Merge remote-tracking branch 'github/master' into feat/agent-event-payload

# Conflicts:
#	docs/core-data-structures/core.i18n.yaml

.agents/notes/implemented/architecture/2026-08-03-pi-ai-declared-provider-catalog.i18n.yaml

@@ -0,0 +1,6 @@
+# Bilingual-pair consistency record (docs/i18n/README.md): the git blob hash of each
+# side as of the last confirmed-consistent state. Both languages carry equal authority;
+# after editing either side, bring the other along and re-record with:
+#   pnpm run verify-translation-pairing --write .agents/notes/implemented/architecture/2026-08-03-pi-ai-declared-provider-catalog.md
+2026-08-03-pi-ai-declared-provider-catalog.md: d75b6bdb91d60026636bf320f8c6625590849a41
+2026-08-03-pi-ai-declared-provider-catalog.zh.md: f8dba9900b1a7a3abcb16c70a35cc18f0c44219f

.agents/notes/implemented/architecture/2026-08-03-pi-ai-declared-provider-catalog.md

@@ -0,0 +1,67 @@
+# Agent Note: pi-ai routes are declared providers, not catalog lookups
+
+Status: implemented
+
+English | [中文](2026-08-03-pi-ai-declared-provider-catalog.zh.md)
+
+## Problem
+
+`dsh-llm-pi-ai` treated the pi-ai package's generated catalog as the boundary of what could be configured. A route key had to name an installed provider (`resolveProfiles` rejected anything else), model listing returned `getBuiltinModels(provider)` verbatim, and request-time model resolution looked the id up in that same catalog and overrode only `baseURL`. Three consequences followed, and all three were dead ends rather than gaps: an OpenAI-compatible gateway, a self-hosted server, or a provider newer than the installed catalog could not be configured at all; a model the catalog had not caught up with failed with `UNKNOWN_MODEL` even against a correct endpoint; and a model's context window and output cap were whatever the pinned pi-ai release said, so a deployment could neither correct a stale value nor supply one for a model pi-ai had never described. Upgrading the package was the only way to move any of it.
+
+The adapter also streamed through `streamSimple` from `@earendil-works/pi-ai/compat`, an entry point whose own module documentation declares it a temporary compatibility surface — its catalog reads are `@deprecated`, and it is deleted when pi-ai finishes its `ModelManager` migration. The three configuration limits and the deprecated dependency have the same fix, because pi-ai's supported runtime (`createModels()` / `createProvider()`) is built around a provider being *declared* rather than looked up.
+
+## Decision
+
+A provider route is a **declaration**, and the installed catalog is its default. `resolveProfiles` no longer checks route keys against `getBuiltinProviders()`. Instead each route resolves to a materialized model list plus the pi-ai `Provider` that serves it:
+
+- `catalog.ts` merges the installed catalog under the profile's own entries. A profile's `models` list *replaces* the route's catalog (an absent or empty list serves it unchanged), and each entry defaults its unset fields from the installed model of the same `id`. Only the fields the harness consumes are configurable — `id`, `name`, `contextWindow`, `maxTokens`. Pricing and input modalities are absent from the surface because nothing reads them: `replay.ts` zeroes pi-ai's cost metadata and `context.ts` keeps only text blocks. Reasoning is absent for a different reason: a bare capability flag would make pi-ai advertise effort levels with no `thinkingLevelMap` to spell them, so it rides the installed entry or is absent. Materialization spreads the installed entry and overrides those four fields, rather than enumerating the result: an enumerated rebuild silently drops every `Model` field this package does not model, which is how `headers` went missing from an nvidia route once already.
+- `provider.ts` builds the route's `Provider`. A catalog route that keeps its catalog protocol **reuses** the installed provider with `getModels()` replaced; every other route is built by `createProvider()` over a protocol table whose entries are the same `@earendil-works/pi-ai/api/*.lazy` factories pi-ai's own provider factories use. That table is narrower than pi-ai's full API set on purpose — it holds only protocols a profile can completely describe with a key, an endpoint, and headers, so Bedrock (SigV4 plus a region), Vertex (project, location, ADC), Azure (provider environment plus an api-version), and Codex (OAuth) are absent rather than offered as routes that cannot authenticate. Catalog routes still reach them through their own provider; only an explicit override is refused.
+- `adapter.ts` turns each resolution into an **immutable snapshot** — the profiles plus a `createModels()` collection holding those providers — and every operation captures a whole snapshot before its first `await`.
+- A model's **explicitly configured** `maxTokens` becomes the seam's `defaultMaxTokens`. The value inherited from the installed catalog does not: pi-ai requires `Model.maxTokens` as the model's output *capability*, while `defaultMaxTokens` is a cap the deployment chose to send on requests that name none, and materializing the former as the latter would start capping every request at a number nobody picked.
+
+### Snapshots, not a shared collection
+
+`Models.streamSimple()` resolves its provider lazily, when the returned stream is first consumed — which is after the adapter has awaited the route's credential. A single collection mutated in place would therefore let a request that started under one configuration finish under another, or fail on a provider that no longer exists, even though `llm.prepareCall()` already froze that step's config and captured its adapter registration. A configuration change builds a *new* collection and leaves the one in use alone, so the seam's per-step freeze holds all the way down: switching models mid-reply takes effect on the next step, never inside the one in flight.
+
+### The directory replaces atomically
+
+The configurable-provider directory follows the profiles, so it changes whenever a declared route appears or leaves. Withdrawing the old registration and making a new one cannot express that: a candidate set the registry refuses — a profile keyed `deepseek-official`, which `llm-deepseek` already declares — would leave this plugin's whole directory withdrawn and the Models page empty, silently, because the settings-change callback contains the failure. `registerConfigurableProviders` therefore returns a handle carrying `replace(entries)` with the same validate-the-candidate-set-first atomicity `registerAdapter` has, and the plugin uses it. A refused swap costs a diagnostic; the previous entries keep serving.
+
+Resolution fails loud and names the route and model at fault: a model the catalog does not describe falls back to the route's own `defaultContextWindow`/`defaultMaxTokens`, so a listing that discloses nothing but ids still yields a serviceable route; a route the catalog does not ship needs `api`, `baseURL`, and a non-empty `models` list. Because the built `Provider` is part of the resolution result, a protocol or model error keeps the last good route set serving, exactly as a bad settings snapshot already did.
+
+The configurable-provider directory is now the installed catalog **joined with** every route the current profiles declare, re-registered when that set changes. Without the join a hand-declared route would have no settings address and no configuration surface could show or edit it.
+
+### A capability whose only level does nothing is reported unavailable
+
+pi-ai reports a model with no reasoning metadata as supporting the single level `off`, and the adapter used to pass that straight through. It reaches the seam as a one-item effort list, which every surface renders as a picker holding one selectable control — and that control is a lie: `off` becomes an *omitted* reasoning option at dispatch, byte-for-byte the request that naming no effort already produces. A provider whose own default is to think keeps thinking while the surface shows `off` selected.
+
+`reasoningInfo` therefore omits the seam's `reasoning` field whenever `model.reasoning` is falsy. The condition is the model's own metadata, not where the model came from, so this covers every hand-declared model **and** the 251 installed-catalog models pi-ai marks as non-reasoning. Those previously offered the lone `off`; they now offer nothing, and the surface shows the provider default alone. Models that do carry reasoning metadata are untouched — their level list still crosses the seam unfiltered, `off` included, because there it selects between real alternatives.
+
+### Credentials stay outside pi-ai
+
+pi-ai's `Models` carries its own credential concept — a `CredentialStore` keyed by provider id, with `envApiKeyAuth` resolving `credential.key ?? env(VAR)`. Adopting it would have created a second credential source of truth beside `ctx.credentials` and, worse, reintroduced the ambient fallback the harness deliberately forbids: a named-but-missing `apiKeyEnv` must fail with `MISSING_CREDENTIAL` rather than authenticate with whatever unrelated key the environment holds.
+
+`ModelsImpl.applyAuth` honours `options.apiKey` as the request's key, but only through a provider that declares an api-key method: `resolveProviderAuth` short-circuits to that method when the override is present, and otherwise falls through to the credential store and then to ambient discovery, returning nothing — and so failing the request with `Provider is not configured` — when the provider has no api-key method at all. The harness therefore resolves the route's key through its own seam, as before, and passes the result as the request's `apiKey`; the collection is constructed with no credential store.
+
+A route's auth follows from that. A catalog route keeps the installed provider's own `auth`, which preserves provider-native ambient discovery for a profile naming no credential, and keeps it through an `api` override too: which environment a provider reads is a property of the provider, not of the wire format its models speak. The exception is a catalog provider with no api-key method — `openai-codex` authenticates through OAuth alone — where a profile that names a credential also gets the harness method beside the provider's own, because otherwise its configured key would be refused before any request went out. A keyless profile on such a route adds nothing and keeps the honest refusal: this adapter holds no OAuth store to resolve through. A hand-declared route gets a harness-owned `ApiKeyAuth` that reports configured-but-keyless rather than unconfigured, leaving the requirement to the protocol — which is where it lives: pi-ai's OpenAI-compatible implementation still demands a key or an `Authorization` header, and says so itself.
+
+## Alternatives considered
+
+- **Keep `createProvider()` but skip the `Models` collection**, streaming through `provider.streamSimple(model, ctx, {apiKey})`. Smallest diff and the credential path is untouched, but `createProvider`'s `auth` is a required field that this path never invokes — a required-by-signature implementation with no caller. It also leaves `refreshModels` needing a hand-built `RefreshModelsContext`, and keeps the adapter off the runtime pi-ai actually supports.
+- **Reuse the installed provider for catalog routes and `createProvider()` only for declared ones**, with no shared resolution. Zero risk to catalog behavior, but catalog materialization, endpoint override, and per-model configuration would each exist twice, and a catalog route that repoints its protocol would have to jump paths mid-resolution. The chosen split confines the asymmetry to provider construction, where it is forced by pi-ai not exposing a built provider's API implementations.
+- **Rebuild every route through `createProvider()`**, including catalog ones. Fully symmetric, but a built `Provider` does not expose its `api`, so the protocol table would become the ceiling on which providers work — Bedrock loads its Smithy module through a separate entry point and would silently stop working.
+- **Expose pi-ai's whole `Model` shape** (cost, input modalities, `thinkingLevelMap`, `compat`). Maximum configurability, but no current consumer reads those fields, so a configured price or modality would change nothing while reading as supported.
+
+- **Keep one mutable `Models` collection and re-sync it.** Fewer allocations, and correct for every operation that resolves synchronously. It is exactly wrong for the one that does not: `stream()` awaits a credential between capturing its model and dispatching it.
+- **Simulate an atomic directory swap with dispose-then-register.** No seam change, and it works whenever the new set is valid — which is the case that never needed atomicity.
+- **A runtime dynamic catalog** — `fetchModels` plus `ModelsStore`, refreshed in the background. Rejected for this change: it makes the model list external mutable state needing cache, invalidation, and an offline path, and the product need is a one-shot discovery action whose result the user adopts into `settings.yaml`. That action belongs to the configuration surface and is deferred with it; `settings.yaml` stays the single source of truth for what a route serves.
+
+## Consequences
+
+Configuring a provider no longer depends on a pi-ai release. A gateway, a self-hosted server, or a model newer than the pinned catalog is a `settings.yaml` edit, and a stale context window can be corrected in place. The deprecated `/compat` import is gone, so pi-ai deleting it is no longer a breaking event. `defaultMaxTokens` now flows from configuration when a deployment states one, without inventing a cap from catalog metadata.
+
+What it costs: `settings.yaml` grows for a declared route, because it must state its endpoint, protocol, and model ids. `api` applies to a whole route, so a mixed-protocol catalog route cannot host a model of the other protocol — splitting it across two route keys is the workaround. Nothing queries a provider's `/models`, so a model list is only as current as its last edit. Reported error shape shifts in one case: a route whose auth resolves to nothing now surfaces pi-ai's own diagnostic as an error `finish` chunk before any network call, where the previous adapter sent a keyless request and surfaced the provider's 401.
+
+## Testing
+
+`tests/catalog.spec.ts` covers the contract end to end against local mock servers: a hand-declared route streaming to its own endpoint with its own credential, its appearance in the configurable-provider directory, per-model overrides defaulting from the installed catalog, a model added to a catalog route, protocol repointing with and without an endpoint override, catalog-only metadata surviving an override, the keyless posture and its `Authorization`-header workaround, an OAuth-only catalog route authenticating with the key its profile names while a keyless one stays unconfigured, a repointed route keeping its catalog auth, and every resolution failure that names a route or model. `tests/catalog.spec.ts` also pins the snapshot and directory contracts: an in-flight request whose route set changes during its credential await still reaches the endpoint it resolved against, the next request picks up the new one, a colliding declared route leaves the directory whole, and a declared route's entry appears and leaves with its profile. `packages/llm/llm/tests/topology.spec.ts` covers `replace` — refusing a candidate another registration owns while keeping the current set, accepting a swap over its own entries, allowing an empty set, and failing after disposal. `tests/sdk-options.spec.ts` re-targets the SDK boundary from the removed `/compat` import to the protocol table's lazy api module, which also pins that a setup failure arrives as a terminal error chunk rather than a throw. The twin's [design-verification role](2026-06-13-twin-llm-adapters.md) is unchanged.

.agents/notes/implemented/architecture/2026-08-03-pi-ai-declared-provider-catalog.zh.md

@@ -0,0 +1,67 @@
+# Agent Note: pi-ai 路由是被声明的提供方，而不是 catalog 查表
+
+Status: implemented
+
+[English](2026-08-03-pi-ai-declared-provider-catalog.md) | 中文
+
+## Problem
+
+`dsh-llm-pi-ai` 把 pi-ai 包生成的 catalog 当成了可配置范围的边界。路由键必须点名一个已安装提供方（`resolveProfiles` 拒绝其余一切），模型列举原样返回 `getBuiltinModels(provider)`，请求期的模型解析又在同一份 catalog 里查这个 id、且只覆盖 `baseURL`。由此产生三个后果，而且三个都是死路而非缺口：OpenAI 兼容网关、自建服务，或比已安装 catalog 更新的提供方，根本无法配置；catalog 尚未跟上的模型即便端点正确也会以 `UNKNOWN_MODEL` 失败；模型的上下文窗口与输出上限完全由锁定的 pi-ai 版本决定，部署既无法更正过期值，也无法为 pi-ai 从未描述过的模型补上。要动其中任何一条，只能升级依赖。
+
+适配器还经 `@earendil-works/pi-ai/compat` 的 `streamSimple` 发起流式请求，而该入口自己的模块文档声明它是临时兼容面——其 catalog 读取标了 `@deprecated`，并会在 pi-ai 完成 `ModelManager` 迁移时被删除。这三条配置限制与这个废弃依赖的解法是同一个，因为 pi-ai 受支持的运行时（`createModels()` / `createProvider()`）正是围绕「提供方是被*声明*出来的，而非查出来的」建立的。
+
+## Decision
+
+提供方路由是一份**声明**，已安装 catalog 是它的默认值。`resolveProfiles` 不再拿路由键去核对 `getBuiltinProviders()`，而是把每条路由解析成一份物化模型列表，外加服务它的 pi-ai `Provider`：
+
+- `catalog.ts` 把已安装 catalog 合并到 profile 自身条目之下。profile 的 `models` 列表*替换*该路由的 catalog（列表缺席或为空则原样服务），每个条目从同 `id` 的已安装模型继承自身未设置的字段。只有 harness 会消费的字段可配置——`id`、`name`、`contextWindow`、`maxTokens`。定价与输入模态不出现在配置面，因为没有任何读取方：`replay.ts` 把 pi-ai 的成本元数据清零，`context.ts` 只保留文本块。推理缺席则是另一个理由：一个孤立的能力布尔量会让 pi-ai 公布出没有 `thinkingLevelMap` 可供拼写的档位，因此它沿用已安装条目或直接缺席。物化时以已安装条目铺底、再覆盖那四个字段，而不是逐字段枚举结果：枚举式重建会静默丢弃本包未建模的每一个 `Model` 字段——`headers` 就是这样从某条 nvidia 路由上消失过一次。
+- `provider.ts` 构造路由的 `Provider`。保持 catalog 协议不变的 catalog 路由会**复用**已安装提供方，只替换 `getModels()`；其余路由都由 `createProvider()` 基于一张协议表构造，表中条目正是 pi-ai 自己的提供方工厂所用的 `@earendil-works/pi-ai/api/*.lazy` factory。该表刻意窄于 pi-ai 的完整 API 集合——只保留 profile 能用密钥、端点与标头完整描述的协议，因此 Bedrock（SigV4 加 region）、Vertex（project、location、ADC）、Azure（提供方环境加 api-version）与 Codex（OAuth）不在其中，而不是被当作无法认证的路由提供出去。catalog 路由仍可经自己的 provider 抵达它们；被拒的只有显式覆盖。
+- `adapter.ts` 把每次解析变成一份**不可变快照**——profiles 加上持有这些 provider 的 `createModels()` 集合——每个操作都在自己第一个 `await` 之前整体捕获一份。
+- 模型**显式配置**的 `maxTokens` 会成为 seam 的 `defaultMaxTokens`；从已安装 catalog 继承来的那份不会：pi-ai 要求 `Model.maxTokens` 表示模型的输出**能力**，而 `defaultMaxTokens` 是部署选定、发给未点名上限的请求的那个值，把前者物化成后者会让每个请求都被一个无人选择的数字封顶。
+
+### 快照，而不是共享集合
+
+`Models.streamSimple()` 惰性解析 provider——在返回的流首次被消费时，而那已在适配器 await 路由凭据之后。因此就地改动的单一集合，会让一个在旧配置下开始的请求在新配置下结束，或者撞上一个已不存在的 provider，尽管 `llm.prepareCall()` 早已冻结了该步的 config 并捕获了其适配器注册。配置变化改为构造**新**集合，正在被使用的那个原封不动，于是 seam 的每步冻结得以贯通到底：回复途中切换模型在下一步生效，绝不影响在途的那一步。
+
+### 目录原子替换
+
+可配置提供方目录跟随 profiles，因此每当一条声明路由出现或离开它都会变化。「撤销旧注册再新建一个」表达不了这件事：注册表拒绝的候选集合——比如一份键为 `deepseek-official` 的 profile，而 `llm-deepseek` 已声明了它——会让本插件的整个目录被撤走、Models 页变空，而且是静默的，因为 settings 变更回调把失败容住了。因此 `registerConfigurableProviders` 改为返回带 `replace(entries)` 的句柄，其「候选集先整体校验」的原子性与 `registerAdapter` 相同，插件改用它。被拒的替换只付出一条诊断；先前的条目继续服务。
+
+解析失败得响亮，并点名出问题的路由与模型：catalog 未描述的模型会回落到该路由自己的 `defaultContextWindow`／`defaultMaxTokens`，因此只公布 id 的列表也能得到可服务的路由；catalog 未提供的路由需要 `api`、`baseURL` 和非空的 `models` 列表。由于构造出的 `Provider` 是解析结果的一部分，协议或模型出错时最后可用的路由集合会继续服务——与此前坏的 settings 快照的行为完全一致。
+
+可配置提供方目录现在是已安装 catalog **与**当前 profile 声明的每条路由的并集，并在该集合变化时重新登记。没有这个并集，手工声明的路由就没有 settings 地址，任何配置界面都无法展示或编辑它。
+
+### 唯一档位什么也做不到的能力，报告为不可用
+
+pi-ai 把没有推理元数据的模型报告为只支持 `off` 一档，而适配器此前原样透传。它抵达 seam 时是一个单元素的 effort 列表，任何界面都会把它渲染成一个只有一项可选控件的选择器——而这个控件在撒谎：`off` 在派发时变成被*省略*的 reasoning 选项，与「不点名任何档位」产出的请求逐字节相同。自身默认就在思考的提供方会继续思考，界面却显示 `off` 已选中。
+
+因此只要 `model.reasoning` 为假，`reasoningInfo` 就省略 seam 的 `reasoning` 字段。判据是模型自身的元数据，而非模型的来源，所以它覆盖每一个手工声明的模型**以及** pi-ai 标记为不具备推理能力的那 251 个已安装 catalog 模型。它们此前提供那个孤零零的 `off`，现在什么也不提供，界面只剩提供方默认。携带推理元数据的模型不受影响——其档位列表仍不经筛选地穿过 seam、`off` 也在内，因为在那里它是在真实备选之间做选择。
+
+### 凭据留在 pi-ai 之外
+
+pi-ai 的 `Models` 自带一套凭据概念——按提供方 id 索引的 `CredentialStore`，配合 `envApiKeyAuth` 解析 `credential.key ?? env(VAR)`。采用它会在 `ctx.credentials` 之外制造第二个凭据事实源，更糟的是会把 harness 明确禁止的环境回落重新引进来：点名了却取不到的 `apiKeyEnv` 必须以 `MISSING_CREDENTIAL` 失败，而不是用环境里恰好持有的某个无关密钥完成认证。
+
+`ModelsImpl.applyAuth` 会把 `options.apiKey` 当作该请求的密钥，但这条路必须经由一个声明了 api-key 方法的提供方：`resolveProviderAuth` 在覆盖存在时短路到该方法，否则依次落到凭据存储与环境发现；若提供方压根没有 api-key 方法，它返回空，请求随即以 `Provider is not configured` 失败。因此 harness 一如既往经自身 seam 解析路由密钥，并把结果作为请求的 `apiKey` 传入；该集合构造时不带任何凭据存储。
+
+路由的 auth 由此推出。catalog 路由保留已安装提供方自己的 `auth`，从而为不点名凭据的 profile 保住其提供方原生环境发现，且在 `api` 覆盖之下同样保留：提供方读哪个环境是提供方自身的属性，而非其模型所讲协议格式的属性。例外是没有 api-key 方法的 catalog 提供方——`openai-codex` 只走 OAuth——此时点名了凭据的 profile 会在提供方原有 auth 之外再获得 harness 的方法，否则它配置的密钥会在任何请求发出之前被拒。这类路由上不点名凭据的 profile 什么也不加、并保留那句诚实的拒绝：本适配器没有可供解析的 OAuth 存储。手工声明的路由则获得一个 harness 自有的 `ApiKeyAuth`，它报告「已配置但无密钥」而非「未配置」，把该要求留给协议——那才是它真正所在的位置：pi-ai 的 OpenAI 兼容实现仍要求密钥或 `Authorization` 标头，并且会自己说出来。
+
+## Alternatives considered
+
+- **保留 `createProvider()` 但不建 `Models` 集合**，改由 `provider.streamSimple(model, ctx, {apiKey})` 发起。改动最小且凭据路径原封不动，但 `createProvider` 的 `auth` 是必填字段，这条路上它永远不会被调用——一份因签名而必填、却没有调用方的实现。它还让 `refreshModels` 需要手工构造 `RefreshModelsContext`，并使适配器始终不在 pi-ai 真正支持的运行时上。
+- **catalog 路由复用已安装提供方，只有声明式路由走 `createProvider()`**，且两者不共享解析。对 catalog 行为零风险，但 catalog 物化、端点覆盖与每模型配置这三件事都要各写两遍，而改指协议的 catalog 路由还得在解析中途跳到另一条路径。已采纳的拆法把不对称收敛在提供方构造这一处——那里的不对称是 pi-ai 不暴露已构造提供方的 API 实现所强加的。
+- **让每条路由都经 `createProvider()` 重建**，包括 catalog 路由。完全对称，但已构造的 `Provider` 不暴露自己的 `api`，于是协议表会成为「哪些提供方能用」的天花板——Bedrock 经独立入口加载其 Smithy 模块，会因此静默失效。
+- **完整暴露 pi-ai 的 `Model` 形状**（成本、输入模态、`thinkingLevelMap`、`compat`）。可配置性最大，但这些字段当前没有任何读取方，因此配了价格或模态什么也不会改变，却看起来像是受支持的。
+
+- **保留单个可变 `Models` 集合并重新同步。** 分配更少，且对每个同步完成解析的操作都是正确的；唯独对那个不同步的操作恰恰是错的：`stream()` 会在捕获模型与派发模型之间 await 一次凭据。
+- **用「先 dispose 再注册」模拟目录原子替换。** 无需改 seam，且在新集合有效时确实可用——而那正是从不需要原子性的那种情形。
+- **运行时动态 catalog**——`fetchModels` 加 `ModelsStore`，后台刷新。本次变更拒绝：它把模型列表变成需要缓存、失效与离线路径的外部可变状态，而产品需求是一次性的发现动作、其结果由用户采纳进 `settings.yaml`。该动作属于配置界面，与之一并暂缓；`settings.yaml` 始终是「路由服务什么」的唯一事实源。
+
+## Consequences
+
+配置一个提供方不再取决于 pi-ai 的发布节奏。网关、自建服务，或比锁定 catalog 更新的模型，都是一次 `settings.yaml` 编辑，过期的上下文窗口也能就地更正。废弃的 `/compat` 导入已经消失，因此 pi-ai 删除它不再是破坏性事件。`defaultMaxTokens` 现在只在部署明确给出时才自配置流出，不会从 catalog 元数据里发明一个上限。
+
+代价是：声明式路由会让 `settings.yaml` 变长，因为它必须自报端点、协议与模型 id。`api` 作用于整条路由，因此混合协议的 catalog 路由无法承载另一种协议的模型——把它拆成两个路由键是变通办法。没有任何环节查询提供方的 `/models`，因此模型列表的新鲜度只到最近一次编辑为止。有一种情形下报错形状发生变化：auth 解析不出任何值的路由，现在会在任何网络调用之前把 pi-ai 自己的诊断作为错误 `finish` 分片呈现，而此前的适配器会发出无密钥请求并呈现提供方的 401。
+
+## Testing
+
+`tests/catalog.spec.ts` 针对本地 mock 服务器端到端覆盖该契约：手工声明的路由带着自己的凭据流向自己的端点、它在可配置提供方目录中的出现、每模型覆盖从已安装 catalog 继承默认值、向 catalog 路由添加模型、带与不带端点覆盖的协议改指、catalog 独有元数据在覆盖后存活、无密钥姿态及其 `Authorization` 标头变通、只走 OAuth 的 catalog 路由用 profile 点名的密钥完成认证而无密钥者保持未配置、改指协议的路由保留其 catalog auth，以及每一种点名路由或模型的解析失败。`tests/catalog.spec.ts` 还钉住了快照与目录两项契约：在途请求即便其路由集在 credential await 期间改变，仍抵达它解析时对应的端点；下一个请求取用新配置；冲突的声明路由让目录保持完好；声明路由的条目随其 profile 出现与离开。`packages/llm/llm/tests/topology.spec.ts` 覆盖 `replace`——拒绝他人已拥有的候选同时保住当前集合、接受对自身条目的替换、允许空集合，以及 dispose 之后失败。`tests/sdk-options.spec.ts` 把 SDK 边界从已移除的 `/compat` 导入改指到协议表的 lazy api 模块，同时钉住「setup 失败以终止性错误分片而非抛出的形式抵达」。twin 的[设计验证角色](2026-06-13-twin-llm-adapters.md)不变。

.agents/notes/implemented/architecture/2026-08-04-declaring-a-provider-from-the-models-page.i18n.yaml

@@ -0,0 +1,6 @@
+# Bilingual-pair consistency record (docs/i18n/README.md): the git blob hash of each
+# side as of the last confirmed-consistent state. Both languages carry equal authority;
+# after editing either side, bring the other along and re-record with:
+#   pnpm run verify-translation-pairing --write .agents/notes/implemented/architecture/2026-08-04-declaring-a-provider-from-the-models-page.md
+2026-08-04-declaring-a-provider-from-the-models-page.md: 53996e488c467e00754837b83b7a33994d4813ee
+2026-08-04-declaring-a-provider-from-the-models-page.zh.md: fa61c48492eabf51f3d325078ceffa84ac52d12c

.agents/notes/implemented/architecture/2026-08-04-declaring-a-provider-from-the-models-page.md

@@ -0,0 +1,43 @@
+# Agent Note: Declaring a provider from the Models page
+
+Status: implemented
+
+English | [中文](2026-08-04-declaring-a-provider-from-the-models-page.zh.md)
+
+## Problem
+
+The two layers below made a pi-ai route [a declaration](2026-08-03-pi-ai-declared-provider-catalog.md) and gave the host a way to [interrogate a draft endpoint](2026-08-04-draft-provider-endpoint-interrogation.md). Neither reached a person who does not edit YAML: the Models page still offered one API-key field per provider and a fold with a base URL, so adding a gateway meant opening `$DSH_HOME/settings.yaml` and knowing the profile shape, and correcting a stale context window meant the same. The capability existed and the surface did not expose it.
+
+Two things were missing, and they are not the same shape. Editing an existing route's models is a *field* on a card that already exists. Declaring a route is a *create*: the route id is being chosen, so until it is chosen there is no settings address to edit.
+
+## Decision
+
+The model list is a component shared by both flows; the create is its own card.
+
+`ModelListEditor` edits a profile's `models` array — one row per model with id, display name, context window, and output cap — and owns the fetch action. An empty list means "serve this route's built-in catalog", so a row is only ever added deliberately; clearing an optional field drops it rather than storing a value the schema would reject, and a capacity that is not a positive integer is not stored at all.
+
+Fetching asks about the endpoint **the form currently shows** — a base URL edited but unsaved, a key typed but unstored — so adding a provider is one pass instead of save-then-return. The reply opens a picker rather than being written: candidates already configured start unchecked, so adopting a selection never overwrites a capacity the user corrected. A provider that cannot be interrogated is a detour, not a dead end; the adapter's own message appears beside rows that stay editable by hand.
+
+`CustomProviderCard` declares a route pi-ai does not ship. It is a separate card because the route id is chosen here: one `settings.mutate` sets the whole profile at `providers.<route>`, and the key travels separately through `credentials.set` under the same `<ROUTE>_API_KEY` derivation an existing provider uses. The three facts a hand-declared route cannot default — endpoint, protocol, and at least one model — gate the create button, so a failure names the field while the user is still looking at it.
+
+The protocol choices come from the namespace's **own schema**, read through the settings descriptor the page already fetches (`providers.*.api` is a union of the adapter's `supportedProtocols()`). No new wire field, no constant in the client, and no way for the offered choices to drift from the accepted ones.
+
+## Alternatives considered
+
+**Declare a provider through `ProviderEditor` with extra fields.** One card instead of two, but the editor is addressed by `settingsPath`, and a route being named has no path yet. Recomputing the path per keystroke would remount the card and discard the draft; deferring it would mean the editor's whole write path no longer described what it was editing.
+
+**Add a wire field for the protocol list.** Explicit, and the obvious first instinct. But the settings schema already crosses the wire and already contains the union, so a second copy could disagree with the first — and the one the adapter enforces is the schema.
+
+**Fetch against the stored profile instead of the live form.** No key would leave the form for an unsaved provider. But the flow that needs fetching most is the one where nothing is stored yet, and a form whose endpoint was edited would quietly interrogate the old one.
+
+**Write adopted candidates straight into the list.** Fewer clicks, but a fetch would then overwrite capacities the user had corrected, and a listing that discloses only ids would replace real numbers with nothing.
+
+## Consequences
+
+A gateway, a self-hosted server, or a model newer than the installed catalog is now configurable without leaving the browser, and the endpoint itself supplies the model ids where it can. The page grew two components and one shared list editor; the editor card's pi-ai fold grew from two fields to a list.
+
+What it costs: only pi-ai routes can be hand-declared, because `llm-pi-ai` is the one namespace whose profiles describe a whole provider — a `llm-deepseek` route stays a composition fact. Interrogation reaches only OpenAI-compatible endpoints, so a gateway speaking another protocol reports that it cannot be asked and its models are typed in. And the page now holds a key in component state for the duration of a fetch, which is the same exposure `credentials.set` already has and no longer than the card lives.
+
+## Testing
+
+`packages/client/ui-models/tests/provider-form.spec.tsx` drives the rendered page over a scripted wire face: adding, editing, and removing rows; a cleared optional field leaving the profile and a non-integer capacity never entering it; the interrogation carrying the edited endpoint, the unsaved key, and the profile's protocol; the picker's default selection, toggling, cancel, and adopt-keeps-tuned-rows; the empty, refused, and rejected-transport paths; the create writing one profile plus its credential; every gate on the create button; and the read-only posture. `protocolChoices` is covered against a schema that declares the union and one that does not.

.agents/notes/implemented/architecture/2026-08-04-declaring-a-provider-from-the-models-page.zh.md

@@ -0,0 +1,43 @@
+# Agent Note: 在 Models 页上声明一个提供方
+
+Status: implemented
+
+[English](2026-08-04-declaring-a-provider-from-the-models-page.md) | 中文
+
+## Problem
+
+下面两层已经让 pi-ai 路由变成[一份声明](2026-08-03-pi-ai-declared-provider-catalog.md)，并给了 host [询问草稿端点](2026-08-04-draft-provider-endpoint-interrogation.md)的能力。但两者都没有抵达不编辑 YAML 的人：Models 页仍然只为每个提供方提供一个 API 密钥输入框和一个装着 API 地址的折叠区，因此接入一个网关意味着打开 `$DSH_HOME/settings.yaml` 并知道 profile 的形状，更正一个过期的上下文窗口也是如此。能力已经存在，界面却没有暴露它。
+
+缺的是两件事，而它们的形状并不相同。编辑既有路由的模型，是一张已经存在的卡片上的一个*字段*；声明一条路由则是一次*创建*：路由 id 正在此处被选定，而在选定之前根本没有可编辑的 settings 地址。
+
+## Decision
+
+模型列表是两条流程共用的组件；创建则是它自己的卡片。
+
+`ModelListEditor` 编辑 profile 的 `models` 数组——一行一个模型，含 id、显示名称、上下文窗口与输出上限——并持有获取动作。空列表意味着「使用该路由的内置 catalog」，因此每一行都只会被刻意添加；清空某个可选字段会丢弃它，而不是存入一个 schema 会拒绝的值，不是正整数的容量则根本不会被存下。
+
+获取会询问表单**当前显示**的端点——已修改但未保存的 API 地址、已键入但未存储的密钥——因此新增一个提供方是一趟走完，而不是「先保存再回来」。回复会打开一个选择框而不是直接写入：已配置过的候选默认不勾选，因此采纳一次选择绝不会覆盖用户已更正的容量。无法被询问的提供方只是绕路而非死路；适配器自己的消息会出现在各行旁边，而这些行仍可手工编辑。
+
+`CustomProviderCard` 声明 pi-ai 未提供的路由。它之所以是独立卡片，正因为路由 id 是在这里选定的：一次 `settings.mutate` 在 `providers.<route>` 上设置整个 profile，密钥则经 `credentials.set` 单独传递，使用与既有提供方相同的 `<ROUTE>_API_KEY` 派生。手工声明的路由无法默认的三件事——端点、协议、至少一个模型——会门控创建按钮，因此失败会在用户仍看着该字段时点名它。
+
+协议选项来自该 namespace **自己的 schema**，经页面本就会获取的 settings 描述符读出（`providers.*.api` 是适配器 `supportedProtocols()` 的一个 union）。没有新增协议字段，客户端里没有常量，提供的选项也无从与被接受的集合发生漂移。
+
+## Alternatives considered
+
+**在 `ProviderEditor` 上加字段来声明提供方。** 两张卡片变一张，但编辑器由 `settingsPath` 寻址，而正在被命名的路由还没有路径。逐次按键重算路径会让卡片重新挂载并丢掉草稿；推迟计算则意味着编辑器的整条写入路径不再描述它正在编辑的东西。
+
+**为协议列表新增一个协议字段。** 显式，也是最直觉的第一反应。但 settings schema 本来就会跨越协议层、本来就含有那个 union，因此第二份副本可能与第一份不一致——而适配器强制执行的是 schema 那一份。
+
+**针对已存 profile 而非实时表单发起获取。** 对尚未保存的提供方来说，密钥就不会离开表单。但最需要获取的恰恰是「什么都还没存」的那条流程，而端点已修改的表单会悄悄去询问旧地址。
+
+**把采纳的候选直接写进列表。** 点击更少，但一次获取就会覆盖用户已更正的容量，而只公布 id 的列表会把真实数字替换成空。
+
+## Consequences
+
+网关、自建服务，或比已安装 catalog 更新的模型，如今无需离开浏览器就能配置，而模型 id 在端点能提供时由端点自己给出。页面多了两个组件和一个共用的列表编辑器；编辑卡片的 pi-ai 折叠区从两个字段长成了一个列表。
+
+代价是：只有 pi-ai 路由可以手工声明，因为 `llm-pi-ai` 是唯一一个其 profile 描述整个提供方的 namespace——`llm-deepseek` 路由仍是组合面的事实。询问只覆盖 OpenAI 兼容端点，因此讲其他协议的网关会报告自己无法被询问，其模型需手工键入。另外，页面在一次获取期间会把密钥保存在组件状态里，这与 `credentials.set` 已有的暴露面相同，且不长于卡片的存活时间。
+
+## Testing
+
+`packages/client/ui-models/tests/provider-form.spec.tsx` 在脚本化的协议面之上驱动渲染后的页面：添加、编辑与移除行；被清空的可选字段离开 profile、非整数容量从不进入；询问携带已修改的端点、未保存的密钥，以及 profile 自身的协议；选择框的默认选中、勾选切换、取消，以及「采纳保留已调优的行」；空列表、被拒、传输被拒三条路径；创建写入一份 profile 加其凭据；创建按钮上的每一道门控；以及只读姿态。`protocolChoices` 针对「声明了该 union」与「没有声明」两种 schema 都有覆盖。

.agents/notes/implemented/architecture/2026-08-04-draft-provider-endpoint-interrogation.i18n.yaml

@@ -0,0 +1,6 @@
+# Bilingual-pair consistency record (docs/i18n/README.md): the git blob hash of each
+# side as of the last confirmed-consistent state. Both languages carry equal authority;
+# after editing either side, bring the other along and re-record with:
+#   pnpm run verify-translation-pairing --write .agents/notes/implemented/architecture/2026-08-04-draft-provider-endpoint-interrogation.md
+2026-08-04-draft-provider-endpoint-interrogation.md: 65545098cd1063c40081481c1ac8f0afdb4fb390
+2026-08-04-draft-provider-endpoint-interrogation.zh.md: cb09042904f4ab1558c0c214d275a934234955ac

.agents/notes/implemented/architecture/2026-08-04-draft-provider-endpoint-interrogation.md

@@ -0,0 +1,50 @@
+# Agent Note: Interrogating a draft provider endpoint
+
+Status: implemented
+
+English | [中文](2026-08-04-draft-provider-endpoint-interrogation.zh.md)
+
+## Problem
+
+Once a pi-ai route became [a declaration rather than a catalog lookup](2026-08-03-pi-ai-declared-provider-catalog.md), a person adding an OpenAI-compatible gateway had to know its model ids before they could configure it. The adapter no longer constrains them to an installed catalog, which is the point, but it also means nothing tells the user what the endpoint actually serves — and most of these endpoints do publish that list at `GET /models`.
+
+The obvious answer, a dynamic runtime catalog refreshed in the background, was rejected with the layer below it: it makes a route's model list external mutable state needing a cache, an invalidation story, and an offline path, while the product need is narrower. What is needed is a *question asked once*, whose answer the user adopts into `settings.yaml` — so `settings.yaml` remains the only thing deciding what a route serves.
+
+The awkward part is that the question is about something that does not exist yet. The provider being added has no route, no stored profile, and no stored credential; the endpoint and key are values in a form the user is still typing. Every existing seam operation is keyed by a registered provider route, so none of them can carry this.
+
+## Decision
+
+Interrogation is keyed by **settings namespace**, not by provider route:
+
+- `ctx.llm.registerModelDiscovery(settingsNs, discover)` lets an adapter plugin offer to interrogate endpoints for the namespace it owns, and `ctx.llm.discoverModels(settingsNs, request)` asks. There is no way to enumerate which namespaces registered: a surface that cannot interrogate learns it from the refusal, and a list nothing consumed would be a required wire field doing nothing. The namespace is the right key because a configuration surface already holds it from the configurable-provider directory, and because a provider being added has no route to name.
+- `LlmModelDiscoveryRequest` carries the draft — an optional `provider`, an optional `baseURL`, an optional `api`, an optional `apiKey`, and a signal — and needs at least one of `provider` or `baseURL` to have anything to answer about. `provider` exists because a route the adapter already describes is answered from its own registry with no network call at all; only a route it does not describe reaches an endpoint. Nothing in this path writes settings or credentials. The one read is the credential of a route the request names: a configuration surface holds a redacted descriptor rather than the stored secret, so the draft's `apiKey` is present only while the user is typing one, and without that read an already-configured route would be interrogated unauthenticated and answer 401. The typed key wins, being the one under test.
+- `LlmDiscoveredModel` makes every field but `id` optional, because most listings disclose an id and nothing else. The reply is candidates, not a catalog: a surface adopting one still owes the capacities the adapter requires.
+- `llm.discoverModels` carries the same draft over the wire. Its `apiKey` is the third and last payload on which a secret may ride, alongside `settings.update`/`mutate` and `credentials.set`, and it is never stored or echoed back. It does ride the client's outgoing envelope like every other secret-bearing payload, where a `subscribeEnvelopes()` observer can see it; redacting that tap is a configuration-plane-wide change, not this method's to make alone. The method is loopback-only for a second reason besides the key: it makes the host issue a GET to a caller-chosen URL and reports the outcome, which is a probe an anonymous LAN caller must not have. Every refusal folds into `model-discovery-failed`, whose message is the adapter's own text and whose details name the endpoint asked but never the credential offered.
+
+`dsh-llm-pi-ai` implements the wire path as a plain `GET {baseURL}/models`, reading `openai-completions` and `openai-responses`: their `GET /models` shape with bearer auth is the one a gateway, a self-hosted server, and the official endpoints all agree on. Azure is excluded despite its OpenAI lineage — it authenticates with an `api-key` header and requires an `api-version` query — and Codex uses OAuth; both would have reported an authentication failure as a provider with no models. Every other protocol answers `DISCOVERY_UNSUPPORTED`, so the surface falls back to hand-entry rather than reporting a guessed response shape as an empty provider. `baseURL` is treated as a prefix rather than a URL to resolve against, so a deployment path such as `https://gateway.example/openai/v1` keeps its segments. The reply is read under a four-megabyte ceiling enforced on the bytes actually received — the endpoint is a URL the user typed, so a declared `content-length` is checked first as a courtesy but never trusted as the bound, matching `dsh-web-fetch`'s two-stage shape for its own caller-supplied URLs.
+
+### Why not pi-ai's own refresh machinery
+
+pi-ai supplies `createProvider({ fetchModels })` plus `Models.refresh()` and a `ModelsStore`, and the layer below already builds pi-ai `Provider` objects. Routing interrogation through them would have meant constructing a throwaway provider and collection per question, with a store whose entire purpose — persisting a catalog across runs — contradicts the decision that `settings.yaml` owns the catalog. It would also have bought nothing: **no built-in pi-ai provider implements `fetchModels`**, so the HTTP call and its response parsing are this package's code either way. A direct fetch says what is actually happening. The route's stored credential is resolved by the plugin's own per-request resolver, and only on the branch that reaches the network, so a catalog route answers without touching credentials and never fails over one the question did not need.
+
+## Alternatives considered
+
+**Key interrogation by provider route.** Symmetric with every other seam operation, and it would let the request omit the endpoint. But the case that motivates the feature — adding a provider — has no route, so the operation would only work for providers already configured, which are the ones that need it least.
+
+**Put the capability on `LlmAdapter`.** Adapters are reached through a route registration, so this has the same problem, plus it would make an adapter instance answer questions about endpoints it does not serve.
+
+**Have the host read the stored profile instead of accepting a draft.** No secret would cross the wire for an already-configured provider. But adding a provider would then require saving an unusable configuration first, and a form whose endpoint was edited but not yet saved would silently interrogate the old one. Accepting the draft keeps what the user sees and what is asked identical — with the credential as the one exception, because it is the one field a surface is never shown and so can never put in the draft.
+
+**Interrogate every pi-ai protocol.** Anthropic's listing happens to share OpenAI's envelope, and Google's does not. Supporting the ones that are easy would make coverage arbitrary and, worse, make a wrong guess at a response shape indistinguishable from a provider with no models. A protocol that says it cannot be interrogated sends the user to hand-entry, which is the documented fallback.
+
+**Buffer the reply with `response.text()` and check its length.** Simpler, but the bound would arrive after the bytes did, and the endpoint is whatever URL the user typed.
+
+## Consequences
+
+A person adding a gateway can ask it what it serves instead of hunting through its documentation, and the answer arrives as candidates they choose from rather than as configuration written behind their back. The seam gained a registry that is deliberately small: one offer per namespace, no storage, no lifecycle beyond the fiber.
+
+What it costs: the wire gained a third secret-carrying payload, so the configuration plane's write-only surface is now three methods rather than two. Discovery coverage is protocol-shaped rather than provider-shaped — an Anthropic-compatible gateway must be filled in by hand even though its listing would parse. And because nothing re-runs the question, a model list is still only as current as its last edit; that is the same trade the layer below made deliberately.
+
+## Testing
+
+`packages/llm/llm/tests/topology.spec.ts` covers the registry: one offer per namespace, disposal with the fiber, normalization that drops duplicate and unusable ids without inventing capacities, and the `NO_DISCOVERY`/`INVALID_DISCOVERY` refusals. `packages/llm/llm-pi-ai/tests/discovery.spec.ts` drives the probe against local HTTP servers — a listing with and without disclosed capacities, a preserved deployment path, an absent credential, a configured route supplying its own where the draft has none and a typed key winning over it, a catalog route answering without resolving one at all, dropped rows, 401/403 versus a server fault, a non-listing and a non-JSON body, an unreachable endpoint, caller cancellation, an unsupported protocol, and the size ceiling in both its declared-length and streamed forms. `packages/host/apiproxy/tests/api-proxy-config.spec.ts` covers the RPC over a real proxy: the draft reaching its namespace whole, absent fields staying absent, no namespace or credential being written, and a failure surfacing as `model-discovery-failed` with the credential absent from the serialized error.

.agents/notes/implemented/architecture/2026-08-04-draft-provider-endpoint-interrogation.zh.md

@@ -0,0 +1,50 @@
+# Agent Note: 询问草稿中的提供方端点
+
+Status: implemented
+
+[English](2026-08-04-draft-provider-endpoint-interrogation.md) | 中文
+
+## Problem
+
+当 pi-ai 路由变成[一份声明而非 catalog 查表](2026-08-03-pi-ai-declared-provider-catalog.md)之后，要接入一个 OpenAI 兼容网关的人，必须先知道它的模型 id 才能完成配置。适配器不再把人限制在已安装 catalog 里——这正是那次改动的目的——但也意味着没有任何东西告诉用户该端点究竟服务什么，而这类端点大多在 `GET /models` 上公布了这份列表。
+
+显而易见的答案——后台刷新的运行时动态 catalog——已随下层一并被拒绝：它会把路由的模型列表变成需要缓存、失效语义与离线路径的外部可变状态，而产品需求要窄得多。真正需要的是**只问一次**，其答案由用户采纳进 `settings.yaml`——从而让 `settings.yaml` 始终是唯一决定路由服务什么的东西。
+
+麻烦之处在于，被问的对象还不存在。正在新增的提供方没有路由、没有已存 profile、也没有已存凭据；端点与密钥都是用户尚在输入的表单值。而现有的每个 seam 操作都以已注册的提供方路由为键，因此没有一个能承载它。
+
+## Decision
+
+询问以 **settings namespace** 为键，而不是提供方路由：
+
+- `ctx.llm.registerModelDiscovery(settingsNs, discover)` 让适配器插件为自己拥有的 namespace 提供「询问端点」的能力，`ctx.llm.discoverModels(settingsNs, request)` 发起询问。没有任何办法枚举哪些 namespace 注册过：询问不了的界面会从那句拒绝里知道，而一份无人消费的列表只会变成一个什么都不做的必填协议字段。以 namespace 为键是对的，因为配置界面已经从可配置提供方目录里拿到了它，也因为正在新增的提供方没有路由可点名。
+- `LlmModelDiscoveryRequest` 携带草稿——可选的 `provider`、可选的 `baseURL`、可选的 `api`、可选的 `apiKey`，以及一个 signal——且 `provider` 与 `baseURL` 至少要有一个，才有东西可答。`provider` 之所以存在，是因为适配器已经描述过的路由直接由它自己的注册表作答、完全不联网；只有它未描述的路由才会抵达某个端点。这条路径不写 settings 与 credentials。唯一的读取是请求所点名路由的凭据：配置界面拿到的是脱敏描述符而非已存的机密，因此草稿里的 `apiKey` 只在用户正键入时才存在；没有这次读取，已配置好的路由就会被不带认证地询问，只换回一个 401。键入的密钥优先，因为那正是被测试的那一把。
+- `LlmDiscoveredModel` 除 `id` 外每个字段都可选，因为大多数列表只公布 id。回复是候选而非 catalog：采纳其中一条的界面仍要补上适配器所需的容量。
+- `llm.discoverModels` 把同一份草稿送过协议层。它的 `apiKey` 是 secret 可以搭乘的第三个、也是最后一个载荷（另两个是 `settings.update`/`mutate` 与 `credentials.set`），且绝不被存储或回显。它确实会像其他承载机密的载荷一样随客户端外发信封同行，`subscribeEnvelopes()` 观察者看得到；把那个抽头脱敏是整个配置面的改动，不该由这一个方法独自决定。除密钥之外它被钉在回环还有第二个理由：它让宿主向调用方选定的 URL 发起 GET 并回报结果，这是匿名 LAN 调用者不该拥有的探测能力。每一种拒绝都折叠为 `model-discovery-failed`，其消息是适配器自己的文本，details 点名被询问的端点，绝不点名所提供的凭据。
+
+`dsh-llm-pi-ai` 的实现只是一次朴素的 `GET {baseURL}/models`，且仅限 OpenAI 兼容协议。它们的列表形状是网关、自建服务与官方端点三方一致认可的那一种，而这正是该动作存在的场景。其余协议一律以 `DISCOVERY_UNSUPPORTED` 回答，让界面回退到手工填写，而不是把猜错的响应形状报成一个空提供方。`baseURL` 按前缀而非待解析 URL 处理，因此 `https://gateway.example/openai/v1` 这类部署路径会保留其路径段。回复在四兆字节上限下读取，且上限落在实际收到的字节上——端点是用户自己填的 URL，因此会先看声明的 `content-length` 作为善意提示，但绝不把它当作边界；这与 `dsh-web-fetch` 面对自己的调用方提供 URL 时所用的两段式形状一致。
+
+### 为什么不用 pi-ai 自己的 refresh 机制
+
+pi-ai 提供了 `createProvider({ fetchModels })` 加上 `Models.refresh()` 与 `ModelsStore`，而下层本来就在构造 pi-ai `Provider` 对象。把询问接到它们上面，意味着每问一次就要构造一个用完即弃的 provider 与集合，而那个 store 的全部目的——跨运行持久化 catalog——恰恰与「`settings.yaml` 拥有 catalog」的决定相抵触。而且它什么也换不来：**没有任何一个 pi-ai 内置 provider 实现了 `fetchModels`**，因此 HTTP 调用及其响应解析无论如何都是本包的代码。直接 fetch 才如实说出正在发生的事。路由已存的凭据由本插件自己那套逐请求解析器取出，且只在真正要联网的那条分支上进行，因此 catalog 路由作答时既不触碰凭据，也不会因为一把这次询问根本用不上的密钥而失败。
+
+## Alternatives considered
+
+**以提供方路由为键。** 与其他每个 seam 操作对称，也能让请求省去端点。但催生该功能的场景——新增提供方——没有路由，于是这个操作只对已配置好的提供方可用，而它们恰恰最不需要它。
+
+**把能力挂在 `LlmAdapter` 上。** 适配器要经由路由注册才能抵达，因此问题相同；而且这会让一个适配器实例去回答它并不服务的端点的问题。
+
+**让 host 读已存 profile，而不是接受草稿。** 对已配置好的提供方来说，不会有 secret 跨越协议层。但这样一来新增提供方就必须先保存一份不可用的配置，而端点已改却尚未保存的表单会静默地去询问旧地址。接受草稿让用户看见的与被询问的保持一致——凭据是唯一的例外，因为它是界面从不被展示、因而永远无法放进草稿的那个字段。
+
+**询问 pi-ai 的每一种协议。** Anthropic 的列表恰好与 OpenAI 共用同一层信封，而 Google 的不是。只支持容易的那几种会让覆盖范围变得任意；更糟的是，猜错的响应形状会与「该提供方没有模型」无法区分。一个明说自己无法被询问的协议，会把用户送去手工填写——那正是既定的回退路径。
+
+**用 `response.text()` 缓冲整个回复再判断长度。** 更简单，但上限会在字节已经到达之后才生效，而端点是用户随手填的任意 URL。
+
+## Consequences
+
+接入网关的人可以直接问它服务什么，而不必去翻它的文档；答案以候选形式抵达，由用户自己挑选，而不是被背着写进配置。seam 因此多了一个刻意保持很小的注册表：每个 namespace 一份、不存储、除 fiber 外没有生命周期。
+
+代价是：协议层多了第三个承载 secret 的载荷，配置面的只写接口从两个方法变成三个。发现能力按协议而非按提供方划分——一个 Anthropic 兼容网关即便其列表能被解析，也仍须手工填写。而且由于没有任何环节会重跑该询问，模型列表的新鲜度依旧只到最近一次编辑为止；这与下层刻意做出的取舍是同一个。
+
+## Testing
+
+`packages/llm/llm/tests/topology.spec.ts` 覆盖注册表：每个 namespace 一份、随 fiber dispose、丢弃重复与不可用 id 且不凭空补容量的归一化，以及 `NO_DISCOVERY`/`INVALID_DISCOVERY` 两种拒绝。`packages/llm/llm-pi-ai/tests/discovery.spec.ts` 针对本地 HTTP 服务器驱动探测——含与不含公布容量的列表、被保留的部署路径、无凭据、草稿没带密钥时已配置路由自行取用凭据且键入的密钥压过它、catalog 路由完全不解析凭据即作答、被丢弃的行、401/403 与服务器故障之别、非列表与非 JSON 响应、不可达端点、调用方取消、不支持的协议，以及尺寸上限的「声明长度」与「流式」两种形态。`packages/host/apiproxy/tests/api-proxy-config.spec.ts` 在真实 proxy 上覆盖该 RPC：草稿完整抵达其 namespace、缺席字段保持缺席、没有 namespace 或凭据被写入，以及失败以 `model-discovery-failed` 呈现且序列化后的错误里不含凭据。

.agents/notes/implemented/architecture/2026-08-06-web-markdown-incremental-ast-renderer.i18n.yaml

@@ -0,0 +1,6 @@
+# Bilingual-pair consistency record (docs/i18n/README.md): the git blob hash of each
+# side as of the last confirmed-consistent state. Both languages carry equal authority;
+# after editing either side, bring the other along and re-record with:
+#   pnpm run verify-translation-pairing --write .agents/notes/implemented/architecture/2026-08-06-web-markdown-incremental-ast-renderer.md
+2026-08-06-web-markdown-incremental-ast-renderer.md: 3599bfcc78dc4eefe5e82f461a469bdba15f3aae
+2026-08-06-web-markdown-incremental-ast-renderer.zh.md: 2e00977da58ef29a77c45abcecf0f3bb62737929

.agents/notes/implemented/architecture/2026-08-06-web-markdown-incremental-ast-renderer.md

@@ -0,0 +1,33 @@
+# Agent Note: Incremental streaming markdown through a direct mdast renderer
+
+Status: implemented
+
+English | [中文](2026-08-06-web-markdown-incremental-ast-renderer.zh.md)
+
+## Problem
+
+`MarkdownText` re-parsed the whole accumulated reply on every streaming publish: react-markdown's string-only API builds a fresh unified processor per render and runs micromark → mdast → hast → React over the full text, so per-chunk main-thread work grew linearly with the reply and the stream's cumulative cost grew quadratically. The existing mitigations (frame batching, the isolated streaming tail, the plain fence arm) bounded how often and how widely that work ran, never how much text each run re-parsed. Fixing it needs AST-level input — freezing settled blocks and re-parsing only the source tail — which the string-only wrapper structurally cannot express.
+
+## Decision
+
+`MarkdownText` renders mdast directly and parses incrementally while streaming:
+
+- **Grammars** ([parse.ts](../../../../packages/client/ui-primitives/src/markdown/parse.ts)): `parseGfm` (streaming arm and `extractMarkdownPlainText`) and `parseGfmWithMath` (settled arm) call `mdast-util-from-markdown` with the same micromark extensions the replaced remark plugins wrapped, so block boundaries are identical everywhere. `mathCompatibility` (ex `remarkMathCompatibility`) now exports its micromark extension directly.
+- **Incremental parsing** ([incremental.ts](../../../../packages/client/ui-primitives/src/markdown/incremental.ts)): CommonMark block parsing is line-based, so appended text reshapes only the parse frontier. `IncrementalMarkdownParser` keeps the trailing two blocks unstable (the last block is the frontier; the second-to-last is safety margin), freezes everything before them, and re-parses only the source tail from the last frozen block's `position.end.offset` — the parser's own offsets, no bespoke source scanning. Each source region parses O(1) times per stream instead of once per chunk; a single giant block (an unclosed fence) degrades to the old full-reparse cost and no worse. Non-append input resets the state under a bumped generation.
+- **Rendering** ([render.tsx](../../../../packages/client/ui-primitives/src/markdown/render.tsx), [katex.tsx](../../../../packages/client/ui-primitives/src/markdown/katex.tsx)): one switch over mdast node types replaces remark-rehype + react-markdown, reproducing the replaced pipeline's DOM byte-for-byte — table alignment as `text-align` styles, tight-list paragraph unwrapping, task-list classes and checkbox spacing, the footnote section (whose in-page anchors the protocol allowlist already reduced to plain text), literal raw HTML, the separator newlines that surface next to literal HTML text, and rehype-katex's three-arm error chain with KaTeX HTML mapped to React through the browser's own `DOMParser` (no wrapper element, so first/last-child margin rules still reach `.katex-display`; React 18 puts the `.katex-mathml` subtree in the HTML namespace exactly as the replaced pipeline did — a pre-existing limitation outside this parity contract, invisible to the visual `.katex-html` arm). Frozen blocks cache their React elements and keep source-offset keys, so crossing the freeze boundary reconciles instead of remounting; `MarkdownText` is memoized.
+
+The DOM is pinned by `tests/fixtures/markdown-dom`: fixtures recorded from the react-markdown implementation before the swap, which the new renderer must reproduce under a whitespace-normalizing serializer. A fixture diff is a user-visible markdown style change to review, never to re-record for a refactor. `tests/markdown-incremental.spec.tsx` holds the equivalence property — at every appended prefix, chunked at 1/3/7/16 bytes, the live component's DOM equals a fresh mount's — plus freeze-boundary DOM-node identity and reset behavior.
+
+This reverses the [assistant-markdown note](../feature/2026-07-23-web-assistant-markdown.md)'s rejected alternative ("maintain a custom React walker"): the incremental requirement is new evidence, the walker's security-sensitive branches (URL allowlist, image policy, inert HTML) were already product-owned functions, and the dependency no longer deleted owned code — it blocked the architecture. That note's untrusted-output policy and renderer selection are unchanged.
+
+## Alternatives considered
+
+**Keep react-markdown and split the source into per-segment `<ReactMarkdown>` instances.** Zero renderer ownership, but each frame parses the tail twice (boundary detection + render), settled math still re-parses everything, hast construction and the per-render processor remain, and blocks remount when crossing the freeze boundary because element trees cannot be cached across instances.
+
+**Render cached mdast through `mdast-util-to-hast` + `hast-util-to-jsx-runtime`.** Keeps upstream's node mappings for free, but retains the hast intermediate per frame and two new direct dependencies for a pipeline whose mapping surface is small, closed, and now pinned by fixtures.
+
+**Parse KaTeX output with `hast-util-from-html-isomorphic` (as rehype-katex does).** Pulls a parse5-based HTML parser into the bundle to parse trusted, vocabulary-constrained KaTeX output the browser's `DOMParser` (with the spec's SVG/MathML attribute adjustments) already parses identically.
+
+## Consequences
+
+Streaming per-chunk work now tracks the unstable tail instead of the whole reply, and react-markdown, remark-gfm, remark-math, rehype-katex, unified, and the hast chain left the browser bundle (`mdast-util-math` and `micromark-util-sanitize-uri` became direct dependencies; both were already transitive). The package owns ~25 node mappings, their tests, and the KaTeX DOM conversion — priced against the fixture contract that freezes their output. Two behavioral deviations, both healed by the settled full parse at finalize: a reference-style link or footnote whose definition lands on the other side of a freeze boundary renders literally while streaming, and a footnote reference can flash back to literal text when its definition freezes while the referencing block is still unstable. This module and KaTeX conversion assume a browser DOM (`DOMParser`), which the client-only package already did.

.agents/notes/implemented/architecture/2026-08-06-web-markdown-incremental-ast-renderer.zh.md

@@ -0,0 +1,33 @@
+# Agent Note: 经由直接 mdast 渲染器的增量流式 Markdown
+
+Status: implemented
+
+[English](2026-08-06-web-markdown-incremental-ast-renderer.md) | 中文
+
+## Problem
+
+`MarkdownText` 在每次流式发布时都重新解析整个已累积的回复:react-markdown 的纯字符串 API 每次渲染都新建 unified processor,并对全文跑完 micromark → mdast → hast → React,因此每个 chunk 的主线程工作量随回复长度线性增长,整个流的累计成本随之二次增长。既有缓解手段(帧级合并、隔离的流式尾部、围栏 plain 臂)约束的是这份工作跑多频繁、波及多广,从未约束每次重新解析多少文本。修复它需要 AST 级输入——冻结已定型的块、只重新解析源文本尾部——这是纯字符串封装在结构上无法表达的。
+
+## Decision
+
+`MarkdownText` 直接渲染 mdast,并在流式期间增量解析:
+
+- **语法**([parse.ts](../../../../packages/client/ui-primitives/src/markdown/parse.ts)):`parseGfm`(流式臂与 `extractMarkdownPlainText`)和 `parseGfmWithMath`(定稿臂)以被替换的 remark 插件所包装的同一组 micromark 扩展调用 `mdast-util-from-markdown`,因此各处块边界完全一致。`mathCompatibility`(原 `remarkMathCompatibility`)现在直接导出其 micromark 扩展。
+- **增量解析**([incremental.ts](../../../../packages/client/ui-primitives/src/markdown/incremental.ts)):CommonMark 块解析按行推进,追加文本只会重塑解析前沿。`IncrementalMarkdownParser` 保留末尾两个块不稳定(最后一块是前沿;倒数第二块是安全裕量),冻结其前的所有块,只从最后一个冻结块的 `position.end.offset` 起重新解析源尾部——用的是解析器自己的偏移量,没有任何自制源扫描。每个源区间在整个流中解析 O(1) 次而非每 chunk 一次;单个巨型块(未闭合围栏)退化为旧的全量重解析成本,不会更差。非追加输入在递增的 generation 下重置状态。
+- **渲染**([render.tsx](../../../../packages/client/ui-primitives/src/markdown/render.tsx)、[katex.tsx](../../../../packages/client/ui-primitives/src/markdown/katex.tsx)):一个对 mdast 节点类型的 switch 取代 remark-rehype + react-markdown,逐字节复刻被替换管线的 DOM——表格对齐渲染为 `text-align` 样式、紧凑列表段落解包、任务列表类名与复选框空格、脚注区(其页内锚点本就被协议白名单降为纯文本)、字面 raw HTML、会与字面 HTML 文本相邻显形的分隔换行,以及 rehype-katex 的三臂容错链,KaTeX HTML 经浏览器自带的 `DOMParser` 映射为 React(无包裹元素,首/末子元素的 margin 规则仍能作用于 `.katex-display`;React 18 会把 `.katex-mathml` 子树放进 HTML 命名空间,与被替换管线完全一致——既有限制,不在本对等性契约范围内,对承担视觉渲染的 `.katex-html` 臂不可见)。冻结块缓存其 React 元素并保持源偏移 key,跨过冻结边界时走 reconcile 而非重挂载;`MarkdownText` 已 memo 化。
+
+DOM 由 `tests/fixtures/markdown-dom` 钉死:fixture 录制自替换前的 react-markdown 实现,新渲染器必须在空白规整序列化器下复现。fixture 差异即用户可见的 markdown 样式变更,必须按此评审,绝不能为重构而重录。`tests/markdown-incremental.spec.tsx` 承载等价性性质——以 1/3/7/16 字节分块,在每个追加前缀处,常驻组件的 DOM 都等于全新挂载——外加冻结边界的 DOM 节点同一性与重置行为。
+
+这推翻了[助手 Markdown Note](../feature/2026-07-23-web-assistant-markdown.md) 中被否决的备选("维护一个自定义 React walker"):增量需求是当时不存在的新证据,walker 的安全敏感分支(URL 白名单、图片策略、惰性 HTML)本就是产品自有函数,而该依赖不再删减自有代码——它阻塞了架构。该 Note 的不可信输出策略与渲染器选型不变。
+
+## Alternatives considered
+
+**保留 react-markdown,把源文本切成逐段 `<ReactMarkdown>` 实例。** 渲染器零自有成本,但每帧对尾部解析两次(边界检测 + 渲染),定稿数学仍要全量重解析,hast 构建与逐渲染 processor 依旧存在,且块跨过冻结边界时会重挂载——元素树无法跨实例缓存。
+
+**用 `mdast-util-to-hast` + `hast-util-to-jsx-runtime` 渲染缓存的 mdast。** 白拿上游节点映射,但每帧保留 hast 中间层,并为一个映射面小、封闭、且已被 fixture 钉死的管线引入两个新直接依赖。
+
+**用 `hast-util-from-html-isomorphic` 解析 KaTeX 输出(rehype-katex 的做法)。** 为解析可信、词汇受限的 KaTeX 输出把基于 parse5 的 HTML 解析器拉进 bundle,而浏览器自带的 `DOMParser`(带规范的 SVG/MathML 属性调整)解析结果完全相同。
+
+## Consequences
+
+流式的每 chunk 工作量现在跟随不稳定尾部而非整个回复,react-markdown、remark-gfm、remark-math、rehype-katex、unified 及 hast 链退出浏览器 bundle(`mdast-util-math` 与 `micromark-util-sanitize-uri` 成为直接依赖;两者原本就是传递依赖)。包自有约 25 个节点映射、其测试以及 KaTeX DOM 转换——代价由冻结其输出的 fixture 契约对冲。两个行为偏差,均在定稿的全量解析处自愈:定义落在冻结边界另一侧的引用式链接或脚注在流式期间渲染为字面文本;当脚注定义先冻结而引用块仍不稳定时,脚注引用可能闪回字面文本。本模块与 KaTeX 转换假定浏览器 DOM(`DOMParser`),这个 client-only 包本就如此。

.agents/notes/implemented/architecture/2026-08-06-web-shell-dist-chunk-layout.i18n.yaml

@@ -0,0 +1,6 @@
+# Bilingual-pair consistency record (docs/i18n/README.md): the git blob hash of each
+# side as of the last confirmed-consistent state. Both languages carry equal authority;
+# after editing either side, bring the other along and re-record with:
+#   pnpm run verify-translation-pairing --write .agents/notes/implemented/architecture/2026-08-06-web-shell-dist-chunk-layout.md
+2026-08-06-web-shell-dist-chunk-layout.md: 1c7b4273dc243685317b149e2fd7fddf2a6c18d1
+2026-08-06-web-shell-dist-chunk-layout.zh.md: 6f4b94e0bd7412e480458e34922273b389aa8892

.agents/notes/implemented/architecture/2026-08-06-web-shell-dist-chunk-layout.md

@@ -0,0 +1,50 @@
+# Agent Note: Web shell dist chunk split and directory layout
+
+Status: implemented
+
+English | [中文](2026-08-06-web-shell-dist-chunk-layout.zh.md)
+
+## Problem
+
+The apps/web shell previously built into a single ~1.2 MB (minified) index chunk, roughly 80% of it vendor bytes — KaTeX, the boot grammars and the shiki engine, react-dom, the markdown pipeline — fused with all the workspace shell code (about one fifth). Any one-line shell change rehashed the whole chunk, forcing returning clients to redownload everything; `dist/assets/` was a flat single-level spread of 100-plus files (the main chunk, 23 lazy-loaded grammar chunks, 59 KaTeX font faces, and sourcemaps intermixed), impossible to navigate.
+
+## Decision
+
+`apps/web/vite.config.ts` splits the shell into two initial chunks via `manualChunks` and sorts the output into directories via naming functions; the entire configuration contains zero regexes — an exact-package-name Set, a filename list, an extension list.
+
+**Membership** (`VENDOR_PACKAGES`, by exact npm package name):
+
+- `vendor` = the three heavy rendering families: math (katex), highlight (shiki), markdown (the micromark/mdast parse pipeline — the incremental React renderer above it is workspace code and not part of this). The live membership is `VENDOR_PACKAGES`; the list is the packages workspace code **imports directly**: the remaining private transitive dependencies (the oniguruma family, @shikijs/core, character tables, dozens more) are referenced only by listed members, so rollup's chunk coloring pulls them into vendor automatically; dependencies shared with the index side fall back to index, diluting it by a few KB — not a correctness issue.
+- **Every vendor member must be react-free (the boundary invariant)**: rollup folds a module shared between the entry and a manual chunk into the manual chunk — one listed package importing react/jsx-runtime would drag the single shared react copy into vendor, away from index. The React side of markdown/math rendering is workspace code and naturally lives in index, so the whole react family stays pinned to index.
+- `index` (the default chunk) = the react family (react, react-dom, scheduler, use-sync-external-store), vendored cordis, all workspace code, and the unlisted small pieces (anser, clsx).
+- `@shikijs/langs` is special-cased: the boot grammars (`BOOT_GRAMMAR_FILES`: typescript, shellscript, json — the three that highlight.ts statically imports, all self-contained data modules with zero internal imports) go into vendor; the remaining 23 lazy-loaded grammars get no assignment and each keeps its own on-demand chunk.
+- `index.html` is wired up automatically by vite: index loads via `<script>` and vendor via `<link rel="modulepreload">`, so the two chunks fetch in parallel with no waterfall.
+
+**Directory layout** (`chunkFileNames` + `assetFileNames`):
+
+- The `assets/` root keeps only the index and vendor js (with their adjacent sourcemaps) and css.
+- Grammar chunks go under `assets/langs/`. The criterion is whether a chunk's `moduleIds` include an `@shikijs/langs` member, not the facade: the shared chunks of embedded grammars (php/ruby/mdx embed html+javascript, which rollup splits out for sharing) **have no facade**, so a facade criterion would miss them; index and vendor are excluded by name, because vendor legitimately carries the three boot grammars.
+- Fonts go under `assets/fonts/` (`FONT_EXTENSIONS`: woff2/woff/ttf; today all of them are KaTeX faces referenced by vendor.css — katex.min.css is imported by an index-side component, but CSS modules go through manualChunks like any module and follow `katex` into vendor.css; the browser fetches only woff2, on demand and only when a formula renders).
+- Sourcemaps need no arrangement: rollup writes each `.map` next to its js and references it by bare relative filename, so when a chunk moves directories its map follows automatically.
+
+All cross-directory references (index's dynamic imports into `langs/`, same-directory relative references among grammar chunks, vendor.css's relative references into `fonts/`) are emitted by the bundler, so the runtime needs zero accompanying changes; the host-side webserver serves the nested paths verbatim under its static prefix.
+
+## Alternatives considered
+
+- **Serving react and the other vendors from a CDN**: dsh web targets local/intranet hosts (often without internet access), so a CDN is simply unavailable; react is the platform seed external of every plugin bundle (the shell is its sole supplier), and switching to the CDN global-variable form would touch three places — the platform manifest, the seed, and the module table; the caching benefit is already delivered by the vendor split.
+- **An inverse catch-all rule (everything in node_modules except the react family goes to vendor)**: membership cannot be read off the configuration, and small pieces like anser/clsx get misassigned to vendor; superseded by the positive exact-package-name list.
+- **Regex family matching**: hard to read; exact package names plus rollup's automatic coloring of transitive dependencies make pattern matching unnecessary.
+- **Identifying grammar chunks by facadeModuleId**: the facade-less shared chunks of embedded grammars would go undetected and fall back to the root directory; the `moduleIds` membership criterion covers both shapes.
+- **Sheltering a react-edged rendering facade in vendor** (the historical react-markdown was one): rollup's shared-module folding would drag the single react copy into vendor, breaking the "react belongs to index" boundary; the constraint is codified as the list's boundary invariant.
+- **Lazy-loading KaTeX wholesale, or turning the boot TypeScript grammar lazy**: either would change first-frame rendering behavior (the fallback for formulas / the first code block); that trade-off is independent of the dist layout and is decided separately.
+
+## Verification
+
+The audit tool ships with the repository: `node scripts/attribute-chunk-bytes.mjs <chunk.js>` (zero-dependency sourcemap VLQ byte attribution, aggregated by npm package / workspace directory). It verifies that vendor contains no workspace bytes, that the react family (including react/jsx-runtime) sits entirely in index, and that the npm side of index retains only the react family plus anser/clsx; the lazy grammar chunk count matches the `LAZY_GRAMMARS` table one to one; the browser keyless replay case is verbatim-identical to the pre-change baseline (apart from environment-specific local reds), so the two-chunk shell loads and renders with no regression.
+
+## Consequences
+
+- A shell code change rehashes only index (about one third of the dist output); vendor (about two thirds) stays cache-stable across shell releases and is invalidated only by dependency upgrades.
+- `dist/assets/` is navigable: two js/css pairs at the root, on-demand grammars in `langs/`, fonts in `fonts/`.
+- Maintenance cost: when workspace code adds a direct import of a rendering family's facade package, `VENDOR_PACKAGES` must be updated alongside (an omission merely dilutes index, nothing breaks); when the boot grammar set grows in highlight.ts without `BOOT_GRAMMAR_FILES` following, that grammar silently lands in index, visible only to a dist audit.
+- The webserver's static surface has no compression yet, so the gzip size win is still on the table; transport-layer compression is a separate, independent decision.

.agents/notes/implemented/architecture/2026-08-06-web-shell-dist-chunk-layout.zh.md

@@ -0,0 +1,50 @@
+# Agent Note: Web 壳产物的 chunk 切分与目录布局
+
+Status: implemented
+
+[English](2026-08-06-web-shell-dist-chunk-layout.md) | 中文
+
+## Problem
+
+apps/web 的壳此前打成单一约 1.2 MB（minified）的 index chunk，其中约八成是 vendor 字节——KaTeX、boot 语法与 shiki 引擎、react-dom、markdown 管线——与全部 workspace 壳代码（约五分之一）熔在一起。任何一行壳代码改动都让整个 chunk 换哈希，回头客户端全量重新下载；`dist/assets/` 是 100 多个文件的单层平铺（主 chunk、23 个懒加载语法 chunk、59 个 KaTeX 字体面、sourcemap 混居），无从导航。
+
+## Decision
+
+`apps/web/vite.config.ts` 以 `manualChunks` 把壳切成两个初始 chunk，并以输出命名函数归类目录；整套配置零正则——精确包名 Set、文件名清单、扩展名清单。
+
+**成员归属**（`VENDOR_PACKAGES`，按精确 npm 包名）：
+
+- `vendor` = 三个重渲染家族：math（katex）、highlight（shiki）、markdown（micromark/mdast 解析管线——其上的增量 React 渲染器是 workspace 代码，不在此列）。成员以 `VENDOR_PACKAGES` 为活口径，清单 = workspace 代码**直接 import** 的包：其余私有传递依赖（oniguruma 系、@shikijs/core、字符表等数十个）只被清单成员引用，rollup 的 chunk 着色自动将其并入 vendor；与 index 侧共享的依赖回落 index，只稀释几 KB，不构成正确性问题。
+- **vendor 全员必须 react-free（边界不变量）**：rollup 会把入口与 manual chunk 共享的模块并入 manual chunk——清单里出现任何 import react/jsx-runtime 的包，唯一一份 react 副本就会被拽进 vendor、脱离 index。markdown/math 的 React 渲染侧是 workspace 代码天然住 index，react 族因此全部钉在 index。
+- `index`（默认 chunk）= react 族（react、react-dom、scheduler、use-sync-external-store）、vendored cordis、全部 workspace 代码及未列入的小件（anser、clsx）。
+- `@shikijs/langs` 特判：boot 语法（`BOOT_GRAMMAR_FILES`：typescript、shellscript、json——highlight.ts 静态 import 的三件，均为零内部 import 的自含数据模块）进 vendor；其余 23 个懒加载语法不做指派，各自保持按需 chunk。
+- `index.html` 由 vite 自动接线：index 走 `<script>`、vendor 走 `<link rel="modulepreload">`，两 chunk 并行拉取，无瀑布。
+
+**目录布局**（`chunkFileNames` + `assetFileNames`）：
+
+- `assets/` 根只留 index 与 vendor 的 js（含随行 sourcemap）与 css。
+- 语法 chunk 归 `assets/langs/`。判据是 chunk 的 `moduleIds` 含 `@shikijs/langs` 成员，而非 facade：内嵌语法共享 chunk（php/ruby/mdx 内嵌 html+javascript，被 rollup 拆出共享）**没有 facade**，facade 判据会漏；index/vendor 按名排除，因 vendor 合法携带 boot 三语法。
+- 字体归 `assets/fonts/`（`FONT_EXTENSIONS`：woff2/woff/ttf；今日全部为 vendor.css 引用的 KaTeX 字面——katex.min.css 虽由 index 侧组件 import，css 模块同样经 manualChunks 归属、随 `katex` 落入 vendor.css；浏览器按需只拉 woff2，且仅在公式渲染时）。
+- sourcemap 无需安排：rollup 把 `.map` 写在各自 js 旁并以裸相对文件名引用，chunk 挪目录 map 自动跟随。
+
+跨目录引用（index 的动态 import 指向 `langs/`、语法 chunk 间同目录相对引用、vendor.css 相对引用 `fonts/`）均由构建器生成，运行时零配套改动；host 侧 webserver 按静态前缀原样服务嵌套路径。
+
+## Alternatives considered
+
+- **react 等 vendor 走 CDN**：dsh web 面向本机/内网主机（常无外网），CDN 直接不可用；react 是全部插件 bundle 的 platform seed external（壳是唯一供给方），改 CDN 全局变量形态需牵动 platform 清单/seed/模块表三处；缓存收益由 vendor 切分即可取得。
+- **反向兜底规则（node_modules 除 react 族全归 vendor）**：成员从配置上读不出来，且把 anser/clsx 类小件错归 vendor；被正向精确包名清单取代。
+- **正则家族匹配**：可读性差；精确包名 + rollup 对传递依赖的自动着色使模式匹配没有必要。
+- **以 facadeModuleId 识别语法 chunk**：无 facade 的内嵌语法共享 chunk 会漏检落回根目录；`moduleIds` 成员判据覆盖两种形态。
+- **在 vendor 里收留带 react 边的渲染门面**（历史上的 react-markdown 属此类）：会经 rollup 的共享模块归并把唯一 react 副本拽进 vendor，破坏「react 归 index」的边界；该约束已成文为清单的边界不变量。
+- **KaTeX 整体懒加载、boot TypeScript 语法转懒**：会改变首帧渲染行为（公式/首个代码块的回退），是独立于产物布局的取舍，另行决策。
+
+## Verification
+
+审计工具随库：`node scripts/attribute-chunk-bytes.mjs <chunk.js>`（零依赖 sourcemap VLQ 字节归属，按 npm 包/workspace 目录聚合）。以其复核：vendor 不含任何 workspace 字节、react 族（含 react/jsx-runtime）全量位于 index、index 的 npm 侧仅剩 react 族与 anser/clsx；懒语法 chunk 数量与 `LAZY_GRAMMARS` 表一一对应；浏览器 keyless replay 用例与改动前基线逐字一致（本机环境性红除外），两 chunk 壳装载渲染无回归。
+
+## Consequences
+
+- 壳代码改动只重哈希 index（约为产物三分之一）；vendor（约三分之二）跨壳版本缓存稳定，仅依赖升级时失效。
+- `dist/assets/` 可导航：根两对 js/css，`langs/` 按需语法，`fonts/` 字体。
+- 维护成本：workspace 代码新增对某渲染家族门面包的直接 import 时需同步 `VENDOR_PACKAGES`（漏列仅稀释 index，不致坏）；在 highlight.ts 扩 boot 语法集而未同步 `BOOT_GRAMMAR_FILES` 时，该语法静默落入 index，仅产物审计可见。
+- webserver 静态面尚无压缩，gzip 体量是潜在值；传输层压缩是另一项独立决策。

.agents/notes/implemented/feature/2026-07-23-web-assistant-markdown.i18n.yaml

@@ -2,5 +2,5 @@
 # side as of the last confirmed-consistent state. Both languages carry equal authority;
 # after editing either side, bring the other along and re-record with:
 #   pnpm run verify-translation-pairing --write .agents/notes/implemented/feature/2026-07-23-web-assistant-markdown.md
-2026-07-23-web-assistant-markdown.md: c21f2dc3c0aff98dd6ea6a88ea7a4742198c92d6
-2026-07-23-web-assistant-markdown.zh.md: e6677a6a22050b59343a67d6b282b92695c24009
+2026-07-23-web-assistant-markdown.md: ad86559e3b5294b6bd67d69ff5c6ce37a5172008
+2026-07-23-web-assistant-markdown.zh.md: b47375db843ea1c81b913e3f7ac8cd1bbc278830

.agents/notes/implemented/feature/2026-07-23-web-assistant-markdown.md

@@ -12,9 +12,9 @@ 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 raw-HTML parsing. A micromark attention extension reuses the CommonMark resolver while letting runs of at least two asterisks close after Unicode punctuation when followed immediately by CJK text. This exception covers punctuation-terminated strong emphasis in whitespace-free CJK prose during streaming and after settlement; single-asterisk emphasis, non-CJK adjacency, escaped source, code, and math retain upstream 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.
+`MarkdownText` parses with `mdast-util-from-markdown` plus the GFM micromark extensions and renders the mdast tree through the package's own renderer, parsing incrementally while a turn streams (the [incremental AST renderer note](../architecture/2026-08-06-web-markdown-incremental-ast-renderer.md) owns that mechanism and its DOM-parity contract). It covers CommonMark blocks plus GFM tables, task lists, strikethrough, and autolinks without raw-HTML parsing. A micromark attention extension reuses the CommonMark resolver while letting runs of at least two asterisks close after Unicode punctuation when followed immediately by CJK text. This exception covers punctuation-terminated strong emphasis in whitespace-free CJK prose during streaming and after settlement; single-asterisk emphasis, non-CJK adjacency, escaped source, code, and math retain upstream 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). When one inline-code token consists entirely of an absolute HTTP(S) URL, its code chrome contains the same keyboard-focusable safe external anchor as an ordinary link; port, path, and query text remain unchanged, while commands, partial URLs, other schemes, and fenced code stay inert. `CodeBlock` ships a language banner and a copy control (`复制` / `复制成功`). Finalized text renders KaTeX through `remark-math` and `rehype-katex`; `remarkMathCompatibility` maps `\(...\)`, `\[...\]`, and block-level same-line `$$...$$` to the same standard math AST nodes. This is a narrow parser compatibility layer, not a regex rewrite or malformed-model-output repair. Streaming stays literal until finalization so incomplete formulae do not flash errors. Citation pills, heading anchors, the thinking-small markdown variant, and custom □/☑ task markers remain out of scope; GFM task lists keep native checkboxes.
+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). When one inline-code token consists entirely of an absolute HTTP(S) URL, its code chrome contains the same keyboard-focusable safe external anchor as an ordinary link; port, path, and query text remain unchanged, while commands, partial URLs, other schemes, and fenced code stay inert. `CodeBlock` ships a language banner and a copy control (`复制` / `复制成功`). Finalized text renders KaTeX through the settled grammar's math extensions; `mathCompatibility` maps `\(...\)`, `\[...\]`, and block-level same-line `$$...$$` to the same standard math AST nodes. This is a narrow parser compatibility layer, not a regex rewrite or malformed-model-output repair. Streaming stays literal until finalization so incomplete formulae do not flash errors. Citation pills, heading anchors, the thinking-small markdown variant, and custom □/☑ task markers remain out of scope; 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.
 
@@ -26,7 +26,7 @@ Fenced code and GFM tables own horizontal overflow so long content cannot widen
 
 ## Alternatives considered
 
-**Promote the existing mdast and micromark development dependencies and maintain a custom React walker.** This avoids a new parser family but makes the product own every node mapping, GFM extension, and security-sensitive rendering branch. The dedicated React renderer keeps that traversal upstream while preserving an AST-to-React path.
+**Promote the existing mdast and micromark development dependencies and maintain a custom React walker.** This avoids a new parser family but makes the product own every node mapping, GFM extension, and security-sensitive rendering branch. The dedicated React renderer keeps that traversal upstream while preserving an AST-to-React path. *Later reversed on new evidence — incremental streaming parsing needs AST-level input the string-only wrapper cannot provide; the [incremental AST renderer note](../architecture/2026-08-06-web-markdown-incremental-ast-renderer.md) owns that decision.*
 
 **Replace `MessageText` with Markdown rendering.** This formats user prompts and steering as a side effect. Those authored surfaces remain literal until the product chooses that behavior explicitly.
 
@@ -42,4 +42,4 @@ Fenced code and GFM tables own horizontal overflow so long content cannot widen
 
 ## 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. URL-shaped inline code becomes navigable without changing its visible literal, while unsafe schemes and mixed code remain non-interactive. Code fences share one chrome and copy path with tool and details surfaces. The initial Web shell includes the Markdown parser, GFM runtime, KaTeX, and shiki allowlist; citation, anchor, and thinking-small surfaces remain deferred.
+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 only the unstable tail after each accumulated update; incomplete Markdown can temporarily change the tail's structure, but the isolated tail bounds React invalidation and the final event does not switch renderers. URL-shaped inline code becomes navigable without changing its visible literal, while unsafe schemes and mixed code remain non-interactive. Code fences share one chrome and copy path with tool and details surfaces. The initial Web shell includes the Markdown parser, GFM runtime, KaTeX, and shiki allowlist; citation, anchor, and thinking-small surfaces remain deferred.

.agents/notes/implemented/feature/2026-07-23-web-assistant-markdown.zh.md

@@ -12,9 +12,9 @@ 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。一个 micromark attention 扩展复用 CommonMark resolver，同时允许至少两个星号组成的连续序列在 Unicode 标点后闭合，前提是其后紧邻 CJK 文本。这一例外涵盖流式输出期间与完成后无空格 CJK 文本中以标点结尾的粗体；单星号强调、紧邻非 CJK 文本的情况、已转义源文本、代码与数学公式仍沿用上游解析行为。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。
+`MarkdownText` 以 `mdast-util-from-markdown` 加 GFM micromark 扩展解析，并经包内自有渲染器渲染 mdast 树，轮次流式输出期间增量解析（[增量 AST 渲染器 Note](../architecture/2026-08-06-web-markdown-incremental-ast-renderer.md) 拥有该机制及其 DOM 一致性契约）。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。一个 micromark attention 扩展复用 CommonMark resolver，同时允许至少两个星号组成的连续序列在 Unicode 标点后闭合，前提是其后紧邻 CJK 文本。这一例外涵盖流式输出期间与完成后无空格 CJK 文本中以标点结尾的粗体；单星号强调、紧邻非 CJK 文本的情况、已转义源文本、代码与数学公式仍沿用上游解析行为。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）。当单个行内代码 token 完全由绝对 HTTP(S) URL 构成时，其代码外框会包含一个与普通链接相同、可通过键盘聚焦的安全外链锚点；端口、路径与查询文本保持不变，而命令、非完整 URL、其他 scheme 与围栏代码仍不会成为链接。`CodeBlock` 提供语言横幅与复制控件（`复制` / `复制成功`）。已完成的文本通过 `remark-math` 和 `rehype-katex` 渲染 KaTeX；`remarkMathCompatibility` 将 `\(...\)`、`\[...\]` 和块级同一行 `$$...$$` 映射为同一套标准数学 AST 节点。这是一层小范围的解析器兼容层，不是正则重写，也不修复格式错误的模型输出。流式输出在完成前保持按字面渲染，避免不完整公式闪现错误。引用胶囊、标题锚点、thinking-small markdown 变体，以及自定义 □/☑ 任务标记仍不在范围内；GFM 任务列表继续使用原生复选框。
+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）。当单个行内代码 token 完全由绝对 HTTP(S) URL 构成时，其代码外框会包含一个与普通链接相同、可通过键盘聚焦的安全外链锚点；端口、路径与查询文本保持不变，而命令、非完整 URL、其他 scheme 与围栏代码仍不会成为链接。`CodeBlock` 提供语言横幅与复制控件（`复制` / `复制成功`）。已完成的文本通过定稿语法的数学扩展渲染 KaTeX；`mathCompatibility` 将 `\(...\)`、`\[...\]` 和块级同一行 `$$...$$` 映射为同一套标准数学 AST 节点。这是一层小范围的解析器兼容层，不是正则重写，也不修复格式错误的模型输出。流式输出在完成前保持按字面渲染，避免不完整公式闪现错误。引用胶囊、标题锚点、thinking-small markdown 变体，以及自定义 □/☑ 任务标记仍不在范围内；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。
 
@@ -26,7 +26,7 @@ assistant 生成的链接目标地址仅限绝对 HTTP、HTTPS 与 mailto URL。
 
 ## 考虑过的替代方案
 
-**将现有的 mdast 与 micromark 开发依赖提升为正式依赖，并维护自定义 React walker。**此方案避免引入新的解析器体系，但产品需要自行负责每种节点映射、GFM 扩展和安全敏感的渲染分支。专用 React 渲染器将这套遍历交由上游维护，同时保留 AST 到 React 的处理路径。
+**将现有的 mdast 与 micromark 开发依赖提升为正式依赖，并维护自定义 React walker。**此方案避免引入新的解析器体系，但产品需要自行负责每种节点映射、GFM 扩展和安全敏感的渲染分支。专用 React 渲染器将这套遍历交由上游维护，同时保留 AST 到 React 的处理路径。*后因新证据被推翻——增量流式解析需要纯字符串封装无法提供的 AST 级输入；该决策由[增量 AST 渲染器 Note](../architecture/2026-08-06-web-markdown-incremental-ast-renderer.md) 拥有。*
 
 **将 `MessageText` 替换为 Markdown 渲染。**这会产生格式化用户提示词与 steering 的副作用。在产品明确选择此行为之前，这两类输入内容仍按字面渲染。
 
@@ -42,4 +42,4 @@ assistant 生成的链接目标地址仅限绝对 HTTP、HTTPS 与 mailto URL。
 
 ## 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。URL 形态的行内代码会在不改变其可见字面文本的情况下变得可导航，而采用不安全 scheme 或混有其他内容的代码仍不可交互。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、KaTeX 与 shiki 允许列表；citation、anchor 和 thinking-small 表层仍暂缓。
+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，流式输出只重新解析不稳定的尾部；未完成的 Markdown 可能暂时改变尾部结构，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。URL 形态的行内代码会在不改变其可见字面文本的情况下变得可导航，而采用不安全 scheme 或混有其他内容的代码仍不可交互。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、KaTeX 与 shiki 允许列表；citation、anchor 和 thinking-small 表层仍暂缓。

.agents/notes/implemented/feature/2026-08-05-web-preview-product-badge.i18n.yaml

@@ -0,0 +1,6 @@
+# Bilingual-pair consistency record (docs/i18n/README.md): the git blob hash of each
+# side as of the last confirmed-consistent state. Both languages carry equal authority;
+# after editing either side, bring the other along and re-record with:
+#   pnpm run verify-translation-pairing --write .agents/notes/implemented/feature/2026-08-05-web-preview-product-badge.md
+2026-08-05-web-preview-product-badge.md: c20dedf8caa497d17a577cf46261c6a24c09a1ce
+2026-08-05-web-preview-product-badge.zh.md: c428dabf2a9a2ece2f90b0c3a527835a494c7cb7

.agents/notes/implemented/feature/2026-08-05-web-preview-product-badge.md

@@ -0,0 +1,33 @@
+# Agent Note: Web preview product badge
+
+Status: implemented
+
+English | [中文](2026-08-05-web-preview-product-badge.zh.md)
+
+## Problem
+
+The Web empty state does not identify the product as a preview. Users can enter the main session surface without seeing that the product is pre-release, while a deployment setting would misrepresent a product-wide lifecycle decision as an operator choice.
+
+## Decision
+
+The empty hero always renders a localized `Preview` / `预览版` badge beneath the headline. It has no configuration switch: preview status is one product identity shared by every deployment, not a deployment-varying tunable.
+
+The badge keeps the business-tertiary background so both themes retain the product-blue context, and uses the theme's primary label token for text. That pairing gives ordinary 12px text sufficient contrast in both light and dark themes; the business-primary foreground is reserved for larger or non-text accents because it does not reach the required contrast on this background.
+
+The badge leaves the product when the first tagged release removes the repository's pre-release stance, or when the owning product decision declares the preview phase complete. That change removes the badge and its locale key together rather than adding a runtime toggle.
+
+## Alternatives considered
+
+**Make preview status configurable.** Rejected because two deployments of the same pre-release product must not present different lifecycle identities, and a configuration field would turn product release state into an unsupported operator choice.
+
+**Use business-primary text on the business-tertiary background.** Rejected because the resulting light- and dark-theme contrast is below the 4.5:1 requirement for the badge's 12px text.
+
+**Hide the badge from the accessibility tree.** Rejected because preview status is product information rather than decoration; the accessible headline therefore includes the badge text.
+
+## Consequences
+
+Every new session exposes the same localized preview identity in visual and accessibility output. Removing preview status is an explicit product-release edit, and the badge favors readable neutral text over an all-blue treatment while retaining the business-tinted background.
+
+## Testing
+
+The conversation component test covers both localized badge values, and the Web lifecycle snapshots pin the English badge in the assembled empty hero.

.agents/notes/implemented/feature/2026-08-05-web-preview-product-badge.zh.md

@@ -0,0 +1,33 @@
+# Agent Note：Web 预览版产品徽标
+
+状态：已实现
+
+[English](2026-08-05-web-preview-product-badge.md) | 中文
+
+## 问题
+
+Web 空状态没有标明产品处于预览版阶段。用户可以在未看到产品尚未正式发布的情况下进入主会话界面；若改用部署设置，则会把面向整个产品的生命周期决策误表述为操作者的选择。
+
+## 决策
+
+空状态主视觉区始终在标题下方渲染本地化的 `Preview` / `预览版` 徽标。它没有配置开关：预览状态是所有部署共同的一项产品身份，而不是随部署变化的可调参数。
+
+徽标沿用 business-tertiary 背景，使两套主题都保留产品蓝的视觉语境；文字则使用主题的 primary label token。这一组合让普通 12px 文字在浅色与暗色主题下都有足够的对比度。business-primary 前景色仅留给较大字号文本或非文本强调元素，因为它在该背景上达不到要求的对比度。
+
+首个 tagged release 取消仓库的预发布立场时，或归属产品方明确决定预览阶段结束时，产品会移除该徽标。这一改动会同时移除徽标及其 locale key，而不是增加运行时开关。
+
+## 曾考虑的替代方案
+
+**让预览状态可配置。** 不予采纳：同一个预发布产品的两套部署不得展示不同的生命周期身份，配置字段还会把产品发布状态变成一项不受支持的操作者选择。
+
+**在 business-tertiary 背景上使用 business-primary 文字。** 不予采纳：由此产生的浅色与暗色主题对比度低于徽标 12px 文字所要求的 4.5:1。
+
+**在无障碍树中隐藏徽标。** 不予采纳：预览状态是产品信息而非装饰，因此无障碍标题会包含徽标文字。
+
+## 后果
+
+每个新会话都会在视觉与无障碍输出中呈现相同的本地化预览版身份。移除预览状态是一项显式的产品发布改动；徽标保留业务蓝色调背景，同时采用可读的中性色文字，而不是全蓝色处理方案。
+
+## 测试
+
+会话组件测试覆盖两个本地化徽标值，Web 生命周期快照则固定组装后空状态主视觉区中的英文徽标。