fix(llm): answer a catalog route's models from pi-ai's own registry

Clicking "fetch available models" on a built-in provider went to the
network. That is the wrong source: pi-ai's registry is the authoritative
list for its own providers, and it carries the context windows and output
caps a `GET /models` listing does not disclose. Asking api.deepseek.com
what DeepSeek serves is both slower and worse, and against an endpoint
that answers a different shape it failed outright.

Interrogation is still keyed by settings namespace — the provider being
added has no route — but the request may now name the route it is
editing. An adapter that already describes that route answers from what
it knows, needs no endpoint at all, and never touches the network; only a
route the catalog does not describe reaches the wire, and one naming no
endpoint is told to set one or enter its models by hand.

`ConfigurableProviderView` gained `supportsDiscovery` so a surface offers
the action where a namespace can answer instead of hardcoding an adapter
family.

Three narrower corrections ride along. Discovery no longer claims Azure
or Codex: Azure authenticates with an `api-key` header and an
`api-version` query despite its OpenAI lineage, and Codex uses OAuth, so
both reported an authentication failure as a provider with no models.
Cancellation during the body read escaped as the raw abort reason rather
than a coded ABORTED. And the schema comment claiming the probe key is
never logged overstated it: the host neither stores nor returns it, but
it rides the client's outgoing envelope like every other secret-bearing
payload, and redacting that tap is a configuration-plane-wide change.

packages/host/apiproxy/README.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 packages/host/apiproxy/README.md
-README.md: 633c8fe39d989802e8debc350279137b66979593
-README.zh.md: 5e5102840319900f6607acc17288882ccf3c0075
+README.md: 70d0ff258d5ed55678789ef6c7e6c8e64e822db3
+README.zh.md: 3259c03b3d3ca19658d20040024dc40c5c3f287a

packages/host/apiproxy/README.md

@@ -38,7 +38,7 @@ Directory picking delegates to the composed `ctx.directoryPicker` backend ([the
 
 The `command.*` and `skill.*` domains expose the host command registry and skill catalog to clients. `command.*` addresses an ordinary session's Agent and resumes a cold ordinary session when needed, while `skill.list` resolves the project root from the session header without touching the Agent registry. `skill.list` serves the browser's user-selected model-reference path, so it returns only skills that are both model-invocable and user-invocable; this domain has no direct skill-loading RPC. `command.execute` runs a slash-command line host-side with pure admission semantics: the response reports whether the line resolved to a handler plus the minted lifecycle `commandId` when it did (correlating the acknowledgment with the flow node), while the outcome rides the durably logged `command/run`/`command/done` lifecycle pair broadcast on the mux stream. Command handlers may legitimately outlast the 30-second transport health deadline, so `command.execute` carries only caller/connection cancellation; that signal cancels the running handler. `host/commands-changed` is the catalog invalidation frame: clients refetch `command.list` instead of diffing.
 
-The `settings.*`, `credentials.*`, and `llm.*` domains are the configuration-page wire. The settings domain serves the namespaces addressed by registered configurable providers (`ctx.llm.listConfigurableProviders()`) plus a small explicit allowlist — the Web preference `permission` and the product-owned `ui-onboarding`; adding a Settings registration alone never makes it remotely readable or writable. Any other namespace answers `settings-not-exposed` — the same answer an unregistered namespace gets, so no caller can enumerate the registry by probing. `settings.describe` returns each exposed namespace's serialized schemastery schema, redacted layered values (resolved/`base`/`user` — a field's presence in `user` marks it user-overridden), the `secrets` slot list, the section's `revision`, and the boolean `hasDocument` capability flag. The browser receives no Host path: pathless `settings.openDocument` asks the provider to materialize its document and then hands the Host-resolved result to the native opener, so no browser payload can select an arbitrary filesystem target. `settings.update`/`settings.replace` write the user layer; `settings.mutate` applies path ops (`set`/`unset`) against the section as stored, which is the removal path for a client holding the redacted view — rebuilding a section from it and replacing wholesale would delete the secrets the wire never returned. Any write may carry `expectedRevision`; a stale one answers `settings-conflict` with both revisions rather than overwriting the writer that landed first, and every other seam refusal folds into `settings-rejected`. Secret-role values never ride any response in any layer; a secret crosses the wire in exactly one direction — inside an `update`/`mutate` payload or `credentials.set`. `credentials.describe` returns value-free views (`configured`/`source`/`writable`), and `credentials.set`/`credentials.unset` map a shadowed-reference refusal onto `credential-rejected`. `llm.providers` merges the configurable-provider directory with live routes (dormant entries carry `active: false`; undeclared live routes append with no settings address) and `llm.models` is the session-independent catalog. `llm.discoverModels` interrogates a provider endpoint the page is still drafting: `settingsNs` selects the adapter family that knows how to read the listing, and the endpoint, protocol, and key come from the form rather than from storage. It writes nothing — the reply is candidates, and only a later `settings.mutate` decides what a route serves — so its `apiKey` is the third and last payload a secret may ride, alongside `settings.update`/`mutate` and `credentials.set`, and is never stored, logged, or echoed. Every refusal (an unserved namespace, a protocol with no readable listing, an unreachable endpoint, a rejected credential) 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. Three invalidation frames keep every surface converged without polling: `host/settings-changed {ns}` (`settings/document-updated` passthrough, so a raw change whose resolved value is unchanged still reaches clients), `host/credentials-changed {ref}` (reference names only, never values), and `host/models-changed` — fired by `llm/adapters-updated` and by a change to a configurable-provider namespace, whose settings carry that provider's catalog and endpoint; a `permission` or `ui-onboarding` change emits only its settings invalidation. The browser carrier restricts the whole configuration plane, reads and native actions included (`settings.describe`/`openDocument`/`update`/`replace`/`mutate`, `credentials.describe`/`set`/`unset`), to loopback same-origin requests — the `host.pickDirectory` privileged set. A composition without a settings or credential provider answers those domains with an actionable `internal` error naming the missing plugin.
+The `settings.*`, `credentials.*`, and `llm.*` domains are the configuration-page wire. The settings domain serves the namespaces addressed by registered configurable providers (`ctx.llm.listConfigurableProviders()`) plus a small explicit allowlist — the Web preference `permission` and the product-owned `ui-onboarding`; adding a Settings registration alone never makes it remotely readable or writable. Any other namespace answers `settings-not-exposed` — the same answer an unregistered namespace gets, so no caller can enumerate the registry by probing. `settings.describe` returns each exposed namespace's serialized schemastery schema, redacted layered values (resolved/`base`/`user` — a field's presence in `user` marks it user-overridden), the `secrets` slot list, the section's `revision`, and the boolean `hasDocument` capability flag. The browser receives no Host path: pathless `settings.openDocument` asks the provider to materialize its document and then hands the Host-resolved result to the native opener, so no browser payload can select an arbitrary filesystem target. `settings.update`/`settings.replace` write the user layer; `settings.mutate` applies path ops (`set`/`unset`) against the section as stored, which is the removal path for a client holding the redacted view — rebuilding a section from it and replacing wholesale would delete the secrets the wire never returned. Any write may carry `expectedRevision`; a stale one answers `settings-conflict` with both revisions rather than overwriting the writer that landed first, and every other seam refusal folds into `settings-rejected`. Secret-role values never ride any response in any layer; a secret crosses the wire in exactly one direction — inside an `update`/`mutate` payload or `credentials.set`. `credentials.describe` returns value-free views (`configured`/`source`/`writable`), and `credentials.set`/`credentials.unset` map a shadowed-reference refusal onto `credential-rejected`. `llm.providers` merges the configurable-provider directory with live routes (dormant entries carry `active: false`; undeclared live routes append with no settings address) and `llm.models` is the session-independent catalog. `llm.discoverModels` interrogates a provider endpoint the page is still drafting: `settingsNs` selects the adapter family that knows how to read the listing, and the endpoint, protocol, and key come from the form rather than from storage. It writes nothing — the reply is candidates, and only a later `settings.mutate` decides what a route serves — so its `apiKey` is the third payload on which a secret may ride, alongside `settings.update`/`mutate` and `credentials.set`. The host never stores or returns it; like the other two it does ride the client's outgoing envelope, which `subscribeEnvelopes()` observers can see, and redacting that tap is a configuration-plane-wide change rather than this method's to make alone. Every refusal (an unserved namespace, a protocol with no readable listing, an unreachable endpoint, a rejected credential) 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. Three invalidation frames keep every surface converged without polling: `host/settings-changed {ns}` (`settings/document-updated` passthrough, so a raw change whose resolved value is unchanged still reaches clients), `host/credentials-changed {ref}` (reference names only, never values), and `host/models-changed` — fired by `llm/adapters-updated` and by a change to a configurable-provider namespace, whose settings carry that provider's catalog and endpoint; a `permission` or `ui-onboarding` change emits only its settings invalidation. The browser carrier restricts the whole configuration plane, reads and native actions included (`settings.describe`/`openDocument`/`update`/`replace`/`mutate`, `credentials.describe`/`set`/`unset`), to loopback same-origin requests — the `host.pickDirectory` privileged set. A composition without a settings or credential provider answers those domains with an actionable `internal` error naming the missing plugin.
 
 The `subagent.*` domain addresses direct children by `{parentSessionId, childSessionId}`. `subagent.list` projects the complete durable one-shot and continuable catalog from `ctx.subagents.listChildren`, including each healthy row's origin-classified `hasChildren` hint, replaces corpus activity with the exact child Agent driver's running state, and includes an exact-live-parent hint; `subagent.history` verifies a healthy direct-child entry and reads its persisted log through `ctx.sessionQuery` without resuming an Agent. `subagent.prompt` accepts only continuable addresses, requires that exact live parent, delivers human content through `ctx.subagents.followup()` with the request `rpcId` as attribution, and returns the accepted inbox `messageId`. Typed errors preserve catalog diagnostics, parent availability, resumability, authorization, and not-delivered distinctions without exposing the model-hidden continuation descriptor. See the [Web subagent conversations Agent Note](../../../.agents/notes/implemented/feature/2026-07-27-web-subagent-conversations.md).
 

packages/host/apiproxy/README.zh.md

@@ -38,7 +38,7 @@ Workspace 列表与 Session 列表是相互独立的重连基线。`workspace.cr
 
 `command.*` 与 `skill.*` 领域向客户端暴露宿主命令注册表和 skill（技能）目录。`command.*` 寻址普通会话的 Agent，并在需要时恢复冷态普通会话；`skill.list` 则从会话头解析项目根目录，不触碰 Agent 注册表。`skill.list` 服务于浏览器中由用户选择的模型引用路径，因此仅返回模型和用户均可调用的 skill；该领域没有直接加载 skill 的 RPC。`command.execute` 在宿主侧运行一条斜杠命令行，语义为纯准入：响应报告该行是否解析到处理器，并在解析到时回带生成的生命周期 `commandId`（将本次确认与流节点关联）；结局经由持久落账并在 mux 流广播的 `command/run`/`command/done` 生命周期事件对承载。命令处理器运行超过 30 秒的传输健康时限仍属正常，因此 `command.execute` 仅携带调用方／连接取消信号；该信号可取消正在运行的处理器。`host/commands-changed` 是目录失效帧：客户端重新拉取 `command.list` 而不是做差分。
 
-`settings.*`、`credentials.*` 与 `llm.*` 领域是配置页协议。settings 领域服务于已注册可配置提供方所指向的 namespace（`ctx.llm.listConfigurableProviders()`），并额外服务于一份小型、显式的 allowlist——Web 偏好 `permission` 与产品持有的 `ui-onboarding`；仅新增一项 Settings 注册，绝不会使其可被远程读取或写入。其他任何 namespace 都只会得到 `settings-not-exposed`——未注册的 namespace 得到的是同一个答复，因此没有调用方能靠逐个探测把注册表枚举出来。`settings.describe` 为每个已暴露 namespace 提供其序列化 schemastery schema、脱敏后的分层值（resolved/`base`/`user`——字段出现在 `user` 中即标记其被用户覆盖）、`secrets` 槽位列表、该分节的 `revision`，以及布尔型 `hasDocument` 能力标志。浏览器不会收到 Host 路径：无路径参数的 `settings.openDocument` 会请求提供方准备文档，再把由 Host 解析出的结果交给原生打开器，因此任何浏览器载荷都无法选择任意文件系统目标。`settings.update`/`settings.replace` 写入用户层；`settings.mutate` 则在已存分节上施加路径 op（`set`/`unset`），这是持有脱敏视图的客户端的删除路径——据此重建分节再整体替换，会删掉协议从未回传过的那些机密。任何写入都可携带 `expectedRevision`；陈旧的期望值会以 `settings-conflict` 连同两个 revision 作答，而不是覆盖先落地的那个写方，其余每种 seam 拒绝则折叠为 `settings-rejected`。secret 角色的值绝不在任何一层搭乘任何响应；secret 只沿一个方向跨越协议——在 `update`/`mutate` 载荷或 `credentials.set` 之内。`credentials.describe` 返回不含值的视图（`configured`/`source`/`writable`），`credentials.set`/`credentials.unset` 则把被遮蔽引用的拒绝映射为 `credential-rejected`。`llm.providers` 把可配置提供方目录与存活路由合并（休眠条目携带 `active: false`；未声明的存活路由追加在后，不带 settings 地址），`llm.models` 则是与会话无关的目录。`llm.discoverModels` 询问页面尚在起草的提供方端点：`settingsNs` 选出懂得读取该列表的适配器家族，端点、协议与密钥则来自表单而非存储。它什么都不写——回复是候选，只有随后的 `settings.mutate` 才决定路由服务什么——因此其 `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、记录或回显。每一种拒绝（无人服务的 namespace、没有可读列表的协议、不可达端点、被拒凭据）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。三个失效帧让每个面无需轮询即保持收敛：`host/settings-changed {ns}`（`settings/document-updated` 透传，因此解析值未变的原始变更同样能到达客户端）、`host/credentials-changed {ref}`（只带引用名，绝不带值），以及 `host/models-changed`——它由 `llm/adapters-updated` 和可配置提供方 namespace 的变更触发，因为该提供方的设置正承载着它的目录与端点；`permission` 或 `ui-onboarding` 变更只会发出自身的 settings 失效通知。浏览器载体把整个配置面（含读取与原生操作：`settings.describe`/`openDocument`/`update`/`replace`/`mutate` 与 `credentials.describe`/`set`/`unset`）限制为仅接受来自回环地址的同源请求——即 `host.pickDirectory` 所在的特权集合。未装 settings 或凭据 provider 的组合会以指名缺失插件、包含解决建议的 `internal` 错误应答这些领域。
+`settings.*`、`credentials.*` 与 `llm.*` 领域是配置页协议。settings 领域服务于已注册可配置提供方所指向的 namespace（`ctx.llm.listConfigurableProviders()`），并额外服务于一份小型、显式的 allowlist——Web 偏好 `permission` 与产品持有的 `ui-onboarding`；仅新增一项 Settings 注册，绝不会使其可被远程读取或写入。其他任何 namespace 都只会得到 `settings-not-exposed`——未注册的 namespace 得到的是同一个答复，因此没有调用方能靠逐个探测把注册表枚举出来。`settings.describe` 为每个已暴露 namespace 提供其序列化 schemastery schema、脱敏后的分层值（resolved/`base`/`user`——字段出现在 `user` 中即标记其被用户覆盖）、`secrets` 槽位列表、该分节的 `revision`，以及布尔型 `hasDocument` 能力标志。浏览器不会收到 Host 路径：无路径参数的 `settings.openDocument` 会请求提供方准备文档，再把由 Host 解析出的结果交给原生打开器，因此任何浏览器载荷都无法选择任意文件系统目标。`settings.update`/`settings.replace` 写入用户层；`settings.mutate` 则在已存分节上施加路径 op（`set`/`unset`），这是持有脱敏视图的客户端的删除路径——据此重建分节再整体替换，会删掉协议从未回传过的那些机密。任何写入都可携带 `expectedRevision`；陈旧的期望值会以 `settings-conflict` 连同两个 revision 作答，而不是覆盖先落地的那个写方，其余每种 seam 拒绝则折叠为 `settings-rejected`。secret 角色的值绝不在任何一层搭乘任何响应；secret 只沿一个方向跨越协议——在 `update`/`mutate` 载荷或 `credentials.set` 之内。`credentials.describe` 返回不含值的视图（`configured`/`source`/`writable`），`credentials.set`/`credentials.unset` 则把被遮蔽引用的拒绝映射为 `credential-rejected`。`llm.providers` 把可配置提供方目录与存活路由合并（休眠条目携带 `active: false`；未声明的存活路由追加在后，不带 settings 地址），`llm.models` 则是与会话无关的目录。`llm.discoverModels` 询问页面尚在起草的提供方端点：`settingsNs` 选出懂得读取该列表的适配器家族，端点、协议与密钥则来自表单而非存储。它什么都不写——回复是候选，只有随后的 `settings.mutate` 才决定路由服务什么——因此其 `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。host 从不存储或回传它；与另两者一样，它确实会搭乘客户端的出站信封，`subscribeEnvelopes()` 的观察者能看到——为该 tap 做脱敏是整个配置面的改动，而非本方法一家的事。每一种拒绝（无人服务的 namespace、没有可读列表的协议、不可达端点、被拒凭据）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。三个失效帧让每个面无需轮询即保持收敛：`host/settings-changed {ns}`（`settings/document-updated` 透传，因此解析值未变的原始变更同样能到达客户端）、`host/credentials-changed {ref}`（只带引用名，绝不带值），以及 `host/models-changed`——它由 `llm/adapters-updated` 和可配置提供方 namespace 的变更触发，因为该提供方的设置正承载着它的目录与端点；`permission` 或 `ui-onboarding` 变更只会发出自身的 settings 失效通知。浏览器载体把整个配置面（含读取与原生操作：`settings.describe`/`openDocument`/`update`/`replace`/`mutate` 与 `credentials.describe`/`set`/`unset`）限制为仅接受来自回环地址的同源请求——即 `host.pickDirectory` 所在的特权集合。未装 settings 或凭据 provider 的组合会以指名缺失插件、包含解决建议的 `internal` 错误应答这些领域。
 
 `subagent.*` 领域通过 `{parentSessionId, childSessionId}` 寻址直接 child。`subagent.list` 从 `ctx.subagents.listChildren` 投影包含 one-shot 与可继续条目的完整持久化目录、每个健康行基于 origin 分类的 `hasChildren` 提示，并把语料活动状态替换为确切 child Agent driver 的运行状态，同时提供确切 parent 是否存活的提示；`subagent.history` 先验证健康的直接 child 条目，再通过 `ctx.sessionQuery` 读取其持久化日志，且不恢复 Agent。`subagent.prompt` 只接受可继续地址，要求该确切 parent 已存活，通过 `ctx.subagents.followup()` 投递用户内容，以请求 `rpcId` 作为来源信息，并返回已接纳消息的 inbox `messageId`。类型化错误保留目录诊断、parent 可用性、可恢复性、授权和未投递等区别，同时不暴露对模型隐藏的继续执行描述符。见 [Web subagent 对话 Agent Note](../../../.agents/notes/implemented/feature/2026-07-27-web-subagent-conversations.md)。
 

packages/host/apiproxy/src/api-proxy.ts

@@ -2563,12 +2563,14 @@ export function createApiProxy(ctx: Context, defaults: ApiProxyDefaults): ApiPro
         const active = new Set(registered.map(provider => provider.id))
         const directory = ctx.llm.listConfigurableProviders()
         const declared = new Set(directory.map(entry => entry.provider))
+        const discoverable = new Set(ctx.llm.listModelDiscoveryNamespaces())
         const views = directory.map(entry => ({
           provider: entry.provider,
           displayName: entry.displayName,
           settingsNs: entry.settingsNs,
           settingsPath: [...entry.settingsPath],
           active: active.has(entry.provider),
+          supportsDiscovery: discoverable.has(entry.settingsNs),
         }))
         // Routes registered without a directory declaration still appear —
         // they exist and serve models — just with no settings address.
@@ -2580,6 +2582,7 @@ export function createApiProxy(ctx: Context, defaults: ApiProxyDefaults): ApiPro
             settingsNs: '',
             settingsPath: [],
             active: true,
+            supportsDiscovery: false,
           })
         }
         return Promise.resolve(ok(request, { providers: views }))
@@ -2590,10 +2593,11 @@ export function createApiProxy(ctx: Context, defaults: ApiProxyDefaults): ApiPro
       },
 
       async discoverModels(request, signal) {
-        const { settingsNs, baseURL, api, apiKey } = request.payload
+        const { settingsNs, provider, baseURL, api, apiKey } = request.payload
         try {
           const models = await ctx.llm.discoverModels(settingsNs, {
-            baseURL,
+            ...provider === undefined ? {} : { provider },
+            ...baseURL === undefined ? {} : { baseURL },
             ...api === undefined ? {} : { api },
             ...apiKey === undefined ? {} : { apiKey },
             ...signal === undefined ? {} : { signal },
@@ -2607,7 +2611,7 @@ export function createApiProxy(ctx: Context, defaults: ApiProxyDefaults): ApiPro
           return err(request, {
             code: 'model-discovery-failed',
             message: error instanceof Error ? error.message : String(error),
-            details: { settingsNs, baseURL },
+            details: { settingsNs, ...baseURL === undefined ? {} : { baseURL } },
           })
         }
       },

packages/host/apiproxy/src/api/llm.schema.ts

@@ -16,6 +16,7 @@ export const configurableProviderViewSchema = z.object({
   settingsNs: z.string(),
   settingsPath: z.array(z.string()),
   active: z.boolean(),
+  supportsDiscovery: z.boolean(),
 }) satisfies z.ZodType<Wire<ConfigurableProviderView>>
 
 /** llm.providers request payload. */
@@ -46,11 +47,14 @@ export const discoveredModelViewSchema = z.object({
 /** llm.discoverModels request payload. */
 export const llmDiscoverModelsRequestSchema = z.object({
   settingsNs: z.string().min(1),
-  baseURL: z.string().min(1),
+  provider: z.string().min(1).optional(),
+  baseURL: z.string().min(1).optional(),
   api: z.string().min(1).optional(),
-  // Write-only: the host uses it for this one interrogation and never stores,
-  // logs, or returns it. Kept out of any redacted echo for the same reason
-  // `credentials.set` never reads a value back.
+  // Write-only at the host: used for this one interrogation, never stored and
+  // never returned. It does ride the client's outgoing envelope like every
+  // other secret-bearing payload (`credentials.set`, `settings.update`), which
+  // `subscribeEnvelopes()` observers can see — redacting that tap is a
+  // configuration-plane-wide change, not this method's to make alone.
   apiKey: z.string().min(1).optional(),
 }) satisfies z.ZodType<Wire<RequestPayload<'llm.discoverModels'>>>
 

packages/host/apiproxy/src/api/llm.ts

@@ -22,6 +22,12 @@ export interface ConfigurableProviderView {
   settingsPath: string[]
   /** Whether the route is currently registered (its models are requestable). */
   active: boolean
+  /**
+   * Whether `llm.discoverModels` can answer for this entry's namespace. A
+   * surface offers the action only where it works instead of naming an adapter
+   * family it would have to hardcode.
+   */
+  supportsDiscovery: boolean
 }
 
 /** Llm-domain unary methods (the map keys llm.* of RpcMethodMap). */
@@ -46,17 +52,22 @@ export interface LlmApi {
    * drafting, and return the models it advertises for the user to adopt.
    *
    * The payload is the draft, not a stored route: `settingsNs` selects the
-   * adapter family that knows how to read the listing, and the endpoint,
-   * protocol, and key come from the form. Nothing is written — the reply is
-   * candidates, and only a later `settings.mutate` decides what a route
-   * serves. `apiKey` is therefore accepted here but never stored, logged, or
-   * echoed back; a provider whose key is already stored omits it and the
-   * endpoint answers unauthenticated or refuses.
+   * adapter family that answers, and the rest comes from the form. `provider`
+   * names the route being edited when there is one — an adapter that already
+   * describes that route answers from its own registry, with better metadata
+   * and no network call, and needs no endpoint. A route it does not describe is
+   * asked over the wire, which is what `baseURL`, `api`, and `apiKey` are for.
+   *
+   * Nothing is written — the reply is candidates, and only a later
+   * `settings.mutate` decides what a route serves. `apiKey` is accepted here
+   * but never stored or returned; a provider whose key is already stored omits
+   * it and the endpoint answers unauthenticated or refuses.
    */
   discoverModels(
     request: RpcRequest<{
       settingsNs: string
-      baseURL: string
+      provider?: string
+      baseURL?: string
       api?: string
       apiKey?: string
     }>,

packages/host/apiproxy/src/api/rpc.schema.ts

@@ -55,7 +55,7 @@ export const rpcErrorSchema: z.ZodType<RpcError> = z.discriminatedUnion('code',
   z.object({ code: z.literal('settings-not-exposed'), message: z.string(), details: z.object({ ns: z.string() }) }),
   z.object({ code: z.literal('settings-conflict'), message: z.string(), details: z.object({ ns: z.string(), expected: z.number(), actual: z.number() }) }),
   z.object({ code: z.literal('credential-rejected'), message: z.string(), details: z.object({ ref: z.string() }) }),
-  z.object({ code: z.literal('model-discovery-failed'), message: z.string(), details: z.object({ settingsNs: z.string(), baseURL: z.string() }) }),
+  z.object({ code: z.literal('model-discovery-failed'), message: z.string(), details: z.object({ settingsNs: z.string(), baseURL: z.string().optional() }) }),
   z.object({ code: z.literal('title-invalid'), message: z.string(), details: z.object({ sessionId: z.string() }) }),
   z.object({ code: z.literal('fork-unavailable'), message: z.string(), details: z.object({ sessionId: z.string() }) }),
   z.object({ code: z.literal('subagent-parent-unavailable'), message: z.string(), details: z.object({ parentSessionId: z.string() }) }),

packages/host/apiproxy/src/api/rpc.ts

@@ -78,7 +78,7 @@ export interface RpcErrorDetailsMap {
    * it is what the form shows before falling back to hand-entry — and the
    * details name the endpoint asked, never the credential offered.
    */
-  'model-discovery-failed': { settingsNs: string; baseURL: string }
+  'model-discovery-failed': { settingsNs: string; baseURL?: string }
   'title-invalid': { sessionId: SessionId }
   'fork-unavailable': { sessionId: SessionId }
   'subagent-parent-unavailable': { parentSessionId: SessionId }

packages/host/apiproxy/tests/api-proxy-config.spec.ts

@@ -523,12 +523,17 @@ describe('llm domain', () => {
     ])
     ctx.llm.registerAdapter(['deepseek-official'], new CatalogAdapter('DeepSeek', ['deepseek-v4-flash']))
     ctx.llm.registerAdapter(['undeclared'], new CatalogAdapter('Undeclared', ['u-1']))
+    // Only one namespace can answer an interrogation, so the flag follows the
+    // entry's namespace rather than being assumed for every row.
+    ctx.llm.registerModelDiscovery('llm-pi-ai', () => Promise.resolve([]))
     const api = createApiProxy(ctx, DEFAULTS)
     const value = expectOk(await api.llm.providers(request({})))
     expect(value.providers).toEqual([
-      { provider: 'deepseek-official', displayName: 'DeepSeek', settingsNs: 'llm-deepseek', settingsPath: [], active: true },
-      { provider: 'openai', displayName: 'openai', settingsNs: 'llm-pi-ai', settingsPath: ['providers', 'openai'], active: false },
-      { provider: 'undeclared', displayName: 'Undeclared', settingsNs: '', settingsPath: [], active: true },
+      { provider: 'deepseek-official', displayName: 'DeepSeek', settingsNs: 'llm-deepseek', settingsPath: [], active: true, supportsDiscovery: false },
+      { provider: 'openai', displayName: 'openai', settingsNs: 'llm-pi-ai', settingsPath: ['providers', 'openai'], active: false, supportsDiscovery: true },
+      // An undeclared live route has no settings address, so nothing can be
+      // interrogated on its behalf either.
+      { provider: 'undeclared', displayName: 'Undeclared', settingsNs: '', settingsPath: [], active: true, supportsDiscovery: false },
     ])
   })
 
@@ -596,6 +601,25 @@ describe('llm.discoverModels', () => {
       .not.toContain('llm-pi-ai')
   })
 
+  it('carries the route being edited so an adapter can answer from its own registry', async () => {
+    const ctx = await harness()
+    let probe: unknown
+    ctx.llm.registerModelDiscovery('llm-pi-ai', (request_) => {
+      probe = request_
+      return Promise.resolve([{ id: 'from-registry', contextWindow: 65_536, maxTokens: 4096 }])
+    })
+    const api = createApiProxy(ctx, DEFAULTS)
+
+    const value = expectOk(await api.llm.discoverModels(request({
+      settingsNs: 'llm-pi-ai',
+      provider: 'deepseek',
+    })))
+
+    // No endpoint at all: a route the adapter already describes needs none.
+    expect(probe).toEqual({ provider: 'deepseek' })
+    expect(value.models).toEqual([{ id: 'from-registry', contextWindow: 65_536, maxTokens: 4096 }])
+  })
+
   it('omits a credential and protocol the draft does not name', async () => {
     const ctx = await harness()
     let probe: unknown

packages/host/apiproxy/tests/client-handler.spec.ts

@@ -677,6 +677,7 @@ describe('config unary surface', () => {
       settingsNs: 'llm-pi-ai',
       settingsPath: ['providers', 'openai'],
       active: false,
+      supportsDiscovery: true,
     }
     const group = { id: 'deepseek-official', name: 'DeepSeek', models: [{ id: 'deepseek-v4-flash', name: 'Flash' }] }
     const api = scriptedApi({