Merge branch 'master' into sdk/ts-client-and-subagent

2026-07-27 17:01:20 +08:00
parent 2b31e236f9 4918872c5c
commit 24d2384294
@@ -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
-2026-06-13-twin-llm-adapters.md: 5c3308b281ce71407002e95dd6e794da2a421fa8
-2026-06-13-twin-llm-adapters.zh.md: 93b084973bccaeb802508e4e939a259c281f2608
+2026-06-13-twin-llm-adapters.md: b922891d4438553fd96a7f4f4226f378e66e8ad2
+2026-06-13-twin-llm-adapters.zh.md: d98b57a0a2e7c92453046022e8cb50aa52994c0f
@@ -12,10 +12,10 @@ English | [中文](2026-06-13-twin-llm-adapters.zh.md)

 Ship **two** adapters against the one contract from the start, deliberately built on different internals:

- `dsh-llm-deepseek` — hand-rolled `fetch` + SSE parsing against the DeepSeek API.
+- `dsh-llm-deepseek` — direct `fetch` + in-repo translation against the DeepSeek API; SSE framing is delegated to `eventsource-parser` ([the SSE-parser swap](../simplification/2026-07-26-eventsource-parser-for-deepseek-sse.md)). The twin identity is owning the fetch/translate internals rather than delegating to a full provider SDK, not hand-rolling transport plumbing.
 - `dsh-llm-pi-ai` — the same endpoint through the `@earendil-works/pi-ai` library (its own event vocabulary).

-The rule they enforce: **anything the StreamChunk vocabulary cannot express for BOTH implementations is a core-vocabulary bug**, caught immediately rather than at the next provider. The pair pinned down conventions now documented on `StreamChunk` in `dsh-llm/src/types.ts`: usage emitted before finish, nothing after finish, tool-call `arguments` as raw JSON strings end-to-end, and the two sanctioned error paths (throw from `stream()` *or* end with `finish {kind:'error'|'aborted'}`) that a consumer must handle on both sides — a divergence the library-backed adapter surfaced that a single hand-rolled adapter would have hidden.
+The rule they enforce: **anything the StreamChunk vocabulary cannot express for BOTH implementations is a core-vocabulary bug**, caught immediately rather than at the next provider. The pair pinned down conventions now documented on `StreamChunk` in `dsh-llm/src/types.ts`: usage emitted before finish, nothing after finish, tool-call `arguments` as raw JSON strings end-to-end, and the two sanctioned error paths (throw from `stream()` *or* end with `finish {kind:'error'|'aborted'}`) that a consumer must handle on both sides — a divergence the library-backed adapter surfaced that a single direct-fetch adapter would have hidden.

 ## Alternatives considered

@@ -24,4 +24,4 @@ The rule they enforce: **anything the StreamChunk vocabulary cannot express for

 ## Consequences

-The twin doubles adapter and key-gated e2e maintenance—both cover V4 Flash and Pro across representative reasoning modes—in exchange for continuous seam-neutrality validation and a second implementation example. Both use `apiKey`, `baseURL`, and `models`; the hand-rolled adapter exposes `thinking`/`reasoningEffort`, while pi-ai exposes one `reasoning` level. A future conformance suite could justify retiring one adapter through a superseding Agent Note.
+The twin doubles adapter and key-gated e2e maintenance—both cover V4 Flash and Pro across representative reasoning modes—in exchange for continuous seam-neutrality validation and a second implementation example. Both use `apiKey`, `baseURL`, and `models`; the direct-fetch adapter exposes `thinking`/`reasoningEffort`, while pi-ai exposes one `reasoning` level. A future conformance suite could justify retiring one adapter through a superseding Agent Note.
@@ -12,10 +12,10 @@ Status: implemented

 从一开始就针对同一份契约交付**两个**适配器，刻意基于不同的内部实现构建：

- `dsh-llm-deepseek`：手写 `fetch` + SSE（Server-Sent Events）解析，直接对接 DeepSeek API。
+- `dsh-llm-deepseek`：直接 `fetch` + 仓库内翻译逻辑对接 DeepSeek API；SSE（Server-Sent Events）分帧委托给 `eventsource-parser`（[SSE 解析器替换](../simplification/2026-07-26-eventsource-parser-for-deepseek-sse.md)）。孪生身份在于自行持有 fetch/translate 内部实现而非委托给完整的提供方 SDK，不在于手写传输层管道。
 - `dsh-llm-pi-ai`：通过 `@earendil-works/pi-ai` 库访问同一端点（该库有自己的事件词汇）。

-二者共同执行的规则是：**凡 StreamChunk 词汇无法为两个实现同时表达的内容，都是核心词汇的缺陷**——立即暴露，而非等到下一个提供方接入时才发现。这对孪生体确定了现已记录在 `dsh-llm/src/types.ts` 中 `StreamChunk` 上的约定：usage 在 finish 之前发出、finish 之后不再有任何事件、工具调用的 `arguments` 全程以原始 JSON 字符串传递，以及消费方必须在两侧都处理的两条合法错误路径（`stream()` 抛异常，*或者*以 `finish {kind:'error'|'aborted'}` 结束）。后一项分歧正是由基于库的适配器暴露出来的，单一手写适配器会将其隐藏。
+二者共同执行的规则是：**凡 StreamChunk 词汇无法为两个实现同时表达的内容，都是核心词汇的缺陷**——立即暴露，而非等到下一个提供方接入时才发现。这对孪生体确定了现已记录在 `dsh-llm/src/types.ts` 中 `StreamChunk` 上的约定：usage 在 finish 之前发出、finish 之后不再有任何事件、工具调用的 `arguments` 全程以原始 JSON 字符串传递，以及消费方必须在两侧都处理的两条合法错误路径（`stream()` 抛异常，*或者*以 `finish {kind:'error'|'aborted'}` 结束）。后一项分歧正是由基于库的适配器暴露出来的，单一直接 fetch 适配器会将其隐藏。

 ## 曾考虑的替代方案

@@ -24,4 +24,4 @@ Status: implemented

 ## 后果

-孪生体使适配器和需要密钥的 e2e 维护量翻倍——两者都覆盖 V4 Flash 和 Pro 在各代表性推理（reasoning）模式下的行为——换来的是持续的 seam 中立性验证和第二份实现示例。两个适配器均使用 `apiKey`、`baseURL` 和 `models`；手写适配器暴露 `thinking`/`reasoningEffort`，pi-ai 适配器暴露一个 `reasoning` 级别。未来如果有一致性测试套件，可以通过后续 Agent Note 论证退役其中一个适配器。
+孪生体使适配器和需要密钥的 e2e 维护量翻倍——两者都覆盖 V4 Flash 和 Pro 在各代表性推理（reasoning）模式下的行为——换来的是持续的 seam 中立性验证和第二份实现示例。两个适配器均使用 `apiKey`、`baseURL` 和 `models`；直接 fetch 适配器暴露 `thinking`/`reasoningEffort`，pi-ai 适配器暴露一个 `reasoning` 级别。未来如果有一致性测试套件，可以通过后续 Agent Note 论证退役其中一个适配器。
@@ -2,5 +2,5 @@
 # side as of the last confirmed-consistent state. Both languages carry equal authority;
 # after editing either side, bring the other along and re-record with:
 #   pnpm run verify-translation-pairing --write
-2026-07-20-routed-model-context-and-compaction-policy.md: b637ba24d4ba5fc25c8cdd515a821ee97883a326
-2026-07-20-routed-model-context-and-compaction-policy.zh.md: 084e762ec29ddc0aecb0bf422c147b9d3122726b
+2026-07-20-routed-model-context-and-compaction-policy.md: b00c744e4ac79983a1f492d710e7b1542278c4f4
+2026-07-20-routed-model-context-and-compaction-policy.zh.md: 88e26aa7a99e8deb434d9a8fd80784be410a999a
@@ -14,9 +14,9 @@ Neither obvious configuration owner is sufficient. Compact-basic is optional and

 ### Adapters own exact-route capacity

-`LlmAdapter.resolveModelContext(provider, model)` optionally returns `LlmModelContext` for one exact route. `LlmService.resolveModelContext()` selects the registered route owner, validates a positive integer `contextWindow`, and returns a detached value. The query is independent of `listModels()`: an unlisted dynamic model may have capacity metadata, and `undefined` means only that the adapter cannot describe capacity.
+`LlmAdapter.resolveModel(provider, model, signal?)` returns aggregate metadata for one exact route, with optional `LlmModelContext` under its `context` field. `LlmService.resolveModelInfo()` selects the registered route owner, validates a positive integer `contextWindow`, and returns detached metadata. The query is independent of `listModels()`: an unlisted dynamic model may have capacity metadata, and an absent `context` means only that the adapter cannot describe capacity.

-The hand-rolled DeepSeek adapter accepts optional `contextWindow` on each configured model plus an adapter-wide `defaultContextWindow`. Exact model capacity wins; an entry without capacity and an unlisted pass-through id inherit the adapter default, or return `undefined` when it is absent. The two built-in model entries each publish an exact 128,000-token capacity. The pi-ai adapter resolves capacity from the same catalog descriptor that authoritatively resolves the request model.
+The hand-rolled DeepSeek adapter accepts optional `contextWindow` on each configured model plus an adapter-wide `defaultContextWindow`. Exact model capacity wins; an entry without capacity and an unlisted pass-through id inherit the adapter default, or omit `context` when it is absent. The two built-in model entries each publish an exact 128,000-token capacity. The pi-ai adapter resolves capacity from the same catalog descriptor that authoritatively resolves the request model.

 ### Token measurement remains model-agnostic

@@ -14,9 +14,9 @@ Status: implemented

 ### 适配器拥有精确路由容量

-`LlmAdapter.resolveModelContext(provider, model)` 可以为一条精确路由返回 `LlmModelContext`。`LlmService.resolveModelContext()` 选择已注册的路由所属方，验证 `contextWindow` 为正整数，并返回分离值。该查询独立于 `listModels()`：不在目录中的动态模型也可以拥有容量元数据，而 `undefined` 只表示适配器无法描述容量。
+`LlmAdapter.resolveModel(provider, model, signal?)` 返回一条精确路由的聚合元数据，其中可选的 `LlmModelContext` 位于 `context` 字段下。`LlmService.resolveModelInfo()` 选择已注册的路由所属方，验证 `contextWindow` 为正整数，并返回分离的元数据。该查询独立于 `listModels()`：不在目录中的动态模型也可以拥有容量元数据，而缺少 `context` 只表示适配器无法描述容量。

-手写 DeepSeek 适配器允许每个已配置模型提供可选 `contextWindow`，并支持适配器级 `defaultContextWindow`。精确模型容量优先；未提供容量的模型项与未列出的透传 id 会继承适配器默认值，若默认值也不存在则返回 `undefined`。两个内置模型项都公开精确的 128,000 token 容量。pi-ai 适配器从同一个目录描述符解析容量，该描述符也用于权威解析请求模型。
+手写 DeepSeek 适配器允许每个已配置模型提供可选 `contextWindow`，并支持适配器级 `defaultContextWindow`。精确模型容量优先；未提供容量的模型项与未列出的透传 id 会继承适配器默认值，若默认值也不存在则省略 `context`。两个内置模型项都公开精确的 128,000 token 容量。pi-ai 适配器从同一个目录描述符解析容量，该描述符也用于权威解析请求模型。

 ### Token 计量保持模型无关

@@ -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
-2026-07-23-toolview-dissolution.md: a420c5945d0272cf8087d5f623e9c383c286d7c2
-2026-07-23-toolview-dissolution.zh.md: 47c1f392f5f7ddbf4e6c686b2574faa7987e6126
+2026-07-23-toolview-dissolution.md: 80c2688b152d1afe1236d4815633a5bf024db1d2
+2026-07-23-toolview-dissolution.zh.md: 928c5f445d601b2246d3ae2f9360232643814468
@@ -14,7 +14,7 @@ After the view ring dissolved into the slot system, the client kept exactly one

 The tool ring is gone as independent infrastructure: a tool row is a **keyed child slot each view declares for itself**, and the client has exactly one registration model. The justification above was hollow — a keyed slot's *key space* is already runtime-open (SlotMap declares slots, never keys; the ask-user composer's `key: 'question'` was the precedent), so the open tool-name set fits `entryKey` dispatch natively.

-Shipped shape (current-state narrative also in the [architecture note](2026-07-19-gui-web-client-architecture.md)): the chat entry's `children` table declares `'conversation.chat.toolview'` (keyed/session); the render site dispatches per row via `entryKey: toolName` with `GenericToolCard` as the call-site `fallback` (the default card is domain property; the fallback option is ordinary renderSlot grammar). The owner payload is the uniform `ToolRowOwnerProps` (`callId`/`toolName`/`block`/`openDetails` — details being a session-level facility, not chat-private), and `ToolRowProps` pre-composes it with the session standard kit for registrant components. A registrant is a plain plugin: `ctx.slots.register({ name: 'conversation.chat.toolview', key: '<tool>', inject? }, Row)` with `inject: ['slots', 'conversation']` as the load-order seam — apply mounts `ConversationService` *after* the chat registration, so the service being present guarantees the slot is declared, by construction. Session-dimension differentiation happens inside the component (`useSessions` reading `parentId` — the decision sits where all the information already is); the bash sample is the third-party-posture exemplar. Trajectory/waterfall toolview slots share this exact shape (names fixed by the slot-naming discipline `<domain>.<entry>.<hole>`, one shared owner type) and land with their own row render sites — RendersCheck rejects a declaration nobody renders, so the type system, not convention, blocks early empty declarations.
+Shipped shape (current-state narrative also in the [architecture note](2026-07-19-gui-web-client-architecture.md)): the chat entry's `children` table declares `'conversation.chat.toolview'` (keyed/session); the render site dispatches per row via `entryKey: toolName` with `GenericToolCard` as the call-site `fallback` (the default card is domain property; the fallback option is ordinary renderSlot grammar). The owner payload is the uniform `ToolRowOwnerProps` (`callId`/`toolName`/`block`/`openDetails` — details being a session-level facility, not chat-private), and `ToolRowProps` pre-composes it with the session standard kit for registrant components. A registrant is a plain plugin: `ctx.slots.register({ name: 'conversation.chat.toolview', key: '<tool>', inject? }, Row)` with `inject: ['slots', 'conversation']` as the load-order seam — apply mounts `ConversationService` *after* the chat registration, so the service being present guarantees the slot is declared, by construction. Session-dimension differentiation happens inside the component (`useSessions` reading `parentId` — the decision sits where all the information already is); the bash sample is the third-party-posture exemplar and paints the same ToolRow chrome as Think (`Bash · {description}`, with a scoped badge only in child sessions). Trajectory/waterfall toolview slots share this exact shape (names fixed by the slot-naming discipline `<domain>.<entry>.<hole>`, one shared owner type) and land with their own row render sites — RendersCheck rejects a declaration nobody renders, so the type system, not convention, blocks early empty declarations.

 Registry-era responsibilities all have successor homes: inject caching and row error isolation ride the framework renderer (entry×scope cache, per-entry `SlotErrorBoundary`); subscribe/getVersion ride the slot core's per-key version machinery; the future "store seat" is the ordinary store seat keyed slots already have (interaction-draft durability is its first named consumer); miss fallback is the call-site `fallback` option.

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

 工具环作为独立基础设施已消失：工具行是**各视图为自己声明的 keyed 子槽**，client 全域只剩一种注册模型。上述理由是空的——keyed slot 的 *key 空间*本就运行时开放（SlotMap 声明槽、从不声明 key；ask-user composer 的 `key: 'question'` 即先例），开放的 tool 名集合天然适配 `entryKey` 分发。

-落地形态（现状叙述同见[架构注](2026-07-19-gui-web-client-architecture.md)）：chat 条目的 `children` 表声明 `'conversation.chat.toolview'`（keyed/session）；渲染点逐行以 `entryKey: toolName` 分发、以 `GenericToolCard` 作调用点 `fallback`（默认卡片是域产权；fallback 选项就是普通 renderSlot 文法）。owner 载荷是统一的 `ToolRowOwnerProps`（`callId`/`toolName`/`block`/`openDetails`——details 是会话级设施，非 chat 私货），`ToolRowProps` 把它与 session 标配 kit 预组合供注册方组件取用。注册方就是普通插件：`ctx.slots.register({ name: 'conversation.chat.toolview', key: '<tool>', inject? }, Row)`，以 `inject: ['slots', 'conversation']` 作加载序缝——apply 把 `ConversationService` 挂在 chat 注册*之后*，故服务在场即保证槽已声明，构造使然。会话维差异化在组件内完成（`useSessions` 读 `parentId`——决策放在已有全部信息的地方）；bash 样例即第三方姿态的样板。trajectory/waterfall 的 toolview 槽共用这套形状（槽名按槽名纪律 `<域>.<条目>.<孔位>` 定死，共用一张 owner 类型），随各自的行渲染点落地——RendersCheck 拒绝无人渲染的声明，挡住提前空声明的是类型系统而非约定。
+落地形态（现状叙述同见[架构注](2026-07-19-gui-web-client-architecture.md)）：chat 条目的 `children` 表声明 `'conversation.chat.toolview'`（keyed/session）；渲染点逐行以 `entryKey: toolName` 分发、以 `GenericToolCard` 作调用点 `fallback`（默认卡片是域产权；fallback 选项就是普通 renderSlot 文法）。owner 载荷是统一的 `ToolRowOwnerProps`（`callId`/`toolName`/`block`/`openDetails`——details 是会话级设施，非 chat 私货），`ToolRowProps` 把它与 session 标配 kit 预组合供注册方组件取用。注册方就是普通插件：`ctx.slots.register({ name: 'conversation.chat.toolview', key: '<tool>', inject? }, Row)`，以 `inject: ['slots', 'conversation']` 作加载序缝——apply 把 `ConversationService` 挂在 chat 注册*之后*，故服务在场即保证槽已声明，构造使然。会话维差异化在组件内完成（`useSessions` 读 `parentId`——决策放在已有全部信息的地方）；bash 样例即第三方姿态的样板，并与 Think 绘制同一套 ToolRow chrome（`Bash · {description}`，scoped badge 仅出现在子会话）。trajectory/waterfall 的 toolview 槽共用这套形状（槽名按槽名纪律 `<域>.<条目>.<孔位>` 定死，共用一张 owner 类型），随各自的行渲染点落地——RendersCheck 拒绝无人渲染的声明，挡住提前空声明的是类型系统而非约定。

 registry 时代的职责各有后继居所：inject 缓存与行错误隔离乘框架渲染器（entry×scope 缓存、per-entry `SlotErrorBoundary`）；subscribe/getVersion 乘 slot core 的 per-key 版本机；将来的「store 席位」就是 keyed slot 本就拥有的普通 store 席位（交互草稿耐久性是其首个具名消费者）；miss 兜底即调用点 `fallback` 选项。

@@ -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
+2026-07-24-adapter-owned-reasoning-effort-capabilities.md: cc66e4ec151fcc04445a91f4a3527cbddd130c33
+2026-07-24-adapter-owned-reasoning-effort-capabilities.zh.md: e0d28e1aca370068478e8fb1704defeaac3ab351
@@ -0,0 +1,33 @@
+# Agent Note: Adapter-owned reasoning effort capabilities
+
+Status: implemented
+
+English | [中文](2026-07-24-adapter-owned-reasoning-effort-capabilities.zh.md)
+
+## Problem
+
+Reasoning strength was adapter configuration only, so a conversation could not discover or change the selected model's supported levels between requests. Promoting one adapter's level union into `dsh-llm` would make every provider and model adopt names it may not support, while a provider-specific options bag would make the loop unable to validate or durably reconstruct the effective request.
+
+## Decision
+
+`dsh-llm` represents a reasoning effort as the opaque branded `ReasoningEffortId`. One adapter-owned `resolveModel(provider, model, signal?)` query returns `LlmResolvedModelInfo`: exact model identity plus optional context and reasoning metadata. `LlmService.resolveModelInfo()` validates and detaches that aggregate. When present, `reasoning.efforts` is a non-empty ordered list of ids with display metadata and may name one configured default. The core requires an explicit or configured effort to appear exactly in that list and never clamps or aliases a value.
+
+`LlmCallConfig` and `GenerateOptions` carry the optional effort. The agent loop prepares the post-`agent/request` config under the active turn signal before writing `request/header`, so defaults and dynamic changes are model-visible only after becoming durable facts. The prepared call retains the exact adapter registration across asynchronous exact-model resolution, durable header logging, and dispatch; direct `LlmService.stream()` calls likewise capture their final registration before awaiting resolution. A route with no registered adapter retains its proposed config so an `llm/stream` middleware can own and short-circuit it; terminal dispatch still rejects an unhandled route. A resumed loop retains the logged effort only when its initial provider/model route is unchanged; a route change discards the previous model's opaque id.
+
+The native DeepSeek adapter advertises `off`, `high`, and `max` when deployment policy permits thinking, and defaults to the configured effort or `high`. Its adapter-owned `off` maps to `thinking.type: disabled` with no `reasoning_effort`; `high` and `max` enable thinking and carry their official wire effort. A `thinking: disabled` deployment publishes only `off` and rejects attempts to enable thinking before provider I/O. The pi-ai adapter publishes each exact model's `getSupportedThinkingLevels()` result unchanged, including `off`, preserves an absent profile default as a provider default, and leaves provider wire-value mapping inside pi-ai. Its common stream options represent `off` by omitting `reasoning`, as required by pi-ai's own API.
+
+## Alternatives considered
+
+**Define the pi-ai `ThinkingLevel` union in core.** Rejected because current pi-ai canonical names are an adapter implementation detail; a future provider can expose a different identifier without requiring a core release.
+
+**Carry an untyped provider options object.** Rejected because the loop could neither validate a selected value nor put a stable provider-neutral fact in the request header.
+
+**Clamp unsupported levels.** Rejected because a silent substitution makes the user's selected control differ from the logged request intent and hides stale deployment configuration.
+
+**Normalize every adapter to a core-owned level list or remove `off`.** Rejected because the selectable vocabulary belongs to the exact model capability. A client can render an adapter's `off` option without requiring every adapter to expose it.
+
+## Consequences
+
+Clients can query one exact route once and render its identity, context capacity, and adapter-owned reasoning choices without knowing a global enum or synthesizing `off`. Adapter configuration remains the deployment-default and policy owner, while `agent/request` can replace the effective effort on each step within that policy. Invalid exact identity, context, or reasoning metadata fails with `INVALID_MODEL_INFO`, `INVALID_MODEL_CONTEXT`, or `INVALID_MODEL_REASONING`; unsupported explicit or configured values fail with `UNSUPPORTED_REASONING_EFFORT` before provider I/O.
+
+The aggregate exact-model query is asynchronous and may fail for adapters backed by authoritative catalogs. Its optional signal is the caller's cancellation boundary; an asynchronous adapter must settle promptly after abort so loop disposal can reach quiescence. Keyless service, adapter, loop, session, and request-header tests pin validation, defaulting, dynamic changes, logging, resume behavior, HMR registration ownership, and cancellation; runnable snapshots pin the resolved effort in real assembled request headers, while key-gated adapter tests exercise provider serialization.
@@ -0,0 +1,33 @@
+# Agent Note：适配器持有的推理强度能力
+
+Status: implemented
+
+[English](2026-07-24-adapter-owned-reasoning-effort-capabilities.md) | 中文
+
+## 问题
+
+推理强度过去只能在适配器中配置，因此对话无法在多次请求之间发现或更改所选模型支持的等级。若将某个适配器的等级联合类型提升到 `dsh-llm`，所有提供方和模型都必须采用一套自身可能并不支持的名称；若改用提供方特有的 options 对象，主循环又无法校验最终生效的请求，也无法通过持久化记录准确重建该请求。
+
+## 决策
+
+`dsh-llm` 使用不透明的品牌类型 `ReasoningEffortId` 表示推理强度。由适配器持有的单次 `resolveModel(provider, model, signal?)` 查询返回 `LlmResolvedModelInfo`，其中包含确切模型身份以及可选的上下文和推理元数据。`LlmService.resolveModelInfo()` 会校验该聚合结果并返回分离值。`reasoning.efforts` 存在时，是包含展示元数据的非空有序 ID 列表，并可指定一个由配置确定的默认值。核心要求显式指定或配置指定的推理强度与列表中的某个 ID 完全一致，且绝不自动调整或为值提供别名。
+
+`LlmCallConfig` 和 `GenerateOptions` 携带可选的推理强度。agent loop（智能体循环）在活跃轮次信号的控制下准备 `agent/request` 处理完成后的配置，再写入 `request/header`，因此默认值和动态变更只有成为持久化事实后才对模型可见。准备完成的调用在异步确切模型解析、请求头持久记录和分派全程保留同一项确切的适配器注册；直接调用 `LlmService.stream()` 时，也会在等待解析前捕获最终的适配器注册。没有已注册适配器的路由会保留原定配置，使 `llm/stream` 中间件可以接管并短路该请求；若仍未得到处理，最终分发会拒绝该路由。恢复后的主循环仅在初始提供方/模型路由未变时保留日志中记录的推理强度；如果路由发生变化，则丢弃上一模型的不透明 ID。
+
+当部署策略允许思考时，原生 DeepSeek 适配器声明 `off`、`high` 和 `max`，默认使用配置指定的推理强度，若未配置则使用 `high`。由适配器持有的 `off` 映射为 `thinking.type: disabled`，且不带 `reasoning_effort`；`high` 和 `max` 会启用思考并携带各自的官方协议强度值。配置为 `thinking: disabled` 的部署仅声明 `off`，并会在提供方 I/O 前拒绝启用思考的尝试。pi-ai 适配器原样发布每个确切模型的 `getSupportedThinkingLevels()` 结果，其中包括 `off`；profile 未指定默认值时保留提供方默认行为，并将提供方协议值的映射留在 pi-ai 内部。按照 pi-ai 自身 API 的要求，其通用流选项通过省略 `reasoning` 来表示 `off`。
+
+## 备选方案
+
+**在核心中定义 pi-ai 的 `ThinkingLevel` 联合类型。** 不予采纳：pi-ai 当前的规范名称属于适配器实现细节；未来的提供方可以暴露不同的标识符，而无需为此发布新的核心版本。
+
+**携带无类型约束的提供方 options 对象。** 不予采纳：主循环既无法校验选定值，也无法在请求头中写入稳定且与提供方无关的事实。
+
+**自动调整不支持的等级。** 不予采纳：静默替换会导致用户选定的控制项与日志记录的请求意图不一致，还会掩盖陈旧的部署配置。
+
+**将每个适配器规范化为核心持有的等级列表，或移除 `off`。** 不予采纳：可选值集合属于确切模型的能力。客户端可以渲染某个适配器的 `off` 选项，而无需要求所有适配器都暴露该选项。
+
+## 影响
+
+客户端只需查询一次确切路由，即可渲染其身份、上下文容量和由适配器持有的推理选项，而无需了解全局枚举或自行合成 `off`。适配器配置仍是部署默认值和策略的归属方，`agent/request` 则可以在该策略范围内为每个步骤替换实际生效的推理强度。确切身份、上下文或推理元数据无效时，分别抛出 `INVALID_MODEL_INFO`、`INVALID_MODEL_CONTEXT` 或 `INVALID_MODEL_REASONING`；显式指定或配置指定的值不受支持时，会在提供方 I/O 前抛出 `UNSUPPORTED_REASONING_EFFORT`。
+
+确切模型元数据的聚合查询采用异步方式，并且对于由权威目录支持的适配器可能失败。可选信号构成调用方的取消边界；异步适配器必须在信号中止后迅速完成结算，使主循环的资源释放达到完全停稳。无密钥的服务、适配器、主循环、会话和请求头测试为校验、默认值解析、动态变更、日志记录、恢复行为、HMR（热模块替换）期间的注册所有权和取消提供回归保障；可运行快照锁定实际组装请求头中的已解析推理强度，仅在有密钥时运行的适配器测试则覆盖提供方序列化。
@@ -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
-2026-06-29-todo-write-tool.md: df1bee2801b0e01b290b63f6edbe2e5b1be80cb7
-2026-06-29-todo-write-tool.zh.md: 7fa5cb2aad2b32ef0662df04ff6576be14a3a8e7
+2026-06-29-todo-write-tool.md: 760373d64f462e3717a174d5793e6d47ab76b0b4
+2026-06-29-todo-write-tool.zh.md: fd29e6049e6a729c2c7e78860ad7c065422da60e
@@ -10,7 +10,7 @@ The harness gives the model bash and subagent tools but no way to record a struc

 ## Decision

-Add a model-facing `todo_write(todos: [{ content, status }])` tool whose whole-list state lives on the event-sourced session log as a new `todo/write` `SessionEventMap` variant. Interactive hosts render from the durable event; the TUI folds it directly, while the [automation-only ACP bridge](../simplification/2026-07-23-acp-automation-only-protocol.md) deliberately omits todo presentation.
+Add a model-facing `todo_write(todos: [{ content, status }])` tool whose whole-list state lives on the event-sourced session log as a new `todo/write` `SessionEventMap` variant. Interactive hosts render from the durable event: the TUI folds it directly, the web client projects it into `ConversationSnapshot.todos` ([web todo display](2026-07-23-web-todo-display.md)), while the [automation-only ACP bridge](../simplification/2026-07-23-acp-automation-only-protocol.md) deliberately omits todo presentation.

 ### Whole-list replace, three-state status

@@ -18,7 +18,7 @@ The model sends the entire list every call; the new list replaces the old (last-

 ### State on the session log, not a service

-The list is appended as a `todo/write` event carrying the full `{ todos }` snapshot. The harness is event-sourced — the LLM history, tool calls, and turn structure all live on the log — so the todo list lives there too. This buys durability, replay, and resume reconstruction for free: a reopened session re-derives the current list from the latest `todo/write`, with no separate persistence backend, in-memory service to rehydrate, or extra wiring. An in-memory `ctx.todos` service would have to reinvent all of that.
+The list is appended as a `todo/write` event carrying the full `{ todos }` snapshot. The harness is event-sourced — the LLM history, tool calls, and turn structure all live on the log — so the todo list lives there too. This buys durability, replay, and resume reconstruction for free: a reopened session re-derives the current list from the latest `todo/write`, with no separate persistence backend, in-memory service to rehydrate, or extra wiring. An in-memory `ctx.todos` service would have to reinvent all of that. (Full-log consumers get this reconstruction outright; the web client's paged window gets it from the tail history page's host-computed projection — see the [web todo display note](2026-07-23-web-todo-display.md).)

 ### NOT a surface event

@@ -10,7 +10,7 @@ harness 为模型提供了 bash 和 subagent 工具，却没有办法记录结

 ## 决策

-新增一个面向模型的 `todo_write(todos: [{ content, status }])` 工具，其整列表状态作为新的 `todo/write` `SessionEventMap` 变体存储在事件溯源的会话日志上。交互式宿主从持久事件渲染；TUI 直接折叠它，而[仅面向自动化的 ACP（Agent Client Protocol）桥接层](../simplification/2026-07-23-acp-automation-only-protocol.md)有意省略 todo 展示。
+新增一个面向模型的 `todo_write(todos: [{ content, status }])` 工具，其整列表状态作为新的 `todo/write` `SessionEventMap` 变体存储在事件溯源的会话日志上。交互式宿主从持久事件渲染：TUI 直接折叠它，web 客户端将其投影进 `ConversationSnapshot.todos`（[web todo 展示](2026-07-23-web-todo-display.md)），而[仅面向自动化的 ACP（Agent Client Protocol）桥接层](../simplification/2026-07-23-acp-automation-only-protocol.md)有意省略 todo 展示。

 ### 整列表替换，三态 status

@@ -18,7 +18,7 @@ harness 为模型提供了 bash 和 subagent 工具，却没有办法记录结

 ### 状态在会话日志上，而非服务

-列表作为 `todo/write` 事件追加到日志，携带完整的 `{ todos }` 快照。harness 是事件溯源的——LLM（大语言模型）历史、工具调用和轮次结构都在日志上——所以 todo 列表也在那里。这免费获得了持久性、回放和恢复重建：重新打开的会话从最新的 `todo/write` 重新推导当前列表，无需独立的持久化后端、无需重新注水的内存服务、无需额外接线。一个内存中的 `ctx.todos` 服务需要重新发明以上所有。
+列表作为 `todo/write` 事件追加到日志，携带完整的 `{ todos }` 快照。harness 是事件溯源的——LLM（大语言模型）历史、工具调用和轮次结构都在日志上——所以 todo 列表也在那里。这免费获得了持久性、回放和恢复重建：重新打开的会话从最新的 `todo/write` 重新推导当前列表，无需独立的持久化后端、无需重新注水的内存服务、无需额外接线。一个内存中的 `ctx.todos` 服务需要重新发明以上所有。（全量 log 消费者直接获得这份重建；web 客户端的分页窗口则从尾页 history 携带的 host 计算投影获得——见 [web todo 展示 Note](2026-07-23-web-todo-display.md)。）

 ### 不是 surface 事件

@@ -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
-2026-07-16-persistent-pty-sessions.md: 148d4a2f47689e38a3ec83a7a41e4f75c4b73d95
-2026-07-16-persistent-pty-sessions.zh.md: 9a9d9cd4b0f61e8abaf011996ecd8739d13851f8
+#   pnpm run verify-translation-pairing --write .agents/notes/implemented/feature/2026-07-16-persistent-pty-sessions.md
+2026-07-16-persistent-pty-sessions.md: 43c87bb159cfe1ab9f8d3a80c2adf25a57ae6e3b
+2026-07-16-persistent-pty-sessions.zh.md: 8afc2103447cc58b1fcbc1062b9564e8ed643477
@@ -154,7 +154,7 @@ The package ships concise tool guidance explaining persistent state, owner isola

 - Per-file coverage pins owner fencing, concurrent reservations, unpublished-spawn cancellation and awaited teardown, sandbox-mode change rejection, retriable lifecycle cleanup, readiness tiers, sanitizer carry state, complete UTF-8 bounds, task integration, schemas, and exact render intents.
 - Linux process fixtures cover non-leader and non-main-thread stdin waits, zombie quiescence, unreadable process state, supported syscall tables, unsupported architectures, and false-positive rejection; macOS inspector logic is injected into the same unit suite.
- Real `node-pty` tests exercise shell state, shared sandbox policy, environment scrubbing, raw-mode foreground `SIGINT`, a TERM-ignoring descendant, and immediate post-disposal quiescence on supported hosts.
+- Real `node-pty` tests exercise shell state, shared sandbox policy, environment scrubbing, raw-mode foreground `SIGINT` after deliberately delayed child readiness under scenario-owned timing bounds, a TERM-ignoring descendant, and immediate post-disposal quiescence on supported hosts.
 - A Loader-driven `cordis.yml` test mounts the real three-package composition. ACP and headless snapshots pin the six schemas, bounded results, and errors through opt-in overlays; TUI snapshots pin terminal and generic card presentation.
 - Package contracts, the architecture map, core data structures, generated catalogs, and the website API describe the same shipped surface.
 - The repository CI-equivalent sequence owns type, lint, coverage, snapshot, documentation, build, hygiene, demo, and built-entry verification.
@@ -154,7 +154,7 @@ plugins:

 - 每文件覆盖率固定 owner 隔离、并发预留、未发布 spawn 的取消与等待式 teardown、沙箱模式变更拒绝、可重试的生命周期清理、就绪层级、sanitizer carry state、完整 UTF-8 结果上限、task 集成、schema 和精确 render intent。
 - Linux 进程 fixture 覆盖非 leader 与非主线程的 stdin 等待、僵尸进程静止性、不可读进程状态、受支持的 syscall 表、不支持的架构和误报拒绝；同一单元测试套件通过注入覆盖 macOS 检查器逻辑。
- 真实 `node-pty` 测试在受支持宿主上覆盖 shell 状态、共享沙箱策略、环境清洗、raw mode 下的前台 `SIGINT`、忽略 `SIGTERM` 的子进程，以及 dispose 返回后立即静默。
+- 真实 `node-pty` 测试在受支持宿主上覆盖 shell 状态、共享沙箱策略、环境清洗、在由场景掌控的时间界限内先有意延迟子进程就绪，再对 raw mode 前台进程发送 `SIGINT`、忽略 `SIGTERM` 的子进程，以及 dispose 返回后立即完全停稳。
 - Loader 驱动的 `cordis.yml` 测试挂载真实三包组合。ACP 与 headless 快照通过 opt-in overlay 固定 6 个 schema、有界结果和错误；TUI 快照固定 terminal 与 generic 卡片展示。
 - 包契约、架构图、核心数据结构、生成目录和 website API 描述同一个已发布接口。
 - 仓库 CI 等价序列负责类型、lint、覆盖率、快照、文档、构建、hygiene、demo 和 built-entry 验证。
@@ -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
-2026-07-17-dedicated-full-screen-tui-front-door.md: aac67ffec89606d04d5abfd233d0469e2241b102
-2026-07-17-dedicated-full-screen-tui-front-door.zh.md: 0f64e2ce14b18d315f75913b82c731e77758e377
+2026-07-17-dedicated-full-screen-tui-front-door.md: 8d7c7b00c8d9b15ea3f2419ed44ca88209e60dad
+2026-07-17-dedicated-full-screen-tui-front-door.zh.md: 49e9541c9f98c9f3beba11945ff452fc38bd9ede
@@ -22,9 +22,9 @@ The selected front door receives the exact generated or resumed `SessionId` used

 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, 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. The idle footer derives context occupancy from `tokenMeter` and shows the selected model; 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 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; agent behavior and answer logging remain owned by their existing services.
+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. 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; agent behavior and answer logging remain owned by their existing services.

-The `/model` command presents the advisory `ctx.llm` catalog as a keyboard selector and changes only this TUI session's target; argument forms remain available for direct selection. Agent-scoped prompt-assembly and request waterfalls snapshot one provider/model pair per step, so `{{provider}}` / `{{model}}` interpolation and request routing cannot split when a command arrives during assembly. The latest logged request header restores a used target; a selection that never reaches a request remains process-local.
+The `/model` command presents the advisory `ctx.llm` catalog as a keyboard selector and changes only this TUI session's target; argument forms remain available for direct selection. Each model row owns the adapter-advertised reasoning-effort order and default: Shift+Tab cycles that row's efforts, includes provider-default behavior when the adapter advertises no default, and leaves models without selectable metadata unchanged. Agent-scoped prompt-assembly and request waterfalls snapshot one provider/model/reasoning-effort target per step, so `{{provider}}` / `{{model}}` interpolation and request routing cannot split when a command arrives during assembly. The latest logged request header restores a used target; a selection that never reaches a request remains process-local.

 ### Terminal ownership

@@ -49,4 +49,4 @@ The implemented [TUI terminal-state snapshot Agent Note](../testing/2026-07-18-t
 - 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.
 - Tool packages extend terminal cards through their existing presentation methods without adding tool-specific branches to the TUI.
- Model selection uses adapter-advertised metadata without turning catalog membership into request validation; unused selections are not durable state.
+- Model and reasoning-effort selection use adapter-advertised metadata without turning catalog membership into request validation; unused selections are not durable state.
@@ -22,9 +22,9 @@ DeepSeek Harness 将 [`@deepseek-ai/dsh-tui`](../../../../packages/ui/tui/README

 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、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。空闲态页脚根据 `tokenMeter` 得出上下文占用率，并显示所选模型；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；agent 行为和答案日志仍由既有服务负责。
+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。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；agent 行为和答案日志仍由既有服务负责。

-`/model` 命令将建议性的 `ctx.llm` 目录呈现为键盘选择器，并且只更改当前 TUI 会话的目标；带参数的形式仍可直接选择目标。agent 作用域内的 prompt 组装和请求两条 waterfall（瀑布式事件）会为每个 step 快照一次同一个提供方/模型字段组合，因此即使命令在组装期间到达，`{{provider}}` / `{{model}}` 插值与请求路由也不会分裂。系统通过日志中最新的请求头恢复已经使用过的目标；未被请求使用的选择只保留在当前进程中。
+`/model` 命令将建议性的 `ctx.llm` 目录呈现为键盘选择器，并且只更改当前 TUI 会话的目标；带参数的形式仍可直接选择目标。每个模型行都持有适配器公布的推理强度顺序和默认值：按 Shift+Tab 可循环切换该行的推理强度；如果适配器没有公布默认值，循环中还会包含提供方默认行为；没有可选元数据的模型则保持不变。agent 作用域内的 prompt 组装和请求两条 waterfall（瀑布式事件）会为每个步骤快照一次同一个提供方/模型/推理强度目标，因此即使命令在组装期间到达，`{{provider}}` / `{{model}}` 插值与请求路由也不会分裂。系统通过日志中最新的请求头恢复已经使用过的目标；未被请求使用的选择只保留在当前进程中。

 ### 终端所有权

@@ -49,4 +49,4 @@ agent 空闲时，编辑器输入调用 `agent.send()`；轮次运行中则调
 - 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。
 - 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。
- 模型选择使用适配器提供的目录元数据，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。
+- 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。
@@ -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
-2026-07-23-web-assistant-markdown.md: ce98a16fa43e2743c18826ee7f2344c38e7c70e7
-2026-07-23-web-assistant-markdown.zh.md: 0d6fd2f9e6b91f76830586ecf4c29774e5d6978a
+2026-07-23-web-assistant-markdown.md: 38d193271d88b3a8f32ba1b191e8a6d432176281
+2026-07-23-web-assistant-markdown.zh.md: be3cd041c6012af142fc27934fda125dfc4cf6de
@@ -12,13 +12,17 @@ The Web conversation preserves assistant Markdown source through session events,

 `@deepseek-ai/dsh-client-ui-primitives` exports `MarkdownText` as the untrusted assistant-text renderer, and `ui-conversation` selects it only for assistant `text` blocks. Finalized history, the streaming tail, and interrupted partials already share `AssistantMarkdown`, so they receive the same renderer without changing events or snapshots. User and steering messages keep `MessageText` and remain literal.

-`MarkdownText` uses `react-markdown` with `remark-gfm` to build React elements from an AST. It covers CommonMark blocks plus GFM tables, task lists, strikethrough, and autolinks without `dangerouslySetInnerHTML`, raw-HTML parsing, or syntax highlighting. The dependency is explicit in `ui-primitives`; because that pure library is seeded by the Web shell, the parser is part of the initial browser bundle.
+`MarkdownText` uses `react-markdown` with `remark-gfm` to build React elements from an AST. It covers CommonMark blocks plus GFM tables, task lists, strikethrough, and autolinks without raw-HTML parsing. Fenced code routes through the shared `CodeBlock`, which highlights registered grammars with the client's shiki singleton (`--shiki-*` tokens) and falls back to plain monospace otherwise. While a turn streams, fences stay on the plain arm so growing fences are not retokenized every chunk.
+
+Visual spacing, tables, links, blockquotes, inline code, and code-block chrome follow deepsuite `@deepseek/md` (`markdown.css` / `code-block.css`) and the same `--dsw-alias-markdown-*`, `--dsw-font-markdown-*`, `--dsw-alias-border-l*`, and `--dsw-alias-label-*` tokens. Links use `--dsw-alias-state-business-primary` (deepsuite's sheet uses `--dsw-alias-brand-text`, which is blue only under newDesign; design-platform keeps brand-text near-black and is not retuned here). `CodeBlock` ships a language banner and a copy control (`复制` / `复制成功`). Citation pills, KaTeX, heading anchors, the thinking-small markdown variant, and custom □/☑ task markers are out of scope until matching product DOM exists; GFM task lists keep native checkboxes.
+
+The dependency is explicit in `ui-primitives`; because that pure library is seeded by the Web shell, the parser and highlighter are part of the initial browser bundle.

 ## Untrusted output policy

-Assistant-authored destinations are restricted to absolute HTTP, HTTPS, and mailto URLs. HTTP(S) links open in a new tab with `rel="noopener noreferrer"`; relative destinations and other protocols render as non-navigable text. Markdown images render only their alt text, so model output cannot initiate a remote image request. Raw HTML remains inert source text because no HTML parser enters the pipeline.
+Assistant-authored destinations are restricted to absolute HTTP, HTTPS, and mailto URLs. HTTP(S) links open in a new tab with `rel="noopener noreferrer"`; relative destinations and other protocols render as non-navigable text. Markdown images render only their alt text, so model output cannot initiate a remote image request. Raw HTML remains inert source text because no HTML parser enters the pipeline. Shiki output is a static span tree generated from the fence text (no scripts or user HTML).

-The renderer uses existing `--dsw-*` typography and color tokens. Fenced code and GFM tables own horizontal overflow so long content cannot widen the conversation column.
+Fenced code and GFM tables own horizontal overflow so long content cannot widen the conversation column.

 ## Alternatives considered

@@ -30,6 +34,8 @@ The renderer uses existing `--dsw-*` typography and color tokens. Fenced code an

 **Enable raw HTML or remote images with sanitization.** Neither capability has a current product need, while both enlarge the executable or network privacy boundary. They remain disabled rather than adding sanitizer and image-policy dependencies.

+**Port deepsuite Prism `highlight.css` and the mdast pipeline.** Appearance parity is owned by CSS Modules and shared `--dsw-*` tokens; highlighting stays on the existing shiki allowlist so the client does not take a second highlighter or Prism class contract.
+
 ## Consequences

-Assistant replies render semantic Markdown consistently during streaming and replay, while tool cards, reasoning rows, interactions, user bubbles, and the host protocol remain unchanged. Streaming reparses the current text after each accumulated update; incomplete Markdown can temporarily change structure, but the isolated tail bounds React invalidation and the final event does not switch renderers. The initial Web shell grows by the Markdown parser and GFM runtime, and future extensions such as syntax highlighting or remote media require a separate bundle and security decision.
+Assistant replies render semantic Markdown consistently during streaming and replay, while tool cards, reasoning rows, interactions, user bubbles, and the host protocol remain unchanged. Streaming reparses the current text after each accumulated update; incomplete Markdown can temporarily change structure, but the isolated tail bounds React invalidation and the final event does not switch renderers. Code fences share one chrome and copy path with tool and details surfaces. The initial Web shell includes the Markdown parser, GFM runtime, and shiki allowlist; cite/math/anchor/thinking-small surfaces remain deferred.
@@ -12,13 +12,17 @@ Web 对话通过会话事件、历史回放与流式累积保留 assistant Markd

 `@deepseek-ai/dsh-client-ui-primitives` 导出 `MarkdownText`，用作不受信任的 assistant 文本渲染器；`ui-conversation` 仅为 assistant `text` 块选择该渲染器。已完成的历史消息、流式输出尾部与被中断的部分输出已经共用 `AssistantMarkdown`，因此无需更改事件或快照，它们便会采用同一渲染器。用户消息与 steering 消息继续使用 `MessageText`，并保持按字面渲染。

-`MarkdownText` 使用 `react-markdown` 与 `remark-gfm`，从 AST 构建 React 元素。它支持 CommonMark 块，以及 GFM 表格、任务列表、删除线与自动链接，但不使用 `dangerouslySetInnerHTML`，不解析原始 HTML，也不进行语法高亮。`ui-primitives` 显式声明该依赖；由于这一纯库由 Web shell 预置，解析器会成为初始浏览器 bundle 的一部分。
+`MarkdownText` 使用 `react-markdown` 与 `remark-gfm`，从 AST 构建 React 元素。它覆盖 CommonMark 块，以及 GFM 表格、任务列表、删除线与自动链接，且不解析原始 HTML。围栏代码经共享的 `CodeBlock` 路由；该组件用客户端的 shiki 单例（`--shiki-*` token）高亮已注册语法，否则回退为纯等宽文本。轮次流式输出期间，围栏停留在纯文本分支，以免每收到一个分片就对增长中的围栏重新分词。
+
+视觉间距、表格、链接、引用块、行内代码与代码块外框遵循 deepsuite `@deepseek/md`（`markdown.css` / `code-block.css`），并使用同一套 `--dsw-alias-markdown-*`、`--dsw-font-markdown-*`、`--dsw-alias-border-l*` 与 `--dsw-alias-label-*` token。链接使用 `--dsw-alias-state-business-primary`（deepsuite 的样式表使用 `--dsw-alias-brand-text`，仅在 newDesign 下为蓝色；design-platform 将 brand-text 保持为近黑色，此处不做重新调色）。`CodeBlock` 提供语言横幅与复制控件（`复制` / `复制成功`）。引用胶囊、KaTeX、标题锚点、thinking-small markdown 变体，以及自定义 □/☑ 任务标记均不在范围内，直至存在匹配的产品 DOM；GFM 任务列表继续使用原生复选框。
+
+该依赖在 `ui-primitives` 中显式声明；由于这一纯库由 Web shell 预置，解析器与高亮器会成为初始浏览器 bundle 的一部分。

 ## 不受信任输出策略

-assistant 生成的目标地址仅限绝对 HTTP、HTTPS 与 mailto URL。HTTP(S) 链接会在新标签页中打开，并带有 `rel="noopener noreferrer"`；相对目标地址与其他协议会渲染为不可导航的文本。Markdown 图片仅渲染替代文本，因此模型输出无法发起远程图片请求。由于管线中未引入 HTML 解析器，原始 HTML 仍是不会生效的源文本。
+assistant 生成的目标地址仅限绝对 HTTP、HTTPS 与 mailto URL。HTTP(S) 链接会在新标签页中打开，并带有 `rel="noopener noreferrer"`；相对目标地址与其他协议会渲染为不可导航的文本。Markdown 图片仅渲染替代文本，因此模型输出无法发起远程图片请求。由于管线中未引入 HTML 解析器，原始 HTML 仍是不会生效的源文本。Shiki 输出是由围栏文本生成的静态 span 树（不含脚本或用户 HTML）。

-渲染器使用现有的 `--dsw-*` 排版与颜色 token。围栏代码块与 GFM 表格各自处理横向溢出，因此较长内容无法撑宽对话栏。
+围栏代码与 GFM 表格各自处理横向溢出，因此较长内容无法撑宽对话栏。

 ## 考虑过的替代方案

@@ -30,6 +34,8 @@ assistant 生成的目标地址仅限绝对 HTTP、HTTPS 与 mailto URL。HTTP(S

 **通过净化启用原始 HTML 或远程图片。**当前产品并不需要这两项功能，但二者都会扩大可执行行为或网络隐私边界。因此它们保持禁用，无需增加净化器与图片策略依赖。

+**移植 deepsuite 的 Prism `highlight.css` 与 mdast 管线。**外观一致性由 CSS Modules 与共享的 `--dsw-*` token 负责；高亮仍走现有的 shiki 允许列表，使客户端不必引入第二套高亮器或 Prism class 契约。
+
 ## 后果

-assistant 回复在流式输出与回放期间都会一致地渲染为语义化 Markdown，而工具卡片、推理行、交互、用户气泡和宿主协议保持不变。每次累积更新后，流式输出都会重新解析当前文本；未完成的 Markdown 可能暂时改变结构，但独立的尾部会限定 React 失效范围，最终事件也不会切换渲染器。初始 Web shell 的体积会因加入 Markdown 解析器与 GFM 运行时而增大；语法高亮或远程媒体等后续扩展需要另行作出 bundle 与安全决策。
+assistant 回复在流式输出与回放期间都会一致地渲染为语义化 Markdown，而工具卡片、推理行、交互、用户气泡和宿主协议保持不变。每次累积更新后，流式输出都会重新解析当前文本；未完成的 Markdown 可能暂时改变结构，但独立的尾部会限定 React 失效范围，最终事件也不会切换渲染器。代码围栏与工具及详情表层共用同一外框与复制路径。初始 Web shell 包含 Markdown 解析器、GFM 运行时与 shiki 允许列表；cite/math/anchor/thinking-small 表层仍暂缓。
@@ -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/2026-07-23-web-todo-display.md
+2026-07-23-web-todo-display.md: 5fe08cc40c1d23ff3a9b8c6d766fea6d3694c30d
+2026-07-23-web-todo-display.zh.md: c121ffc27e3d0a93707c2c22b2f180023ebae5be
@@ -0,0 +1,36 @@
+# Agent Note: Web todo display — snapshot side-effect channel + two render surfaces
+
+Status: implemented
+
+English | [中文](2026-07-23-web-todo-display.zh.md)
+
+## Problem
+
+`todo_write` appends `todo/write` whole-list snapshots to the session log; the TUI renders a persistent plan panel (the automation-only ACP bridge deliberately omits todo presentation). The web client dropped the event entirely: the host mux stream already forwards every session event, but `todo/write` is not a surface type (it never folds into `ConversationSnapshot.nodes`), and no side-effect branch accumulated it — the browser had no consumption point and no display surface.
+
+## Decision
+
+Consume `todo/write` as a Session side effect, not a surface node, and render it on two surfaces matching the split the TUI already draws.
+
+### Side-effect channel, converging with window replay
+
+`applyEventSideEffects` gains a `todo/write` case (whole list, last write wins). Unlike partial/openCalls, `rebuildDerivedFromWindow` deliberately does NOT reset it: the value is session-level — taken from the tail history page's full-log projection — and an arbitrary window may not contain the latest write, so an older-page prepend keeps it and only an in-window or live write overwrites it. Every `installWindow` caller is a tail request (`doOpen`, its gap re-pull, `repairGap`; `loadOlder` prepends without it), which the host answers with the projection or omits it only when the full log holds no `todo/write` — so an absent field is the authoritative empty list and is assigned as such. That distinction matters on rollback: a live write whose host crashed before persisting leaves the log empty, and preserving the prior value instead would strand the rolled-back plan on screen indefinitely. `ConversationSnapshot.todos` is the read surface. This follows the event's own contract ("log-only UI state; never derived history"): surfacing each write as a conversation node would render superseded lists as if they were still standing.
+
+### TodoPanel: the durable list as a persistent strip
+
+The panel mounts through the `conversation.input.dock` slot (a plain registrant plugin, `todoDockEntry`, the QueueDock posture: `inject: ['slots', 'conversation']` as the load-order seam, `order: -1` above the queue rows), hidden while empty, collapsible to a header of title + `"<done>/<total> tasks · <n> in progress"` (no in-progress content hint when collapsed). Status glyphs are the figma todo set (green check ring / blue fading ring / dashed pending ring) on a tip-surface card (`--dsw-specific-tip`, 14px radius, `width: calc(100% - 88px)` / `max-width: 776px` centered; InputBar top pad 6px is the gap to the composer card). It reads `snapshot.todos` via the standard-kit `useSession` hook the dock entry receives — no store, no service, no ctx. The inner component stays props-complete and framework-free; the dock adapter is a one-line wrapper.
+
+### TodoRow: the per-call row through the keyed toolview slot
+
+The dedicated `todo_write` chat row is a plain registrant plugin (`todoToolview`, mounted from `apply`) that registers into the keyed `conversation.chat.toolview` slot via `ctx.slots.register` — the same seam and load-order posture as the bash sample (`inject: ['slots', 'conversation']`), but a product registration. The summary derives from call args (`N/M done · active item`); unparseable args fall back to the generic row summary; clicking opens the details column with the raw args. No `ToolEventView` is added for todo — presentation is client-owned, and the durable list renders from the session event, not the tool card.
+
+## Alternatives considered
+
+- **Fold todo writes into `nodes` as surface entries** — replayed windows would render every superseded list; the event is deliberately not a surface type.
+- **Hardcoding the panel inside `ConversationRoot`** — the original landing spot before the input-dock slot existed; the dock is the architecture's home for always-on strips above the composer, and a hardcode bypasses the slot registry's disposal and ordering.
+- **Details column for the panel** — the details slot is single-occupant and selection-driven, a different lifetime than an always-on strip.
+- **Host-computed view (a todo `ToolEventView`)** — presentation belongs to the client; the wire already carries the whole snapshot in the event payload.
+
+## Consequences
+
+Replay correctness is owned by one code path: any future change to window rebuild keeps todos consistent for free, and the fixture (fx-alpha turn 65) plus the assembled keyless snapshot (`apps/web/tests/todo-display.snapshot.ts`) pin the full chain (row summary and state, dock panel content, collapse round-trip) over the built client graph. `todos` is a required `ConversationSnapshot` field, so scripted fakes in specs must carry it. The TUI panel is untouched (the automation-only ACP bridge deliberately omits todo presentation); the web surfaces render the same event, adding one wire field and no new event type. That field is how cold-load reconstruction stays host-backed: the tail history page carries `todos` — the full-log latest `todo/write`, computed independently of the page window (the same backscan posture the view pairing uses) — so a reopened session restores the plan even when the last write precedes the window; that value survives an older-page prepend, is overwritten by any later write, and resets to empty when a tail response carries no projection.
@@ -0,0 +1,36 @@
+# Agent Note: Web todo 展示——快照副作用通道 + 两个渲染面
+
+Status: implemented
+
+[English](2026-07-23-web-todo-display.md) | 中文
+
+## Problem
+
+`todo_write` 把 `todo/write` 的整份列表快照追加进会话日志；TUI 渲染一块常驻的 plan 面板（自动化专用的 ACP 桥接刻意不做 todo 呈现）。Web 客户端把这个事件整个丢弃了：host mux 流本已转发每一个会话事件，但 `todo/write` 不是 surface 类型（它从不 fold 进 `ConversationSnapshot.nodes`），也没有任何副作用分支累积它——浏览器既无消费点，也无展示面。
+
+## Decision
+
+把 `todo/write` 当作 Session 副作用消费，而非 surface 节点，并在两个面上渲染它，这两个面正对应 TUI 已经绘制的那套划分。
+
+### 副作用通道，与窗口回放收敛
+
+`applyEventSideEffects` 新增一个 `todo/write` 分支（整份列表，后写覆盖先写）。与 partial/openCalls 不同，`rebuildDerivedFromWindow` 刻意不重置它：该值是会话级的——取自尾页 history 携带的全量 log 投影——而任意窗口未必包含最近一次写入，因此往前翻页保留它，只有窗口内或实时的写入才会覆盖。`installWindow` 的每个调用方都是尾页请求（`doOpen`、其补洞重拉、`repairGap`；`loadOlder` 只往前拼接、不走它），而 host 对尾页请求要么带上投影、要么仅在全量 log 没有任何 `todo/write` 时省略——因此字段缺失就是权威的空列表，直接照此赋值。这个区分在回滚场景上要紧：实时写入若在 host 持久化前崩溃，log 里就是空的，此时保留旧值会让已回滚的计划永远留在屏幕上。`ConversationSnapshot.todos` 是读取面。这遵循事件自身的契约（「仅日志 UI 状态，绝非派生历史」）：把每次写入作为对话节点呈现，会让已被取代的列表看起来仍然有效。
+
+### TodoPanel：长驻列表作为一条常驻横条
+
+面板经 `conversation.input.dock` slot 挂载（普通注册者插件 `todoDockEntry`，QueueDock 同款姿势：`inject: ['slots', 'conversation']` 载序 seam，`order: -1` 排在队列条上方），空列表时隐藏，可折叠为标题加 `"<已完成>/<总数> tasks · <n> in progress"` 的表头（折叠态不再附带进行中条目正文）。状态图标为 figma todo 套件（绿色勾选环／蓝色渐隐环／虚线未开始环），卡片使用 tip 表面（`--dsw-specific-tip`、14px 圆角、`width: calc(100% - 88px)`／`max-width: 776px` 居中；InputBar 顶部 6px 内边距是到输入卡的间距）。它经 dock entry 收到的标准件 `useSession` hook 读取 `snapshot.todos`——无 store、无 service、无 ctx。内部组件保持 props 完备且框架无关；dock 适配件只是一行包装。
+
+### TodoRow：经 keyed toolview slot 的逐调用行
+
+专用的 `todo_write` 对话行是一个普通注册者插件（`todoToolview`，由 `apply` 挂载），经 `ctx.slots.register` 注册进 keyed 的 `conversation.chat.toolview` slot——与 bash 样例同一接缝、同一载序姿态（`inject: ['slots', 'conversation']`），但属产品级注册。摘要由调用 args 推导（`N/M done · active item`）；无法解析的 args 回退到通用行摘要；点击会以原始 args 打开 details 列。todo 不新增任何 `ToolEventView`——呈现归客户端所有，常驻列表从会话事件渲染，而非工具卡。
+
+## Alternatives considered
+
+- **把 todo 写入作为 surface 条目折叠进 `nodes`**——回放的窗口会渲染每一份已被取代的列表；该事件被刻意设计成非 surface 类型。
+- **面板硬编码进 `ConversationRoot`**——input-dock slot 出现之前的原始落点；dock 是本架构给"composer 上方常开横条"安排的家，硬编码绕开了 slot 注册表的 disposal 与定序。
+- **面板放进 details 列**——details slot 单占用且由选中驱动，生命周期不同于一条常开横条。
+- **host 计算的视图（一个 todo `ToolEventView`）**——呈现属于客户端；协议已在事件载荷里携带整份快照。
+
+## Consequences
+
+回放正确性由一条代码路径掌管：未来对窗口重建的任何改动都免费保持 todos 一致；fx-alpha 第 65 轮的 fixture（测试前置数据）加 assembled keyless snapshot（`apps/web/tests/todo-display.snapshot.ts`）在构建产物客户端全图上钉住整条链（行摘要与状态、dock 面板内容、折叠往返）。`todos` 是 `ConversationSnapshot` 的必填字段，所以 spec 里脚本化的 fake 必须带上它。TUI 面板未受改动（自动化专用的 ACP 桥接刻意不做 todo 呈现）；Web 各面渲染同一个事件，只新增一个协议字段，不新增事件类型。冷加载重建正是靠这个字段由 host 兜底：history 尾页附带 `todos`——全量 log 上最新一次 `todo/write` 的投影，独立于分页窗口计算（与 view 配对同一种 backscan 姿势）——因此重开会话时即使最后一次写入落在窗口之前，计划也照常恢复；该值跨往前翻页保留，之后的任何写入照常覆盖，而尾页响应不带投影时复位为空。
@@ -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
-2026-07-26-turndown-for-tool-web-html-markdown.md: 7f25e51bf6e6fc9313a880abee737bca80a472af
-2026-07-26-turndown-for-tool-web-html-markdown.zh.md: 3a59b08e13fd392e4f34ac543f32f5b4648f3c1c
+2026-07-27-user-message-icon-actions.md: 869e7a2518a3ec927c0689a10816a410dc5f0862
+2026-07-27-user-message-icon-actions.zh.md: 353e5ac765bb2fbab9932449cf2247768a1f412f
@@ -0,0 +1,27 @@
+# Agent Note: User-message IconActions under the bubble
+
+Status: implemented
+
+English | [中文](2026-07-27-user-message-icon-actions.zh.md)
+
+## Problem
+
+The chat user bubble had no under-bubble action chrome. The Harness design (figma `User_Bubble/message_container`) shows three IconActions — copy, branch in new chat, and edit — right-aligned under the bubble, matching the product action-bar pattern used elsewhere.
+
+## Decision
+
+`MessageItem` owns the actions for `kind: 'user'` only. Layout is a column (`align-items: flex-end`, 6px gap): bubble, then a 28px action row with 10px gaps and 28px circular icon buttons (`IconCopyOutline16`, `IconBranchOutline16`, `IconEditOutline16`). Tooltips carry Chinese labels. Actions stay visible by default; `@media (hover: hover)` hides them until the row is hovered or focus-within, so touch / `hover: none` devices keep discoverable controls (opacity alone still hit-tests).
+
+Copy writes the bubble's joined text blocks to the clipboard (`navigator.clipboard.writeText`, with an `execCommand` fallback). Branch and edit are present chrome with no handlers yet — they reserve the design seats without inventing session-fork or edit-resubmit behavior.
+
+Steering bubbles keep the badge-only form and do not show these actions.
+
+## Alternatives considered
+
+**Wire branch/edit to real session fork and draft-edit now.** Rejected for this change: those product flows are not specified; shipping inert buttons matches the requested scope and avoids half-built mutation paths.
+
+**Always hide with `opacity: 0` outside hover.** Rejected for touch: without `@media (hover: hover)`, idle opacity still hit-tests while looking empty. Hover-capable pointers keep the fade; others keep the actions visible.
+
+## Consequences
+
+User messages expose copy immediately; branch/edit remain clickable stubs until a later decision owns their behavior. Tests pin the three buttons, copy payload, and steering exclusion.
@@ -0,0 +1,27 @@
+# Agent Note: 用户消息气泡下方的 IconActions
+
+Status: implemented
+
+[English](2026-07-27-user-message-icon-actions.md) | 中文
+
+## 问题
+
+聊天用户气泡下方没有操作栏。Harness 设计稿（figma `User_Bubble/message_container`）在气泡下方右对齐展示三个 IconActions——复制、在新对话中分支、编辑——与产品其他位置使用的操作栏模式一致。
+
+## 决策
+
+仅当 `kind: 'user'` 时，`MessageItem` 拥有这些操作。布局为纵向列（`align-items: flex-end`，间距 6px）：先是气泡，再是高度 28px 的操作行；行内间距 10px，圆形图标按钮尺寸为 28px（`IconCopyOutline16`、`IconBranchOutline16`、`IconEditOutline16`）。Tooltip 承载中文标签。操作默认保持可见；`@media (hover: hover)` 下在悬停或 focus-within 前隐藏，以便触摸／`hover: none` 设备仍能发现控件（仅靠 opacity 仍会命中测试）。
+
+复制将气泡内拼接后的文本块写入剪贴板（`navigator.clipboard.writeText`，并以 `execCommand` 作为回退）。分支与编辑目前仅有外观、尚无处理函数——它们预留设计席位，但不发明会话 fork 或编辑重提交流程。
+
+steering（中途引导）气泡保持仅徽章形态，不展示这些操作。
+
+## 考虑过的替代方案
+
+**现在就把分支／编辑接到真实的会话 fork 与草稿编辑。**本次变更不予采纳：这些产品流程尚未定稿；交付无行为按钮符合请求范围，也避免半成品的变更路径。
+
+**在悬停外始终以 `opacity: 0` 隐藏。**因触摸不予采纳：若无 `@media (hover: hover)`，空闲 opacity 看起来空白但仍会命中测试。具备悬停能力的指针保留淡入；其他设备保持操作可见。
+
+## 后果
+
+用户消息立即可用复制；分支／编辑仍为可点击的占位，直至后续决策明确其行为。测试钉死三个按钮、复制载荷，以及对 steering 的排除。
@@ -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
-2026-07-02-bilingual-docs-and-pairing-gate.md: bebff600ca27763c04bdecea78ceb916542f7eca
-2026-07-02-bilingual-docs-and-pairing-gate.zh.md: 66c9b75b1bab47558bb63b7e97cf6d7c7610b0d5
+#   pnpm run verify-translation-pairing --write .agents/notes/implemented/process/2026-07-02-bilingual-docs-and-pairing-gate.md
+2026-07-02-bilingual-docs-and-pairing-gate.md: 8e9d0ab8653528517346b198c97814d8d8979440
+2026-07-02-bilingual-docs-and-pairing-gate.zh.md: cf37d56be0180cc742b0dabefeb1f8f3b54ba557
@@ -11,7 +11,7 @@ This repo's documentation corpus is read by people and agents inside and outside
 ## Decision

 - **Paired sibling files with equal authority.** A documentation pair is three sibling files: English `foo.md`, Chinese `foo.zh.md`, and a consistency record `foo.i18n.yaml`. Neither language is canonical — a document may be authored and reviewed Chinese-first and translated to English afterwards, or the reverse; what binds the pair is that both sides must say the same thing, and pairs merge whole (both languages plus the record, never one alone). Policy: [docs/i18n/README.md](../../../../docs/i18n/README.md); translation rules: [docs/i18n/translation-rules.md](../../../../docs/i18n/translation-rules.md); terminology source of truth: [docs/i18n/terminology.md](../../../../docs/i18n/terminology.md).
- **A sidecar record of both blob hashes makes consistency checkable.** `foo.i18n.yaml` holds the full git blob hash of each side as of the last confirmed-consistent state. An edit to either side without re-confirming the pair is then mechanically detectable as a pure content comparison — no history lookup — and the hashes are computable for files edited in the same PR, which a commit-hash record is not. Re-recording (`verify-translation-pairing --write`) produces a reviewable yaml diff: confirming consistency is an explicit, visible act in the PR.
+- **A sidecar record of both blob hashes makes consistency checkable.** `foo.i18n.yaml` holds the full git blob hash of each side as of the last confirmed-consistent state. An edit to either side without re-confirming the pair is then mechanically detectable as a pure content comparison — no history lookup — and the hashes are computable for files edited in the same PR, which a commit-hash record is not. Re-recording (`verify-translation-pairing --write <pair>`, which requires naming the confirmed pairs — bulk re-record is an explicit `--write --all`) produces a reviewable yaml diff: confirming consistency is an explicit, visible act in the PR.
 - **`verify-translation-pairing` joins `doc-sync`.** The gate ([scripts/verify-translation-pairing.ts](../../../../scripts/verify-translation-pairing.ts)) enforces: every discovered, non-excluded source has a complete pair; every existing pair is complete (all three files) and consistent (both hashes match, switcher links both ways, structural signatures identical); and excluded generated, instruction, or bilingual-by-construction files stay unpaired. [scripts/translation-pairing.manifest.json](../../../../scripts/translation-pairing.manifest.json) contains only explicit exclusions, so no requirement can bypass discovery and receive a weaker check. Source-oriented code gates consume a `.zh.md` fence sequence as a derivative only when its unsuffixed sibling has the same tracked fences in the same order with byte-identical bodies; an incomplete, reordered, reclassified, or changed sequence stays independent, so the owning code gate or pairing gate reports the mismatch.
 - **One corpus-wide requirement.** Every document in scope requires a complete pair from creation; the policy has no per-file rollout state, date cutoff, or README-specific class. README discovery covers every case-insensitive README basename outside vendored, dependency, and ignored build-output trees, including future top-level directories. A site-published pair uses `pairedPages()` so the root locale projects `.zh.md` and `/en/` projects `.md`; creating a counterpart alone does not publish it.
 - **Pairing records are metadata, not Cordis Loader configuration.** Cordis configuration discovery accepts actual `.cordis.yml` and `.cordis.yaml` files while excluding `*.i18n.yaml`, even when the document name contains `cordis`. This preserves validation of executable Loader entries without parsing translation hashes as configuration.
@@ -41,4 +41,4 @@ Paired sibling files with locale suffixes are the dominant Chinese big-tech conv
 - When the two sides disagree, no mechanical rule picks a winner — the PR review does. That is the price of equal authority, accepted deliberately: the alternative (a canonical language) forbids Chinese-first authoring.
 - Generated docs (`cordis-catalog/`, `tool-catalog/`, `module-graph.md`) are excluded for now; the planned follow-up is to teach their generators to emit Chinese alongside English, at which point they leave the exclusion list.
 - The exclusions-only manifest makes every current and future in-scope document mandatory through the same path. There is no explicit requirement, cutoff, or class entry that can fall outside discovery while appearing enforced.
- The recorded hashes double as the update tool (`git cat-file -p <hash>` recovers either side's last-confirmed text for a minimal diff-based update), so re-translation of whole files is never forced by the mechanism.
+- The recorded hashes double as the update tool: [gen-translation-brief](2026-07-26-briefed-minimal-translation-updates.md) recovers either side's last-confirmed text from them and assembles the minimal-update briefing, so re-translation of whole files is never forced by the mechanism.
@@ -11,7 +11,7 @@ Status: implemented
 ## 决策

 - **配对兄弟文件，两种语言同权。** 一对文档由三个兄弟文件组成：英文 `foo.md`、中文 `foo.zh.md`，以及一份一致性记录 `foo.i18n.yaml`。没有哪种语言是正典：一篇文档可以先用中文撰写和评审、之后再译成英文，反之亦可；约束配对的是：两侧必须表达相同的内容，且配对整体合并（两种语言加记录，绝不单独落一侧）。政策见 [docs/i18n/README.md](../../../../docs/i18n/README.md)；翻译规则见 [docs/i18n/translation-rules.md](../../../../docs/i18n/translation-rules.md)；术语真源见 [docs/i18n/terminology.md](../../../../docs/i18n/terminology.md)。
- **伴随记录保存两侧 blob hash，使一致性可检查。** `foo.i18n.yaml` 保存两侧文件在上一次确认一致时各自的完整 git blob hash。此后修改了任一侧而未重新确认配对，都能被机械检测出来（纯内容比较，无需查询历史），而且同一个 PR（Pull Request）内改动的文件也能计算出 hash，commit hash 式的记录做不到这一点。重新记录（`verify-translation-pairing --write`）会产生一份可评审的 yaml diff：确认一致在 PR 中是一个显式、可见的动作。
+- **伴随记录保存两侧 blob hash，使一致性可检查。** `foo.i18n.yaml` 保存两侧文件在上一次确认一致时各自的完整 git blob hash。此后修改了任一侧而未重新确认配对，都能被机械检测出来（纯内容比较，无需查询历史），而且同一个 PR（Pull Request）内改动的文件也能计算出 hash，commit hash 式的记录做不到这一点。重新记录（`verify-translation-pairing --write <pair>`，要求点名所确认的配对；批量重新记录是显式的 `--write --all`）会产生一份可评审的 yaml diff：确认一致在 PR 中是一个显式、可见的动作。
 - **`verify-translation-pairing` 加入 `doc-sync`。** 门禁（[scripts/verify-translation-pairing.ts](../../../../scripts/verify-translation-pairing.ts)）强制执行以下规则：每个已发现且未排除的源文档都有完整配对；每个现有配对都完整（三个文件齐全）且一致（两个 hash 匹配、切换行双向互链、结构签名一致）；被排除的生成文档、指令文档或本身即双语的文档不得配对。[scripts/translation-pairing.manifest.json](../../../../scripts/translation-pairing.manifest.json) 只包含显式排除项，因此任何要求都无法绕过发现流程而接受较弱的检查。只有当 `.zh.md` 围栏序列与其无后缀兄弟文件拥有顺序相同、正文按字节一致的同一组受跟踪围栏时，面向源码的代码门禁才会将其作为派生内容消费；不完整、顺序变更、重分类或已改动的序列仍会独立受检，因此由其所属的代码门禁或配对门禁报告不匹配。
 - **全语料统一要求。** 范围内的每篇文档从创建起就必须有完整配对；政策没有逐文件推进状态、日期分界或 README 专用类别。README 发现会覆盖 vendor 源码、依赖目录与被忽略的构建产物目录之外所有文件名不区分大小写匹配 README 的文件，包括今后新增的顶层目录。发布到文档站的配对使用 `pairedPages()`，由根 locale 投影 `.zh.md`，由 `/en/` 投影 `.md`；仅创建对侧文件并不会发布它。
 - **配对记录是元数据，而不是 Cordis Loader 配置。** Cordis 配置发现会接受实际的 `.cordis.yml` 和 `.cordis.yaml` 文件，同时排除 `*.i18n.yaml`，即使文档名中包含 `cordis` 也不例外。这样既能继续校验可执行的 Loader 配置项，又不会把翻译 hash 当作配置来解析。
@@ -41,4 +41,4 @@ Status: implemented
 - 两侧说法冲突时，没有机械规则裁决谁赢，由 PR 评审裁决。这是同权的代价，且是有意接受的：另一个选项（正典语言）会禁止中文先行撰写。
 - 生成文档（`cordis-catalog/`、`tool-catalog/`、`module-graph.md`）暂被排除；计划中的后续工作是让生成器在输出英文的同时输出中文，届时将这些文件移出排除清单。
 - 只含排除项的 manifest 通过同一路径，要求当前及今后纳入范围的每篇文档都必须配对。不存在显式要求、分界或类别条目可以落在发现范围之外，却看似已经强制执行。
- 记录的 hash 兼作更新工具（`git cat-file -p <hash>` 能还原任一侧上次确认的文本，用于基于 diff 的最小更新），因此这套机制从不强迫整篇重译。
+- 记录的 hash 兼作更新工具：[gen-translation-brief](2026-07-26-briefed-minimal-translation-updates.md) 会从中还原任一侧上次确认的文本并组装最小更新简报，因此这套机制从不强迫整篇重译。
@@ -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
-2026-07-21-serial-cross-platform-ci-reference.md: 5433d2c51831ce61d06a16ee0b0ed982911f9218
-2026-07-21-serial-cross-platform-ci-reference.zh.md: 041d53d13e14354c995e4b65defce94a97646b0a
+#   pnpm run verify-translation-pairing --write .agents/notes/implemented/process/2026-07-21-serial-cross-platform-ci-reference.md
+2026-07-21-serial-cross-platform-ci-reference.md: 5eac1bc1c47c7309942b5615bc98a7fed893f346
+2026-07-21-serial-cross-platform-ci-reference.zh.md: 35fb761023fe7be081bf7d9591a53ed98b6e3abc
@@ -24,7 +24,7 @@ The macOS reference runs the ordinary Vitest project in forked processes. Node 2

 Master reference jobs are diagnostic and do not participate in the pull request's required `all checks passed` result. A pull request runs only its required jobs; a master push runs only the three serial references. Performance is evaluated from completed hosted-job timestamps and reported as a measurement; it is not encoded as a `timeout-minutes` value.

-The portable reference uses GitHub's standard `ubuntu-latest`, `macos-latest`, and `windows-2025` labels. Required pull-request jobs use the same portable Linux and Windows capacity under the [required-CI decision](2026-07-23-portable-required-pull-request-ci.md). Higher-core hosted runners remain manual benchmarks because a correctness path must remain runnable without repository-external runner configuration.
+The portable reference uses GitHub's standard `ubuntu-latest`, `macos-latest`, and `windows-2025` labels; `serial / windows` is the one remaining native-Windows job, the complete-kernel oracle behind the Wine-hosted pull-request lane ([Wine lane decision](2026-07-27-wine-windows-gates-experiment.md)). Required pull-request jobs use portable standard capacity under the [required-CI decision](2026-07-23-portable-required-pull-request-ci.md). Higher-core hosted runners remain manual benchmarks because a correctness path must remain runnable without repository-external runner configuration.

 ## Alternatives considered

@@ -24,7 +24,7 @@ macOS 参考流程使用 fork 进程运行常规 Vitest 项目。macOS arm64 上

 master 分支的参考作业仅用于诊断，不参与拉取请求所要求的 `all checks passed` 结果。拉取请求只运行其必需作业；向 master 推送时只运行三个串行参考作业。系统根据已完成托管作业的时间戳评估性能，并将其报告为测量结果，而不是写成 `timeout-minutes` 值。

-可移植的参考流程使用 GitHub 标准的 `ubuntu-latest`、`macos-latest` 和 `windows-2025` 标签。依据[必需 CI 决策](2026-07-23-portable-required-pull-request-ci.md)，拉取请求必需作业使用相同的可移植 Linux 和 Windows 容量。更高核心数的托管运行器仍仅用于手动基准测试，因为正确性路径必须无需仓库外部的运行器配置即可运行。
+可移植的参考流程使用 GitHub 标准的 `ubuntu-latest`、`macos-latest` 和 `windows-2025` 标签；`serial / windows` 是仅存的原生 Windows 作业，是 Wine 托管拉取请求通道背后的完整内核标尺（[Wine 通道决策](2026-07-27-wine-windows-gates-experiment.md)）。依据[必需 CI 决策](2026-07-23-portable-required-pull-request-ci.md)，拉取请求必需作业使用可移植的标准容量。更高核心数的托管运行器仍仅用于手动基准测试，因为正确性路径必须无需仓库外部的运行器配置即可运行。

 ## 曾考虑的替代方案

@@ -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
-2026-07-23-portable-required-pull-request-ci.md: d1002c7d9db7cd8bbed3bdfda8a773a4b124bf16
-2026-07-23-portable-required-pull-request-ci.zh.md: fedfc6b9c982ace5ece430c52db23c22ec5119d4
+#   pnpm run verify-translation-pairing --write .agents/notes/implemented/process/2026-07-23-portable-required-pull-request-ci.md
+2026-07-23-portable-required-pull-request-ci.md: 1a6939e8386e381cba114a7be71993a644457a45
+2026-07-23-portable-required-pull-request-ci.zh.md: cf0af769f9e740a2c9285caf4be05023371578d9
@@ -12,9 +12,9 @@ Billing health, a runner definition's `Ready` state, and a large autoscaling cei

 ## Decision

-[CI](../../../../.github/workflows/ci.yml) runs the required primary Node 24 jobs, plus the stable `all checks passed` aggregate, on repo-restricted enterprise 32-core pools. The aggregate performs no checkout or repository gate, but sharing the enterprise pool prevents the required verdict from introducing a separate standard-hosted billing dependency after its substantive jobs have already succeeded. The required Windows job runs on standard `windows-2025` with single-worker bounds, keeping the complete Windows contract independent of enterprise Windows allocation. Standard `ubuntu-latest` jobs retain Node 22.19, Node 26, and Python SDK compatibility, and `master` runs complete serial Linux, macOS, and Windows references. Those standard-hosted jobs keep the portable execution boundary observable without duplicating the primary inventory on every pull request.
+[CI](../../../../.github/workflows/ci.yml) runs the required primary Node 24 jobs, plus the stable `all checks passed` aggregate, on repo-restricted enterprise 32-core pools. The aggregate performs no checkout or repository gate, but sharing the enterprise pool prevents the required verdict from introducing a separate standard-hosted billing dependency after its substantive jobs have already succeeded. The required Windows job runs Windows Node under Wine on standard `ubuntu-latest` for the blocking surfaces ([Wine lane decision](2026-07-27-wine-windows-gates-experiment.md)), keeping the pull-request Windows contract independent of any Windows runner allocation; the complete native-kernel Windows inventory lives in the master serial reference. Standard `ubuntu-latest` jobs retain Node 22.19, Node 26, and Python SDK compatibility, and `master` runs complete serial Linux, macOS, and Windows references. Those standard-hosted jobs keep the portable execution boundary observable without duplicating the primary inventory on every pull request.

-The two Linux primary jobs, Node compatibility, Python SDK, and `windows node 24 / complete` remain dependencies of `all checks passed`; branch protection continues to require `e2e` and `all checks passed`. There is no automatic fallback when a remaining enterprise Linux label cannot allocate: the standard jobs continue to report their own contracts, but they cannot manufacture the missing required result.
+The two Linux primary jobs, Node compatibility, Python SDK, and `windows node 24 / wine blocking` remain dependencies of `all checks passed`; branch protection continues to require `e2e` and `all checks passed`. There is no automatic fallback when a remaining enterprise Linux label cannot allocate: the standard jobs continue to report their own contracts, but they cannot manufacture the missing required result.

 The [larger-runner decision](2026-07-22-evidence-based-larger-hosted-runners.md) owns the current primary topology and its measurements. The [serial cross-platform reference](2026-07-21-serial-cross-platform-ci-reference.md) remains the independent standard-hosted completeness check, and the manual larger-runner suites retain size comparisons without expanding the ordinary required matrix.

@@ -30,6 +30,6 @@ The [larger-runner decision](2026-07-22-evidence-based-larger-hosted-runners.md)

 ## Consequences

-Ordinary pull requests spend enterprise capacity on the Linux critical path while standard Windows trades longer runtime for independent allocation. A live exact-head run proves the same commands that branch protection consumes; queue delay is reported separately from each job's `startedAt` to `completedAt` execution interval.
+Ordinary pull requests spend enterprise capacity on the Linux critical path while the Wine-hosted Windows job keeps its verdict on standard Linux allocation. A live exact-head run proves the same commands that branch protection consumes; queue delay is reported separately from each job's `startedAt` to `completedAt` execution interval.

 Standard compatibility and required Windows jobs remain useful when enterprise allocation is degraded, but they do not make a blocked required Linux job or aggregate green. Recovering Linux availability may require restoring the complete standard-hosted topology; changing a pool definition's status alone is insufficient evidence that it can receive work.
@@ -12,9 +12,9 @@ Status: implemented

 ## 决策

-[CI](../../../../.github/workflows/ci.yml) 在仅限本仓库使用的企业级 32 核运行器池上运行必需的主 Node 24 作业，以及稳定的 `all checks passed` 聚合流程。该聚合流程不执行代码检出或仓库门禁；但让它与所依赖的实质性作业共用企业级运行器池，可以避免这些作业已经成功后，必需判定结果又引入一项单独的标准托管计费依赖。必需的 Windows 作业在标准 `windows-2025` 上运行，并采用单工作线程上限，使完整的 Windows 契约不依赖企业级 Windows 运行器分配。标准 `ubuntu-latest` 作业保留 Node 22.19、Node 26 和 Python SDK 兼容性，`master` 则运行完整的 Linux、macOS 和 Windows 串行参考流程。这些标准托管作业让可移植执行边界保持可观测，而不必在每个拉取请求中重复主清单。
+[CI](../../../../.github/workflows/ci.yml) 在仅限本仓库使用的企业级 32 核运行器池上运行必需的主 Node 24 作业，以及稳定的 `all checks passed` 聚合流程。该聚合流程不执行代码检出或仓库门禁；但让它与所依赖的实质性作业共用企业级运行器池，可以避免这些作业已经成功后，必需判定结果又引入一项单独的标准托管计费依赖。必需的 Windows 作业在标准 `ubuntu-latest` 上通过 Wine 运行 Windows Node 以覆盖阻断表面（[Wine 通道决策](2026-07-27-wine-windows-gates-experiment.md)），使拉取请求的 Windows 契约不依赖任何 Windows 运行器分配；完整的原生内核 Windows 清单归 master 串行参考流程所有。标准 `ubuntu-latest` 作业保留 Node 22.19、Node 26 和 Python SDK 兼容性，`master` 则运行完整的 Linux、macOS 和 Windows 串行参考流程。这些标准托管作业让可移植执行边界保持可观测，而不必在每个拉取请求中重复主清单。

-两项 Linux 主作业、Node 兼容性、Python SDK 和 `windows node 24 / complete` 继续作为 `all checks passed` 的依赖项；分支保护继续要求 `e2e` 和 `all checks passed`。剩余的企业级 Linux 运行器标签无法分配运行器时没有自动后备机制：标准作业会继续报告各自的契约，但无法产出缺失的必需结果。
+两项 Linux 主作业、Node 兼容性、Python SDK 和 `windows node 24 / wine blocking` 继续作为 `all checks passed` 的依赖项；分支保护继续要求 `e2e` 和 `all checks passed`。剩余的企业级 Linux 运行器标签无法分配运行器时没有自动后备机制：标准作业会继续报告各自的契约，但无法产出缺失的必需结果。

 当前主拓扑及其测量结果由[大型运行器决策](2026-07-22-evidence-based-larger-hosted-runners.md)记录。[跨平台串行参考流程](2026-07-21-serial-cross-platform-ci-reference.md)继续作为独立的标准托管完整性检查，手动大型运行器套件则保留规格比较，同时不扩大普通必需矩阵。

@@ -30,6 +30,6 @@ Status: implemented

 ## 后果

-普通拉取请求会将企业级运行器容量用于 Linux 关键路径，而标准托管 Windows 作业则以更长的运行时间换取不依赖企业池的运行器分配。一次实际的分支头精确运行能够证明分支保护使用的同一组命令；排队延迟与每个作业从 `startedAt` 到 `completedAt` 的执行区间分开报告。
+普通拉取请求会将企业级运行器容量用于 Linux 关键路径，而 Wine 托管的 Windows 作业让其判定保持在标准 Linux 运行器分配上。一次实际的分支头精确运行能够证明分支保护使用的同一组命令；排队延迟与每个作业从 `startedAt` 到 `completedAt` 的执行区间分开报告。

 企业级运行器分配能力下降时，标准兼容性作业和必需的 Windows 作业仍能提供有用证据，但无法让受阻的必需 Linux 作业或聚合流程变绿。恢复 Linux 可用性时，可能需要恢复完整的标准托管拓扑；仅改变运行器池定义的状态，不足以证明它可以接收作业。
@@ -0,0 +1,6 @@
+# Bilingual-pair consistency record (docs/i18n/README.md): the git blob hash of each
+# side as of the last confirmed-consistent state. Both languages carry equal authority;
+# after editing either side, bring the other along and re-record with:
+#   pnpm run verify-translation-pairing --write .agents/notes/implemented/process/2026-07-26-briefed-minimal-translation-updates.md
+2026-07-26-briefed-minimal-translation-updates.md: 63f25d5c36caecba435e9192534e0b494e1e0105
+2026-07-26-briefed-minimal-translation-updates.zh.md: 17cf2f895b187fb794e691c2b14d6e0bff78363d
@@ -0,0 +1,50 @@
+# Agent Note: Briefed minimal translation updates
+
+Status: implemented
+
+English | [中文](2026-07-26-briefed-minimal-translation-updates.zh.md)
+
+## Problem
+
+The [bilingual pairing contract](2026-07-02-bilingual-docs-and-pairing-gate.md) already prescribed minimal counterpart updates — diff the edited side against its last-confirmed state, patch the counterpart, never re-translate — but the committed workflow made every update pay whole-document overheads. The translating subagent loaded the full guidance corpus (skill, pairing contract, translation rules, the 192-line terminology table, style samples, prose standard) before touching a two-line diff; it re-derived the last-confirmed diff by hand through `git cat-file`; and each iteration re-ran the corpus-wide pairing gate, which parses every pair in the tree to validate one. A small English prose edit routinely cost tens of times its proportional share of tokens and minutes, which taxes exactly the behavior the contract wants — bringing the counterpart along in the same PR.
+
+## Decision
+
+Pair updates run on a generated briefing instead of the guidance corpus; only new pairs still run the whole-document workflow, which is unchanged.
+
+- **`pnpm run gen-translation-brief [--apply] [pair...]`** ([scripts/gen-translation-brief.ts](../../../../scripts/gen-translation-brief.ts), assembly in [scripts/translation-brief.ts](../../../../scripts/translation-brief.ts)) prints, per out-of-sync pair, the authored side's diff from its recorded last-confirmed blob to the working tree plus the change mapped at the narrowest safely aligned granularity, deterministically widening on mapping failure: a change confined to the pair's byte-identical code fences is computed outright (`--apply` splices it into the counterpart and validates the result against the pairing gate's structural signature before writing); otherwise changed Markdown units (headings, paragraphs, table rows, list items, code fences, block quotes, HTML blocks, thematic breaks, link definitions — matched by container-scoped kind sequences) each carry their last-confirmed source, current source, and current counterpart text with line numbers; units that do not align fall back to depth-matched heading sections; and when sections do not align either, or both sides drifted, the briefing says so and withholds the mapping instead of guessing. Terminology rows are matched against the changed spans only (word-boundary English matching with plural inflections), and for Chinese targets the briefing tracks each relevant term's document-wide first occurrence — when an edit moves it, the vacated and receiving spans join the briefing with an explanatory note, since the 首次出现 annotation must move with it. The unit mapping, code splice, and first-occurrence mechanics adopt the planner design from the [incremental prompt-pipeline work](https://github.com/deepseek-harness/deepseek-harness/pull/684), whose provider-backed bake-off independently validated the same scope ladder for the automated pipeline. The briefing is the translator's whole working set; the full sources of truth remain the escalation path for decisions the briefing cannot answer.
+- **The update path in [dsh-translate-docs](../../../skills/dsh-translate-docs/SKILL.md)** consumes the briefing: mechanical (code-fence-only) changes are applied with `--apply`, no subagent; prose diffs go to a subagent whose prompt is the briefing, not the corpus; verification is clause-by-clause on the changed spans, not the whole document.
+- **The pairing gate takes pair arguments.** `verify-translation-pairing [pair...]` checks just the named pairs (any of a pair's three files, or the bare stem, names it); the corpus-wide sweep remains the no-argument form that `doc-sync` and CI run. `--write` now requires naming the confirmed pairs — bare `--write` refuses, and re-recording everything is an explicit `--write --all` — because the old bare form silently blessed every drifted pair in the tree, including ones the caller never looked at, and a prose-only drift would then stay green forever. Each record's comment names its own scoped command.
+
+## Benchmark
+
+The decision followed a controlled replay of ten real pair updates from this repo's history (July 2026; 1-64 changed English lines each, READMEs, RFCs, Agent Notes, and user docs). Each example was reconstructed in a scratch repo at its true last-confirmed state with the English edit uncommitted, then run through competing workflows with fresh subagents: the status-quo corpus-loading path, the briefed path, a no-guidance control, whole-document re-translation, the briefed path on a small model, and a three-pairs-per-agent batch. Outputs were gated mechanically and scored blind by judges who also received the real historical update and the untouched stale counterpart as controls.
+
+- The briefed path matched the status-quo path on judged faithfulness, preservation, and fluency — both at or above the real historical updates — while spending roughly a third of the tokens and wall clock on the stall-free examples (medians across all ten: 276k vs 595k relative token-cost units, 14 vs 32 turns).
+- Re-translation was confirmed harmful, not merely wasteful: judged preservation collapsed (4.4/10 vs 9.8) because it discards reviewed phrasing, it drifted established terminology the update arms kept (the counterpart's own text carries the renderings), and it was the most expensive arm.
+- The no-guidance control held quality too — the binding context for an update is the diff plus the counterpart's own reviewed text, not the corpus — but the briefing buys a fixed working set, inline terminology, and the both-sides-drifted warning at negligible cost over it.
+- On the briefing, a small model performed at parity with the large one, so the update path no longer assumes a frontier translator.
+- Batching three pairs into one subagent showed no reliable saving over three briefed runs and couples unrelated failures; it was rejected.
+
+A second head-to-head replay on the same ten examples compared this note's shipped briefing against its earlier section-only form (no unit tier, no computed mechanical path, counterpart-only context, no first-occurrence tracking). Prose quality and cost were at parity — pairwise blind verdicts split with only stylistic margins — and the shipped form won on two objective outcomes: the two code-fence-only examples were completed byte-identical to the human-reviewed historical updates in under a second with no model tokens, and on the example whose edit moved a term's document-wide first occurrence, the shipped briefing's flagged move reproduced the human-reviewed gloss relocation while the section-only form left a 首次出现 violation for review to catch.
+
+## Alternatives considered
+
+- **Keep the workflow, just scope the gate** — the gate scan was the smaller cost; the corpus loads and archaeology dominated. Scoping alone would have left the ~3x overhead in place.
+- **Whole-document re-translation as the update path** (what a naive pipeline does) — rejected on benchmark evidence: preservation collapse, terminology drift, highest cost. The contract's minimal-update rule survives with data behind it.
+- **Batching several pairs per subagent** — rejected: no measured saving (briefings already deduplicate the fixed content), and one stalled or confused pair holds the others hostage.
+- **Per-paragraph translation-memory records in the sidecar** (segment hashes instead of whole-file hashes) — rejected: paragraph boundaries may legitimately differ across the pair, either side can be authored first, and the records would bloat and conflict in merges. Span mapping computed on demand from the existing whole-file hashes recovers the same alignment when it is trustworthy and says so when it is not.
+- **An update mode in the automated prompt pipeline (prompt-v5)** — deferred, not designed here: nothing drives [scripts/translation-prompt.ts](../../../../scripts/translation-prompt.ts) today, and the agent path was the live cost center. The pipeline keeps its whole-document v4 contract until it has a consumer.
+
+## Consequences
+
+- A small prose edit's counterpart update now costs a briefing generation plus one small focused task — no corpus reads, no archaeology, no corpus-wide scans inside the loop — and the same PR obligation holds; the cheap path and the correct path point the same way.
+- The briefing generator is a second consumer of the consistency records: recorded blob hashes now also drive diff recovery and section mapping, strengthening the incentive to keep records honest.
+- `--write` without arguments no longer works; muscle-memory callers must name pairs or pass `--all`. That is the point — the bulk bless is now a visible, deliberate act.
+- Scoped checks mean an update loop can be green while an unrelated pair elsewhere is red; the corpus-wide check in `doc-sync`/CI still owns the tree-level invariant.
+- Span mapping trusts an alignment only when the kind sequences match across the last-confirmed source, current source, and current counterpart; a mapping failure widens deterministically (units → sections → whole document) rather than guessing, so a restructured document gets an explicit "locate the regions yourself" briefing, never a wrong map.
+- A first-occurrence move can enlarge a briefing beyond the directly changed spans; that cost is an explicit consequence of the 首次出现 contract, not an alignment heuristic.
+
+## Testing
+
+[scripts/translation-brief.spec.ts](../../../../scripts/translation-brief.spec.ts) pins unit and section span extraction (container-scoped kinds, depth-only section alignment so translated heading text still maps, preamble), alignment and changed-index detection, the mechanical code splice and each of its refusal conditions, terminology row matching in both directions with word-boundary and plural-inflection discipline, first-occurrence movement tracking, fence escalation, and the rendered briefing's contract (unit bundles with three-way context, mechanical/sections/document scopes, per-direction digests, scoped finish commands). [scripts/translation-pairing.spec.ts](../../../../scripts/translation-pairing.spec.ts) pins argument normalization (any pair file or bare stem to the anchor) and the CLI matrix: scoped check, bare `--write` refusal, `--write <pair>`, `--write --all`, `--list` exclusivity, unknown flags.
@@ -0,0 +1,50 @@
+# Agent Note: 基于简报的最小化翻译更新
+
+Status: implemented
+
+[English](2026-07-26-briefed-minimal-translation-updates.md) | 中文
+
+## 问题
+
+[双语配对契约](2026-07-02-bilingual-docs-and-pairing-gate.md)早已规定对侧文件按最小幅度更新：把被改的一侧与其上次确认状态做 diff，据此修补对侧文件，绝不整篇重译；但仓库内置的工作流让每次更新都付出整篇文档级别的开销。负责翻译的 subagent 在动手处理一个两行的 diff 之前，要先加载完整的指导语料（guidance corpus），即 skill（技能）、配对契约、翻译规则、192 行的术语表、语体样例与行文标准；要通过 `git cat-file` 手工重新推导上次确认状态以来的 diff；每轮迭代还要重跑全语料配对门禁，而该门禁为校验一个配对要解析整棵树里的每一个配对。一次小的英文行文修改，动辄花掉数十倍于其应得份额的 token 用量与分钟数，被惩罚的恰恰是契约想要的行为：在同一个 PR（Pull Request）里把对侧文件一并带上。
+
+## 决策
+
+配对更新基于生成的简报（briefing）运行，而非基于指导语料；只有新建配对仍走整篇文档工作流，后者保持不变。
+
+- **`pnpm run gen-translation-brief [--apply] [pair...]`**（[scripts/gen-translation-brief.ts](../../../../scripts/gen-translation-brief.ts)，组装逻辑在 [scripts/translation-brief.ts](../../../../scripts/translation-brief.ts)）针对每个失去同步的配对，打印被改一侧从其记录在案的上次确认 blob 到当前工作区的 diff，并附上以能安全对齐的最窄粒度映射的这次改动，映射失败时粒度确定性地逐级放宽：仅落在配对中逐字节一致的围栏代码块内的改动会直接算出（`--apply` 会把它拼接进对侧文件，并在写入前用配对门禁的结构签名校验所得结果）；否则，每个有改动的 Markdown 单元（标题、段落、表格行、列表项、围栏代码块、块引用、HTML 块、分隔线、链接定义；匹配依据是以容器为作用域的种类序列）都带上各自的上次确认源文、当前源文与当前对侧文本及行号；无法对齐的单元回退到按深度匹配的标题章节；当章节也无法对齐或两侧同时漂移时，简报会明说这一点并省略映射，而不是靠猜。术语表行只与改动块匹配（英文术语按词边界匹配，含复数变形）；当目标侧是中文时，简报还会跟踪每个相关术语在整篇文档中的首次出现：一旦某次编辑使其移位，腾出的与接收的两处区间就会附一条解释性说明加入简报，因为「首次出现」括注必须随之移动。单元映射、代码拼接与首次出现机制采纳了[增量提示词流水线工作](https://github.com/deepseek-harness/deepseek-harness/pull/684)中的规划器设计；该项工作中接入提供方的对比评测，已为自动流水线独立验证了同一套范围阶梯。简报就是译者的全部工作集；简报回答不了的决策，仍以完整的真源文档作为升级求证路径。
+- **[dsh-translate-docs](../../../skills/dsh-translate-docs/SKILL.md) 中的更新路径**消费这份简报：机械类改动（只涉及围栏代码块）用 `--apply` 应用，不动用 subagent；行文类 diff 交给 subagent，其提示词就是简报本身，而非指导语料；核验只对改动块逐句进行，不覆盖整篇文档。
+- **配对门禁接受配对参数。**`verify-translation-pairing [pair...]` 只检查被点名的配对（配对三个文件中的任意一个，或其裸词干，都能指代该配对）；全语料扫描仍是 `doc-sync`（文档同步门禁）与 CI 运行的无参数形式。`--write` 现在要求点名已确认的配对：裸 `--write` 会拒绝执行，重新记录全部配对必须显式写 `--write --all`；原因是旧的裸形式会默默为树中每一个漂移的配对背书，包括调用者从未看过的那些，纯行文层面的漂移于是可以永远保持绿灯。每份记录的注释都写明针对该配对自身的按对命令。
+
+## 基准测试
+
+该决策来自对本仓库历史上十次真实配对更新的受控回放（2026 年 7 月；每例改动 1 到 64 行英文，涵盖 README、RFC、Agent Note（agent 决策记录）与用户文档）。每个样例都在临时仓库中重建到其真实的上次确认状态，英文改动保持未提交，再用全新的 subagent 分别跑过相互竞争的各条工作流：维持现状的语料加载路径、简报路径、无指导对照组、整篇重译、小模型上的简报路径，以及每个 agent（智能体）一次处理三对文档的批量方案。产出先经机械门禁把关，再由评委盲评打分；评委还同时收到真实的历史更新与原样未动的陈旧对侧文件作为对照。
+
+- 简报路径在盲评的忠实性、保留度与流畅度上与现状路径打平（两者都达到或超过真实历史更新的水平），而在未发生停滞的样例上只花费约三分之一的 token 用量与墙钟时间（全部十例的中位数：相对 token 成本单位 276k 对 595k，轮次数 14 对 32）。
+- 整篇重译被证实有害，而不只是浪费：它丢弃经评审的措辞，盲评保留度因此崩塌（4.4/10 对 9.8）；它还使各更新组保持住的既定术语发生漂移（既定译法本就写在对侧文件自身的正文里）；而且它是成本最高的一组。
+- 无指导对照组的质量同样立得住（对一次更新有约束力的上下文，是 diff 加上对侧文件自身经评审的正文，而非指导语料），但简报以几乎可忽略的额外成本，换来固定的工作集、内联的术语，以及两侧同时漂移的警告。
+- 以简报为输入，小模型的表现与大模型持平，因此更新路径不再假定翻译必须由前沿模型完成。
+- 把三对文档合并给同一个 subagent，相比三次各自带简报的运行没有可靠的节省，还把互不相关的失败耦合在一起；该方案被否决。
+
+在同样这十个样例上进行的第二次正面对比回放，把本文最终交付的简报与其早前仅按章节的形态（没有单元层级、没有直接算出的机械路径、上下文只含对侧文件、不跟踪首次出现）相对照。行文质量与成本两相持平（两两盲评裁定各有胜负，差距仅在文风），而最终交付的形态在两项客观结果上胜出：两个只涉及围栏代码块的样例在一秒之内完成且不消耗任何模型 token，产出与经人工评审的历史更新逐字节一致；而在那个编辑使某术语在整篇文档中的首次出现发生移位的样例上，最终交付的简报所标记的移位复现了经人工评审的括注迁移，仅按章节的形态则留下一处「首次出现」违例，留待评审去捕捉。
+
+## 曾考虑的替代方案
+
+- **保留原工作流，只让门禁支持按对检查**：门禁扫描本是较小的开销，大头在语料加载与翻查历史。只收窄检查范围，约 3 倍的开销仍会原地保留。
+- **把整篇重译作为更新路径**（朴素流水线的做法）：依据基准测试证据否决，理由是保留度崩塌、术语漂移、成本最高。契约的最小更新规则得以延续，且从此有数据支撑。
+- **每个 subagent 批量处理多对文档**：否决。没有实测出节省（简报本身已对固定内容做了去重），而且一对文档停滞或陷入混乱会把其余配对一并拖住。
+- **在伴随记录中保存逐段的翻译记忆条目**（用分段 hash 取代整文件 hash）：否决。配对两侧的段落边界可以合理地不同，任一侧都可能先撰写，这类条目还会不断膨胀并在合并时产生冲突。基于现有整文件 hash 按需计算的区间映射，在对齐可信时能恢复同样的对齐关系，不可信时会明确说明。
+- **给自动提示词流水线加一个更新模式（prompt-v5）**：推迟，本文不做设计。今天没有任何调用方在驱动 [scripts/translation-prompt.ts](../../../../scripts/translation-prompt.ts)，实际的成本中心是 agent 路径。流水线在拥有消费方之前，维持其整篇文档的 v4 契约。
+
+## 后果
+
+- 一次小的行文修改，其对侧更新如今只需生成一份简报，外加一个小而聚焦的任务（不读指导语料、不翻查历史、循环内不做全语料扫描），同一 PR 内完成更新的义务保持不变；低成本的路径与正确的路径指向同一个方向。
+- 简报生成器成为一致性记录的第二个消费方：记录的 blob hash 如今还驱动 diff 还原与章节映射，这进一步强化了如实维护记录的动机。
+- 不带参数的 `--write` 不再可用；靠肌肉记忆的调用者必须点名配对或传 `--all`。这正是目的所在：批量背书如今是一个可见的、有意为之的动作。
+- 按对检查意味着一个更新循环可以在别处某个无关配对处于红灯时自己保持绿灯；`doc-sync`/CI 中的全语料检查仍然承载树级不变式。
+- 区间映射只在上次确认源文、当前源文与当前对侧文本三方的种类序列一致时才信任一处对齐；映射失败时粒度确定性地逐级放宽（单元 → 章节 → 整篇文档）而不是靠猜，因此被重构过的文档拿到的是一份明确写着「请自行定位相关区域」的简报，绝不会是一张错误的地图。
+- 「首次出现」的一次移位可能让简报扩大到直接改动块之外；这一成本是「首次出现」契约的明确后果，而非对齐启发式。
+
+## 测试
+
+[scripts/translation-brief.spec.ts](../../../../scripts/translation-brief.spec.ts) 固定单元与章节的区间提取（以容器为作用域的种类、只按深度对齐章节从而让已翻译的标题文字仍能映射、首个标题前的序言）、对齐与改动索引检测、机械代码拼接及其每一个拒绝条件、带词边界与复数变形约束的双向术语行匹配、首次出现移位跟踪、围栏升级，以及渲染后简报的契约（带三方上下文的单元条目、机械／章节／整篇文档三种范围、分方向的规则摘要、按对的收尾命令）。[scripts/translation-pairing.spec.ts](../../../../scripts/translation-pairing.spec.ts) 固定参数归一化（配对的任一文件或裸词干都归一到锚点）与 CLI（命令行界面）用例矩阵：按对检查、裸 `--write` 拒绝执行、`--write <pair>`、`--write --all`、`--list` 的互斥性、未知标志。
@@ -0,0 +1,6 @@
+# Bilingual-pair consistency record (docs/i18n/README.md): the git blob hash of each
+# side as of the last confirmed-consistent state. Both languages carry equal authority;
+# after editing either side, bring the other along and re-record with:
+#   pnpm run verify-translation-pairing --write .agents/notes/implemented/process/2026-07-27-wine-windows-gates-experiment.md
+2026-07-27-wine-windows-gates-experiment.md: aab8aecdfca06c1f15641044a071015f543a84b6
+2026-07-27-wine-windows-gates-experiment.zh.md: 5239b185e1e0c63aa626ee3f20f3f298c0c8579d
@@ -0,0 +1,45 @@
+# Agent Note: Wine-run Windows blocking gates on Linux runners
+
+Status: implemented
+
+English | [中文](2026-07-27-wine-windows-gates-experiment.zh.md)
+
+## Problem
+
+The pull-request Windows lane exists to prove the two blocking win32 surfaces — the workspace build and the production site — and it ran on hosted `windows-2025`, the slowest job in the required matrix: 7–9 minutes against 1.5–2.5 for the Linux jobs, so the Windows VM's boot, setup, and filesystem costs dominated every pull request's critical path.
+
+The question the experiment answered: can a plain Linux runner produce an equivalent win32 signal for the blocking surfaces at Linux wall clock, so no Windows VM sits on the pull-request path at all?
+
+## Decision
+
+The required pull-request `windows` job in [ci.yml](../../../../.github/workflows/ci.yml) (`windows node 24 / wine blocking`) runs the blocking gate commands on `ubuntu-latest` under Wine with real Windows binaries: a checksum-verified win-x64 Node.js executes `tsc -b`, `tsdown`, and the VitePress production build, so the win32 branches of the toolchain — backslash path handling, `CreateProcess` spawn semantics, PE loading of `@esbuild/win32-x64`, and the rolldown/rollup MSVC `.node` addons — actually execute. The master `serial-windows` job is untouched: the complete native-kernel inventory, including the observational portability gates this lane does not run, still executes on real `windows-2025` on every master push.
+
+Dependencies install natively on Linux with `supportedArchitectures` extended to win32-x64, which materializes the Windows platform packages in the same store; the cmd-shim layer is bypassed by invoking each tool's JavaScript entrypoint directly, the same processes `run-gates` ultimately spawns. `nodeLinker: hoisted` is load-bearing, not stylistic: the independent prototype in [PR #689](https://github.com/deepseek-harness/deepseek-harness/pull/689) kept pnpm's default isolated layout — including a faithful offline Windows-pnpm re-install over a Linux-prefetched store — and Windows Node under Wine still could not resolve `@esbuild/win32-x64` or load the koffi prebuild through the isolated symlink chain, failing before any repository gate ran. A flat layout with real files is what makes the gates reachable at all; #689's checksum pinning is adopted, while its Windows-pnpm-installs-the-tree goal is explicitly given up (the install contract stays Linux-tested here).
+
+The lane holds the wall clock of the Linux CI jobs through four levers: the master-refreshed pnpm store cache (restore-only, same key as the Linux jobs), Wine provisioning (apt install, Windows Node download, `wineboot`) running concurrently with `pnpm install`, the two blocking surfaces running concurrently — the same shape `run-gates` gives them on native Windows — and an apt-archive cache keyed on the runner image, seeded from master by the `wine apt cache` job so every pull request restores from the default-branch scope.
+
+Four environment constraints shape the job, each found as a red run: Ubuntu's `wine64` package alone puts nothing on PATH (install `wine`, the dispatcher); Node under Wine cannot attach stdio to the Actions runner's pipes (`Socket open EBADF` at bootstrap — every invocation routes stdio through a file); Wine does not realpath pnpm's isolated-layout Unix symlinks (the hoisted layout above); and Wine cannot create Windows symlinks (`ENOTSUP` from VitePress's `linkVue` — the `vue` link is laid down host-side before the gate).
+
+## Measured results
+
+Measured on 2026-07-27, warm caches, pull-request trigger, standard 2-core `ubuntu-latest`: 2m46s end-to-end — setup and cache restores ~17s, concurrent install+provision 33s, concurrent gates 110s — against 1.5–2.5 minutes for the Linux CI jobs and 7–9 minutes for the replaced `windows-2025` job. A cold-cache run pays roughly one extra minute. An 8-core benchmark leg was defined during the experiment but never left the restricted `dsh-ubuntu-*` pool's queue; the standard-runner number met the target, so no larger box is used.
+
+## Alternatives considered
+
+**Keep the hosted `windows-2025` pull-request job (status quo).** Nothing wrong with its signal, only its latency: 7–9 minutes for two build commands, the slowest required job in the matrix. It survives as the master serial reference, where completeness matters more than latency.
+
+**A full Windows guest under QEMU/KVM inside the Linux runner.** Real NT kernel, so full fidelity including case-insensitive NTFS and ConPTY — but tens of minutes of image download and unattended install before the first gate runs (40m19s measured end-to-end on the sibling experiment branch `exp/kvm-windows-ci`). Promotable only with disk-image caching that pressures the Actions cache budget.
+
+**Windows pnpm performing the install under Wine ([PR #689](https://github.com/deepseek-harness/deepseek-harness/pull/689)).** The higher-fidelity variant of this same idea: MinGit and pnpm staged into the prefix, a Linux prefetch filling the store, then `pnpm install --offline` run by Windows Node so the install contract itself executes as win32. It reached the install but not the gates — Wine's networking could not reach the registry directly, and the isolated `node_modules` layout defeated resolution of the Windows platform packages even after a clean offline install. This lane trades that fidelity away (hoisted layout, Linux-side install) to reach the gates; the two records are complementary halves of the same verdict.
+
+**Filesystem-semantics lanes on Linux (casefolded ext4, filename lint).** Catches the highest-frequency Windows breakage class for near-zero cost but proves nothing about win32 binaries. Explored as the sibling experiment branch `exp/casefold-windows-ci`; complementary to, not competitive with, this lane.
+
+**Windows containers.** Not possible: Windows containers require a Windows host kernel; a hosted Linux runner cannot run them.
+
+**Dropping the Windows lane.** Rejected — win32 is a first-class product target: the koffi-backed DACL and durable-namespace modules, ConPTY-based PTY sessions, and Windows path policy all ship in `packages/`.
+
+## Consequences
+
+Every pull request's Windows verdict now arrives in Linux-job time on free standard capacity, and no Windows VM allocation sits on the pull-request critical path; `all checks passed` consumes the same `windows` job id it always did.
+
+What the trade costs: Wine reimplements Win32 over a case-sensitive ext4 — NTFS case-insensitivity, real DACLs, ConPTY, and crash-durability semantics are not proved here, and the observational portability inventory (duplication, publint, node-next types, built-package invariants on win32) no longer runs on pull requests at all. The master `serial-windows` reference owns all of that: a Wine-green pull request can still fail the native-kernel master run, and that failure mode is accepted as post-merge. The lane also inherits Wine-specific divergences as permanent job structure — file-routed stdio, the host-side `vue` link, the hoisted layout — so a future toolchain change that depends on isolated-layout semantics or in-process symlink creation will surface here first as a Wine failure rather than a product failure, and triage must classify it as such. If Wine reds ever recur without product cause, the recorded fallback is reverting the `windows` job to the pre-Wine `windows-2025` definition preserved in git history.
@@ -0,0 +1,45 @@
+# Agent Note: 在 Linux runner 上用 Wine 运行 Windows 阻断门禁
+
+Status: implemented
+
+[English](2026-07-27-wine-windows-gates-experiment.md) | 中文
+
+## 问题
+
+Pull request 的 Windows 通道存在的意义是证明两个阻断性 win32 表面——workspace 构建与生产站点——它此前运行在托管 `windows-2025` 上，是必需矩阵中最慢的作业：7–9 分钟，对照 Linux 作业的 1.5–2.5 分钟，因此 Windows VM 的启动、准备与文件系统开销主导了每个 pull request 的关键路径。
+
+实验回答的问题是：一台普通 Linux runner 能否以 Linux 墙钟为阻断表面产出等效的 win32 信号，让 pull request 路径上完全没有 Windows VM？
+
+## 决策
+
+[ci.yml](../../../../.github/workflows/ci.yml) 中必需的 pull request `windows` 作业（`windows node 24 / wine blocking`）在 `ubuntu-latest` 上通过 Wine 用真实 Windows 二进制运行阻断门禁命令：校验和验证过的 win-x64 Node.js 执行 `tsc -b`、`tsdown` 与 VitePress 生产构建，因此工具链的 win32 分支——反斜杠路径处理、`CreateProcess` 派生语义、`@esbuild/win32-x64` 的 PE 加载、以及 rolldown/rollup 的 MSVC `.node` 插件——都真正执行。master 的 `serial-windows` 作业原封不动：完整的原生内核清单，包括本通道不运行的观察性可移植性门禁，仍在每次 master push 时于真实 `windows-2025` 上执行。
+
+依赖在 Linux 上原生安装，`supportedArchitectures` 扩展到 win32-x64，使 Windows 平台包物化进同一个 store；通过直接调用各工具的 JavaScript 入口绕开 cmd-shim 层，这正是 `run-gates` 最终派生的那些进程。`nodeLinker: hoisted` 是承重的，不是风格问题：[PR #689](https://github.com/deepseek-harness/deepseek-harness/pull/689) 的独立原型保留了 pnpm 默认的 isolated 布局——包括在 Linux 预取的 store 上忠实地用 Windows pnpm 离线重装——而 Wine 下的 Windows Node 依然无法穿过 isolated 符号链接链解析 `@esbuild/win32-x64` 或加载 koffi 预编译产物，在任何仓库门禁运行前就失败了。扁平的真实文件布局才让门禁变得可达；本通道采纳了 #689 的校验和固定，同时明确放弃其"Windows pnpm 安装依赖树"的目标（安装契约在此仍由 Linux 侧验证）。
+
+该通道靠四个杠杆保持 Linux CI 作业的墙钟：master 刷新的 pnpm store 缓存（只恢复，与 Linux 作业同键）、Wine 供给（apt 安装、Windows Node 下载、`wineboot`）与 `pnpm install` 并发运行、两个阻断表面并发运行——与 `run-gates` 在原生 Windows 上给它们的形状相同——以及按 runner 镜像为键的 apt 归档缓存，由 master 的 `wine apt cache` 作业播种，使每个 pull request 都能从默认分支作用域恢复。
+
+四条环境约束塑造了该作业，每条都以一次红色运行被发现：Ubuntu 的 `wine64` 包本身不往 PATH 放任何东西（要装 `wine` 调度器）；Wine 下的 Node 无法把 stdio 接到 Actions runner 的管道上（引导期 `Socket open EBADF`——所有调用都经文件中转 stdio）；Wine 不对 pnpm isolated 布局的 Unix 符号链接做 realpath（即上文的 hoisted 布局）；Wine 无法创建 Windows 符号链接（VitePress 的 `linkVue` 报 `ENOTSUP`——`vue` 链接在门禁前由宿主侧铺好）。
+
+## 实测结果
+
+2026-07-27 实测，热缓存，pull request 触发，标准 2 核 `ubuntu-latest`：端到端 2 分 46 秒——准备与缓存恢复约 17 秒，并发安装+供给 33 秒，并发门禁 110 秒——对照 Linux CI 作业的 1.5–2.5 分钟与被替换的 `windows-2025` 作业的 7–9 分钟。冷缓存约多付一分钟。实验期间定义过 8 核基准腿，但它从未离开受限 `dsh-ubuntu-*` 池的队列；标准 runner 的数字已达标，故不使用更大的机器。
+
+## 考虑过的替代方案
+
+**保留托管 `windows-2025` 的 pull request 作业（现状）。** 其信号没有问题，问题只在延迟：为两条构建命令花 7–9 分钟，是必需矩阵中最慢的作业。它作为 master 串行参照存续——在那里完整性比延迟更重要。
+
+**在 Linux runner 内用 QEMU/KVM 跑完整 Windows 客户机。** 真实 NT 内核，保真度完整，包括大小写不敏感的 NTFS 与 ConPTY——但首个门禁运行前要花数十分钟下载镜像并做无人值守安装（兄弟实验分支 `exp/kvm-windows-ci` 实测端到端 40 分 19 秒）。只有配上会挤压 Actions 缓存预算的磁盘镜像缓存才可晋升。
+
+**在 Wine 下由 Windows pnpm 执行安装（[PR #689](https://github.com/deepseek-harness/deepseek-harness/pull/689)）。** 同一想法的更高保真度变体：把 MinGit 与 pnpm 放进 prefix，用 Linux 预取填充 store，再由 Windows Node 运行 `pnpm install --offline`，让安装契约本身以 win32 身份执行。它到达了安装但没到达门禁——Wine 的网络无法直接访问 registry，且 isolated 的 `node_modules` 布局即便在干净的离线安装后也挫败了 Windows 平台包的解析。本通道用掉这份保真度（hoisted 布局、Linux 侧安装）来换取门禁可达；两份记录是同一裁决互补的两半。
+
+**Linux 上的文件系统语义通道（casefold ext4、文件名 lint）。** 以近零成本捕获最高频的 Windows 破坏类别，但对 win32 二进制什么也证明不了。作为兄弟实验分支 `exp/casefold-windows-ci` 探索；与本通道互补而非竞争。
+
+**Windows 容器。** 不可行：Windows 容器要求 Windows 宿主内核；托管 Linux runner 无法运行。
+
+**砍掉 Windows 通道。** 已否决——win32 是一等产品目标：基于 koffi 的 DACL 与持久命名空间模块、基于 ConPTY 的 PTY 会话、以及 Windows 路径策略都随 `packages/` 交付。
+
+## 结果
+
+每个 pull request 的 Windows 裁决现在以 Linux 作业的时间在免费标准容量上到达，pull request 关键路径上不再有任何 Windows VM 分配；`all checks passed` 消费的仍是原来的 `windows` 作业 id。
+
+这笔交易的代价：Wine 在大小写敏感的 ext4 之上重实现 Win32——NTFS 大小写不敏感、真实 DACL、ConPTY 与崩溃持久性语义在此都未被证明，且观察性可移植性清单（duplication、publint、node-next 类型、win32 上的构建包不变量）完全不再于 pull request 上运行。master 的 `serial-windows` 参照拥有这一切：Wine 绿灯的 pull request 仍可能在原生内核的 master 运行上失败，该失败模式被接受为合并后处理。该通道还把 Wine 特有的分歧继承为永久的作业结构——文件中转的 stdio、宿主侧的 `vue` 链接、hoisted 布局——因此未来依赖 isolated 布局语义或进程内符号链接创建的工具链变更会先在这里以 Wine 失败而非产品失败的形式浮现，分诊必须如此归类。若 Wine 红灯在无产品原因的情况下反复出现，记录在案的退路是把 `windows` 作业还原为 git 历史中保存的 Wine 之前的 `windows-2025` 定义。
@@ -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
-2026-07-26-consolidate-gate-scripts-on-existing-deps.md: 0bf32e01ea407e2718f8ec39ca962587a37df9cc
-2026-07-26-consolidate-gate-scripts-on-existing-deps.zh.md: ba9f157a61ffdc415acf9b2a61857026fc2c8bf1
+2026-07-26-consolidate-gate-scripts-on-existing-deps.md: 6370c8f92eff7296327e941e698ec4f733100bb2
+2026-07-26-consolidate-gate-scripts-on-existing-deps.zh.md: 3587c92e0e9655d8b38d24d184c1c68c44b131d4
@@ -0,0 +1,33 @@
+# Agent Note: Consolidate gate scripts on already-present deps and builtins
+
+Status: implemented
+
+English | [中文](2026-07-26-consolidate-gate-scripts-on-existing-deps.zh.md)
+
+## Problem
+
+The `scripts/` gates mostly used the right tools (`node:fs` `globSync` in 15+ gates, mdast/micromark in the markdown gates), but a handful of stragglers hand-rolled what a sibling gate already did with an existing dependency or builtin:
+
+- **Duplicated fence scanners.** `scripts/md-fences.ts` (~55 lines, consumed by `doc-typecheck.ts`) and `extractEquivBlocks` in `scripts/verify-type-equiv.ts` (~39 lines) were two copies of the same regex line-scanner for fenced code blocks, while `scripts/verify-mermaid.ts` already extracted fences by visiting mdast `code` nodes — and `markdownProseLines` in `scripts/markdown.ts` itself parsed to mdast but then hand-tracked fence state with a second regex. The regex scanners only recognized backtick fences at column 0, so they silently disagreed with the mdast-based gates on tilde and indented fences.
+- **Hand-rolled argv parsing.** `parseOptions` in `scripts/publint-all.ts` and its near-identical copy in `scripts/verify-built-package-invariants.mjs` (~26 lines) stepped argv indexes manually, while sibling scripts (`verify-runtime-closure.ts`, `build-exe-for-python-sdk.ts`, `packages/sdk/scripts/src/args.ts`) already used the `node:util` `parseArgs` builtin.
+- **Hand-rolled directory walks.** Five sites re-derived nested `readdirSync` walks that `globSync` covers: `verify-runtime-closure.ts` (packages + vendor manifests), `dev-web.ts` `discoverPluginDirs`, `verify-package-paths.ts` `realPackageNames`, `verify-client-domain-graph.ts` `listSources`, and `publint-all.ts` `addPath` (~55–65 lines total). `scripts/package-invariants.ts` shows the one-line `globSync` template.
+
+No new dependency was needed anywhere; every replacement is an existing devDep or a Node builtin.
+
+## Decision
+
+- A shared mdast fence helper, `markdownFences` in `scripts/markdown.ts`, visits `code` nodes for the language, full info string, body, and 1-based opening-fence line; `doc-typecheck.ts` and `verify-type-equiv.ts` extract fences through it. `md-fences.ts` and the duplicated `extractEquivBlocks` scanner are deleted, and `markdownProseLines` derives fenced lines from the parsed `code` nodes' positions instead of a second regex.
+- Both CLIs parse argv via `parseArgs`; unknown options and missing values still fail loud, with `parseArgs`'s own error text instead of the bespoke usage strings.
+- The five straggler walks use `globSync`. The walks in `check-workspace-constraints.ts` and `clean.ts` stay: they need dirent-level detail to diagnose malformed trees, which glob-by-pattern cannot report.
+
+## Alternatives considered
+
+- **A new glob/walking dependency (`tinyglobby`, `fdir`).** Rejected: the builtin already won repo-wide; these were stragglers, not a gap.
+- **`p-map` for `publint-all.ts`'s ~19-line ordered worker pool.** Deliberately left out: one new devDep for one small deletion is at the edge of the [dependency policy](../process/2026-07-26-dependencies-over-hand-rolling.md) bar, and the pool's requirements (bounded workers, deterministic order, env override) are documented in the [parallel-gates note](../process/2026-07-06-parallel-pre-push-gates.md). Fold it in only if `p-map` earns a second consumer.
+- **Leaving the fence scanners.** Rejected: two drifting copies of a parser beside a third correct implementation is exactly the duplication the shared `markdown.ts` helper exists to prevent, and the column-0-backtick-only limitation was a latent inconsistency between sibling gates.
+
+## Consequences
+
+- One fence parser: every markdown gate now classifies fences through mdast, so tilde, indented, and 4-backtick container fences behave identically everywhere. The docs tree contained no fence shape the regex scanners mishandled, so gate results are unchanged on the tree that landed the swap: `pnpm run doc-sync` and each rewritten gate ran before and after with byte-identical output (`doc-typecheck` block/opt-out counts, `verify-type-equiv` match counts, `publint`, `verify-built-package-invariants`, `verify-runtime-closure`, `verify-package-paths`, `verify-client-domain-graph`, and both package-README prose gates).
+- `verify-type-equiv` still rejects an unterminated type-equivalence fence: mdast silently closes an unterminated block at end-of-file (its comparisons could then pass), so the shared helper reports whether a closing delimiter exists and the gate errors on an unclosed block, preserving the removed scanner's rejection. The `doc-typecheck` scanner never had that error path.
+- `parseArgs` keeps the last value of a duplicated option instead of erroring — a dev-tool edge case the tests don't pin, accepted in exchange for deleting the two bespoke parsers. (Strict mode still rejects a `--`-prefixed token where a value is expected, matching the replaced parsers.)
@@ -0,0 +1,33 @@
+# Agent Note: 把门禁脚本统一到已有依赖与内置模块上
+
+Status: implemented
+
+[English](2026-07-26-consolidate-gate-scripts-on-existing-deps.md) | 中文
+
+## 问题
+
+`scripts/` 下的门禁大多本已在用正确的工具（15 个以上的门禁使用 `node:fs` 的 `globSync`，markdown 门禁使用 mdast/micromark），但少数几个掉队的脚本曾手写同类门禁早已用既有依赖或内置模块完成的事情：
+
+- **重复的围栏扫描器。**`scripts/md-fences.ts`（约 55 行，由 `doc-typecheck.ts` 消费）和 `scripts/verify-type-equiv.ts` 中的 `extractEquivBlocks`（约 39 行）曾是同一个围栏代码块正则行扫描器的两份拷贝，而 `scripts/verify-mermaid.ts` 早已通过访问 mdast `code` 节点来提取代码围栏；`scripts/markdown.ts` 自己的 `markdownProseLines` 也曾先解析成 mdast，再用第二个正则手工跟踪围栏状态。这两个正则扫描器只识别第 0 列的反引号围栏，因此在波浪线围栏和缩进围栏上与基于 mdast 的门禁悄悄不一致。
+- **手写的 argv 解析。**`scripts/publint-all.ts` 中的 `parseOptions` 和 `scripts/verify-built-package-invariants.mjs` 中与之几乎相同的拷贝（约 26 行）曾手工推进 argv 下标，而同类脚本（`verify-runtime-closure.ts`、`build-exe-for-python-sdk.ts`、`packages/sdk/scripts/src/args.ts`）早已在使用 `node:util` 的内置 `parseArgs`。
+- **手写的目录遍历。**五处代码曾各自重写 `globSync` 已覆盖的嵌套 `readdirSync` 遍历：`verify-runtime-closure.ts` 对 packages 与 vendor manifest（元数据清单）的扫描、`dev-web.ts` 的 `discoverPluginDirs`、`verify-package-paths.ts` 的 `realPackageNames`、`verify-client-domain-graph.ts` 的 `listSources`，以及 `publint-all.ts` 的 `addPath`（合计约 55–65 行）。`scripts/package-invariants.ts` 展示了一行式的 `globSync` 模板。
+
+所有替换都不需要引入新依赖；每一处替换用的都是既有的 devDependency 或 Node 内置模块。
+
+## 决策
+
+- `scripts/markdown.ts` 中的共享 mdast 围栏辅助函数 `markdownFences` 访问 `code` 节点，读取语言、完整 info string、块体以及以 1 起始的开围栏行号；`doc-typecheck.ts` 和 `verify-type-equiv.ts` 通过它提取代码围栏。`md-fences.ts` 和重复的 `extractEquivBlocks` 扫描器已删除，`markdownProseLines` 也改为从解析出的 `code` 节点位置推导围栏内的行，而不再用第二个正则。
+- 两个 CLI 都改用 `parseArgs` 解析 argv；未知选项和缺失取值仍然大声失败，只是错误文案换成了 `parseArgs` 自带的文本，而非原先手写的用法字符串。
+- 那五处掉队的目录遍历改用 `globSync`。`check-workspace-constraints.ts` 和 `clean.ts` 中的遍历保留：它们需要 dirent 级别的细节来诊断结构异常的目录树，按模式匹配的 glob 报告不了这些信息。
+
+## 曾考虑的替代方案
+
+- **新的 glob/目录遍历依赖（`tinyglobby`、`fdir`）。**不予采纳：内置模块已在全仓库范围内胜出；这几处只是掉队者，不是能力缺口。
+- **用 `p-map` 替换 `publint-all.ts` 中约 19 行的有序 worker 池。**刻意未纳入：为一次小删除引入一个新 devDependency，正处在[依赖策略](../process/2026-07-26-dependencies-over-hand-rolling.md)门槛的边缘，而且该池的需求（worker 数量有界、确定性顺序、环境变量覆盖）已记录在[并行 pre-push 门禁决策记录](../process/2026-07-06-parallel-pre-push-gates.md)中。仅当 `p-map` 赢得第二个消费方时再顺带纳入。
+- **保留这两个围栏扫描器。**不予采纳：在第三个正确实现旁边放着两份逐渐漂移的解析器拷贝，正是共享的 `markdown.ts` 辅助函数要防止的那种重复；「只认第 0 列反引号」的限制也是同类门禁之间的潜在不一致。
+
+## 后果
+
+- 只剩一个围栏解析器：所有 markdown 门禁现在都经由 mdast 归类代码围栏，因此波浪线围栏、缩进围栏和四反引号容器围栏在各处的行为完全一致。文档树中不存在正则扫描器处理有误的围栏形态，所以在落地这次替换的代码树上门禁结果不变：`pnpm run doc-sync` 及每个被改写的门禁在改动前后各跑一遍，输出逐字节相同（`doc-typecheck` 的块数/opt-out 计数、`verify-type-equiv` 的匹配计数、`publint`、`verify-built-package-invariants`、`verify-runtime-closure`、`verify-package-paths`、`verify-client-domain-graph`，以及两个包 README 散文门禁）。
+- `verify-type-equiv` 仍然拒绝未闭合的类型等价围栏：mdast 会在文件末尾静默闭合未闭合的代码块（其比较随后可能通过），因此共享辅助函数会报告闭合定界符是否存在，门禁在块未闭合时报错，保留了被删扫描器的这条拒绝路径。`doc-typecheck` 的扫描器本来就没有这条错误路径。
+- `parseArgs` 对重复出现的选项保留最后一个值而不报错——一个测试未固定的开发工具边缘用例，作为删除两份手写解析器的交换被接受。（严格模式下，需要取值处遇到以 `--` 开头的 token 仍会拒绝，与被替换的解析器行为一致。）
@@ -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
-2026-07-26-eventsource-parser-for-deepseek-sse.md: 8a93b7f6c7aa0d428f25e87c44e1d29e884ecc81
-2026-07-26-eventsource-parser-for-deepseek-sse.zh.md: b16109d9458f487c7e463cf02e6b2d22fbbde015
+2026-07-26-eventsource-parser-for-deepseek-sse.md: e7835bc738b3dec5aefd6011848525f6604e852e
+2026-07-26-eventsource-parser-for-deepseek-sse.zh.md: 933b993d479026d8f2bd2dc3173abd9e60823806
@@ -0,0 +1,28 @@
+# Agent Note: Replace the hand-rolled SSE parser in llm-deepseek with eventsource-parser
+
+Status: implemented
+
+English | [中文](2026-07-26-eventsource-parser-for-deepseek-sse.zh.md)
+
+## Problem
+
+`packages/llm/llm-deepseek/src/sse.ts` hand-implemented Server-Sent Events parsing: a streaming `TextDecoder`, event-block splitting on `\r?\n\r?\n`, `data:` payload extraction and joining, comment/field skipping, the `[DONE]` sentinel, a `STREAM_CLOSED` error on EOF without it, and a flush of a final unterminated event block. The file was ~67 lines with ~108 lines of dedicated tests (`tests/sse.spec.ts`) re-proving SSE spec behavior — UTF-8 split across chunks, CRLF handling, multi-`data:` joining, no-space-after-colon — that a maintained parser already guarantees. Its only consumer is `adapter.ts` (`yield* translate(parseSse(response.body))`).
+
+This is exactly the surface `eventsource-parser` owns: the de-facto standard SSE parser (it underlies the Vercel AI SDK and the MCP SDK), zero-dependency, actively maintained, and already present in this repo's lockfile transitively via `@modelcontextprotocol/sdk` — so adopting it directly adds no new supply-chain surface in practice.
+
+## Decision
+
+`sse.ts` delegates SSE framing to `EventSourceParserStream` from `eventsource-parser/stream`: `parseSse` pipes the response body through `new TextDecoderStream()` then `new EventSourceParserStream()` and keeps only the DeepSeek protocol shim — yield each event's `data`, terminate on `[DONE]`, and throw `LlmError('STREAM_CLOSED')` when the stream ends without the sentinel. All required builtins (`TextDecoderStream`, `pipeThrough`, async-iterable `ReadableStream`) exist at the Node ^22.19 engine floor. The spec-conformance tests are gone; `tests/sse.spec.ts` pins only the `[DONE]`/`STREAM_CLOSED`/EOF contract. `eventsource-parser` is `llm-deepseek`'s second runtime dependency after schemastery. The [twin-adapters note](../architecture/2026-06-13-twin-llm-adapters.md) and the `dsh-llm` JSDoc that branded this adapter "hand-rolled fetch + SSE parsing" now describe it as direct fetch with library-framed SSE.
+
+The library also strips a leading BOM (the hand-rolled parser would fail to match `data:` after one) and offers `maxBufferSize` hardening the hand-rolled parser lacked.
+
+## Alternatives considered
+
+- **Keep the hand-rolled parser.** Defensible under the [twin-adapters decision](../architecture/2026-06-13-twin-llm-adapters.md): the adapter is deliberately the hand-rolled design-verification twin of the pi-ai adapter. But that note's load-bearing distinction is owning the fetch/translate internals versus delegating to a full provider SDK; a ~700-byte SSE micro-parser is transport plumbing, not the design under verification. The twin-adapters note now states that reading explicitly.
+- **`createParser({onEvent})` callback API instead of the stream.** Works fed by a manual `TextDecoder` loop, but the `pipeThrough` composition deletes more of the hand-rolled code.
+
+## Consequences
+
+- The remaining shim only encodes the DeepSeek `[DONE]`/`STREAM_CLOSED` protocol; SSE framing edge cases are eventsource-parser's contract and are no longer re-proven here.
+- One deliberate robustness deviation is dropped: the hand-rolled parser flushed a final event block that lacked its terminating blank line, so a trailing `data: [DONE]` without `\n\n` still yielded DONE. eventsource-parser is spec-strict and only dispatches on the blank line, so that shape is now `STREAM_CLOSED`. Real providers and `dsh-llm-mock-server` always terminate events properly — the flush was a robustness nicety, not an observed provider shape — and `tests/sse.spec.ts` pins the new truncation verdict for that tail.
+- The documented "hand-rolled" identity of the twin adapter narrows to the fetch/translate internals; the twin-adapters note was updated in the same change rather than leaving the claim stale.
@@ -0,0 +1,28 @@
+# Agent Note: 用 eventsource-parser 替换 llm-deepseek 中手写的 SSE 解析器
+
+Status: implemented
+
+[English](2026-07-26-eventsource-parser-for-deepseek-sse.md) | 中文
+
+## 问题
+
+`packages/llm/llm-deepseek/src/sse.ts` 曾手写实现 SSE（Server-Sent Events）解析：一个流式 `TextDecoder`、按 `\r?\n\r?\n` 切分事件块、提取并拼接 `data:` 载荷、跳过注释与其他字段、`[DONE]` 哨兵、在未见哨兵即 EOF 时抛出 `STREAM_CLOSED` 错误，以及对最后一个未终结事件块的 flush。该文件约 67 行，另有约 108 行专属测试（`tests/sse.spec.ts`）重复验证 SSE 规范行为——UTF-8 字符被切分到多个分片、CRLF 处理、多条 `data:` 拼接、冒号后无空格——而这些行为，持续维护的解析器早已有保证。它唯一的消费方是 `adapter.ts`（`yield* translate(parseSse(response.body))`）。
+
+这恰好是 `eventsource-parser` 负责的接口面：事实标准的 SSE 解析器（Vercel AI SDK 和 MCP SDK 都构建在它之上），零依赖，持续维护，并且已通过 `@modelcontextprotocol/sdk` 作为传递依赖出现在本仓库的 lockfile 中——因此直接采用它实际上不增加新的供应链接触面。
+
+## 决策
+
+`sse.ts` 将 SSE 分帧委托给 `eventsource-parser/stream` 的 `EventSourceParserStream`：`parseSse` 把响应 body 依次管道接入 `new TextDecoderStream()` 和 `new EventSourceParserStream()`，只保留 DeepSeek 协议垫层——逐个产出事件的 `data`，遇到 `[DONE]` 终止，流在未见哨兵时结束则抛出 `LlmError('STREAM_CLOSED')`。所需的全部内置能力（`TextDecoderStream`、`pipeThrough`、可异步迭代的 `ReadableStream`）在 Node ^22.19 引擎下限即已存在。规范符合性测试已删除；`tests/sse.spec.ts` 只固定 `[DONE]`/`STREAM_CLOSED`/EOF 契约。`eventsource-parser` 是 `llm-deepseek` 继 schemastery 之后的第二个运行时依赖。曾把该适配器标为「手写 fetch + SSE 解析」的[孪生适配器 Agent Note（agent 决策记录）](../architecture/2026-06-13-twin-llm-adapters.md)与 `dsh-llm` JSDoc，现在将其描述为直接 fetch 加库分帧的 SSE。
+
+该库还会剥离开头的 BOM（手写解析器在 BOM 之后会无法匹配 `data:`），并提供手写解析器缺少的 `maxBufferSize` 加固能力。
+
+## 曾考虑的替代方案
+
+- **保留手写解析器。** 依据[孪生适配器决策](../architecture/2026-06-13-twin-llm-adapters.md)，这一选择有辩护余地：该适配器有意作为 pi-ai 适配器的手写设计验证孪生体。但那份 Agent Note 起支撑作用的区分在于「自行持有 fetch/translate 内部实现」与「委托给完整的提供方 SDK」；一个约 700 字节的 SSE 微型解析器属于传输层管道，不是被验证的设计本身。孪生适配器 Agent Note 现已明确写出这一解读。
+- **改用 `createParser({onEvent})` 回调 API 而非流。** 配合手动的 `TextDecoder` 循环可以工作，但 `pipeThrough` 组合方式能删除更多手写代码。
+
+## 后果
+
+- 剩下的垫层只编码 DeepSeek 的 `[DONE]`/`STREAM_CLOSED` 协议；SSE 分帧边界情形属于 eventsource-parser 的契约，不再在这里重复验证。
+- 放弃了一处有意为之的健壮性偏离：手写解析器会 flush 缺少终结空行的最后一个事件块，因此末尾的 `data: [DONE]` 即使没有 `\n\n` 也仍产出 DONE。eventsource-parser 严格遵循规范，只在空行处分发事件，所以这种形态现在是 `STREAM_CLOSED`。真实提供方和 `dsh-llm-mock-server` 总是正确终结事件——该 flush 只是健壮性上的锦上添花，并非实际观测到的提供方形态——`tests/sse.spec.ts` 固定了对该尾部的新截断判定。
+- 孪生适配器有文档记录的「手写」身份收窄到 fetch/translate 内部实现；孪生适配器 Agent Note 在同一次变更中更新，而不是让声明陈旧下去。
@@ -0,0 +1,6 @@
+# Bilingual-pair consistency record (docs/i18n/README.md): the git blob hash of each
+# side as of the last confirmed-consistent state. Both languages carry equal authority;
+# after editing either side, bring the other along and re-record with:
+#   pnpm run verify-translation-pairing --write .agents/notes/implemented/simplification/2026-07-26-turndown-for-tool-web-html-markdown.md
+2026-07-26-turndown-for-tool-web-html-markdown.md: 0e387021e3d3be3011cc0d64d37864b30aec4fdf
+2026-07-26-turndown-for-tool-web-html-markdown.zh.md: 44e7d08c1db40a1203e8cda955b521774335e774
@@ -0,0 +1,37 @@
+# Agent Note: Replace tool-web's regex HTML-to-markdown converter with turndown
+
+Status: implemented
+
+English | [中文](2026-07-26-turndown-for-tool-web-html-markdown.zh.md)
+
+## Problem
+
+`dsh-tool-web`'s `src/html.ts` (~86 lines, ~40 lines of dedicated tests; deleted by this change) converted fetched HTML to markdown with regexes: strip script/style/noscript/comments, convert `<a>`/`<h1-6>`/`<li>`, decode numeric entities plus a 12-entry named-entity table, collapse whitespace. The module's own JSDoc said "A richer converter can replace it without changing the seam or tool schema", and the README's Known Limitations documented it as "a minimal regex converter, not an HTML parser — tables, images, and nested formatting are lost." The [web capability seam note](../architecture/2026-06-24-web-capability-seam.md) assigns HTML→markdown to this package as presentation, so the swap point was exactly here. The converter's output is model-visible on every fetched HTML page; no keyless snapshot exercised `web_fetch`, so no expected outputs pinned it.
+
+## Decision
+
+`packages/web/tool-web/src/fetch.ts` owns a module-level [`turndown`](https://github.com/mixmark-io/turndown) instance (`headingStyle: 'atx'`, `codeBlockStyle: 'fenced'`, `bulletListMarker: '-'` — fixed model-facing presentation, not deployment tunables) with `@joplin/turndown-plugin-gfm`'s composite `gfm` plugin for tables/strikethrough and `remove(['script', 'style', 'noscript'])` replacing the old wholesale drops. `formatFetchOutput` limits both the source prefix converted synchronously and the complete rendered output with `fetchMaxOutputChars` (default 200,000), so a custom provider cannot make conversion work unbounded before the output cap applies. The HTML arm then guards conversion twice: a conservative linear lexical pass treats comment contents conservatively, skips raw-text bodies, honors quoted tag text, and passes a body through raw when its stack crosses 512 levels; a try/catch also falls back to raw HTML when turndown rejects markup the guard cannot model. The GFM cell rule is overridden to ignore `colspan`, which Markdown cannot represent, rather than letting an untrusted numeric attribute synthesize arbitrary empty cells. `html.ts` and its conversion tests are deleted; the source/output bounds, fallback, and status-header/truncation-footer formatting are tested in `tests/tool-web.spec.ts`, and the README's Known Limitations trades the regex-converter caveat for the bounded degradation cases. The gfm plugin ships no types; `src/turndown-plugin-gfm.d.ts` declares the one imported export over `@types/turndown` (a devDependency).
+
+The dependency-weight question the proposal flagged resolves in favor of the swap: `@deepseek-ai/dsh-tool-web` is in the single-file-executable closure ([single-exe note](../architecture/2026-07-10-single-file-executable-sdk-runtime-distribution.md)), and the exe's asset globs would pack ~7.9 MB of the three packages as published — but ~6 MB of that is `@mixmark-io/domino`'s test corpus (`test/**`), with runtime `lib/` at ~550 KB against a ~174 MB artifact, under 0.5% either way.
+
+## Snapshot coverage
+
+The previously-missing keyless `web_fetch` snapshot ships with the change as the acp-agent scenario `web-fetch`: `examples/acp-agent/web.cordis.yml` composes the web seam, the real `dsh-web-fetch-local` provider, `tool-web` with `search: false`, and `web-fetch-fixture-server.mjs` — a loopback HTTP fixture on a fixed port (the fetched URL is part of the recorded transcript) serving deterministic HTML with named entities, a GFM table, and nested formatting. Recording and keyless replay both drive the real HTTP fetch and conversion; the pinned tool result is the turndown output, and the scenario pins the `web` header class (the `web_fetch` schema and guidance).
+
+## Alternatives considered
+
+- **`@mozilla/readability` + a DOM.** Solves a different problem (content extraction, not conversion) and drags a heavier DOM dependency; the seam only asks for markdown rendering of whatever the fetch returned.
+- **Keep the regex converter.** It was an explicit v1 placeholder per its own JSDoc; keeping it meant model-visible quality (tables, images, nested formatting) stayed lost for the cost of maintaining bespoke entity tables.
+- **The minimal `entities`-only variant.** The proposal's fallback position: replace only the entity-decoding third of `html.ts` with the zero-dependency `entities` package, deleting less but avoiding the dependency-weight question. Not taken because the closure math above made the weight immaterial while the full swap deletes the whole hand-rolled converter and its documented quality gaps.
+- **`turndown-plugin-gfm` (the original) instead of `@joplin/turndown-plugin-gfm`.** The original is unmaintained (last publish 2018); the Joplin fork is current against turndown 7 and actively released.
+
+## Consequences
+
+- **Bought**: standards-based model-visible markdown — ordinary tables, images, strikethrough, nested emphasis, fenced code blocks, and the complete named-entity set — plus the deletion of the bespoke converter and its entity tables.
+- **Paid**: two runtime dependencies (`turndown` → `@mixmark-io/domino`) enter tool-web and therefore the exe closure (~550 KB of runtime code as measured above); overlong input is converted only through a bounded prefix, pathological nesting falls back to raw HTML, and spanning table cells are flattened because GFM has no corresponding syntax.
+- Model-visible output changed on every fetched HTML page; nothing pinned the old output, and the new snapshot pins the new one.
+
+## Testing
+
+- `packages/web/tool-web/tests/tool-web.spec.ts` covers the turndown conversion surface (entities, links, tables, nesting, script/style/noscript removal), ignored table spans, source-prefix and complete-output bounds, fast raw-HTML passthrough for deep or deceptively closed nesting, linear handling of malformed tags, the residual converter-throw fallback, and exact and tiny output budgets; per-file coverage on the package src is 100%.
+- The `web-fetch` acp-agent snapshot pins the assembled behavior keylessly end to end (real Loader composition, real HTTP fetch, real conversion).
@@ -0,0 +1,37 @@
+# Agent Note: 用 turndown 替换 tool-web 的正则 HTML 转 markdown 转换器
+
+Status: implemented
+
+[English](2026-07-26-turndown-for-tool-web-html-markdown.md) | 中文
+
+## 问题
+
+`dsh-tool-web` 的 `src/html.ts`（约 86 行，另有约 40 行专属测试；已由本变更删除）曾用正则表达式把抓取到的 HTML 转成 markdown：剥离 script、style、noscript 标签与注释，转换 `<a>`/`<h1-6>`/`<li>`，解码数字实体外加一张 12 项的命名实体表，并折叠空白。该模块自身的 JSDoc 写明「A richer converter can replace it without changing the seam or tool schema」，README 的 Known Limitations 章节也把它记载为「a minimal regex converter, not an HTML parser — tables, images, and nested formatting are lost」。[web 能力 seam 决策记录](../architecture/2026-06-24-web-capability-seam.md)把 HTML 转 markdown 作为呈现职责划归本包（package），因此替换点恰好就在这里。每个抓取到的 HTML 页面上，该转换器的输出都对模型可见；此前没有任何无密钥快照执行到 `web_fetch`，因此没有预期输出固定它的行为。
+
+## 决策
+
+`packages/web/tool-web/src/fetch.ts` 持有一个模块级 [`turndown`](https://github.com/mixmark-io/turndown) 实例（`headingStyle: 'atx'`、`codeBlockStyle: 'fenced'`、`bulletListMarker: '-'`——固定的面向模型呈现方式，不是部署可调项），配合 `@joplin/turndown-plugin-gfm` 的组合 `gfm` 插件提供表格／删除线支持，并用 `remove(['script', 'style', 'noscript'])` 替代旧实现的整体剥离。`formatFetchOutput` 通过 `fetchMaxOutputChars`（默认 200,000）同时限制同步转换的源前缀和完整渲染输出，因此自定义提供方无法在输出上限生效前造成无界的转换工作。随后，HTML 分支对转换做双重防护：保守的线性词法扫描会保守处理注释内容，跳过原始文本元素的内容，正确处理标签内的引号文本，并在栈深超过 512 层时将主体作为原始 HTML 直接透传；当 turndown 拒绝守卫无法建模的标记时，try/catch 同样回退为原始 HTML。GFM 单元格规则被覆写为忽略 `colspan`；Markdown 无法表示它，这也避免了不受信任的数值属性凭空合成任意数量的空单元格。`html.ts` 及其转换测试已删除；源／输出上限、回退以及状态头／截断页脚格式化均在 `tests/tool-web.spec.ts` 中有测试覆盖，README 的 Known Limitations 用有界降级情形替换了正则转换器警示。gfm 插件不带类型声明；`src/turndown-plugin-gfm.d.ts` 基于 `@types/turndown`（devDependency）声明了唯一被导入的导出。
+
+提案标记的依赖体积问题的裁决结果支持替换：`@deepseek-ai/dsh-tool-web` 在单文件可执行文件闭包内（[single-exe 决策记录](../architecture/2026-07-10-single-file-executable-sdk-runtime-distribution.md)），可执行文件的资产 glob 会把这三个包按发布原样打入约 7.9 MB——但其中约 6 MB 是 `@mixmark-io/domino` 的测试语料（`test/**`），运行时 `lib/` 仅约 550 KB，相对约 174 MB 的产物，两种口径都不到 0.5%。
+
+## 快照覆盖
+
+此前缺失的无密钥 `web_fetch` 快照随本变更以 acp-agent 场景 `web-fetch` 落地：`examples/acp-agent/web.cordis.yml` 组合了 web seam、真实的 `dsh-web-fetch-local` 提供方、`search: false` 的 `tool-web`，以及 `web-fetch-fixture-server.mjs`——一个固定端口（抓取的 URL 是录制 transcript（文本记录）的一部分）上的回环 HTTP fixture，提供包含命名实体、GFM 表格与嵌套格式的确定性 HTML。录制与无密钥回放都驱动真实的 HTTP 抓取与转换；固定住的工具结果就是 turndown 的输出，该场景同时固定 `web` header 类（`web_fetch` 的 schema 与指引）。
+
+## 曾考虑的替代方案
+
+- **`@mozilla/readability` 加一个 DOM。** 它解决的是另一个问题（内容提取，而非格式转换），还会拖入更重的 DOM 依赖；这个 seam 只要求把抓取返回的内容渲染成 markdown。
+- **保留正则转换器。** 按其自身 JSDoc 的说法，它本来就是明确的 v1 占位实现；保留它意味着模型可见的质量（表格、图片、嵌套格式）继续缺失，代价还是维护一套自制实体表。
+- **仅引入 `entities` 的最小变体。** 提案中的退守方案：只用零依赖的 `entities` 包替换 `html.ts` 中的实体解码部分，删得更少但完全避开依赖体积问题。未采纳：上述闭包测算表明体积无关紧要，而完整替换能删掉整个手写转换器及其记录在案的质量缺口。
+- **用原版 `turndown-plugin-gfm` 而非 `@joplin/turndown-plugin-gfm`。** 原版已无人维护（最后发布于 2018 年）；Joplin 分叉与 turndown 7 保持同步并持续发布。
+
+## 后果
+
+- **收益**：基于标准的模型可见 markdown——普通表格、图片、删除线、嵌套强调、围栏代码块以及完整的命名实体集——并删除了自制转换器及其实体表。
+- **代价**：两个运行时依赖（`turndown` → `@mixmark-io/domino`）进入 tool-web 进而进入可执行文件闭包（如上实测约 550 KB 运行时代码）；超长输入只转换有界前缀，病态嵌套回退为原始 HTML，跨列表格单元格会被展平，因为 GFM 没有对应语法。
+- 每个抓取到的 HTML 页面上模型可见的输出都已变化；旧输出本无任何固定，新快照固定了新输出。
+
+## 测试
+
+- `packages/web/tool-web/tests/tool-web.spec.ts` 覆盖 turndown 转换面（实体、链接、表格、嵌套、script/style/noscript 移除）、被忽略的表格跨列、源前缀与完整输出上限、深层或带欺骗性闭合嵌套的快速原始 HTML 透传、畸形标签的线性处理、残余的转换器抛错回退，以及恰好达到上限和极小的输出预算；该包 src 的逐文件覆盖率为 100%。
+- acp-agent 的 `web-fetch` 快照无密钥地端到端固定组装后的行为（真实 Loader 组合、真实 HTTP 抓取、真实转换）。
@@ -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
-2026-07-18-tui-terminal-state-snapshots.md: 8e86588f69fdb9d615232252ecf57309d440f1cd
-2026-07-18-tui-terminal-state-snapshots.zh.md: b70a46830f44e9da663e30745fcdb7ad281592da
+#   pnpm run verify-translation-pairing --write .agents/notes/implemented/testing/2026-07-18-tui-terminal-state-snapshots.md
+2026-07-18-tui-terminal-state-snapshots.md: 18c79bc2d0dabf4d78887354f30a2cdc083899e1
+2026-07-18-tui-terminal-state-snapshots.zh.md: d1d4a6ca859e94a153e0bf645a17f03c0dac234b
@@ -33,7 +33,7 @@ The live-model fixtures use `DSH_SNAPSHOT=record`; record mode rewrites their pr

 ### Semantic terminal projection

-The package-local `HeadlessTerminal` implements the same pi-tui `Terminal` interface as the process terminal and feeds every ANSI write into the pinned `@xterm/headless` parser. Snapshot code waits for synchronized frames to quiesce before reading state, so a checkpoint represents a completed screen rather than a timer-dependent write prefix.
+The package-local `HeadlessTerminal` implements the same pi-tui `Terminal` interface as the process terminal and feeds every ANSI write into the pinned `@xterm/headless` parser. Snapshot code waits for synchronized frames to quiesce before reading state. The streaming checkpoint freezes the loader interval while allowing real wall-clock delay across one animation tick, so it pins semantic status rather than whichever spinner glyph the scheduler happened to render.

 Each expected output projects dimensions, active-buffer and viewport coordinates, lifecycle and cursor state, rows, wrap markers, and non-default style ranges into text. Scroll-heavy cards capture the used buffer; overlays capture the visible viewport. Text and style remain separate so a reviewer can distinguish content changes from presentation changes without decoding ANSI bytes.

@@ -33,7 +33,7 @@ TUI 覆盖分为四个互补层次：

 ### 语义终端投影

-包内的 `HeadlessTerminal` 实现与进程终端相同的 pi-tui `Terminal` 接口，并把每次 ANSI 写入交给固定版本的 `@xterm/headless` 解析器。读取状态前，快照代码会等待同步帧稳定，因此每个检查点表示已经完成的画面，而不是依赖计时的写入前缀。
+包内的 `HeadlessTerminal` 实现与进程终端相同的 pi-tui `Terminal` 接口，并把每次 ANSI 写入交给固定版本的 `@xterm/headless` 解析器。读取状态前，快照代码会等待同步帧稳定。流式输出检查点会冻结 loader 的 interval，同时保留跨过一次动画 tick 的真实墙钟等待，从而固定语义状态，而非调度器碰巧渲染出的某个加载动画字形。

 每份预期输出把终端尺寸、活动缓冲区和视口坐标、生命周期与光标状态、各行、换行标记以及非默认样式区间投影为文本。滚动内容较多的卡片捕获已使用缓冲区；浮层捕获可见视口。文本和样式相互分离，评审人无需解码 ANSI 字节即可区分内容变化与呈现变化。

@@ -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
-2026-07-24-web-gui-browser-e2e-lane.md: c4e34b3f44162c7021cb25681eea7e49ac78f672
-2026-07-24-web-gui-browser-e2e-lane.zh.md: 466e1c0fc16aac21b87b68cfedaec4fb22a417e2
+#   pnpm run verify-translation-pairing --write .agents/notes/implemented/testing/2026-07-24-web-gui-browser-e2e-lane.md
+2026-07-24-web-gui-browser-e2e-lane.md: d9e0a9660ecd6aeb75e835e68f92c0a268423872
+2026-07-24-web-gui-browser-e2e-lane.zh.md: e8c7d1c4596f20d88bd08423549fb6a9f7b0654b
@@ -18,7 +18,7 @@ A plain shared-fixture module (the [testing-policy sanctioned shape](../../../..

 `launchWebScaffold()` boots the real web composition from the shipped `apps/cli/cordis.yml` through the vendored Loader's include mechanism — the same tree and mechanism `AppCLIEntry` drives for `dsh web`. Divergences ride include patches over that tree, the ACP `cordis.snapshot.yml` pattern expressed in-process: temp `persistenceRoot`, `workspace-context` disabled (recorded fixtures must not embed this repo's AGENTS.md), `session-title-llm` disabled (its fire-and-forget title call would race the loop for the session's replay cursor), the webserver row pinned to port 0 with the built dist, and in keyless modes `llm-deepseek` disabled. A patch id that stops matching a row fails the boot sweep loudly instead of drifting. The boot runs `chdir`'d to the temp workspace so the api-gateway's `process.cwd()` session default, tool cwds, and fixtures agree; the `dsh web` bin's own glue (argv, profile json, AppCLIEntry) stays held by the keyless CLI smokes in `smoke-real.e2e.ts`. Setup rollback and ordinary close both dispose the Cordis tree before removing the two owned temp roots, attempt every cleanup independently, and report cleanup failures without masking the setup failure.

-Keyless model displacement is the disabled adapter row plus `installLlmReplay` filling the open seam on the settled root ctx in providers-catalog mode — never catch-all: with the adapter row disabled no adapter exists, so catch-all would leave `resolveModelContext` unroutable and `compact-basic`'s post-step pressure check would warn every step instead of being provably inert (the published 128k `contextWindow` keeps it inert for small fixtures). The direct install rather than an inserted replay plugin row is deliberate: it returns the `ReplayHandle` the teardown consumption check needs. A scenario with no fixture leaves the seam empty, so a stray stream fails loud with NO_ADAPTER.
+Keyless model displacement is the disabled adapter row plus `installLlmReplay` filling the open seam on the settled root ctx in providers-catalog mode — never catch-all: with the adapter row disabled no adapter exists, so catch-all would leave `resolveModelInfo` unroutable and `compact-basic`'s post-step pressure check would warn every step instead of being provably inert (the published 128k `contextWindow` keeps it inert for small fixtures). The direct install rather than an inserted replay plugin row is deliberate: it returns the `ReplayHandle` the teardown consumption check needs. A scenario with no fixture leaves the seam empty, so a stray stream fails loud with NO_ADAPTER.

 `seedSession()` seeds cold sessions through the real persistence API — a throwaway `Context` mounting `SessionStore` + `SessionPersistenceJsonl` against the host's root, `create()` + `append()`, one `utimes` backdate for deterministic sidebar order (the `semantic-checkpoint.snapshot.ts` precedent) — never raw file writes, so the seeder knows nothing of bucket hashing, filename encoding, or compression, and the host's zstd default needs no boot knob. Seeds are validated at seed time (parseable, ending in `turn/end` — an open final turn would be mutated by resume's crash repair).

@@ -18,7 +18,7 @@ Web GUI 以一条真实组装链交付——chromium 页面 → client 插件 bu

 `launchWebScaffold()` 通过 vendored Loader 的 include 机制，从交付的 `apps/cli/cordis.yml` 启动真实 web 组合——与 `AppCLIEntry` 为 `dsh web` 驱动的是同一棵树、同一套机制。差异全部经 include patch 覆盖在这棵树上，即 ACP `cordis.snapshot.yml` 模式的进程内表达：临时 `persistenceRoot`；禁用 `workspace-context`（录制的 fixture 不得嵌入本仓库的 AGENTS.md）；禁用 `session-title-llm`（其发后不管的标题调用会与循环争抢会话的回放游标）；webserver 行钉到端口 0 加已构建 dist；无密钥模式下禁用 `llm-deepseek`。patch 的 id 一旦不再匹配任何行，boot 扫描会大声失败而不是漂移。boot 在临时工作区 `chdir` 下运行，使 api-gateway 的 `process.cwd()` 会话默认值、工具 cwd 与 fixture 一致；`dsh web` bin 自身的胶水（argv、profile json、AppCLIEntry）仍由 `smoke-real.e2e.ts` 中的无密钥 CLI 冒烟把守。初始化回滚和正常关闭都会先对 Cordis 树执行 dispose（资源释放），再删除 scaffold 持有的两个临时根目录；每项清理都会独立尝试，并会报告清理失败而不掩盖初始化失败。

-无密钥的模型替换 = 禁用适配器行的 patch 加 `installLlmReplay` 在停稳的根 ctx 上以提供方目录（providers-catalog）模式填充开放的 seam——绝不用 catch-all：适配器行被禁用后不存在任何适配器，catch-all 会让 `resolveModelContext` 无路由可走，`compact-basic` 的步后压力检查将步步告警，而不是被可证明地闲置（发布的 128k `contextWindow` 使该路径对小 fixture 保持闲置）。选择直接安装而非插入回放插件行是刻意的：直接安装返回收尾消费检查所需的 `ReplayHandle`。没有 fixture 的场景让 seam 保持空置，任何离群的流式调用都会以 NO_ADAPTER 大声失败。
+无密钥的模型替换 = 禁用适配器行的 patch 加 `installLlmReplay` 在停稳的根 ctx 上以提供方目录（providers-catalog）模式填充开放的 seam——绝不用 catch-all：适配器行被禁用后不存在任何适配器，catch-all 会让 `resolveModelInfo` 无路由可走，`compact-basic` 的步后压力检查将步步告警，而不是被可证明地闲置（发布的 128k `contextWindow` 使该路径对小 fixture 保持闲置）。选择直接安装而非插入回放插件行是刻意的：直接安装返回收尾消费检查所需的 `ReplayHandle`。没有 fixture 的场景让 seam 保持空置，任何离群的流式调用都会以 NO_ADAPTER 大声失败。

 `seedSession()` 通过真实持久化 API 播种冷会话——一次性 `Context` 挂载 `SessionStore` + `SessionPersistenceJsonl` 指向 host 的根目录，`create()` + `append()`，一次 `utimes` 回拨保证侧栏顺序确定（`semantic-checkpoint.snapshot.ts` 先例）——绝不裸写文件，因此播种器对桶哈希、文件名编码、压缩一无所知，host 的 zstd 默认值也无需任何启动开关。种子在播种时即校验（可解析、以 `turn/end` 结尾——未闭合的最终轮次会被恢复（resume）的崩溃修复改写）。

@@ -1,38 +0,0 @@
-# Agent Note: Consolidate gate scripts on already-present deps and builtins
-
-Status: proposed
-
-English | [中文](2026-07-26-consolidate-gate-scripts-on-existing-deps.zh.md)
-
-## Problem
-
-The `scripts/` gates mostly use the right tools (`node:fs` `globSync` in 15+ gates, mdast/micromark in the markdown gates), but a handful of stragglers hand-roll what a sibling gate already does with an existing dependency or builtin:
-
- **Duplicated fence scanners.** `scripts/md-fences.ts` (~55 lines, consumed by `doc-typecheck.ts`) and `extractEquivBlocks` in `scripts/verify-type-equiv.ts` (~39 lines) are two copies of the same regex line-scanner for fenced code blocks, while `scripts/verify-mermaid.ts` already extracts fences by visiting mdast `code` nodes via the shared `scripts/markdown.ts` helpers — and `markdownProseLines` in `markdown.ts` itself parses to mdast but then hand-tracks fence state with a second regex. The regex scanners only recognize backtick fences at column 0, so they silently disagree with the mdast-based gates on tilde and indented fences.
- **Hand-rolled argv parsing.** `parseOptions` in `scripts/publint-all.ts` and its near-identical copy in `scripts/verify-built-package-invariants.mjs` (~26 lines) step argv indexes manually, while sibling scripts (`verify-runtime-closure.ts`, `build-exe-for-python-sdk.ts`, `packages/sdk/scripts/src/args.ts`) already use the `node:util` `parseArgs` builtin.
- **Hand-rolled directory walks.** Five sites re-derive nested `readdirSync` walks that `globSync` covers: `verify-runtime-closure.ts` (packages + vendor manifests), `dev-web.ts` `discoverPluginDirs`, `verify-package-paths.ts` `realPackageNames`, `verify-client-domain-graph.ts` `listSources`, and `publint-all.ts` `addPath` (~55–65 lines total). `scripts/package-invariants.ts` shows the one-line `globSync` template.
-
-No new dependency is needed anywhere; every replacement is an existing devDep or a Node builtin.
-
-## Proposal
-
- Extract a shared ~10–15-line mdast fence helper (visiting `code` nodes for `lang`, `meta`, `value`, `position.start.line`) into `scripts/markdown.ts`; rewrite `doc-typecheck.ts` and `verify-type-equiv.ts` onto it; delete `md-fences.ts` and the duplicated scanner; drop the redundant fence regex in `markdownProseLines`.
- Replace both `parseOptions` copies with `parseArgs`.
- Replace the five straggler walks with `globSync`. Keep the walks in `check-workspace-constraints.ts` and `clean.ts`: they need dirent-level detail to diagnose malformed trees, which glob-by-pattern cannot report.
-
-## Alternatives considered
-
- **A new glob/walking dependency (`tinyglobby`, `fdir`).** Rejected: the builtin already won repo-wide; these are stragglers, not a gap.
- **`p-map` for `publint-all.ts`'s ~19-line ordered worker pool.** Deliberately left out: one new devDep for one small deletion is at the edge of the [dependency policy](../../implemented/process/2026-07-26-dependencies-over-hand-rolling.md) bar, and the pool's requirements (bounded workers, deterministic order, env override) are documented in the [parallel-gates note](../../implemented/process/2026-07-06-parallel-pre-push-gates.md). Fold it in only if `p-map` earns a second consumer.
- **Leaving the fence scanners.** Rejected: two drifting copies of a parser beside a third correct implementation is exactly the duplication the shared `markdown.ts` helper exists to prevent, and the column-0-backtick-only limitation is a latent inconsistency between sibling gates.
-
-## Acceptance criteria
-
- `md-fences.ts` is gone; `doc-typecheck` and `verify-type-equiv` extract fences through `scripts/markdown.ts`; `pnpm run doc-sync` passes with unchanged results on the current tree (any delta traces to a fence shape the regex scanners mishandled).
- Both CLIs parse via `parseArgs`; unknown options still fail loud.
- The five walk sites use `globSync`; the gates they feed pass unchanged.
-
-## Risks
-
- Behavioral deltas on pathological markdown: mdast honors tilde/indented fences the regex scanners ignored, so `doc-typecheck`'s opt-out ratio could shift if any stray fence shape exists in the docs tree; verify by running `doc-sync` before/after.
- `parseArgs` keeps the last value of a duplicated option instead of erroring — a dev-tool edge case the tests don't pin. (Strict mode still rejects a `--`-prefixed token where a value is expected, matching the current parsers.)
@@ -1,38 +0,0 @@
-# Agent Note: 把门禁脚本统一到已有依赖与内置模块上
-
-Status: proposed
-
-[English](2026-07-26-consolidate-gate-scripts-on-existing-deps.md) | 中文
-
-## 问题
-
-`scripts/` 下的门禁大多已经在用正确的工具（15 个以上的门禁使用 `node:fs` 的 `globSync`，markdown 门禁使用 mdast/micromark），但少数几个掉队的脚本仍在手写同类门禁早已用既有依赖或内置模块完成的事情：
-
- **重复的围栏扫描器。**`scripts/md-fences.ts`（约 55 行，由 `doc-typecheck.ts` 消费）和 `scripts/verify-type-equiv.ts` 中的 `extractEquivBlocks`（约 39 行）是同一个围栏代码块正则行扫描器的两份拷贝，而 `scripts/verify-mermaid.ts` 已经通过共享的 `scripts/markdown.ts` 辅助函数访问 mdast `code` 节点来提取代码围栏；`markdown.ts` 自己的 `markdownProseLines` 也是先解析成 mdast，再用第二个正则手工跟踪围栏状态。这两个正则扫描器只识别第 0 列的反引号围栏，因此在波浪线围栏和缩进围栏上与基于 mdast 的门禁悄悄不一致。
- **手写的 argv 解析。**`scripts/publint-all.ts` 中的 `parseOptions` 和 `scripts/verify-built-package-invariants.mjs` 中与之几乎相同的拷贝（约 26 行）手工推进 argv 下标，而同类脚本（`verify-runtime-closure.ts`、`build-exe-for-python-sdk.ts`、`packages/sdk/scripts/src/args.ts`）已经在使用 `node:util` 的内置 `parseArgs`。
- **手写的目录遍历。**五处代码各自重写了 `globSync` 已覆盖的嵌套 `readdirSync` 遍历：`verify-runtime-closure.ts` 对 packages 与 vendor manifest（元数据清单）的扫描、`dev-web.ts` 的 `discoverPluginDirs`、`verify-package-paths.ts` 的 `realPackageNames`、`verify-client-domain-graph.ts` 的 `listSources`，以及 `publint-all.ts` 的 `addPath`（合计约 55–65 行）。`scripts/package-invariants.ts` 展示了一行式的 `globSync` 模板。
-
-所有替换都不需要引入新依赖；每一处替换用的都是既有的 devDependency 或 Node 内置模块。
-
-## 提案
-
- 在 `scripts/markdown.ts` 中提取一个约 10–15 行的共享 mdast 围栏辅助函数（访问 `code` 节点，读取 `lang`、`meta`、`value`、`position.start.line`）；把 `doc-typecheck.ts` 和 `verify-type-equiv.ts` 改写到它上面；删除 `md-fences.ts` 和重复的扫描器；去掉 `markdownProseLines` 中冗余的围栏正则。
- 用 `parseArgs` 替换两份 `parseOptions` 拷贝。
- 用 `globSync` 替换那五处掉队的目录遍历。保留 `check-workspace-constraints.ts` 和 `clean.ts` 中的遍历：它们需要 dirent 级别的细节来诊断结构异常的目录树，按模式匹配的 glob 报告不了这些信息。
-
-## 曾考虑的替代方案
-
- **新的 glob/目录遍历依赖（`tinyglobby`、`fdir`）。**不予采纳：内置模块已在全仓库范围内胜出；这几处只是掉队者，不是能力缺口。
- **用 `p-map` 替换 `publint-all.ts` 中约 19 行的有序 worker 池。**刻意未纳入：为一次小删除引入一个新 devDependency，正处在[依赖策略](../../implemented/process/2026-07-26-dependencies-over-hand-rolling.md)门槛的边缘，而且该池的需求（worker 数量有界、确定性顺序、环境变量覆盖）已记录在[并行 pre-push 门禁决策记录](../../implemented/process/2026-07-06-parallel-pre-push-gates.md)中。仅当 `p-map` 赢得第二个消费方时再顺带纳入。
- **保留这两个围栏扫描器。**不予采纳：在第三个正确实现旁边放着两份逐渐漂移的解析器拷贝，正是共享的 `markdown.ts` 辅助函数要防止的那种重复；「只认第 0 列反引号」的限制也是同类门禁之间的潜在不一致。
-
-## 验收标准
-
- `md-fences.ts` 已删除；`doc-typecheck` 与 `verify-type-equiv` 通过 `scripts/markdown.ts` 提取代码围栏；`pnpm run doc-sync` 在当前代码树上通过且结果不变（如有差异，必须能追溯到正则扫描器处理有误的某种围栏形态）。
- 两个 CLI 都改用 `parseArgs` 解析；未知选项仍然大声失败。
- 五处遍历代码改用 `globSync`；它们供给的门禁保持原样通过。
-
-## 风险
-
- 病态 markdown 上的行为差异：mdast 会承认正则扫描器忽略的波浪线围栏和缩进围栏，因此如果文档树中存在任何零散的此类围栏形态，`doc-typecheck` 的 opt-out 比例可能变化；应在改动前后分别运行 `doc-sync` 加以验证。
- `parseArgs` 对重复出现的选项保留最后一个值而不报错——一个测试未固定的开发工具边缘用例。（严格模式下，需要取值处遇到以 `--` 开头的 token 仍会拒绝，与现有解析器行为一致。）
@@ -1,33 +0,0 @@
-# Agent Note: Replace the hand-rolled SSE parser in llm-deepseek with eventsource-parser
-
-Status: proposed
-
-English | [中文](2026-07-26-eventsource-parser-for-deepseek-sse.zh.md)
-
-## Problem
-
-`packages/llm/llm-deepseek/src/sse.ts` hand-implements Server-Sent Events parsing: a streaming `TextDecoder`, event-block splitting on `\r?\n\r?\n`, `data:` payload extraction and joining, comment/field skipping, the `[DONE]` sentinel, a `STREAM_CLOSED` error on EOF without it, and a flush of a final unterminated event block. The file is ~67 lines with ~108 lines of dedicated tests (`tests/sse.spec.ts`) re-proving SSE spec behavior — UTF-8 split across chunks, CRLF handling, multi-`data:` joining, no-space-after-colon — that a maintained parser already guarantees. Its only consumer is `adapter.ts` (`yield* translate(parseSse(response.body))`).
-
-This is exactly the surface `eventsource-parser` owns: the de-facto standard SSE parser (it underlies the Vercel AI SDK and the MCP SDK), zero-dependency, actively maintained, and already present in this repo's lockfile transitively via `@modelcontextprotocol/sdk` — so adopting it directly adds no new supply-chain surface in practice.
-
-## Proposal
-
-Replace `sse.ts` with `EventSourceParserStream` from `eventsource-parser/stream`: `response.body.pipeThrough(new TextDecoderStream()).pipeThrough(new EventSourceParserStream())`, keeping only the DeepSeek protocol shim (~10–25 lines): yield each event's `data`, terminate on `[DONE]`, and throw `LlmError('STREAM_CLOSED')` when the stream ends without the sentinel. All required builtins (`TextDecoderStream`, `pipeThrough`, async-iterable `ReadableStream`) exist at the Node ^22.19 engine floor. Delete the spec-conformance tests; keep the `[DONE]`/`STREAM_CLOSED`/EOF contract tests. Add `eventsource-parser` to `llm-deepseek`'s dependencies (its second runtime dep after schemastery). Update the [twin-adapters note](../../implemented/architecture/2026-06-13-twin-llm-adapters.md) and the `dsh-llm` JSDoc that brand this adapter "hand-rolled fetch + SSE parsing" in the same PR.
-
-The library also strips a leading BOM (the hand-rolled parser would fail to match `data:` after one) and offers `maxBufferSize` hardening the current parser lacks.
-
-## Alternatives considered
-
- **Keep the hand-rolled parser.** Defensible under the [twin-adapters decision](../../implemented/architecture/2026-06-13-twin-llm-adapters.md): the adapter is deliberately the hand-rolled design-verification twin of the pi-ai adapter. But the note's load-bearing distinction is owning the fetch/translate internals versus delegating to a full provider SDK; a ~700-byte SSE micro-parser is transport plumbing, not the design under verification. Whether that reading stands is the twin-note owner's call — this proposal explicitly needs their sign-off.
- **`createParser({onEvent})` callback API instead of the stream.** Works fed by a manual `TextDecoder` loop, but the `pipeThrough` composition deletes more of the hand-rolled code.
-
-## Acceptance criteria
-
- `sse.ts`'s parsing internals are gone; the remaining shim only encodes the DeepSeek `[DONE]`/`STREAM_CLOSED` protocol.
- `llm-deepseek` unit tests and the real-API e2e suite pass; keyless snapshots are unchanged (parsing is transport-internal and payload extraction is equivalent).
- The twin-adapters note and `dsh-llm` JSDoc no longer claim hand-rolled SSE parsing.
-
-## Risks
-
- One deliberate robustness deviation is lost: the hand-rolled parser flushes a final event block that lacks its terminating blank line, and `tests/sse.spec.ts` pins that a trailing `data: [DONE]` without `\n\n` still yields DONE. eventsource-parser is spec-strict and only dispatches on the blank line, so that shape becomes `STREAM_CLOSED`. Real providers and `dsh-llm-mock-server` always terminate events properly, so the pinned behavior is a robustness nicety, not an observed provider shape — drop the test, or keep a tiny buffer-tail check if the deviation is judged load-bearing.
- Dilutes the documented "hand-rolled" identity of the twin adapter; mitigated by updating the note in the same change rather than leaving the claim stale.
@@ -1,33 +0,0 @@
-# Agent Note: 用 eventsource-parser 替换 llm-deepseek 中手写的 SSE 解析器
-
-Status: proposed
-
-[English](2026-07-26-eventsource-parser-for-deepseek-sse.md) | 中文
-
-## 问题
-
-`packages/llm/llm-deepseek/src/sse.ts` 手写实现了 SSE（Server-Sent Events）解析：一个流式 `TextDecoder`、按 `\r?\n\r?\n` 切分事件块、提取并拼接 `data:` 载荷、跳过注释与其他字段、`[DONE]` 哨兵、在未见哨兵即 EOF 时抛出 `STREAM_CLOSED` 错误，以及对最后一个未终结事件块的 flush。该文件约 67 行，另有约 108 行专属测试（`tests/sse.spec.ts`）重复验证 SSE 规范行为——UTF-8 字符被切分到多个分片、CRLF 处理、多条 `data:` 拼接、冒号后无空格——而这些行为，持续维护的解析器早已有保证。它唯一的消费方是 `adapter.ts`（`yield* translate(parseSse(response.body))`）。
-
-这恰好是 `eventsource-parser` 负责的接口面：事实标准的 SSE 解析器（Vercel AI SDK 和 MCP SDK 都构建在它之上），零依赖，持续维护，并且已通过 `@modelcontextprotocol/sdk` 作为传递依赖出现在本仓库的 lockfile 中——因此直接采用它实际上不增加新的供应链接触面。
-
-## 提案
-
-用 `eventsource-parser/stream` 的 `EventSourceParserStream` 替换 `sse.ts`：`response.body.pipeThrough(new TextDecoderStream()).pipeThrough(new EventSourceParserStream())`，只保留 DeepSeek 协议垫层（约 10–25 行）：逐个产出事件的 `data`，遇到 `[DONE]` 终止，流在未见哨兵时结束则抛出 `LlmError('STREAM_CLOSED')`。所需的全部内置能力（`TextDecoderStream`、`pipeThrough`、可异步迭代的 `ReadableStream`）在 Node ^22.19 引擎下限即已存在。删除规范符合性测试；保留 `[DONE]`/`STREAM_CLOSED`/EOF 契约测试。将 `eventsource-parser` 加入 `llm-deepseek` 的依赖（这是它继 schemastery 之后的第二个运行时依赖）。在同一个 PR（Pull Request）中更新[孪生适配器 Agent Note（agent 决策记录）](../../implemented/architecture/2026-06-13-twin-llm-adapters.md)以及 `dsh-llm` 中把该适配器标为「手写 fetch + SSE 解析」的 JSDoc。
-
-该库还会剥离开头的 BOM（手写解析器在 BOM 之后会无法匹配 `data:`），并提供当前解析器缺少的 `maxBufferSize` 加固能力。
-
-## 曾考虑的替代方案
-
- **保留手写解析器。** 依据[孪生适配器决策](../../implemented/architecture/2026-06-13-twin-llm-adapters.md)，这一选择有辩护余地：该适配器有意作为 pi-ai 适配器的手写设计验证孪生体。但那份 Agent Note 起支撑作用的区分在于「自行持有 fetch/translate 内部实现」与「委托给完整的提供方 SDK」；一个约 700 字节的 SSE 微型解析器属于传输层管道，不是被验证的设计本身。这一解读是否成立由孪生 Agent Note 的所有者裁定——本提案明确需要其签署确认。
- **改用 `createParser({onEvent})` 回调 API 而非流。** 配合手动的 `TextDecoder` 循环可以工作，但 `pipeThrough` 组合方式能删除更多手写代码。
-
-## 验收标准
-
- `sse.ts` 的解析内部实现消失；剩下的垫层只编码 DeepSeek 的 `[DONE]`/`STREAM_CLOSED` 协议。
- `llm-deepseek` 单元测试与真实 API 的 e2e 套件通过；无密钥快照不变（解析属于传输层内部，载荷提取等价）。
- 孪生适配器 Agent Note 与 `dsh-llm` 的 JSDoc 不再声称手写 SSE 解析。
-
-## 风险
-
- 会失去一处有意为之的健壮性偏离：手写解析器会 flush 缺少终结空行的最后一个事件块，`tests/sse.spec.ts` 固定了「末尾的 `data: [DONE]` 即使没有 `\n\n` 也仍产出 DONE」这一行为。eventsource-parser 严格遵循规范，只在空行处分发事件，因此这种形态会变成 `STREAM_CLOSED`。真实提供方和 `dsh-llm-mock-server` 总是正确终结事件，所以被固定的行为只是健壮性上的锦上添花，并非实际观测到的提供方形态：可以删除该测试；若判定该偏离确有支撑作用，也可以保留一个小型的缓冲区尾部检查。
- 稀释了孪生适配器有文档记录的「手写」身份；缓解方式是在同一次变更中更新那份 Agent Note，而不是让声明陈旧下去。
@@ -1,32 +0,0 @@
-# Agent Note: Replace tool-web's regex HTML-to-markdown converter with turndown
-
-Status: proposed
-
-English | [中文](2026-07-26-turndown-for-tool-web-html-markdown.zh.md)
-
-## Problem
-
-`packages/web/tool-web/src/html.ts` (~86 lines, ~40 lines of dedicated tests) converts fetched HTML to markdown with regexes: strip script/style/noscript/comments, convert `<a>`/`<h1-6>`/`<li>`, decode numeric entities plus a 12-entry named-entity table, collapse whitespace. The module's own JSDoc says "A richer converter can replace it without changing the seam or tool schema", and the README's Known Limitations documents it as "a minimal regex converter, not an HTML parser — tables, images, and nested formatting are lost." The [web capability seam note](../../implemented/architecture/2026-06-24-web-capability-seam.md) assigns HTML→markdown to this package as presentation, so the swap point is exactly here. The converter's output is model-visible on every fetched HTML page; no keyless snapshot currently exercises `web_fetch`, so no expected outputs pin it.
-
-## Proposal
-
-Replace `htmlToMarkdown` with `turndown` (`new TurndownService().turndown(html)`), optionally with `turndown-plugin-gfm` for tables. The consumer switch in `fetch.ts` and the status-header/truncation-footer formatting stay. Wrap the call in try/catch falling back to the raw text path: the regex version could never throw; turndown on pathological HTML could. Delete `html.ts` and its conversion tests; keep tests for the fallback and the surrounding formatting. Update the README's Known Limitations to drop the regex-converter caveat.
-
-If the "deliberately minimal fallback" stance is preferred instead, a minimal variant still deletes the worst part: replace the entity-decoding third of the file (~30 lines: `decodeEntities`, `NAMED_ENTITIES`, `safeFromCodePoint`) with the zero-dependency `entities` package (already in the lockfile transitively), erasing the documented "about a dozen entities" limitation at near-zero risk.
-
-## Alternatives considered
-
- **`@mozilla/readability` + a DOM.** Solves a different problem (content extraction, not conversion) and drags a heavier DOM dependency; the seam only asks for markdown rendering of whatever the fetch returned.
- **Keep the regex converter.** It was an explicit v1 placeholder per its own JSDoc; keeping it means model-visible quality (tables, images, nested formatting) stays lost for the cost of maintaining bespoke entity tables.
- **The minimal `entities`-only variant.** Kept in the proposal as the fallback position; it deletes less but avoids the dependency-weight question entirely.
-
-## Acceptance criteria
-
- `web_fetch` renders tables/nested formatting via turndown (or, minimal variant: decodes all named entities), with the README limitation updated.
- Unit tests cover the fallback path; `pnpm run test` passes for the package.
- A keyless snapshot exercising `web_fetch` markdown rendering is added per testing policy (the missing snapshot coverage is part of the change, and it pins the new output).
-
-## Risks
-
- Model-visible output changes on every fetched HTML page — transcript drift is acceptable pre-release, and nothing currently pins the old output.
- Dependency weight: turndown's one dependency (`@mixmark-io/domino`) is a ~200 KB DOM that would enter the single-file-executable closure if tool-web ships in it ([single-exe note](../../implemented/architecture/2026-07-10-single-file-executable-sdk-runtime-distribution.md)); the minimal `entities` variant avoids this if closure size is the deciding factor.
@@ -1,32 +0,0 @@
-# Agent Note: 用 turndown 替换 tool-web 的正则 HTML 转 markdown 转换器
-
-Status: proposed
-
-[English](2026-07-26-turndown-for-tool-web-html-markdown.md) | 中文
-
-## 问题
-
-`packages/web/tool-web/src/html.ts`（约 86 行，另有约 40 行专属测试）用正则表达式把抓取到的 HTML 转成 markdown：剥离 script、style、noscript 标签与注释，转换 `<a>`/`<h1-6>`/`<li>`，解码数字实体外加一张 12 项的命名实体表，并折叠空白。该模块自身的 JSDoc 写明「A richer converter can replace it without changing the seam or tool schema」，README 的 Known Limitations 章节也把它记载为「a minimal regex converter, not an HTML parser — tables, images, and nested formatting are lost」。[web 能力 seam 决策记录](../../implemented/architecture/2026-06-24-web-capability-seam.md)把 HTML 转 markdown 作为呈现职责划归本包（package），因此替换点恰好就在这里。每个抓取到的 HTML 页面上，该转换器的输出都对模型可见；当前没有任何无密钥快照执行到 `web_fetch`，因此没有预期输出固定它的行为。
-
-## 提案
-
-用 `turndown` 替换 `htmlToMarkdown`（`new TurndownService().turndown(html)`），可选择配合 `turndown-plugin-gfm` 支持表格。`fetch.ts` 中的消费方分支与状态头、截断页脚的格式化保持不变。把调用包在 try/catch 中，失败时回退到原始文本路径：正则版本从不可能抛异常，而 turndown 处理病态 HTML 时可能抛出。删除 `html.ts` 及其转换测试；保留回退路径与外围格式化的测试。更新 README 的 Known Limitations 章节，移除正则转换器的警示说明。
-
-如果更倾向于「刻意保持最小回退实现」的立场，最小变体仍能删掉最糟的部分：用零依赖的 `entities` 包（已通过传递依赖存在于 lockfile 中）替换文件中占三分之一的实体解码部分（约 30 行：`decodeEntities`、`NAMED_ENTITIES`、`safeFromCodePoint`），以近乎为零的风险抹掉文档记载的「about a dozen entities」限制。
-
-## 曾考虑的替代方案
-
- **`@mozilla/readability` 加一个 DOM。** 它解决的是另一个问题（内容提取，而非格式转换），还会拖入更重的 DOM 依赖；这个 seam 只要求把抓取返回的内容渲染成 markdown。
- **保留正则转换器。** 按其自身 JSDoc 的说法，它本来就是明确的 v1 占位实现；保留它意味着模型可见的质量（表格、图片、嵌套格式）继续缺失，代价还是维护一套自制实体表。
- **仅引入 `entities` 的最小变体。** 已作为退守方案保留在提案中；它删得更少，但完全避开了依赖体积问题。
-
-## 验收标准
-
- `web_fetch` 经由 turndown 渲染表格与嵌套格式（或在最小变体下：解码全部命名实体），README 中的限制说明同步更新。
- 单元测试覆盖回退路径；该包的 `pnpm run test` 通过。
- 按测试政策补充一个执行 `web_fetch` markdown 渲染的无密钥快照（缺失的快照覆盖是本变更的一部分，它同时固定新输出）。
-
-## 风险
-
- 模型可见的输出在每个抓取到的 HTML 页面上都会变化：预发布阶段的 transcript（文本记录）漂移可以接受，且当前没有任何东西固定旧输出。
- 依赖体积：turndown 的唯一依赖（`@mixmark-io/domino`）是一个约 200 KB 的 DOM 实现，若 tool-web 进入单文件可执行文件，它会一并进入闭包（[single-exe 决策记录](../../implemented/architecture/2026-07-10-single-file-executable-sdk-runtime-distribution.md)）；若闭包体积是决定因素，最小的 `entities` 变体可以避开这一点。
@@ -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
-2026-07-26-dependency-swaps-rejected-by-nih-audit.md: a1d15b89f85b41e1044d9597dee6a1a0190240e6
-2026-07-26-dependency-swaps-rejected-by-nih-audit.zh.md: 4893784effdee7605f9194a80010b5a5033edbc1
+2026-07-26-dependency-swaps-rejected-by-nih-audit.md: f55bc2a9b7fb2a9599760734fca2a295665b95a2
+2026-07-26-dependency-swaps-rejected-by-nih-audit.zh.md: f740e0717cffd1b2ff4f3f5db8c9775afdfe79f0
@@ -18,7 +18,7 @@ Adopt the following dependency swaps. Rejected — per-item evidence below; a fu
 - **`vscode-languageserver-types` for lsp-local's wire-type subset**: ~80 type lines and ~45 guard lines, but upstream guards differ in both directions (accept `uri: undefined` the repo must reject; require `targetRange` the repo tolerates absent), and the initialize-result shapes live in `vscode-languageserver-protocol`, dragging `vscode-jsonrpc` in as a runtime dep — ~1 MB for 80 spec-exact lines.
 - **`json-rpc-2.0` for `dsh-jsonrpc`**: deletable correlation/dispatch is real (~100–130 lines) but the NDJSON wire must stay bit-identical for the hand-rolled Python SDK client, the package is single-maintainer, and the [GUI RPC note](../../implemented/architecture/2026-07-19-gui-layering-and-rpc-protocol.md) already treats this package as a frozen narrow surface. `vscode-jsonrpc` is a worse fit still (Content-Length framing, cancellation vocabulary the protocol lacks).
 - **`jsonrpcclient` for the Python SDK client**: v4 builds/parses messages only — ~20 lines — while the 500 lines that matter (subprocess lifecycle, threaded reader, id correlation, bidirectional server-role responses) stay; the library is in low-maintenance mode.
- **`eventsource-parser` for apiproxy's `readSse`**: only ~15 lines of framing are deletable, both wire ends are in-repo so spec conformance is moot, and it would add a dep to a browser-safe package. (Contrast with the [llm-deepseek proposal](../../proposed/simplification/2026-07-26-eventsource-parser-for-deepseek-sse.md), where a real provider sits across the wire.)
+- **`eventsource-parser` for apiproxy's `readSse`**: only ~15 lines of framing are deletable, both wire ends are in-repo so spec conformance is moot, and it would add a dep to a browser-safe package. (Contrast with the [llm-deepseek proposal](../../implemented/simplification/2026-07-26-eventsource-parser-for-deepseek-sse.md), where a real provider sits across the wire.)

 **Retry, timers, async:**

@@ -18,7 +18,7 @@ Status: rejected — 下列每一项替换在证据上都未达到净简化门
 - **以 `vscode-languageserver-types` 承担 lsp-local 的协议类型子集**：约 80 行类型加约 45 行守卫，但上游守卫在两个方向上都与本仓库不一致（接受本仓库必须拒绝的 `uri: undefined`；强制要求本仓库容忍缺失的 `targetRange`），而且 initialize 结果的形状住在 `vscode-languageserver-protocol` 里，会把 `vscode-jsonrpc` 拖成运行时依赖——为 80 行严格贴合规范的代码付出约 1 MB。
 - **以 `json-rpc-2.0` 替换 `dsh-jsonrpc`**：可删除的关联/分发代码确实存在（约 100–130 行），但 NDJSON 协议格式（wire format）必须与手写的 Python SDK 客户端逐位一致，该包只有单一维护者，且 [GUI RPC 决策](../../implemented/architecture/2026-07-19-gui-layering-and-rpc-protocol.md)已把这个包当作冻结的窄接口面对待。`vscode-jsonrpc` 更不合适（Content-Length 分帧、该协议并不具备的取消词汇）。
 - **以 `jsonrpcclient` 承担 Python SDK 客户端**：v4 只做消息的构造/解析——约 20 行——而真正要紧的 500 行（子进程生命周期、线程化读取器、id 关联、双向的服务端角色应答）全都保留；该库处于低维护模式。
- **以 `eventsource-parser` 替换 apiproxy 的 `readSse`**：可删除的分帧只有约 15 行，线路两端都在仓库内，规范符合性无关紧要，而且这会给一个浏览器安全的包添加依赖。（对比 [llm-deepseek 提案](../../proposed/simplification/2026-07-26-eventsource-parser-for-deepseek-sse.md)：那里线路对面是真实的提供方。）
+- **以 `eventsource-parser` 替换 apiproxy 的 `readSse`**：可删除的分帧只有约 15 行，线路两端都在仓库内，规范符合性无关紧要，而且这会给一个浏览器安全的包添加依赖。（对比 [llm-deepseek 提案](../../implemented/simplification/2026-07-26-eventsource-parser-for-deepseek-sse.md)：那里线路对面是真实的提供方。）

 **重试、定时器与异步：**

@@ -46,4 +46,4 @@ Apply the ordered relocate-condense-raise policy in [docs/AGENTS.md](../../../do

 ## Validation and PR hygiene

-Run at least `pnpm run doc-sync`, `pnpm run lint`, and `git diff --check`; JSDoc changes may regenerate catalogs. If a paired doc changed, follow [dsh-translate-docs](../dsh-translate-docs/SKILL.md) and run `pnpm run verify-translation-pairing --write`. The PR body should give word deltas, explain any deliberately long exception, and list checks.
+Run at least `pnpm run doc-sync`, `pnpm run lint`, and `git diff --check`; JSDoc changes may regenerate catalogs. If a paired doc changed, follow [dsh-translate-docs](../dsh-translate-docs/SKILL.md) and run `pnpm run verify-translation-pairing --write <pair>`. The PR body should give word deltas, explain any deliberately long exception, and list checks.
@@ -5,17 +5,33 @@ description: Use when creating or updating the bilingual counterpart of a doc in

 # Translating DeepSeek-Harness docs

-## Delegate to a subagent
-
-When this skill fires and translations need to be written, do not translate yourself: spawn a subagent to do the translation work. If you are that delegated subagent, skip this section; the sections from here on address the agent actually writing the translation.
-
 ## What this skill is

 **This skill is guidance, not a translation memory.** It is the workflow map for keeping `foo.md ↔ foo.zh.md` pairs consistent and natural in both languages. Both languages carry equal authority — a change is authored in either one, and that side is the source for that update. You are the translator: the rules below say what must hold, not how to phrase any particular sentence — phrasing judgment is yours, terminology is not.

-## Sources of truth (read, don't re-summarize)
+## Triage by change type — this decides everything else

-These are authoritative; read them at the source so this skill never drifts out of sync.
+- **Update** (pair exists, one side edited): follow [the update path](#the-update-path-briefing-driven). It is briefing-driven and deliberately cheap: no guidance-corpus reading, no git archaeology, smallest counterpart edit. Never re-translate a whole document to apply an update — a minimal update preserves the reviewed phrasing of everything that didn't change; a re-translation throws that review away.
+- **New pair** (no counterpart yet): follow [the whole-document path](#the-whole-document-path-new-pairs).
+- **Deleted or renamed doc**: delete or rename the counterpart and the `.i18n.yaml` alongside it — the gate reports an incomplete pair otherwise.
+
+Frozen Agent Notes under `.agents/notes/archived/` are not translation work. Their complete triplets are sealed by the archive verifier; never update, re-record, or repair either side after archival.
+
+## The update path (briefing-driven)
+
+Benchmarked on real pair updates from this repo's history, the briefing-driven path costs a fraction of a guidance-corpus-loading run at equal measured quality; the [briefed-updates Agent Note](../../notes/implemented/process/2026-07-26-briefed-minimal-translation-updates.md) holds the evidence.
+
+1. **Generate the briefing**: `pnpm run gen-translation-brief <any file of the pair>` (no arguments briefs every out-of-sync pair). The briefing maps the change at the narrowest safely aligned granularity — changed Markdown units (paragraph, table row, list item, heading), then whole heading sections, then whole document — and contains the authored side's diff since the last confirmed-consistent state, each changed unit's last-confirmed source, current source, and current counterpart text (with line numbers), the terminology rows the change touches, first-occurrence movement notes, and a digest of the binding update rules.
+2. **Mechanical-only diff? `--apply` it.** When every change lies inside code fences that the pair shares byte-identically, the briefing says so; `pnpm run gen-translation-brief --apply <pair>` splices the edited fences into the counterpart and structure-validates the result before writing — no subagent, no hand-editing.
+3. **Prose diff? Delegate to a subagent, passing the briefing** (or the command to generate it). The briefing is the translator's whole working set — the subagent does not re-read the guidance corpus (the rules digest, terminology rows, and each changed unit's three-way context are inline) and does not re-derive the diff. It escalates to the whole-document path's sources of truth only when the briefing leaves a specific decision genuinely unanswerable — an unlisted term with no precedent in the surrounding text, or a whole-document briefing (`BOTH sides changed`, or neither units nor sections align), which always means reconciling by hand under [translation-rules.md](../../../docs/i18n/translation-rules.md).
+4. **Smallest edit that covers the diff.** Preserve the reviewed phrasing of everything the diff does not touch, then verify the changed hunks clause by clause against the source: nothing added, nothing dropped, terminology per the inline rows, code spans verbatim.
+5. **Record and verify, scoped**: `pnpm run verify-translation-pairing --write <pair>` then `pnpm run verify-translation-pairing <pair>`. `--write` names exactly the pairs you confirmed — it refuses to run bare so a bulk re-record is always an explicit `--all`. The corpus-wide check still runs in `doc-sync`/CI; do not run it per-update.
+
+## The whole-document path (new pairs)
+
+When translations need to be written from scratch, the orchestrating agent does not translate: spawn a subagent to do the translation work. The translator reads the sources of truth below first, then translates the whole file into the other language — section by section for long documents, keeping each section's structure locked to the source as you go rather than fixing structure at the end.
+
+### Sources of truth (read, don't re-summarize)

 - **[docs/i18n/README.md](../../../docs/i18n/README.md)** — the pairing contract: the three-file pair (`foo.md`, `foo.zh.md`, `foo.i18n.yaml`), the consistency record's both-side blob hashes, the language-switcher lines, scope, and exclusions.
 - **[docs/i18n/translation-rules.md](../../../docs/i18n/translation-rules.md)** — how to translate: faithfulness, structure preservation, terminology discipline, typography (MUST/SHOULD levels).
@@ -23,29 +39,7 @@ These are authoritative; read them at the source so this skill never drifts out
 - **[docs/i18n/translation-prompt.md](../../../docs/i18n/translation-prompt.md)** — the automated pipeline's calibrated machine-consumed template. Agents using this skill do not render it; the terminology table is the only repository file the automated renderer injects, while this skill and `translation-rules.md` remain binding for agent-authored translations.
 - **[dsh-prose-standard](../dsh-prose-standard/SKILL.md)** — required prose coverage and editorial judgment. Apply it to both sides without adding or dropping source propositions.

-## Find the work
-
- `pnpm run verify-translation-pairing --list` prints every in-scope document as missing / out-of-sync / ok. Missing and out-of-sync rows are contract violations; the normal check rejects them.
- In a PR that edits paired docs, the work list is the diff itself: every changed side of a pair needs its counterpart updated and the pair re-recorded in the same PR, and the gate goes red if you forget.
-
-## Triage by change type
-
-Do not process every file the same way:
-
- **New pair** (no counterpart yet): whichever language exists — English or Chinese — translate the whole file into the other, section by section for long documents, keeping each section's structure locked to the source as you go rather than fixing structure at the end.
- **Update** (pair exists, one side edited): do NOT re-translate. The consistency record names the exact last-confirmed text of both sides — recover the edited side's previous state and diff:
-
-  ```sh
-  git cat-file -p <hash-from-i18n-yaml> > /tmp/last-confirmed.md
-  git diff --no-index /tmp/last-confirmed.md docs/foo.md
-  ```
-
-  Apply the smallest counterpart edits that cover that diff. A minimal update preserves the reviewed phrasing of everything that didn't change; a re-translation throws that review away.
- **Deleted or renamed doc**: delete or rename the counterpart and the `.i18n.yaml` alongside it — the gate reports an incomplete pair otherwise.
-
-Frozen Agent Notes under `.agents/notes/archived/` are not translation work. Their complete triplets are sealed by the archive verifier; never update, re-record, or repair either side after archival.
-
-## Translate
+### Translate

 - **Pass 1 — write, don't transpose.** Read a semantic unit, then restate it as a native technical author in the nearest [style sample's](../../../docs/i18n/style-samples.md) register. Preserve the required frame without forcing sentence-by-sentence correspondence.
 - **Pass 2 — verify against the source, clause by clause.** Fidelity is checked here, not written in: confirm nothing was added or dropped, every term follows the table, and each code span survived verbatim. Fix by rewriting the sentence natively, not by patching words into it.
@@ -54,15 +48,19 @@ Frozen Agent Notes under `.agents/notes/archived/` are not translation work. The
 - Code blocks are byte-identical across the pair, comments included. Relative links keep their `.md` targets; only the switcher line links `.zh.md`.
 - The pairing gate checks heading depths, fenced blocks, table row and column counts, list kinds, ordered-list starts, list item counts, and link targets. In Pass 2, manually verify list and table order, noncanonical list numbering, inline code, emphasis, meaning, terminology, and tone.

+## Find the work
+
+- `pnpm run verify-translation-pairing --list` prints every in-scope document as missing / out-of-sync / ok. Missing and out-of-sync rows are contract violations; the normal check rejects them.
+- `pnpm run gen-translation-brief` with no arguments prints the briefing for every out-of-sync pair.
+- In a PR that edits paired docs, the work list is the diff itself: every changed side of a pair needs its counterpart updated and the pair re-recorded in the same PR, and the gate goes red if you forget.
+
 ## Finish the pair

 1. Switcher: `[English](foo.md) | 中文` immediately after the Chinese file's H1, `English | [中文](foo.zh.md)` after the English file's H1 — add both if this is a new pair.
-2. Record consistency: `pnpm run verify-translation-pairing --write` recomputes and records both sides' full blob hashes in `foo.i18n.yaml`. The yaml diff in your PR is the reviewable statement "I confirmed these two say the same thing" — only run it after you actually have.
+2. Record consistency: `pnpm run verify-translation-pairing --write <pair>` recomputes and records both sides' full blob hashes in `foo.i18n.yaml`. The yaml diff in your PR is the reviewable statement "I confirmed these two say the same thing" — only run it after you actually have.
 3. No manifest entry is needed for an ordinary document: every in-scope source requires a pair. Change [scripts/translation-pairing.manifest.json](../../../scripts/translation-pairing.manifest.json) only when the owning policy documents a genuine generated, instructional, or bilingual-by-construction exclusion.
-
-## Verify the mechanical and human halves
-
-Run `pnpm run verify-translation-pairing`, then the rest of the Markdown gates (`pnpm run verify-md-wrap && pnpm run verify-md-links`, or full `pnpm run doc-sync` before the PR). Fix what they report and manually verify the obligations listed in Pass 2 that the gates do not encode. Keep the PR reviewable: state which pairs are new versus minimally updated and list 「待定术语」 prominently.
+4. Before the PR: the touched pairs are green under the scoped check; `pnpm run doc-sync` (which includes the corpus-wide pairing check plus `verify-md-wrap`/`verify-md-links`) runs once at PR level per [dsh-pre-push-checks](../dsh-pre-push-checks/SKILL.md), not inside each translation task.
+5. Keep the PR reviewable: state which pairs are new versus minimally updated and list 「待定术语」 prominently.

 ## How to respond to translation review

@@ -1,3 +1,3 @@
 # AGENTS.md — GitHub Actions

-Run Windows jobs under native `pwsh`.
+Run jobs on Windows runners (`windows-*` labels) under native `pwsh`. The pull-request `windows` job is not one of them: it runs Windows Node under Wine on hosted Linux, so its steps are bash — see the [Wine lane Agent Note](../.agents/notes/implemented/process/2026-07-27-wine-windows-gates-experiment.md).
@@ -281,41 +281,224 @@ jobs:
      - name: Run complete keyless Python suite
        run: uv run --python 3.10 --group test --project python/sdk pytest

-  # One standard Windows box shares setup across the required build/site checks
-  # and the observational portability inventory. Serial worker bounds keep this
-  # recovery path portable; Linux owns duplicate lint, coverage, and snapshots.
+  # The required pull-request Windows signal: the two blocking win32 surfaces
+  # (workspace build, production site) execute with real, checksum-verified
+  # Windows Node under Wine on standard hosted Linux. The master
+  # serial-windows job below keeps the complete native-kernel inventory —
+  # including the observational portability gates this lane does not run —
+  # on real windows-2025. Direct tool entrypoints stand in for pnpm's cmd
+  # shims, which a Linux-side install does not create; layout, fidelity
+  # limits, and measured timings live in
+  # .agents/notes/implemented/process/2026-07-27-wine-windows-gates-experiment.md
  windows:
    if: github.event_name == 'pull_request'
-    runs-on: windows-2025
-    name: windows node 24 / complete
+    runs-on: ubuntu-latest
+    name: windows node 24 / wine blocking
+    timeout-minutes: 15
    env:
-      DSH_COVERAGE_MAX_WORKERS: '1'
-      DSH_GATE_CONCURRENCY: '1'
-      DSH_PUBLINT_CONCURRENCY: '1'
+      WINEDEBUG: '-all'
+      WINEARCH: win64
+      # Skip Wine Mono / Gecko installers: Node needs neither.
+      WINEDLLOVERRIDES: 'mscoree,mshtml='
    steps:
      - uses: actions/checkout@v6
-
-      - name: Enable Developer Mode (symlink support)
-        shell: pwsh
-        run: >-
-          reg add "HKLM\SOFTWARE\Microsoft\Windows\CurrentVersion\AppModelUnlock"
-          /t REG_DWORD /f /v "AllowDevelopmentWithoutDevLicense" /d "1"
+        with:
+          persist-credentials: false

      - uses: actions/setup-node@v6
        with:
          node-version: ${{ env.PRIMARY_NODE_VERSION }}

-      # Extracting the many-file pnpm store cache is slower than a clean install,
-      # and saving it adds more latency after gates.
-      - name: Enable corepack and install (immutable)
-        shell: pwsh
+      - uses: actions/cache/restore@v4
+        with:
+          path: /home/runner/.local/share/pnpm/store/v11
+          key: ${{ runner.os }}-node-${{ env.PRIMARY_NODE_VERSION }}-pnpm-${{ hashFiles('pnpm-lock.yaml') }}
+          restore-keys: |
+            ${{ runner.os }}-node-${{ env.PRIMARY_NODE_VERSION }}-pnpm-
+
+      # Master's wine-apt-cache job seeds the default-branch scope every pull
+      # request can read; a save from this job only reaches reruns of the
+      # same merge ref.
+      - name: Compose Wine apt cache key
+        id: wine-cache-key
+        run: echo "key=wine-debs-${ImageOS:-linux}-${ImageVersion:-v0}" >> "$GITHUB_OUTPUT"
+
+      - uses: actions/cache@v4
+        with:
+          path: ~/wine-debs
+          key: ${{ steps.wine-cache-key.outputs.key }}
+
+      - name: Install dependencies and provision Wine concurrently
        run: |
          corepack enable
-          pnpm install --frozen-lockfile

-      - name: Run blocking and observational Windows gates concurrently
-        shell: pwsh
-        run: pnpm run check:ci:windows-complete
+          # Windows-lane install-time overrides. supportedArchitectures
+          # additionally materializes the win32-x64 platform packages
+          # (@esbuild/win32-x64, rolldown and rollup MSVC bindings) the
+          # Windows toolchain resolves at runtime; nodeLinker: hoisted lays
+          # node_modules out flat with real files because Windows Node under
+          # Wine does not realpath pnpm's isolated-layout symlinks. Neither
+          # override is recorded in the lockfile, so --frozen-lockfile stays
+          # valid. --ignore-scripts skips Linux lifecycle scripts no gate in
+          # this lane loads; the win32 binaries ship prebuilt.
+          cat >> pnpm-workspace.yaml <<'EOF'
+
+          nodeLinker: hoisted
+          supportedArchitectures:
+            os: [current, win32]
+            cpu: [current, x64]
+          EOF
+
+          pnpm install --frozen-lockfile --ignore-scripts &
+          install_pid=$!
+
+          provision_wine() {
+            set -euo pipefail
+            # Wine from the apt cache when present; else download the full
+            # dependency closure once and keep it for the next run. The
+            # `wine` dispatcher package (not bare `wine64`) is what puts a
+            # binary on PATH.
+            if compgen -G "$HOME/wine-debs/*.deb" > /dev/null; then
+              sudo apt-get install -y --no-install-recommends "$HOME"/wine-debs/*.deb
+            else
+              sudo apt-get update
+              sudo apt-get install -y --no-install-recommends --download-only wine
+              mkdir -p "$HOME/wine-debs"
+              cp /var/cache/apt/archives/*.deb "$HOME/wine-debs/" 2>/dev/null || true
+              sudo apt-get install -y --no-install-recommends wine
+            fi
+            WINE_BIN=''
+            for candidate in "$(command -v wine || true)" "$(command -v wine64 || true)" /usr/lib/wine/wine64; do
+              if [ -n "$candidate" ] && [ -x "$candidate" ]; then WINE_BIN="$candidate"; break; fi
+            done
+            [ -n "$WINE_BIN" ] || { echo '::error::no wine binary found after install'; exit 1; }
+            echo "WINE_BIN=$WINE_BIN" >> "$GITHUB_ENV"
+
+            # Windows Node for the repo's primary line, checksum-verified
+            # against the same dist directory.
+            version=$(curl -fsSL https://nodejs.org/dist/index.json \
+              | jq -r --arg p "v${PRIMARY_NODE_VERSION}." '[.[] | select(.version | startswith($p))][0].version')
+            echo "Windows Node: $version"
+            curl -fsSL -o "$RUNNER_TEMP/node-win.zip" \
+              "https://nodejs.org/dist/${version}/node-${version}-win-x64.zip"
+            curl -fsSL "https://nodejs.org/dist/${version}/SHASUMS256.txt" \
+              | awk -v a="node-${version}-win-x64.zip" '$2 == a { print $1 "  '"$RUNNER_TEMP"'/node-win.zip" }' \
+              | sha256sum --check -
+            unzip -q "$RUNNER_TEMP/node-win.zip" -d "$RUNNER_TEMP/node-win"
+            echo "NODE_WIN=$RUNNER_TEMP/node-win/node-${version}-win-x64/node.exe" >> "$GITHUB_ENV"
+
+            "$WINE_BIN" wineboot --init || true
+            wineserver -w || true
+          }
+          provision_wine &
+          wine_pid=$!
+
+          install_status=0
+          wait "$install_pid" || install_status=$?
+          wine_status=0
+          wait "$wine_pid" || wine_status=$?
+          if (( install_status != 0 )); then exit "$install_status"; fi
+          exit "$wine_status"
+
+      - name: Resolve entrypoints, link vue, smoke Windows Node
+        run: |
+          # Node under Wine cannot attach stdio to the Actions runner's pipes
+          # (Socket open EBADF at bootstrap), so every invocation runs through
+          # this wrapper: stdio to a regular file, replayed after exit.
+          cat > "$RUNNER_TEMP/wine-node.sh" <<'SH'
+          #!/usr/bin/env bash
+          set -u
+          log="$1"; shift
+          "$WINE_BIN" "$NODE_WIN" "$@" < /dev/null > "$log" 2>&1
+          status=$?
+          tail -n 300 "$log"
+          exit "$status"
+          SH
+          chmod +x "$RUNNER_TEMP/wine-node.sh"
+
+          resolve() {
+            local name="$1"; shift
+            for p in "$@"; do
+              if [ -f "$p" ]; then echo "$name=$PWD/$p" >> "$GITHUB_ENV"; return 0; fi
+            done
+            echo "::error::$name not found at any of: $*"; return 1
+          }
+          resolve TSC_JS node_modules/typescript/bin/tsc
+          resolve TSDOWN_JS node_modules/tsdown/dist/run.mjs
+          resolve VITEPRESS_JS website/node_modules/vitepress/bin/vitepress.js node_modules/vitepress/bin/vitepress.js
+
+          # VitePress links vue into the site's node_modules at build time;
+          # Wine cannot CREATE Windows symlinks (ENOTSUP) but follows
+          # pre-existing Unix ones, so lay the link down host-side.
+          if [ -d node_modules/vue ] && [ ! -e website/node_modules/vue ]; then
+            mkdir -p website/node_modules
+            ln -s ../../node_modules/vue website/node_modules/vue
+          fi
+
+          "$RUNNER_TEMP/wine-node.sh" "$RUNNER_TEMP/smoke.log" -p "'smoke: ' + process.platform + ' ' + process.arch + ' ' + process.version"
+
+      # The two blocking surfaces run concurrently, the same shape run-gates
+      # gives ci-windows-blocking on native Windows: `build` = tsc -b then
+      # tsdown, `production site` = the VitePress build. Both statuses are
+      # captured so one failure cannot hide the other's result.
+      - name: Run blocking Windows gates concurrently under Wine
+        run: |
+          build_gate() {
+            "$RUNNER_TEMP/wine-node.sh" "$RUNNER_TEMP/tsc.log" "$TSC_JS" -b --pretty false || return $?
+            "$RUNNER_TEMP/wine-node.sh" "$RUNNER_TEMP/tsdown.log" "$TSDOWN_JS"
+          }
+          site_gate() {
+            cd website
+            "$RUNNER_TEMP/wine-node.sh" "$RUNNER_TEMP/site.log" "$VITEPRESS_JS" build .
+          }
+          start=$SECONDS
+          build_gate > "$RUNNER_TEMP/build-gate.out" 2>&1 &
+          build_pid=$!
+          site_gate > "$RUNNER_TEMP/site-gate.out" 2>&1 &
+          site_pid=$!
+          build_status=0
+          wait "$build_pid" || build_status=$?
+          site_status=0
+          wait "$site_pid" || site_status=$?
+          echo "== build gate (exit $build_status, $((SECONDS - start))s elapsed) =="
+          tail -n 120 "$RUNNER_TEMP/build-gate.out"
+          echo "== production site gate (exit $site_status, $((SECONDS - start))s elapsed) =="
+          tail -n 120 "$RUNNER_TEMP/site-gate.out"
+          if (( build_status != 0 )); then exit "$build_status"; fi
+          exit "$site_status"
+
+      - name: Shut down wineserver
+        if: always()
+        run: wineserver -k 2>/dev/null || true
+
+  # Master seeds the Wine apt-archive cache in the default-branch scope,
+  # which every pull request's windows job can restore; saves from
+  # pull-request runs are scoped to their own merge ref and help nobody
+  # else. Runs in seconds when the image version already has a cache.
+  wine-apt-cache:
+    if: github.event_name == 'push' && github.ref == 'refs/heads/master'
+    name: wine apt cache
+    runs-on: ubuntu-latest
+    timeout-minutes: 10
+    steps:
+      - name: Compose Wine apt cache key
+        id: wine-cache-key
+        run: echo "key=wine-debs-${ImageOS:-linux}-${ImageVersion:-v0}" >> "$GITHUB_OUTPUT"
+
+      - uses: actions/cache@v4
+        id: wine-cache
+        with:
+          path: ~/wine-debs
+          key: ${{ steps.wine-cache-key.outputs.key }}
+
+      - name: Download the Wine dependency closure
+        if: steps.wine-cache.outputs.cache-hit != 'true'
+        run: |
+          sudo apt-get update
+          sudo apt-get install -y --no-install-recommends --download-only wine
+          mkdir -p "$HOME/wine-debs"
+          cp /var/cache/apt/archives/*.deb "$HOME/wine-debs/"
+          du -sh "$HOME/wine-debs"

  # Master pushes run only the serial reference jobs below.
  # Each host executes the complete, unsharded primary Node aggregate with one
@@ -693,8 +876,8 @@ jobs:
  # 'cancelled' and 'skipped'.
  all-checks-passed:
    name: all checks passed
-    # The required verdict must not add a separate standard-hosted billing dependency.
-    runs-on: dsh-enterprise-ubuntu-latest-32core-test
+    # This bookkeeping-only verdict must not depend on custom-pool provisioning.
+    runs-on: ubuntu-latest
    needs: [node-24, node-24-coverage, node-24-consumers, node-compat, python-sdk, windows]
    if: always() && github.event_name == 'pull_request'
    steps:
@@ -12,7 +12,7 @@ DeepSeek Harness SDK is a plugin-based agent harness on vendored Cordis: **every
 vendor/      Vendored Cordis source — manifest + sync procedure in vendor/README.md
 packages/    @deepseek-ai/dsh-<pkg> workspaces at packages/<group>/<pkg>/
  core/        product API spine: session, system-prompt, tools, agent, agent-loop
-  llm/         LLM seam + DeepSeek adapters (hand-rolled + pi-ai design twin)
+  llm/         LLM seam + DeepSeek adapters (direct-fetch + pi-ai design twin)
  bash/        bash executor seam + local impl + model-facing bash tools
  pty/         persistent PTY seam/backend/tools
  fs/          filesystem seam + local impl + policy gate + read/write/edit tools
@@ -144,7 +144,7 @@ it('renders the fixture run_code turn: code parent row, nested sub-rows, error s
      "errorSubRow": true,
      "parentRow": "CodeRead the notes files and summarize",
      "subRows": [
-        "$List notes",
+        "BashList notes",
        "Readnotes/demo.txt",
        "Readnotes/missing.txt",
      ],
@@ -213,9 +213,9 @@ it('trajectory and waterfall surface the run_code sub-calls with real timing', a
  }).toMatchInlineSnapshot(`
    {
      "subCells": [
-        "#53Subbash · {"command":"ls notes","description":"List notes"}+0.8s",
-        "#54Subread · {"path":"notes/demo.txt"}+0.8s",
-        "#55Subread · {"path":"notes/missing.txt"}+0.8s",
+        "#51Subbash · {"command":"ls notes","description":"List notes"}+0.8s",
+        "#52Subread · {"path":"notes/demo.txt"}+0.8s",
+        "#53Subread · {"path":"notes/missing.txt"}+0.8s",
      ],
    }
  `)
@@ -64,7 +64,7 @@ const CONFIG_PATH = join(REPO_ROOT, 'apps/cli/cordis.yml')

 // Replay publishes the provider catalog the gateway routes to (providers
 // mode, never catch-all: with llm-deepseek disabled no adapter exists, so a
-// catch-all would leave resolveModelContext unroutable and compact-basic's
+// catch-all would leave resolveModelInfo unroutable and compact-basic's
 // post-step pressure check would warn every step). The published
 // contextWindow keeps that pressure path provably inert for small fixtures.
 const REPLAY_PROVIDERS = [{ id: 'deepseek', name: 'DeepSeek', models: [{ id: 'deepseek-v4-flash', contextWindow: 128_000 }] }]
@@ -0,0 +1,191 @@
+// @vitest-environment jsdom
+// Todo display snapshot over the BUILT client graph (the code-mode-fixture
+// idiom: real bundles via AppWebEntry, keyless FixtureApiClient transport).
+// Opens the fixture history session and pins the todo_write turn's two
+// surfaces: the dedicated TodoRow in the chat flow (keyed toolview, summary
+// derived from the call args) and the TodoPanel plan strip riding the
+// 'conversation.input.dock' slot (fed by ConversationSnapshot.todos, seeded
+// by the tail history page), including the collapse interaction.
+import { readFileSync } from 'node:fs'
+import { join } from 'node:path'
+import { act, cleanup, fireEvent, screen, waitFor, within } from '@testing-library/react'
+import { afterEach, beforeEach, expect, it, vi } from 'vitest'
+import type { WebBootEntry } from '@deepseek-ai/dsh-client-modules/client'
+import { AppWebEntry } from '@deepseek-ai/dsh-client-web'
+
+const PLUGINS: readonly (WebBootEntry & { dir: string })[] = [
+  { id: '@deepseek-ai/dsh-client-connection', dir: 'connection', url: '/plugins/connection.js', rev: 'fx', inject: [], immediately: true },
+  { id: '@deepseek-ai/dsh-client-runtime', dir: 'runtime', url: '/plugins/runtime.js', rev: 'fx', inject: ['@deepseek-ai/dsh-client-connection'], immediately: true },
+  { id: '@deepseek-ai/dsh-client-ui-theme', dir: 'ui-theme', url: '/plugins/ui-theme.js', rev: 'fx', inject: [], immediately: true },
+  { id: '@deepseek-ai/dsh-client-locale', dir: 'locale', url: '/plugins/locale.js', rev: 'fx', inject: [], immediately: true },
+  { id: '@deepseek-ai/dsh-client-ui-layout', dir: 'ui-layout', url: '/plugins/ui-layout.js', rev: 'fx', inject: ['@deepseek-ai/dsh-client-runtime'] },
+  { id: '@deepseek-ai/dsh-client-ui-sidebar', dir: 'ui-sidebar', url: '/plugins/ui-sidebar.js', rev: 'fx', inject: ['@deepseek-ai/dsh-client-ui-layout'] },
+  { id: '@deepseek-ai/dsh-client-ui-conversation', dir: 'ui-conversation', url: '/plugins/ui-conversation.js', rev: 'fx', inject: ['@deepseek-ai/dsh-client-ui-layout'] },
+  {
+    id: '@deepseek-ai/dsh-client-ui-workspace',
+    dir: 'ui-workspace',
+    url: '/plugins/ui-workspace.js',
+    rev: 'fx',
+    inject: [
+      '@deepseek-ai/dsh-client-runtime',
+      '@deepseek-ai/dsh-client-ui-conversation',
+      '@deepseek-ai/dsh-client-ui-sidebar',
+    ],
+  },
+]
+
+const bundles = new Map(PLUGINS.map(plugin => [
+  plugin.url,
+  readFileSync(join(process.cwd(), 'packages/client', plugin.dir, 'lib/client.js'), 'utf8'),
+]))
+
+interface FixtureWindow extends Window {
+  __DSH_BOOT__?: { rev: string; entries: WebBootEntry[] }
+  __ModuleLoader__?: unknown
+}
+
+class ResizeObserverStub {
+  observe(): void {}
+  disconnect(): void {}
+  unobserve(): void {}
+}
+
+const win = window as FixtureWindow
+let unmount: (() => void) | undefined
+
+beforeEach(() => {
+  localStorage.clear()
+  document.title = 'DeepSeek Harness'
+  vi.stubGlobal('ResizeObserver', ResizeObserverStub)
+  vi.stubGlobal('requestAnimationFrame', (callback: FrameRequestCallback) =>
+    setTimeout(() => { callback(0) }, 0) as unknown as number)
+  vi.stubGlobal('cancelAnimationFrame', (id: number) => { clearTimeout(id) })
+})
+
+afterEach(() => {
+  act(() => { unmount?.() })
+  unmount = undefined
+  cleanup()
+  delete win.__DSH_BOOT__
+  delete win.__ModuleLoader__
+  delete (globalThis as Record<string, unknown>).__fxTiming
+  document.body.innerHTML = ''
+  document.head.querySelectorAll('style[data-plugin]').forEach((style) => { style.remove() })
+  document.title = ''
+  history.replaceState(null, '', '/')
+  vi.unstubAllGlobals()
+})
+
+/** Boot the complete built client graph against the populated fixture branch. */
+function boot(): void {
+  history.replaceState(null, '', '/?fixture')
+  const root = document.createElement('div')
+  root.id = 'root'
+  document.body.appendChild(root)
+  win.__DSH_BOOT__ = { rev: 'fx', entries: PLUGINS.map(({ dir: _dir, ...plugin }) => plugin) }
+  act(() => {
+    const entry = new AppWebEntry(root, {
+      fetchBundle: (url) => {
+        const code = bundles.get(url)
+        return code === undefined ? Promise.reject(new Error(`missing built bundle ${url}`)) : Promise.resolve(code)
+      },
+      executeBundle: (code) => { (0, eval)(code) },
+    })
+    void entry.run()
+    unmount = () => { entry.dispose() }
+  })
+}
+
+/** Collapse decorative whitespace while preserving the text a user sees. */
+function visibleText(element: Element): string {
+  return (element.textContent ?? '').replace(/\s+/g, ' ').trim()
+}
+
+/** Open the fixture history session (the alpha log carrying the todo_write turn) and wait for its tail. */
+async function openFixtureSession(): Promise<void> {
+  const tree = await screen.findByRole('tree', { name: 'Sessions' }, { timeout: 10_000 })
+  // Anchor on the expandable Workspace group row: the title and the blank
+  // session row can both read "fixture", and the session-count meta shifts
+  // when a blank session joins the group.
+  const group = (await within(tree).findAllByText('fixture'))
+    .map(el => el.closest<HTMLElement>('[role="treeitem"]'))
+    .find(el => el?.getAttribute('aria-expanded') !== null)
+  if (group === null || group === undefined) throw new Error('fixture Workspace group missing')
+  if (group.getAttribute('aria-expanded') === 'false') {
+    fireEvent.click(within(group).getByText('fixture'))
+    await waitFor(() => {
+      expect(group.getAttribute('aria-expanded')).toBe('true')
+    })
+  }
+  const session = await within(tree).findByText('Fixture 历史会话')
+  fireEvent.click(session)
+  await waitFor(() => {
+    expect(document.querySelector('[data-sample="todo-row"]')).not.toBeNull()
+  }, { timeout: 10_000 })
+}
+
+it('renders the todo_write turn: dedicated tool row + the dock plan strip', async () => {
+  boot()
+  await openFixtureSession()
+
+  const row = document.querySelector('[data-sample="todo-row"]')
+  if (row === null) throw new Error('todo row missing')
+  const panel = document.querySelector('[data-testid="todo-panel"]')
+  if (panel === null) throw new Error('todo panel missing from the input dock')
+
+  // Header spans are adjacent inline nodes; textContent joins "To-dos" +
+  // "1/3…" with no space (visual gap is CSS gap: 10px, not a text node).
+  expect({
+    row: visibleText(row),
+    rowState: row.getAttribute('data-state'),
+    panelHeader: visibleText(panel.querySelector('button') ?? panel),
+    panelItems: [...panel.querySelectorAll('li')].map(item => ({
+      status: item.getAttribute('data-status'),
+      text: visibleText(item),
+    })),
+  }).toMatchInlineSnapshot(`
+    {
+      "panelHeader": "To-dos1/3 tasks · 1 in progress",
+      "panelItems": [
+        {
+          "status": "completed",
+          "text": "梳理需求",
+        },
+        {
+          "status": "in_progress",
+          "text": "实现 fixture 样本",
+        },
+        {
+          "status": "pending",
+          "text": "浏览器验收",
+        },
+      ],
+      "row": "☰更新任务清单1/3 已完成 · 实现 fixture 样本",
+      "rowState": "ok",
+    }
+  `)
+})
+
+it('collapses the plan strip to the count summary and restores it', async () => {
+  boot()
+  await openFixtureSession()
+
+  const panel = document.querySelector('[data-testid="todo-panel"]')
+  if (panel === null) throw new Error('todo panel missing from the input dock')
+  const header = panel.querySelector('button')
+  if (header === null) throw new Error('todo panel header missing')
+
+  fireEvent.click(header)
+  expect({
+    collapsedHeader: visibleText(header),
+    listGone: panel.querySelector('ul') === null,
+  }).toMatchInlineSnapshot(`
+    {
+      "collapsedHeader": "To-dos1/3 tasks · 1 in progress",
+      "listGone": true,
+    }
+  `)
+
+  fireEvent.click(header)
+  expect(panel.querySelectorAll('li')).toHaveLength(3)
+})
@@ -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
-architecture.md: b426891c0483f42a64b597632cf1871aff79ca2d
-architecture.zh.md: 13feefb6854e79ddee38602902d325a789fd7744
+architecture.md: 34e56dd955e6ac5ad8fad236bdf7ae9cfc810a83
+architecture.zh.md: ea105b869eba799d00dfaa134ed34d14c7505fc8
@@ -91,7 +91,7 @@ forever:
      agent/pre-step
      snapshot the derived messages (the reconstruction boundary)
      'step/start'
-      agent/request (config only) -> log request/header -> checkpoint -> llm/stream (frozen)
+      agent/request -> prepare reasoning/default under turn signal -> log request/header -> checkpoint -> llm/stream (frozen, registration-bound)
      on final adapter-path or terminal in-band failure:
        'step/end'
        agent/request-error(original error, failure facts, immutable prior failures, signal)
@@ -125,7 +125,7 @@ Pruning precedes summaries; overflow retries require durable progress. Bounded r

 Adapter failures close the step before `agent/request-error` with exact `Error`, `LlmFailure`, and history. Retries open steps; success clears history; exhaustion stores failure on `turn/end`. Failed chunks commit nothing.

-Other failures use `agent/error`. Cancellation and disposal beat recovery; undispatched tools get synthetic `tool/call`/`ABORTED_BEFORE_DISPATCH` pairs. The signal retires before `turn/end`. Effective `cancel()` emits its cause, clears queues, and aborts; observers cannot veto, idle calls emit nothing, and durability records `aborted`. Disposal awaits quiescence ([decision](../.agents/notes/implemented/architecture/2026-07-16-explicit-turn-cancellation.md)).
+Other failures use `agent/error`. Cancellation and disposal beat recovery; the turn signal also cancels asynchronous model-capability preparation before any request header is committed, and undispatched tools get synthetic `tool/call`/`ABORTED_BEFORE_DISPATCH` pairs. The signal retires before `turn/end`. Effective `cancel()` emits its cause, clears queues, and aborts; observers cannot veto, idle calls emit nothing, and durability records `aborted`. Disposal awaits quiescence ([decision](../.agents/notes/implemented/architecture/2026-07-16-explicit-turn-cancellation.md)).

 Session events are turn-enclosed; reload closes an interrupted tail with a synthetic `interrupted` turn end. Post-close failures use `agent/error`. Each turn has one [TurnEndReason](core-data-structures/session.md#why-a-turn-ended-turnendreasonmap).

@@ -91,7 +91,7 @@ forever:
      agent/pre-step
      snapshot the derived messages (the reconstruction boundary)
      'step/start'
-      agent/request (config only) -> log request/header -> checkpoint -> llm/stream (frozen)
+      agent/request -> prepare reasoning/default under turn signal -> log request/header -> checkpoint -> llm/stream (frozen, registration-bound)
      on final adapter-path or terminal in-band failure:
        'step/end'
        agent/request-error(original error, failure facts, immutable prior failures, signal)
@@ -125,7 +125,7 @@ forever:

 适配器故障会先关闭步骤，再进入 `agent/request-error`；该事件会收到准确的 `Error`、`LlmFailure` 和历史记录。重试会开启步骤；成功会清除历史记录；重试耗尽后，故障存入 `turn/end`。失败分片不会提交任何内容。

-其他故障使用 `agent/error`。取消和资源释放均优先于恢复；尚未分派的工具会得到合成的 `tool/call`/`ABORTED_BEFORE_DISPATCH` 对。信号会在 `turn/end` 前失效。实际生效的 `cancel()` 会发出原因、清空队列并中止；观察方不能否决该操作，空闲状态下的调用不发出任何事件，持久化会记录 `aborted`。dispose（资源释放）会等待系统停稳（[决策](../.agents/notes/implemented/architecture/2026-07-16-explicit-turn-cancellation.md)）。
+其他故障使用 `agent/error`。取消和资源释放均优先于恢复；轮次信号还会在提交任何请求头之前取消异步模型能力准备，尚未分派的工具会得到合成的 `tool/call`/`ABORTED_BEFORE_DISPATCH` 对。信号会在 `turn/end` 前失效。实际生效的 `cancel()` 会发出原因、清空队列并中止；观察方不能否决该操作，空闲状态下的调用不发出任何事件，持久化会记录 `aborted`。dispose（资源释放）会等待系统停稳（[决策](../.agents/notes/implemented/architecture/2026-07-16-explicit-turn-cancellation.md)）。

 会话事件均位于轮次边界内；重新加载会用合成的 `interrupted` 轮次结束事件闭合中断的日志尾部。关闭后的故障使用 `agent/error`。每个轮次有一个 [TurnEndReason](core-data-structures/session.md#why-a-turn-ended-turnendreasonmap)。

@@ -565,18 +565,19 @@ Requires: `llm`
 /**
 * Plugin config, validated by the same-named schemastery schema. Every field
 * is optional in yml: credentials/endpoint fall back to the environment (a
- * missing API key fails plugin load, not the first call), and omitted
- * thinking fields send nothing on the wire, so the provider default applies.
+ * missing API key fails plugin load, not the first call), omitted thinking
+ * mode uses the provider default, and omitted reasoning effort resolves to
+ * `high`.
 */
 export interface Config {
  /** API key; falls back to $DEEPSEEK_API_KEY. Required one way or the other. */
  apiKey?: string
  /** Endpoint base; falls back to $DEEPSEEK_BASE_URL, then the public API. */
  baseURL?: string
-  /** Thinking-mode default for every request (provider default: enabled). */
+  /** Deployment thinking policy; `disabled` limits every conversation request to `off`. */
  thinking?: 'enabled' | 'disabled'
-  /** Thinking effort (only meaningful with thinking enabled). */
-  reasoningEffort?: 'high' | 'max'
+  /** Default thinking effort (default `high`); `off` disables thinking per request. */
+  reasoningEffort?: 'off' | 'high' | 'max'
  /** Positive context capacity used when the selected model has no exact value. */
  defaultContextWindow?: number
  /** Advisory models shown by discovery consumers; defaults to V4 Flash and V4 Pro. */
@@ -585,7 +586,7 @@ export interface Config {
  streamIdleTimeoutMs?: number
 }

-/** One optional model entry advertised by the hand-written adapter. */
+/** One optional model entry advertised by the direct-fetch adapter. */
 export interface DeepSeekCatalogModel {
  /** Wire model id accepted by the configured endpoint. */
  id: string
@@ -598,7 +599,7 @@ export interface DeepSeekCatalogModel {
 }
 ```

-Source: [`packages/llm/llm-deepseek/src/index.ts:34`](../packages/llm/llm-deepseek/src/index.ts)
+Source: [`packages/llm/llm-deepseek/src/index.ts:35`](../packages/llm/llm-deepseek/src/index.ts)

 ## `@deepseek-ai/dsh-llm-pi-ai`

@@ -622,7 +623,7 @@ export interface PiAiProviderProfile {
  /** Provider request headers; Harness attribution wins reserved names. */
  headers?: Record<string, string>
  /** Provider-neutral pi-ai reasoning level. */
-  reasoning?: ThinkingLevel
+  reasoning?: ModelThinkingLevel
  /** Token budgets used by reasoning providers that support them. */
  thinkingBudgets?: ThinkingBudgets
  /** Prompt-cache retention preference. */
@@ -638,7 +639,7 @@ export interface PiAiProviderProfile {
 }
 ```

-Depends on: `CacheRetention` (`@earendil-works/pi-ai`) · `ThinkingBudgets` (`@earendil-works/pi-ai`) · `ThinkingLevel` (`@earendil-works/pi-ai`) · `Transport` (`@earendil-works/pi-ai`)
+Depends on: `CacheRetention` (`@earendil-works/pi-ai`) · `ModelThinkingLevel` (`@earendil-works/pi-ai`) · `ThinkingBudgets` (`@earendil-works/pi-ai`) · `Transport` (`@earendil-works/pi-ai`)

 Source: [`packages/llm/llm-pi-ai/src/config.ts:48`](../packages/llm/llm-pi-ai/src/config.ts)

@@ -688,7 +689,7 @@ export interface ReplayModelConfig {
 }
 ```

-Source: [`packages/support/llm-replay/src/index.ts:590`](../packages/support/llm-replay/src/index.ts)
+Source: [`packages/support/llm-replay/src/index.ts:598`](../packages/support/llm-replay/src/index.ts)

 ## `@deepseek-ai/dsh-llm-retry`

@@ -1705,7 +1706,7 @@ Source: [`packages/tasks/tool-tasks/src/index.ts:23`](../packages/tasks/tool-tas
 Requires: `tools` · `web` · `systemPrompt`

 ```ts config-catalog
-/** Plugin config: which web tools to register, the source cap, and per-tool budgets. */
+/** Plugin config: which web tools to register, the source cap, per-tool budgets, and the fetch output cap. */
 export interface Config {
  /** Register `web_search`. Defaults to true. */
  search?: boolean
@@ -1717,10 +1718,12 @@ export interface Config {
  fetchTimeoutMs?: number
  /** Cooperative timeout budget (ms) for `web_search`. Defaults to 30000. */
  searchTimeoutMs?: number
+  /** Cap on source characters converted and complete `web_fetch` output characters. Defaults to 200000. */
+  fetchMaxOutputChars?: number
 }
 ```

-Source: [`packages/web/tool-web/src/index.ts:29`](../packages/web/tool-web/src/index.ts)
+Source: [`packages/web/tool-web/src/index.ts:35`](../packages/web/tool-web/src/index.ts)

 ## `@deepseek-ai/dsh-tool-workflow`

@@ -1830,7 +1833,7 @@ export interface TuiConfig {
 }
 ```

-Source: [`packages/ui/tui/src/index.ts:270`](../packages/ui/tui/src/index.ts)
+Source: [`packages/ui/tui/src/index.ts:273`](../packages/ui/tui/src/index.ts)

 ## `@deepseek-ai/dsh-tui-demo`

@@ -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
-adding-an-llm-adapter.md: f20442b8c2ce823452a3ea13409f202185d12d04
-adding-an-llm-adapter.zh.md: 2864dd1e18742c7449e24f22504a5a38976ab450
+#   pnpm run verify-translation-pairing --write docs/cookbook/adding-an-llm-adapter.md
+adding-an-llm-adapter.md: a7f9dced70041653a0cb815147a07b6386d79e3e
+adding-an-llm-adapter.zh.md: 3515927585201326b713bb03cd863886ce7846bd
@@ -2,7 +2,7 @@

 English | [中文](adding-an-llm-adapter.zh.md)

-How to connect a new model provider. Reference implementations: `packages/llm/llm-deepseek` (hand-rolled HTTP/SSE) and `packages/llm/llm-pi-ai` (wrapping an LLM library). Read the `StreamChunk` doc in `packages/llm/llm/src/types.ts` first — it records the protocol conventions both adapters were verified against.
+How to connect a new model provider. Reference implementations: `packages/llm/llm-deepseek` (direct HTTP, SSE framed by `eventsource-parser`) and `packages/llm/llm-pi-ai` (wrapping an LLM library). Read the `StreamChunk` doc in `packages/llm/llm/src/types.ts` first — it records the protocol conventions both adapters were verified against.

 ## The shape

@@ -32,7 +32,7 @@ Registration is effect-based (HMR-safe); one adapter per provider route — dupl
 - A `GenerateOptions` field your provider cannot honor (e.g. a `stop` list on a provider without stop sequences): throw `LlmError(..., 'UNSUPPORTED')` rather than silently dropping it.
 - If the provider requires response ids, signatures, or other native metadata on follow-up calls, emit the minimal lossless-JSON projection as `finish.replayState`. Validate it when rebuilding history. `LlmService` passes it only when the historical provider route and target provider route are currently owned by the exact same adapter instance; your adapter decides whether same-model, cross-model, or cross-provider restoration is legal. Never infer native replay from provider/model names alone when state is absent.

-Provider-specific request knobs (thinking modes, effort levels) belong in the ADAPTER's Config, not in `GenerateOptions` — the core vocabulary stays provider-neutral.
+Provider-specific thinking-mode toggles remain in the adapter's Config. Exact model metadata uses one provider-neutral capability seam: implement `resolveModel()` with provider/model identity and optional `context` and `reasoning` fields, declare a configured `defaultEffort` only when one exists, and honor the resolver's optional `AbortSignal`. Reasoning efforts are ordered opaque ids mapped to provider requests by the adapter. Preserve the adapter's authoritative selectable list, including an adapter-defined `off` when supported, without exposing final wire spellings or clamping unsupported values; an id need not equal its wire representation.

 ## Structure that worked

@@ -2,7 +2,7 @@

 [English](adding-an-llm-adapter.md) | 中文

-如何接入一个新的模型提供方。参考实现：`packages/llm/llm-deepseek`（手写 HTTP/SSE）与 `packages/llm/llm-pi-ai`（封装 LLM 库）。请先阅读 `packages/llm/llm/src/types.ts` 中的 `StreamChunk` 文档——它记录了两个适配器都经过验证的协议约定。
+如何接入一个新的模型提供方。参考实现：`packages/llm/llm-deepseek`（直接 HTTP，SSE 由 `eventsource-parser` 分帧）与 `packages/llm/llm-pi-ai`（封装 LLM 库）。请先阅读 `packages/llm/llm/src/types.ts` 中的 `StreamChunk` 文档——它记录了两个适配器都经过验证的协议约定。

 ## 基本形态

@@ -32,7 +32,7 @@ export function apply(ctx: Context, config: Config) {
 - 如果 `GenerateOptions` 中某个字段你的提供方无法支持（例如提供方不支持 stop sequences 时收到 `stop` 列表）：抛出 `LlmError(..., 'UNSUPPORTED')`，而非静默丢弃。
 - 如果提供方在后续调用中需要响应 ID、签名或其他原生元数据，请将其最小无损 JSON 投影作为 `finish.replayState` 发出。重建历史时验证该状态。只有历史提供方路由和目标提供方路由当前由完全相同的适配器实例拥有时，`LlmService` 才会传递该状态；由适配器决定同模型、跨模型或跨提供方恢复是否合法。状态缺失时，切勿仅根据提供方/模型名称推断原生回放。

-提供方特有的请求旋钮（thinking 模式、effort 级别）放在**适配器**的 Config 中，而非 `GenerateOptions` 中——核心词汇保持提供方无关。
+提供方特有的 thinking 模式开关仍放在适配器的 Config 中。确切模型元数据使用一处提供方无关的能力 seam：实现 `resolveModel()`，返回提供方/模型身份以及可选的 `context` 和 `reasoning` 字段；仅当存在配置指定的默认值时才声明 `defaultEffort`；响应传给解析器的可选 `AbortSignal`。推理强度是由适配器映射到提供方请求的有序不透明 ID。请保留适配器给出的权威可选列表，包括适配器在支持时定义的 `off`；不得暴露最终协议值的具体拼写，也不得自动调整不支持的值。ID 无需与其协议表示相同。

 ## 经验证有效的结构

@@ -616,7 +616,7 @@ Waterfall around every streaming model call (retry, replay, routing). Bound to t

 Types: [GenerateOptions](../core-data-structures/core.md) · [LlmService](../core-data-structures/llm-streaming.md) · [StreamChunk](../core-data-structures/llm-streaming.md)

-Source: [`packages/llm/llm/src/index.ts:52`](../../packages/llm/llm/src/index.ts)
+Source: [`packages/llm/llm/src/index.ts:53`](../../packages/llm/llm/src/index.ts)

 ## `session/*`

@@ -730,33 +730,57 @@ listProviders(): LlmProviderInfo[]
 async listModels(provider: string): Promise<LlmModelInfo[]>

 /**
- * Resolve context capacity from the adapter that owns one exact route.
- * This query is independent of the advisory model catalog: an unlisted model
- * may return metadata, while `undefined` never rejects later routing.
+ * Resolve and validate all metadata from the adapter that owns one exact
+ * route. The result is detached from adapter-owned objects; catalog
+ * membership remains advisory and does not control request routing.
 * @param provider - registered provider route to inspect.
 * @param model - exact model id passed to the adapter.
- * @returns detached context metadata, or `undefined` when the adapter has none.
+ * @param signal - optional cancellation for adapter-owned asynchronous lookup.
+ * @returns exact model identity plus available context and reasoning metadata.
 */
-async resolveModelContext( provider: string, model: string, ): Promise<LlmModelContext | undefined>
+async resolveModelInfo( provider: string, model: string, signal?: AbortSignal, ): Promise<LlmResolvedModelInfo>
+
+/**
+ * Validate a conversation call config against its exact model capability and
+ * materialize an adapter-configured default. Unsupported explicit efforts
+ * reject before provider I/O; no clamping or aliasing is performed. This
+ * standalone query does not bind a later dispatch; use {@link prepareCall}
+ * when logging and streaming must share one adapter registration.
+ * @param config - provider/model route and optional request controls.
+ * @param signal - optional cancellation for adapter-owned capability lookup.
+ * @returns a detached config only when a default must be materialized.
+ */
+async resolveCallConfig(config: LlmCallConfig, signal?: AbortSignal): Promise<LlmCallConfig>
+
+/**
+ * Resolve one call under its current adapter registration. The returned
+ * one-shot handle keeps that registration across header logging and dispatch,
+ * so HMR cannot combine one adapter's capability result with another adapter.
+ * @param config - provider/model route and optional request controls.
+ * @param signal - optional cancellation for adapter-owned capability lookup.
+ * @returns a prepared config and its registration-bound stream entry point.
+ */
+async prepareCall(config: LlmCallConfig, signal?: AbortSignal): Promise<PreparedLlmCall>

 /**
 * Stream one model call as raw chunks (token-level deltas). Throws
 * `LlmError` with code `NO_ADAPTER` if no adapter is registered for
 * `options.provider`. Replay state is retained only when the same adapter
 * instance owns its historical provider and the target provider. Final
- * adapter selection, dispatch, and iteration failures retain their original
- * Error identity and are tagged in a call-local scope for narrow agent-loop
- * request recovery; middleware and nested-call failures remain untagged for
- * the outer call.
+ * adapter selection remains fixed through asynchronous exact-model resolution
+ * and dispatch. Selection, dispatch, and iteration failures retain their
+ * original Error identity and are tagged in a call-local scope for narrow
+ * agent-loop request recovery; middleware and nested-call failures remain
+ * untagged for the outer call.
 * @param options - the full request; `options.provider` selects the adapter.
 * @returns the chunk stream, possibly wrapped by `llm/stream` listeners.
 */
 stream(options: GenerateOptions): AsyncIterable<StreamChunk>
 ```

-Types: [GenerateOptions](../core-data-structures/core.md) · [LlmAdapter](../core-data-structures/llm-streaming.md) · [LlmModelContext](../core-data-structures/core.md) · [LlmModelInfo](../core-data-structures/core.md) · [LlmProviderInfo](../core-data-structures/core.md) · [StreamChunk](../core-data-structures/llm-streaming.md)
+Types: [GenerateOptions](../core-data-structures/core.md) · [LlmAdapter](../core-data-structures/llm-streaming.md) · [LlmCallConfig](../core-data-structures/core.md) · [LlmModelInfo](../core-data-structures/core.md) · [LlmProviderInfo](../core-data-structures/core.md) · [LlmResolvedModelInfo](../core-data-structures/core.md) · [PreparedLlmCall](../core-data-structures/llm-streaming.md) · [StreamChunk](../core-data-structures/llm-streaming.md)

-Source: [`packages/llm/llm/src/index.ts:159`](../../packages/llm/llm/src/index.ts)
+Source: [`packages/llm/llm/src/index.ts:177`](../../packages/llm/llm/src/index.ts)

 ## `ctx.permission` — `PermissionService`

@@ -1342,7 +1366,7 @@ fork(source: SessionForkSource, boundary?: number, childSessionId?: SessionId):

 Types: [CreateSessionOptions](../core-data-structures/persistence.md) · [OutOfBandSessionEventType](../core-data-structures/session.md) · [Session](../core-data-structures/session.md) · [SessionEvent](../core-data-structures/core.md) · [SessionEventMap](../core-data-structures/session.md) · [SessionId](../core-data-structures/core.md) · [TurnTrigger](../core-data-structures/session.md)

-Source: [`packages/core/session/src/index.ts:606`](../../packages/core/session/src/index.ts)
+Source: [`packages/core/session/src/index.ts:611`](../../packages/core/session/src/index.ts)

 ## `ctx.sessionTitle` — `SessionTitleService`

@@ -1878,7 +1902,7 @@ The concrete provider retains pi-tui, focus, and terminal lifecycle state. Plugi
 abstract openOverlay(request: TuiOverlayRequest): TuiOverlaySession
 ```

-Source: [`packages/ui/tui/src/index.ts:150`](../../packages/ui/tui/src/index.ts)
+Source: [`packages/ui/tui/src/index.ts:153`](../../packages/ui/tui/src/index.ts)

 ## `ctx.userInteraction` — `UserInteractionService`

@@ -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
-core.md: 781267cccdb5bbda33e5be6a9e807fdbe47dbc83
-core.zh.md: d0f67983b98b0cf679a8e599a5f8ab3c64490dd0
+core.md: 86aea325f5401d728a9b4aa147d78db9163a32c1
+core.zh.md: d09fb48419c14959eefb5a1df1c593b36c188cf5
@@ -199,7 +199,7 @@ interface LlmModelInfo {
 }
 ```

-Correctness-sensitive model capacity is queried separately from the advisory catalog and is owned by the adapter serving the exact route.
+Correctness-sensitive metadata is resolved separately from the advisory catalog and is owned by the adapter serving the exact route. Context capacity and reasoning choices share one exact-model result so consumers do not repeat authoritative model resolution.

 ```ts type-equiv
 /** Provider-owned context capacity for one exact provider/model route. */
@@ -209,12 +209,56 @@ interface LlmModelContext {
 }
 ```

+Reasoning effort is another exact-route capability. The core brands identifiers but does not enumerate their values; each adapter owns the ordered set, display names, and optional deployment default.
+
+```ts type-equiv
+/** Adapter-owned identifier for one model's selectable reasoning effort. */
+type ReasoningEffortId = Branded<'ReasoningEffortId'>
+```
+
+```ts type-equiv
+/** Display metadata for one adapter-owned reasoning effort. */
+interface LlmReasoningEffortInfo {
+  /** Opaque stable value accepted by {@link GenerateOptions.reasoningEffort}. */
+  id: ReasoningEffortId
+  /** Human-readable effort name for selectors and diagnostics. */
+  name: string
+  /** Optional user-facing distinction from otherwise similar efforts. */
+  description?: string
+}
+```
+
+```ts type-equiv
+/** Selectable reasoning efforts for one exact provider/model route. */
+interface LlmModelReasoningInfo {
+  /** Supported efforts in adapter-preferred display order. */
+  efforts: readonly LlmReasoningEffortInfo[]
+  /**
+   * Adapter-configured default materialized into requests when callers omit
+   * an effort. Absence preserves the provider's own default.
+   */
+  defaultEffort?: ReasoningEffortId
+}
+```
+
+```ts type-equiv
+/** Exact-route model metadata resolved by its owning adapter. */
+interface LlmResolvedModelInfo extends LlmModelInfo {
+  /** Provider-owned context capacity when known. */
+  context?: LlmModelContext
+  /** Adapter-owned selectable reasoning levels when exposed. */
+  reasoning?: LlmModelReasoningInfo
+}
+```
+
 ```ts type-equiv
 /** A single model request, fully assembled. */
 interface GenerateOptions {
  /** Registered provider route selecting the adapter instance. */
  provider: string
  model: string
+  /** Adapter-owned reasoning effort selected for this exact model. */
+  reasoningEffort?: ReasoningEffortId
  /**
   * Ordered conversation messages, exactly as the provider sees them (after
   * the `system` slot). A loop-built request assembles them as
@@ -291,21 +335,23 @@ The model-facing `ToolSchema` is the wire shape; the registered `ToolDefinition`

 The loop builds each request from logged state. `EpochHeader` records call config, rendered prompt, authoritative returned tool order (configured by `toolOrder`, or lexicographic when unset), and session prefix through full `request/header` snapshots. Together with derived history, this makes the request reconstructable from the session log. See [session.md](session.md#the-request-header-event-requestheader) and the [reconstructability Agent Note](../../.agents/notes/implemented/architecture/2026-07-05-reconstructable-requests.md).

-`agent/request` receives a frozen call-config seed and may return a replacement to switch provider, model, or sampling. `agent/session-prefix` composes request-only prefix messages once per loop instance, and the header records the exact result used. Requests reaching `llm/stream` are deep-frozen, so mutation throws, and carry a process-local loop identity so observers do not confuse separately logged frozen auxiliary calls with conversation requests.
+`agent/request` receives a frozen call-config seed and may return a replacement to switch provider, model, reasoning effort, or sampling. After the waterfall, the loop prepares the exact model capability under the turn signal, rejects unsupported explicit effort ids without clamping, materializes an adapter-configured default, and logs the effective value. The prepared call keeps one adapter registration through dispatch. `agent/session-prefix` composes request-only prefix messages once per loop instance, and the header records the exact result used. Requests reaching `llm/stream` are deep-frozen, so mutation throws, and carry a process-local loop identity so observers do not confuse separately logged frozen auxiliary calls with conversation requests.

 On the wire, a loop-built request reads in this order: the `system` slot (the rendered prompt assembly) → `messagePrefix` (the frozen session prefix) → the derived history — the boundary snapshot, whose tail is the newest `user/message` on a turn's first step and the previous step's tool results on later steps. The prefix never enters the derived history; its durable record is the header events, and the dev invariant recomputes exactly this equation against every loop-built request.

-FIXME(call-config-shape): revisit the exact definition of this type — which fields are genuinely epoch-level for cache purposes (`model` certainly; the sampling scalars sit here out of caution), and where provider-specific extras (reasoning options, extra body params) belong when an adapter needs them.
+FIXME(call-config-shape): revisit which remaining fields are genuinely epoch-level for cache purposes (`model` and the model-owned reasoning effort are explicit; the sampling scalars sit here out of caution).

 ```ts type-equiv
 /**
- * Provider + model + sampling scalars of one conversation's requests. Every field maps
- * 1:1 onto the same-named `GenerateOptions` field; the loop builds requests
- * from the logged header rather than accepting these per call.
+ * Provider, model, reasoning effort, and sampling scalars of one conversation's
+ * requests. Every field maps 1:1 onto the same-named `GenerateOptions` field;
+ * the loop builds requests from the logged header rather than accepting these
+ * per call.
 */
 interface LlmCallConfig {
  provider: string
  model: string
+  reasoningEffort?: ReasoningEffortId
  temperature?: number
  maxTokens?: number
  stop?: string[]
@@ -205,7 +205,7 @@ interface LlmModelInfo {
 }
 ```

-对正确性敏感的模型容量与参考目录分开查询，并归服务该确切路由的适配器所有。
+对正确性敏感的元数据与参考目录分开解析，并归服务该确切路由的适配器所有。上下文容量和推理选项共用同一个确切模型结果，消费方因而无需重复执行权威模型解析。

 ```ts type-equiv
 /** Provider-owned context capacity for one exact provider/model route. */
@@ -215,12 +215,56 @@ interface LlmModelContext {
 }
 ```

+推理强度是另一项针对确切路由的能力。核心为标识符添加品牌类型，但不枚举其值；有序集合、展示名称和可选的部署默认值均由各适配器持有。
+
+```ts type-equiv
+/** Adapter-owned identifier for one model's selectable reasoning effort. */
+type ReasoningEffortId = Branded<'ReasoningEffortId'>
+```
+
+```ts type-equiv
+/** Display metadata for one adapter-owned reasoning effort. */
+interface LlmReasoningEffortInfo {
+  /** Opaque stable value accepted by {@link GenerateOptions.reasoningEffort}. */
+  id: ReasoningEffortId
+  /** Human-readable effort name for selectors and diagnostics. */
+  name: string
+  /** Optional user-facing distinction from otherwise similar efforts. */
+  description?: string
+}
+```
+
+```ts type-equiv
+/** Selectable reasoning efforts for one exact provider/model route. */
+interface LlmModelReasoningInfo {
+  /** Supported efforts in adapter-preferred display order. */
+  efforts: readonly LlmReasoningEffortInfo[]
+  /**
+   * Adapter-configured default materialized into requests when callers omit
+   * an effort. Absence preserves the provider's own default.
+   */
+  defaultEffort?: ReasoningEffortId
+}
+```
+
+```ts type-equiv
+/** Exact-route model metadata resolved by its owning adapter. */
+interface LlmResolvedModelInfo extends LlmModelInfo {
+  /** Provider-owned context capacity when known. */
+  context?: LlmModelContext
+  /** Adapter-owned selectable reasoning levels when exposed. */
+  reasoning?: LlmModelReasoningInfo
+}
+```
+
 ```ts type-equiv
 /** A single model request, fully assembled. */
 interface GenerateOptions {
  /** Registered provider route selecting the adapter instance. */
  provider: string
  model: string
+  /** Adapter-owned reasoning effort selected for this exact model. */
+  reasoningEffort?: ReasoningEffortId
  /**
   * Ordered conversation messages, exactly as the provider sees them (after
   * the `system` slot). A loop-built request assembles them as
@@ -297,21 +341,23 @@ interface ToolSchema {

 循环从已记录状态构建每个请求。`EpochHeader` 通过完整的 `request/header` 快照记录调用配置、渲染后的提示词、权威返回工具顺序（由 `toolOrder` 配置；未配置时按字典序）以及会话前缀。结合派生历史，请求便可由会话日志重建。见 [session.md](session.md#the-request-header-event-requestheader) 与[可重建性 Agent Note（agent 决策记录）](../../.agents/notes/implemented/architecture/2026-07-05-reconstructable-requests.md)。

-`agent/request` 接收冻结的调用配置种子，并可返回替代值以切换提供方、模型或采样参数。`agent/session-prefix` 为每个循环实例组合一次仅用于请求的 prefix 消息，header 记录实际使用的确切结果。到达 `llm/stream` 的请求会被深度冻结，因此变更会抛异常；请求还携带进程本地循环标识，使观察者不会把单独记录的冻结辅助调用误认成对话请求。
+`agent/request` 接收冻结的调用配置种子，并可返回替代值以切换提供方、模型、推理强度或采样参数。waterfall 结束后，循环会在轮次信号控制下完成确切模型的能力准备，拒绝显式指定但不受支持的推理强度 ID（不自动调整），填入适配器配置的默认值，并记录最终生效值。准备完成的调用直至分派完成始终持有同一项适配器注册。`agent/session-prefix` 为每个循环实例组合一次仅用于请求的 prefix 消息，header 记录实际使用的确切结果。到达 `llm/stream` 的请求会被深度冻结，因此变更会抛异常；请求还携带进程本地循环标识，使观察者不会把单独记录的冻结辅助调用误认成对话请求。

 在协议格式上，循环构建的请求按此顺序读取：`system` 槽位（渲染后的提示词组装）→ `messagePrefix`（冻结的会话前缀）→ 派生历史——边界快照，其尾部在轮次首步是最新的 `user/message`，在后续步骤是上一步的工具结果。前缀从不进入派生历史；它的持久记录是 header 事件，开发不变式针对每个循环构建的请求精确重算此等式。

-FIXME(call-config-shape)：重新审视此类型的精确定义——出于缓存目的，哪些字段确实属于 epoch 层级（`model` 肯定属于；采样标量目前出于谨慎放在这里），以及适配器需要时，提供方特有的额外项（推理选项、额外 body 参数）应归属何处。
+FIXME(call-config-shape)：重新审视其余哪些字段出于缓存目的确实属于 epoch 层级（`model` 和模型持有的推理强度已明确属于；采样标量目前出于谨慎保留在此）。

 ```ts type-equiv
 /**
- * Provider + model + sampling scalars of one conversation's requests. Every field maps
- * 1:1 onto the same-named `GenerateOptions` field; the loop builds requests
- * from the logged header rather than accepting these per call.
+ * Provider, model, reasoning effort, and sampling scalars of one conversation's
+ * requests. Every field maps 1:1 onto the same-named `GenerateOptions` field;
+ * the loop builds requests from the logged header rather than accepting these
+ * per call.
 */
 interface LlmCallConfig {
  provider: string
  model: string
+  reasoningEffort?: ReasoningEffortId
  temperature?: number
  maxTokens?: number
  stop?: string[]
@@ -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
-llm-streaming.md: fb97e74a9ec01e62ba940295112fb157cc73bd1d
-llm-streaming.zh.md: fda59a64aeef1c037372d69dbc15ee0de48de222
+#   pnpm run verify-translation-pairing --write docs/core-data-structures/llm-streaming.md
+llm-streaming.md: 9151ba569af6b640144c1850d616b6f762b68a8b
+llm-streaming.zh.md: ae4a6843b9bf4f09169b17271c7edea76f8f2051
@@ -67,7 +67,7 @@ Every adapter MUST obey these, and every consumer may rely on them:
 - **Every provider HTTP request carries the app-attribution header.** Adapters send `attributionHeaders()` (below) - the `User-Agent` baseline - and prove it with a wire-level test (mock server asserting the received header, or the library's header hook for a library-backed adapter).
 - **Replay state is adapter-owned.** A successful `finish` may carry lossless-JSON state needed to reconstruct a native provider response. The loop stores it with the assembled assistant message unless an `agent/step-result` listener rewrote the content. On a later request, `LlmService` passes the state only when the historical provider and target provider are currently registered to the exact same adapter instance. That adapter validates the state and owns any cross-model or cross-provider conversion; other adapters receive the provider-neutral content and provenance without the private state.

-This contract is pinned down by two deliberately independent implementations: `dsh-llm-deepseek` (hand-rolled fetch/SSE) and `dsh-llm-pi-ai` (a generic multi-provider adapter through `@earendil-works/pi-ai`). The library-backed adapter exercises the finish-chunk error path, while transport-boundary tests prove each idle watchdog stops its actual request.
+This contract is pinned down by two deliberately independent implementations: `dsh-llm-deepseek` (direct fetch, SSE framing via `eventsource-parser`) and `dsh-llm-pi-ai` (a generic multi-provider adapter through `@earendil-works/pi-ai`). The library-backed adapter exercises the finish-chunk error path, while transport-boundary tests prove each idle watchdog stops its actual request.

 ## `AppIdentity` — app attribution

@@ -157,14 +157,30 @@ declare class BlockAssembler {

 ## The seam

-`LlmAdapter` is the provider seam: subclass, implement `stream()`, and register one adapter instance with `ctx.llm.registerAdapter(providers, adapter)`. `GenerateOptions.provider` selects the registered adapter; `GenerateOptions.model` is passed to that adapter and need not be registered at lifecycle start. Duplicate provider routes fail atomically. Optional `providerInfo()` and asynchronous `listModels()` methods feed `LlmService.listProviders()` / `listModels()` with detached selector metadata. That catalog is advisory rather than a request whitelist: the adapter remains authoritative and may accept unlisted model ids. The separate `resolveModelContext()` query exposes correctness-sensitive capacity for an exact route without making catalog membership authoritative; absence means unknown metadata, not invalid routing. Adapter lookup happens at the terminal continuation of the `llm/stream` waterfall, so a listener may short-circuit the call or route a mutable one-shot request before lookup. The `block-start` / `block-end` `index` correlation and the assembler together mean an adapter only has to emit well-formed chunks — block reassembly is not each adapter's problem. The consumer surface (`ctx.llm.stream()`) and the `llm/stream` waterfall are described in [architecture.md § Content blocks and streaming](../architecture.md#content-blocks-and-streaming-dsh-llm).
+`LlmAdapter` is the provider seam: subclass, implement `stream()`, and register one adapter instance with `ctx.llm.registerAdapter(providers, adapter)`. `GenerateOptions.provider` selects the registered adapter; `GenerateOptions.model` is passed to that adapter and need not be registered at lifecycle start. Duplicate provider routes fail atomically. Optional `providerInfo()` and asynchronous `listModels()` methods feed `LlmService.listProviders()` / `listModels()` with detached selector metadata. That catalog is advisory rather than a request whitelist: the adapter remains authoritative and may accept unlisted model ids. One asynchronous `resolveModel()` query returns exact model identity plus optional correctness-sensitive context capacity and ordered model-owned reasoning ids with an optional deployment default; absent fields mean unavailable metadata or capability, not invalid catalog membership. The resolver receives optional cancellation and must settle promptly after abort. `LlmService.resolveModelInfo()` validates and detaches the aggregate. The service validates and materializes reasoning through `resolveCallConfig()` at the final adapter boundary, so direct calls cannot bypass unsupported-effort rejection; direct dispatch captures one registration before awaiting that resolution. The agent loop instead uses `prepareCall()` to keep the same registration across model resolution, durable header logging, and dispatch. Adapter lookup happens at the terminal continuation of the `llm/stream` waterfall, so a listener may short-circuit the call or route a mutable one-shot request before lookup. The `block-start` / `block-end` `index` correlation and the assembler together mean an adapter only has to emit well-formed chunks — block reassembly is not each adapter's problem. The consumer surface (`ctx.llm.stream()`) and the `llm/stream` waterfall are described in [architecture.md § Content blocks and streaming](../architecture.md#content-blocks-and-streaming-dsh-llm).
+
+```ts type-equiv
+/** One model call whose config and adapter registration were resolved together. */
+interface PreparedLlmCall {
+  /** Detached, deep-frozen config with any adapter-owned default materialized. */
+  readonly config: LlmCallConfig
+  /**
+   * Dispatch this call once through the registration captured during
+   * preparation. The request's call-config fields must match {@link config};
+   * reuse or mismatch fails with `INVALID_PREPARED_CALL`.
+   * @param options - fully assembled request carrying the prepared config.
+   * @returns the chunk stream, including the `llm/stream` waterfall.
+   */
+  stream(options: GenerateOptions): AsyncIterable<StreamChunk>
+}
+```

 ```ts public-api
 /**
 * Provider-wire adapter for the harness message and stream vocabulary. Register implementations
 * with `ctx.llm.registerAdapter(providers, adapter)`. Every provider HTTP request must include
- * `attributionHeaders()`; prove that at the wire or library header-hook boundary. The hand-rolled
- * DeepSeek and pi-ai adapters intentionally exercise this contract through different internals.
+ * `attributionHeaders()`; prove that at the wire or library header-hook boundary. The direct-fetch
+ * DeepSeek and library-backed pi-ai adapters intentionally exercise this contract through different internals.
 */
 declare abstract class LlmAdapter {
  /**
@@ -182,16 +198,19 @@ declare abstract class LlmAdapter {
   */
  listModels(_provider: string): Promise<readonly LlmModelInfo[]>;
  /**
-   * Resolve context capacity for one model accepted by this adapter. Absence
-   * means the adapter does not know the capacity, not that routing is invalid.
-   * @param _provider - one provider route owned by this adapter.
-   * @param _model - exact model id passed to {@link GenerateOptions.model}.
-   * @returns provider-owned context metadata, or `undefined` when unavailable.
+   * Resolve all metadata available for one exact model. This query is
+   * independent of the advisory catalog and does not validate request routing.
+   * @param provider - one provider route owned by this adapter.
+   * @param model - exact model id passed to {@link GenerateOptions.model}.
+   * @param _signal - cancellation for this exact-model lookup; asynchronous
+   *   implementations must settle promptly after it aborts.
+   * @returns provider/model identity plus any context and reasoning metadata.
   */
-  resolveModelContext(
-    _provider: string,
-    _model: string,
-  ): Promise<LlmModelContext | undefined>;
+  resolveModel(
+    provider: string,
+    model: string,
+    _signal?: AbortSignal,
+  ): Promise<LlmResolvedModelInfo>;
  /**
   * Stream one model call as raw chunks. The only required method.
   * @param options - the fully-assembled request; implementations must honor `options.signal`.
@@ -67,7 +67,7 @@ interface LlmFailure {
 - **每个提供方 HTTP 请求都携带应用归属头。** 适配器发送 `attributionHeaders()`（见下文）作为 `User-Agent` 基线，并通过协议级测试加以证明（mock 服务器断言收到的 header，或对基于库的适配器使用库的 header 钩子）。
 - **回放状态归适配器所有。** 成功的 `finish` 可以携带重建提供方原生响应所需的无损 JSON 状态。除非 `agent/step-result` listener 改写了内容，否则循环会将其与组装后的 assistant 消息一起存储。后续请求中，仅当历史提供方与目标提供方当前注册到完全相同的适配器实例时，`LlmService` 才会传递该状态。该适配器负责校验状态并拥有所有跨模型或跨提供方转换；其他适配器只会收到提供方无关的内容与 provenance，不会收到私有状态。

-该契约由两个有意保持独立的实现锁定：`dsh-llm-deepseek`（手写 fetch/SSE（Server-Sent Events））和 `dsh-llm-pi-ai`（通过 `@earendil-works/pi-ai` 实现的通用多提供方适配器）。基于库的适配器覆盖 finish 分片错误路径，而传输边界测试证明每个空闲 watchdog 都会停止其实际请求。
+该契约由两个有意保持独立的实现锁定：`dsh-llm-deepseek`（直接 fetch，SSE（Server-Sent Events）分帧经由 `eventsource-parser`）和 `dsh-llm-pi-ai`（通过 `@earendil-works/pi-ai` 实现的通用多提供方适配器）。基于库的适配器覆盖 finish 分片错误路径，而传输边界测试证明每个空闲 watchdog 都会停止其实际请求。

 ## `AppIdentity`：应用归属

@@ -157,14 +157,30 @@ declare class BlockAssembler {

 ## seam

-`LlmAdapter` 是提供方 seam：创建子类、实现 `stream()`，再用 `ctx.llm.registerAdapter(providers, adapter)` 注册一个适配器实例。`GenerateOptions.provider` 选择已注册适配器；`GenerateOptions.model` 会传给该适配器，无需在生命周期启动时注册。重复提供方路由会原子失败。可选的 `providerInfo()` 与异步 `listModels()` 方法为 `LlmService.listProviders()` / `listModels()` 提供分离的 selector 元数据。该目录仅供参考，不是请求白名单：适配器仍是权威，并可接受未列出的模型 id。单独的 `resolveModelContext()` 查询会暴露确切路由上对正确性敏感的容量信息，但不会让目录成员关系具有权威性；缺失表示元数据未知，而不是路由无效。适配器查找发生在 `llm/stream` waterfall（瀑布式事件）的终端 continuation，因此 listener 可以在查找前短路调用，或路由一个可变的一次性请求。`block-start` / `block-end` 的 `index` 关联与 assembler 共同意味着适配器只需 emit 格式正确的分片——块重组不是每个适配器各自的问题。消费方 surface（`ctx.llm.stream()`）与 `llm/stream` waterfall 见 [architecture.md § 内容块与流式传输](../architecture.md#content-blocks-and-streaming-dsh-llm)。
+`LlmAdapter` 是提供方 seam：创建子类、实现 `stream()`，再用 `ctx.llm.registerAdapter(providers, adapter)` 注册一个适配器实例。`GenerateOptions.provider` 选择已注册适配器；`GenerateOptions.model` 会传给该适配器，无需在生命周期启动时注册。重复提供方路由会原子失败。可选的 `providerInfo()` 与异步 `listModels()` 方法为 `LlmService.listProviders()` / `listModels()` 提供分离的 selector 元数据。该目录仅供参考，不是请求白名单：适配器仍是权威，并可接受未列出的模型 id。单次异步 `resolveModel()` 查询返回确切模型身份，以及可选的对正确性敏感的上下文容量、由模型持有的有序推理强度 ID 和部署默认值；字段缺失表示元数据或能力不可用，而不表示目录成员关系无效。解析器会接收可选的取消信号，并且必须在信号中止后迅速完成结算。`LlmService.resolveModelInfo()` 会校验聚合结果并返回分离值。服务通过最终适配器边界的 `resolveCallConfig()` 校验推理强度并填入默认值，因此直接调用也无法绕过对不支持推理强度的拒绝；直接分派会在等待解析前捕获一项适配器注册。agent loop 则使用 `prepareCall()`，使模型解析、请求头持久记录和分派全程使用同一项注册。适配器查找发生在 `llm/stream` waterfall（瀑布式事件）的终端 continuation，因此 listener 可以在查找前短路调用，或路由一个可变的一次性请求。`block-start` / `block-end` 的 `index` 关联与 assembler 共同意味着适配器只需 emit 格式正确的分片——块重组不是每个适配器各自的问题。消费方 surface（`ctx.llm.stream()`）与 `llm/stream` waterfall 见 [architecture.md § 内容块与流式传输](../architecture.md#content-blocks-and-streaming-dsh-llm)。
+
+```ts type-equiv
+/** One model call whose config and adapter registration were resolved together. */
+interface PreparedLlmCall {
+  /** Detached, deep-frozen config with any adapter-owned default materialized. */
+  readonly config: LlmCallConfig
+  /**
+   * Dispatch this call once through the registration captured during
+   * preparation. The request's call-config fields must match {@link config};
+   * reuse or mismatch fails with `INVALID_PREPARED_CALL`.
+   * @param options - fully assembled request carrying the prepared config.
+   * @returns the chunk stream, including the `llm/stream` waterfall.
+   */
+  stream(options: GenerateOptions): AsyncIterable<StreamChunk>
+}
+```

 ```ts public-api
 /**
 * Provider-wire adapter for the harness message and stream vocabulary. Register implementations
 * with `ctx.llm.registerAdapter(providers, adapter)`. Every provider HTTP request must include
- * `attributionHeaders()`; prove that at the wire or library header-hook boundary. The hand-rolled
- * DeepSeek and pi-ai adapters intentionally exercise this contract through different internals.
+ * `attributionHeaders()`; prove that at the wire or library header-hook boundary. The direct-fetch
+ * DeepSeek and library-backed pi-ai adapters intentionally exercise this contract through different internals.
 */
 declare abstract class LlmAdapter {
  /**
@@ -182,16 +198,19 @@ declare abstract class LlmAdapter {
   */
  listModels(_provider: string): Promise<readonly LlmModelInfo[]>;
  /**
-   * Resolve context capacity for one model accepted by this adapter. Absence
-   * means the adapter does not know the capacity, not that routing is invalid.
-   * @param _provider - one provider route owned by this adapter.
-   * @param _model - exact model id passed to {@link GenerateOptions.model}.
-   * @returns provider-owned context metadata, or `undefined` when unavailable.
+   * Resolve all metadata available for one exact model. This query is
+   * independent of the advisory catalog and does not validate request routing.
+   * @param provider - one provider route owned by this adapter.
+   * @param model - exact model id passed to {@link GenerateOptions.model}.
+   * @param _signal - cancellation for this exact-model lookup; asynchronous
+   *   implementations must settle promptly after it aborts.
+   * @returns provider/model identity plus any context and reasoning metadata.
   */
-  resolveModelContext(
-    _provider: string,
-    _model: string,
-  ): Promise<LlmModelContext | undefined>;
+  resolveModel(
+    provider: string,
+    model: string,
+    _signal?: AbortSignal,
+  ): Promise<LlmResolvedModelInfo>;
  /**
   * Stream one model call as raw chunks. The only required method.
   * @param options - the fully-assembled request; implementations must honor `options.signal`.
@@ -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
-session.md: 2cbbac8042d04522fea0b1ed7a66c503e4b63f4e
-session.zh.md: e932c8f99f684f1b8985b006ab4ddb145db966bd
+#   pnpm run verify-translation-pairing --write docs/core-data-structures/session.md
+session.md: 59bcf027c3ec728c5583478c7d17bd6eef2dc2fa
+session.zh.md: 01b6f080b5fcf7b83ca46b99a8d80113a6ebe8bf
@@ -175,7 +175,7 @@ The request envelope — the `EpochHeader` (call config + rendered system prompt
 * canonical empty optional fields are absent.
 */
 interface EpochHeader {
-  /** The conversation's call configuration (provider, model, and sampling scalars). */
+  /** The conversation's call configuration (provider, model, reasoning effort, and sampling scalars). */
  config: LlmCallConfig
  /** Rendered system prompt text; absent for a system-less request. */
  system?: string
@@ -177,7 +177,7 @@ interface TodoItem {
 * canonical empty optional fields are absent.
 */
 interface EpochHeader {
-  /** The conversation's call configuration (provider, model, and sampling scalars). */
+  /** The conversation's call configuration (provider, model, reasoning effort, and sampling scalars). */
  config: LlmCallConfig
  /** Rendered system prompt text; absent for a system-less request. */
  system?: string
@@ -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
-development.md: c46d84740e6f0a1f67158f39f9ea421cb57165d4
-development.zh.md: 9e13ab258e1db5406f84ece61959a995110578ae
+#   pnpm run verify-translation-pairing --write docs/development.md
+development.md: fd7f39ae7b5aac2d44572979ca8c8f1d2df0de6f
+development.zh.md: 7dd6209bad75d605e0056d2465a35b08aa091780
@@ -112,6 +112,7 @@ pnpm run verify-md-wrap  # fail on hard-wrapped prose paragraphs in docs/README
 pnpm run verify-mermaid  # fail if a ```mermaid diagram has invalid Mermaid syntax
 pnpm run verify-type-equiv  # fail if a ```ts type-equiv doc block drifts from its source type
 pnpm run verify-doc-budgets  # fail if a budgeted standing doc exceeds its word ceiling
+pnpm run gen-translation-brief   # print the minimal-update briefing for out-of-sync translation pairs (--apply splices code-only edits)
 pnpm run doc-sync       # all Markdown/doc gates, scheduled concurrently; the doc-sync leaf list in scripts/run-gates.ts is the full list
 pnpm run gen-module-graph     # regenerate docs/module-graph.md from package peerDeps
 pnpm run verify-module-graph  # fail if docs/module-graph.md is stale
@@ -112,6 +112,7 @@ pnpm run verify-md-wrap  # fail on hard-wrapped prose paragraphs in docs/README
 pnpm run verify-mermaid  # fail if a ```mermaid diagram has invalid Mermaid syntax
 pnpm run verify-type-equiv  # fail if a ```ts type-equiv doc block drifts from its source type
 pnpm run verify-doc-budgets  # fail if a budgeted standing doc exceeds its word ceiling
+pnpm run gen-translation-brief   # print the minimal-update briefing for out-of-sync translation pairs (--apply splices code-only edits)
 pnpm run doc-sync       # all Markdown/doc gates, scheduled concurrently; the doc-sync leaf list in scripts/run-gates.ts is the full list
 pnpm run gen-module-graph     # regenerate docs/module-graph.md from package peerDeps
 pnpm run verify-module-graph  # fail if docs/module-graph.md is stale
@@ -33,7 +33,7 @@ This matrix shows which packages dispatch each harness-owned event and which pac
 | `fs/observed` | `emit` | [`packages/fs/fs/src/index.ts:71`](../packages/fs/fs/src/index.ts) | [`tool-fs`](../packages/fs/tool-fs) (`emit`) | [`fs-policy`](../packages/fs/fs-policy) |
 | `fs/write-intent` | `waterfall` | [`packages/fs/fs/src/index.ts:54`](../packages/fs/fs/src/index.ts) | [`tool-fs`](../packages/fs/tool-fs) (`waterfall`) | [`fs-policy`](../packages/fs/fs-policy) |
 | `goal/changed` | `emit` | [`packages/goal/goal/src/types.ts:167`](../packages/goal/goal/src/types.ts) | [`goal`](../packages/goal/goal) (`emit`) | [`goal-session`](../packages/goal/goal-session) |
-| `llm/stream` | `waterfall` | [`packages/llm/llm/src/index.ts:52`](../packages/llm/llm/src/index.ts) | [`llm`](../packages/llm/llm) (`waterfall`) | [`agent-loop`](../packages/core/agent-loop), [`llm`](../packages/llm/llm), [`llm-replay`](../packages/support/llm-replay), [`session-checkpoint-policy`](../packages/session-persistence/session-checkpoint-policy), [`session-title`](../packages/session-title/session-title) |
+| `llm/stream` | `waterfall` | [`packages/llm/llm/src/index.ts:53`](../packages/llm/llm/src/index.ts) | [`llm`](../packages/llm/llm) (`waterfall`) | [`agent-loop`](../packages/core/agent-loop), [`llm`](../packages/llm/llm), [`llm-replay`](../packages/support/llm-replay), [`session-checkpoint-policy`](../packages/session-persistence/session-checkpoint-policy), [`session-title`](../packages/session-title/session-title) |
 | `session/created` | `emit` | [`packages/core/session/src/index.ts:79`](../packages/core/session/src/index.ts) | [`session`](../packages/core/session) (`events.dispatch`) | `apiproxy`, [`compact`](../packages/compact/compact), [`goal`](../packages/goal/goal), [`hook-protocol`](../packages/hooks/hook-protocol), [`jsonrpc`](../packages/ui/jsonrpc), [`llm-retry`](../packages/llm/llm-retry), [`session`](../packages/core/session), [`session-persistence`](../packages/session-persistence/session-persistence), [`user-approval`](../packages/ui/user-approval) |
 | `session/disposed` | `emit` | [`packages/core/session/src/index.ts:89`](../packages/core/session/src/index.ts) | [`session`](../packages/core/session) (`events.dispatch`) | [`agent-loop`](../packages/core/agent-loop), `apiproxy`, [`session-persistence`](../packages/session-persistence/session-persistence), [`session-title`](../packages/session-title/session-title) |
 | `session/event` | `emit` | [`packages/core/session/src/index.ts:101`](../packages/core/session/src/index.ts) | [`session`](../packages/core/session) (`events.dispatch`) | [`acp`](../packages/acp/acp), `apiproxy`, [`cli-demo`](../packages/examples/cli-demo), [`compact`](../packages/compact/compact), [`goal`](../packages/goal/goal), [`goal-session`](../packages/goal/goal-session), [`hook-protocol`](../packages/hooks/hook-protocol), [`jsonrpc`](../packages/ui/jsonrpc), [`session`](../packages/core/session), [`session-persistence`](../packages/session-persistence/session-persistence), [`session-title`](../packages/session-title/session-title), [`token-meter`](../packages/llm/token-meter), [`tui`](../packages/ui/tui), [`user-approval`](../packages/ui/user-approval), [`workspace-context`](../packages/context/workspace-context) |
@@ -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
-README.md: daddd35f981879b539ec76f0c21cf232f4688036
-README.zh.md: c36728240690edb7ae35f33eaa7595a5d320860f
+#   pnpm run verify-translation-pairing --write docs/i18n/README.md
+README.md: 9fe01174909f0400b94c70b95759526aec783d2e
+README.zh.md: 2a9a20ae4702e205879f0fbfb6d76d3a761d2446