feat(native-llm): route Anthropic API-key models through native runtime (#28271)

packages/opencode/test/session/llm-native-recorded.test.ts

@@ -1,7 +1,7 @@
 import { NodeFileSystem } from "@effect/platform-node"
 import { HttpRecorder, Redactor } from "@opencode-ai/http-recorder"
 import { describe, expect } from "bun:test"
-import { tool } from "ai"
+import { tool, type ModelMessage, type JSONValue } from "ai"
 import { Effect, Layer, Stream } from "effect"
 import { FetchHttpClient } from "effect/unstable/http"
 import path from "node:path"
@@ -12,6 +12,7 @@ import { Plugin } from "@/plugin"
 import { Provider } from "@/provider/provider"
 import { ModelID, ProviderID } from "@/provider/schema"
 import { Filesystem } from "@/util/filesystem"
+import { LLMEvent, LLMResponse } from "@opencode-ai/llm"
 import { LLMClient, RequestExecutor } from "@opencode-ai/llm/route"
 import { RuntimeFlags } from "@/effect/runtime-flags"
 import type { Agent } from "../../src/agent/agent"
@@ -22,22 +23,105 @@ import type { ModelsDev } from "@opencode-ai/core/models-dev"
 import { TestInstance } from "../fixture/fixture"
 import { testEffect } from "../lib/effect"
 
-const OPENAI_CASSETTE = "session/native-openai-tool-call"
-const ZEN_CASSETTE = "session/native-zen-tool-call"
 const FIXTURES_DIR = path.join(import.meta.dir, "../fixtures/recordings")
-const OPENAI_API_KEY = process.env.OPENCODE_RECORD_OPENAI_API_KEY ?? process.env.OPENAI_API_KEY
-const CONSOLE_TOKEN = process.env.OPENCODE_RECORD_CONSOLE_TOKEN
-const ZEN_ORG_ID = process.env.OPENCODE_RECORD_ZEN_ORG_ID
-const ZEN_API_URL =
-  process.env.OPENCODE_RECORD_ZEN_API_URL ?? "https://console.opencode.ai/proxy/connections/fixture/v1"
+
+const zenURL = (connection: string) => `https://console.opencode.ai/proxy/connections/${connection}/v1`
+
+type ProviderSpec = {
+  readonly providerID: ProviderID
+  readonly modelID: string
+  readonly cassette: string
+  readonly protocol: string
+  readonly tags: ReadonlyArray<string>
+  readonly canRecord: boolean
+  readonly config: (model: ModelsDev.Provider["models"][string]) => Partial<Config.Info>
+}
+
+const cloneModel = (model: ModelsDev.Provider["models"][string]) =>
+  structuredClone(model) as NonNullable<NonNullable<Config.Info["provider"]>[string]["models"]>[string]
+
+const PROVIDERS = {
+  openai: {
+    providerID: ProviderID.openai,
+    modelID: "gpt-4.1-mini",
+    cassette: "session/native-openai-tool-loop",
+    protocol: "openai-responses",
+    tags: ["opencode", "native", "tool-loop"],
+    canRecord: Boolean(process.env.OPENCODE_RECORD_OPENAI_API_KEY ?? process.env.OPENAI_API_KEY),
+    config: (model) => ({
+      enabled_providers: ["openai"],
+      provider: {
+        openai: {
+          name: "OpenAI",
+          env: ["OPENAI_API_KEY"],
+          npm: "@ai-sdk/openai",
+          api: "https://api.openai.com/v1",
+          models: { [model.id]: cloneModel(model) },
+          options: {
+            apiKey: process.env.OPENCODE_RECORD_OPENAI_API_KEY ?? process.env.OPENAI_API_KEY ?? "fixture-openai-key",
+            baseURL: "https://api.openai.com/v1",
+          },
+        },
+      },
+    }),
+  },
+  opencode: {
+    providerID: ProviderID.opencode,
+    modelID: "gpt-5.2-codex",
+    cassette: "session/native-zen-tool-loop",
+    protocol: "openai-responses",
+    tags: ["opencode", "zen", "native", "tool-loop"],
+    canRecord: Boolean(process.env.OPENCODE_RECORD_CONSOLE_TOKEN && process.env.OPENCODE_RECORD_ZEN_ORG_ID),
+    config: (model) => ({
+      enabled_providers: ["opencode"],
+      provider: {
+        opencode: {
+          name: "OpenCode Zen",
+          env: ["OPENCODE_CONSOLE_TOKEN"],
+          npm: "@ai-sdk/openai-compatible",
+          // The connection slug is account-specific; the cassette redactor
+          // normalizes it to {connection} for replay. Set during recording.
+          api: zenURL(process.env.OPENCODE_RECORD_ZEN_CONNECTION ?? "fixture"),
+          models: { [model.id]: cloneModel(model) },
+          options: {
+            apiKey: process.env.OPENCODE_RECORD_CONSOLE_TOKEN ?? "fixture-console-token",
+            headers: { "x-org-id": process.env.OPENCODE_RECORD_ZEN_ORG_ID ?? "fixture-org" },
+          },
+        },
+      },
+    }),
+  },
+  anthropic: {
+    providerID: ProviderID.anthropic,
+    modelID: "claude-haiku-4-5-20251001",
+    cassette: "session/native-anthropic-tool-loop",
+    protocol: "anthropic-messages",
+    tags: ["opencode", "native", "tool-loop"],
+    canRecord: Boolean(process.env.OPENCODE_RECORD_ANTHROPIC_API_KEY ?? process.env.ANTHROPIC_API_KEY),
+    config: (model) => ({
+      enabled_providers: ["anthropic"],
+      provider: {
+        anthropic: {
+          name: "Anthropic",
+          env: ["ANTHROPIC_API_KEY"],
+          npm: "@ai-sdk/anthropic",
+          api: "https://api.anthropic.com/v1",
+          models: { [model.id]: cloneModel(model) },
+          options: {
+            apiKey:
+              process.env.OPENCODE_RECORD_ANTHROPIC_API_KEY ?? process.env.ANTHROPIC_API_KEY ?? "fixture-anthropic-key",
+            baseURL: "https://api.anthropic.com/v1",
+          },
+        },
+      },
+    }),
+  },
+} satisfies Record<string, ProviderSpec>
 
 const shouldRecord = process.env.RECORD === "true"
-const canRunOpenAI = shouldRecord
-  ? Boolean(OPENAI_API_KEY)
-  : HttpRecorder.hasCassetteSync(OPENAI_CASSETTE, { directory: FIXTURES_DIR })
-const canRunZen = shouldRecord
-  ? Boolean(CONSOLE_TOKEN && ZEN_ORG_ID)
-  : HttpRecorder.hasCassetteSync(ZEN_CASSETTE, { directory: FIXTURES_DIR })
+
+const canRun = (spec: ProviderSpec) =>
+  shouldRecord ? spec.canRecord : HttpRecorder.hasCassetteSync(spec.cassette, { directory: FIXTURES_DIR })
 
 async function loadFixture(providerID: string, modelID: string) {
   const data = await Filesystem.readJson<Record<string, ModelsDev.Provider>>(
@@ -50,234 +134,140 @@ async function loadFixture(providerID: string, modelID: string) {
   return model
 }
 
-const openAIConfig = (model: ModelsDev.Provider["models"][string]): Partial<Config.Info> => ({
-  enabled_providers: ["openai"],
-  provider: {
-    openai: {
-      name: "OpenAI",
-      env: ["OPENAI_API_KEY"],
-      npm: "@ai-sdk/openai",
-      api: "https://api.openai.com/v1",
-      models: {
-        [model.id]: JSON.parse(JSON.stringify(model)) as NonNullable<
-          NonNullable<Config.Info["provider"]>[string]["models"]
-        >[string],
-      },
-      options: {
-        apiKey: OPENAI_API_KEY ?? "fixture-openai-key",
-        baseURL: "https://api.openai.com/v1",
-      },
-    },
-  },
-})
-
-const zenConfig = (model: ModelsDev.Provider["models"][string]): Partial<Config.Info> => ({
-  enabled_providers: ["opencode"],
-  provider: {
-    opencode: {
-      name: "OpenCode Zen",
-      env: ["OPENCODE_CONSOLE_TOKEN"],
-      npm: "@ai-sdk/openai-compatible",
-      api: ZEN_API_URL,
-      models: {
-        [model.id]: JSON.parse(JSON.stringify(model)) as NonNullable<
-          NonNullable<Config.Info["provider"]>[string]["models"]
-        >[string],
-      },
-      options: {
-        apiKey: CONSOLE_TOKEN ?? "fixture-console-token",
-        headers: {
-          "x-org-id": ZEN_ORG_ID ?? "fixture-org",
-        },
-      },
-    },
-  },
-})
-
-function recordedNativeLLMLayer(cassette: string, metadata: Record<string, unknown>) {
-  const cassetteService = HttpRecorder.Cassette.fileSystem({ directory: FIXTURES_DIR }).pipe(
-    Layer.provide(NodeFileSystem.layer),
-  )
+function recordedNativeLLMLayer(spec: ProviderSpec) {
   // Only the HTTP client is recorded; RequestExecutor and the opencode LLM stack remain real.
-  const recorder = HttpRecorder.recordingLayer(cassette, {
-    mode: shouldRecord ? "record" : "replay",
-    metadata,
-    redactor: Redactor.compose(
-      Redactor.defaults({
-        url: {
-          transform: (url) => url.replace(/\/proxy\/connections\/[^/]+\/v1/, "/proxy/connections/{connection}/v1"),
-        },
-      }),
-      {
-        response: (snapshot) => ({ ...snapshot, body: snapshot.body.replace(/wrk_[A-Z0-9]+/g, "wrk_redacted") }),
-      },
+  const recordedClient = LLMClient.layer.pipe(
+    Layer.provide(RequestExecutor.layer),
+    Layer.provide(
+      HttpRecorder.recordingLayer(spec.cassette, {
+        mode: shouldRecord ? "record" : "replay",
+        metadata: { provider: spec.providerID, protocol: spec.protocol, route: spec.protocol, tags: spec.tags },
+        redactor: Redactor.compose(
+          Redactor.defaults({
+            url: {
+              transform: (url) => url.replace(/\/proxy\/connections\/[^/]+\/v1/, "/proxy/connections/{connection}/v1"),
+            },
+          }),
+          {
+            response: (snapshot) => ({ ...snapshot, body: snapshot.body.replace(/wrk_[A-Z0-9]+/g, "wrk_redacted") }),
+          },
+        ),
+      }).pipe(Layer.provide(FetchHttpClient.layer)),
     ),
-  }).pipe(Layer.provide(FetchHttpClient.layer))
-  const executor = RequestExecutor.layer.pipe(Layer.provide(recorder))
-  const client = LLMClient.layer.pipe(Layer.provide(executor))
-
-  const providerLayer = Provider.defaultLayer.pipe(
-    Layer.provide(Auth.defaultLayer),
-    Layer.provide(Config.defaultLayer),
-    Layer.provide(Plugin.defaultLayer),
-  )
-  const llmLayer = LLM.layer.pipe(
-    Layer.provide(Auth.defaultLayer),
-    Layer.provide(Config.defaultLayer),
-    Layer.provide(Provider.defaultLayer),
-    Layer.provide(Plugin.defaultLayer),
-    Layer.provide(client),
-    Layer.provide(cassetteService),
-    Layer.provide(RuntimeFlags.layer({ experimentalNativeLlm: true })),
   )
 
-  return Layer.mergeAll(providerLayer, llmLayer)
+  return Layer.mergeAll(
+    Provider.defaultLayer.pipe(
+      Layer.provide(Auth.defaultLayer),
+      Layer.provide(Config.defaultLayer),
+      Layer.provide(Plugin.defaultLayer),
+    ),
+    LLM.layer.pipe(
+      Layer.provide(Auth.defaultLayer),
+      Layer.provide(Config.defaultLayer),
+      Layer.provide(Provider.defaultLayer),
+      Layer.provide(Plugin.defaultLayer),
+      Layer.provide(recordedClient),
+      Layer.provide(HttpRecorder.Cassette.fileSystem({ directory: FIXTURES_DIR }).pipe(Layer.provide(NodeFileSystem.layer))),
+      Layer.provide(RuntimeFlags.layer({ experimentalNativeLlm: true })),
+    ),
+  )
 }
 
-const openAIIt = testEffect(
-  recordedNativeLLMLayer(OPENAI_CASSETTE, {
-    provider: "openai",
-    protocol: "openai-responses",
-    route: "openai-responses",
-    tags: ["opencode", "native", "tool-call"],
-  }),
-)
-const zenIt = testEffect(
-  recordedNativeLLMLayer(ZEN_CASSETTE, {
-    provider: "opencode",
-    protocol: "openai-responses",
-    route: "openai-responses",
-    tags: ["opencode", "zen", "native", "tool-call"],
-  }),
-)
-const recordedOpenAIInstance = canRunOpenAI ? openAIIt.instance : openAIIt.instance.skip
-const recordedZenInstance = canRunZen ? zenIt.instance : zenIt.instance.skip
-
-const writeConfig = (
-  directory: string,
-  model: ModelsDev.Provider["models"][string],
-  config: (model: ModelsDev.Provider["models"][string]) => Partial<Config.Info> = openAIConfig,
-) =>
+const writeConfig = (directory: string, spec: ProviderSpec, model: ModelsDev.Provider["models"][string]) =>
   Effect.promise(() =>
     Bun.write(
       path.join(directory, "opencode.json"),
-      JSON.stringify({ $schema: "https://opencode.ai/config.json", ...config(model) }),
+      JSON.stringify({ $schema: "https://opencode.ai/config.json", ...spec.config(model) }),
     ),
   )
 
-const getModel = (providerID: ProviderID, modelID: ModelID) =>
-  Effect.gen(function* () {
-    const provider = yield* Provider.Service
-    return yield* provider.getModel(providerID, modelID)
-  })
-
 const collect = (input: LLM.StreamInput) =>
   Effect.gen(function* () {
     const llm = yield* LLM.Service
     return Array.from(yield* llm.stream(input).pipe(Stream.runCollect))
   })
 
-describe("session.llm native recorded", () => {
-  recordedOpenAIInstance("uses real RequestExecutor with HTTP recorder for native OpenAI tools", () =>
-    Effect.gen(function* () {
-      const test = yield* TestInstance
-      const model = yield* Effect.promise(() => loadFixture("openai", "gpt-4.1-mini"))
-      yield* writeConfig(test.directory, model)
+const WEATHER_RESULT = { temperature: 22, condition: "sunny" } as const
+const WEATHER_SYSTEM =
+  "Use the get_weather tool exactly once to look up Paris, then reply with exactly: Paris is sunny."
+const WEATHER_USER = "What is the weather in Paris?"
 
-      const sessionID = SessionID.make("session-recorded-native-tool")
-      const agent = {
-        name: "test",
-        mode: "primary",
-        prompt: "Call tools exactly as instructed.",
-        options: {},
-        permission: [{ permission: "*", pattern: "*", action: "allow" }],
-        temperature: 0,
-      } satisfies Agent.Info
-      const resolved = yield* getModel(ProviderID.openai, ModelID.make(model.id))
-      let executed: unknown
-
-      const events = yield* collect({
-        user: {
-          id: MessageID.make("msg_user-recorded-native-tool"),
-          sessionID,
-          role: "user",
-          time: { created: 0 },
-          agent: agent.name,
-          model: { providerID: ProviderID.make("openai"), modelID: ModelID.make(model.id) },
-        } satisfies MessageV2.User,
-        sessionID,
-        model: resolved,
-        agent,
-        system: ["You must call the lookup tool exactly once with query weather. Do not answer in text."],
-        messages: [{ role: "user", content: "Use lookup." }],
-        toolChoice: "required",
-        tools: {
-          lookup: tool({
-            description: "Lookup data.",
-            inputSchema: z.object({ query: z.string() }),
-            execute: async (args, options) => {
-              executed = { args, toolCallId: options.toolCallId }
-              return { output: "looked up" }
-            },
-          }),
-        },
-      })
-
-      expect(events.filter((event) => event.type === "step-finish")).toHaveLength(1)
-      expect(events.filter((event) => event.type === "finish")).toHaveLength(1)
-      expect(events.some((event) => event.type === "tool-result")).toBe(true)
-      expect(executed).toMatchObject({ args: { query: "weather" }, toolCallId: expect.any(String) })
-    }),
-  )
-
-  recordedZenInstance("uses console-managed Zen config with native OpenAI-compatible tools", () =>
-    Effect.gen(function* () {
-      const test = yield* TestInstance
-      const model = yield* Effect.promise(() => loadFixture("opencode", "gpt-5.2-codex"))
-      yield* writeConfig(test.directory, model, zenConfig)
-
-      const sessionID = SessionID.make("session-recorded-native-zen-tool")
-      const agent = {
-        name: "test",
-        mode: "primary",
-        prompt: "Call tools exactly as instructed.",
-        options: {},
-        permission: [{ permission: "*", pattern: "*", action: "allow" }],
-      } satisfies Agent.Info
-      const resolved = yield* getModel(ProviderID.opencode, ModelID.make(model.id))
-      let executed: unknown
-
-      const events = yield* collect({
-        user: {
-          id: MessageID.make("msg_user-recorded-native-zen-tool"),
-          sessionID,
-          role: "user",
-          time: { created: 0 },
-          agent: agent.name,
-          model: { providerID: ProviderID.opencode, modelID: ModelID.make(model.id) },
-        } satisfies MessageV2.User,
-        sessionID,
-        model: resolved,
-        agent,
-        system: ["You must call the lookup tool exactly once with query weather. Do not answer in text."],
-        messages: [{ role: "user", content: "Use lookup." }],
-        toolChoice: "required",
-        tools: {
-          lookup: tool({
-            description: "Lookup data.",
-            inputSchema: z.object({ query: z.string() }),
-            execute: async (args, options) => {
-              executed = { args, toolCallId: options.toolCallId }
-              return { output: "looked up" }
-            },
-          }),
-        },
-      })
-
-      expect(events.filter((event) => event.type === "step-finish")).toHaveLength(1)
-      expect(events.filter((event) => event.type === "finish")).toHaveLength(1)
-      expect(events.some((event) => event.type === "tool-result")).toBe(true)
-      expect(executed).toMatchObject({ args: { query: "weather" }, toolCallId: expect.any(String) })
-    }),
-  )
+const weatherTool = tool({
+  description: "Get the current weather for a city.",
+  inputSchema: z.object({ city: z.string() }),
+  execute: async () => WEATHER_RESULT,
+})
+
+const toolRoundtrip = (
+  call: { readonly id: string; readonly name: string; readonly input: unknown },
+  result: JSONValue,
+): ModelMessage[] => [
+  { role: "assistant", content: [{ type: "tool-call", toolCallId: call.id, toolName: call.name, input: call.input }] },
+  {
+    role: "tool",
+    content: [{ type: "tool-result", toolCallId: call.id, toolName: call.name, output: { type: "json", value: result } }],
+  },
+]
+
+const driveToolLoop = (spec: ProviderSpec) =>
+  Effect.gen(function* () {
+    const test = yield* TestInstance
+    const model = yield* Effect.promise(() => loadFixture(spec.providerID, spec.modelID))
+    yield* writeConfig(test.directory, spec, model)
+
+    const sessionID = SessionID.make(`session-recorded-${spec.providerID}-loop`)
+    const modelID = ModelID.make(model.id)
+    const agent = {
+      name: "test",
+      mode: "primary",
+      prompt: "Answer using tools when appropriate.",
+      options: {},
+      permission: [{ permission: "*", pattern: "*", action: "allow" }],
+      temperature: 0,
+    } satisfies Agent.Info
+    const provider = yield* Provider.Service
+    const resolved = yield* provider.getModel(spec.providerID, modelID)
+
+    const userMessage = { role: "user", content: WEATHER_USER } satisfies ModelMessage
+    const base = {
+      user: {
+        id: MessageID.make(`msg_user-recorded-${spec.providerID}-loop`),
+        sessionID,
+        role: "user",
+        time: { created: 0 },
+        agent: agent.name,
+        model: { providerID: spec.providerID, modelID },
+      } satisfies MessageV2.User,
+      sessionID,
+      model: resolved,
+      agent,
+      system: [WEATHER_SYSTEM],
+      tools: { get_weather: weatherTool },
+    }
+
+    const turn1 = yield* collect({ ...base, messages: [userMessage] })
+    const toolCall = turn1.find(LLMEvent.is.toolCall)
+    expect(toolCall).toBeDefined()
+    expect(turn1.find(LLMEvent.is.toolResult)).toBeDefined()
+    expect(toolCall!.name).toBe("get_weather")
+    expect(toolCall!.input).toMatchObject({ city: expect.stringMatching(/Paris/i) })
+    expect(turn1.filter(LLMEvent.is.stepFinish)).toHaveLength(1)
+
+    const turn2 = yield* collect({
+      ...base,
+      messages: [userMessage, ...toolRoundtrip(toolCall!, WEATHER_RESULT)],
+    })
+
+    expect(LLMResponse.text({ events: turn2 })).toMatch(/Paris is sunny/i)
+    expect(turn2.filter(LLMEvent.is.finish)).toHaveLength(1)
+    expect(turn2.filter(LLMEvent.is.toolCall)).toHaveLength(0)
+  })
+
+describe("session.llm native recorded", () => {
+  for (const [name, spec] of Object.entries(PROVIDERS)) {
+    const it = testEffect(recordedNativeLLMLayer(spec))
+    const instance = canRun(spec) ? it.instance : it.instance.skip
+    instance(`${name}: drives a tool loop to a final text answer`, () => driveToolLoop(spec))
+  }
 })

packages/opencode/test/session/llm-native.test.ts

@@ -262,11 +262,11 @@ describe("session.llm-native.request", () => {
     })
     expect(
       LLMNativeRuntime.status({
-        model: { ...baseModel, providerID: ProviderID.make("anthropic") },
-        provider: { ...providerInfo, id: ProviderID.make("anthropic") },
+        model: { ...baseModel, providerID: ProviderID.make("google") },
+        provider: { ...providerInfo, id: ProviderID.make("google") },
         auth: undefined,
       }),
-    ).toEqual({ type: "unsupported", reason: "provider is not openai or opencode" })
+    ).toEqual({ type: "unsupported", reason: "provider is not openai, opencode, or anthropic" })
     expect(
       LLMNativeRuntime.status({
         model: baseModel,
@@ -277,11 +277,11 @@ describe("session.llm-native.request", () => {
 
     expect(
       LLMNativeRuntime.status({
-        model: { ...baseModel, api: { ...baseModel.api, npm: "@ai-sdk/anthropic" } },
+        model: { ...baseModel, api: { ...baseModel.api, npm: "@ai-sdk/google" } },
         provider: providerInfo,
         auth: undefined,
       }),
-    ).toEqual({ type: "unsupported", reason: "provider package is not OpenAI" })
+    ).toEqual({ type: "unsupported", reason: "provider package is not OpenAI or Anthropic" })
 
     expect(
       LLMNativeRuntime.status({
@@ -289,7 +289,27 @@ describe("session.llm-native.request", () => {
         provider: { ...providerInfo, options: {} },
         auth: undefined,
       }),
-    ).toEqual({ type: "unsupported", reason: "OpenAI API key is not configured" })
+    ).toEqual({ type: "unsupported", reason: "API key is not configured" })
+  })
+
+  test("enables native runtime for Anthropic API-key models", () => {
+    expect(
+      LLMNativeRuntime.status({
+        model: {
+          ...baseModel,
+          providerID: ProviderID.make("anthropic"),
+          api: { ...baseModel.api, npm: "@ai-sdk/anthropic", url: "https://api.anthropic.com/v1" },
+        },
+        provider: {
+          ...providerInfo,
+          id: ProviderID.make("anthropic"),
+          name: "Anthropic",
+          env: ["ANTHROPIC_API_KEY"],
+          options: { apiKey: "test-anthropic-key" },
+        },
+        auth: undefined,
+      }),
+    ).toMatchObject({ type: "supported", apiKey: "test-anthropic-key" })
   })
 
   test("prefers console provider api key over stored opencode auth", () => {