fix(provider): select DeepSeek adapter for Azure deployments

packages/core/src/plugin/provider/azure.ts

@@ -3,6 +3,14 @@ import { define } from "../internal"
 import { ProviderV2 } from "../../provider"
 
 function selectLanguage(sdk: any, modelID: string, useChat: boolean) {
+  // Azure DeepSeek deployments need the DeepSeek-specific adapter for reasoning
+  // effort and reasoning_content across agentic turns; it is chat-completions
+  // based, so it also applies when useCompletionUrls is set.
+  // Note: this predicate (any "deepseek" deployment) deliberately differs from
+  // the "deepseek-v4" gate in transform.ts reasoningVariants, which only
+  // exposes a max effort for v4 models.
+  // Keep in sync with packages/opencode/src/provider/provider.ts selectAzureLanguageModel.
+  if (modelID.toLowerCase().includes("deepseek") && sdk.deepseek) return sdk.deepseek(modelID)
   if (useChat && sdk.chat) return sdk.chat(modelID)
   if (sdk.responses) return sdk.responses(modelID)
   if (sdk.messages) return sdk.messages(modelID)

packages/core/test/plugin/provider-azure.test.ts

@@ -55,6 +55,7 @@ function fakeSelectorSdk(calls: string[]) {
     responses: make("responses"),
     messages: make("messages"),
     chat: make("chat"),
+    deepseek: make("deepseek"),
     languageModel: make("languageModel"),
   }
 }
@@ -261,6 +262,64 @@ describe("AzurePlugin", () => {
     }),
   )
 
+  it.effect("selects deepseek adapter for DeepSeek deployments", () =>
+    Effect.gen(function* () {
+      const plugin = yield* PluginV2.Service
+      const aisdk = yield* AISDK.Service
+      const calls: string[] = []
+      yield* addPlugin()
+      yield* aisdk.runLanguage({
+        model: ModelV2.Info.make({
+          ...ModelV2.Info.empty(ProviderV2.ID.azure, ModelV2.ID.make("deepseek-v4-pro")),
+          api: { id: ModelV2.ID.make("deepseek-v4-pro"), type: "aisdk", package: "test-provider" },
+        }),
+        sdk: fakeSelectorSdk(calls),
+        options: {},
+      })
+      expect(calls).toEqual(["deepseek:deepseek-v4-pro"])
+    }),
+  )
+
+  it.effect("keeps the deepseek adapter when useCompletionUrls is set", () =>
+    Effect.gen(function* () {
+      const plugin = yield* PluginV2.Service
+      const aisdk = yield* AISDK.Service
+      const calls: string[] = []
+      yield* addPlugin()
+      yield* aisdk.runLanguage({
+        model: ModelV2.Info.make({
+          ...ModelV2.Info.empty(ProviderV2.ID.azure, ModelV2.ID.make("deepseek-v4-pro")),
+          api: { id: ModelV2.ID.make("deepseek-v4-pro"), type: "aisdk", package: "test-provider" },
+        }),
+        sdk: fakeSelectorSdk(calls),
+        options: { useCompletionUrls: true },
+      })
+      expect(calls).toEqual(["deepseek:deepseek-v4-pro"])
+    }),
+  )
+
+  it.effect("falls back to responses when the sdk has no deepseek adapter", () =>
+    Effect.gen(function* () {
+      const plugin = yield* PluginV2.Service
+      const aisdk = yield* AISDK.Service
+      const calls: string[] = []
+      const make = (method: string) => (id: string) => {
+        calls.push(`${method}:${id}`)
+        return { modelId: id, provider: method, specificationVersion: "v3" }
+      }
+      yield* addPlugin()
+      yield* aisdk.runLanguage({
+        model: ModelV2.Info.make({
+          ...ModelV2.Info.empty(ProviderV2.ID.azure, ModelV2.ID.make("deepseek-v4-pro")),
+          api: { id: ModelV2.ID.make("deepseek-v4-pro"), type: "aisdk", package: "test-provider" },
+        }),
+        sdk: { responses: make("responses"), chat: make("chat"), languageModel: make("languageModel") },
+        options: {},
+      })
+      expect(calls).toEqual(["responses:deepseek-v4-pro"])
+    }),
+  )
+
   it.effect("falls back through the legacy Azure selector order", () =>
     Effect.gen(function* () {
       const plugin = yield* PluginV2.Service

packages/opencode/src/provider/provider.ts

@@ -158,6 +158,14 @@ type CustomDep = {
 }
 
 function selectAzureLanguageModel(sdk: any, modelID: string, useChat: boolean) {
+  // Azure DeepSeek deployments need the DeepSeek-specific adapter for reasoning
+  // effort and reasoning_content across agentic turns; it is chat-completions
+  // based, so it also applies when useCompletionUrls is set.
+  // Note: this predicate (any "deepseek" deployment) deliberately differs from
+  // the "deepseek-v4" gate in transform.ts reasoningVariants, which only
+  // exposes a max effort for v4 models.
+  // Keep in sync with packages/core/src/plugin/provider/azure.ts selectLanguage.
+  if (modelID.toLowerCase().includes("deepseek") && sdk.deepseek) return sdk.deepseek(modelID)
   if (useChat && sdk.chat) return sdk.chat(modelID)
   if (sdk.responses) return sdk.responses(modelID)
   if (sdk.messages) return sdk.messages(modelID)

packages/opencode/src/provider/transform.ts

@@ -941,8 +941,13 @@ export function variants(model: Provider.Model): Record<string, Record<string, a
     case "@ai-sdk/azure":
       // https://v5.ai-sdk.dev/providers/ai-sdk-providers/azure
       if (id === "o1-mini") return {}
+      const azureEfforts = [...openaiReasoningEfforts(id, model.release_date)]
+      // Only deepseek-v4 deployments expose a max effort; the adapter
+      // selection in the azure loaders gates on any "deepseek" deployment,
+      // which is deliberately broader.
+      if (model.api.id.toLowerCase().includes("deepseek-v4") && !azureEfforts.includes("max")) azureEfforts.push("max")
       return Object.fromEntries(
-        openaiReasoningEfforts(id, model.release_date).map((effort) => [
+        azureEfforts.map((effort) => [
           effort,
           {
             reasoningEffort: effort,

packages/opencode/test/provider/provider.test.ts

@@ -809,6 +809,71 @@ it.instance("getSmallModel skips inferred models for Azure Cognitive Services",
   }),
 )
 
+const azureTestConfig = {
+  provider: {
+    azure: {
+      npm: "@ai-sdk/azure",
+      options: { resourceName: "test-resource", apiKey: "test-key" },
+      models: {
+        "deepseek-v4-pro": { name: "DeepSeek V4 Pro" },
+        "gpt-5": { name: "GPT-5" },
+      },
+    },
+  },
+}
+
+it.instance(
+  "azure deepseek deployments select the DeepSeek adapter",
+  Effect.gen(function* () {
+    yield* set("AZURE_RESOURCE_NAME", "test-resource")
+    yield* set("AZURE_API_KEY", "test-key")
+    const provider = yield* Provider.Service
+    const model = yield* provider.getModel(ProviderV2.ID.azure, ModelV2.ID.make("deepseek-v4-pro"))
+    const language = yield* provider.getLanguage(model)
+    expect(language.provider).toBe("azure.deepseek")
+  }),
+  { config: azureTestConfig },
+)
+
+it.instance(
+  "azure deepseek deployments keep the DeepSeek adapter with useCompletionUrls",
+  Effect.gen(function* () {
+    yield* set("AZURE_RESOURCE_NAME", "test-resource")
+    yield* set("AZURE_API_KEY", "test-key")
+    const provider = yield* Provider.Service
+    const model = yield* provider.getModel(ProviderV2.ID.azure, ModelV2.ID.make("deepseek-v4-pro"))
+    const language = yield* provider.getLanguage({ ...model, options: { ...model.options, useCompletionUrls: true } })
+    expect(language.provider).toBe("azure.deepseek")
+  }),
+  { config: azureTestConfig },
+)
+
+it.instance(
+  "azure non-deepseek deployments keep responses routing",
+  Effect.gen(function* () {
+    yield* set("AZURE_RESOURCE_NAME", "test-resource")
+    yield* set("AZURE_API_KEY", "test-key")
+    const provider = yield* Provider.Service
+    const model = yield* provider.getModel(ProviderV2.ID.azure, ModelV2.ID.make("gpt-5"))
+    const language = yield* provider.getLanguage(model)
+    expect(language.provider).toBe("azure.responses")
+  }),
+  { config: azureTestConfig },
+)
+
+it.instance(
+  "azure non-deepseek deployments keep chat routing with useCompletionUrls",
+  Effect.gen(function* () {
+    yield* set("AZURE_RESOURCE_NAME", "test-resource")
+    yield* set("AZURE_API_KEY", "test-key")
+    const provider = yield* Provider.Service
+    const model = yield* provider.getModel(ProviderV2.ID.azure, ModelV2.ID.make("gpt-5"))
+    const language = yield* provider.getLanguage({ ...model, options: { ...model.options, useCompletionUrls: true } })
+    expect(language.provider).toBe("azure.chat")
+  }),
+  { config: azureTestConfig },
+)
+
 it.instance(
   "getSmallModel respects config small_model override",
   Effect.gen(function* () {

packages/opencode/test/provider/transform.test.ts

@@ -3702,6 +3702,35 @@ describe("ProviderTransform.variants", () => {
     expect(result).toEqual({})
   })
 
+  test("azure deepseek-v4 deployments include a max reasoning effort variant", () => {
+    const model = createMockModel({
+      id: "azure/deepseek-v4-pro",
+      providerID: "azure",
+      api: {
+        id: "deepseek-v4-pro",
+        url: "https://azure.com",
+        npm: "@ai-sdk/azure",
+      },
+    })
+    const result = ProviderTransform.variants(model)
+    expect(Object.keys(result)).toEqual(["low", "medium", "high", "max"])
+    expect(result.max).toMatchObject({ reasoningEffort: "max" })
+  })
+
+  test("azure non-deepseek deployments do not include a max reasoning effort variant", () => {
+    const model = createMockModel({
+      id: "azure/gpt-5",
+      providerID: "azure",
+      api: {
+        id: "gpt-5",
+        url: "https://azure.com",
+        npm: "@ai-sdk/azure",
+      },
+    })
+    const result = ProviderTransform.variants(model)
+    expect(result.max).toBeUndefined()
+  })
+
   test("minimax returns empty object", () => {
     const model = createMockModel({
       id: "minimax/minimax-model",