命令的作者写结果时无需知道由哪个界面渲染，任何界面也都不必再去重。代价是分派参数离开了折叠行——命令仍在执行时，行上只有名字和 `执行中…`——以及"不加题头"这条规则是靠评审执行的约定，而非门禁。`/permission` 的文案由 permission 包的命令测试钉住，装配后的行文案由 [seeded-history](../../../../apps/web/tests/snapshots/seeded-history/command-row.expected.md) web 预期输出钉住：因为 `/permission` 完全在 host 上执行，它能无密钥地抵达一条真实的落定命令行。

.agents/notes/implemented/architecture/2026-07-30-config-plane-boundaries.i18n.yaml

@@ -0,0 +1,6 @@
+# Bilingual-pair consistency record (docs/i18n/README.md): the git blob hash of each
+# side as of the last confirmed-consistent state. Both languages carry equal authority;
+# after editing either side, bring the other along and re-record with:
+#   pnpm run verify-translation-pairing --write .agents/notes/implemented/architecture/2026-07-30-config-plane-boundaries.md
+2026-07-30-config-plane-boundaries.md: 8a29dcc934126d6e3dfa9c0a6a308ef006017a5a
+2026-07-30-config-plane-boundaries.zh.md: c858b7dd5b6fcd61936c33f1f09d7d2e89a3cfc7

.agents/notes/implemented/architecture/2026-07-30-config-plane-boundaries.md

@@ -0,0 +1,41 @@
+# Agent Note: what the configuration plane exposes, and who may overwrite what
+
+Status: implemented
+
+English | [中文](2026-07-30-config-plane-boundaries.zh.md)
+
+> Scope: the review round over the [web configuration plane](2026-07-30-web-config-plane.md) — which namespaces reach the wire, which callers reach them, and how an editor holding a partial, possibly stale view writes without destroying what it cannot see.
+
+## Problem
+
+The plane worked and was reachable by more callers, and with more authority, than its design claimed.
+
+`trustedHosts` gated only writes, so a declared LAN client could call `settings.describe` — every exposed namespace's configuration — and `credentials.describe`, which reports whether an arbitrary environment-variable name is configured and where it resolves from. That fence is a DNS-rebinding defense and says so; treating it as an authorization boundary for reads was a category error. Separately, the proxy served every registered namespace: the settings seam is deliberately general, so the first plugin to call `settings.register()` for its own configuration would silently become remotely readable and writable, without passing anywhere near a review of the web surface.
+
+The editor was worse than reachable — it was destructive. It reads the redacted descriptor, which by construction omits `role('secret')` fields. Clearing one field rebuilt the whole user section from that redacted copy and sent `settings.replace`, so a stored literal `apiKey` the wire had never returned was deleted as a side effect. Reproduced directly: `{baseURL, reasoning}` in, `apiKey` gone. Row removal took the same path. And nothing carried a version, so two tabs editing one namespace silently overwrote each other; the seam's per-namespace write queue orders writes but cannot tell a fresh writer from one replaying a stale snapshot.
+
+Three smaller defects sat beside them. `llm/adapters-updated` documented contained observer failures but only caught synchronous ones, so an async listener's rejection escaped as an unhandled rejection. llm-deepseek's retry-policy swap disposed its registration before re-registering, publishing an empty route set between the two — an observer saw the provider disappear and come back, despite a comment claiming no such window. And a transport rejection during the page's credential enrichment escaped `load()`, stranding the page in `loading` with no error shown.
+
+## Decision
+
+**Reading configuration is as privileged as writing it.** `settings.describe` and `credentials.describe` join the loopback-only set, so the whole configuration plane stays same-origin until real authentication exists. The model catalog (`llm.providers`, `llm.models`) deliberately does not: it carries provider ids, display names, and model lists — no endpoints, no key state — and a LAN client's model picker needs it. The boundary is asserted over a real HTTP server rather than a hand-assembled request, because the `Host` header a browser actually sends is what decides it.
+
+**The plane serves exactly the namespaces a registered model provider addresses.** `ctx.llm.listConfigurableProviders()` is the allow-list, so the product boundary is enforced rather than inferred from today's plugin set, and a future namespace becomes web-configurable only by joining that directory. An unregistered namespace and an unexposed one answer identically (`settings-not-exposed`), so probing cannot enumerate the registry.
+
+**A caller with a partial view names the field it means.** `Settings.mutate(ns, ops)` applies `set`/`unset` path ops to the section as it stands at the front of the write queue. The client builds ops by diffing its opening snapshot against its draft, so it mentions only fields it can see: a secret absent from both sides produces no op and survives by construction, not by care. `replace` remains the deliberate wholesale reset.
+
+**Staleness is detected, not ordered away.** Each namespace carries a monotonic `revision` over its RAW section; writes may carry `expectedRevision`, and a mismatch rejects with `SettingsConflictError` → `settings-conflict` on the wire, both revisions attached. The editor captures the revision it opened at and, on conflict, tells the user to reopen rather than replaying its snapshot.
+
+**The raw layer gets its own event.** `settings/updated` stays gated on the resolved value — that is what a consumer means by change. `settings/document-updated (ns, revision)` fires on any raw-section change, because a configuration surface must learn that a field went from inherited to overridden (same resolved value, different meaning) and that its held revision is stale. The host frame `host/settings-changed` now rides this event, and a change to an exposed provider namespace also emits `host/models-changed`: that namespace holds the provider's catalog, which no route change announces.
+
+## Alternatives considered
+
+- **A deployment-declared namespace allowlist on the proxy config** — more general, but it moves the product boundary to whoever writes cordis.yml, and an empty default would break the shipped page until every deployment opted in. The provider directory already states exactly which namespaces are model configuration.
+- **Opt-in metadata at `settings.register()`** — the most honest semantics (the namespace's owner declares its own exposure), and the largest change: the seam's public interface, both LLM plugins, and their docs. Recorded as the shape to adopt if a non-LLM namespace ever needs the plane.
+- **Distinguishing "unregistered" from "registered but unexposed"** — better diagnostics, and a namespace-enumeration oracle. The uniform answer is deliberate.
+- **Detecting conflicts by diffing instead of a revision** — comparing the submitted base against storage would work for whole-section writes, but the editor holds a REDACTED section: it cannot produce a comparable base, which is the same reason it cannot safely `replace`. A counter needs neither.
+- **Fixing the redaction gaps in this round** — `redactSecrets` walks only `object`/`dict`/`array`, so a secret behind a union, intersection, or transform is returned verbatim with an empty `secrets` list; `schema.toJSON()` carries a secret field's `.default(...)`; write-rejection messages return schema text that may quote the input; the client rehydrates the envelope through schemastery's `new Function`; and pi-ai's plain-string `headers` dict can legitimately hold `Authorization`. All confirmed, all deliberately left for a fail-closed `describeForWire()` that refuses a schema it cannot prove safe. They are recorded as `TODO(settings-wire-redaction)` and in the owning READMEs' Known Limitations rather than half-fixed here.
+
+## Consequences
+
+A LAN client on a `trustedHosts` deployment can no longer render the settings page at all; loopback is the configuration surface. A plugin that registers a settings namespace is not web-configurable until it also registers a configurable provider — deliberate, and the reason `settings-not-exposed` names the boundary in its message. `SettingsDescriptor` gained a required `revision`, so any programmatic constructor of a descriptor-shaped value must supply it, and `settings/document-updated` is a new event any provider-side listener may now observe. Clients that ignore `expectedRevision` keep last-write-wins semantics unchanged. Deferred: the fail-closed wire describe (with the `headers` and envelope-sanitization work it carries), and a non-executable browser schema protocol.

.agents/notes/implemented/architecture/2026-07-30-config-plane-boundaries.zh.md

@@ -0,0 +1,41 @@
+# Agent Note：配置面暴露什么，以及谁有权覆盖什么
+
+Status: implemented
+
+[English](2026-07-30-config-plane-boundaries.md) | 中文
+
+> 范围：针对 [Web 配置面](2026-07-30-web-config-plane.md)的评审轮——哪些 namespace 能抵达协议、哪些调用方能抵达它们，以及一个只持有局部、且可能过期视图的编辑器该如何写入，才不会毁掉它看不见的东西。
+
+## 问题
+
+这个面能用，但能触达它的调用方、以及它们所拥有的权限，都比设计声称的更多。
+
+`trustedHosts` 只拦住了写入，因此一个已声明的 LAN 客户端可以调用 `settings.describe`——拿到每个已暴露 namespace 的配置——以及 `credentials.describe`，后者会报告任意一个环境变量名是否已配置、又从何处解析。那道 fence 是 DNS 重绑定防御，它自己也是这么写的；把它当作读取的授权边界，是一次范畴错误。另一件事是：代理服务于每一个已注册的 namespace。settings seam 是刻意做成通用的，因此第一个为自身配置调用 `settings.register()` 的插件，就会悄无声息地变成可远程读写，而完全不必经过任何针对 Web 表层的评审。
+
+编辑器比"可触达"更糟——它是破坏性的。它读到的是脱敏后的 descriptor，后者按构造省略了 `role('secret')` 字段。清空其中一个字段，会用这份脱敏副本重建整个用户分节并发出 `settings.replace`，于是一个协议从未回传过的已存字面 `apiKey` 被顺带删除。这一点被直接复现：输入 `{baseURL, reasoning}`，输出时 `apiKey` 消失。删除整行走的是同一条路径。而且没有任何东西携带版本，因此两个标签页编辑同一个 namespace 会静默互相覆盖；seam 的逐 namespace 写队列只排定写入次序，分辨不出一个新写方与一个重放过期快照的写方。
+
+另有三个较小的缺陷与之并列。`llm/adapters-updated` 的文档写着观察者失败会被收容，却只捕获同步失败，于是异步 listener 的 rejection 作为 unhandled rejection 逃逸。llm-deepseek 的重试策略换路由先释放注册、再重新注册，在两者之间发布了一个空路由集——观察者会看到该提供方消失又回来，尽管注释宣称不存在这样的空窗。还有，页面做凭据增强时的传输层 rejection 会逃出 `load()`，把页面卡在 `loading` 且不显示任何错误。
+
+## 决策
+
+**读配置与写配置同样特权。**`settings.describe` 与 `credentials.describe` 加入仅限回环的集合，因此在真正的认证层出现之前，整个配置面都保持同源。模型目录（`llm.providers`、`llm.models`）刻意不在其中：它携带的是提供方 id、显示名与模型列表——没有端点、没有密钥状态——而 LAN 客户端的模型选择器正需要它。这条边界由一台真实 HTTP 服务器来断言，而不是手工拼装的请求，因为真正决定它的，是浏览器实际发出的那个 `Host` 头。
+
+**这个面恰好服务于已注册模型提供方所指向的那些 namespace。**`ctx.llm.listConfigurableProviders()` 就是允许列表，于是产品边界是被执行的，而不是从今天的插件集合里推断出来的；将来的 namespace 只有加入该目录才会变得可在 Web 上配置。未注册的 namespace 与未暴露的 namespace 得到完全相同的答复（`settings-not-exposed`），因此探测无法枚举注册表。
+
+**持有局部视图的调用方，点名它真正要改的字段。**`Settings.mutate(ns, ops)` 会把 `set`/`unset` 路径 op 施加在写入排到队首那一刻的分节上。客户端通过对比自己打开时的快照与草稿来构造 op，因此它只提及自己看得见的字段：两侧都没有的机密不会产生任何 op，它的留存是构造使然，而非小心使然。`replace` 仍是那个刻意的整体重置。
+
+**过期是被检测出来的，而不是靠排序绕过去的。**每个 namespace 都带有一个针对其**原始**分节的单调 `revision`；写入可携带 `expectedRevision`，不匹配即以 `SettingsConflictError` 拒绝——在协议上是 `settings-conflict`，并附上两个 revision。编辑器记住自己打开时的 revision，冲突时请用户重新打开，而不是把自己的快照重放上去。
+
+**原始层拥有自己的事件。**`settings/updated` 仍以解析值为门槛——那才是消费方所说的"变化"。`settings/document-updated (ns, revision)` 则在任何原始分节变化时触发，因为配置界面必须知道某个字段从继承变成了覆盖（解析值相同，含义不同），也必须知道自己持有的 revision 已经过期。host 帧 `host/settings-changed` 现在搭乘这个事件；而已暴露提供方 namespace 的变更还会额外发出 `host/models-changed`：该 namespace 正持有这个提供方的目录，而没有任何路由变更会宣告它。
+
+## 曾考虑的替代方案
+
+- **在代理配置上做部署声明式的 namespace 白名单**——更通用，但它把产品边界交给了写 cordis.yml 的人，而空的默认值会让已交付的页面在每个部署显式开启之前直接失效。提供方目录本就精确地说明了哪些 namespace 属于模型配置。
+- **在 `settings.register()` 处 opt-in metadata**——语义最正（由 namespace 的属主自行声明其暴露与否），改动也最大：seam 的公共接口、两个 LLM 插件，以及它们的文档。记录为：一旦某个非 LLM 的 namespace 确实需要这个面，就采用这个形状。
+- **区分"未注册"与"已注册但未暴露"**——诊断更好，同时也是一台 namespace 枚举预言机。统一答复是刻意为之。
+- **用 diff 而非 revision 来检测冲突**——对整分节写入而言，拿提交时的基线与存储比对是可行的，但编辑器持有的是**脱敏后**的分节：它给不出可比对的基线，这与它不能安全地 `replace` 是同一个原因。计数器两者都不需要。
+- **本轮就修掉脱敏的缺口**——`redactSecrets` 只遍历 `object`/`dict`/`array`，因此藏在 union、intersection 或 transform 之后的机密会被原样返回，且 `secrets` 列表为空；`schema.toJSON()` 会带上 secret 字段的 `.default(...)`；写入拒绝的消息返回的是可能引用了输入的 schema 文本；客户端通过 schemastery 的 `new Function` 重建信封；而 pi-ai 那个纯字符串的 `headers` 字典完全可以合法地放下 `Authorization`。全部经确认属实，也全部刻意留给一个 fail-closed 的 `describeForWire()`——它会拒绝自己无法证明安全的 schema。它们被记录为 `TODO(settings-wire-redaction)` 以及各属主 README 的 Known Limitations，而不是在这里做一半。
+
+## 影响
+
+`trustedHosts` 部署下的 LAN 客户端已经完全无法渲染设置页；配置表层就是回环。注册了 settings namespace 的插件，在它同时注册可配置提供方之前不会变得可在 Web 上配置——这是刻意的，也正是 `settings-not-exposed` 要在消息里点明这条边界的原因。`SettingsDescriptor` 新增了必填的 `revision`，因此以编程方式构造 descriptor 形状值的地方都必须提供它；`settings/document-updated` 是一个新事件，provider 侧的任何 listener 现在都可以观察它。忽略 `expectedRevision` 的客户端，其后写胜出的语义完全不变。延后事项：fail-closed 的协议 describe（连同它所承载的 `headers` 与信封净化工作），以及一套客户端无法执行的浏览器 schema 协议。

.agents/notes/implemented/architecture/2026-07-30-credential-boundaries-and-atomic-registration.i18n.yaml

@@ -0,0 +1,6 @@
+# Bilingual-pair consistency record (docs/i18n/README.md): the git blob hash of each
+# side as of the last confirmed-consistent state. Both languages carry equal authority;
+# after editing either side, bring the other along and re-record with:
+#   pnpm run verify-translation-pairing --write .agents/notes/implemented/architecture/2026-07-30-credential-boundaries-and-atomic-registration.md
+2026-07-30-credential-boundaries-and-atomic-registration.md: 6fe5f554acbfd804db9625fcaa794d513c8799c4
+2026-07-30-credential-boundaries-and-atomic-registration.zh.md: 3eb3b022064124aad2a389abba3063af4e2110fa

.agents/notes/implemented/architecture/2026-07-30-credential-boundaries-and-atomic-registration.md

@@ -0,0 +1,36 @@
+# Agent Note: credential boundaries, whole-snapshot requests, and atomic route registration
+
+Status: implemented
+
+English | [中文](2026-07-30-credential-boundaries-and-atomic-registration.zh.md)
+
+> Scope: the third review round over the [request-level LLM configuration seam](2026-07-29-request-level-llm-config-credentials.md) — where a stored credential lives and who can read it, how one request's facts stay one generation, and how a route set changes without a window. Companion to the [settings write-path note](2026-07-30-settings-write-path-integrity.md), whose provider fixes this round applies to `credentials-local` and whose writer lock it promotes into `dsh-atomic-write`.
+
+## Problem
+
+Review found the credential path leaking across boundaries it had drawn. The shipped surfaces hoisted `$DSH_HOME/.env` into `process.env` before cordis booted, so on the next run `credentials-local` classified every key it had stored itself as a read-only ambient launch override: `describe()` reported `source: 'env'` with `writable: false`, `set`/`unset` rejected as shadowed, and a key stored from the web page or TUI became unrotatable and undeletable while the adapter kept using the value captured at launch. The store's own write path repeated the settings-local defects that same review round fixed (two independent chains, whole-file render from a stale cache), plus editor bugs of its own: a physical line inside another key's quoted multi-line value read as an assignment, CRLF endings degraded to LF, a multi-line entry reported `writable: true` while `set` always threw, and `credentials/updated` was emitted bare after the commit, so one broken observer made a durable write look failed. On the read side, the file's `0600` mode stops other OS users but not the model, whose bash and filesystem tools run as the same user.
+
+Two request-path defects sat beside them. DeepSeek's per-request resolution kept connection facts in a last-good snapshot but re-read the literal `apiKey` from the raw configuration, so a settings generation the resolver rejected could still put its key on the previous generation's endpoint. pi-ai handed the SDK `undefined` when a configured `apiKeyEnv` resolved to nothing, letting pi-ai's own environment discovery authenticate with an unrelated provider key — another tenant, silently billed. And its route swap disposed the old registration before creating the new one: a route another adapter owned dropped every existing route, after which the facts cache could equal the registry's, so restoring the working configuration never re-applied.
+
+## Decision
+
+**`$DSH_HOME/.env` belongs to the credential provider alone.** No surface loads it into `process.env`. The genuine launch environment and the invoking directory's `.env` (loaded by the bin) stay the read-only ambient layer, so a composition without the provider resolves keys exactly as before, while a stored key stays file-sourced and writable across restarts — proven by a real restart in the loader composition rather than by a unit assertion about `describe()`.
+
+**The stored credential has no boundary against the model, and the READMEs say so.** `0600` under a `0700` directory stops other OS users; the model's bash and filesystem tools run as that same user, and the shipped default confines nothing. What the harness does hold to is narrower and stated as exactly that: no surface hoists the document into `process.env`, and the model is never handed a resolved path to it, so reaching the value takes a deliberate read of a path it was not given. An OS-keychain provider — a store the model's processes cannot read at all — is recorded as the real answer rather than implied by a partial one.
+
+**One request, one generation.** DeepSeek's resolved snapshot carries the credential facts (literal key and reference) beside the endpoint, and `resolveApiKey` receives that snapshot instead of re-reading configuration. A rejected generation now contributes nothing at all. pi-ai defers to provider-native discovery only for a profile naming no credential; a configured reference that misses fails with `MISSING_CREDENTIAL` naming the route and the reference. The boot-time credential probe is deleted: it could run before the credentials service mounted and reported every failure as a missing key, while the first request already gives the accurate error.
+
+**Route replacement is a registry operation, not a caller sequence.** `registerAdapter` returns a handle carrying `replace(providers)`: the candidate set is validated in full first (conflicts, names, provider metadata), then swapped in one synchronous section. A refused replacement leaves the previous routes registered and serving, and the caller's facts cache only advances after the registry actually holds the new set, so reverting to a working configuration re-applies. pi-ai's registration facts are sorted by provider, so a settings document that merely reorders its keys is no longer a route change.
+
+**Contained publication for committed credential writes.** `Credentials.notifyUpdated` fans `credentials/updated` out one listener at a time; sync throws and async rejections are logged without changing the committed operation's outcome, and `INVARIANT`-coded failures rethrow after every listener ran — the same shape the settings seam uses for `settings/updated`. `installSettingsSection`'s cleanup now distinguishes its two triggers: a provider detaching still falls back to the composition entry and re-derives, while the consumer's own unload returns immediately instead of re-registering routes during teardown.
+
+## Alternatives considered
+
+- **A sandbox read-denial naming `$DSH_HOME/.env`** — implemented as a `readDenyPaths` policy field (a trailing SBPL `deny file-read* file-write*`, a `/dev/null` bwrap bind) and withdrawn on its own evidence. bwrap must create that bind's mount point inside a tree its profile has already made read-only, so it refuses the entire confinement whenever the parent directory is absent — every host that has not stored a credential yet, including a fresh install; Landlock cannot subtract from its own `/` read grant, so every confined call would report `partial` for a file it never hid. A protection that breaks confinement where it works and misreports it where it does not is worse than a documented absence. Denying the whole harness home was rejected earlier for a separate reason: it also covers `sessions/`, and `DSH_SESSION_JSONL` is a documented model-visible capability.
+- **Removing `DSH_HOME` from the model's bash environment** — considered as defense in depth and rejected as theater with a real cost: the default home is a documented convention the agent can reconstruct, while the variable is how legitimate tooling finds harness state. There is no boundary here for it to complement; hiding the pointer would only make the absence harder to see.
+- **Shipping the OS-keychain provider in this round** — it is the only design where the model's processes genuinely cannot read the secret, and it is a sibling package with three platform backends. Sizing it against the rest of this review round would have delayed every other fix; it is recorded as the deferred answer, not as a maybe.
+- **A `replaceRegistration(previous, next)` service method** — the review's shape, but it makes the caller carry the previous handle and lets it pass a mismatched one. Hanging `replace` on the registration handle makes ownership structural: only the registration that holds routes can replace them.
+
+## Consequences
+
+`update()`-adjacent behavior gained documented failure modes: a credential write can now fail on the lock deadline or on an unparsable on-disk document, and `describe()` reports `writable: false` for multi-line entries it will not rewrite. `LlmAdapter` registrants keep working unchanged (the handle is still callable as the disposer), and `DeepSeekConnectionOptions` gained credential fields, so a programmatic constructor of the adapter must supply `apiKeyEnv`. Deferred: the OS-keychain credential provider, and per-value revision checks for two writers editing one reference (last-write-wins remains the documented resolution).

.agents/notes/implemented/architecture/2026-07-30-credential-boundaries-and-atomic-registration.zh.md

@@ -0,0 +1,40 @@
+# Agent Note: 凭据边界、按整份快照发起的请求与原子路由注册
+
+Status: implemented
+
+[English](2026-07-30-credential-boundaries-and-atomic-registration.md) | 中文
+
+> 范围：对[请求级 LLM（大语言模型）配置 seam](2026-07-29-request-level-llm-config-credentials.md)的第三轮评审——存下来的凭据落在哪里、谁能读到它，一次请求的事实如何保持为同一代，以及一组路由如何在不留空窗的前提下更换。本 note 与 [settings 写路径 note](2026-07-30-settings-write-path-integrity.md) 配套：本轮把那篇 note 的提供方修复套用到 `credentials-local`，并把其中的写锁提升进 `dsh-atomic-write`。
+
+## 问题
+
+评审发现，凭据路径正在越过它自己划下的边界泄漏。已交付的各个面在 Cordis 启动之前就把 `$DSH_HOME/.env` 提升进了 `process.env`，于是下一次运行时，`credentials-local` 会把它自己存下的每个键都判成来自环境的只读启动覆盖：`describe()` 报告 `source: 'env'` 且 `writable: false`，`set`/`unset` 以被遮蔽为由拒绝，从 web 页面或 TUI 存入的密钥既无法轮换也无法删除，而适配器还在继续使用启动时捕获的那个值。
+
+存储自身的写路径重演了同一轮评审在 settings-local 修掉的那些缺陷（两条相互独立的链、从陈旧缓存渲染整份文件），还叠加了编辑器自己的缺陷：另一个键的带引号多行值内部的一条物理行会被读成赋值，CRLF 行尾会退化成 LF，多行条目报告 `writable: true` 而 `set` 总是抛错，`credentials/updated` 又在提交之后裸发，于是一个出错的观察者就能让一次已经落盘的写入看起来失败。
+
+在读取一侧，文件的 `0600` 权限挡得住其他 OS 用户，却挡不住模型：它的 bash 与文件系统工具就以同一个用户身份运行。
+
+与之并排的还有两个请求路径缺陷。DeepSeek 的按请求解析把连接事实保存在最后可用快照里，却仍从原始配置重新读取字面 `apiKey`，于是被 resolver 拒绝的那一代设置，照样能把自己的密钥送到上一代的端点上。配置了 `apiKeyEnv` 却解析不到值时，pi-ai 会把 `undefined` 交给 SDK，让 pi-ai 自己的环境发现拿一个毫不相干的提供方密钥完成鉴权——那是另一个租户，账单还悄悄记在它头上。而且它的路由替换是先释放旧注册、再创建新注册：只要有一条路由已被别的适配器占有，现有路由就会被全部丢掉，此后事实缓存可能与注册表中的事实相等，于是把配置改回可用状态也不会重新生效。
+
+## 决策
+
+**`$DSH_HOME/.env` 只归凭据提供方所有。**没有任何一个面会把它加载进 `process.env`。真正的启动环境，以及调用目录中由 bin 加载的 `.env`，仍然是那一层只读的环境来源，因此不挂载该提供方的组合，解析密钥的方式与从前完全一致，而存下的密钥跨重启仍然来源于文件、仍然可写——这一点由 Loader 组合中的一次真实重启来证明，而不是靠对 `describe()` 的单元断言。
+
+**存下的凭据对模型没有边界，而 README 就是这么写的。**`0700` 目录下的 `0600` 挡得住其他 OS 用户；模型的 bash 与文件系统工具正是以同一用户身份运行，而已交付的默认值不约束任何东西。harness 真正守住的更窄，也就照这个宽度写下来：没有任何一个面会把该文档提升进 `process.env`，模型也从不会拿到它的解析后路径，因此要拿到这个值，需要刻意去读一条并未交给它的路径。OS 钥匙串（keychain）提供方——一个模型的进程根本读不到的存储——被记录为真正的答案，而不是靠一个残缺的方案去暗示它。
+
+**一次请求，一代设置。**DeepSeek 解析出的快照在端点旁一并携带凭据事实（字面密钥与引用），`resolveApiKey` 接收这份快照，而不再重新读取配置。被拒绝的那一代如今完全不再贡献任何东西。只有当一个 profile 完全没有点名凭据时，pi-ai 才交给提供方原生的发现流程；配置了引用却解析不到，就以 `MISSING_CREDENTIAL` 失败，并点名该路由与该引用。启动时的凭据探测被删除：它可能在凭据服务挂载之前就运行，并把每一种失败都报成密钥缺失，而第一次请求本就会给出准确的错误。
+
+**路由替换是注册表的操作，不是调用方的一串步骤。**`registerAdapter` 返回一个携带 `replace(providers)` 的句柄：候选集合先被完整校验（冲突、名称、提供方元数据），再在一个同步区段内完成替换。被拒绝的替换会让先前的路由保持注册并继续服务，而调用方的事实缓存只有在注册表确实持有新集合之后才会推进，因此改回可用配置时会重新生效。pi-ai 的注册事实按提供方排序，因此仅仅调换键顺序的设置文档不再算作路由变更。
+
+**已提交的凭据写入采用收容式发布。**`Credentials.notifyUpdated` 逐个监听器扇出 `credentials/updated`；同步抛错与异步 rejection 都只记日志，不改变已提交操作的结果，而带 `INVARIANT` 代码的失败会在每个监听器都运行完之后重抛——与 settings seam 处理 `settings/updated` 的形状相同。`installSettingsSection` 的清理现在会区分它的两个触发来源：提供方脱离时仍回退到组合的 entry 配置并重新推导，而消费方自身卸载时立即返回，不再在拆卸过程中重新注册路由。
+
+## 曾考虑的替代方案
+
+- **用沙箱点名拒读 `$DSH_HOME/.env`**——已按 `readDenyPaths` 策略字段实现过（末尾一条 SBPL `deny file-read* file-write*`、一条 `/dev/null` 的 bwrap bind），又被它自己的证据推翻。bwrap 必须在自己 profile 已经置为只读的目录树内部创建该 bind 的挂载点，因此只要父目录不存在，它就会拒绝整次约束——那是每一台还没有存过凭据的主机，包括全新安装；Landlock 无法从它自己对 `/` 的读取授权中减去任何东西，于是每一次受限调用都会为一个它其实从未藏起的文件报 `partial`。一项在生效之处破坏约束、在不生效之处误报的保护，比一条写明的「没有保护」更糟。至于拒掉整个 harness home，早先另有理由被否：它同时覆盖 `sessions/`，而 `DSH_SESSION_JSONL` 是一项成文的、模型可见的能力。
+- **把 `DSH_HOME` 从模型的 bash 环境中移除**——作为纵深防御考虑过，最终按「有真实代价的表演」不予采纳：默认 home 是 agent（智能体）能自行重建的成文约定，而这个变量正是正当工具链定位 harness 状态的途径。这里并不存在一条需要它来补强的边界，藏起指针只会让这份缺席更难被看见。
+- **本轮就交付 OS 钥匙串提供方**——只有这个设计能让模型的进程真正读不到机密，而它是一个带三种平台后端的兄弟包（package）。把它与本轮评审的其余工作放在一起评估体量，会拖慢其他每一项修复；它被记录为那个延后的答案，而不是一个「也许」。
+- **做成 `replaceRegistration(previous, next)` 服务方法**——这是评审给出的形状，但它要求调用方自行携带上一个句柄，也允许它传入一个不匹配的句柄。把 `replace` 挂在注册句柄上，让归属关系变成结构性的：只有持有路由的那一项注册才能替换它们。
+
+## 后果
+
+`update()` 邻近的行为多了成文的失败模式：凭据写入现在可能因锁截止时间到期、或磁盘文档无法解析而失败，`describe()` 对它不会改写的多行条目报告 `writable: false`。`LlmAdapter` 的注册方无需改动即可继续工作（句柄本身仍可当作释放器调用），`DeepSeekConnectionOptions` 则新增了凭据字段，因此以编程方式构造该适配器必须提供 `apiKeyEnv`。延后事项：OS 钥匙串凭据提供方，以及针对两个写方编辑同一引用的逐值修订号检查（后写胜出仍是成文的解决方式）。

.agents/notes/implemented/architecture/2026-07-30-session-end-seed-log-boundary.i18n.yaml

@@ -2,5 +2,5 @@
 # side as of the last confirmed-consistent state. Both languages carry equal authority;
 # after editing either side, bring the other along and re-record with:
 #   pnpm run verify-translation-pairing --write .agents/notes/implemented/architecture/2026-07-30-session-end-seed-log-boundary.md
-2026-07-30-session-end-seed-log-boundary.md: 837531ba0bd3ecf404eb47ee933438546c682a54
-2026-07-30-session-end-seed-log-boundary.zh.md: 33680c1845364de62e5b53ead13de418a389f908
+2026-07-30-session-end-seed-log-boundary.md: 268646e192d0b8e0a5dde03957a18ef155b7038e
+2026-07-30-session-end-seed-log-boundary.zh.md: dca87e16de5e567ff85d2b32b8243f76ebed1c4a

.agents/notes/implemented/architecture/2026-07-30-session-end-seed-log-boundary.md

@@ -14,13 +14,13 @@ Crash repair does not close the gap and must not: `interruptedTurnClosers` synth
 
 ## Decision
 
-`Session`'s constructor appends the log-only `session/end-seed` event immediately after a non-empty constructor seed, as the seeded session's first live write at the seq `firstLiveSeq` names. The event is the durable projection of that field: `firstLiveSeq` answers where this lifecycle's writes start for a consumer holding the object, while `session/end-seed` answers the same question for one holding only stored bytes. Its payload is empty — position and `time` carry the whole meaning — and it is not a `SurfaceEventType`, so it produces no message and cannot perturb derived history.
+`Session`'s constructor appends the log-only `session/end-seed` event immediately after an explicitly supplied constructor seed, including an empty one, as the seeded session's first live write at the seq `firstLiveSeq` names. The event is the durable projection of that field: `firstLiveSeq` answers where this lifecycle's writes start for a consumer holding the object, while `session/end-seed` answers the same question for one holding only stored bytes. Its payload is empty — position and `time` carry the whole meaning — and it is not a `SurfaceEventType`, so it produces no message and cannot perturb derived history. The seq-0 marker distinguishes an empty resumed session from a genuinely fresh session, preventing new-session defaults from being applied during resume.
 
 A bracket owner reads it positionally: an unmatched opening marker before `session/end-seed` has a smaller seq, came from the constructor seed, and belongs to a lifecycle that has ended. Core writes the boundary and reads nothing from it; each bracket's vocabulary stays with its owning plugin, so no core predicate helper ships without a consumer to shape it.
 
 The constructor is the placement because it is the single waist every seeded session passes through. All six entry points reach it: `agents.resume()`, config-driven startup on a persisted id (`restoreOrCreateConfigured`), `sessions.fork()`, a subagent fork child, `coordinator.adopt()`'s live-prefix path, and a bare `sessions.create(id, {seed})`. A boundary written at persistence load would miss both fork paths — and a forked child inheriting a still-running parent's open `compact/start` is precisely the case that must be classifiable. A boundary written at loop start would miss `fork()` and `adopt()`, and would have to fire on `SessionStartSource: 'startup'`, which is what a fork child publishes, so that field would stop discriminating.
 
-Two guards keep the marker from becoming noise. An empty seed writes nothing because there is no seed to end. A seed already ending in one is not re-marked, which makes the write idempotent. Idempotence is load-bearing rather than tidiness — `agentFor()` resumes a cold session on first touch, so merely opening one in a client is a pickup, and without the guard browsing would grow a log by one event per visit.
+Two guards keep the marker precise. An omitted seed writes nothing because the session is fresh. A seed already ending in one is not re-marked, which makes the write idempotent. Idempotence is load-bearing rather than tidiness — `agentFor()` resumes a cold session on first touch, so merely opening one in a client is a pickup, and without the guard browsing would grow a log by one event per visit.
 
 ## Persistence needs no changes
 
@@ -48,7 +48,7 @@ The predicate holds for a bracket *this* session inherited, not as a liveness si
 
 Bought: one boundary, written in one place, correct for all six seeded-start paths — including the fork gap the persistence-layer version could not reach. The persistence packages keep a pure read path. `firstLiveSeq` gains a durable twin rather than a second, competing notion of the same boundary.
 
-Cost: a seeded session's log is one event longer, which moved seq expectations in tests across nine packages (session, agent-loop, persistence contract, jsonl, session-query, session-title, subagent-inprocess, telemetry, token-meter). Two of those updates are load-bearing rather than mechanical: telemetry's adoption tests now assert the boundary IS exported, because it is this lifecycle's own write, and the property suite's replay invariant is restated as "seed reproduced verbatim, plus one log-only boundary" with idempotence added as its own property.
+Cost: a seeded session's log is one event longer, including an empty resumed log. Seq expectations move with that boundary. Two updates are load-bearing rather than mechanical: telemetry's adoption tests assert the boundary IS exported, because it is this lifecycle's own write, and the property suite's replay invariant is "seed reproduced verbatim, plus one log-only boundary" with idempotence as its own property.
 
 `session/end-seed` joins the on-disk vocabulary. Under the pre-release stance (`SESSION_FORMAT_VERSION` pinned at `0`, no compatibility promise) older logs simply lack it, and a log without a boundary correctly classifies nothing as constructor-seed history.
 

.agents/notes/implemented/architecture/2026-07-30-session-end-seed-log-boundary.zh.md

@@ -14,13 +14,13 @@ Status: implemented
 
 ## Decision
 
-`Session` 的构造函数紧接非空构造种子之后追加仅日志事件 `session/end-seed`，作为带种子会话的第一次实时写入，位置正是 `firstLiveSeq` 指出的 seq。该事件是那个字段的持久投影：`firstLiveSeq` 为持有对象的消费方回答本生命周期的写入从哪里开始，`session/end-seed` 则为只持有存储字节的消费方回答同一问题。它的 payload 为空——位置与 `time` 承载全部含义——并且不是 `SurfaceEventType`，因此不产生消息，也无法扰动派生历史。
+`Session` 的构造函数紧接显式传入的构造种子（包括空种子）之后追加仅日志事件 `session/end-seed`，作为带种子会话的第一次实时写入，位置正是 `firstLiveSeq` 指出的 seq。该事件是那个字段的持久投影：`firstLiveSeq` 为持有对象的消费方回答本生命周期的写入从哪里开始，`session/end-seed` 则为只持有存储字节的消费方回答同一问题。它的 payload 为空——位置与 `time` 承载全部含义——并且不是 `SurfaceEventType`，因此不产生消息，也无法扰动派生历史。这个 seq-0 标记把从空日志恢复的会话与真正的全新会话区分开来，从而防止恢复期间应用新会话默认值。
 
 括号所有方按位置读取它：在 `session/end-seed` 之前的未配对开启标记具有更小的 seq，来自构造种子，并且属于一个已结束的生命周期。核心写入该边界但不从中读取任何内容；每个括号的词汇表仍归其所属插件，因此在没有消费方来塑形之前，核心不会先发布谓词辅助函数。
 
 选择构造函数，是因为它是每一个带种子会话都必经的唯一收窄处。全部六个入口都会到达它：`agents.resume()`、在已持久化 id 上的配置驱动启动（`restoreOrCreateConfigured`）、`sessions.fork()`、子代理 fork 子会话、`coordinator.adopt()` 的实时前缀路径，以及裸的 `sessions.create(id, {seed})`。在持久化加载时写入的边界会漏掉两条 fork 路径——而一个继承了仍在运行的父会话开放 `compact/start` 的 fork 子会话，恰恰是必须可判定的场景。在 loop 启动时写入的边界会漏掉 `fork()` 与 `adopt()`，并且不得不在 `SessionStartSource: 'startup'` 上触发——那正是 fork 子会话发布的取值，于是该字段将不再具有区分力。
 
-两条守卫让这个标记不至于变成噪声。空种子不写入任何内容，因为没有种子需要结束。种子本身已以该事件结尾时不会重复标记，这让写入具备幂等性。幂等性是承重的，而不是为了整洁——`agentFor()` 会在首次触碰时恢复一个冷会话，因此在客户端里仅仅打开一个会话就是一次接手；没有这条守卫，浏览会让日志每访问一次就增长一个事件。
+两条守卫让这个标记保持精确。省略种子时不写入任何内容，因为这是全新会话。种子本身已以该事件结尾时不会重复标记，这让写入具备幂等性。幂等性是承重的，而不是为了整洁——`agentFor()` 会在首次触碰时恢复一个冷会话，因此在客户端里仅仅打开一个会话就是一次接手；没有这条守卫，浏览会让日志每访问一次就增长一个事件。
 
 ## 持久化无需任何改动
 
@@ -48,7 +48,7 @@ Status: implemented
 
 买到的：一条边界，在一处写入，对全部六条带种子启动路径都正确——包括持久化层方案触及不到的 fork 缺口。持久化各包保留纯读取路径。`firstLiveSeq` 获得一个持久孪生体，而不是关于同一边界的第二套彼此竞争的概念。
 
-代价：带种子会话的日志长了一个事件，这在九个包（session、agent-loop、持久化契约、jsonl、session-query、session-title、subagent-inprocess、telemetry、token-meter）里挪动了 seq 期望。其中两处更新是承重的而非机械的：telemetry 的收养测试现在断言该边界*会*被导出，因为它是本生命周期的自有写入；而属性测试套件的重放不变式被重述为"种子逐字节复现，外加一个仅日志边界"，并把幂等性补成一条独立属性。
+代价：带种子会话的日志长了一个事件，空日志恢复也包括在内。seq 期望会随这条边界移动。两处更新是承重的而非机械的：telemetry 的收养测试断言该边界*会*被导出，因为它是本生命周期的自有写入；属性测试套件的回放不变式则是"种子逐字节复现，外加一个仅日志边界"，并把幂等性作为独立属性。
 
 `session/end-seed` 加入了落盘词汇表。在预发布立场下（`SESSION_FORMAT_VERSION` 固定为 `0`，不作兼容承诺），更旧的日志只是没有它，而没有边界的日志会正确地判定没有任何内容属于构造种子历史。
 

.agents/notes/implemented/architecture/2026-07-30-settings-write-path-integrity.i18n.yaml

@@ -0,0 +1,6 @@
+# Bilingual-pair consistency record (docs/i18n/README.md): the git blob hash of each
+# side as of the last confirmed-consistent state. Both languages carry equal authority;
+# after editing either side, bring the other along and re-record with:
+#   pnpm run verify-translation-pairing --write .agents/notes/implemented/architecture/2026-07-30-settings-write-path-integrity.md
+2026-07-30-settings-write-path-integrity.md: 7bd50adc8a812759c3ae3f80a50978d75baa712a
+2026-07-30-settings-write-path-integrity.zh.md: da07745ef1de1b694fc3fd1ee3d04322cdadc992

.agents/notes/implemented/architecture/2026-07-30-settings-write-path-integrity.md

@@ -0,0 +1,35 @@
+# Agent Note: settings write-path integrity and observer lifecycle
+
+Status: implemented
+
+English | [中文](2026-07-30-settings-write-path-integrity.zh.md)
+
+> Scope: the third review round over `packages/settings/` — write-path data integrity in `dsh-settings-local` (operation chain, read-modify-write, cross-process writer lock, diff-shaped YAML edits) and observer lifecycle in `dsh-settings` (watch disposal, async listener containment, the JSON-shape write boundary). This note reverses one deferral recorded in the [user-settings seam note](2026-07-28-user-settings-seam.md): the cross-process lockfile now ships.
+
+## Problem
+
+Review found the provider's write path could destroy state it never observed, and the seam's observer lifecycle leaked past disposal. Concretely: watcher reloads and document writes ran on two independent promise chains while every write rendered the whole next document from the cached text, so an external edit still inside the debounce window was overwritten — and the follow-up reload no-oped because the post-rename content matched the cache, erasing the edit without a trace. The initial `load()` raced the watcher's own setup, leaving a startup window whose changes never fire an event. Two processes sharing a harness home rendered from independent caches, last writer winning whole namespaces. On the seam side, a `watch()` disposer only removed the observer from its set — an invocation already chained onto the watcher tail still ran after disposal, and nothing drained started invocations at service dispose; the `settings/updated` manual fan-out caught only synchronous throws, so an async listener's rejection escaped as an unhandled rejection; and `structuredClone` admitted Dates, Maps, BigInts, and cycles that YAML/JSON storage silently distorts on the reload round-trip (a Date lands as a timestamp string, a BigInt as a plain number). YAML writes replaced the whole namespace node, deleting every comment inside the section a comment-preserving provider had promised to keep.
+
+## Decision
+
+**One operation chain, and every write is a read-modify-write.** Watcher refreshes and persists from every namespace queue share a single settled chain, and `persistSection` begins by reconciling the on-disk text into the seam — publishing any unobserved difference first — before rendering against that fresh text. A write can no longer resurrect a stale document, and an on-disk document that turned invalid fails the write loud rather than being overwritten (the reload path keeps its warn-and-keep-last-good policy; the shared `reconcileFromDisk` throws and each caller picks its policy). The watcher's `ready` signal queues one extra reconcile, closing the startup gap between the initial load and the watcher becoming active.
+
+**Writes hold a `wx`-created `<file>.lock` sibling.** The read-render-rename cycle runs under a cross-process writer lock with exponential backoff and a 2 s acquisition deadline. A contender times out without removing the existing lock because age cannot distinguish a crashed owner from a paused live writer; orphan recovery is an operator action. Readers never lock — the rename commit is atomic — so contention is writer-only. The retry and deadline constants are protocol invariants, not deployment config.
+
+**Observer disposal is quiescent.** Watchers carry an `active` flag checked when a queued invocation would start, so a disposer that ran while the invocation waited prevents the start entirely; started invocations register in a service-level `pendingTails` set that the dispose drain awaits beside the write queues. The `settings/updated` fan-out contains a returned thenable's rejection through the same listener diagnostic as a sync throw, and the event contract now states that the `INVARIANT` rethrow serves synchronous listeners only — invariant companions must stay sync, which the shipped companion already is.
+
+**The write boundary admits JSON data only.** The call-time snapshot is a single `cloneJsonShaped` walk that detaches the patch and rejects any non-JSON value — Date, Map, BigInt, non-finite number, function, symbol, class instance, `undefined` array entry, circular reference — with its `$`-rooted path before anything persists. Object entries that are explicitly `undefined` still skip (the sparse-patch contract), now enforced at the boundary instead of inside `mergeLayers`.
+
+**YAML edits are leaf-level diffs.** `renderYaml` diffs the stored section against the next one and applies only `setIn` for changed values and `deleteIn` for removed keys, recursing through maps. Comments, anchors, and formatting survive on every untouched node and on the key node of every changed pair; arrays and other non-map values replace wholesale when unequal (`deepEqualJson` is the shared predicate), taking comments inside them along.
+
+## Alternatives considered
+
+- **`proper-lockfile` instead of a hand-rolled lock** — the dependency-over-hand-rolling policy was weighed: the library is barely maintained, its ownership and retry policy is broader than this one-file protocol needs, and the shipped lock is a small exclusive-create loop with deterministic contention tests. The policy favors dependencies that delete owned code; this one would replace a narrow protocol with an opaque peer.
+- **Revision/CAS instead of a lock** — rename cannot express compare-and-swap, so a CAS needs a version sidecar or content re-hash and a retry loop in every writer; the lock achieves the same serialization with one primitive and keeps readers free.
+- **Merging external edits into the in-flight write's own section** — the seam merges patches over the state visible at call time, so a same-namespace external edit racing a write still resolves last-write-wins; folding it in would need three-way merge semantics no consumer has asked for. The write publishes the external state first, so the loser is at least observed before being superseded.
+- **Declaring async `settings/updated` listeners unsupported** — the typed signature is `void` and lint flags misused promises, but an unlinted JS plugin can still register an async listener; a contract note cannot un-throw an unhandled rejection, so containment is the only defense that holds at runtime.
+- **Keeping `structuredClone` and validating in the provider** — the seam is the durable boundary's owner (every provider stores JSON-shaped documents), and rejecting at call time gives the caller the offending path; a provider-side check would reject after merge, blaming the merged section instead of the caller's value.
+
+## Consequences
+
+`update()` has documented failure modes for the lock deadline and an invalid on-disk document, and rejection messages carry `$`-rooted paths. A crashed holder can leave a lock that requires verified operator removal; automatic age-based takeover would permit overlapping writers. Remaining, documented in the provider README: same-namespace concurrent edits stay last-write-wins (no per-value merge or revision check), a watcher event the OS never delivers leaves the cache stale until the next signal or write, and comments inside replaced arrays or attached inline to changed scalar values go with the value they described. The [user-settings seam note](2026-07-28-user-settings-seam.md)'s deferred-lockfile alternative is superseded by this note. The same defect classes exist in `dsh-credentials-local` (two chains over one `.env`, cached whole-file write-back, post-persist emit) and in the `llm/adapters-updated` fan-out on the stacked branches; those fixes belong to the PRs that introduce the packages and follow this template on merge-up.

.agents/notes/implemented/architecture/2026-07-30-settings-write-path-integrity.zh.md

@@ -0,0 +1,41 @@
+# Agent Note: settings 写路径完整性与观察者生命周期
+
+Status: implemented
+
+[English](2026-07-30-settings-write-path-integrity.md) | 中文
+
+> 范围：对 `packages/settings/` 的第三轮评审——`dsh-settings-local` 的写路径数据完整性（操作链、读-改-写、跨进程写锁、diff 形态的 YAML 编辑）与 `dsh-settings` 的观察者生命周期（watch 的 dispose（资源释放）、异步监听器收容、JSON 形态写入边界）。本 note 推翻了[用户设置 seam note](2026-07-28-user-settings-seam.md)所记录的一项延后决定：跨进程锁文件现已交付。
+
+## 问题
+
+评审发现，提供方的写路径可能销毁它从未观察到的状态，而 seam 的观察者生命周期会泄漏到 dispose 之后。具体而言：watcher 重载与文档写入跑在两条相互独立的 promise 链上，而每次写入都从缓存文本渲染出完整的下一份文档，于是仍处于防抖窗口内的外部编辑会被覆盖——随后的重载又因 rename 后的内容与缓存一致而成为空操作，这次编辑就被无痕抹去。初始 `load()` 与 watcher 自身的建立过程存在竞态，留下一个启动窗口：落在这个窗口内的变更永远不会触发事件。共享同一 harness home 的两个进程各自从独立的缓存渲染，后写者以整个 namespace 为单位胜出。
+
+在 seam 一侧，`watch()` 的释放器只把观察者从集合中移除——已经接到 watcher 链尾的调用在 dispose 之后照常运行，服务 dispose 时也没有任何环节排空已启动的调用；`settings/updated` 的手动扇出只捕获同步抛错，异步监听器的 rejection 会以 unhandled rejection 的形式逃逸；`structuredClone` 则放行 Date、Map、BigInt 与循环引用，而 YAML/JSON 存储会在重载往返中悄悄扭曲这些值（Date 会变成时间戳字符串，BigInt 会变成普通数字）。
+
+YAML 写入则整体替换 namespace 节点，把分节内的每条注释都删掉——而这个保注释的提供方承诺过要保住它们。
+
+## 决策
+
+**单一操作链，且每次写入都是读-改-写。**watcher 的刷新与来自各 namespace 队列的持久化共享同一条结算链；`persistSection` 会先把磁盘上的文本对账进 seam——任何未被观察到的差异都先发布出去——然后才对照这份新鲜文本渲染。写入不再可能复活一份陈旧文档；磁盘上已变非法的文档会让写入响亮失败，而不是被覆盖（重载路径保持其“告警并保留最后可用值”策略；共享的 `reconcileFromDisk` 抛错，各调用方自选策略）。watcher 的 `ready` 信号会额外排入一次对账，弥合初始加载与 watcher 生效之间的启动缺口。
+
+**写入持有以 `wx` 创建的同目录 `<file>.lock`。**读-渲染-rename 循环在一把跨进程写锁下运行，采用指数退避与 2 s 获取期限。竞争者会超时，但不会移除现有锁，因为锁龄无法区分已经崩溃的所有者与被暂停但仍存活的写入方；遗留锁恢复须由操作者执行。读方从不加锁——rename 提交是原子的——因此竞争只发生在写方之间。重试与期限常量是协议不变式，而非部署配置。
+
+**观察者 dispose 达到完全停稳。**watcher 携带一个 `active` 标志，排队的调用即将启动时先检查它，因此在调用等待期间已经运行过的释放器能让这次启动彻底不发生；已启动的调用会登记进服务级的 `pendingTails` 集合，dispose 排空除了等待各写队列，还会等待该集合。`settings/updated` 扇出会把监听器返回的 thenable 的 rejection 收容进与同步抛错相同的监听器诊断；事件契约现已写明 `INVARIANT` 重抛只服务同步监听器——不变式配套插件必须保持同步，而已交付的那个配套插件本就是同步的。
+
+**写入边界只放行 JSON 数据。**调用时刻的快照就是一次 `cloneJsonShaped` 遍历：它把 patch 从调用方分离出来，并在任何内容持久化之前拒绝一切非 JSON 值——Date、Map、BigInt、非有限数值、函数、symbol、类实例、值为 `undefined` 的数组元素、循环引用——拒绝时附带该值以 `$` 为根的路径。显式为 `undefined` 的对象条目仍会跳过（稀疏 patch 契约），这一契约如今在边界处强制执行，而不再放在 `mergeLayers` 内部。
+
+**YAML 编辑是叶子级 diff。**`renderYaml` 对比已存储分节与下一份分节，只对变化的值应用 `setIn`、对移除的键应用 `deleteIn`，并沿 map 递归。注释、锚点与格式在每个未触碰节点上、以及每个被改键值对的键节点上全部保留；数组等非 map 值在不相等时整体替换（`deepEqualJson` 是共享的判定谓词），其内部注释随之一并被带走。
+
+## 曾考虑的替代方案
+
+- **用 `proper-lockfile` 取代手写锁**——按“依赖优先于手写”政策做过权衡：该库几乎无人维护，其所有权与重试策略比这个单文件协议所需的更宽泛，而已交付的锁只是一个小型独占创建循环，带确定性的竞争测试。该政策偏向能删除自有代码的依赖；这个依赖只会把一个窄协议换成不透明的等价物。
+- **用修订号/CAS 取代锁**——rename 表达不了 compare-and-swap，因此 CAS 需要一个版本伴随文件或内容重哈希，外加每个写方里的一个重试循环；锁用一个原语实现同样的串行化，还让读方完全免锁。
+- **把外部编辑合并进正在进行的写入自身的分节**——seam 是在调用时刻可见的状态之上合并 patch 的，因此与写入竞态的同 namespace 外部编辑仍按后写胜出解决；要把外部编辑并进来，需要三方合并语义，而没有任何消费方提出过这种需求。写入会先发布外部状态，落败一方至少在被取代之前被观察到。
+- **宣布不支持异步 `settings/updated` 监听器**——类型签名是 `void`，lint 也会标记误用的 promise，但未经 lint 的 JS 插件仍能注册异步监听器；契约里的一句说明无法收回已经抛出的 unhandled rejection，收容是唯一在运行时守得住的防线。
+- **保留 `structuredClone`、在提供方里做校验**——seam 才是持久化边界的所有者（每个提供方存储的都是 JSON 形态文档），而且在调用时刻拒绝能把违规值的路径给到调用方；提供方侧的检查要到合并之后才拒绝，归咎的是合并后的分节，而不是调用方传入的值。
+
+## 后果
+
+`update()` 对锁获取期限与磁盘文档非法都有成文的失败模式，rejection 消息携带以 `$` 为根的路径。持有者崩溃后可能留下锁，需要操作者核实后移除；若按锁龄自动接管，则会允许多个写入方重叠。仍然存在、且已记录在提供方 README 中的有：同 namespace 并发编辑仍是后写胜出（没有逐值合并，也没有修订号检查）；OS 从未投递的 watcher 事件会让缓存保持陈旧，直到下一个信号或下一次写入；被替换数组内部的注释、以及行内附着在被改标量值上的注释，会随其描述的值一起消失。
+
+[用户设置 seam note](2026-07-28-user-settings-seam.md)里“延后锁文件”那条替代方案已被本 note 取代。同类缺陷还存在于 `dsh-credentials-local`（两条链共用一个 `.env`、按缓存整文件写回、持久化之后才发事件）与堆叠分支上的 `llm/adapters-updated` 扇出；这些修复归引入相应包（package）的那些 PR（Pull Request）所有，向上合并时按本模板处理。

.agents/notes/implemented/architecture/2026-07-30-web-config-plane.i18n.yaml

@@ -0,0 +1,6 @@
+# Bilingual-pair consistency record (docs/i18n/README.md): the git blob hash of each
+# side as of the last confirmed-consistent state. Both languages carry equal authority;
+# after editing either side, bring the other along and re-record with:
+#   pnpm run verify-translation-pairing --write .agents/notes/implemented/architecture/2026-07-30-web-config-plane.md
+2026-07-30-web-config-plane.md: 6d1a8c242c1888ee4fca9e21ebc814f7a345d633
+2026-07-30-web-config-plane.zh.md: c3255cacfdd1f06d12f7bb2631f95273536b7ef9

.agents/notes/implemented/architecture/2026-07-30-web-config-plane.md

@@ -0,0 +1,36 @@
+# Agent Note: the web configuration plane
+
+Status: implemented
+
+English | [中文](2026-07-30-web-config-plane.zh.md)
+
+> Scope: the wire face and web UI deferred from the [request-level LLM configuration note](2026-07-29-request-level-llm-config-credentials.md) — the `settings.*`/`credentials.*`/`llm.*` RPC domains with pushed invalidations, layered+redacted `describe()`, the llm configurable-provider directory and topology event, the standalone `dsh-client-schema-form` model layer, and the Models settings page with its hand-written provider editor. The `deepseek` → `deepseek-official` provider-route rename rides along as the enabling breaking change.
+
+## Problem
+
+PR1 made LLM adapter configuration restart-free at the seam, but the only writer was a text editor on `settings.yaml`: the web client had no wire access to settings, credentials, or provider topology, so "store a key, prompt again" still meant leaving the product. Three gaps blocked a config page rather than one: `describe()` returned only the merged effective value (a form cannot tell a user override from a composition default, and serializing it would have shipped `role('secret')` values to every browser), nothing enumerated the providers an adapter *could* run (a bare-mounted `llm-pi-ai` was invisible until configured), and the two adapters both wanted a `deepseek` route key, so the directory could not attribute routes to owning namespaces unambiguously. Hand-maintaining a form per provider was rejected outright — the schemas already exist as schemastery `Config` values, and a second source of field truth drifts.
+
+## Decision
+
+**Wire domains on the compiled RPC map, rejections as codes, invalidations as frames.** `settings.describe/update/replace`, `credentials.describe/set/unset`, `llm.providers`, and `llm.models` (claiming the reserved `host.listModels` surface) join `RpcMethodMap`, so the seven compiler-locked wiring sites keep contract, schema, handler, and client in lockstep. Seam rejections fold into `settings-rejected {ns}` / `credential-rejected {ref}` business errors (HTTP stays a carrier), and three `HostFrame`s — `host/settings-changed {ns}`, `host/credentials-changed {ref}`, `host/models-changed` — follow the `host/commands-changed` shape so every client converges without polling. Writes join `pickDirectory`/`openPath` in the connection guard's privileged set: loopback + same-origin or 403, because a LAN-exposed dsh web must not accept config mutation from another origin.
+
+**`describe()` grows layers and structural secret redaction.** `SettingsDescriptor` carries `base`/`user` beside the effective value, so the form marks "overridden" by presence in the user layer, not value inequality (an override *equal* to the base is still an override). `describe({ redactSecrets: true })` — mandatory at every wire face — strips `role('secret')` subtrees from all three layers via a pure structural walk of the schema (object/dict/array containers; a secret-role subtree is one opaque leaf) and enumerates the stripped slots as `{path, set}`, so a page can render write-only inputs without ever receiving a value.
+
+**The llm seam declares configurability and announces topology.** `registerConfigurableProviders()` is an all-or-nothing, fiber-scoped directory of `{provider, displayName, settingsNs, settingsPath}` — the addressing a config page needs to open the right settings subtree for a route that may not exist yet; `listConfigurableProviders()` merges with live routes in the wire handler so undeclared live routes still report active. The zero-payload `'llm/adapters-updated'` event fires from all four registration/unregistration commit points with contained listener dispatch (INVARIANT rethrow), following the settings/commands precedent. `llm-deepseek`'s route renamed to `deepseek-official` because the pi-ai catalog legitimately owns `deepseek` as an aggregator entry; pre-release stance, no alias.
+
+**A hand-written editor over a schema model layer.** `dsh-client-schema-form` rehydrates the wire's `toJSON()` envelope into live schemastery nodes for validation, path resolution, and immutable draft editing — but no generic rendering: the first cut shipped a full schema-driven form renderer, and the resulting page was an unstyled schema dump (every advanced field flattened onto the card, raw field names as labels, the `retryPolicy` unsupported-fallback in the main flow). The user chose the hand-written direction over adding a hint/grouping system, and a second round removed the reference input entirely: the card's primary field is one **API key** input, a whole-section provider without a configured key opens as its setup card, and the collapsed 自定义设置 fold carries the curated per-family extras (`baseURL` for both families, plus `reasoningEffort` for deepseek / `reasoning` for pi-ai), with every other field owned by `settings.yaml`. Validation still runs the rehydrated schema before writing, so a hand-coded field that drifts from its schema fails loud on save rather than silently.
+
+**The Models page is a three-domain join with seam-shaped apply semantics.** Rows are configured providers; the add card's select is the dormant directory remainder. Route liveness still gates readiness and invalidates the join, but the page does not render it as provider status because configuration presence and runtime availability are distinct. The key path stays reference-shaped without ever showing a reference: a typed key stores **write-only** through `credentials.set` under the profile's `apiKeyEnv`, deriving `<ROUTE>_API_KEY` when none exists (the pi-ai profile records the derivation), so `settings.yaml` never carries a key value. Profile edits and removals land as minimal path-addressed `settings.mutate` operations against the redacted user section, which never names a secret the page did not receive. Removing a user-layer provider first opens a localized model-provider confirmation dialog; cancellation, its close button, and its mask leave the profile untouched, while the destructive confirmation submits the single unset and blocks duplicate submission until it settles.
+
+## Alternatives considered
+
+- **Serving JSON Schema over the wire** — schemastery's `toJSON()` envelope round-trips `role()`/meta and rehydrates into the validator the client already ships for drafts; converting to JSON Schema loses exactly the role annotations the credential control and secret redaction key on.
+- **A generic schema-driven form renderer** — implemented first, then replaced: field truth without visual hierarchy produced an ugly, unusable card, and making it good meant building a hint vocabulary (primary/advanced grouping, per-field descriptions, array item cards) rivaling the hand-written editor in cost while still fitting no mockup exactly. Two schemas exist today (the deepseek `Config` and the shared pi-ai profile), so hand-writing is two thin namespace-keyed layouts; the drift risk is bounded by save-time schema validation and by unknown fields staying untouched in the document.
+- **Masking secrets per-field with sentinel backfill on `replace`** — the PR1 decision (secrets are references) already deleted the stored-literal case for the product default; structural redaction plus a write-only credential path handles the residue without teaching every writer a sentinel protocol.
+- **Storing the typed key as a literal `apiKey` setting** — the v1 "one API key input" requirement could have written the literal into the profile, but every UI removal path rebuilds the user section from the *redacted* layers, so any reset or row deletion would silently drop stored sibling keys; deriving a reference keeps the input single-field while keeping `settings.yaml` secret-free and every replace safe.
+- **A `models` bridge plugin owning provider configuration** — same rejection as PR1: per-plugin namespaces plus a four-field directory declaration give the UI everything it needs; the bridge's unified dict re-imports the adapter-mapping indirection.
+- **Page-side polling instead of pushed frames** — the mux already carries `host/commands-changed`; three more frames cost one shape each and make a second tab, an external `settings.yaml` edit, and a settings-born route converge at event speed.
+
+## Consequences
+
+The whole loop is pinned keyless in the browser lane (`apps/web/tests/models-settings.e2e.ts`): the add card offers the dormant pi-ai catalog, adding `minimax-cn` with a typed key writes the reference-only profile into `settings.yaml`, stores the value into the harness home's `.env` under the derived `MINIMAX_CN_API_KEY`, registers the route live on the topology frame, and the customized fold merges `reasoning` beside the reference — zero model calls, ARIA goldens for the add-card, configured, and delete-confirmation states, plus a scaffold `harnessHome` so tests never touch a real `~/.dsh` (the provider under test is one whose derived reference cannot collide with a developer's exported keys). The removal scenario proves cancellation leaves the profile intact, confirmation removes it, and the intentionally retained credential survives. The rename touched 239 files (fixtures, goldens, docs, python) in one commit with no compatibility alias. The renderer replacement cost one commit and no wire change: apply semantics, redaction, and the directory join were renderer-agnostic all along. Deferred: a per-row models preview (the picker already lists models), a page address for live routes that never declared configurability, and explicit removal of a provider's retained credential.

.agents/notes/implemented/architecture/2026-07-30-web-config-plane.zh.md

@@ -0,0 +1,36 @@
+# Agent Note：web 配置平面
+
+Status: implemented
+
+[English](2026-07-30-web-config-plane.md) | 中文
+
+> 范围：[请求级 LLM 配置 note](2026-07-29-request-level-llm-config-credentials.md) 中延后的 wire 面与 web UI——带推送式失效的 `settings.*`/`credentials.*`/`llm.*` RPC 领域、分层且脱敏的 `describe()`、llm 可配置提供方目录与拓扑事件、独立的 `dsh-client-schema-form` 模型层，以及带手写提供方编辑器的 Models 设置页。`deepseek` → `deepseek-official` 提供方路由重命名作为解锁前提的破坏性变更一并搭车合入。
+
+## 问题
+
+PR1 让 LLM（大语言模型）适配器配置在 seam 层面免重启，但唯一的写入方还是直接编辑 `settings.yaml` 的文本编辑器：web 客户端没有触达设置、凭据或提供方拓扑的任何 wire 通道，「存入密钥、再次发起提示」于是仍意味着离开产品本身。挡住配置页的缺口不是一个，而是三个：`describe()` 只返回合并后的生效值（表单分不清用户覆盖与组合默认值，而且照原样序列化会把 `role('secret')` 的值发到每一个浏览器）；没有任何东西枚举适配器*可以*运行的提供方（裸挂载的 `llm-pi-ai` 在配置之前完全不可见）；两个适配器又都想要 `deepseek` 这个路由键，目录因此无法无歧义地把路由归到拥有它的 namespace 名下。为每个提供方手工维护一份表单被直接否决——schema 已经以 schemastery `Config` 值的形式存在，第二份字段真源注定漂移。
+
+## 决策
+
+**wire 领域挂上编译期 RPC 映射，拒绝落为错误码，失效落为帧。**`settings.describe/update/replace`、`credentials.describe/set/unset`、`llm.providers` 与 `llm.models`（认领预留的 `host.listModels` 面）一同加入 `RpcMethodMap`，七处由编译器锁定的接线位点因此让契约、schema、处理器与客户端保持步调一致。seam 侧的拒绝折叠为 `settings-rejected {ns}`/`credential-rejected {ref}` 业务错误（HTTP 仍只是载体），三个 `HostFrame`——`host/settings-changed {ns}`、`host/credentials-changed {ref}`、`host/models-changed`——沿用 `host/commands-changed` 的形状，因此每个客户端都无需轮询即可收敛。写入与 `pickDirectory`/`openPath` 一起进入连接守卫的特权集合：回环 + 同源，否则 403，因为暴露在局域网上的 dsh web 绝不能接受来自其他源的配置修改。
+
+**`describe()` 增加分层与结构化 secret 脱敏。**`SettingsDescriptor` 在生效值之外携带 `base`/`user`，表单据此按「字段是否出现在用户层」来标记「已覆盖」，而非按值是否不等（与 base *相等*的覆盖仍然是覆盖）。`describe({ redactSecrets: true })`——在每个 wire 面都强制启用——经由对 schema 的纯结构遍历（object/dict/array 容器；secret 角色子树整体是一个不透明叶节点）从全部三层剥除 `role('secret')` 子树，并把剥除的槽位枚举为 `{path, set}`，页面因此不必收到任何值就能渲染只写输入框。
+
+**llm seam 声明可配置性并公布拓扑。**`registerConfigurableProviders()` 是一个全有或全无、以 fiber 为作用域的目录，条目为 `{provider, displayName, settingsNs, settingsPath}`——这正是配置页要为一条可能尚不存在的路由打开正确设置子树时所需要的寻址；`listConfigurableProviders()` 在 wire 处理器里与存活路由合并，未声明的存活路由因此仍报告为激活。零负载的 `'llm/adapters-updated'` 事件从全部四个注册／注销提交点触发，listener 派发带异常隔离（INVARIANT 重抛），沿用 settings/commands 的先例。`llm-deepseek` 的路由重命名为 `deepseek-official`，因为 pi-ai catalog 名正言顺地拥有 `deepseek` 这个聚合器条目；依预发布立场，不设别名。
+
+**架在 schema 模型层之上的手写编辑器。**`dsh-client-schema-form` 把 wire 的 `toJSON()` 信封还原（rehydrate）为活的 schemastery 节点，用于校验、路径解析与不可变草稿编辑——但不做通用渲染：第一版交付了完整的 schema 驱动表单渲染器，得到的却是一个未加样式、把 schema 原样倾倒出来的页面（每个进阶字段都平铺到卡片上、原始字段名直接充当标签、`retryPolicy` 的「不支持」回退落在主流程里）。用户没有再加一套提示／分组系统，而是选择了手写方向，第二轮又把引用输入框整个移除：卡片的主字段是一个 **API 密钥**输入框，未配置密钥的整分节提供方会以其设置卡片的形式打开，收起的「自定义设置」折叠区承载按家族精选的额外字段（两个家族都有 `baseURL`，另加 deepseek 的 `reasoningEffort`／pi-ai 的 `reasoning`），其余每个字段都归 `settings.yaml` 所有。校验仍会在写入前运行还原出的 schema，因此偏离其 schema 的手写字段会在保存时大声失败，而非静默失败。
+
+**Models 页是一次三领域联接，应用语义与 seam 同形。**每一行是一个已配置的提供方；「新增」卡片的选择框是可配置提供方目录中剩余的休眠条目。路由存活状态仍用于就绪判定，并会使该联接失效，但页面不将其渲染为提供方状态，因为配置存在与运行时可用性是两个不同概念。密钥通道保持引用形态，却从不展示任何引用：键入的密钥经 `credentials.set` **只写**存入 profile 的 `apiKeyEnv` 之下，引用不存在时便派生 `<ROUTE>_API_KEY`（pi-ai profile 会记录该派生），因此 `settings.yaml` 从不携带密钥值。profile 的编辑和删除会针对脱敏后的用户分节，以按路径寻址的最小 `settings.mutate` 操作落地，绝不会点名页面未收到的机密。删除用户层提供方时，会先打开本地化的模型提供方确认对话框；取消操作、关闭按钮和遮罩均不会改动 profile，而破坏性确认会提交唯一一条 unset，并在其完成前阻止重复提交。
+
+## 曾考虑的替代方案
+
+- **在 wire 上改发 JSON Schema**——schemastery 的 `toJSON()` 信封能往返保留 `role()`/meta，并还原成客户端为草稿校验本就自带的那个校验器；转换成 JSON Schema 丢掉的恰恰是凭据控件与 secret 脱敏所依赖的角色注解。
+- **通用的 schema 驱动表单渲染器**——先实现、后被替换：如实呈现字段却缺失视觉层级，产出的卡片丑陋且不可用；要把它做好，就意味着构建一套提示词汇（主要／进阶分组、逐字段描述、数组项卡片），成本堪比手写编辑器，却仍无法与任何设计稿完全吻合。今天存在两份 schema（deepseek 的 `Config` 与共享的 pi-ai profile），手写因此就是两套以 namespace 为键的薄布局；漂移风险由保存时的 schema 校验以及未知字段在文档中的原样保留共同约束。
+- **逐字段脱敏机密并在 `replace` 时回填哨兵值**——PR1 的决策（机密是引用）已经为产品默认形态删掉了「存储字面量」这种情况；结构化脱敏加上只写的凭据通道足以处理残余情形，无需让每个写入方都学会一套哨兵协议。
+- **把键入的密钥存成字面 `apiKey` 设置**——v1「单个 API 密钥输入框」的需求本可以把字面量直接写进 profile，但 UI 的每条删除路径都会从*脱敏后的*各层重建用户分节，任何重置或整行删除都会静默丢掉已存储的兄弟密钥；派生引用让输入保持单字段，同时让 `settings.yaml` 不含机密、每一次 replace 都安全。
+- **由 `models` 桥接插件持有提供方配置**——与 PR1 相同的否决理由：按插件划分的 namespace 加上四字段的目录声明已经给了 UI 需要的一切；桥接层的统一字典会把适配器映射那层间接重新引进来。
+- **页面侧轮询而非推送帧**——mux 已经承载 `host/commands-changed`；再加三个帧各自只多一个形状的成本，就让第二个标签页、外部的 `settings.yaml` 编辑和由设置催生的路由都以事件速度收敛。
+
+## 后果
+
+整条闭环以无密钥方式固定在浏览器测试通道（`apps/web/tests/models-settings.e2e.ts`）：「新增」卡片提供休眠的 pi-ai catalog，携键入的密钥添加 `minimax-cn` 会把只含引用的 profile 写入 `settings.yaml`、把密钥值存入 harness 家目录 `.env` 中派生的 `MINIMAX_CN_API_KEY` 之下、路由随拓扑帧注册为存活，「自定义设置」折叠区则把 `reasoning` 合并到引用旁边——全程零模型调用，「新增」卡片态、已配置态与删除确认态各有 ARIA golden，另有脚手架式的 `harnessHome`，测试绝不触碰真实的 `~/.dsh`（受测提供方是派生引用不可能与开发者已导出密钥相撞的那一个）。删除场景证明：取消后 profile 保持原样，确认后会将其删除，而刻意保留的凭据依然存在。这次重命名在一次提交中触及 239 个文件（fixture（测试前置数据）、golden、文档、python），未保留兼容别名。替换渲染器只花了一次提交，且没有任何 wire 变更：应用语义、脱敏与目录联接从一开始就与渲染器无关。延后事项：每行的模型预览（选择器已能列出模型）、为从未声明可配置性的存活路由提供页面地址，以及显式删除提供方所保留的凭据。

.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。

.agents/notes/implemented/feature/2026-06-30-session-store-fork-api.i18n.yaml

@@ -2,5 +2,5 @@
 # side as of the last confirmed-consistent state. Both languages carry equal authority;
 # after editing either side, bring the other along and re-record with:
 #   pnpm run verify-translation-pairing --write .agents/notes/implemented/feature/2026-06-30-session-store-fork-api.md
-2026-06-30-session-store-fork-api.md: 5342deba8ca879026d32ee1420cb3c0fdf67c500
-2026-06-30-session-store-fork-api.zh.md: 51a3e0ce50aff10a9812c91d24dc6e78a56c43ba
+2026-06-30-session-store-fork-api.md: 69ff85e1f137f4f263bf951af0a3f655411c606a
+2026-06-30-session-store-fork-api.zh.md: 3304a6f384c9004b3572c95881f832a4aa21b77c

.agents/notes/implemented/feature/2026-06-30-session-store-fork-api.md

@@ -28,6 +28,12 @@ class SessionStore extends Service {
 
 An empty prefix is forkable; any non-empty boundary must be a safe existing sequence outside an open turn. Typed errors distinguish missing sources, stale objects, duplicate child ids, invalid boundaries, and prefixes ending during execution. Broader log validation and crash repair remain with their existing owners.
 
+### Host and browser adaptation
+
+The Host `session.fork` RPC accepts `atSeq` as an anchor within the desired turn rather than as the store's inclusive safe boundary. It selects the first `turn/end` at or after that anchor; an omitted or past-end anchor selects the last completed turn. An anchor already in the log but not followed by a matching `turn/end` returns `fork-unavailable` and never falls back to an earlier turn, so a message action cannot silently omit the clicked message.
+
+The Host creates the child through the agent registry with the selected seed and lineage, and pre-publication setup installs the latest logged provider, model, and reasoning target before the child can run. It then attaches the child to the source Workspace. An attachment failure returns `workspace-attach-failed` with the already-published child id; the client reconciles that child into its summary list before surfacing the error. The Session-row action uses the last completed turn, while a message action supplies its event seq; both open the child after success, and lineage expansion makes it visible beneath the source.
+
 ## Alternatives considered
 
 **Separate `ctx.sessionFork` service.** This was the first implementation, but review showed it overfit the capability-seam pattern. The code had no swappable backend, no extra event surface, no independent ownership lifecycle, and no durable behavior beyond `ctx.sessions.create({ seed, meta })`. Keeping a separate package would make callers discover and install a second service just to perform policy around a session-store primitive.
@@ -40,4 +46,4 @@ An empty prefix is forkable; any non-empty boundary must be a safe existing sequ
 
 The public surface stays small and discoverable: live session branching is part of `ctx.sessions`, next to `create({ seed })`, rather than a standalone service or a two-step helper pair. Persistence continues to work through existing `session/created` and `session/flush` behavior: a forked child starts life with seeded events, so existing backends persist that seed once and preserve `parentSession` / `seedLength` in the header.
 
-The v1 scope still excludes ACP `session/fork`, unloaded persisted-session forking, model-facing tools, and subagent refactors. If a future ACP method is added, it should advertise the capability only after it has protocol and snapshot coverage; this Agent Note adds no ACP wire behavior, so no ACP snapshot is required. Fork-child replay remains covered by the existing [seed-boundary testing Agent Note](../testing/2026-06-22-fork-child-replay-seed-boundary.md), while this API gets focused `dsh-session` unit tests plus JSONL persistence coverage.
+The v1 scope still excludes ACP `session/fork`, unloaded persisted-session forking, model-facing tools, and subagent refactors. If a future ACP method is added, it should advertise the capability only after it has protocol and snapshot coverage; this Agent Note adds no ACP wire behavior, so no ACP snapshot is required. Fork-child replay remains covered by the existing [seed-boundary testing Agent Note](../testing/2026-06-22-fork-child-replay-seed-boundary.md); focused store, Host, carrier, and client tests pin the boundary and reconciliation contracts, while the real Chromium scenario pins the assembled message action and lineage tree.

.agents/notes/implemented/feature/2026-06-30-session-store-fork-api.zh.md

@@ -28,6 +28,12 @@ class SessionStore extends Service {
 
 空前缀可以被 fork；任何非空边界都必须是位于开放轮次之外且安全、已存在的序号。类型化的错误区分源缺失、对象陈旧、子 id 重复、边界无效和前缀结束于执行过程中等情况。更广泛的日志校验与崩溃恢复仍由其现有的负责方处理。
 
+### Host 与浏览器适配
+
+Host 的 `session.fork` RPC 接受 `atSeq`，并将其视为所需轮次内的锚点，而非 store 中包含该序号的安全边界。它选择该锚点处或其后的首个 `turn/end`；锚点省略或超过末尾时，选择最后一个已完成轮次。若锚点已在日志中，但从该锚点起找不到匹配的 `turn/end`，则返回 `fork-unavailable`，绝不回退到更早的轮次，因此消息操作不会静默遗漏所点击的消息。
+
+Host 通过 agent（智能体）注册表，以选定的种子和谱系创建子会话；发布前 setup 会先安装日志中最新的提供方、模型和推理（reasoning）目标，子会话才能运行。随后，Host 将子会话附加到源 Workspace。若附加失败，则返回 `workspace-attach-failed` 及已发布的子会话 id；客户端先将该子会话对账到摘要列表，再向调用方报告错误。Session 行操作使用最后一个已完成轮次，消息操作则提供其事件 seq；两者都会在成功后打开子会话，展开谱系后可在源会话下看到它。
+
 ## 曾考虑的替代方案
 
 **独立的 `ctx.sessionFork` 服务。** 这是最初的实现，但评审表明它过度套用了 capability-seam 模式。代码没有可替换的后端、没有额外的事件面、没有独立的所有权生命周期，也没有超出 `ctx.sessions.create({ seed, meta })` 的持久化行为。保留独立包会迫使调用方为了在会话存储原语之上执行一层策略而去发现并安装第二个服务。
@@ -40,4 +46,4 @@ class SessionStore extends Service {
 
 公开接口保持精简且易于发现：活跃会话分支是 `ctx.sessions` 的一部分，紧邻 `create({ seed })`，而非一个独立服务或一对两步辅助函数。持久化继续通过现有的 `session/created` 和 `session/flush` 行为运作：fork 出的子会话以种子事件开始生命，因此现有后端只需持久化该种子一次，并在 header 中保存 `parentSession`／`seedLength`。
 
-v1 范围仍然排除 ACP（Agent Client Protocol） `session/fork`、对未加载的已持久化会话的 fork、面向模型的工具，以及 subagent 重构。如果未来添加 ACP 方法，应在具备协议与快照覆盖后才广播该能力；本 Agent Note 不添加任何 ACP 协议行为，因此不需要 ACP 快照。fork 子会话的回放仍由现有的[种子边界测试 Agent Note](../testing/2026-06-22-fork-child-replay-seed-boundary.md) 覆盖，而本 API 则获得专门的 `dsh-session` 单元测试加 JSONL 持久化覆盖。
+v1 范围仍然排除 ACP（Agent Client Protocol） `session/fork`、对未加载的已持久化会话的 fork、面向模型的工具，以及 subagent 重构。如果未来添加 ACP 方法，应在具备协议与快照覆盖后才广播该能力；本 Agent Note 不添加任何 ACP 协议行为，因此不需要 ACP 快照。fork 子会话的回放仍由现有的[种子边界测试 Agent Note](../testing/2026-06-22-fork-child-replay-seed-boundary.md) 覆盖；store、Host、载体与客户端的专项测试固定边界和对账契约，真实 Chromium 场景则固定组装后的消息操作与谱系树。

.agents/notes/implemented/feature/2026-07-17-dedicated-full-screen-tui-front-door.i18n.yaml

@@ -2,5 +2,5 @@
 # side as of the last confirmed-consistent state. Both languages carry equal authority;
 # after editing either side, bring the other along and re-record with:
 #   pnpm run verify-translation-pairing --write .agents/notes/implemented/feature/2026-07-17-dedicated-full-screen-tui-front-door.md
-2026-07-17-dedicated-full-screen-tui-front-door.md: 10564b110cc2e56615bde86830c0d39d5a7cee38
-2026-07-17-dedicated-full-screen-tui-front-door.zh.md: 4624a4db3db598793eee257f829a080f4d7ad711
+2026-07-17-dedicated-full-screen-tui-front-door.md: c011a0284ea0efe59693785c038f814a866068ac
+2026-07-17-dedicated-full-screen-tui-front-door.zh.md: 5aea4ac0c5a3b29c098c297e514fab49caf643ff

.agents/notes/implemented/feature/2026-07-17-dedicated-full-screen-tui-front-door.md

@@ -20,7 +20,7 @@ The selected front door receives the exact generated or resumed `SessionId` used
 
 ### Session projection and interaction
 
-The TUI rebuilds the transcript from the active `session.surface` and reprojects it whenever an event carries a `surfaceOp`, so resumed and compacted history matches the model-visible conversation. It renders Markdown text and reasoning, including fenced code with hidden Markdown markers, a dim optional language label, and a code-colored body, token totals, the latest `todo/write` plan, and tool cards produced through each tool definition's `presentCall` and `presentResult` methods. Long card bodies retain a configurable head/tail preview with the hidden-line count; one terminal control expands or collapses every card. Pending chunks and tool calls update the same components that completed events settle.
+The TUI rebuilds the transcript from the append-origin session events, so resumed history keeps every message the reader already saw; a compacted range stays readable behind one marker instead of matching the model-visible conversation ([append-origin transcript](../bug-fix/2026-07-29-human-transcript-append-origin.md)). It renders Markdown text and reasoning, including fenced code with hidden Markdown markers, a dim optional language label, and a code-colored body, token totals, the latest `todo/write` plan, and tool cards produced through each tool definition's `presentCall` and `presentResult` methods. Long card bodies retain a configurable head/tail preview with the hidden-line count; one terminal control expands or collapses every card. Pending chunks and tool calls update the same components that completed events settle.
 
 Editor input calls `agent.send()` while idle and `agent.steer()` while a turn is running. Cancellation, reasoning visibility, tool-card expansion, redraw, transcript clearing, and exit are terminal-only controls. `/exit` and `/quit` share the same exit path: they cancel an active turn, wait for idle, and then restore and close the terminal. The idle footer derives context occupancy from `tokenMeter` and shows the selected model and explicit reasoning effort; during a run, elapsed activity and the Escape interrupt hint replace that summary. `/status` remains available in either state and appends a detailed terminal-only snapshot: session identity and timestamps, selected model, reasoning effort/default state and reasoning visibility, lifecycle counts folded from the event log, the same deduplicated usage buckets and KV-cache rate as the footer, and context use from `tokenMeter` plus the selected model's advertised capacity. The plugin registers the shared `userInteraction` provider and presents queued questions in a wide bottom-left keyboard panel with batch progress, numbered options, and aligned descriptions; the panel's controls hint lists only actions meaningful for the current option count, omitting navigation when exactly one option is shown; agent behavior and answer logging remain owned by their existing services.
 
@@ -47,6 +47,6 @@ The implemented [TUI terminal-state snapshot Agent Note](../testing/2026-07-18-t
 
 - Interactive terminal work has a stateful Markdown, card, plan, and question interface with no second terminal protocol to keep aligned.
 - The TUI carries a pi-tui dependency and a strict TTY requirement; non-TTY deployments use the Headless app or a structured protocol.
-- Session projection makes resume and compaction consistent with the durable conversation, but one configured session owns the transcript and editor.
+- Session projection makes resume consistent with the durable conversation, but one configured session owns the transcript and editor.
 - Tool packages extend terminal cards through their existing presentation methods without adding tool-specific branches to the TUI.
 - Model and reasoning-effort selection use adapter-advertised metadata without turning catalog membership into request validation; unused selections are not durable state.

.agents/notes/implemented/feature/2026-07-17-dedicated-full-screen-tui-front-door.zh.md

@@ -20,7 +20,7 @@ DeepSeek Harness 将 [`@deepseek-ai/dsh-tui`](../../../../packages/ui/tui/README
 
 ### 会话投影与交互
 
-TUI 从活跃的 `session.surface` 重建 transcript（文本记录），并在事件携带 `surfaceOp` 时重新投影，因此恢复或压缩后的历史与模型可见会话保持一致。TUI 渲染 Markdown 文本与推理（其中围栏代码块隐藏 Markdown 标记、保留一个暗色的可选语言标签，并使用代码配色的正文）、token 用量、最新 `todo/write` 计划，以及各工具定义通过 `presentCall` 和 `presentResult` 方法生成的工具卡片。较长的工具卡片正文会保留可配置的头尾预览，并显示隐藏行数；一个终端控制可以展开或收起全部卡片。进行中的分片与工具调用会更新同一组组件，随后由完成事件收束状态。
+TUI 从追加来源的会话事件重建 transcript（文本记录），因此恢复后的历史会保留读者已经看到的每条消息；被压缩的范围不再与模型可见会话保持一致，而是留在一行标记之后仍可阅读（[追加来源的 transcript](../bug-fix/2026-07-29-human-transcript-append-origin.md)）。TUI 渲染 Markdown 文本与推理（其中围栏代码块隐藏 Markdown 标记、保留一个暗色的可选语言标签，并使用代码配色的正文）、token 用量、最新 `todo/write` 计划，以及各工具定义通过 `presentCall` 和 `presentResult` 方法生成的工具卡片。较长的工具卡片正文会保留可配置的头尾预览，并显示隐藏行数；一个终端控制可以展开或收起全部卡片。进行中的分片与工具调用会更新同一组组件，随后由完成事件收束状态。
 
 agent 空闲时，编辑器输入调用 `agent.send()`；轮次运行中则调用 `agent.steer()`。取消、推理显隐、工具卡片展开、重绘、清空 transcript 和退出都只是终端控制。`/exit` 和 `/quit` 共用同一条退出路径：先取消进行中的轮次，等待 agent 空闲，然后恢复并关闭终端。空闲态页脚根据 `tokenMeter` 得出上下文占用率，并显示所选模型和显式选定的推理强度；agent 运行期间，该摘要会替换为带已用时长的活动指示和 Escape 中断提示。`/status` 在这两种状态下均可用，并会追加一份仅在终端显示的详细快照，其中包括会话标识与时间戳、所选模型、推理强度（或默认状态）及推理显隐状态、从事件日志归并得出的生命周期计数、与页脚一致的去重用量分项和 KV 缓存命中率，以及 `tokenMeter` 给出的上下文用量和所选模型公布的容量。插件注册共享的 `userInteraction` 提供方，在左下角宽幅键盘操作面板中呈现排队的问题，面板显示批次进度、带编号的选项和对齐的描述；面板的操作提示只列出在当前选项数量下有意义的操作，仅有一个选项时不显示导航项；agent 行为和答案日志仍由既有服务负责。
 
@@ -47,6 +47,6 @@ agent 空闲时，编辑器输入调用 `agent.send()`；轮次运行中则调
 
 - 交互式终端拥有带状态的 Markdown、卡片、计划和提问界面，无需再对齐第二套终端协议。
 - TUI 会引入 pi-tui 依赖并严格要求 TTY；非 TTY 部署使用 Headless app 或结构化协议。
-- 会话投影使恢复和压缩与持久会话保持一致，但只有一个已配置会话拥有 transcript 和编辑器。
+- 会话投影使恢复与持久会话保持一致，但只有一个已配置会话拥有 transcript 和编辑器。
 - 工具包通过既有呈现方法扩展终端卡片，无需在 TUI 中增加工具专用分支。
 - 模型和推理强度选择使用适配器公布的元数据，但不会把目录成员关系变成请求校验；未使用的选择不属于持久化状态。

.agents/notes/implemented/feature/2026-07-25-session-list-browsing-and-manual-order.i18n.yaml

@@ -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
-2026-07-25-session-list-browsing-and-manual-order.md: 586995bf459aeaee88672863977f7acf2a7061a3
-2026-07-25-session-list-browsing-and-manual-order.zh.md: 432d5167a57d30bc04a0b4faf213e4341f07bd2f
+#   pnpm run verify-translation-pairing --write .agents/notes/implemented/feature/2026-07-25-session-list-browsing-and-manual-order.md
+2026-07-25-session-list-browsing-and-manual-order.md: 831aa53e532a75392690c330837482bb0f9c32b1
+2026-07-25-session-list-browsing-and-manual-order.zh.md: 9ad074d59c13585aa4fca46ae4d40e2deb15cde6

.agents/notes/implemented/feature/2026-07-25-session-list-browsing-and-manual-order.md

@@ -12,14 +12,14 @@ Two existing mechanisms stood in the way. First, the host durably promoted the a
 
 ## Decision
 
-### Flat view and viewing state
+### Flat rows and viewing state
 
-The group-by menu offers two modes, WorkSpace / In one list. Flat mode renders every session (fork children included) as a top-level row, strictly newest-first by `updatedAt`, with no parent/child adjacency; the Intent placeholder renders as the first row. The mode choice persists in the browser (`dsh.workspace.view`) across reloads.
+The group-by menu offers two modes, WorkSpace / In one list. WorkSpace mode renders peer session rows within each group in the manual order from `WorkspaceView.sessionIds`; In one list combines every session and sorts them strictly newest-first by `updatedAt`. Neither mode projects `parentId` into a list hierarchy; fork lineage remains session data only. [Web session fork actions](2026-07-27-web-session-fork-actions.md) define the complete fork behavior. The mode choice persists in the browser (`dsh.workspace.view`) across reloads.
 
 ### Row interactions
 
 - Session rows show a detail card after a 500ms hover dwell (full title / relative time / status line; the status line has only running/idle until the wire grows a status field). The card and the row menu are mutually exclusive: no card while a menu is open or a drag is in flight.
-- Session-row … menu: Rename / Fork session / Delete session, visual-only this iteration; workspace-header … menu: Rename (wired) / Delete workspace (visual-only). Menus close when the pointer leaves them.
+- Session-row … menu: Rename / Fork session / Delete session; Rename and Fork are wired, while Delete remains visual-only. The workspace-header … menu's Rename / Delete workspace actions are both wired. Menus close when the pointer leaves them.
 - Supporting primitives: `Menu` gains label entries, danger rows, and `closeOnPointerLeave`; a new `HoverCard` (portaled placement, open delay, disabled guard).
 
 ### workspace.rename
@@ -30,7 +30,7 @@ The group-by menu offers two modes, WorkSpace / In one list. Flat mode renders e
 
 The `session/event` → `touchSession` activity-pinning chain is deleted wholesale; the workspace account order is now manually owned — new sessions prepend at attach, and explicit reordering goes through `workspace.insertSessionBefore({ workspaceId, sessionId, beforeSessionId? })` (DOM insertBefore semantics: with an anchor it inserts before it, omitted appends to the end). The entity throws a typed `WorkspaceMoveInvalidError` only for unaccounted session/anchor ids; the handler maps exactly that to the business code `workspace-move-invalid`, while storage failures stay internal.
 
-The UI is HTML5 drag on root rows inside a group (workspace grouping only, outside search; fork children ride with their parent and are not draggable). Order authority stays entirely host-side: drop only sends the RPC, the client performs zero local reordering, and the view refreshes from the response upsert and the changed frame; a failed move changes nothing. The client's upsert rejects snapshots older (`updatedAt`) than the installed projection so a late unary response cannot roll back a newer frame.
+The UI is HTML5 drag on session rows inside a group (workspace grouping only, outside search; fork children and their source sessions are ordered independently). Order authority stays entirely host-side: drop only sends the RPC, the client performs zero local reordering, and the view refreshes from the response upsert and the changed frame; a failed move changes nothing. The client's upsert rejects snapshots older (`updatedAt`) than the installed projection so a late unary response cannot roll back a newer frame.
 
 ### Shell/region split
 
@@ -46,15 +46,15 @@ ui-sidebar shrinks to the column-geometry shell: brand row, fold state machine,
 
 **Keep the rename dialog in ui-sidebar (smallest change)** — that is the problem itself: workspace-domain dialogs scattered in a borrowed slot, with each addition (the Delete confirmation is coming) repeating the cross-package wiring. Review first considered moving only the rename modal; the ruling was to give the whole browsing region to ui-workspace and leave the shell geometry-only.
 
-**Keep parent/child adjacency in flat mode** — contradicts strict recency (a child newer than its parent's sibling cannot slot adjacently), and the flat view's purpose is dropping the hierarchy; flattening fully and disabling drag in flat mode (no persistence carrier) is more consistent.
+**Nest sessions by fork lineage in WorkSpace mode** — nesting makes the current child visible only while its ancestors are expanded and limits in-group manual ordering to root nodes; `parentId` is lineage data, not a list-navigation structure. Flattening all sessions into peer rows lets each row be opened, searched, and ordered independently; In one list still disables drag because it has no workspace persistence carrier.
 
 ## Consequences
 
 - Manual order is the sole authority over the workspace account: an order the user arranges is never scrambled by activity; the cost is losing float-to-top-on-activity, whose signal now rides the row status dot and time label. The `WorkspaceView.sessionIds` wire contract is reworded to the manual-order semantics.
 - The two-fact shell/region contract funnels every future workspace-domain feature (Delete confirmation, cross-group moves, Ungrouped adoption) into the single ui-workspace package; ui-sidebar no longer evolves with session-list features.
 - Flat mode supports neither reordering nor a create-in-workspace entry point (switching back to grouped view is required) — an accepted scope reduction.
-- Wiring the three session-menu items and workspace Delete, and growing the wire status enum, remain future iterations.
+- Wiring session Delete and growing the wire status enum remain future iterations.
 
 ## Testing
 
-Package-level suites cover the derivations (deriveGroups/deriveFlat), row components, both apply registrations and passthroughs, host entity move semantics, and the rename/insertSessionBefore RPC implementations with their fixture stubs; the `apps/web` keyless snapshots regress the assembled application; delivery acceptance additionally runs a 12-item playwright (chromium headless) checklist (grouped default, flat switch and persistence, hover-card appearance and suppression, both menus, the full rename chain, drag persistence) and drives the real host over the wire for rename success / duplicate rejection / `workspace-move-invalid`.
+Package-level suites cover the derivations (deriveGroups/deriveFlat), peer session rows, both apply registrations and passthroughs, host entity move semantics, and the rename/insertSessionBefore RPC implementations with their fixture stubs; the `apps/web` keyless snapshots regress the assembled application and pin that a fork does not introduce session expansion controls.

.agents/notes/implemented/feature/2026-07-25-session-list-browsing-and-manual-order.zh.md

@@ -12,14 +12,14 @@ Status: implemented
 
 ## Decision
 
-### 平铺视图与浏览态
+### 平铺行与浏览态
 
-group-by 菜单提供 WorkSpace / In one list 两种模式。平铺模式把所有 session(含 fork 子)一律作为顶层行,严格按 `updatedAt` 新→旧排序,不保持父子相邻;Intent 占位行渲染在列表首行。模式选择持久化在浏览器(`dsh.workspace.view`),刷新保持。
+group-by 菜单提供 WorkSpace / In one list 两种模式。WorkSpace 模式按 `WorkspaceView.sessionIds` 的手动序在各组内展示同级 session 行；In one list 把所有 session 合并后严格按 `updatedAt` 新→旧排序。两种模式都不把 `parentId` 投影成列表层级，fork 谱系只保留为 session 数据；完整 fork 行为由 [Web session fork 操作](2026-07-27-web-session-fork-actions.md)定义。模式选择持久化在浏览器(`dsh.workspace.view`),刷新保持。
 
 ### 行交互
 
 - session 行悬停 500ms 出详情卡(全名/相对时间/状态行;状态本期只有 running/idle 两态,枚举扩展待 wire 增补 status 字段)。卡片与行菜单互斥:菜单开启或拖拽进行中不出卡。
-- session 行 … 菜单:Rename / Fork session / Delete session,本期纯视觉;workspace 组头 … 菜单:Rename(已接线)/ Delete workspace(纯视觉)。菜单鼠标移出即关。
+- session 行 … 菜单:Rename / Fork session / Delete session，其中 Rename 与 Fork 已接线，Delete 仍为纯视觉；workspace 组头 … 菜单的 Rename / Delete workspace 均已接线。菜单鼠标移出即关。
 - 支撑件:`Menu` 新增 label 条目、danger 行、`closeOnPointerLeave`;新增 `HoverCard`(portal 定位、开启延时、disabled 守卫)。
 
 ### workspace.rename
@@ -30,7 +30,7 @@ group-by 菜单提供 WorkSpace / In one list 两种模式。平铺模式把所
 
 `session/event` → `touchSession` 活动置顶链整体删除;workspace 账本序改为纯手动拥有——新 session attach 时前插,显式重排走 `workspace.insertSessionBefore({ workspaceId, sessionId, beforeSessionId? })`(DOM insertBefore 语义:锚给了插锚前,缺省 append 到末尾)。实体只对不在账的 session/锚抛类型化的 `WorkspaceMoveInvalidError`,handler 仅把它映射为业务码 `workspace-move-invalid`,存储故障保持 internal。
 
-UI 为组内 root 行的 HTML5 拖拽(仅 workspace 分组、非搜索态;fork 子随父不单独拖)。顺序权威完全在 host:drop 只发 RPC,client 零本地重排,视图靠响应体 upsert 与 changed 帧刷新;失败即无事发生。client 的 upsert 拒绝比已装载投影更旧(`updatedAt`)的快照,防迟到的一元响应回滚更新的帧。
+UI 为组内 session 行的 HTML5 拖拽(仅 workspace 分组、非搜索态；fork 子与源会话一样独立排序)。顺序权威完全在 host:drop 只发 RPC,client 零本地重排,视图靠响应体 upsert 与 changed 帧刷新;失败即无事发生。client 的 upsert 拒绝比已装载投影更旧(`updatedAt`)的快照,防迟到的一元响应回滚更新的帧。
 
 ### 壳/区域切分
 
@@ -46,15 +46,15 @@ ui-sidebar 缩为列几何壳:品牌行、折叠状态机、New Session、Settin
 
 **rename 对话框留在 ui-sidebar(最小改动)** —— 正是问题本身:workspace 域的对话框散落在借来的坑里,每加一个(Delete 确认框将至)都重演跨包接线。评审中先议了「只挪 rename Modal」的中间态,最终裁定整个浏览区域归 ui-workspace,壳只留几何。
 
-**平铺模式保持父子相邻成组** —— 与「严格按时间」矛盾(子新于兄则插不进相邻位),且平铺本意就是取消层级;拉平并禁用平铺下的拖拽(无持久化载体)更一致。
+**WorkSpace 模式按 fork 谱系嵌套 session** —— 嵌套会让当前子会话依赖祖先展开态才能可见，也让组内手动序只能移动根节点；`parentId` 是 lineage 数据，不是列表导航结构。所有 session 拍平成同级行后，每行都可独立打开、搜索与排序；In one list 仍因没有 workspace 持久化载体而禁用拖拽。
 
 ## Consequences
 
 - 手动序是唯一的 workspace 账本序权威:用户排好的顺序不再被活动打乱;代价是「最近活跃浮到最上」的行为消失,活跃感知转由行内状态点与时间标签承担。`WorkspaceView.sessionIds` 的 wire 契约随之改为手动序措辞。
 - 壳/区域两事实契约把 workspace 域的后续功能(Delete 确认、跨组移动、Ungrouped 收编)全部收进 ui-workspace 单包;ui-sidebar 不再随 session 列表功能演进。
 - 平铺模式不支持排序与分组入口(建到指定 workspace 需切回分组视图),是拍板接受的范围收窄。
-- session 菜单三项与 workspace Delete 的功能接线、状态枚举扩 wire,留待后续迭代。
+- session Delete 的功能接线与状态枚举扩 wire,留待后续迭代。
 
 ## Testing
 
-包级用例覆盖派生(deriveGroups/deriveFlat)、行组件、两处 apply 注册与透传、host 实体移位语义、rename/insertSessionBefore 的 RPC 实现与 fixture 桩;`apps/web` keyless snapshot 回归覆盖装配后的应用;交付验收另以 playwright(chromium headless)过 12 项清单(分组默认、平铺切换与持久化、hover 卡出现与抑制、双菜单、rename 全链、拖拽落盘),并对真 host 直打 wire 验证 rename 成功/重名拒绝/`workspace-move-invalid` 三径。
+包级用例覆盖派生(deriveGroups/deriveFlat)、同级 session 行、两处 apply 注册与透传、host 实体移位语义、rename/insertSessionBefore 的 RPC 实现与 fixture 桩；`apps/web` keyless snapshot 回归覆盖装配后的应用，并钉住 fork 后没有 session 展开控件。

.agents/notes/implemented/feature/2026-07-27-web-session-fork-actions.i18n.yaml

@@ -0,0 +1,6 @@
+# Bilingual-pair consistency record (docs/i18n/README.md): the git blob hash of each
+# side as of the last confirmed-consistent state. Both languages carry equal authority;
+# after editing either side, bring the other along and re-record with:
+#   pnpm run verify-translation-pairing --write .agents/notes/implemented/feature/2026-07-27-web-session-fork-actions.md
+2026-07-27-web-session-fork-actions.md: b5dc7e820de069a68b38ed87c7d29ffbdb4867bc
+2026-07-27-web-session-fork-actions.zh.md: 774cd74d69eb02d43ca01c8ec7ba1cf94c24b1dc

.agents/notes/implemented/feature/2026-07-27-web-session-fork-actions.md

@@ -0,0 +1,33 @@
+# Agent Note: Web session fork actions
+
+Status: implemented
+
+English | [中文](2026-07-27-web-session-fork-actions.zh.md)
+
+## Problem
+
+The Session store already provides a fork primitive that creates a child session from a completed-turn prefix, but the Web client has no unified interaction contract. The Session-row menu can express only “branch from the latest completed turn,” while message IconActions need to express “branch from the turn containing this message”; if the two entry points independently interpret the boundary, switching, and failure behavior, the same user action acquires two sets of semantics. Nesting a fork child beneath its source session also makes the newly selected child visible only while its ancestors are expanded and weakens the workspace manual-order model.
+
+## Decision
+
+The Web Session-row menu and message IconActions share the client runtime's `sessions.fork` action. A Session row passes `{ sessionId, increaseTitle: true }`, so it forks at the source session's last completed turn; a user message or settled assistant content message passes `{ sessionId, atSeq: node.seq, increaseTitle: true }`, so it forks at the turn containing that event. Only the client consumes `increaseTitle`: after adding the child session to its local list, the client increments a trailing `(N)` or `（N）` in the source session's persisted title without changing bracket style, appends ` (1)` to an unnumbered title, and skips the rename when no persisted title exists; the Host fork request still contains only `sessionId` and the optional `atSeq`. The caller opens the child only after the rename succeeds; a fork or rename failure leaves the source session and current selection unchanged, while a child created before a rename failure remains in the list.
+
+`forkAt(seq)` touches the session service only in ui-conversation's apply injection layer; message components report only the event `seq`. Session rows likewise initiate the operation only through ui-workspace's injected callback. Neither presentation package owns session mutation state or duplicates the host's boundary evaluation.
+
+Session lineage is not projected into a list hierarchy. WorkSpace mode displays source sessions and all fork children as peer rows in the manual order from `WorkspaceView.sessionIds`; every row can be opened, searched, and dragged independently. In one list mode continues to sort strictly by `updatedAt`; the Ungrouped group also sorts by recency when no workspace ledger is available. `parentId` remains available for lineage, tool presentation, and later queries, but does not control session-list visibility.
+
+## Alternatives considered
+
+**Wire only the Session-row menu.** Rejected: at a message, the user has already selected more precise context; forcing them back to the list can only degrade the boundary to the latest completed turn, while the visible message branch icon would remain non-responsive.
+
+**Allow branching only from user messages.** Rejected: settled assistant content also has a stable event `seq`, and the host places it in its containing completed turn; making only one of two visually identical branch buttons work would create an invisible behavioral difference.
+
+**Nest fork children beneath their source by `parentId`.** Rejected: lineage is not navigation ownership; nesting requires automatic ancestor expansion to reveal the current item and prevents children from participating in the workspace's peer manual order.
+
+**Call the session service directly from message components.** Rejected: client components must not touch `ctx` or business services; injected callbacks keep mutation in the apply world and leave components driven purely by props.
+
+## Consequences
+
+Users can create forks from Session rows, user messages, or settled assistant content messages; all three entry points ultimately use the same runtime/host operation. Message entry points preserve the exact event boundary, while the list entry point preserves the “latest completed turn” shortcut. Successive fork titles increment through `(1)`, `(2)`, and so on instead of repeatedly appending `(1)`; titles with fullwidth parentheses retain that style. Every fork child immediately appears as an ordinary peer row, so the list no longer needs session expansion state, recursive nodes, or twist controls.
+
+Fork and child-rename failures stay silent and preserve the source selection, preventing a derivation action from disrupting the current reading position; this tradeoff also means the UI does not yet expose a failure reason or retry entry point. Package tests separately pin the two message `seq` paths, title increments, and the peer-list derivation; `apps/web/tests/message-actions.e2e.ts` exercises assistant-message branching and Session-row menu branching through the assembled application.

.agents/notes/implemented/feature/2026-07-27-web-session-fork-actions.zh.md

@@ -0,0 +1,33 @@
+# Agent Note: Web session fork 操作
+
+Status: implemented
+
+[English](2026-07-27-web-session-fork-actions.md) | 中文
+
+## Problem
+
+Session store 已提供按完成轮前缀创建子会话的 fork 原语，但 Web 端没有一份统一的交互契约。Session 行菜单只能表达「从最新完成轮分支」，消息 IconActions 还需要表达「从这条消息所在轮分支」；如果两处各自解释边界、切换与失败行为，同一个用户动作会形成两套语义。把 fork 子会话嵌套在源会话下还会让新选中的子会话依赖祖先展开态才能看见，并削弱 workspace 的手动排序模型。
+
+## Decision
+
+Web 的 session 行菜单与消息 IconActions 共用 client runtime 的 `sessions.fork` 操作。Session 行传 `{ sessionId, increaseTitle: true }`，因此在源会话最后一个已完成轮次处分支；用户消息与已定稿 assistant 内容消息传 `{ sessionId, atSeq: node.seq, increaseTitle: true }`，因此在包含该事件的轮次处分支。`increaseTitle` 只由 client 消费：子会话进入本地列表后，client 把源会话持久化标题尾部的 `(N)` 或 `（N）` 递增并保留括号样式，无编号时追加 ` (1)`，没有持久化标题时不改名；Host fork 请求仍只有 `sessionId` 与可选的 `atSeq`。改名成功后调用方才打开子会话；fork 或改名失败时保持源会话与当前选择不变，改名失败时已创建的子会话仍留在列表中。
+
+`forkAt(seq)` 只在 ui-conversation 的 apply 注入层接触 session 服务，消息组件只回传事件 `seq`。Session 行同理只通过 ui-workspace 的注入回调发起操作；两个呈现包都不持有 session mutation 状态，也不复制 host 的边界求值。
+
+Session lineage 不投影成列表层级。WorkSpace 模式按 `WorkspaceView.sessionIds` 的手动序把源会话与所有 fork 子会话显示为同级行，每行都可独立打开、搜索和拖拽；In one list 模式继续按 `updatedAt` 严格排序；Ungrouped 组在没有 workspace 账本时也按 recency 排序。`parentId` 仍用于 lineage、工具呈现和后续查询，但不控制 session 列表可见性。
+
+## Alternatives considered
+
+**只接 session 行菜单。** 否决：用户在消息处已经选择了更精确的上下文，强迫其回到列表只能退化为最新完成轮，且已展示的消息分支图标会成为无响应控件。
+
+**只允许用户消息分支。** 否决：已定稿 assistant 内容同样有稳定事件 `seq`，host 会把它归入所属完成轮；让两个外观相同的分支按钮只有一个可用会制造不可见的行为差异。
+
+**按 `parentId` 把 fork 子会话嵌套在源会话下。** 否决：lineage 不是导航所有权；嵌套要求自动展开祖先才能看见当前项，并让子会话无法参与 workspace 的同级手动排序。
+
+**由消息组件直接调用 session 服务。** 否决：client 组件不得接触 `ctx` 或业务服务；注入回调让 mutation 留在 apply 世界，组件保持纯 props。
+
+## Consequences
+
+用户可从 session 行、用户消息或已定稿 assistant 内容消息创建分支，三处最终走同一个 runtime/host 操作；消息点位保留精确事件边界，列表点位保留「最新完成轮」快捷语义。连续 fork 的标题按 `(1)`、`(2)` 递增，而不是重复追加 `(1)`；全角括号标题保持全角样式。所有 fork 子会话立即作为普通同级行出现，列表不再需要 session 展开状态、递归节点或 twist 控件。
+
+Fork 与子会话改名失败都保持静默并保留源选择，避免一个派生操作破坏当前阅读位置；该取舍也意味着 UI 暂不提供失败原因或重试入口。Package tests 分别钉住两种消息 `seq`、标题递增与同级列表派生，`apps/web/tests/message-actions.e2e.ts` 通过装配后的应用执行 assistant 消息分支与 session 行菜单分支。

.agents/notes/implemented/feature/2026-07-27-web-session-search.i18n.yaml

@@ -0,0 +1,6 @@
+# Bilingual-pair consistency record (docs/i18n/README.md): the git blob hash of each
+# side as of the last confirmed-consistent state. Both languages carry equal authority;
+# after editing either side, bring the other along and re-record with:
+#   pnpm run verify-translation-pairing --write .agents/notes/implemented/feature/2026-07-27-web-session-search.md
+2026-07-27-web-session-search.md: 9a634c586a4793d1c6986a7e7c0b0c1157b5b687
+2026-07-27-web-session-search.zh.md: 5ec2baf7443aaaaa75abc348ee426df9c14fbaa2

.agents/notes/implemented/feature/2026-07-27-web-session-search.zh.md

@@ -0,0 +1,44 @@
+# Agent Note: Web 历史会话搜索
+
+Status: implemented
+
+[English](2026-07-27-web-session-search.md) | 中文
+
+## 问题
+
+Web 侧边栏会展示会话标题及其 Workspace 归属，但无法根据只出现在消息中的词语检索历史对话。在浏览器中扫描历史记录，需要附加或加载每个会话，重复实现现有的索引搜索服务，也会让冷态持久化会话的检索既缓慢又容易遗漏。产品还需要一条可预测的故障路径：派生索引不可用时，不得抹去客户端能够在本地计算出的标题匹配结果。
+
+## 决策
+
+Web 与 headless 共用的组合会使用 `openAt: first-search` 和内存数据库挂载 [`@deepseek-ai/dsh-session-query-sqlite`](../../../../packages/session-query/session-query-sqlite/README.md)。服务启动时处于 ACTIVE 状态，而其 `node:sqlite` 模块与连接私有句柄分别要到首次内容查询才会导入和打开。这让 Node 22 的启动输出在使用搜索前不会出现 SQLite 实验性警告，但并不承诺在首次搜索导入该模块时抑制警告。每个服务实例都独占自己的索引，因此并行 CLI 或 Web 调用可维持 SQLite 后端的单一所有者契约，又不会留下进程级派生文件。数据库从空状态启动，并在该首次查询时惰性对齐实时会话与持久化会话。它仍是与规范 JSONL 持久化相互独立的可丢弃派生索引。
+
+宿主网关通过现有的类型化 RPC 栈公开 `session.search`。它根据 `session.list` 使用的同一组可见摘要推导授权集合，向 `ctx.sessionQuery.searchSessions` 请求全局排序后的当前 surface `user/message`、`assistant/message` 和 `steering/message` 匹配项，并持续消费提供方分页，直到获得 20 个已授权会话及一个前瞻项，或结果流耗尽。首个提供方页面请求 20 个命中；如果第一页返回 `SESSION_QUERY_INVALID_LIMIT`，页面大小会依次折半为 10、5、2、1，并在续传和陈旧世代重启中沿用探测所得的大小。每个命中的会话 id、最佳匹配会话 id、surface 和事件类型都会经过重新校验，其 snippet 才能离开宿主。发出的 snippet 最多包含 240 个 Unicode 码点；宿主与传输 schema 共用协议边界及码点安全的截断辅助函数，而传输 schema 会在客户端解析时独立强制执行 snippet 上限。将可能很大的授权集合排除在 SQLite 绑定之外，可避开可移植变量上限，同时保持全局排序。响应仍只有一个有界页面；`hasMore` 会指示 UI 提示用户缩小查询范围，而不是公开分页能力。陈旧的续传会丢弃当前尝试的部分结果、去重条目和游标，然后依据原始可见性快照从第一页重新开始。上限探测与陈旧重试共用 100 次提供方调用的限制（因此最多检查 2,000 个命中）；如果某页命中数超过其请求的上限、续传游标重复，或用尽该调用预算后结果流仍未耗尽，都会直接返回 `internal` 业务错误，不返回部分结果。载体信号会取消已被取代的工作，包括持久化列表枚举、分批受限执行的冷会话元数据 stat，以及每一次提供方调用；即使同时收到上限拒绝或陈旧拒绝，也以取消为准。查询服务缺失或索引／查询故障无法恢复时，仍作为业务错误处理，不会修改规范会话存储。
+
+[`WorkspaceBrowser`](../../../../packages/client/ui-workspace/README.md) 有意将元数据搜索与内容搜索保持独立。其默认界面文案为英文；输入框及防御性请求路径会移除 NUL，将查询限制在请求 schema 规定的 500 个 UTF-16 code unit 内且不会拆分 surrogate pair。非空白查询会立即从会话列表中计算不区分大小写的标题和 Workspace 子串匹配，在 250 ms 防抖后发起内容请求，在查询变化时中止前一请求，并忽略陈旧的完成结果。它先按新近程度排列本地匹配，再合并由后端排序且仅匹配内容的结果，按会话 id 去重；无论常规分组模式如何，最终都渲染为扁平列表。每一行显示标题、Workspace，并在存在时显示一行摘要片段。选择某一行只会打开对应会话，并保留查询条件；不会跳转至确切事件。
+
+结果上限是单一协议常量，而非逐连接状态。`SESSION_SEARCH_RESULT_LIMIT` 位于 `dsh-host-apiproxy` 中强制执行它的响应 schema 旁边，`SessionsService.searchResultLimit` 则把该常量重新公开给呈现插件。功能包要取用它，必须显式扩展 sessions 域的对外面：`ISessions`（即注入为 `ctx.sessions` 的那个面，也因此是测试运行时的 sessions 替身必须实现的面）在该上限旁声明了搜索动作。连接 handle 不携带它：逐连接字段会暗示该上限随传输层变化或由服务端协商，而 schema 固定的 `max` 恰恰禁止这一点，并且会让同一事实在同一模块内拥有两处归属。
+
+内容匹配沿用 SQLite 后端经过规范化的字面 token／短语语义。共享语义投影会排除推理（reasoning）块，因此 UI 搜索绝不会将模型的私有推理作为命中或 snippet 返回；派生索引的 schema 版本会随之前进，使现有持久化索引重建并移除先前的这些文档。FTS5 运算符只作为数据处理，此搜索界面不提供拼写错误纠正、模糊匹配、前缀匹配或任意子串扩展。特别是，`unicode61` 分词器可能将一段连续中文视作单个 token，因此不保证 `搜索` 之类的较短查询能匹配 `会话搜索功能` 的内部片段。标题与 Workspace 匹配仍采用普通的客户端子串匹配。
+
+## 故障与可见性契约
+
+搜索绝不会扩大会话可见范围：没有可供服务的 cwd 的冷会话会被排除，原因与它们不出现在 `session.list` 中相同；只有 id 位于这条基线中的提供方命中才能离开宿主。被遮蔽事件和纯日志事件、消息内容之外的工具事件、错误、待办事项及其他追踪记录都不会产生 UI 命中结果。
+
+首个或后续内容请求仍在处理期间，UI 会保留即时元数据匹配结果，并显示历史搜索状态。如果后端失败，这些行会保持不变，并显示警告说明内容搜索不可用。合并后没有任何行时，界面会显示明确的空状态。候选行超过 20 条时，界面会提示用户缩小查询范围。
+
+## 曾考虑的替代方案
+
+- **在浏览器中扫描每个会话的历史记录**：不予采纳，因为这会让 UI 承担传输与折叠开销；除非加载冷态日志，否则还会漏掉这些日志；并会重复实现已经由 `ctx.sessionQuery` 负责的语义提取与源对齐。
+- **首版即加入 trigram 或模糊搜索**：不予采纳，因为这会改变索引大小、排序、短查询行为与产品预期。trigram 本身也无法解决双字查询。首版沿用现有后端契约，将召回扩展留作另一项基于度量结果的决策。
+- **返回事件地址并跳转至确切匹配位置**：本版不予采纳，因为对话虚拟化与稳定的事件导航需要单独的 UI 契约。会话级导航本身已有价值，无需让搜索与这项工作耦合。
+- **在侧边栏公开游标分页**：不予采纳，改为固定显示前 20 条结果并提示缩小查询范围；这样可使交互与取消状态保持有界。
+
+## 后果
+
+无需预先打开，即可检索到历史持久化对话，同时宿主仍只保留一条可见性边界和一套语义索引实现。即时本地结果掩盖了大部分请求延迟，取消机制可防止已作废查询重新渲染列表，后端故障则会降级为内容搜索尚不可用时已有的行为。
+
+首次内容查询可能耗时更长，因为它要先导入并打开 SQLite，再承担惰性对齐的开销。搜索质量采用 token／短语召回，而不是模糊召回或任意子串召回，并受上述连续中文限制。结果粒度为会话，最多 20 条，不支持分页，也不能跳转到具体消息。如果有效但选择性极差或反复陈旧的提供方尝试未能在 100 次调用内完成，系统会进入仅保留元数据匹配的故障路径，而不是无限制地继续处理。
+
+## 测试
+
+宿主测试将请求与响应校验、可见会话过滤、事件和 surface 过滤、结果与 snippet 边界、共享调用预算内的自适应提供方上限、沿用探测所得上限的陈旧世代重启、游标与跨页去重行为、取消优先级及故障映射固定为契约。SQLite 生命周期测试将启动时激活、首次搜索时的打开与失败、共享就绪状态以及未打开状态下的处置固定为契约；语义提取测试与 SQLite／fixture 搜索测试将排除仅存在于推理中的文本固定为契约。Node 22 兼容性门禁会构建 CLI 与 Web 产物，在移除环境级警告抑制并采用隔离的临时 home／提供方环境后，以普通 Node 启动随产品交付的 `dsh web`／`AppCLIEntry` 组合，等待启动完成并稳定，再沿随产品交付的信号路径对其执行 dispose（资源释放）。fixture（测试前置数据）、运行时与 UI 测试将以匹配位置为中心的有界 snippet、无状态委托、500 个 code unit 的查询边界、防抖／中止／陈旧响应行为、本地回退、合并顺序、去重、英文文案、ARIA 树成员关系、行渲染与导航语义固定为契约。无密钥的组装层 Web 测试会在保留惰性打开配置的同时，播种一段尚未打开的持久化对话，通过 SQLite 索引按可见消息内容找到它，捕获侧边栏结果，打开该会话，并验证查询条件仍然保留。

.agents/notes/implemented/feature/2026-07-28-sdk-max-output-tokens.i18n.yaml

@@ -2,5 +2,5 @@
 # side as of the last confirmed-consistent state. Both languages carry equal authority;
 # after editing either side, bring the other along and re-record with:
 #   pnpm run verify-translation-pairing --write .agents/notes/implemented/feature/2026-07-28-sdk-max-output-tokens.md
-2026-07-28-sdk-max-output-tokens.md: 5db48f21892d56addea7b72f73319f9dbfd1e71f
-2026-07-28-sdk-max-output-tokens.zh.md: 38b172718716d100726188163ff22be9ea0a7325
+2026-07-28-sdk-max-output-tokens.md: 3ba3e226d64b7d3d192d67bd88af463d2b0d9dc5
+2026-07-28-sdk-max-output-tokens.zh.md: aec566011d2d7a311b4de509c47ebba383c3b0d7

.agents/notes/implemented/feature/2026-07-28-sdk-max-output-tokens.md

@@ -12,7 +12,7 @@ The Python and TypeScript SDKs could select a provider and model but could not b
 
 The high-level SDKs expose one optional process-wide output cap: Python names it `max_tokens`, TypeScript names it `maxTokens`, and the shared `initialize` wire payload carries `maxTokens`. The JSON-RPC server rejects values that are not positive safe integers and stores the accepted cap with its provider/model route.
 
-Each SDK-created root Agent receives the cap through `AgentOptions.maxTokens`. Agent Loop places that value in the initial `LlmCallConfig`, logs it in the request header, and reconstructs every dispatched conversation request from that durable header. Omitting the option leaves `maxTokens` absent so the selected provider retains its default.
+Each SDK-created root Agent receives the cap through `AgentOptions.maxTokens`. Agent Loop places that value in the initial `LlmCallConfig`; final call preparation preserves the explicit value or materializes an exact-model adapter default, logs the effective cap in the request header, and reconstructs every dispatched conversation request from that durable header. Omitting the SDK option therefore allows the selected adapter or provider route default to apply.
 
 In-process subagents inherit the parent's provider, model, and output cap. An explicit `SubagentStartRequest.agentOptions.maxTokens`, including one configured by `dsh-tool-subagent`, overrides the inherited value for that child and its descendants. Out-of-process providers own the configuration of their separate runtime; `subagent-dsh-sdk` therefore exposes its own optional `maxTokens` and forwards it through that child runtime's SDK handshake.
 
@@ -20,7 +20,7 @@ Compaction, session-title generation, web search, and other auxiliary calls keep
 
 ## Alternatives considered
 
-**Set an adapter environment variable.** This would be DeepSeek-adapter-specific, invisible in the session request header, ineffective for intercepted or alternate adapters, and easy to confuse with a provider default. The cap belongs in provider-neutral request configuration.
+**Set only an adapter environment variable.** A serializer-private fallback would be DeepSeek-adapter-specific, invisible in the session request header, ineffective for intercepted or alternate adapters, and easy to confuse with a provider default. Adapter-owned defaults may instead be exposed as exact-model metadata and materialized into provider-neutral request configuration before logging.
 
 **Add `maxTokens` to every `session/prompt`.** Per-turn mutation would enlarge the wire and introduce request-config transitions that callers do not need for the current evaluation use case. A runtime initialization option gives every session in one SDK process the same reproducible budget.
 

.agents/notes/implemented/feature/2026-07-28-sdk-max-output-tokens.zh.md

@@ -12,7 +12,7 @@ Python 与 TypeScript SDK 可以选择提供方和模型，却无法限制对话
 
 高层 SDK 公开一个可选的进程级输出上限：Python 命名为 `max_tokens`，TypeScript 命名为 `maxTokens`，共享的 `initialize` 线载荷使用 `maxTokens`。JSON-RPC 服务端拒绝非正安全整数，并将通过校验的上限与提供方／模型路由一同保存。
 
-每个由 SDK 创建的根 Agent 都通过 `AgentOptions.maxTokens` 获得该上限。Agent Loop 将它放入初始 `LlmCallConfig`、记录到请求 header，并从该持久化 header 重建每次分派的对话请求。省略该选项时，`maxTokens` 保持缺失，由所选提供方保留默认值。
+每个由 SDK 创建的根 Agent 都通过 `AgentOptions.maxTokens` 获得该上限。Agent Loop 将它放入初始 `LlmCallConfig`；最终调用准备会保留显式值，或填入确切模型的适配器默认值，再将生效上限记录到请求 header，并从该持久化 header 重建每次分派的对话请求。因此，省略 SDK 选项时会应用所选适配器或提供方路由的默认值。
 
 进程内 subagent 继承父级的提供方、模型和输出上限。显式的 `SubagentStartRequest.agentOptions.maxTokens`（包括通过 `dsh-tool-subagent` 配置的值）会覆盖该子级及其后代的继承值。进程外提供方自行持有其独立运行时的配置；因此 `subagent-dsh-sdk` 公开独立的可选 `maxTokens`，并通过该子运行时自己的 SDK 握手传入。
 
@@ -20,7 +20,7 @@ Python 与 TypeScript SDK 可以选择提供方和模型，却无法限制对话
 
 ## Alternatives considered
 
-**设置适配器环境变量。** 这种方式仅适用于 DeepSeek 适配器，不会出现在会话请求 header 中，对被拦截请求或其他适配器无效，也容易与提供方默认值混淆。该上限属于提供方无关的请求配置。
+**仅设置适配器环境变量。** 序列化器私有回退仅适用于 DeepSeek 适配器，不会出现在会话请求 header 中，对被拦截请求或其他适配器无效，也容易与提供方默认值混淆。适配器持有的默认值可以改为通过确切模型元数据公开，并在记录前填入提供方无关的请求配置。
 
 **在每个 `session/prompt` 上增加 `maxTokens`。** 按轮次修改会扩大线协议，并引入当前评测用例不需要的请求配置转换。运行时初始化选项可让一个 SDK 进程中的每个会话拥有相同、可重现的预算。
 

.agents/notes/implemented/feature/2026-07-29-addressable-queue-operations.i18n.yaml

@@ -2,5 +2,5 @@
 # side as of the last confirmed-consistent state. Both languages carry equal authority;
 # after editing either side, bring the other along and re-record with:
 #   pnpm run verify-translation-pairing --write .agents/notes/implemented/feature/2026-07-29-addressable-queue-operations.md
-2026-07-29-addressable-queue-operations.md: 78a7d346163bb7e5e76c989c6e93576b4a6cee64
-2026-07-29-addressable-queue-operations.zh.md: 050b9755ad4ebe70e2bdcafb711ef279331e27af
+2026-07-29-addressable-queue-operations.md: 7a08b889c958e583dc430d33a1855fe3725f3d48
+2026-07-29-addressable-queue-operations.zh.md: 701b028c7494fd7cb608d05a5d170c9075b155d7

.agents/notes/implemented/feature/2026-07-29-addressable-queue-operations.md

@@ -18,7 +18,7 @@ The Web queue rendered pending messages but could not edit or delete one row. `M
 
 **Queue addresses require a live Agent.** `session.updateQueue` queries only the mounted Agent registry and never resumes a cold session: an `InboxItemId` is process-local and cannot name work after restart or disposal. A missing Agent and a driver-claimed occurrence both return `queue-item-not-found`.
 
-**Web actions address Queue only.** The Host excludes pending steering from `session/queue`; steering retains its existing durable transcript path after consumption. QueueDock exposes edit and delete, but no send-now control. The UI derives queue row and mutation types from the runtime `SessionFace` contract rather than importing the connection plugin, so plugin cooperation continues through services and snapshots. Edit is available only when all content blocks are text; the editor cannot silently drop non-text blocks. An editing row exposes only save and cancel, with Enter and Escape as their keyboard equivalents. Delete removes the exact occurrence.
+**Web actions address Queue only.** The Host excludes pending steering from `session/queue`; steering retains its existing durable transcript path after consumption. QueueDock hides while empty, renders one pending occurrence directly, and defaults two or more occurrences to a collapsed `"<n> 条排队消息"` header that expands or collapses the complete list. The header exposes `aria-expanded` and `aria-controls`; the expanded list scrolls within a 180px height bound. An active edit or mutation keeps its rows visible, and emptying the queue restores the collapsed default for the next queue. Visible rows expose edit and delete, but no send-now control. The UI derives queue row and mutation types from the runtime `SessionFace` contract rather than importing the connection plugin, so plugin cooperation continues through services and snapshots. Edit is available only when all content blocks are text; the editor cannot silently drop non-text blocks. An editing row exposes only save and cancel, with Enter and Escape as their keyboard equivalents. Delete removes the exact occurrence.
 
 ## Alternatives considered
 
@@ -34,7 +34,7 @@ The Web queue rendered pending messages but could not edit or delete one row. `M
 
 ## Verification
 
-AgentLoop contract tests hold prompt admission while editing and removing exact queued occurrences, reject mutations of steering occurrences, and verify the resulting independent turn and terminal lifecycle events. Host schema and proxy tests cover queued-only authoritative snapshots, synchronous re-entrant mutation order, reconnect, cold-Agent rejection, typed not-found errors, and the RPC transport. Client runtime and QueueDock tests cover non-optimistic projection, text-only editing, save and cancel affordances, removal, retirement races, and disabled mixed-content editing. Keyless browser scenarios drive the exposed edit and delete actions through the built Web composition and real HTTP/SSE wire.
+AgentLoop contract tests hold prompt admission while editing and removing exact queued occurrences, reject mutations of steering occurrences, and verify the resulting independent turn and terminal lifecycle events. Host schema and proxy tests cover queued-only authoritative snapshots, synchronous re-entrant mutation order, reconnect, cold-Agent rejection, typed not-found errors, and the RPC transport. Client runtime and QueueDock tests cover non-optimistic projection, single-row presentation, default multi-row collapse, interaction-forced visibility, reset after emptying, expansion, text-only editing, save and cancel affordances, removal, retirement races, and disabled mixed-content editing. Keyless browser scenarios capture the default collapsed header before expanding the queue and driving its exposed edit and delete actions through the built Web composition and real HTTP/SSE wire.
 
 ## Consequences
 

.agents/notes/implemented/feature/2026-07-29-addressable-queue-operations.zh.md

@@ -18,7 +18,7 @@ Web 队列能够渲染待处理消息，但无法编辑或删除其中某一行
 
 **Queue 寻址要求 Agent 存活。** `session.updateQueue` 只查询已挂载的 Agent 注册表，绝不恢复冷会话：`InboxItemId` 属于进程本地标识，无法在重启或资源释放后继续指向工作。Agent 缺失和单次入队项已被驱动器认领这两种情况都返回 `queue-item-not-found`。
 
-**Web 操作只面向 Queue。** Host 从 `session/queue` 中排除待处理 steering；steering 消费后仍沿用既有的持久 transcript（文本记录）路径。QueueDock 暴露编辑和删除，不提供立即发送控件。UI 从运行时 `SessionFace` 契约派生队列行与变更类型，而不是导入连接插件，因此插件仍通过服务和快照协作。仅当所有内容块都是文本时才提供编辑功能；编辑器不能静默丢弃非文本块。编辑中的行只展示保存和取消操作，对应的键盘操作分别是 Enter 和 Escape。删除会移除对应的精确入队项。
+**Web 操作只面向 Queue。** Host 从 `session/queue` 中排除待处理 steering；steering 消费后仍沿用既有的持久 transcript（文本记录）路径。QueueDock 在队列为空时隐藏，只有一个待处理项时直接渲染该行，存在两个或更多待处理项时则默认收起为可展开或收起完整列表的 `"<n> 条排队消息"` 表头。表头暴露 `aria-expanded` 和 `aria-controls`；展开后的列表以 180px 为高度上限，并可滚动。存在进行中的编辑或变更时，列表行会保持可见；队列清空后，下一次出现队列时会恢复默认收起状态。可见行暴露编辑和删除操作，不提供立即发送控件。UI 从运行时 `SessionFace` 契约派生队列行与变更类型，而不是导入连接插件，因此插件仍通过服务和快照协作。仅当所有内容块都是文本时才提供编辑功能；编辑器不能静默丢弃非文本块。编辑中的行只展示保存和取消操作，对应的键盘操作分别是 Enter 和 Escape。删除会移除对应的精确入队项。
 
 ## 考虑过的替代方案
 
@@ -34,7 +34,7 @@ Web 队列能够渲染待处理消息，但无法编辑或删除其中某一行
 
 ## 验证
 
-AgentLoop 契约测试会在编辑和移除精确 queued 入队项时阻塞提示词接纳，拒绝对 steering 入队项的变更，并验证所得独立轮次及终态生命周期事件。Host schema 与代理测试覆盖仅含 queued 项的权威快照、同步可重入变更顺序、重连、拒绝冷 Agent、类型化 not-found 错误和 RPC 传输。客户端运行时和 QueueDock 测试覆盖非乐观投影、仅文本编辑、保存与取消入口、移除、退役竞态，以及禁用混合内容编辑。无密钥浏览器场景会通过构建后的 Web 组合和真实 HTTP／SSE 协议操作公开的编辑和删除。
+AgentLoop 契约测试会在编辑和移除精确 queued 入队项时阻塞提示词接纳，拒绝对 steering 入队项的变更，并验证所得独立轮次及终态生命周期事件。Host schema 与代理测试覆盖仅含 queued 项的权威快照、同步可重入变更顺序、重连、拒绝冷 Agent、类型化 not-found 错误和 RPC 传输。客户端运行时和 QueueDock 测试覆盖非乐观投影、单行呈现、多行默认收起、交互期间强制保持可见、清空后重置、展开、仅文本编辑、保存与取消入口、移除、退役竞态，以及禁用混合内容编辑。无密钥浏览器场景会先捕获默认收起的表头，再展开队列，并通过构建后的 Web 组合和真实 HTTP／SSE 协议操作其公开的编辑和删除。
 
 ## 后果
 

.agents/notes/implemented/feature/2026-07-29-directory-picker-adaptive-default.i18n.yaml

@@ -0,0 +1,6 @@
+# Bilingual-pair consistency record (docs/i18n/README.md): the git blob hash of each
+# side as of the last confirmed-consistent state. Both languages carry equal authority;
+# after editing either side, bring the other along and re-record with:
+#   pnpm run verify-translation-pairing --write .agents/notes/implemented/feature/2026-07-29-directory-picker-adaptive-default.md
+2026-07-29-directory-picker-adaptive-default.md: 7ff6529bb8e445f63343b1019ac520f56b19d5e4
+2026-07-29-directory-picker-adaptive-default.zh.md: a2a2d8ec4c91a347eedfc3aa3413b091ee847934

.agents/notes/implemented/feature/2026-07-29-directory-picker-adaptive-default.md

@@ -0,0 +1,30 @@
+# Agent Note: Adaptive default for the directory-picker interaction
+
+Status: implemented
+
+English | [中文](2026-07-29-directory-picker-adaptive-default.zh.md)
+
+## Problem
+
+The [directory-picker seam](../architecture/2026-07-28-directory-picker-capability-seam.md) made the interaction a `cordis.yml` swap point, but the shipped composition still had to pin one backend: `-browse` everywhere meant a local operator never got the OS chooser, `-native` everywhere breaks every remote deployment. The right default depends on facts only the running host knows — where the server binds, whether the process was launched over SSH, whether a display session exists — so no static row is correct for all deployments.
+
+## Decision
+
+A third sibling package, **`dsh-host-directory-picker-auto`**: a node-half-only *chooser* that owns no picking code and no UI. Its `apply` samples the host facts exactly once at boot — bind host from the injected `httpServer` (a new `host` getter mirrors the existing `port`), `SSH_CONNECTION`/`SSH_TTY`, platform, `DISPLAY`/`WAYLAND_DISPLAY`, and a `PATH` probe for a Linux chooser binary (zenity/kdialog) — resolves them through one exported pure function, and mounts the chosen dual-face backend with `ctx.loader.create({name})` into the Loader's **in-memory root tree**; the effect's disposer removes the entry and joins the backend fiber's teardown (`remove()` alone only starts it), so unloading the chooser settles only after the backend quiesced. `native` requires every attended-and-servable signal: loopback bind ∧ no SSH markers ∧ a display session the native backend can drive — assumed on darwin/win32, requiring `DISPLAY`/`WAYLAND_DISPLAY` plus a chooser binary on linux, and never true elsewhere (the native backend supports exactly darwin/win32/linux). Anything ambiguous resolves to `browse`, which works everywhere. `apps/cli` now mounts `-auto` as its `directory-picker` row; composing `-native` or `-browse` directly remains the pin.
+
+Why entry-level mounting is the load-bearing mechanism: the client module table (`dsh-client-modules`) reconciles **Loader entries** reactively over `internal/plugin`, so a backend mounted as a real entry gets its browser half discovered exactly as a config-row's would be — the seam's one-row-swaps-both-faces invariant survives adaptivity with zero duplicated client code. The dev HMR row (`AppCLIEntry`) is the mechanism precedent. Root-tree targeting matters: the root tree's `write()` is a no-op, so the resolved row can never be persisted back into `cordis.yml` (the Include subtree *does* write).
+
+## Alternatives considered
+
+- **Boot-glue resolution in `AppCLIEntry`** (ship both rows with static `disabled`, patch `disabled` from a `--directory-picker=auto|native|browse` flag). Works — `PatchOptions` patches metadata, and the modules scan skips disabled rows — but leaves the decision app-private where every future composition re-implements it; the chooser plugin gives any `cordis.yml` the same one-row adaptivity. Reintroduce the flag only when a deployment needs to *force* a backend without editing its yml.
+- **One merged plugin branching per call** (client tries `pick`, falls back to the browse dialog on `directory-picker-unavailable`). Rejected: the client would need both flows in one bundle — the bundle-purity gate forbids cross-plugin value imports and jscpd forbids copying the dialog — and per-call probing pays a doomed RPC on every open of a browse host.
+- **Resurrecting the wire advertisement** so both client flows mount and branch on the host's kind. Rejected: reverses the seam note's deletion for no consumer the chooser doesn't already serve, and collides with the `single` directory-flow holes.
+- **Per-connection adaptivity** (native for a loopback browser, browse for a remote one, same server). Deferred: needs a per-client capability, the advertisement above, and both flows mounted; no deployment serves both operator shapes at once today.
+
+## Consequences
+
+- The shipped web GUI adapts out of the box: attended local host → OS chooser; SSH launch, all-interfaces bind, headless host, unsupported platform, or Linux without a chooser binary → in-app browser. Detection infers operator location from launch context, which no launch-side signal can prove: a detached tmux session loses `SSH_*`; a non-Aqua darwin process still counts as displayed; and the `ssh -L` shape (a workstation-local launch later reached through a forwarded port, arriving from `127.0.0.1`) resolves `native` and opens the chooser on the unattended workstation — per-connection adaptivity could not fix that last case either. A wrong `native` choice degrades to the backend's existing retryable failure dialog; deployments in these shapes compose `-browse` directly.
+- The chooser mounts backends by runtime string (`BACKEND_PACKAGES`, exported), which yml-row scanning cannot see; `verify-cordis-config` therefore requires every composition mounting `-auto` to declare both backends as dependencies, so keyless Linux CI (which only ever resolves `browse`) cannot hide a dropped `-native` dependency. The shipped-tree web e2e/snapshot lane (`apps/web/tests/scaffold.ts`) pins `-browse` by disable+insert patch — its goldens are interaction-specific and must not depend on the host running the suite.
+- One resolution per boot keeps the seam's capability-stability contract; per-connection shapes remain out of scope until a deployment demands them.
+- Mounting the chooser **and** a backend row together fails loud (duplicate `directoryPicker` service; duplicate flow in the `single` holes).
+- The host typecheck aggregate now references the two backend projects (declarations only, node entries carry no client merge) so the chooser's REAL-composition test can mount them — the mirror of the client aggregate's `webserver` reference.

.agents/notes/implemented/feature/2026-07-29-directory-picker-adaptive-default.zh.md

@@ -0,0 +1,30 @@
+# Agent Note：目录选择交互的自适应默认值
+
+状态：已实现
+
+[English](2026-07-29-directory-picker-adaptive-default.md) | 中文
+
+## 问题
+
+[目录选择 seam](../architecture/2026-07-28-directory-picker-capability-seam.md)把交互形态做成了 `cordis.yml` 的切换点，但随附的组合仍必须固定一个后端：处处用 `-browse` 意味着本地操作者永远得不到 OS 选择器，处处用 `-native` 则弄坏所有远程部署。正确的默认值取决于只有运行中的宿主才知道的事实——服务器绑定在哪里、进程是否经 SSH 启动、是否存在显示会话——因此没有哪一静态行对所有部署都正确。
+
+## 决策
+
+第三个同级包 **`dsh-host-directory-picker-auto`**：一个只有 node 半侧的*选择器*，不持有任何选取代码，也没有 UI。它的 `apply` 在启动时恰好采样一次宿主事实——从注入的 `httpServer` 读绑定宿主（新增的 `host` getter 与既有的 `port` 对称）、`SSH_CONNECTION`／`SSH_TTY`、平台、`DISPLAY`／`WAYLAND_DISPLAY`、以及对 Linux 选择器二进制（zenity／kdialog）的一次 `PATH` 探查——经由一个导出的纯函数判定，再用 `ctx.loader.create({name})` 把选中的双面后端挂进 Loader 的**内存根树**；该 effect 的 disposer 会移除该条目并汇入后端 fiber 的拆卸（单靠 `remove()` 只是启动拆卸），因此卸载选择器要到后端静止之后才落定。`native` 要求全部“有人值守且可服务”信号：回环绑定 ∧ 无 SSH 标记 ∧ native 后端能驱动的显示会话——darwin／win32 上视为存在，linux 上要求 `DISPLAY`／`WAYLAND_DISPLAY` 外加一个选择器二进制，其余平台一律不成立（native 后端恰好支持 darwin／win32／linux）。任何含糊情形都判定为处处可用的 `browse`。`apps/cli` 现在把 `-auto` 挂为它的 `directory-picker` 行；直接组合 `-native` 或 `-browse` 仍是固定交互的方式。
+
+条目级挂载之所以是承重机制：client 模块表（`dsh-client-modules`）基于 `internal/plugin` 对 **Loader 条目**做响应式协调，因此以真实条目挂载的后端，其 browser half 被发现的方式与配置行完全相同——seam 的“一行同时换两面”不变式在自适应下依然成立，且没有一行重复的 client 代码。开发环境的 HMR 行（`AppCLIEntry`）是该机制的先例。瞄准根树很关键：根树的 `write()` 是 no-op，因此判定出的行绝不会被持久化回 `cordis.yml`（Include 子树*会*写回）。
+
+## 曾考虑的替代方案
+
+- **在 `AppCLIEntry` 里做启动胶水判定**（随附两行并带静态 `disabled`，由 `--directory-picker=auto|native|browse` 标志修补 `disabled`）。可行——`PatchOptions` 能修补元数据，模块扫描也会跳过禁用行——但把决策留成应用私有，此后每个组合都要重新实现；选择器插件让任何 `cordis.yml` 都获得同样的一行自适应。只有当某个部署需要不改自己的 yml 就*强制*指定后端时，才重新引入该标志。
+- **合并成一个按调用分支的插件**（client 先试 `pick`，收到 `directory-picker-unavailable` 再回退到浏览对话框）。否决：client 得把两套流程装进同一个 bundle——bundle 纯净门禁禁止跨插件的值导入，jscpd 禁止复制对话框——而且按调用探测让 browse 宿主每次打开都付出一次注定失败的 RPC。
+- **复活 wire 广播**，让两套 client 流程都挂载并按宿主的 kind 分支。否决：推翻 seam Agent Note 的那次删除，却服务不了任何选择器尚未服务的消费方，还与 `single` 目录流洞相冲突。
+- **按连接自适应**（同一台服务器，回环浏览器用 native、远程浏览器用 browse）。延期：需要按客户端的能力对象、上述广播，以及同时挂载两套流程；今天没有部署同时服务两种操作者形态。
+
+## 后果
+
+- 随附的 web GUI 开箱即自适应：有人值守的本地宿主 → OS 选择器；SSH 启动、全网卡绑定、无头宿主、不支持的平台，或没有选择器二进制的 Linux → 应用内浏览器。探测是从启动上下文推断操作者位置，而任何启动侧信号都无法证明这一点：脱离的 tmux 会话会丢失 `SSH_*`；非 Aqua 的 darwin 进程仍被算作有显示；而 `ssh -L` 形态（在工作站本地启动、之后经转发端口访问，从 `127.0.0.1` 到达）会判定 `native`，把选择器弹在无人值守的工作站上——即便按连接自适应也修不了最后这一情形。错误的 `native` 选择会退化为后端既有的可重试失败对话框；处于这些形态的部署直接组合 `-browse`。
+- 选择器按运行时字符串（已导出的 `BACKEND_PACKAGES`）挂载后端，yml 行扫描看不到这一点；因此 `verify-cordis-config` 要求每个挂载 `-auto` 的组合把两个后端都声明为依赖，使无密钥的 Linux CI（它永远只会判定出 `browse`）无法掩盖被丢掉的 `-native` 依赖。随附树的 web e2e／快照通道（`apps/web/tests/scaffold.ts`）以 disable+insert 补丁固定 `-browse`——其 golden 是交互特定的，绝不能依赖运行该套件的宿主。
+- 每次启动只判定一次，维持 seam 的能力稳定性契约；按连接的形态在有部署提出需求前仍不在范围内。
+- 同时挂载选择器**和**某个后端行会大声失败（重复的 `directoryPicker` 服务；`single` 洞中的重复流程）。
+- host 类型检查聚合现在引用两个后端项目（仅声明，node 入口不携带 client 合并），使选择器的 REAL-composition 测试能挂载它们——与 client 聚合对 `webserver` 的引用互为镜像。

.agents/notes/implemented/feature/2026-07-29-web-message-icon-actions-and-clock.i18n.yaml

@@ -2,5 +2,5 @@
 # side as of the last confirmed-consistent state. Both languages carry equal authority;
 # after editing either side, bring the other along and re-record with:
 #   pnpm run verify-translation-pairing --write .agents/notes/implemented/feature/2026-07-29-web-message-icon-actions-and-clock.md
-2026-07-29-web-message-icon-actions-and-clock.md: e79662056792c3ab413468ad038dec40455be767
-2026-07-29-web-message-icon-actions-and-clock.zh.md: 72d3b4e0cda19438f2f46fd402b3b76de3726ae5
+2026-07-29-web-message-icon-actions-and-clock.md: f43f7f9c9687e4494993d7e225d11cf6446a9954
+2026-07-29-web-message-icon-actions-and-clock.zh.md: 866fae79f6ad3ea2cb80e5443d2cf5f763562d29

.agents/notes/implemented/feature/2026-07-29-web-message-icon-actions-and-clock.md

@@ -10,9 +10,9 @@ The web chat user bubble already had copy / branch / edit IconActions but no clo
 
 ## Decision
 
-**User bubbles prepend a date-aware local clock to the existing IconActions row; finalized assistant *content* nodes (non-empty text blocks) append a copy / branch / clock row with `margin-top: 16px`; both seats stay visible whenever mounted and re-format at the next local midnight.**
+**User bubbles prepend a date-aware local clock to the existing IconActions row; the last content-text assistant of each turn appends a copy / branch / clock row with `margin-top: 16px`; both seats stay visible whenever mounted and re-format at the next local midnight.**
 
-Both seats format `node.time` through `formatMessageClock`: same calendar day → `HH:mm`, earlier this year → `M月D日 HH:mm`, other years → `YYYY年M月D日 HH:mm`. `useCalendarDay` is a component-local day tick (timeout to the next local midnight) so memoized rows re-render when the calendar day changes without a new framework hook. `MessageItem` places the label before copy (figma `388:20051`). `AssistantMarkdown` places it after branch (figma `43:32997`) only when `streaming` is false, the event time is known, and the node has non-empty text content; Think-only nodes and the streaming tail omit the row. Copy writes joined text blocks. Branch stays a chrome stub. Clipboard write and the clock helpers live in `message-chrome.ts`. The assembled surface is pinned by `apps/web/tests/message-actions.e2e.ts` (cold-seeded history + aria golden); aria normalization collapses every clock shape to `{{clock}}`.
+Both seats format `node.time` through `formatMessageClock`: same calendar day → `HH:mm`, earlier this year → `M月D日 HH:mm`, other years → `YYYY年M月D日 HH:mm`. `useCalendarDay` is a component-local day tick (timeout to the next local midnight) so memoized rows re-render when the calendar day changes without a new framework hook. `MessageItem` places the label before copy (figma `388:20051`). `ChatView` derives turn-tail seqs via `assistantActionsSeqs` and withholds `time` for mid-turn content; `AssistantMarkdown` places the row after branch (figma `43:32997`) only when `streaming` is false, the event time is known, and the node has non-empty text content. Think-only nodes, mid-turn narration, and the streaming tail omit the row. Copy writes joined text blocks. Both message rows pass their event's `seq` to the same fork callback; [Web session fork actions](2026-07-27-web-session-fork-actions.md) define the real mutation contract. Clipboard write and the clock helpers live in `message-chrome.ts`. The assembled surface is pinned by `apps/web/tests/message-actions.e2e.ts` (cold-seeded history + aria golden); aria normalization collapses every clock shape to `{{clock}}`.
 
 ## Alternatives considered
 
@@ -20,12 +20,14 @@ Both seats format `node.time` through `formatMessageClock`: same calendar day
 
 **Put IconActions under every finalized assistant node (including Think-only).** Rejected: copy has nothing useful to write without text content, and repeating the chrome under every step/Think row clutters the flow; only content output owns the seat.
 
+**Put IconActions under every content-text assistant in a multi-step turn.** Rejected: mid-turn narration (text before tools) is not the settled answer; repeating copy/branch/clock under each step clutters the flow. Only the last content assistant of the turn owns the seat.
+
 **Hover-reveal the action row on hover-capable pointers.** Rejected: once the row exists it should stay discoverable; opacity hiding made the chrome easy to miss and required parent hover selectors that duplicated the mount gate.
 
-**Wire branch to a real session fork.** Rejected for this change: same rationale as the archived [user IconActions note](../../archived/feature/2026-07-27-user-message-icon-actions.md) — the mutation path is unspecified; the button reserves the design seat.
+**Let the IconActions decision also define session fork semantics.** Rejected: this note owns only message chrome, clocks, and mount gating; boundary selection, failure behavior, and switching semantics belong to the separate [Web session fork actions](2026-07-27-web-session-fork-actions.md), keeping presentation components from becoming a second home for session mutation.
 
 **Publish the calendar day through a chat store or inject hook.** Rejected: the day tick is presentation-only local state with no cross-entry consumers; a component-local timeout matches the client rule that behavioral hooks may own state that does not subscribe to an external source.
 
 ## Consequences
 
-Settled assistant content answers expose copy and the event clock as soon as the row mounts; Think-only nodes stay chrome-free; branch stays a stub. User and assistant clocks share the same day/year widening rules and refresh after midnight without a message mutation. Per-message paging remains a deferred footer seat in the package README. Package tests pin the three clock shapes, the midnight widen, and the content-only assistant gate; the web e2e scenario pins the assembled IconActions chrome.
+Each turn's last settled content answer exposes copy, branch, and the event clock as soon as the row mounts; mid-turn content and Think-only nodes stay chrome-free. User and assistant clocks share the same day/year widening rules and refresh after midnight without a message mutation. Per-message paging remains a deferred footer seat in the package README. Package tests pin the three clock shapes, the midnight widen, the content-only assistant gate, the turn-tail seq gate, and the respective event `seq` values passed by the user and assistant branch buttons; the web e2e scenario pins the assembled IconActions chrome.

.agents/notes/implemented/feature/2026-07-29-web-message-icon-actions-and-clock.zh.md

@@ -10,9 +10,9 @@ Web 聊天的用户气泡已有复制／分支／编辑 IconActions，但没有
 
 ## 决策
 
-**用户气泡在既有 IconActions 行前追加感知日期的本地时钟；已定稿的 assistant *内容*节点（非空 text 块）在正文下追加带 `margin-top: 16px` 的复制／分支／时钟；两边只要挂载就保持可见，并在下一个本地午夜重新格式化。**
+**用户气泡在既有 IconActions 行前追加感知日期的本地时钟；每个轮次中最后一条带 text 内容的 assistant 在正文下追加带 `margin-top: 16px` 的复制／分支／时钟；两边只要挂载就保持可见，并在下一个本地午夜重新格式化。**
 
-两边都通过 `formatMessageClock` 格式化 `node.time`：同一日历日 → `HH:mm`，同年更早 → `M月D日 HH:mm`，跨年 → `YYYY年M月D日 HH:mm`。`useCalendarDay` 是组件本地的日刻度（定时到下一个本地午夜），因此 memo 行在日历日变化时会重渲染，且不新增框架 hook。`MessageItem` 把标签放在复制之前（figma `388:20051`）。`AssistantMarkdown` 把它放在分支之后（figma `43:32997`），且仅在 `streaming` 为 false、已知事件时间、且节点含非空 text 内容时渲染；纯 Think 节点与流式尾部省略该行。复制写入拼接后的 text 块。分支仍是 chrome stub。剪贴板写入与时钟辅助函数放在 `message-chrome.ts`。组装面由 `apps/web/tests/message-actions.e2e.ts`（冷 seed 历史 + aria golden）钉住；aria 归一化把每种时钟形态折叠为 `{{clock}}`。
+两边都通过 `formatMessageClock` 格式化 `node.time`：同一日历日 → `HH:mm`，同年更早 → `M月D日 HH:mm`，跨年 → `YYYY年M月D日 HH:mm`。`useCalendarDay` 是组件本地的日刻度（定时到下一个本地午夜），因此 memo 行在日历日变化时会重渲染，且不新增框架 hook。`MessageItem` 把标签放在复制之前（figma `388:20051`）。`ChatView` 通过 `assistantActionsSeqs` 推导轮次尾部的 seq，并不为轮次中间的内容传入 `time`；`AssistantMarkdown` 把该行放在分支之后（figma `43:32997`），且仅在 `streaming` 为 false、已知事件时间、且节点含非空 text 内容时渲染。纯 Think 节点、轮次中间的叙述与流式尾部省略该行。复制写入拼接后的 text 块。两种消息行都把自己的事件 `seq` 交给同一个 fork 回调；真实 mutation 契约由 [Web session fork 操作](2026-07-27-web-session-fork-actions.md)定义。剪贴板写入与时钟辅助函数放在 `message-chrome.ts`。组装面由 `apps/web/tests/message-actions.e2e.ts`（冷 seed 历史 + aria golden）钉住；aria 归一化把每种时钟形态折叠为 `{{clock}}`。
 
 ## 曾考虑的方案
 
@@ -20,12 +20,14 @@ Web 聊天的用户气泡已有复制／分支／编辑 IconActions，但没有
 
 **给每个已定稿 assistant 节点（含纯 Think）都挂 IconActions。** 否决：没有 text 内容时复制没有可写内容，且在每一步／Think 下重复 chrome 会打乱流程；只有内容输出拥有该座位。
 
+**给多步骤轮次中的每一条带 text 内容的 assistant 都挂 IconActions。** 否决：轮次中间的叙述（工具调用前的 text）不是已定稿答案；在每一步下重复复制／分支／时钟会打乱流程。只有该轮次中最后一条内容 assistant 拥有该座位。
+
 **在具备 hover 能力的指针上用 hover 才揭示操作行。** 否决：行一旦存在就应保持可发现；用 opacity 隐藏容易漏看，且需要父级 hover 选择器重复挂载门控。
 
-**把分支接到真实的会话 fork。** 本次否决：与已归档的[用户 IconActions 笔记](../../archived/feature/2026-07-27-user-message-icon-actions.md)同一理由——变更路径尚未规定；按钮只预留设计座位。
+**由 IconActions 决策同时定义 session fork 语义。** 否决：本笔记只拥有消息 chrome、时钟与挂载门控；边界选择、失败行为和切换语义属于独立的 [Web session fork 操作](2026-07-27-web-session-fork-actions.md)，避免展示组件成为 session mutation 的第二正家。
 
 **通过 chat store 或 inject hook 发布日历日。** 否决：日刻度只是展示层本地状态，没有跨入口消费者；组件本地 timeout 符合「行为 hook 可拥有不订阅外部源的状态」这一客户端规则。
 
 ## 后果
 
-已定稿的 assistant 内容回答在行挂载后立刻暴露复制与事件时钟；纯 Think 节点不带 chrome；分支仍为 stub。用户与 assistant 时钟共用同一套跨天／跨年加宽规则，并在午夜后无需消息变更即可刷新。逐消息分页仍是包 README 中的暂缓 footer 座位。包级测试钉住三种时钟形态、午夜加宽与 assistant 仅内容门控；Web e2e 场景钉住组装后的 IconActions chrome。
+每个轮次中最后一条已定稿的内容回答在行挂载后立刻暴露复制、分支与事件时钟；轮次中间的内容与纯 Think 节点不带 chrome。用户与 assistant 时钟共用同一套跨天／跨年加宽规则，并在午夜后无需消息变更即可刷新。逐消息分页仍是包 README 中的暂缓 footer 座位。包级测试钉住三种时钟形态、午夜加宽、assistant 仅内容门控、轮次尾部 seq 门控，以及 user/assistant 分支按钮各自传递的事件 `seq`；Web e2e 场景钉住组装后的 IconActions chrome。

.agents/notes/implemented/feature/2026-07-30-deepseek-onboarding-credential-setup.i18n.yaml

@@ -0,0 +1,6 @@
+# Bilingual-pair consistency record (docs/i18n/README.md): the git blob hash of each
+# side as of the last confirmed-consistent state. Both languages carry equal authority;
+# after editing either side, bring the other along and re-record with:
+#   pnpm run verify-translation-pairing --write .agents/notes/implemented/feature/2026-07-30-deepseek-onboarding-credential-setup.md
+2026-07-30-deepseek-onboarding-credential-setup.md: ed53ffe64d3ba27e8746d58ad84d1a4c401f6ce4
+2026-07-30-deepseek-onboarding-credential-setup.zh.md: 340728ab9348e0132954e403f9bdbf561e340247

.agents/notes/implemented/feature/2026-07-30-deepseek-onboarding-credential-setup.md

@@ -0,0 +1,33 @@
+# Agent Note: official DeepSeek first-run credential setup
+
+Status: implemented
+
+English | [中文](2026-07-30-deepseek-onboarding-credential-setup.zh.md)
+
+## Problem
+
+The [web configuration plane](../architecture/2026-07-30-web-config-plane.md) makes provider settings and credentials live-editable, but a first-time user still lands on the empty conversation Hero without an actionable explanation when the shipped `deepseek-official` route has no credential. The Models page can repair that state, yet requiring the user to discover it weakens onboarding. A prompt must not confuse a missing credential with a missing adapter: the browser can store a value for an existing credential reference, but it cannot dynamically mount the `llm-deepseek` Cordis plugin.
+
+## Decision
+
+**One readiness projection owns both Models and onboarding facts.** `ui-models` keeps a single store that joins `llm.providers({})`, redacted `settings.describe({})`, and batched `credentials.describe({refs})`. The onboarding projection selects the `deepseek-official` configurable-provider entry owned by the `llm-deepseek` namespace and empty settings path, reads the effective `apiKeyEnv`, and evaluates the matching credential descriptor. A live route with the same provider id but no matching configurable-provider declaration is adapter-absent for onboarding. A configured literal `apiKey` secret sidecar is also ready, so compatibility configuration does not trigger a false prompt; a configured process-environment credential is ready and remains read-only.
+
+**The settings shell contributes ordering and navigation, not provider policy.** `ui-settings` declares a root-scoped `settings.onboarding` list slot and mounts one ordered step at a time while the current surface is the empty Hero. The active registrant receives `complete()` and a private `openSection(id)` callback; completion transfers ownership to the next entry. `ui-models` registers the DeepSeek step through the same declaration-aware deferred-registration path as its Models section, so plugin load order does not become a contract and independently contributed dialogs cannot stack. The product-wide welcome step that precedes it is owned separately by [the versioned welcome decision](2026-07-30-versioned-gui-welcome-onboarding.md).
+
+**The prompt routes to the one credential editor.** A mounted, active adapter with a resolved, writable, unconfigured reference presents one action that opens Settings on Models. The existing DeepSeek setup card there exclusively owns the password input, `credentials.set({ref, value})`, write failures, and post-write refresh; the onboarding overlay never holds or submits a secret. An unavailable settings or credential capability keeps its deployment diagnostic and routes to the same page, while an absent adapter remains skipped because navigation cannot mount a Cordis plugin.
+
+**Unavailable states do not capture the product.** An absent configurable-provider entry, inactive route, failed initial join, read-only deployment, or unresolved settings or credential capability completes the step without rendering because the onboarding action cannot repair that state. The Models page remains the deployment diagnostic and retry surface. Configure later completes a missing-credential step for the current mounted coordinator pass and writes no completion fact. Settings, credential, provider-topology, and connection invalidations all refresh the shared join, so an external credential update completes an open step without a reload.
+
+## Alternatives considered
+
+**A separate onboarding store and readiness RPC sequence** — rejected because it would create a second client-side interpretation of provider identity, settings paths, secret sidecars, credential references, and invalidation ordering beside the Models page.
+
+**A second API-key editor inside onboarding** — rejected because the Models page already renders its DeepSeek setup card for exactly this state. Duplicating its secret draft, write errors, and configured-state convergence would add a second security-sensitive UI without another user capability.
+
+**Writing the API key into provider settings** — rejected because a literal secret would enter the settings mutation path and whole-section replacement cannot safely reconstruct redacted values. Credential storage is already the product seam and supplies immediate invalidation.
+
+**Showing the prompt when `llm-deepseek` is absent** — rejected because browser navigation has no supported operation that mounts the missing Cordis plugin.
+
+## Consequences
+
+The ordered flow leads from the product notice to the shipped adapter's existing editor without restarting: a keyless browser test boots the real Web composition under an isolated harness home, acknowledges the notice, follows the DeepSeek page to Models, stores a generated key through that page into the home's `.env`, verifies no key reaches DOM, ARIA, or browser console output, and confirms the running page reports configured. The full keyless Web replay lane also pins that a non-configurable replay route with the same provider id does not block unrelated journeys. Pure readiness and React tests pin literal, file, process-environment, missing-provider, missing-capability, navigation, cancellation, external-invalidation, and coordinator-transfer behavior. The flow deliberately inherits the configuration plane's documented base limitations rather than adding local secret storage, redaction, or settings replacement workarounds.

.agents/notes/implemented/feature/2026-07-30-deepseek-onboarding-credential-setup.zh.md

@@ -0,0 +1,33 @@
+# Agent Note: DeepSeek 官方首次使用凭据配置
+
+Status: implemented
+
+[English](2026-07-30-deepseek-onboarding-credential-setup.md) | 中文
+
+## 问题
+
+[web 配置平面](../architecture/2026-07-30-web-config-plane.md)让提供方设置与凭据可以实时编辑，但首次使用的用户仍会进入空白对话 Hero；当随产品提供的 `deepseek-official` 路由缺少凭据时，界面没有给出可采取操作的说明。Models 页能修复该状态，但要求用户自行发现这个入口会削弱首次使用引导。界面不得混淆凭据缺失与适配器缺失：浏览器可以为现有凭据引用存入值，但无法动态挂载 `llm-deepseek` Cordis 插件。
+
+## 决策
+
+**Models 与首次使用引导共享同一个就绪状态投影。**`ui-models` 维护一个 store，把 `llm.providers({})`、脱敏后的 `settings.describe({})` 和批量调用的 `credentials.describe({refs})` 联接为同一份状态。首次使用投影选取由 `llm-deepseek` namespace 与空 settings path 持有的 `deepseek-official` 可配置提供方条目，读取生效的 `apiKeyEnv`，并检查对应的凭据描述符。同 provider id 但没有匹配可配置提供方声明的存活路由，在首次使用引导中视为适配器缺失。若 `apiKey` 字面量对应的 secret 槽位标记为已设置，也会判定为就绪，兼容配置因此不会误触发页面；通过进程环境提供的凭据若已配置，同样判定为就绪并保持只读。
+
+**设置外壳只贡献排序与导航，不持有提供方策略。** `ui-settings` 声明一个根作用域的 `settings.onboarding` list slot，并在当前界面为空白 Hero 时，每次只挂载一个有序步骤。当前注册方会收到 `complete()` 和私有 `openSection(id)` 回调；完成当前步骤后，所有权转交给下一项。`ui-models` 沿用 Models 分区所使用、感知 slot 声明的延迟注册路径来注册 DeepSeek 步骤，因此插件加载顺序不会成为契约，独立贡献的对话框也无法堆叠。排在它之前的产品级欢迎步骤由[版本化欢迎决策](2026-07-30-versioned-gui-welcome-onboarding.md)单独持有。
+
+**首次使用页面只负责跳转到唯一的凭据编辑器。**适配器已挂载且处于活跃状态，其引用可解析、可写但尚未配置时，界面会显示一个操作按钮，用于打开「设置」的 Models 分区。该分区已有的 DeepSeek 设置卡片全权负责密码输入框、`credentials.set({ref, value})`、写入失败处理和写入后刷新；首次使用页面绝不持有或提交 secret。
+
+**不可用状态不会占住产品。** 可配置提供方条目缺失、路由不活跃、初始联接失败、部署只读或设置／凭据能力无法解析时，都会直接完成而不渲染该步骤，因为首次使用引导无法修复这些状态。Models 页仍是部署诊断与重试界面。「稍后配置」只会完成协调器当前这一次缺少凭据的步骤，不写入任何完成状态。设置、凭据、提供方拓扑和连接失效事件都会刷新共享联接，因此外部凭据更新无需重新加载页面即可完成已打开的步骤。
+
+## 曾考虑的替代方案
+
+**为首次使用引导单设 store 与就绪状态 RPC 调用序列**：不予采用，因为这会在 Models 页之外，再建立一套客户端解释，用于判定提供方身份、设置路径、secret 槽位的伴随信息、凭据引用及失效事件顺序。
+
+**在首次使用引导中增设第二个 API key 编辑器**：不予采用，因为 Models 页已为这一状态渲染 DeepSeek 设置卡片。复制其中的 secret 草稿、写入错误处理和已配置状态收敛会增加第二个安全敏感的 UI，却不会带来新的用户能力。
+
+**把 API key 写入提供方设置**：不予采用，因为字面量 secret 会进入设置变更路径，而整个分节替换无法安全重建脱敏值。凭据存储已经是产品 seam，并能立即发出失效事件。
+
+**`llm-deepseek` 缺失时仍显示浮层**：不予采用，因为浏览器导航没有任何受支持的操作可以挂载缺失的 Cordis 插件。
+
+## 后果
+
+有序流程从产品声明页开始，无需重启即可引导用户前往随产品提供的适配器已有的编辑器：无密钥浏览器测试在隔离的 harness 家目录下启动真实 Web 组合，确认声明后依照 DeepSeek 页面前往 Models，通过该页面把生成的密钥存入该目录的 `.env`，验证密钥未进入 DOM、ARIA 或浏览器控制台输出，并确认运行中的页面报告已配置。完整的无密钥 Web 回放也固定了同 id 的不可配置回放路由不会阻塞无关流程。纯就绪状态测试与 React 测试固化了字面量凭据、文件凭据、进程环境凭据、提供方缺失、能力缺失、导航、取消、外部失效和协调器移交行为。该流程直接继承配置平面已记录的基础限制，不会另加局部的机密存储、脱敏或设置替换变通方案。

.agents/notes/implemented/feature/2026-07-30-plan-review-presentation-intent.i18n.yaml

@@ -0,0 +1,6 @@
+# Bilingual-pair consistency record (docs/i18n/README.md): the git blob hash of each
+# side as of the last confirmed-consistent state. Both languages carry equal authority;
+# after editing either side, bring the other along and re-record with:
+#   pnpm run verify-translation-pairing --write .agents/notes/implemented/feature/2026-07-30-plan-review-presentation-intent.md
+2026-07-30-plan-review-presentation-intent.md: aeab12aac308c24aa5c4b5953a60ae0f2cf6c5b5
+2026-07-30-plan-review-presentation-intent.zh.md: 4096018e374212c821675ce6ed3a2355df20edc9

.agents/notes/implemented/feature/2026-07-30-plan-review-presentation-intent.md

@@ -0,0 +1,55 @@
+# Agent Note: Plan review as a decision, not a question
+
+Status: implemented
+
+English | [中文](2026-07-30-plan-review-presentation-intent.zh.md)
+
+## Problem
+
+`exit_plan_mode` presents a finished plan for review through `ctx.userInteraction.ask()`, the same seam `ask_user_question` uses. On the Web GUI that made a plan review render as the generic question flow of [the ask-question Web presentation](2026-07-29-ask-question-web-presentation.md): a `1 / 1` pager, the plan as a question's supporting detail, the two verdicts as numbered radio rows with descriptions, an "Other — enter a custom answer" row, and `Skip this question` / `Submit` in the footer.
+
+Every one of those affordances is wrong for the surface. Reviewing a plan is one decision over one document, and the quiz chrome told the user they were being examined rather than asked to approve work — reported as "让人很困惑以为在做题". The paging controls page a set of one. Skipping is not an outcome the tool accepts (it folds into keep-planning). Worst, the surface gave no hint that this was the plan gate at all, while the adjacent waiting-approval takeover already had exactly the right shape for a decision: a tinted strip naming what is being decided, the subject as the body, and a right-aligned action row.
+
+## Decision
+
+A question may declare a **presentation intent**, and the Web composer renders a declared intent as its own surface. `AskUserQuestionItem` gains `intent?: AskUserQuestionIntent`, a tagged shape whose one member is `{ kind: 'plan-review', approve: string }`; `plan-mode` sets it on the review question, naming `Approve` as the label that approves.
+
+An intent shapes presentation only. The answer protocol is untouched: a UI honouring the intent answers with the same option labels a generic UI would send, so `exit_plan_mode` reads one answer shape regardless of which surface collected it, and a UI that does not know a tag renders the generic flow with nothing lost but the layout.
+
+`approve` names the affirmative option instead of relying on option order, so no UI infers a verdict from a position. Two assertions an intent makes are beyond the types, and `UserInteractionService.ask()` rejects both as `BAD_INTENT` at the asker: an `approve` naming none of that question's own options — before any UI can answer a choice never offered — and an intent on a question with no `detail`, the thing it declares itself a review of, which would ask the user to approve something invisible. On the wire the intent is a discriminated union, so an unrecognised tag is a rejected frame rather than a silently generic render.
+
+`ui-question` renders the intent as `PlanReviewPanel`, in the waiting-approval card language: the amber strip carries `Plan review`, the plan is the scrolling markdown body, and the decision row holds three actions — `Chat about it`, `Refuse`, `Approve`. The question text becomes the card's accessible name rather than a headline, because the buttons already say what the decision is. Approve and Refuse answer with the asker's own option labels and keep the asker's descriptions as tooltips; `Chat about it` cancels the request, which returns the composer so the user can simply say what they want. All copy is bilingual under the existing `question` namespace.
+
+Routing lives inside the single composer entry (`QuestionComposer` chooses the shape) rather than in a second chain registration, and `planReviewOf` claims a request only when the card can send every answer that request allows: one question declaring the intent, the plan as its `detail`, the named approve label offered, and a binary single choice — at most one option besides approve, and not multi-select. A third option or a multi-select batch has answers two buttons cannot express, so the generic flow keeps it, and keeps anything else the card cannot render. "Presentation only" is therefore literal: an intent never costs the user a reachable answer, and the client — downstream of a wire boundary — leaves every request answerable.
+
+Dismissal became its own model-facing outcome. `ASK_CANCELLED` previously reached the model as "the user cancelled ask_user_question", naming a tool it never called; `exit_plan_mode` now reports that the user dismissed the review to speak instead and to stay in plan mode and wait. Every other ask failure — an abort from turn cancel or provider teardown, where no user is coming — keeps its own message.
+
+## Alternatives considered
+
+**Make plan review its own pending kind (`plan-review/requested`).** Rejected as the wrong size for a presentation problem. It buys an honest response shape (approve / decline / discuss instead of an answer batch) at the cost of a third `PendingKind`, new requested/resolved frames and schemas, an api-proxy registry and respond branch, client session and baseline-replay handling, and a new three-package capability seam for a decision the question protocol already expresses. Worth revisiting only if plan review grows outcomes the answer shape cannot carry.
+
+**Route the card on the question's `id` or `header` (`plan-review` / `Plan review`).** Rejected: string-sniffing a foreign package's copy across a wire boundary, which any wording change silently breaks. The intent is the declaration that makes the routing legible.
+
+**Order the options and let the card read position 0 as approve.** Rejected: a positional contract at a package seam, invisible in both the type and the wire frame, and unenforceable — a producer that reorders its options would invert a user's verdict. Naming the label costs one string.
+
+**Register a second composer-chain entry for the plan card.** Rejected: two entries would select over the same pending question carrier, making the surface depend on chain priority and on whether the plan package's client half is composed at all. One entry that picks its own shape cannot race itself, and the generic flow is the built-in fallback.
+
+**Put the panel in `ui-plan` beside the plan chip.** Rejected: the panel's whole behavior is the question carrier's answer encoding (`PendingQuestion`), which `ui-question` owns; the intent is a question-protocol field, not plan-mode's private channel. Rendering declared intents belongs to the package that owns question rendering, as tool render intents belong to the tool renderer.
+
+**Extract a shared takeover card with `ui-conversation`'s `ApprovalPanel`.** Not done: the two takeovers agree on tokens and geometry but not on content — this body is scrolling markdown, that one a headline plus a command line — and the shared shell would be two elements wide. They are kept in step by token, not by component.
+
+**Give `Chat about it` its own protocol outcome.** Rejected: dismissing a request is a verb the generic flow already has (the `×` that cancels the batch). Promoting it to a labelled button is presentation; inventing a fourth wire outcome for it is not.
+
+## Consequences
+
+The question protocol now carries a presentation axis. Adding a second intent is a tag on the union, a producer that sets it, a schema member, and a panel — no new frame, service, or answer shape. The cost is that the question seam knows presentation exists at all, and that `ui-question` knows the word "plan"; both are the price of one entry owning every question surface.
+
+The plan gate reads as a plan gate: the plan is the card's content, the verdict is two labelled buttons, and taking the turn back is a third. The generic flow is untouched for every other question, and its committed goldens did not move.
+
+A deployment whose client half predates this change still shows the quiz layout — correct, answerable, and merely unstyled — because the intent is additive and the fallback is the generic flow.
+
+## Testing
+
+`ui-question` tests pin the narrowing (single-question batch, intent present, plan as detail, named approve label offered, binary single choice, decline absent when only approve is offered) and the panel (strip, markdown plan, accessible name, absence of pager/radio/skip/custom, approve and decline answering with the asker's labels, dismissal cancelling, one-shot latch with re-arm and message on a rejected receipt, tooltips present and absent, both locales). `user-interaction` tests pin both `BAD_INTENT` rejections and intent pass-through; `plan-mode` tests pin the declared intent against its own option list and both failure messages; the apiproxy schema test pins wire acceptance and an unknown tag's rejection.
+
+The `plan-review` Web e2e lane records `/plan` entering plan mode for real, the model calling `exit_plan_mode`, the decision card taking the composer (asserting the generic flow did **not** claim the request), and the card's own Approve completing the turn — two keyless goldens, the waiting card and the approved transcript.

.agents/notes/implemented/feature/2026-07-30-plan-review-presentation-intent.zh.md

@@ -0,0 +1,55 @@
+# Agent Note：计划审阅是一次决定，不是一道题
+
+Status: implemented
+
+[English](2026-07-30-plan-review-presentation-intent.md) | 中文
+
+## 问题
+
+`exit_plan_mode` 通过 `ctx.userInteraction.ask()` 把写好的计划交给用户审阅，而这正是 `ask_user_question` 使用的同一个 seam。在 Web GUI 上，这导致计划审阅渲染为[ask-question Web 呈现](2026-07-29-ask-question-web-presentation.md)里的通用问题流程：一个 `1 / 1` 分页器、计划作为问题的补充说明、两个裁决作为带描述的编号单选行、一行"其他，请填写自定义答案"，以及底部的 `跳过本题` / `提交`。
+
+这些可交互元素对这个界面而言无一正确。审阅一份计划是对一份文档做一次决定，而做题式的界面告诉用户他正在被考试，而不是被请求批准一份工作 —— 实际反馈是"让人很困惑以为在做题"。分页控件在给只有一项的集合分页。跳过并不是该工具接受的结果（它会折叠成继续规划）。最糟的是，这个界面完全没有暗示这就是计划关口，而旁边的等待审批接管早就具备了一次决定该有的形状：一条带色条带说明正在决定什么、主体是决定的对象、右对齐的操作行。
+
+## 决定
+
+一个问题可以声明**呈现意图（presentation intent）**，Web 输入区把已声明的意图渲染为它自己的界面。`AskUserQuestionItem` 新增 `intent?: AskUserQuestionIntent`，一个带标签的形状，目前唯一成员是 `{ kind: 'plan-review', approve: string }`；`plan-mode` 在审阅问题上设置它，并指明 `Approve` 是表示批准的标签。
+
+意图只塑造呈现。回答协议不变：遵循意图的 UI 回答的仍是通用 UI 会发送的那些选项标签，因此无论由哪个界面收集，`exit_plan_mode` 读到的都是同一种回答形状；而不认识某个标签的 UI 渲染通用流程，除布局之外一无所失。
+
+`approve` 指名肯定选项，而不依赖选项顺序，因此没有任何 UI 会从位置推断裁决。意图作出的两项断言超出类型的表达能力，`UserInteractionService.ask()` 都以 `BAD_INTENT` 在提问方一侧拒绝：`approve` 未命中该问题自身的任一选项 —— 早于任何 UI 回答一个从未被提供过的选择；以及意图落在没有 `detail` 的问题上，而 `detail` 正是它自称在审阅的东西，那会让用户去批准一件看不见的事。在协议格式（wire format）上意图是可辨识联合，因此无法识别的标签是被拒绝的帧，而不是静默退回通用渲染。
+
+`ui-question` 把该意图渲染为 `PlanReviewPanel`，沿用等待审批卡片的语言：琥珀色条带写着 `Plan review`，计划是可滚动的 markdown 主体，决定行放三个操作 —— `Chat about it`、`Refuse`、`Approve`。问题文本成为卡片的无障碍名称而非标题，因为按钮已经说明了这次决定是什么。Approve 与 Refuse 用提问方自己的选项标签回答，并把提问方的描述保留为 tooltip；`Chat about it` 取消该请求，从而让输入区归位，用户直接说他想说的话即可。所有文案在既有 `question` 命名空间下双语。
+
+路由住在单一输入区条目内部（由 `QuestionComposer` 选择形状），而不是第二个链式注册；`planReviewOf` 仅在卡片能够发出该请求允许的每一个答案时才接管：只有一个问题且声明了意图、以 `detail` 承载计划、提供了被指名的批准标签，且是二元单选 —— 除批准外最多一个选项，且非多选。出现第三个选项或多选批次时，其答案是两个按钮无法表达的，通用流程保留它，也保留其他任何卡片渲染不了的请求。因此"只塑造呈现"是字面意义上的：意图绝不让用户失去一个可达的答案，而位于协议边界下游的客户端让每个请求都保持可回答。
+
+放弃审阅成为面向模型的独立结果。`ASK_CANCELLED` 以前传到模型的是"the user cancelled ask_user_question"，指名了一个它从未调用的工具；现在 `exit_plan_mode` 报告用户放弃审阅是为了改用说话，并要求留在 plan mode 中等待。其余每一种 ask 失败 —— 轮次取消或提供方拆卸导致的中止，那里并没有用户会来 —— 保留它们自己的消息。
+
+## 备选方案
+
+**让计划审阅成为自己的待处理种类（`plan-review/requested`）。** 否决：对一个呈现问题来说尺寸不对。它换来的是诚实的响应形状（approve / decline / discuss 而非一批回答），代价是第三个 `PendingKind`、新的 requested/resolved 帧与 schema、一个 api-proxy 注册表与响应分支、客户端会话与基线重放处理，以及为一个问题协议已能表达的决定新增一个三包能力 seam。只有当计划审阅长出回答形状承载不了的结果时才值得重新考虑。
+
+**按问题的 `id` 或 `header`（`plan-review` / `Plan review`）路由卡片。** 否决：这是跨协议边界嗅探另一个包的文案字符串，任何措辞改动都会静默破坏它。意图才是让路由可读的那个声明。
+
+**约定选项顺序，让卡片把第 0 个位置读作批准。** 否决：这是包边界上的位置约定，在类型和协议帧里都看不见，也无法强制 —— 生产方一旦重排选项，就会颠倒用户的裁决。指名标签只花一个字符串。
+
+**为计划卡片注册第二个输入区链条目。** 否决：两个条目会对同一个待回答问题载体做选择，使界面取决于链优先级、以及计划包的客户端半边是否被组合。一个自己挑形状的条目不会和自己抢，而通用流程正是内建的回退。
+
+**把面板放在 `ui-plan` 里、紧挨计划状态标签。** 否决：面板的全部行为就是问题载体的回答编码（`PendingQuestion`），那是 `ui-question` 拥有的；意图是问题协议的字段，不是 plan-mode 的私有通道。渲染已声明的意图属于拥有问题渲染的那个包，正如工具渲染意图属于工具渲染方。
+
+**与 `ui-conversation` 的 `ApprovalPanel` 抽出共享的接管卡片。** 未做：两个接管在 token 和几何上一致，但内容不一致 —— 这边的主体是可滚动 markdown，那边是一行标题加一行命令 —— 共享外壳只会剩两个元素宽。它们靠 token 保持一致，而不是靠组件。
+
+**给 `Chat about it` 自己的协议结果。** 否决：放弃一个请求是通用流程已有的动词（取消整批的 `×`）。把它提升为带标签的按钮属于呈现；为它发明第四种协议结果不属于。
+
+## 结果
+
+问题协议从此带有一个呈现轴。新增第二个意图 = 联合上的一个标签、一个设置它的生产方、一个 schema 成员、一个面板 —— 不需要新的帧、服务或回答形状。代价是问题 seam 从此知道"呈现"这件事存在，且 `ui-question` 知道"plan"这个词；两者都是由单一条目拥有全部问题界面所要付的价钱。
+
+计划关口读起来就像计划关口：计划是卡片的内容，裁决是两个带标签的按钮，把轮次拿回来是第三个。通用流程对其他每个问题都未受影响，其已提交的 golden 也没有变动。
+
+客户端半边早于本次改动的部署仍然显示做题式布局 —— 正确、可回答、只是没有专门样式 —— 因为意图是增量的，而回退就是通用流程。
+
+## 测试
+
+`ui-question` 测试钉住收窄（单问题批、意图存在、计划作为 detail、被指名的批准标签确实被提供、二元单选、只提供批准时 decline 缺席）与面板（条带、markdown 计划、无障碍名称、无分页/单选/跳过/自定义、批准与拒绝用提问方的标签回答、放弃触发取消、一次性闭锁在回执被拒时重新武装并给出消息、tooltip 有与无、两种语言）。`user-interaction` 测试钉住两种 `BAD_INTENT` 拒绝与意图透传；`plan-mode` 测试钉住已声明的意图与其自身选项列表的一致、以及两条失败消息；apiproxy schema 测试钉住协议接受与未知标签的拒绝。
+
+`plan-review` Web e2e 通道录制了 `/plan` 真实进入 plan mode、模型调用 `exit_plan_mode`、决定卡片接管输入区（并断言通用流程**没有**接管该请求）、以及卡片自身的 Approve 完成该轮 —— 两份无密钥 golden：等待中的卡片与批准后的会话记录。

.agents/notes/implemented/feature/2026-07-30-versioned-gui-welcome-onboarding.i18n.yaml

@@ -0,0 +1,6 @@
+# Bilingual-pair consistency record (docs/i18n/README.md): the git blob hash of each
+# side as of the last confirmed-consistent state. Both languages carry equal authority;
+# after editing either side, bring the other along and re-record with:
+#   pnpm run verify-translation-pairing --write .agents/notes/implemented/feature/2026-07-30-versioned-gui-welcome-onboarding.md
+2026-07-30-versioned-gui-welcome-onboarding.md: 0705469e02ddb9068722ae5d500c151f077c83fd
+2026-07-30-versioned-gui-welcome-onboarding.zh.md: bdd21d635f824b8c4a4813e6bff7798b34ec9677

.agents/notes/implemented/feature/2026-07-30-versioned-gui-welcome-onboarding.md

@@ -0,0 +1,35 @@
+# Agent Note: Versioned GUI welcome onboarding
+
+Status: implemented
+
+English | [中文](2026-07-30-versioned-gui-welcome-onboarding.zh.md)
+
+## Problem
+
+The GUI's credential onboarding begins with a DeepSeek-specific readiness check, but the internal-test notice applies to every user and must precede provider setup even when a credential is already configured. Treating both as independent overlays permits simultaneous dialogs, while a process-local dismissal cannot distinguish a completed notice from a window closed before acknowledgement or intentionally present revised copy once.
+
+## Decision
+
+**The Settings shell coordinates ordered steps.** `settings.onboarding` remains a root-scoped list, but `ui-settings` projects its entry ids and order into one coordinator and mounts only the first incomplete step. The active registrant receives `complete()` and `openSection(id)`; no later step mounts until ownership transfers. The product welcome registers at order `-100`, while `ui-models` retains only the conditional DeepSeek readiness and credential-routing step at order `0`.
+
+**Ownerless product onboarding belongs to `ui-settings-general`.** `src/onboarding-copy.ts` is the single editable source for the complete Chinese notice, its faithful English counterpart, the Continue labels, and `WELCOME_NOTICE_VERSION`. Runtime locale dictionaries derive their welcome values from that file, and tests import the same owner instead of repeating paragraph text. The notice is browser UI only: it creates no Session event and contributes no model-visible content.
+
+**Acknowledgement is durable per Harness profile.** The Host half registers a `ui-onboarding` section in the user-settings seam, stored under the active `$DSH_HOME/settings.yaml`. The browser shows the notice unless `welcomeNoticeVersion` equals the owner constant exactly. Continue applies one path mutation with the current version and calls `complete()` only after the Host commits it; a failed write leaves the notice open, and closing the page or process writes nothing. Bumping the constant intentionally makes every profile acknowledge the revised copy once.
+
+**Concurrent views converge without stale replacement.** The acknowledgement write omits `expectedRevision` deliberately: every tab writes the same version to one path, so the operation is idempotent and preserves sibling fields instead of rebuilding the section. `settings/document-updated` becomes `host/settings-changed`; an already mounted tab refetches and advances when another tab or an external editor commits the current version. The API proxy exposes this one product namespace through a closed allowlist beside configurable-provider namespaces, without treating its changes as model-catalog invalidations.
+
+**Onboarding temporarily owns the viewport as one continuous stage.** A solid product surface replaces the complete application view through a body-level portal and marks the underlying app root inert; the exact required mask remains mounted behind that surface with `position:absolute`, zero left/right/bottom offsets, `top:80px`, `rgba(0, 0, 0, 0.24)`, and `backdrop-filter: blur(2px)`. Welcome and conditional credential setup render as successive pages in this stage instead of independent modals. Both pages reuse the Web UI's black `BrandWordmark`. The welcome page preserves the four authored paragraphs verbatim under the `内测声明` title; every paragraph uses one 16/28 body scale, and only the requested action clause inside the final paragraph receives a subtle 500 weight. A short staggered opacity/vertical entrance supplies pacing without blocking interaction and disappears under reduced motion. The title receives initial focus, Continue is the sole button, and no close, Escape, or mask-click path exists.
+
+## Alternatives considered
+
+**Browser local storage** — rejected because acknowledgement would follow one browser profile rather than `$DSH_HOME`; a fresh Harness profile could incorrectly inherit a prior acknowledgement, and external profile edits would have no authoritative update stream.
+
+**A second independent modal in `ui-settings-general`** — rejected because list registrants would still stack whenever welcome and credential readiness were both true. Ordered ownership belongs to the shell that declares and renders the list.
+
+**Persisting on render or window close** — rejected because observation is not acknowledgement and close delivery is unreliable. Only the explicit Continue commit may suppress the next launch.
+
+**A generic public settings-exposure flag** — rejected because one product namespace does not justify widening every settings registrant's public configuration surface. The gateway keeps an explicit closed allowlist.
+
+## Consequences
+
+A fresh profile always sees the welcome notice before provider-specific onboarding; an already configured credential skips only the later DeepSeek step. Reloading after Continue stays past the acknowledged version, changing the owner version presents it again, and closing before Continue leaves the next launch unchanged. Focused store and React tests pin exact-version comparison, write failure, sole-action behavior, no-dismiss paths, coordinator ordering, conditional DeepSeek transfer, and HMR cleanup. The real Chromium scenario boots the shipped Web composition with an isolated harness home, verifies the exact mask geometry and computed styles, reloads before and after acknowledgement, continues into missing-credential setup, confirms an acknowledged-version mismatch returns while the credential is configured, and checks the browser console.

.agents/notes/implemented/feature/2026-07-30-versioned-gui-welcome-onboarding.zh.md

@@ -0,0 +1,35 @@
+# Agent Note: 版本化 GUI 欢迎引导
+
+Status: implemented
+
+[English](2026-07-30-versioned-gui-welcome-onboarding.md) | 中文
+
+## 问题
+
+GUI 的凭据引导从 DeepSeek 专用的就绪状态检查开始，但内部测试通知适用于每位用户，即使凭据已经配置，也必须先于提供方设置显示。若把两者作为独立浮层处理，多个对话框可能同时出现；仅存于进程内的关闭标记既无法区分通知已完成确认还是窗口在确认前已关闭，也无法在文案有意修订后重新显示一次通知。
+
+## 决策
+
+**设置外壳协调有序步骤。** `settings.onboarding` 仍是根作用域 list，但 `ui-settings` 会把其中各条目的 id 和顺序投影到一个协调器中，并且只挂载第一个未完成的步骤。当前注册方会收到 `complete()` 和 `openSection(id)`；所有权转移前，不会挂载后续步骤。产品欢迎步骤的顺序为 `-100`，`ui-models` 则只保留顺序为 `0` 的 DeepSeek 条件式就绪状态与凭据跳转步骤。
+
+**不属于单一功能的产品引导由 `ui-settings-general` 持有。** `src/onboarding-copy.ts` 是完整中文通知、忠实英文对侧文案、两种语言的「继续」按钮文案和 `WELCOME_NOTICE_VERSION` 的唯一可编辑来源。运行时 locale 字典从该文件派生欢迎文案，测试也导入同一个所有者，而不重复段落文本。该通知只存在于浏览器 UI：它不会创建会话事件，也不会贡献任何模型可见内容。
+
+**确认状态按 Harness profile 持久化。** 宿主端在 user-settings seam 中注册 `ui-onboarding` 分节，并存入当前 `$DSH_HOME/settings.yaml`。除非 `welcomeNoticeVersion` 与文案所有者文件中的常量精确相等，否则浏览器会显示通知。「继续」会以当前版本执行一次路径变更，并且仅在宿主端提交成功后调用 `complete()`；写入失败时通知保持打开，关闭页面或进程则不会写入任何内容。提升该常量会有意要求每个 profile 对修订后的文案重新确认一次。
+
+**并发视图无需陈旧的整体替换即可收敛。** 确认写入有意省略 `expectedRevision`：每个标签页都向同一路径写入相同版本，因此该操作是幂等的，并会保留同级字段，而不是重建整个分节。`settings/document-updated` 会转为 `host/settings-changed`；另一个标签页或外部编辑器提交当前版本后，已挂载的标签页会重新拉取状态并推进。API 网关在可配置提供方 namespace 之外，通过封闭的允许列表暴露这一个产品 namespace，同时不会把它的变更视为模型目录失效事件。
+
+**引导流程会暂时接管视口，形成一个连续阶段。** 纯色产品界面通过挂载到 `body` 的 portal 取代完整的应用视图，并将底层应用根节点标记为 inert；严格符合要求的遮罩仍挂载在该界面后方，并保留 `position:absolute`、left/right/bottom 偏移量为零、`top:80px`、`rgba(0, 0, 0, 0.24)` 和 `backdrop-filter: blur(2px)`。欢迎页和按条件显示的凭据设置页在这一阶段中依次呈现，而不是各自作为独立的模态窗口。两个页面都复用 Web UI 的黑色 `BrandWordmark`。欢迎页在 `内测声明` 标题下逐字保留既定的四段文案；所有段落统一采用 16/28 的正文字号与行高，只有最后一段中指定的行动语句使用较为克制的 500 字重。短暂的错落式透明度与纵向位移动画营造出舒缓节奏，但不会阻碍交互，并会在用户启用减少动态效果时禁用。初始焦点落在标题上，「继续」是唯一按钮，且不存在关闭、Escape 或点击遮罩的退出路径。
+
+## 曾考虑的替代方案
+
+**浏览器本地存储**：不予采用，因为确认状态会跟随某个浏览器 profile，而不是 `$DSH_HOME`；全新的 Harness profile 可能错误继承此前的确认状态，外部 profile 编辑也没有权威更新流。
+
+**在 `ui-settings-general` 中再增加一个独立模态窗口**：不予采用，因为欢迎通知和凭据就绪状态同时为真时，list 注册方仍会堆叠。声明并渲染该 list 的外壳应当持有有序所有权。
+
+**在渲染或窗口关闭时持久化**：不予采用，因为看见通知不等于确认，窗口关闭事件也无法可靠送达。只有显式提交「继续」才能阻止通知在下次启动时再次显示。
+
+**通用的公开设置暴露标志**：不予采用，因为一个产品 namespace 不足以证明应当扩大每个 settings 注册方的公开配置面。网关保留显式的封闭允许列表。
+
+## 后果
+
+全新 profile 始终会在提供方专用引导之前看到欢迎通知；凭据已经配置时，只会跳过后续 DeepSeek 步骤。点击「继续」后重新加载不会再次显示已确认版本，更改文案所有者文件中的版本值会让通知重新出现，而确认前关闭窗口不会改变下次启动。针对性的 store 与 React 测试固化了精确版本比较、写入失败、单一操作、不可关闭路径、协调器顺序、按条件移交 DeepSeek 步骤和 HMR（热模块替换）清理行为。真实 Chromium 场景会使用隔离的 harness 家目录启动随产品提供的 Web 组合，验证遮罩的精确几何尺寸和计算样式，在确认前后分别重新加载，继续进入凭据缺失设置流程，确认凭据已配置时确认版本不匹配仍会使通知重新出现，并检查浏览器控制台。

.agents/notes/implemented/feature/2026-07-30-web-context-injection-disclosure.i18n.yaml

@@ -0,0 +1,6 @@
+# Bilingual-pair consistency record (docs/i18n/README.md): the git blob hash of each
+# side as of the last confirmed-consistent state. Both languages carry equal authority;
+# after editing either side, bring the other along and re-record with:
+#   pnpm run verify-translation-pairing --write .agents/notes/implemented/feature/2026-07-30-web-context-injection-disclosure.md
+2026-07-30-web-context-injection-disclosure.md: 84c3259f3f226e501a671cc55cacf7d7d96f61fb
+2026-07-30-web-context-injection-disclosure.zh.md: 4d77e06e27badb02fb73ca2ea2a739b33c5804de