diff --git a/packages/core/src/plugin/provider/azure.ts b/packages/core/src/plugin/provider/azure.ts
index ceb07611952..40cdeb01516 100644
--- a/packages/core/src/plugin/provider/azure.ts
+++ b/packages/core/src/plugin/provider/azure.ts
@@ -1,4 +1,5 @@
-import { Clock, Effect, Schema, Semaphore, Stream } from "effect"
+import { Clock, Effect, FiberHandle, Option, Schema, Semaphore, Stream } from "effect"
+import { HttpClient, HttpClientRequest, HttpClientResponse } from "effect/unstable/http"
 import { ChildProcess } from "effect/unstable/process"
 import { define } from "@opencode/plugin/effect/plugin"
 import { Form } from "@opencode/schema/form"
@@ -7,14 +8,22 @@ import { App } from "../../app.js"
 import { Bus } from "../../bus.js"
 import { Credential } from "../../credential.js"
 import { Integration } from "../../integration.js"
+import { IntegrationConnection } from "../../integration/connection.js"
+import { Model } from "../../model.js"
 import { Provider } from "../../provider.js"
 import { iife } from "../../util/iife.js"
 import { which } from "../../util/which.js"
+import type { PluginInternal } from "../internal.js"
 import { configuredSettings } from "./configured.js"
 
 const cognitiveScope = "https://cognitiveservices.azure.com/.default"
 const foundryScope = "https://ai.azure.com/.default"
+const managementScope = "https://management.azure.com/.default"
 const methodID = Integration.MethodID.make("azure-cli")
+// A resource name becomes a hostname label and a query literal, so anything else never leaves the process. Azure
+// allows only letters, digits, and hyphens in it.
+// https://learn.microsoft.com/azure/ai-services/cognitive-services-custom-subdomains
+const resourcePattern = /^[a-zA-Z0-9][a-zA-Z0-9-]*$/
 const decodeJSON = Schema.decodeUnknownEffect(Schema.fromJsonString(Schema.Unknown))
 const decodeToken = Schema.decodeUnknownEffect(
   Schema.Struct({
@@ -23,209 +32,443 @@ const decodeToken = Schema.decodeUnknownEffect(
     expiresOn: Schema.optional(Schema.NonEmptyString),
   }),
 )
-export const AzurePlugin = define({
-  id: "opencode.provider.azure",
-  effect: Effect.fn(function* (ctx) {
-    const configured = yield* configuredSettings(Provider.ID.azure)
-    const processes = yield* AppProcess.Service
-    const bus = yield* Bus.Service
-    const tokens = new Map<string, { access: string; expires: number }>()
-    const loading = Semaphore.makeUnsafe(1)
-    const loaded: { resource?: string } = {}
+const ResourceDeployments = Schema.Struct({ data: Schema.Array(Schema.Unknown) })
+const decodeResourceDeployment = Schema.decodeUnknownOption(
+  Schema.Struct({ id: Schema.NonEmptyString, model: Schema.NonEmptyString, status: Schema.String }),
+)
+const ManagementDeployments = Schema.Struct({
+  value: Schema.Array(Schema.Unknown),
+  nextLink: Schema.optional(Schema.NonEmptyString),
+})
+const decodeManagementDeployment = Schema.decodeUnknownOption(
+  Schema.Struct({
+    name: Schema.NonEmptyString,
+    properties: Schema.Struct({
+      model: Schema.Struct({ name: Schema.NonEmptyString }),
+      provisioningState: Schema.String,
+    }),
+  }),
+)
+const ResourceQuery = Schema.Struct({ query: Schema.String })
+const Resources = Schema.Struct({ data: Schema.Array(Schema.Struct({ id: Schema.NonEmptyString })) })
 
-    const command = (args: string[]) =>
-      processes
-        .run(ChildProcess.make("az", args, { extendEnv: true, stdin: "ignore" }), { timeout: "10 seconds" })
-        .pipe(
-          Effect.flatMap(AppProcess.requireSuccess),
-          Effect.flatMap((result) => decodeJSON(result.stdout.toString("utf8"))),
+type Deployment = { readonly name: string; readonly model: string }
+
+export function make(
+  endpoints = {
+    resource: (name: string) => `https://${name}.openai.azure.com/openai`,
+    management: "https://management.azure.com",
+  },
+) {
+  return define({
+    id: "opencode.provider.azure",
+    effect: Effect.fn(function* (ctx) {
+      const configured = yield* configuredSettings(Provider.ID.azure)
+      const processes = yield* AppProcess.Service
+      const bus = yield* Bus.Service
+      const credentials = yield* Credential.Service
+      const providers = yield* Provider.Service
+      const http = HttpClient.filterStatusOk(yield* HttpClient.HttpClient)
+      const tokens = new Map<string, { access: string; expires: number }>()
+      // A resource keeps its Azure Resource Manager ID until it is deleted, so discovery looks it up once.
+      const resourceIDs = new Map<string, string>()
+      const loading = Semaphore.makeUnsafe(1)
+      const discovery = yield* FiberHandle.make<void, never>()
+      const loaded: {
+        resource?: string
+        url?: string
+        deployments?: readonly Deployment[]
+        connection?: Effect.Success<ReturnType<typeof ctx.integration.connection.active>>
+      } = {}
+
+      const command = (args: string[]) =>
+        processes
+          .run(ChildProcess.make("az", args, { extendEnv: true, stdin: "ignore" }), { timeout: "10 seconds" })
+          .pipe(
+            Effect.flatMap(AppProcess.requireSuccess),
+            Effect.flatMap((result) => decodeJSON(result.stdout.toString("utf8"))),
+          )
+
+      const token = Effect.fn("AzurePlugin.token")(function* (scope: string) {
+        const now = yield* Clock.currentTimeMillis
+        const cached = tokens.get(scope)
+        if (cached && cached.expires - now > 5 * 60_000) return cached
+        const result = yield* command(["account", "get-access-token", "--scope", scope, "--output", "json"]).pipe(
+          Effect.flatMap(decodeToken),
         )
-
-    const token = Effect.fn("AzurePlugin.token")(function* (scope: string) {
-      const now = yield* Clock.currentTimeMillis
-      const cached = tokens.get(scope)
-      if (cached && cached.expires - now > 5 * 60_000) return cached
-      const result = yield* command(["account", "get-access-token", "--scope", scope, "--output", "json"]).pipe(
-        Effect.flatMap(decodeToken),
-      )
-      const expires = result.expires_on !== undefined ? result.expires_on * 1000 : Date.parse(result.expiresOn ?? "")
-      if (!Number.isFinite(expires))
-        return yield* Effect.fail(new Error("Azure CLI returned an invalid token expiration"))
-      const refreshed = { access: result.accessToken, expires }
-      tokens.set(scope, refreshed)
-      return refreshed
-    })
-
-    const available = Boolean(which("az"))
-    const form = () =>
-      iife(() => {
-        if (resolveResourceName(configured) || typeof configured?.baseURL === "string") return
-        return Form.Fields.make([
-          {
-            type: "string",
-            key: "resourceName",
-            title: "Enter Azure Resource Name",
-            placeholder: "e.g. my-models",
-            required: true,
-          },
-        ])
+        const expires = result.expires_on !== undefined ? result.expires_on * 1000 : Date.parse(result.expiresOn ?? "")
+        if (!Number.isFinite(expires))
+          return yield* Effect.fail(new Error("Azure CLI returned an invalid token expiration"))
+        const refreshed = { access: result.accessToken, expires }
+        tokens.set(scope, refreshed)
+        return refreshed
       })
 
-    yield* ctx.integration.transform((editor) => {
-      editor.method.update({
-        integrationID: Provider.ID.azure,
-        method: { type: "key", label: "API key", form: form() },
-      })
-      if (!available) return
-      editor.method.update({
-        integrationID: Provider.ID.azure,
-        method: {
-          id: methodID,
-          type: "oauth",
-          label: "Microsoft Entra ID (Azure CLI)",
-          form: form(),
-        },
-        authorize: (answer) =>
-          Effect.succeed({
-            mode: "auto" as const,
-            url: "",
-            instructions: "Sign in with `az login` before continuing.",
-            callback: Effect.gen(function* () {
-              const resourceName =
-                typeof answer.resourceName === "string" ? answer.resourceName : resolveResourceName(configured)
-              if (!resourceName) return yield* Effect.fail(new Error("Azure resource name is required"))
-              const current = yield* token(cognitiveScope)
-              loaded.resource = resourceName
-              yield* ctx.provider.reload()
-              return Credential.OAuth.make({
-                type: "oauth",
-                methodID,
-                access: current.access,
-                refresh: "azure-cli",
-                expires: current.expires,
-                metadata: { resourceName },
-              })
-            }),
-          }),
-        refresh: (credential) =>
-          token(cognitiveScope).pipe(
-            Effect.map((current) =>
-              Credential.OAuth.make({ ...credential, access: current.access, expires: current.expires }),
+      const management = (request: HttpClientRequest.HttpClientRequest) =>
+        token(managementScope).pipe(
+          Effect.flatMap((current) =>
+            http.execute(
+              request.pipe(
+                HttpClientRequest.bearerToken(current.access),
+                HttpClientRequest.acceptJson,
+                HttpClientRequest.setHeader("User-Agent", App.useragent(ctx.app)),
+              ),
             ),
           ),
-      })
-    })
-
-    const load = Effect.fn("AzurePlugin.load")(function* () {
-      const connection = yield* ctx.integration.connection.active(Provider.ID.azure)
-      const credential = connection
-        ? yield* ctx.integration.connection.resolve(connection).pipe(Effect.orElseSucceed(() => undefined))
-        : undefined
-      if (credential?.type !== "oauth" || credential.methodID !== methodID) {
-        loaded.resource = undefined
-        return
-      }
-      const resource =
-        typeof credential.metadata?.resourceName === "string" ? credential.metadata.resourceName : undefined
-      loaded.resource = resource
-    })
-
-    yield* load()
-    yield* ctx.provider.transform((evt) => {
-      for (const item of evt.list()) {
-        if (
-          item.provider.id !== Provider.ID.azure &&
-          !item.provider.package.startsWith("@opencode/ai/providers/azure/")
         )
-          continue
-        const resourceName = resolveResourceName(item.provider.settings, loaded.resource)
-        const websocket = responsesWebSocketCapable(item.provider)
-        if (!resourceName && !websocket) continue
-        evt.update(item.provider.id, (provider) => {
-          provider.settings = {
-            ...provider.settings,
-            ...(resourceName === undefined ? {} : { resourceName }),
-            ...(websocket ? { transport: provider.settings?.transport ?? "websocket" } : {}),
-            ...(resourceName !== undefined && typeof provider.settings?.baseURL === "string"
-              ? { baseURL: expandResourceName(provider.settings.baseURL, resourceName) }
-              : {}),
-          }
+
+      const available = Boolean(which("az"))
+      const form = () =>
+        iife(() => {
+          if (resolveResourceName(configured) || typeof configured?.baseURL === "string") return
+          return Form.Fields.make([
+            {
+              type: "string",
+              key: "resourceName",
+              title: "Enter Azure Resource Name",
+              placeholder: "e.g. my-models",
+              required: true,
+            },
+          ])
         })
-      }
-    })
-    yield* ctx.model.transform((models) => {
-      for (const item of models.provider.list()) {
-        if (
-          item.provider.id !== Provider.ID.azure &&
-          !item.provider.package.startsWith("@opencode/ai/providers/azure/")
+
+      yield* ctx.integration.transform((editor) => {
+        editor.method.update({
+          integrationID: Provider.ID.azure,
+          method: { type: "key", label: "API key", form: form() },
+        })
+        if (!available) return
+        editor.method.update({
+          integrationID: Provider.ID.azure,
+          method: {
+            id: methodID,
+            type: "oauth",
+            label: "Microsoft Entra ID (Azure CLI)",
+            form: form(),
+          },
+          authorize: (answer) =>
+            Effect.succeed({
+              mode: "auto" as const,
+              url: "",
+              instructions: "Sign in with `az login` before continuing.",
+              callback: Effect.gen(function* () {
+                const resourceName =
+                  (typeof answer.resourceName === "string" ? answer.resourceName.trim() : "") ||
+                  resolveResourceName(configured)
+                if (!resourceName) return yield* Effect.fail(new Error("Azure resource name is required"))
+                const current = yield* token(cognitiveScope)
+                return Credential.OAuth.make({
+                  type: "oauth",
+                  methodID,
+                  access: current.access,
+                  refresh: "azure-cli",
+                  expires: current.expires,
+                  metadata: { resourceName },
+                })
+              }),
+            }),
+          refresh: (credential) =>
+            token(cognitiveScope).pipe(
+              Effect.map((current) =>
+                Credential.OAuth.make({ ...credential, access: current.access, expires: current.expires }),
+              ),
+            ),
+        })
+      })
+
+      const load = Effect.fn("AzurePlugin.load")(function* () {
+        const connection = yield* ctx.integration.connection.active(Provider.ID.azure)
+        // Startup awaits this, so it reads the stored credential: resolving the connection would refresh an
+        // expired token through the Azure CLI.
+        const stored =
+          connection?.type === "credential" ? yield* credentials.get(Credential.ID.make(connection.id)) : undefined
+        return { connection, resource: credentialResource(stored?.value) }
+      })
+
+      // Resource Graph searches every subscription the Azure CLI account can read, not only the selected one.
+      // https://learn.microsoft.com/rest/api/azureresourcegraph/resourcegraph/resources/resources
+      const findResource = Effect.fn("AzurePlugin.findResource")(function* (resource: string) {
+        const response = yield* HttpClientRequest.post(
+          `${endpoints.management}/providers/Microsoft.ResourceGraph/resources?api-version=2022-10-01`,
+        ).pipe(
+          HttpClientRequest.schemaBodyJson(ResourceQuery)({ query: resourceQuery(resource) }),
+          Effect.flatMap(management),
+          Effect.flatMap(HttpClientResponse.schemaBodyJson(Resources)),
+          Effect.timeout("10 seconds"),
         )
-          continue
-        const resourceName = resolveResourceName(item.provider.settings, loaded.resource)
-        for (const model of models.list(item.provider.id)) {
-          models.update(item.provider.id, model.id, (draft) => {
-            if (resourceName && typeof draft.settings?.baseURL === "string")
-              draft.settings.baseURL = expandResourceName(
-                draft.settings.baseURL,
-                resolveResourceName(draft.settings, resourceName) ?? resourceName,
-              )
+        const id = response.data[0]?.id
+        if (!id) return yield* Effect.fail(new Error(`Azure resource "${resource}" was not found`))
+        return id
+      })
+
+      const managementDeployments = Effect.fn("AzurePlugin.managementDeployments")(function* (resource: string) {
+        const key = resource.toLowerCase()
+        const id = resourceIDs.get(key) ?? (yield* findResource(resource))
+        resourceIDs.set(key, id)
+        const origin = new URL(endpoints.management).origin
+        return yield* Stream.paginate(`${endpoints.management}${id}/deployments?api-version=2024-10-01`, (url) =>
+          management(HttpClientRequest.get(url)).pipe(
+            Effect.flatMap(HttpClientResponse.schemaBodyJson(ManagementDeployments)),
+            Effect.timeout("10 seconds"),
+            Effect.flatMap((response) =>
+              // Every page carries the management token, so a page link must stay on the management endpoint.
+              // https://learn.microsoft.com/rest/api/aiservices/accountmanagement/deployments/list
+              response.nextLink !== undefined && URL.parse(response.nextLink)?.origin !== origin
+                ? Effect.fail(new Error("Azure returned a deployment page outside the management endpoint"))
+                : Effect.succeed([
+                    response.value.flatMap((raw): Deployment[] => {
+                      const item = Option.getOrUndefined(decodeManagementDeployment(raw))
+                      return item?.properties.provisioningState === "Succeeded"
+                        ? [{ name: item.name, model: item.properties.model.name }]
+                        : []
+                    }),
+                    Option.fromNullishOr(response.nextLink),
+                  ] as const),
+            ),
+          ),
+        ).pipe(
+          Stream.runCollect,
+          // A moved or recreated resource has a new ID, so the next discovery looks it up again.
+          Effect.tapError(() => Effect.sync(() => resourceIDs.delete(key))),
+        )
+      })
+
+      const resourceDeployments = Effect.fn("AzurePlugin.resourceDeployments")(function* (
+        url: string,
+        credential: Credential.Value,
+      ) {
+        return yield* http
+          .execute(
+            HttpClientRequest.get(url).pipe(
+              HttpClientRequest.acceptJson,
+              HttpClientRequest.setHeader("User-Agent", App.useragent(ctx.app)),
+              credential.type === "key"
+                ? HttpClientRequest.setHeader("api-key", credential.key)
+                : HttpClientRequest.bearerToken(credential.access),
+            ),
+          )
+          .pipe(
+            Effect.flatMap(HttpClientResponse.schemaBodyJson(ResourceDeployments)),
+            Effect.timeout("10 seconds"),
+            Effect.map((response) =>
+              response.data.flatMap((raw): Deployment[] => {
+                const item = Option.getOrUndefined(decodeResourceDeployment(raw))
+                return item?.status === "succeeded" ? [{ name: item.id, model: item.model }] : []
+              }),
+            ),
+          )
+      })
+
+      // Azure documents the management API as the deployment inventory, but only an Azure CLI session can reach it:
+      // Azure Resource Manager accepts Entra ID tokens, never resource keys.
+      // https://learn.microsoft.com/rest/api/aiservices/accountmanagement/deployments/list
+      // The resource's own inventory serves API keys and identities without Azure Resource Manager read access. Only
+      // data-plane version 2022-12-01 has it; later versions dropped `/deployments` and keep `/models`, which lists
+      // models the resource can deploy rather than its deployments.
+      // https://github.com/Azure/azure-rest-api-specs/blob/main/specification/cognitiveservices/data-plane/OpenAIAuthoring/stable/2022-12-01/azureopenai.json
+      const deployments = (url: string, resource: string, credential: Credential.Value) =>
+        credential.type === "oauth"
+          ? managementDeployments(resource).pipe(Effect.catch(() => resourceDeployments(url, credential)))
+          : resourceDeployments(url, credential)
+
+      // Local and quick, so a switch rebinds the provider before discovery for the new connection calls Azure.
+      const rebind = () =>
+        loading.withPermit(
+          Effect.gen(function* () {
+            const current = yield* load()
+            if (
+              IntegrationConnection.key(current.connection) === IntegrationConnection.key(loaded.connection) &&
+              current.resource === loaded.resource
+            )
+              return
+            Object.assign(loaded, current, { url: undefined, deployments: undefined })
+            yield* ctx.provider.reload()
+          }),
+        )
+
+      const discover = Effect.fn("AzurePlugin.discover")(function* () {
+        const connection = loaded.connection
+        const settings = (yield* providers.get(Provider.ID.azure))?.settings
+        const name = loaded.resource ?? resolveResourceName(settings)
+        // A custom endpoint may expose other deployments than the resource does, so it keeps the catalog.
+        const url =
+          connection && name !== undefined && resourcePattern.test(name) && typeof settings?.baseURL !== "string"
+            ? `${endpoints.resource(name)}/deployments?api-version=2022-12-01`
+            : undefined
+        if (loaded.connection !== connection) return
+        // Keep the last inventory through transient failures only for the same connection and resource.
+        if (loaded.url !== url) {
+          loaded.url = url
+          if (loaded.deployments) {
+            loaded.deployments = undefined
+            yield* ctx.model.reload()
+          }
+        }
+        if (!connection || !name || !url) return
+        const credential = yield* ctx.integration.connection
+          .resolve(connection)
+          .pipe(Effect.orElseSucceed(() => undefined))
+        if (!credential || (credential.type === "oauth" && credential.methodID !== methodID)) return
+        const found = yield* deployments(url, name, credential).pipe(
+          // Azure promises no order; normalize it so a reordered response does not rebuild the model list.
+          Effect.map((list) => list.toSorted((a, b) => a.name.localeCompare(b.name))),
+          Effect.catch((cause) =>
+            Effect.logWarning("failed to sync Azure deployments", { cause }).pipe(Effect.as(undefined)),
+          ),
+        )
+        if (!found) return
+        if (
+          loaded.connection !== connection ||
+          IntegrationConnection.key(connection) !==
+            IntegrationConnection.key(yield* ctx.integration.connection.active(Provider.ID.azure))
+        )
+          return
+        if (JSON.stringify(found) === JSON.stringify(loaded.deployments)) return
+        const catalog = new Map(
+          Array.from((yield* providers.snapshot()).records.get(Provider.ID.azure)?.models.keys() ?? [], (id) => [
+            id.toLowerCase(),
+            id,
+          ]),
+        )
+        const unmatched = found.filter((deployment) => !catalogModel(catalog, deployment))
+        if (unmatched.length > 0)
+          yield* Effect.logWarning("Azure deployments of models outside the catalog need explicit configuration", {
+            deployments: unmatched.map((deployment) => deployment.name),
+          })
+        loaded.deployments = found
+        yield* ctx.model.reload()
+      })
+
+      const refresh = () => rebind().pipe(Effect.andThen(FiberHandle.run(discovery, discover())))
+
+      // The connection's resource wins for Azure itself, matching the runtime merge of credentials over settings.
+      const resourceFor = (provider: Provider.Info) =>
+        provider.id === Provider.ID.azure
+          ? (loaded.resource ?? resolveResourceName(provider.settings))
+          : resolveResourceName(provider.settings, loaded.resource)
+
+      Object.assign(loaded, yield* load())
+      yield* ctx.provider.transform((evt) => {
+        for (const item of evt.list()) {
+          if (
+            item.provider.id !== Provider.ID.azure &&
+            !item.provider.package.startsWith("@opencode/ai/providers/azure/")
+          )
+            continue
+          const resourceName = resourceFor(item.provider)
+          const websocket = responsesWebSocketCapable(item.provider)
+          if (!resourceName && !websocket) continue
+          evt.update(item.provider.id, (provider) => {
+            provider.settings = {
+              ...provider.settings,
+              ...(resourceName === undefined ? {} : { resourceName }),
+              ...(websocket ? { transport: provider.settings?.transport ?? "websocket" } : {}),
+              ...(resourceName !== undefined && typeof provider.settings?.baseURL === "string"
+                ? { baseURL: expandResourceName(provider.settings.baseURL, resourceName) }
+                : {}),
+            }
           })
         }
-      }
-    })
+        const item = evt.get(Provider.ID.azure)
+        if (!item) return
+        // Bind resource settings and discovery to their account, so a switch hides them until the rebind.
+        // Keep the full templates here for explicit configuration; the model transform narrows the visible list.
+        evt.add({
+          info: item.provider,
+          models: Array.from(item.models.values()),
+          sourceConnection: loaded.connection,
+        })
+      })
+      yield* ctx.model.transform((models) => {
+        for (const item of models.provider.list()) {
+          if (
+            item.provider.id !== Provider.ID.azure &&
+            !item.provider.package.startsWith("@opencode/ai/providers/azure/")
+          )
+            continue
+          const resourceName = resourceFor(item.provider)
+          for (const model of models.list(item.provider.id)) {
+            models.update(item.provider.id, model.id, (draft) => {
+              if (resourceName && typeof draft.settings?.baseURL === "string")
+                draft.settings.baseURL = expandResourceName(
+                  draft.settings.baseURL,
+                  resolveResourceName(draft.settings, resourceName) ?? resourceName,
+                )
+            })
+          }
+        }
+        if (!loaded.deployments) return
+        // Narrowing here rather than in the provider catalog keeps every catalog model available as the base
+        // of a model the user configures explicitly; those are applied after this transform.
+        const catalog = models.list(Provider.ID.azure)
+        const deployed = deployedModels(loaded.deployments, catalog)
+        for (const model of catalog) {
+          if (!deployed.has(model.id)) models.remove(Provider.ID.azure, model.id)
+        }
+        for (const [id, model] of deployed) {
+          models.update(Provider.ID.azure, id, (draft) => Object.assign(draft, model))
+        }
+      })
 
-    const reload = () => loading.withPermit(load().pipe(Effect.andThen(ctx.provider.reload())))
-    yield* bus.subscribe(Credential.Event.Switched).pipe(
-      Stream.filter((event) => event.data.integrationID === Integration.ID.make("azure")),
-      Stream.runForEach(reload),
-      Effect.forkScoped({ startImmediately: true }),
-    )
+      // A switch interrupts discovery for the previous connection instead of waiting for its Azure calls.
+      yield* bus.subscribe(Credential.Event.Switched).pipe(
+        Stream.filter((event) => event.data.integrationID === Integration.ID.make("azure")),
+        Stream.runForEach(() => refresh()),
+        Effect.forkScoped({ startImmediately: true }),
+      )
+      // Deployments load in the background so startup never waits on Azure; the catalog serves until they arrive.
+      // Later changes load when the connection changes, so a new deployment needs a reconnect or restart.
+      yield* refresh().pipe(Effect.forkScoped)
 
-    // Entra bearer tokens are minted per request from the target URL's scope, so they are injected
-    // at the transport hooks rather than stored as a credential.
-    const bearer = Effect.fn("AzurePlugin.bearer")(function* (url: string) {
-      const connection = yield* ctx.integration.connection.active(Provider.ID.azure)
-      const credential = connection
-        ? yield* ctx.integration.connection.resolve(connection).pipe(Effect.orElseSucceed(() => undefined))
-        : undefined
-      if (credential?.type !== "oauth" || credential.methodID !== methodID) return
-      const target = new URL(url)
-      const scope =
-        target.hostname.endsWith(".services.ai.azure.com") && !target.pathname.startsWith("/models")
-          ? foundryScope
-          : cognitiveScope
-      const current = yield* token(scope).pipe(Effect.orDie)
-      return `Bearer ${current.access}`
-    })
-    yield* ctx.session.hook(
-      "http.request",
-      (evt) =>
-        Effect.gen(function* () {
-          if (evt.model.providerID !== Provider.ID.azure) return
-          const authorization = yield* bearer(evt.request.url)
-          if (!authorization) return
-          evt.request.headers.delete("api-key")
-          evt.request.headers.delete("x-api-key")
-          evt.request.headers.set("authorization", authorization)
-          evt.request.headers.set("user-agent", App.useragent(ctx.app))
-        }),
-      { providerID: Provider.ID.azure },
-    )
-    yield* ctx.session.hook(
-      "experimental.ws.handshake",
-      (evt) =>
-        Effect.gen(function* () {
-          if (evt.model.providerID !== Provider.ID.azure) return
-          const authorization = yield* bearer(evt.url)
-          if (!authorization) return
-          delete evt.headers["api-key"]
-          delete evt.headers["x-api-key"]
-          evt.headers.authorization = authorization
-          evt.headers["user-agent"] = App.useragent(ctx.app)
-        }),
-      { providerID: Provider.ID.azure },
-    )
-  }),
-})
+      // Entra bearer tokens are minted per request from the target URL's scope, so they are injected
+      // at the transport hooks rather than stored as a credential.
+      const bearer = Effect.fn("AzurePlugin.bearer")(function* (url: string) {
+        const connection = yield* ctx.integration.connection.active(Provider.ID.azure)
+        const credential = connection
+          ? yield* ctx.integration.connection.resolve(connection).pipe(Effect.orElseSucceed(() => undefined))
+          : undefined
+        if (credential?.type !== "oauth" || credential.methodID !== methodID) return
+        const target = new URL(url)
+        const scope =
+          target.hostname.endsWith(".services.ai.azure.com") && !target.pathname.startsWith("/models")
+            ? foundryScope
+            : cognitiveScope
+        const current = yield* token(scope).pipe(Effect.orDie)
+        return `Bearer ${current.access}`
+      })
+      yield* ctx.session.hook(
+        "http.request",
+        (evt) =>
+          Effect.gen(function* () {
+            if (evt.model.providerID !== Provider.ID.azure) return
+            const authorization = yield* bearer(evt.request.url)
+            if (!authorization) return
+            evt.request.headers.delete("api-key")
+            evt.request.headers.delete("x-api-key")
+            evt.request.headers.set("authorization", authorization)
+            evt.request.headers.set("user-agent", App.useragent(ctx.app))
+          }),
+        { providerID: Provider.ID.azure },
+      )
+      yield* ctx.session.hook(
+        "experimental.ws.handshake",
+        (evt) =>
+          Effect.gen(function* () {
+            if (evt.model.providerID !== Provider.ID.azure) return
+            const authorization = yield* bearer(evt.url)
+            if (!authorization) return
+            delete evt.headers["api-key"]
+            delete evt.headers["x-api-key"]
+            evt.headers.authorization = authorization
+            evt.headers["user-agent"] = App.useragent(ctx.app)
+          }),
+        { providerID: Provider.ID.azure },
+      )
+    }),
+  } satisfies PluginInternal.InternalPlugin)
+}
+
+export const AzurePlugin = make()
 
 function resolveResourceName(settings: Readonly<Record<string, unknown>> | undefined, fallback?: string) {
   const configured = settings?.resourceName
@@ -239,6 +482,60 @@ function expandResourceName(baseURL: string, resourceName: string) {
     .replaceAll("${AZURE_COGNITIVE_SERVICES_RESOURCE_NAME}", resourceName)
 }
 
+// The Azure CLI method stores the resource as credential metadata, the API key method as its form answer.
+function credentialResource(credential: Credential.Value | undefined) {
+  const resource =
+    credential?.type === "key"
+      ? credential.configuration?.resourceName
+      : credential?.methodID === methodID
+        ? credential.metadata?.resourceName
+        : undefined
+  return typeof resource === "string" && resource.trim() !== "" ? resource : undefined
+}
+
+function resourceQuery(resource: string) {
+  return [
+    "resources",
+    "| where type =~ 'microsoft.cognitiveservices/accounts' and kind in~ ('AIServices', 'OpenAI')",
+    // The custom subdomain is the resource name of every endpoint, and Entra ID authentication requires one.
+    // Subdomains are globally unique, so a name matches at most one resource.
+    // https://learn.microsoft.com/azure/ai-services/cognitive-services-custom-subdomains
+    "| extend resourceName = tostring(properties.customSubDomainName)",
+    `| where resourceName =~ '${resource}'`,
+    "| project id",
+    "| take 1",
+  ].join(" ")
+}
+
+// A deployment's ID is its name, while limits, costs, and routes come from the catalog model it deploys. Azure compares
+// names without case and may return another case, so IDs are lowercase like the catalog's.
+// https://learn.microsoft.com/azure/azure-resource-manager/management/resource-name-rules
+function deployedModels(deployments: readonly Deployment[], catalog: readonly Model.MutableInfo[]) {
+  const models = new Map(catalog.map((model) => [model.id.toLowerCase(), model]))
+  return new Map(
+    deployments.flatMap((deployment) => {
+      const model = catalogModel(models, deployment)
+      if (!model) return []
+      const id = Model.ID.make(deployment.name.toLowerCase())
+      const info: Model.MutableInfo = {
+        ...structuredClone(model),
+        id,
+        modelID: Model.ID.make(deployment.name),
+        name: id === model.id.toLowerCase() ? model.name : `${model.name} (${deployment.name})`,
+      }
+      return [[id, info] as const]
+    }),
+  )
+}
+
+// Azure spells some models unlike the catalog: a model name plus a separate version, such as `gpt-4` for GPT-4 Turbo,
+// `gpt-35-turbo`, or mixed case such as `DeepSeek-V4-Flash`. A deployment named after a catalog model then stands for
+// that model, as it did before discovery; any other is left to explicit configuration.
+// https://learn.microsoft.com/azure/foundry/openai/concepts/retired-models
+function catalogModel<T>(models: ReadonlyMap<string, T>, deployment: Deployment) {
+  return models.get(deployment.model.toLowerCase()) ?? models.get(deployment.name.toLowerCase())
+}
+
 function responsesWebSocketCapable(provider: Provider.Info) {
   if (provider.package !== "@opencode/ai/providers/azure/responses") return false
   const settings = provider.settings
diff --git a/packages/core/test/plugin/provider-azure.test.ts b/packages/core/test/plugin/provider-azure.test.ts
index ddb44295c67..2d4815bd4f0 100644
--- a/packages/core/test/plugin/provider-azure.test.ts
+++ b/packages/core/test/plugin/provider-azure.test.ts
@@ -1,27 +1,35 @@
 import { chmod } from "node:fs/promises"
 import { Agent } from "@opencode/core/agent"
 import { describe, expect } from "bun:test"
-import { Effect, Schedule } from "effect"
+import { Effect, Schedule, Schema } from "effect"
+import { Config } from "@opencode/core/config"
+import { ConfigProviderPlugin } from "@opencode/core/config/plugin/provider"
+import { Bus } from "@opencode/core/bus"
 import { Credential } from "@opencode/core/credential"
 import { Model } from "@opencode/core/model"
 import { Plugin } from "@opencode/core/plugin"
 import { PluginHost } from "@opencode/core/plugin/host"
 import { PluginHooks } from "@opencode/core/plugin/hooks"
-import { AzurePlugin } from "@opencode/core/plugin/provider/azure"
+import { make } from "@opencode/core/plugin/provider/azure"
 import { Provider } from "@opencode/core/provider"
 import { Integration } from "@opencode/core/integration"
 import { Location } from "@opencode/core/location"
 import { Session } from "@opencode/core/session"
+import { Document, Info } from "@opencode/schema/config"
 import { AppProcess } from "@opencode/util/process"
 import { testEffect } from "../lib/effect"
 import { PluginTestLayer } from "./fixture"
 
 const it = testEffect(PluginTestLayer)
+const decodeConfig = Schema.decodeUnknownSync(Info)
 
-const addPlugin = Effect.fn(function* () {
+// Nothing listens here, so tests that are not about deployments never find any.
+const offline = { resource: () => "http://127.0.0.1:1/openai", management: "http://127.0.0.1:1" }
+
+const addPlugin = Effect.fn(function* (endpoints = offline) {
   const plugin = yield* Plugin.Service
   const host = yield* PluginHost.make(plugin)
-  yield* AzurePlugin.effect(host)
+  yield* make(endpoints).effect(host)
 })
 
 function required<T>(value: T | undefined): T {
@@ -29,6 +37,103 @@ function required<T>(value: T | undefined): T {
   return value
 }
 
+function eventually<A, R>(
+  effect: Effect.Effect<A, never, R>,
+  predicate: (value: A) => boolean,
+  remaining = 3000,
+): Effect.Effect<A, Error, R> {
+  return Effect.gen(function* () {
+    const value = yield* effect
+    if (predicate(value)) return value
+    if (remaining === 0) return yield* Effect.fail(new Error("Timed out waiting for value"))
+    yield* Effect.promise(() => Bun.sleep(1))
+    return yield* eventually(effect, predicate, remaining - 1)
+  })
+}
+
+type AzureRequest = {
+  readonly method: string
+  readonly path: string
+  readonly key: string | null
+  readonly authorization: string | null
+  readonly body: string
+}
+
+function withAzure<A, E, R>(
+  respond: (request: AzureRequest) => Response | Promise<Response>,
+  fx: (input: { endpoints: typeof offline; requests: AzureRequest[] }) => Effect.Effect<A, E, R>,
+) {
+  return Effect.acquireUseRelease(
+    Effect.sync(() => {
+      const requests: AzureRequest[] = []
+      const server = Bun.serve({
+        port: 0,
+        fetch: async (request) => {
+          const url = new URL(request.url)
+          const received = {
+            method: request.method,
+            path: url.pathname + url.search,
+            key: request.headers.get("api-key"),
+            authorization: request.headers.get("authorization"),
+            body: await request.text(),
+          }
+          requests.push(received)
+          return respond(received)
+        },
+      })
+      return { requests, server }
+    }),
+    ({ requests, server }) =>
+      fx({
+        requests,
+        endpoints: { resource: () => `${server.url.origin}/openai`, management: server.url.origin },
+      }),
+    ({ server }) => Effect.promise(() => server.stop(true)),
+  )
+}
+
+const seedCatalog = Effect.gen(function* () {
+  const catalog = yield* Provider.Service
+  yield* catalog.transform((editor) => {
+    editor.update(Provider.ID.azure, (provider) => {
+      provider.package = "@opencode/ai/providers/azure/responses"
+    })
+    editor.models.update(Provider.ID.azure, Model.ID.make("gpt-5"), () => {})
+    editor.models.update(Provider.ID.azure, Model.ID.make("gpt-5-mini"), (model) => {
+      model.name = "GPT-5 Mini"
+      model.limit = { context: 400_000, output: 128_000 }
+    })
+    editor.models.update(Provider.ID.azure, Model.ID.make("gpt-5-nano"), (model) => {
+      model.name = "GPT-5 Nano"
+      model.limit = { context: 300_000, output: 64_000 }
+    })
+    editor.models.update(Provider.ID.azure, Model.ID.make("deepseek-v4-flash"), (model) => {
+      model.name = "DeepSeek-V4-Flash"
+      model.package = "@opencode/ai/providers/openai-compatible"
+      model.settings = { baseURL: "https://${AZURE_RESOURCE_NAME}.services.ai.azure.com/models" }
+    })
+  })
+})
+
+const azureModels = Effect.gen(function* () {
+  const models = yield* Model.Service
+  return (yield* models.all())
+    .filter((model) => model.providerID === Provider.ID.azure)
+    .toSorted((a, b) => a.id.localeCompare(b.id))
+})
+
+const cliTokens = (args: readonly string[]) => ({
+  accessToken: `${args[args.indexOf("--scope") + 1]}-token`,
+  expires_on: Math.floor((Date.now() + 60 * 60 * 1000) / 1000),
+})
+
+const account = (name: string) => ({
+  id: `/subscriptions/sub/resourceGroups/rg-${name}/providers/Microsoft.CognitiveServices/accounts/${name}`,
+  resourceName: name,
+  resourceGroup: `rg-${name}`,
+  location: "swedencentral",
+})
+
 function withEnv<A, E, R>(vars: Record<string, string | undefined>, fx: () => Effect.Effect<A, E, R>) {
   return Effect.acquireUseRelease(
     Effect.sync(() => {
@@ -100,6 +205,15 @@ const azureCredential = Effect.gen(function* () {
     }),
   })
 })
+
+const keyCredential = Effect.gen(function* () {
+  const credentials = yield* Credential.Service
+  return yield* credentials.create({
+    integrationID: Integration.ID.make("azure"),
+    value: Credential.Key.make({ type: "key", key: "secret", configuration: { resourceName: "test-resource" } }),
+  })
+})
+
 describe("AzurePlugin", () => {
   it.effect("registers a resource name form when the environment does not provide one", () =>
     withEnv({ AZURE_RESOURCE_NAME: undefined, AZURE_COGNITIVE_SERVICES_RESOURCE_NAME: undefined }, () =>
@@ -216,15 +330,14 @@ describe("AzurePlugin", () => {
               access: "legacy-cli-token",
               metadata: { resourceName: "test-resource" },
             })
-            expect(commands).toEqual([
-              [
-                "account",
-                "get-access-token",
-                "--scope",
-                "https://cognitiveservices.azure.com/.default",
-                "--output",
-                "json",
-              ],
+            // Discovery for the new connection may already be minting its management token afterwards.
+            expect(commands[0]).toEqual([
+              "account",
+              "get-access-token",
+              "--scope",
+              "https://cognitiveservices.azure.com/.default",
+              "--output",
+              "json",
             ])
           }),
       ),
@@ -236,30 +349,562 @@ describe("AzurePlugin", () => {
     return withAzureCommands(
       (args) => {
         commands.push([...args])
-        return []
+        return cliTokens(args)
+      },
+      () =>
+        withAzure(
+          (request) =>
+            request.method === "POST"
+              ? Response.json({ data: [account("test-resource")] })
+              : Response.json({
+                  value: [
+                    { name: "gpt-5-mini", properties: { model: { name: "gpt-5-mini" }, provisioningState: "Succeeded" } },
+                  ],
+                }),
+          ({ endpoints, requests }) =>
+            Effect.gen(function* () {
+              const catalog = yield* Provider.Service
+              const models = yield* Model.Service
+              yield* seedCatalog
+              yield* azureCredential
+              yield* addPlugin(endpoints)
+
+              // Startup serves the whole catalog; the deployments arrive afterwards without blocking it.
+              expect(commands).toEqual([])
+              expect(requests).toEqual([])
+              expect((yield* catalog.get(Provider.ID.azure))?.settings?.resourceName).toBe("test-resource")
+              expect(yield* models.get(Provider.ID.azure, Model.ID.make("gpt-5-mini"))).toBeDefined()
+              expect(yield* models.get(Provider.ID.azure, Model.ID.make("gpt-5-nano"))).toBeDefined()
+
+              const deployed = yield* eventually(azureModels, (list) => list.length === 1)
+              expect(deployed.map((model) => model.id)).toEqual([Model.ID.make("gpt-5-mini")])
+              expect(commands).toEqual([
+                ["account", "get-access-token", "--scope", "https://management.azure.com/.default", "--output", "json"],
+              ])
+            }),
+        ),
+    )
+  })
+
+  it.live("does not refresh an expired Azure CLI token while starting", () => {
+    const commands: string[][] = []
+    return withAzureCommands(
+      (args) => {
+        commands.push([...args])
+        return { accessToken: "refreshed-token", expires_on: Math.floor((Date.now() + 60 * 60 * 1000) / 1000) }
       },
       () =>
         Effect.gen(function* () {
+          const credentials = yield* Credential.Service
+          yield* credentials.create({
+            integrationID: Integration.ID.make("azure"),
+            value: Credential.OAuth.make({
+              type: "oauth",
+              methodID: Integration.MethodID.make("azure-cli"),
+              access: "expired-token",
+              refresh: "azure-cli",
+              expires: Date.now() - 60 * 60 * 1000,
+              metadata: { resourceName: "test-resource" },
+            }),
+          })
           const catalog = yield* Provider.Service
-          const models = yield* Model.Service
           yield* catalog.transform((editor) => {
             editor.update(Provider.ID.azure, (provider) => {
               provider.package = "@opencode/ai/providers/azure/responses"
             })
-            editor.models.update(Provider.ID.azure, Model.ID.make("gpt-5-mini"), () => {})
-            editor.models.update(Provider.ID.azure, Model.ID.make("gpt-5-nano"), () => {})
           })
-          yield* azureCredential
           yield* addPlugin()
 
           expect(commands).toEqual([])
           expect((yield* catalog.get(Provider.ID.azure))?.settings?.resourceName).toBe("test-resource")
-          expect(yield* models.get(Provider.ID.azure, Model.ID.make("gpt-5-mini"))).toBeDefined()
-          expect(yield* models.get(Provider.ID.azure, Model.ID.make("gpt-5-nano"))).toBeDefined()
         }),
     )
   })
 
+  it.live("narrows the catalog to the deployments listed with an API key", () =>
+    withEnv({ AZURE_RESOURCE_NAME: undefined, AZURE_COGNITIVE_SERVICES_RESOURCE_NAME: undefined }, () =>
+      withAzure(
+        () =>
+          Response.json({
+            data: [
+              { id: "gpt-5-mini-eu", model: "gpt-5-mini", status: "succeeded" },
+              { id: "gpt-5-mini", model: "gpt-5-mini", status: "succeeded" },
+              { id: "nano-b", model: "gpt-5-nano", status: "succeeded" },
+              { id: "nano-a", model: "gpt-5-nano", status: "succeeded" },
+              { id: "DeepSeek-V4-Flash", model: "DeepSeek-V4-Flash", status: "succeeded" },
+              { id: "gpt-5-pending", model: "gpt-5", status: "creating" },
+              { id: "fine-tuned", model: "gpt-4o-mini.ft-123", status: "succeeded" },
+              { id: 42 },
+            ],
+          }),
+        ({ endpoints, requests }) =>
+          Effect.gen(function* () {
+            const catalog = yield* Provider.Service
+            yield* seedCatalog
+            yield* keyCredential
+            yield* addPlugin(endpoints)
+
+            const deployed = yield* eventually(azureModels, (list) => list.length === 5)
+            expect(deployed.map((model) => [model.id, model.modelID, model.name])).toEqual([
+              [Model.ID.make("deepseek-v4-flash"), Model.ID.make("DeepSeek-V4-Flash"), "DeepSeek-V4-Flash"],
+              // The deployment named after its model keeps the model ID even though Azure listed another first.
+              [Model.ID.make("gpt-5-mini"), Model.ID.make("gpt-5-mini"), "GPT-5 Mini"],
+              [Model.ID.make("gpt-5-mini-eu"), Model.ID.make("gpt-5-mini-eu"), "GPT-5 Mini (gpt-5-mini-eu)"],
+              // Custom deployment IDs are stable independently of the other deployments of the same model.
+              [Model.ID.make("nano-a"), Model.ID.make("nano-a"), "GPT-5 Nano (nano-a)"],
+              [Model.ID.make("nano-b"), Model.ID.make("nano-b"), "GPT-5 Nano (nano-b)"],
+            ])
+            // A further deployment of a model keeps the catalog facts of that model.
+            expect(deployed[2]?.limit).toEqual({ context: 400_000, output: 128_000 })
+            // The resource name entered with the API key reaches catalog endpoints, not only the request route.
+            expect(deployed[0]?.settings?.baseURL).toBe("https://test-resource.services.ai.azure.com/models")
+            expect((yield* catalog.get(Provider.ID.azure))?.settings?.resourceName).toBe("test-resource")
+            expect(requests).toEqual([
+              {
+                method: "GET",
+                path: "/openai/deployments?api-version=2022-12-01",
+                key: "secret",
+                authorization: null,
+                body: "",
+              },
+            ])
+          }),
+      ),
+    ),
+  )
+
+  it.live("lists Azure CLI deployments through the management API", () => {
+    const commands: string[][] = []
+    const resource = account("test-resource")
+    return withAzureCommands(
+      (args) => {
+        commands.push([...args])
+        return cliTokens(args)
+      },
+      () =>
+        withAzure(
+          (request) => {
+            if (request.path.startsWith("/providers/Microsoft.ResourceGraph/resources"))
+              return Response.json({ data: [resource] })
+            return Response.json({
+              value: [
+                {
+                  name: "gpt-production",
+                  properties: { model: { format: "OpenAI", name: "gpt-5-mini" }, provisioningState: "Succeeded" },
+                },
+                {
+                  name: "gpt-5-nano",
+                  properties: { model: { format: "OpenAI", name: "gpt-5-nano" }, provisioningState: "Failed" },
+                },
+              ],
+            })
+          },
+          ({ endpoints, requests }) =>
+            Effect.gen(function* () {
+              yield* seedCatalog
+              yield* azureCredential
+              yield* addPlugin(endpoints)
+
+              const deployed = yield* eventually(azureModels, (list) => list.length === 1)
+              expect(deployed.map((model) => [model.id, model.modelID])).toEqual([
+                [Model.ID.make("gpt-production"), Model.ID.make("gpt-production")],
+              ])
+              expect(commands).toEqual([
+                ["account", "get-access-token", "--scope", "https://management.azure.com/.default", "--output", "json"],
+              ])
+              expect(requests.map((request) => [request.method, request.path, request.authorization])).toEqual([
+                [
+                  "POST",
+                  "/providers/Microsoft.ResourceGraph/resources?api-version=2022-10-01",
+                  "Bearer https://management.azure.com/.default-token",
+                ],
+                [
+                  "GET",
+                  `${resource.id}/deployments?api-version=2024-10-01`,
+                  "Bearer https://management.azure.com/.default-token",
+                ],
+              ])
+              expect(requests[0]?.body).toContain("resourceName =~ 'test-resource'")
+            }),
+        ),
+    )
+  })
+
+  it.live("falls back to the resource inventory when the Azure CLI cannot read the management API", () =>
+    withAzureCommands(cliTokens, () =>
+      withAzure(
+        (request) =>
+          request.path.startsWith("/openai/deployments")
+            ? Response.json({ data: [{ id: "gpt-5-mini", model: "gpt-5-mini", status: "succeeded" }] })
+            : new Response("Forbidden", { status: 403 }),
+        ({ endpoints, requests }) =>
+          Effect.gen(function* () {
+            yield* seedCatalog
+            yield* azureCredential
+            yield* addPlugin(endpoints)
+
+            const deployed = yield* eventually(azureModels, (list) => list.length === 1)
+            expect(deployed.map((model) => model.id)).toEqual([Model.ID.make("gpt-5-mini")])
+            expect(requests.map((request) => [request.method, request.path, request.authorization])).toEqual([
+              [
+                "POST",
+                "/providers/Microsoft.ResourceGraph/resources?api-version=2022-10-01",
+                "Bearer https://management.azure.com/.default-token",
+              ],
+              ["GET", "/openai/deployments?api-version=2022-12-01", "Bearer stored-token"],
+            ])
+          }),
+      ),
+    ),
+  )
+
+  it.live("keeps the catalog when deployments cannot be listed", () =>
+    withAzure(
+      () => new Response("Unavailable", { status: 503 }),
+      ({ endpoints, requests }) =>
+        Effect.gen(function* () {
+          yield* seedCatalog
+          yield* keyCredential
+          yield* addPlugin(endpoints)
+
+          yield* eventually(Effect.succeed(requests), (list) => list.length === 1)
+          yield* Effect.promise(() => Bun.sleep(25))
+          expect((yield* azureModels).map((model) => model.id)).toEqual([
+            Model.ID.make("deepseek-v4-flash"),
+            Model.ID.make("gpt-5"),
+            Model.ID.make("gpt-5-mini"),
+            Model.ID.make("gpt-5-nano"),
+          ])
+        }),
+    ),
+  )
+
+  it.live("keeps the listed deployments while the Azure CLI cannot refresh its token", () => {
+    const commands: string[][] = []
+    return withAzureCommands(
+      (args) => {
+        commands.push([...args])
+        return new Error("az: please run 'az login' to setup account")
+      },
+      () =>
+        withAzure(
+          () => Response.json({ data: [{ id: "gpt-5-mini", model: "gpt-5-mini", status: "succeeded" }] }),
+          ({ endpoints }) =>
+            Effect.gen(function* () {
+              const credentials = yield* Credential.Service
+              const bus = yield* Bus.Service
+              yield* seedCatalog
+              const credential = yield* azureCredential
+              yield* addPlugin(endpoints)
+              yield* eventually(azureModels, (list) => list.length === 1)
+
+              // A failed refresh of the same account retains its last successful inventory.
+              yield* credentials.update(credential.id, {
+                value: Credential.OAuth.make({
+                  type: "oauth",
+                  methodID: Integration.MethodID.make("azure-cli"),
+                  access: "expired-token",
+                  refresh: "azure-cli",
+                  expires: Date.now() - 60 * 60 * 1000,
+                  metadata: { resourceName: "test-resource" },
+                }),
+              })
+              yield* bus.publish(
+                Credential.Event.Switched,
+                { integrationID: Integration.ID.make("azure"), credentialID: credential.id },
+                { global: true },
+              )
+              yield* eventually(Effect.succeed(commands), (list) => list.length > 0)
+              yield* Effect.promise(() => Bun.sleep(25))
+
+              expect((yield* azureModels).map((model) => model.id)).toEqual([Model.ID.make("gpt-5-mini")])
+            }),
+        ),
+    )
+  })
+
+  it.live("rebinds a switched account without waiting for the previous account's discovery", () => {
+    const pending = Promise.withResolvers<Response>()
+    const calls: AzureRequest[] = []
+    return withAzure(
+      (request) => {
+        calls.push(request)
+        if (calls.length === 2) return pending.promise
+        return Response.json({
+          data: [
+            {
+              id: request.key === "secret" ? "production-a" : "production-b",
+              model: "gpt-5-mini",
+              status: "succeeded",
+            },
+          ],
+        })
+      },
+      ({ endpoints }) =>
+        Effect.gen(function* () {
+          const bus = yield* Bus.Service
+          const credentials = yield* Credential.Service
+          const providers = yield* Provider.Service
+          yield* seedCatalog
+          const first = yield* keyCredential
+          yield* addPlugin(endpoints)
+          const previous = yield* eventually(azureModels, (list) => list.some((model) => model.id === "production-a"))
+
+          yield* bus.publish(
+            Credential.Event.Switched,
+            { integrationID: Integration.ID.make("azure"), credentialID: first.id },
+            { global: true },
+          )
+          yield* eventually(Effect.succeed(calls), (list) => list.length === 2)
+          const next = yield* credentials.create({
+            integrationID: Integration.ID.make("azure"),
+            value: Credential.Key.make({
+              type: "key",
+              key: "other-key",
+              configuration: { resourceName: "other-resource" },
+            }),
+          })
+
+          // The previous account's discovery is still waiting on Azure, yet the new account is already listed.
+          const deployed = yield* eventually(azureModels, (list) => list.length === 1 && list[0]?.id === "production-b")
+          expect(deployed[0]?.settings?.resourceName).toBe("other-resource")
+          expect((yield* providers.available()).some((provider) => provider.id === Provider.ID.azure)).toBe(true)
+          expect((yield* providers.snapshot()).records.get(Provider.ID.azure)?.sourceConnection).toMatchObject({
+            type: "credential",
+            id: next.id,
+          })
+          expect(previous.map((model) => model.id)).toEqual([Model.ID.make("production-a")])
+
+          // The interrupted discovery never publishes the previous account's late answer.
+          pending.resolve(Response.json({ data: [{ id: "stale", model: "gpt-5-nano", status: "succeeded" }] }))
+          yield* Effect.promise(() => Bun.sleep(25))
+          expect((yield* azureModels).map((model) => model.id)).toEqual([Model.ID.make("production-b")])
+        }),
+    )
+  })
+
+  it.live("names models after their deployments and keeps catalog models that Azure spells differently", () =>
+    withAzure(
+      () =>
+        Response.json({
+          data: [
+            // Azure compares names without case and may return another case than the one created.
+            { id: "GPT-5-Nano", model: "gpt-5-nano", status: "succeeded" },
+            // A deployment named after another model is still the model it deploys.
+            { id: "gpt-5", model: "gpt-5-mini", status: "succeeded" },
+            // Azure's model name differs from the catalog, as with `gpt-4` for GPT-4 Turbo.
+            { id: "deepseek-v4-flash", model: "DeepSeek-V4-Flash-2026", status: "succeeded" },
+            { id: "ft-legal", model: "gpt-4o-mini.ft-123", status: "succeeded" },
+          ],
+        }),
+      ({ endpoints }) =>
+        Effect.gen(function* () {
+          yield* seedCatalog
+          yield* keyCredential
+          yield* addPlugin(endpoints)
+
+          const deployed = yield* eventually(azureModels, (list) => list.length === 3)
+          expect(deployed.map((model) => [model.id, model.modelID, model.name, model.limit.context])).toEqual([
+            [Model.ID.make("deepseek-v4-flash"), Model.ID.make("deepseek-v4-flash"), "DeepSeek-V4-Flash", 200_000],
+            [Model.ID.make("gpt-5"), Model.ID.make("gpt-5"), "GPT-5 Mini (gpt-5)", 400_000],
+            [Model.ID.make("gpt-5-nano"), Model.ID.make("GPT-5-Nano"), "GPT-5 Nano", 300_000],
+          ])
+        }),
+    ),
+  )
+
+  it.live("keeps deployment IDs when another deployment of the same model is removed", () => {
+    const inventory = { names: ["nano-a", "nano-b"] }
+    return withAzure(
+      () => Response.json({ data: inventory.names.map((id) => ({ id, model: "gpt-5-nano", status: "succeeded" })) }),
+      ({ endpoints }) =>
+        Effect.gen(function* () {
+          const bus = yield* Bus.Service
+          yield* seedCatalog
+          const credential = yield* keyCredential
+          yield* addPlugin(endpoints)
+          const before = yield* eventually(azureModels, (list) => list.length === 2)
+          expect(before.map((model) => model.id)).toEqual([Model.ID.make("nano-a"), Model.ID.make("nano-b")])
+
+          inventory.names = ["nano-b"]
+          yield* bus.publish(
+            Credential.Event.Switched,
+            { integrationID: Integration.ID.make("azure"), credentialID: credential.id },
+            { global: true },
+          )
+          const after = yield* eventually(azureModels, (list) => list.length === 1)
+          expect(after).toEqual([before[1]])
+        }),
+    )
+  })
+
+  it.live("loads all management deployment pages before publishing the inventory", () => {
+    const pending = Promise.withResolvers<Response>()
+    const state = { url: "" }
+    return withAzureCommands(cliTokens, () =>
+      withAzure(
+        (request) => {
+          if (request.path.startsWith("/openai/deployments")) return new Response("Not found", { status: 404 })
+          if (request.method === "POST") return Response.json({ data: [account("test-resource")] })
+          if (request.path === "/page-2") return pending.promise
+          return Response.json({
+            nextLink: `${state.url}/page-2`,
+            value: [{ name: "mini", properties: { model: { name: "gpt-5-mini" }, provisioningState: "Succeeded" } }],
+          })
+        },
+        ({ endpoints, requests }) =>
+          Effect.gen(function* () {
+            state.url = endpoints.management
+            yield* seedCatalog
+            yield* azureCredential
+            yield* addPlugin(endpoints)
+            yield* eventually(Effect.succeed(requests), (list) => list.length === 3)
+            expect(requests[2]).toMatchObject({
+              path: "/page-2",
+              authorization: "Bearer https://management.azure.com/.default-token",
+            })
+            expect(yield* azureModels).toHaveLength(4)
+            pending.resolve(
+              Response.json({
+                value: [
+                  { name: "nano", properties: { model: { name: "gpt-5-nano" }, provisioningState: "Succeeded" } },
+                ],
+              }),
+            )
+            const deployed = yield* eventually(azureModels, (list) => list.length === 2)
+            expect(deployed.map((model) => model.modelID)).toEqual([Model.ID.make("mini"), Model.ID.make("nano")])
+          }),
+      ),
+    )
+  })
+
+  it.live("keeps the last complete inventory when a later management page fails", () => {
+    const state = { fail: false, url: "" }
+    return withAzureCommands(cliTokens, () =>
+      withAzure(
+        (request) => {
+          if (request.path.startsWith("/openai/deployments")) return new Response("Not found", { status: 404 })
+          if (request.method === "POST") return Response.json({ data: [account("test-resource")] })
+          if (request.path === "/page-2") return new Response("Unavailable", { status: 503 })
+          if (state.fail)
+            return Response.json({
+              nextLink: `${state.url}/page-2`,
+              value: [{ name: "nano", properties: { model: { name: "gpt-5-nano" }, provisioningState: "Succeeded" } }],
+            })
+          return Response.json({
+            value: [{ name: "mini", properties: { model: { name: "gpt-5-mini" }, provisioningState: "Succeeded" } }],
+          })
+        },
+        ({ endpoints, requests }) =>
+          Effect.gen(function* () {
+            state.url = endpoints.management
+            const bus = yield* Bus.Service
+            yield* seedCatalog
+            const credential = yield* azureCredential
+            yield* addPlugin(endpoints)
+            yield* eventually(azureModels, (list) => list.length === 1 && list[0]?.id === "mini")
+            state.fail = true
+            yield* bus.publish(
+              Credential.Event.Switched,
+              { integrationID: Integration.ID.make("azure"), credentialID: credential.id },
+              { global: true },
+            )
+            // The cached resource skips Resource Graph; the failed page then falls back to the resource inventory.
+            yield* eventually(Effect.succeed(requests), (list) => list.length === 5)
+            expect(requests.map((request) => request.path)).toEqual([
+              "/providers/Microsoft.ResourceGraph/resources?api-version=2022-10-01",
+              `${account("test-resource").id}/deployments?api-version=2024-10-01`,
+              `${account("test-resource").id}/deployments?api-version=2024-10-01`,
+              "/page-2",
+              "/openai/deployments?api-version=2022-12-01",
+            ])
+            yield* Effect.promise(() => Bun.sleep(25))
+            expect((yield* azureModels).map((model) => model.id)).toEqual([Model.ID.make("mini")])
+          }),
+      ),
+    )
+  })
+
+  it.live("does not follow a management page outside the management endpoint", () =>
+    withAzureCommands(cliTokens, () =>
+      withAzure(
+        (request) => {
+          if (request.path.startsWith("/openai/deployments")) return new Response("Not found", { status: 404 })
+          if (request.method === "POST") return Response.json({ data: [account("test-resource")] })
+          return Response.json({
+            nextLink: "https://attacker.example/page-2",
+            value: [{ name: "mini", properties: { model: { name: "gpt-5-mini" }, provisioningState: "Succeeded" } }],
+          })
+        },
+        ({ endpoints, requests }) =>
+          Effect.gen(function* () {
+            yield* seedCatalog
+            yield* azureCredential
+            yield* addPlugin(endpoints)
+
+            yield* eventually(Effect.succeed(requests), (list) => list.length === 3)
+            yield* Effect.promise(() => Bun.sleep(25))
+            expect(requests[2]?.path).toBe("/openai/deployments?api-version=2022-12-01")
+            // A partial inventory is never published in place of the catalog.
+            expect(yield* azureModels).toHaveLength(4)
+          }),
+      ),
+    ),
+  )
+
+  it.live("keeps the catalog for a custom endpoint", () =>
+    withAzure(
+      () => Response.json({ data: [{ id: "gpt-5-mini", model: "gpt-5-mini", status: "succeeded" }] }),
+      ({ endpoints, requests }) =>
+        Effect.gen(function* () {
+          const catalog = yield* Provider.Service
+          yield* seedCatalog
+          yield* catalog.transform((editor) => {
+            editor.update(Provider.ID.azure, (provider) => {
+              provider.settings = { baseURL: "https://gateway.example/azure" }
+            })
+          })
+          yield* keyCredential
+          yield* addPlugin(endpoints)
+
+          yield* Effect.promise(() => Bun.sleep(50))
+          expect(requests).toEqual([])
+          expect(yield* azureModels).toHaveLength(4)
+        }),
+    ),
+  )
+
+  it.live("leaves a model that is configured explicitly to the configuration", () =>
+    withAzure(
+      () => Response.json({ data: [{ id: "gpt-5-mini", model: "gpt-5-mini", status: "succeeded" }] }),
+      ({ endpoints }) =>
+        Effect.gen(function* () {
+          yield* seedCatalog
+          yield* keyCredential
+          yield* addPlugin(endpoints)
+          const plugin = yield* Plugin.Service
+          yield* ConfigProviderPlugin.Plugin.effect(yield* PluginHost.make(plugin)).pipe(
+            Effect.provide(
+              Config.testLayer([
+                new Document({
+                  type: "document",
+                  info: decodeConfig({
+                    providers: { azure: { models: { "gpt-5-nano": { modelID: "nano-production" } } } },
+                  }),
+                }),
+              ]),
+            ),
+          )
+
+          const deployed = yield* eventually(azureModels, (list) => list.length === 2)
+          expect(deployed.map((model) => [model.id, model.modelID, model.name, model.limit.context])).toEqual([
+            [Model.ID.make("gpt-5-mini"), Model.ID.make("gpt-5-mini"), "GPT-5 Mini", 400_000],
+            [Model.ID.make("gpt-5-nano"), Model.ID.make("nano-production"), "GPT-5 Nano", 300_000],
+          ])
+        }),
+    ),
+  )
+
   it.effect("uses the correct bearer token audience for Azure and Foundry requests", () =>
     withAzureCommands(
       (args) => {
@@ -319,6 +964,21 @@ describe("AzurePlugin", () => {
     ),
   )
 
+  it.effect("prefers the connection's resource over configuration, as requests do", () =>
+    Effect.gen(function* () {
+      const catalog = yield* Provider.Service
+      yield* catalog.transform((editor) => {
+        editor.update(Provider.ID.azure, (provider) => {
+          provider.package = "@opencode/ai/providers/azure/responses"
+          provider.settings = { resourceName: "from-config" }
+        })
+      })
+      yield* keyCredential
+      yield* addPlugin()
+      expect(required(yield* catalog.get(Provider.ID.azure)).settings?.resourceName).toBe("test-resource")
+    }),
+  )
+
   it.effect("resolves resourceName from env", () =>
     withEnv({ AZURE_RESOURCE_NAME: "from-env" }, () =>
       Effect.gen(function* () {
@@ -473,7 +1133,14 @@ describe("AzurePlugin", () => {
         yield* addPlugin()
 
         expect((yield* catalog.get(Provider.ID.azure))?.settings?.transport).toBe("websocket")
-        for (const modelID of [models.responses, models.chat, models.preview, models.deploymentURL, models.gateway, models.nonAzure]) {
+        for (const modelID of [
+          models.responses,
+          models.chat,
+          models.preview,
+          models.deploymentURL,
+          models.gateway,
+          models.nonAzure,
+        ]) {
           const model = required(yield* service.get(Provider.ID.azure, modelID))
           expect(model.settings?.transport).toBeUndefined()
         }
diff --git a/services/www/src/docs/content/providers.mdx b/services/www/src/docs/content/providers.mdx
index 22588741113..c6ea1b20ea1 100644
--- a/services/www/src/docs/content/providers.mdx
+++ b/services/www/src/docs/content/providers.mdx
@@ -268,7 +268,11 @@ choose **Microsoft Entra ID (Azure CLI)** when connecting Azure.
 az login
 ```
 
-For a resource in another tenant or subscription, select both explicitly.
+Enter the resource name when connecting if it is not already supplied by configuration or the environment. A resource
+name saved with a connection takes precedence over configuration; connect again to change it.
+
+Azure CLI connections use the account selected in the Azure CLI. For a resource in another tenant or subscription,
+select both explicitly.
 
 ```bash
 az login --tenant TENANT_ID
@@ -278,8 +282,32 @@ az account set --subscription NAME_OR_ID
 Instead of configuration, `AZURE_RESOURCE_NAME` supplies the resource name to the OpenCode server. The legacy
 `AZURE_COGNITIVE_SERVICES_RESOURCE_NAME` variable also works.
 
-OpenCode does not query Azure management APIs or discover deployments. If a deployment does not match its catalog
-model name, map an OpenCode model ID to the deployment with `modelID`.
+### Deployments
+
+Azure serves a model only through a deployment, so OpenCode lists the deployments of your resource and shows those
+instead of the whole Azure catalog. The list loads in the background after startup, after connecting, and after
+switching accounts. To pick up a deployment added later, restart OpenCode or connect again.
+
+- Until the list loads for a connection, or if it fails, the whole catalog stays available.
+- A failed reload for the same connection keeps its last complete inventory.
+- Switching accounts never shows the previous account's inventory; the new account starts from the whole catalog.
+- Deployments inherit limits, costs, and capabilities from their catalog models.
+
+Deployment names are their model IDs, in lowercase because Azure ignores case in names, for example
+`azure/gpt-production`. A deployment named after its model, such as `gpt-5-mini`, keeps the catalog ID. These IDs stay
+stable when other deployments are added or removed.
+
+OpenCode matches a deployment to the catalog by the model it deploys, or by its name when Azure spells the model
+differently, as with `gpt-4` for GPT-4 Turbo. A deployment named after another model, such as `gpt-5` deploying
+`gpt-5-mini`, shows that model's name, limits, and costs.
+
+With the Azure CLI, OpenCode lists deployments through the Azure management API, which needs read access to the
+resource. If that fails, and always with an API key, it uses the resource's legacy deployment inventory. Discovery reads
+every page before publishing the list. A custom `settings.baseURL` keeps the catalog.
+
+Configure a deployment explicitly when its model is not in the catalog, such as a fine-tuned model, or to map a model ID
+to a deployment yourself. OpenCode logs a warning naming such deployments. Explicitly configured models are always
+kept.
 
 ```jsonc title="opencode.jsonc"
 {