Preview native LLM runtime stack (#27114)

packages/opencode/test/session/compaction.test.ts

@@ -30,6 +30,7 @@ import { TestConfig } from "../fixture/config"
 import { SyncEvent } from "@/sync"
 import { RuntimeFlags } from "@/effect/runtime-flags"
 import { EventV2Bridge } from "@/event-v2-bridge"
+import { LLMEvent, Usage } from "@opencode-ai/llm"
 
 void Log.init({ print: false })
 
@@ -47,6 +48,10 @@ const ref = {
   modelID: ModelID.make("test-model"),
 }
 
+const usage = (input: ConstructorParameters<typeof Usage>[0]) => new Usage(input)
+
+const basicUsage = () => usage({ inputTokens: 1, outputTokens: 1, totalTokens: 2 })
+
 afterEach(() => {
   mock.restore()
 })
@@ -296,11 +301,11 @@ function readCompactionPart(sessionID: SessionID) {
 
 function llm() {
   const queue: Array<
-    Stream.Stream<LLM.Event, unknown> | ((input: LLM.StreamInput) => Stream.Stream<LLM.Event, unknown>)
+    Stream.Stream<LLMEvent, unknown> | ((input: LLM.StreamInput) => Stream.Stream<LLMEvent, unknown>)
   > = []
 
   return {
-    push(stream: Stream.Stream<LLM.Event, unknown> | ((input: LLM.StreamInput) => Stream.Stream<LLM.Event, unknown>)) {
+    push(stream: Stream.Stream<LLMEvent, unknown> | ((input: LLM.StreamInput) => Stream.Stream<LLMEvent, unknown>)) {
       queue.push(stream)
     },
     layer: Layer.succeed(
@@ -319,54 +324,22 @@ function llm() {
 function reply(
   text: string,
   capture?: (input: LLM.StreamInput) => void,
-): (input: LLM.StreamInput) => Stream.Stream<LLM.Event, unknown> {
+): (input: LLM.StreamInput) => Stream.Stream<LLMEvent, unknown> {
   return (input) => {
     capture?.(input)
     return Stream.make(
-      { type: "start" } satisfies LLM.Event,
-      { type: "text-start", id: "txt-0" } satisfies LLM.Event,
-      { type: "text-delta", id: "txt-0", delta: text, text } as LLM.Event,
-      { type: "text-end", id: "txt-0" } satisfies LLM.Event,
-      {
-        type: "finish-step",
-        finishReason: "stop",
-        rawFinishReason: "stop",
-        response: { id: "res", modelId: "test-model", timestamp: new Date() },
-        providerMetadata: undefined,
-        usage: {
-          inputTokens: 1,
-          outputTokens: 1,
-          totalTokens: 2,
-          inputTokenDetails: {
-            noCacheTokens: undefined,
-            cacheReadTokens: undefined,
-            cacheWriteTokens: undefined,
-          },
-          outputTokenDetails: {
-            textTokens: undefined,
-            reasoningTokens: undefined,
-          },
-        },
-      } satisfies LLM.Event,
-      {
-        type: "finish",
-        finishReason: "stop",
-        rawFinishReason: "stop",
-        totalUsage: {
-          inputTokens: 1,
-          outputTokens: 1,
-          totalTokens: 2,
-          inputTokenDetails: {
-            noCacheTokens: undefined,
-            cacheReadTokens: undefined,
-            cacheWriteTokens: undefined,
-          },
-          outputTokenDetails: {
-            textTokens: undefined,
-            reasoningTokens: undefined,
-          },
-        },
-      } satisfies LLM.Event,
+      LLMEvent.textStart({ id: "txt-0" }),
+      LLMEvent.textDelta({ id: "txt-0", text }),
+      LLMEvent.textEnd({ id: "txt-0" }),
+      LLMEvent.stepFinish({
+        index: 0,
+        reason: "stop",
+        usage: basicUsage(),
+      }),
+      LLMEvent.finish({
+        reason: "stop",
+        usage: basicUsage(),
+      }),
     )
   }
 }
@@ -1204,7 +1177,7 @@ describe("session.compaction.process", () => {
         Stream.fromAsyncIterable(
           {
             async *[Symbol.asyncIterator]() {
-              yield { type: "start" } as LLM.Event
+              yield LLMEvent.stepStart({ index: 0 })
               throw new APICallError({
                 message: "boom",
                 url: "https://example.com/v1/chat/completions",
@@ -1290,55 +1263,62 @@ describe("session.compaction.process", () => {
     { git: true },
   )
 
+  itCompaction.instance(
+    "silently drops reasoning-delta arriving without prior reasoning-start",
+    () => {
+      // Regression: PR initially auto-created a reasoning Part for orphan deltas (no preceding
+      // reasoning-start). Reverted to match dev — drop silently. Pinned here so any future
+      // change to processor.ts reasoning-delta handling triggers this test.
+      const stub = llm()
+      stub.push(
+        Stream.make(
+          LLMEvent.reasoningDelta({ id: "orphan-1", text: "stray reasoning" }),
+          LLMEvent.textStart({ id: "txt-0" }),
+          LLMEvent.textDelta({ id: "txt-0", text: "summary" }),
+          LLMEvent.textEnd({ id: "txt-0" }),
+          LLMEvent.stepFinish({ index: 0, reason: "stop", usage: basicUsage() }),
+          LLMEvent.finish({ reason: "stop", usage: basicUsage() }),
+        ),
+      )
+      return Effect.gen(function* () {
+        const ssn = yield* SessionNs.Service
+        const session = yield* ssn.create({})
+        const msg = yield* createUserMessage(session.id, "hello")
+        const msgs = yield* ssn.messages({ sessionID: session.id })
+        yield* SessionCompaction.use.process({
+          parentID: msg.id,
+          messages: msgs,
+          sessionID: session.id,
+          auto: false,
+        })
+
+        const summary = (yield* ssn.messages({ sessionID: session.id })).find(
+          (item) => item.info.role === "assistant" && item.info.summary,
+        )
+        expect(summary?.parts.some((part) => part.type === "reasoning")).toBe(false)
+        // Sanity: the text part still got through.
+        expect(summary?.parts.some((part) => part.type === "text" && part.text === "summary")).toBe(true)
+      }).pipe(withCompaction({ llm: stub.layer }))
+    },
+    { git: true },
+  )
+
   itCompaction.instance(
     "does not allow tool calls while generating the summary",
     () => {
       const stub = llm()
       stub.push(
         Stream.make(
-          { type: "start" } satisfies LLM.Event,
-          { type: "tool-input-start", id: "call-1", toolName: "_noop" } satisfies LLM.Event,
-          { type: "tool-call", toolCallId: "call-1", toolName: "_noop", input: {} } satisfies LLM.Event,
-          {
-            type: "finish-step",
-            finishReason: "tool-calls",
-            rawFinishReason: "tool_calls",
-            response: { id: "res", modelId: "test-model", timestamp: new Date() },
-            providerMetadata: undefined,
-            usage: {
-              inputTokens: 1,
-              outputTokens: 1,
-              totalTokens: 2,
-              inputTokenDetails: {
-                noCacheTokens: undefined,
-                cacheReadTokens: undefined,
-                cacheWriteTokens: undefined,
-              },
-              outputTokenDetails: {
-                textTokens: undefined,
-                reasoningTokens: undefined,
-              },
-            },
-          } satisfies LLM.Event,
-          {
-            type: "finish",
-            finishReason: "tool-calls",
-            rawFinishReason: "tool_calls",
-            totalUsage: {
-              inputTokens: 1,
-              outputTokens: 1,
-              totalTokens: 2,
-              inputTokenDetails: {
-                noCacheTokens: undefined,
-                cacheReadTokens: undefined,
-                cacheWriteTokens: undefined,
-              },
-              outputTokenDetails: {
-                textTokens: undefined,
-                reasoningTokens: undefined,
-              },
-            },
-          } satisfies LLM.Event,
+          LLMEvent.toolCall({ id: "call-1", name: "_noop", input: {} }),
+          LLMEvent.stepFinish({
+            index: 0,
+            reason: "tool-calls",
+            usage: basicUsage(),
+          }),
+          LLMEvent.finish({
+            reason: "tool-calls",
+            usage: basicUsage(),
+          }),
         ),
       )
       return Effect.gen(function* () {
@@ -1544,20 +1524,7 @@ describe("SessionNs.getUsage", () => {
     const model = createModel({ context: 100_000, output: 32_000 })
     const result = SessionNs.getUsage({
       model,
-      usage: {
-        inputTokens: 1000,
-        outputTokens: 500,
-        totalTokens: 1500,
-        inputTokenDetails: {
-          noCacheTokens: undefined,
-          cacheReadTokens: undefined,
-          cacheWriteTokens: undefined,
-        },
-        outputTokenDetails: {
-          textTokens: undefined,
-          reasoningTokens: undefined,
-        },
-      },
+      usage: usage({ inputTokens: 1000, outputTokens: 500, totalTokens: 1500 }),
     })
 
     expect(result.tokens.input).toBe(1000)
@@ -1571,20 +1538,7 @@ describe("SessionNs.getUsage", () => {
     const model = createModel({ context: 100_000, output: 32_000 })
     const result = SessionNs.getUsage({
       model,
-      usage: {
-        inputTokens: 1000,
-        outputTokens: 500,
-        totalTokens: 1500,
-        inputTokenDetails: {
-          noCacheTokens: 800,
-          cacheReadTokens: 200,
-          cacheWriteTokens: undefined,
-        },
-        outputTokenDetails: {
-          textTokens: undefined,
-          reasoningTokens: undefined,
-        },
-      },
+      usage: usage({ inputTokens: 1000, outputTokens: 500, totalTokens: 1500, cacheReadInputTokens: 200 }),
     })
 
     expect(result.tokens.input).toBe(800)
@@ -1595,20 +1549,7 @@ describe("SessionNs.getUsage", () => {
     const model = createModel({ context: 100_000, output: 32_000 })
     const result = SessionNs.getUsage({
       model,
-      usage: {
-        inputTokens: 1000,
-        outputTokens: 500,
-        totalTokens: 1500,
-        inputTokenDetails: {
-          noCacheTokens: undefined,
-          cacheReadTokens: undefined,
-          cacheWriteTokens: undefined,
-        },
-        outputTokenDetails: {
-          textTokens: undefined,
-          reasoningTokens: undefined,
-        },
-      },
+      usage: usage({ inputTokens: 1000, outputTokens: 500, totalTokens: 1500 }),
       metadata: {
         anthropic: {
           cacheCreationInputTokens: 300,
@@ -1624,20 +1565,7 @@ describe("SessionNs.getUsage", () => {
     // AI SDK v6 normalizes inputTokens to include cached tokens for all providers
     const result = SessionNs.getUsage({
       model,
-      usage: {
-        inputTokens: 1000,
-        outputTokens: 500,
-        totalTokens: 1500,
-        inputTokenDetails: {
-          noCacheTokens: 800,
-          cacheReadTokens: 200,
-          cacheWriteTokens: undefined,
-        },
-        outputTokenDetails: {
-          textTokens: undefined,
-          reasoningTokens: undefined,
-        },
-      },
+      usage: usage({ inputTokens: 1000, outputTokens: 500, totalTokens: 1500, cacheReadInputTokens: 200 }),
       metadata: {
         anthropic: {},
       },
@@ -1651,20 +1579,7 @@ describe("SessionNs.getUsage", () => {
     const model = createModel({ context: 100_000, output: 32_000 })
     const result = SessionNs.getUsage({
       model,
-      usage: {
-        inputTokens: 1000,
-        outputTokens: 500,
-        totalTokens: 1500,
-        inputTokenDetails: {
-          noCacheTokens: undefined,
-          cacheReadTokens: undefined,
-          cacheWriteTokens: undefined,
-        },
-        outputTokenDetails: {
-          textTokens: 400,
-          reasoningTokens: 100,
-        },
-      },
+      usage: usage({ inputTokens: 1000, outputTokens: 500, reasoningTokens: 100, totalTokens: 1500 }),
     })
 
     expect(result.tokens.input).toBe(1000)
@@ -1685,20 +1600,7 @@ describe("SessionNs.getUsage", () => {
     })
     const result = SessionNs.getUsage({
       model,
-      usage: {
-        inputTokens: 0,
-        outputTokens: 1_000_000,
-        totalTokens: 1_000_000,
-        inputTokenDetails: {
-          noCacheTokens: undefined,
-          cacheReadTokens: undefined,
-          cacheWriteTokens: undefined,
-        },
-        outputTokenDetails: {
-          textTokens: 750_000,
-          reasoningTokens: 250_000,
-        },
-      },
+      usage: usage({ inputTokens: 0, outputTokens: 1_000_000, reasoningTokens: 250_000, totalTokens: 1_000_000 }),
     })
 
     expect(result.tokens.output).toBe(750_000)
@@ -1710,20 +1612,7 @@ describe("SessionNs.getUsage", () => {
     const model = createModel({ context: 100_000, output: 32_000 })
     const result = SessionNs.getUsage({
       model,
-      usage: {
-        inputTokens: 0,
-        outputTokens: 0,
-        totalTokens: 0,
-        inputTokenDetails: {
-          noCacheTokens: undefined,
-          cacheReadTokens: undefined,
-          cacheWriteTokens: undefined,
-        },
-        outputTokenDetails: {
-          textTokens: undefined,
-          reasoningTokens: undefined,
-        },
-      },
+      usage: usage({ inputTokens: 0, outputTokens: 0, totalTokens: 0 }),
     })
 
     expect(result.tokens.input).toBe(0)
@@ -1746,20 +1635,7 @@ describe("SessionNs.getUsage", () => {
     })
     const result = SessionNs.getUsage({
       model,
-      usage: {
-        inputTokens: 1_000_000,
-        outputTokens: 100_000,
-        totalTokens: 1_100_000,
-        inputTokenDetails: {
-          noCacheTokens: undefined,
-          cacheReadTokens: undefined,
-          cacheWriteTokens: undefined,
-        },
-        outputTokenDetails: {
-          textTokens: undefined,
-          reasoningTokens: undefined,
-        },
-      },
+      usage: usage({ inputTokens: 1_000_000, outputTokens: 100_000, totalTokens: 1_100_000 }),
     })
 
     expect(result.cost).toBe(3 + 1.5)
@@ -1796,20 +1672,12 @@ describe("SessionNs.getUsage", () => {
     })
     const result = SessionNs.getUsage({
       model,
-      usage: {
+      usage: usage({
         inputTokens: 650_000,
         outputTokens: 100_000,
         totalTokens: 750_000,
-        inputTokenDetails: {
-          noCacheTokens: undefined,
-          cacheReadTokens: 100_000,
-          cacheWriteTokens: undefined,
-        },
-        outputTokenDetails: {
-          textTokens: undefined,
-          reasoningTokens: undefined,
-        },
-      },
+        cacheReadInputTokens: 100_000,
+      }),
     })
 
     expect(result.tokens.input).toBe(550_000)
@@ -1841,20 +1709,7 @@ describe("SessionNs.getUsage", () => {
     })
     const result = SessionNs.getUsage({
       model,
-      usage: {
-        inputTokens: 300_000,
-        outputTokens: 100_000,
-        totalTokens: 400_000,
-        inputTokenDetails: {
-          noCacheTokens: undefined,
-          cacheReadTokens: undefined,
-          cacheWriteTokens: undefined,
-        },
-        outputTokenDetails: {
-          textTokens: undefined,
-          reasoningTokens: undefined,
-        },
-      },
+      usage: usage({ inputTokens: 300_000, outputTokens: 100_000, totalTokens: 400_000 }),
     })
 
     expect(result.cost).toBe(0.9 + 0.4)
@@ -1865,24 +1720,16 @@ describe("SessionNs.getUsage", () => {
     (npm) => {
       const model = createModel({ context: 100_000, output: 32_000, npm })
       // AI SDK v6: inputTokens includes cached tokens for all providers
-      const usage = {
+      const item = usage({
         inputTokens: 1000,
         outputTokens: 500,
         totalTokens: 1500,
-        inputTokenDetails: {
-          noCacheTokens: 800,
-          cacheReadTokens: 200,
-          cacheWriteTokens: undefined,
-        },
-        outputTokenDetails: {
-          textTokens: undefined,
-          reasoningTokens: undefined,
-        },
-      }
+        cacheReadInputTokens: 200,
+      })
       if (npm === "@ai-sdk/amazon-bedrock") {
         const result = SessionNs.getUsage({
           model,
-          usage,
+          usage: item,
           metadata: {
             bedrock: {
               usage: {
@@ -1903,7 +1750,7 @@ describe("SessionNs.getUsage", () => {
 
       const result = SessionNs.getUsage({
         model,
-        usage,
+        usage: item,
         metadata: {
           anthropic: {
             cacheCreationInputTokens: 300,
@@ -1924,20 +1771,7 @@ describe("SessionNs.getUsage", () => {
     const model = createModel({ context: 100_000, output: 32_000, npm: "@ai-sdk/google-vertex/anthropic" })
     const result = SessionNs.getUsage({
       model,
-      usage: {
-        inputTokens: 1000,
-        outputTokens: 500,
-        totalTokens: 1500,
-        inputTokenDetails: {
-          noCacheTokens: 800,
-          cacheReadTokens: 200,
-          cacheWriteTokens: undefined,
-        },
-        outputTokenDetails: {
-          textTokens: undefined,
-          reasoningTokens: undefined,
-        },
-      },
+      usage: usage({ inputTokens: 1000, outputTokens: 500, totalTokens: 1500, cacheReadInputTokens: 200 }),
       metadata: {
         vertex: {
           cacheCreationInputTokens: 300,

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

@@ -0,0 +1,283 @@
+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 { Effect, Layer, Stream } from "effect"
+import { FetchHttpClient } from "effect/unstable/http"
+import path from "node:path"
+import z from "zod"
+import { Auth } from "@/auth"
+import { Config } from "@/config/config"
+import { Plugin } from "@/plugin"
+import { Provider } from "@/provider/provider"
+import { ModelID, ProviderID } from "@/provider/schema"
+import { Filesystem } from "@/util/filesystem"
+import { LLMClient, RequestExecutor } from "@opencode-ai/llm/route"
+import { RuntimeFlags } from "@/effect/runtime-flags"
+import type { Agent } from "../../src/agent/agent"
+import { LLM } from "../../src/session/llm"
+import { MessageV2 } from "../../src/session/message-v2"
+import { MessageID, SessionID } from "../../src/session/schema"
+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 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 })
+
+async function loadFixture(providerID: string, modelID: string) {
+  const data = await Filesystem.readJson<Record<string, ModelsDev.Provider>>(
+    path.join(import.meta.dir, "../tool/fixtures/models-api.json"),
+  )
+  const provider = data[providerID]
+  if (!provider) throw new Error(`Missing provider in fixture: ${providerID}`)
+  const model = provider.models[modelID]
+  if (!model) throw new Error(`Missing model in fixture: ${modelID}`)
+  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),
+  )
+  // 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") }),
+      },
+    ),
+  }).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)
+}
+
+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,
+) =>
+  Effect.promise(() =>
+    Bun.write(
+      path.join(directory, "opencode.json"),
+      JSON.stringify({ $schema: "https://opencode.ai/config.json", ...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 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) })
+    }),
+  )
+})

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

@@ -0,0 +1,357 @@
+import { describe, expect, test } from "bun:test"
+import { ToolFailure } from "@opencode-ai/llm"
+import { LLMClient, RequestExecutor } from "@opencode-ai/llm/route"
+import { jsonSchema, tool, type ModelMessage } from "ai"
+import { Effect } from "effect"
+import { LLMNative } from "@/session/llm/native-request"
+import { LLMNativeRuntime } from "@/session/llm/native-runtime"
+import type { Provider } from "@/provider/provider"
+import { ModelID, ProviderID } from "@/provider/schema"
+
+const baseModel: Provider.Model = {
+  id: ModelID.make("gpt-5-mini"),
+  providerID: ProviderID.make("openai"),
+  api: {
+    id: "gpt-5-mini",
+    url: "https://api.openai.com/v1",
+    npm: "@ai-sdk/openai",
+  },
+  name: "GPT-5 Mini",
+  capabilities: {
+    temperature: true,
+    reasoning: true,
+    attachment: true,
+    toolcall: true,
+    input: {
+      text: true,
+      audio: false,
+      image: true,
+      video: false,
+      pdf: false,
+    },
+    output: {
+      text: true,
+      audio: false,
+      image: false,
+      video: false,
+      pdf: false,
+    },
+    interleaved: false,
+  },
+  cost: {
+    input: 0,
+    output: 0,
+    cache: {
+      read: 0,
+      write: 0,
+    },
+  },
+  limit: {
+    context: 128_000,
+    input: 128_000,
+    output: 32_000,
+  },
+  status: "active",
+  options: {},
+  headers: {
+    "x-model": "model-header",
+  },
+  release_date: "2026-01-01",
+}
+
+const providerInfo: Provider.Info = {
+  id: ProviderID.make("openai"),
+  name: "OpenAI",
+  source: "config",
+  env: ["OPENAI_API_KEY"],
+  options: { apiKey: "test-openai-key" },
+  models: {},
+}
+
+describe("session.llm-native.request", () => {
+  test("maps normalized stream inputs to a native LLM request", () => {
+    const messages: ModelMessage[] = [
+      {
+        role: "system",
+        content: "system from messages",
+      },
+      {
+        role: "user",
+        content: [
+          { type: "text", text: "hello", providerOptions: { openai: { cacheControl: { type: "ephemeral" } } } },
+          { type: "file", mediaType: "image/png", filename: "img.png", data: "data:image/png;base64,Zm9v" },
+        ],
+      },
+      {
+        role: "assistant",
+        content: [
+          { type: "reasoning", text: "thinking", providerOptions: { openai: { encryptedContent: "secret" } } },
+          { type: "text", text: "I'll run it" },
+          {
+            type: "tool-call",
+            toolCallId: "call-1",
+            toolName: "bash",
+            input: { command: "ls" },
+            providerOptions: { openai: { itemId: "item-1" } },
+          },
+        ],
+      },
+      {
+        role: "tool",
+        content: [
+          {
+            type: "tool-result",
+            toolCallId: "call-1",
+            toolName: "bash",
+            output: { type: "text", value: "ok" },
+            providerOptions: { openai: { outputId: "output-1" } },
+          },
+        ],
+      },
+    ]
+
+    const request = LLMNative.request({
+      model: baseModel,
+      system: ["agent system"],
+      messages,
+      tools: {
+        bash: tool({
+          description: "Run a shell command",
+          inputSchema: jsonSchema({
+            type: "object",
+            properties: {
+              command: { type: "string" },
+            },
+            required: ["command"],
+          }),
+        }),
+      },
+      toolChoice: "required",
+      temperature: 0.2,
+      topP: 0.9,
+      topK: 40,
+      maxOutputTokens: 1024,
+      providerOptions: { openai: { store: false } },
+      headers: { "x-request": "request-header" },
+    })
+
+    expect(request.model).toMatchObject({
+      id: "gpt-5-mini",
+      provider: "openai",
+      route: "openai-responses",
+      baseURL: "https://api.openai.com/v1",
+      headers: {
+        "x-model": "model-header",
+        "x-request": "request-header",
+      },
+      limits: {
+        context: 128_000,
+        output: 32_000,
+      },
+    })
+    expect(request.system).toEqual([
+      { type: "text", text: "agent system" },
+      { type: "text", text: "system from messages" },
+    ])
+    expect(request.generation).toMatchObject({
+      temperature: 0.2,
+      topP: 0.9,
+      topK: 40,
+      maxTokens: 1024,
+    })
+    expect(request.providerOptions).toEqual({ openai: { store: false } })
+    expect(request.toolChoice).toMatchObject({ type: "required" })
+    expect(request.tools).toMatchObject([
+      {
+        name: "bash",
+        description: "Run a shell command",
+        inputSchema: {
+          type: "object",
+          properties: {
+            command: { type: "string" },
+          },
+          required: ["command"],
+        },
+      },
+    ])
+    expect(request.messages).toMatchObject([
+      {
+        role: "user",
+        content: [
+          { type: "text", text: "hello", providerMetadata: { openai: { cacheControl: { type: "ephemeral" } } } },
+          { type: "media", mediaType: "image/png", filename: "img.png", data: "data:image/png;base64,Zm9v" },
+        ],
+      },
+      {
+        role: "assistant",
+        content: [
+          { type: "reasoning", text: "thinking", providerMetadata: { openai: { encryptedContent: "secret" } } },
+          { type: "text", text: "I'll run it" },
+          {
+            type: "tool-call",
+            id: "call-1",
+            name: "bash",
+            input: { command: "ls" },
+            providerMetadata: { openai: { itemId: "item-1" } },
+          },
+        ],
+      },
+      {
+        role: "tool",
+        content: [
+          {
+            type: "tool-result",
+            id: "call-1",
+            name: "bash",
+            result: { type: "text", value: "ok" },
+            providerMetadata: { openai: { outputId: "output-1" } },
+          },
+        ],
+      },
+    ])
+  })
+
+  test("selects native routes from existing provider packages", () => {
+    expect(
+      LLMNative.model({ ...baseModel, api: { ...baseModel.api, url: "", npm: "@ai-sdk/anthropic" } }),
+    ).toMatchObject({
+      route: "anthropic-messages",
+      baseURL: "https://api.anthropic.com/v1",
+    })
+    expect(LLMNative.model({ ...baseModel, api: { ...baseModel.api, url: "", npm: "@ai-sdk/google" } })).toMatchObject({
+      route: "gemini",
+      baseURL: "https://generativelanguage.googleapis.com/v1beta",
+    })
+    expect(
+      LLMNative.model({ ...baseModel, api: { ...baseModel.api, npm: "@ai-sdk/openai-compatible" } }),
+    ).toMatchObject({
+      route: "openai-compatible-chat",
+      baseURL: "https://api.openai.com/v1",
+    })
+    expect(
+      LLMNative.model({ ...baseModel, api: { ...baseModel.api, url: "", npm: "@openrouter/ai-sdk-provider" } }),
+    ).toMatchObject({
+      route: "openrouter",
+      baseURL: "https://openrouter.ai/api/v1",
+    })
+  })
+
+  test("fails fast for unsupported provider packages", () => {
+    expect(() =>
+      LLMNative.request({
+        model: { ...baseModel, api: { ...baseModel.api, npm: "unknown-provider" } },
+        messages: [],
+      }),
+    ).toThrow("Native LLM request adapter does not support provider package unknown-provider")
+  })
+
+  test("only enables native runtime for supported OpenAI API-key models", () => {
+    expect(LLMNativeRuntime.status({ model: baseModel, provider: providerInfo, auth: undefined })).toMatchObject({
+      type: "supported",
+      apiKey: "test-openai-key",
+    })
+    expect(
+      LLMNativeRuntime.status({
+        model: { ...baseModel, providerID: ProviderID.make("opencode") },
+        provider: { ...providerInfo, id: ProviderID.make("opencode") },
+        auth: undefined,
+      }),
+    ).toMatchObject({
+      type: "supported",
+      apiKey: "test-openai-key",
+    })
+    expect(
+      LLMNativeRuntime.status({
+        model: { ...baseModel, providerID: ProviderID.make("anthropic") },
+        provider: { ...providerInfo, id: ProviderID.make("anthropic") },
+        auth: undefined,
+      }),
+    ).toEqual({ type: "unsupported", reason: "provider is not openai or opencode" })
+    expect(
+      LLMNativeRuntime.status({
+        model: baseModel,
+        provider: providerInfo,
+        auth: { type: "oauth", refresh: "refresh", access: "access", expires: 1 },
+      }),
+    ).toEqual({ type: "unsupported", reason: "OAuth auth is not supported" })
+
+    expect(
+      LLMNativeRuntime.status({
+        model: { ...baseModel, api: { ...baseModel.api, npm: "@ai-sdk/anthropic" } },
+        provider: providerInfo,
+        auth: undefined,
+      }),
+    ).toEqual({ type: "unsupported", reason: "provider package is not OpenAI" })
+
+    expect(
+      LLMNativeRuntime.status({
+        model: baseModel,
+        provider: { ...providerInfo, options: {} },
+        auth: undefined,
+      }),
+    ).toEqual({ type: "unsupported", reason: "OpenAI API key is not configured" })
+  })
+
+  test("native tool wrapper converts thrown errors into typed ToolFailure", async () => {
+    const wrapped = LLMNativeRuntime.nativeTools(
+      {
+        explode: {
+          description: "always throws",
+          inputSchema: jsonSchema({ type: "object" }),
+          execute: async () => {
+            throw new Error("boom")
+          },
+        } as any,
+      },
+      { messages: [] as ModelMessage[], abort: new AbortController().signal },
+    )
+
+    const failure = await Effect.runPromise(
+      Effect.flip(wrapped.explode!.execute!({}, { id: "call-1", name: "explode" })),
+    )
+    expect(failure).toBeInstanceOf(ToolFailure)
+    expect((failure as ToolFailure).message).toBe("boom")
+  })
+
+  test("native tool wrapper raises ToolFailure when the source tool has no execute handler", async () => {
+    // The AI SDK Tool shape allows execute to be omitted (e.g., client-side / MCP tools).
+    // The native runtime owns execution, so encountering such a tool here means upstream
+    // wiring is wrong; we want a typed failure, not a silent skip or unhandled exception.
+    const wrapped = LLMNativeRuntime.nativeTools(
+      { incomplete: { description: "no execute", inputSchema: jsonSchema({ type: "object" }) } as any },
+      { messages: [] as ModelMessage[], abort: new AbortController().signal },
+    )
+
+    const failure = await Effect.runPromise(
+      Effect.flip(wrapped.incomplete!.execute!({}, { id: "call-1", name: "incomplete" })),
+    )
+    expect(failure).toBeInstanceOf(ToolFailure)
+    expect((failure as ToolFailure).message).toContain("incomplete")
+  })
+
+  test("compiles through the native OpenAI Responses route", async () => {
+    const prepared = await Effect.runPromise(
+      LLMClient.prepare(
+        LLMNative.request({
+          model: baseModel,
+          messages: [{ role: "user", content: "hello" }],
+          providerOptions: { openai: { store: false } },
+          maxOutputTokens: 512,
+          headers: { "x-request": "request-header" },
+        }),
+      ).pipe(Effect.provide(LLMClient.layer), Effect.provide(RequestExecutor.defaultLayer)),
+    )
+
+    expect(prepared).toMatchObject({
+      route: "openai-responses",
+      protocol: "openai-responses",
+      body: {
+        model: "gpt-5-mini",
+        input: [{ role: "user", content: [{ type: "input_text", text: "hello" }] }],
+        max_output_tokens: 512,
+        store: false,
+        stream: true,
+      },
+    })
+  })
+})

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

@@ -1,15 +1,20 @@
 import { afterAll, beforeAll, beforeEach, describe, expect, test } from "bun:test"
 import path from "path"
 import { tool, type ModelMessage } from "ai"
-import { Cause, Effect, Exit, Stream } from "effect"
+import { Cause, Effect, Exit, Layer, Stream } from "effect"
+import { HttpClientRequest, HttpClientResponse } from "effect/unstable/http"
 import z from "zod"
 import { makeRuntime } from "../../src/effect/run-service"
 import { InstanceRef } from "../../src/effect/instance-ref"
 import { LLM } from "../../src/session/llm"
 import type { InstanceContext } from "../../src/project/instance-context"
+import { LLMClient, RequestExecutor } from "@opencode-ai/llm/route"
+import { Auth } from "@/auth"
+import { Config } from "@/config/config"
 import { Provider } from "@/provider/provider"
 import { ProviderTransform } from "@/provider/transform"
 import { ModelsDev } from "@opencode-ai/core/models-dev"
+import { Plugin } from "@/plugin"
 import { ProviderID, ModelID } from "../../src/provider/schema"
 import { Filesystem } from "@/util/filesystem"
 import { tmpdir, withTestInstance } from "../fixture/fixture"
@@ -17,6 +22,33 @@ import type { Agent } from "../../src/agent/agent"
 import { MessageV2 } from "../../src/session/message-v2"
 import { SessionID, MessageID } from "../../src/session/schema"
 import { AppRuntime } from "../../src/effect/app-runtime"
+import { RuntimeFlags } from "@/effect/runtime-flags"
+import { Permission } from "@/permission"
+import { LLMAISDK } from "@/session/llm/ai-sdk"
+import { Session as SessionNs } from "@/session/session"
+
+const openAIConfig = (model: ModelsDev.Provider["models"][string], baseURL: string): Partial<Config.Info> => {
+  const { experimental: _experimental, ...configModel } = model
+  type ConfigModel = NonNullable<NonNullable<Config.Info["provider"]>[string]["models"]>[string]
+  return {
+    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(configModel)) as ConfigModel,
+        },
+        options: {
+          apiKey: "test-openai-key",
+          baseURL,
+        },
+      },
+    },
+  }
+}
 
 async function getModel(providerID: ProviderID, modelID: ModelID, ctx: InstanceContext) {
   const effect = Effect.gen(function* () {
@@ -35,6 +67,26 @@ async function drain(input: LLM.StreamInput, ctx: InstanceContext) {
   })
 }
 
+async function drainWith(layer: Layer.Layer<LLM.Service>, input: LLM.StreamInput, ctx: InstanceContext) {
+  return Effect.runPromise(
+    LLM.Service.use((svc) => svc.stream(input).pipe(Stream.runDrain)).pipe(
+      Effect.provide(layer),
+      Effect.provideService(InstanceRef, ctx),
+    ),
+  )
+}
+
+function llmLayerWithExecutor(executor: Layer.Layer<RequestExecutor.Service>, flags: Partial<RuntimeFlags.Info> = {}) {
+  return LLM.layer.pipe(
+    Layer.provide(Auth.defaultLayer),
+    Layer.provide(Config.defaultLayer),
+    Layer.provide(Provider.defaultLayer),
+    Layer.provide(Plugin.defaultLayer),
+    Layer.provide(LLMClient.layer.pipe(Layer.provide(executor))),
+    Layer.provide(RuntimeFlags.layer(flags)),
+  )
+}
+
 describe("session.llm.hasToolCalls", () => {
   test("returns false for empty messages array", () => {
     expect(LLM.hasToolCalls([])).toBe(false)
@@ -122,6 +174,338 @@ describe("session.llm.hasToolCalls", () => {
   })
 })
 
+describe("session.llm.ai-sdk adapter", () => {
+  type AISDKAdapterEvent = Parameters<typeof LLMAISDK.toLLMEvents>[1]
+
+  const adapt = (events: ReadonlyArray<AISDKAdapterEvent>) => {
+    const state = LLMAISDK.adapterState()
+    return Effect.runPromise(
+      Effect.forEach(events, (event) => LLMAISDK.toLLMEvents(state, event)).pipe(Effect.map((items) => items.flat())),
+    )
+  }
+  // oxlint-disable-next-line typescript-eslint/no-unsafe-type-assertion -- tests defensive adapter branches outside AI SDK's current typed surface
+  const uncheckedAdapterEvent = (input: unknown) => input as AISDKAdapterEvent
+
+  test("maps AI SDK stream chunks without losing session-visible fields", async () => {
+    const metadata = { openai: { itemID: "item-1" } }
+    const events = await adapt([
+      { type: "start" },
+      { type: "start-step", request: {}, warnings: [] },
+      { type: "text-start", id: "text-1", providerMetadata: metadata },
+      { type: "text-delta", id: "text-1", text: "Hel", providerMetadata: { openai: { delta: 1 } } },
+      { type: "text-delta", id: "text-1", text: "lo", providerMetadata: { openai: { delta: 2 } } },
+      { type: "text-end", id: "text-1", providerMetadata: { openai: { done: true } } },
+      { type: "reasoning-start", id: "reasoning-1", providerMetadata: metadata },
+      { type: "reasoning-delta", id: "reasoning-1", text: "Think", providerMetadata: { openai: { delta: 3 } } },
+      { type: "reasoning-end", id: "reasoning-1", providerMetadata: { openai: { done: true } } },
+      { type: "tool-input-start", id: "call-1", toolName: "lookup", providerMetadata: metadata },
+      { type: "tool-input-delta", id: "call-1", delta: '{"query":' },
+      { type: "tool-input-delta", id: "call-1", delta: '"weather"}' },
+      { type: "tool-input-end", id: "call-1", providerMetadata: { openai: { inputDone: true } } },
+      {
+        type: "tool-call",
+        toolCallId: "call-1",
+        toolName: "lookup",
+        input: { query: "weather" },
+        providerExecuted: true,
+        providerMetadata: { openai: { called: true } },
+      },
+      {
+        type: "tool-result",
+        toolCallId: "call-1",
+        toolName: "lookup",
+        input: { query: "weather" },
+        output: { title: "Lookup", output: "sunny", metadata: { ok: true } },
+        providerExecuted: true,
+        providerMetadata: { openai: { result: true } },
+      },
+      {
+        type: "finish-step",
+        response: { id: "response-1", timestamp: new Date(0), modelId: "gpt-test" },
+        finishReason: "other",
+        rawFinishReason: "other",
+        usage: {
+          inputTokens: 10,
+          outputTokens: 5,
+          totalTokens: 15,
+          inputTokenDetails: { noCacheTokens: 5, cacheReadTokens: 3, cacheWriteTokens: 2 },
+          outputTokenDetails: { textTokens: 4, reasoningTokens: 1 },
+        },
+        providerMetadata: { openai: { step: true } },
+      },
+      {
+        type: "finish",
+        finishReason: "other",
+        rawFinishReason: "other",
+        totalUsage: {
+          inputTokens: 11,
+          outputTokens: 6,
+          totalTokens: 17,
+          cachedInputTokens: 4,
+          reasoningTokens: 2,
+          inputTokenDetails: { noCacheTokens: 7, cacheReadTokens: 4, cacheWriteTokens: undefined },
+          outputTokenDetails: { textTokens: 4, reasoningTokens: 2 },
+        },
+      },
+    ])
+
+    expect(events).toMatchObject([
+      { type: "step-start", index: 0 },
+      { type: "text-start", id: "text-1", providerMetadata: metadata },
+      { type: "text-delta", id: "text-1", text: "Hel", providerMetadata: { openai: { delta: 1 } } },
+      { type: "text-delta", id: "text-1", text: "lo", providerMetadata: { openai: { delta: 2 } } },
+      { type: "text-end", id: "text-1", providerMetadata: { openai: { done: true } } },
+      { type: "reasoning-start", id: "reasoning-1", providerMetadata: metadata },
+      { type: "reasoning-delta", id: "reasoning-1", text: "Think", providerMetadata: { openai: { delta: 3 } } },
+      { type: "reasoning-end", id: "reasoning-1", providerMetadata: { openai: { done: true } } },
+      { type: "tool-input-start", id: "call-1", name: "lookup", providerMetadata: metadata },
+      { type: "tool-input-delta", id: "call-1", name: "lookup", text: '{"query":' },
+      { type: "tool-input-delta", id: "call-1", name: "lookup", text: '"weather"}' },
+      { type: "tool-input-end", id: "call-1", name: "lookup", providerMetadata: { openai: { inputDone: true } } },
+      {
+        type: "tool-call",
+        id: "call-1",
+        name: "lookup",
+        input: { query: "weather" },
+        providerExecuted: true,
+        providerMetadata: { openai: { called: true } },
+      },
+      {
+        type: "tool-result",
+        id: "call-1",
+        name: "lookup",
+        result: { type: "json", value: { title: "Lookup", output: "sunny", metadata: { ok: true } } },
+        providerExecuted: true,
+        providerMetadata: { openai: { result: true } },
+      },
+      {
+        type: "step-finish",
+        index: 0,
+        reason: "unknown",
+        usage: {
+          inputTokens: 10,
+          outputTokens: 5,
+          totalTokens: 15,
+          reasoningTokens: 1,
+          cacheReadInputTokens: 3,
+          cacheWriteInputTokens: 2,
+        },
+        providerMetadata: { openai: { step: true } },
+      },
+      {
+        type: "finish",
+        reason: "unknown",
+        usage: {
+          inputTokens: 11,
+          outputTokens: 6,
+          totalTokens: 17,
+          reasoningTokens: 2,
+          cacheReadInputTokens: 4,
+        },
+      },
+    ])
+  })
+
+  test("creates stable block ids when AI SDK omits them", async () => {
+    const events = await adapt([
+      uncheckedAdapterEvent({ type: "text-delta", text: "implicit text" }),
+      uncheckedAdapterEvent({ type: "text-end" }),
+      uncheckedAdapterEvent({ type: "reasoning-delta", text: "implicit reasoning" }),
+      uncheckedAdapterEvent({ type: "reasoning-end" }),
+    ])
+
+    expect(events).toMatchObject([
+      { type: "text-delta", id: "text-0", text: "implicit text" },
+      { type: "text-end", id: "text-0" },
+      { type: "reasoning-delta", id: "reasoning-0", text: "implicit reasoning" },
+      { type: "reasoning-end", id: "reasoning-0" },
+    ])
+  })
+
+  test("explicitly ignores non-session-visible AI SDK chunks", async () => {
+    expect(
+      await adapt([
+        uncheckedAdapterEvent({ type: "abort" }),
+        uncheckedAdapterEvent({ type: "source" }),
+        uncheckedAdapterEvent({ type: "file" }),
+        uncheckedAdapterEvent({ type: "raw" }),
+        uncheckedAdapterEvent({ type: "tool-output-denied" }),
+        uncheckedAdapterEvent({ type: "tool-approval-request" }),
+      ]),
+    ).toEqual([])
+  })
+
+  test("preserves tool-error cause", async () => {
+    const error = new Permission.RejectedError()
+    const events = await Effect.runPromise(
+      LLMAISDK.toLLMEvents(LLMAISDK.adapterState(), {
+        type: "tool-error",
+        toolCallId: "call_123",
+        toolName: "bash",
+        input: {},
+        error,
+      }),
+    )
+
+    expect(events).toHaveLength(1)
+    expect(events[0]).toMatchObject({
+      type: "tool-error",
+      id: "call_123",
+      name: "bash",
+      message: error.message,
+      error,
+    })
+  })
+
+  test("emits undefined usage when every AI SDK usage field is missing", async () => {
+    // If every numeric field is undefined the translator should signal "no usage info"
+    // by emitting undefined, not by polluting the event with usage: {}. Downstream cost
+    // telemetry distinguishes "missing" from "zero," so emitting an empty object causes
+    // false positives ("usage was tracked, just empty") instead of correct nulls.
+    const events = await adapt([
+      {
+        type: "finish-step",
+        response: { id: "response-1", timestamp: new Date(0), modelId: "gpt-test" },
+        finishReason: "stop",
+        rawFinishReason: "stop",
+        providerMetadata: undefined,
+        usage: {
+          inputTokens: undefined,
+          outputTokens: undefined,
+          totalTokens: undefined,
+          reasoningTokens: undefined,
+          cachedInputTokens: undefined,
+          inputTokenDetails: { noCacheTokens: undefined, cacheReadTokens: undefined, cacheWriteTokens: undefined },
+          outputTokenDetails: { textTokens: undefined, reasoningTokens: undefined },
+        },
+      },
+    ])
+
+    expect(events).toHaveLength(1)
+    const stepFinish = events[0]
+    if (stepFinish.type !== "step-finish") throw new Error("expected step-finish")
+    expect(stepFinish.usage).toBeUndefined()
+  })
+
+  test("reuses adapter state cleanly across streams once finish has fired", async () => {
+    // adapterState() is meant to be per-stream, but the only thing finish currently clears
+    // is toolNames — step, text counters, and the current text/reasoning IDs all leak
+    // forward. A caller that reuses a state across two streams sees text-1/reasoning-1/
+    // step index 1 on the second stream's first events. The test pins the intended
+    // contract: after finish, the same state can be reused and starts fresh.
+    const state = LLMAISDK.adapterState()
+    const run = (events: ReadonlyArray<AISDKAdapterEvent>) =>
+      Effect.runPromise(
+        Effect.forEach(events, (event) => LLMAISDK.toLLMEvents(state, event)).pipe(Effect.map((items) => items.flat())),
+      )
+
+    await run([
+      { type: "start-step", request: {}, warnings: [] },
+      uncheckedAdapterEvent({ type: "text-delta", text: "first" }),
+      uncheckedAdapterEvent({ type: "text-end" }),
+      uncheckedAdapterEvent({ type: "reasoning-delta", text: "first reasoning" }),
+      uncheckedAdapterEvent({ type: "reasoning-end" }),
+      {
+        type: "finish-step",
+        response: { id: "r1", timestamp: new Date(0), modelId: "gpt-test" },
+        finishReason: "stop",
+        rawFinishReason: "stop",
+        providerMetadata: undefined,
+        usage: {
+          inputTokens: 1,
+          outputTokens: 1,
+          totalTokens: 2,
+          inputTokenDetails: { noCacheTokens: undefined, cacheReadTokens: undefined, cacheWriteTokens: undefined },
+          outputTokenDetails: { textTokens: undefined, reasoningTokens: undefined },
+        },
+      },
+      {
+        type: "finish",
+        finishReason: "stop",
+        rawFinishReason: "stop",
+        totalUsage: {
+          inputTokens: 1,
+          outputTokens: 1,
+          totalTokens: 2,
+          inputTokenDetails: { noCacheTokens: undefined, cacheReadTokens: undefined, cacheWriteTokens: undefined },
+          outputTokenDetails: { textTokens: undefined, reasoningTokens: undefined },
+        },
+      },
+    ])
+
+    const secondStream = await run([
+      { type: "start-step", request: {}, warnings: [] },
+      uncheckedAdapterEvent({ type: "text-delta", text: "second" }),
+      uncheckedAdapterEvent({ type: "text-end" }),
+      uncheckedAdapterEvent({ type: "reasoning-delta", text: "second reasoning" }),
+      uncheckedAdapterEvent({ type: "reasoning-end" }),
+    ])
+
+    expect(secondStream).toMatchObject([
+      { type: "step-start", index: 0 },
+      { type: "text-delta", id: "text-0", text: "second" },
+      { type: "text-end", id: "text-0" },
+      { type: "reasoning-delta", id: "reasoning-0", text: "second reasoning" },
+      { type: "reasoning-end", id: "reasoning-0" },
+    ])
+  })
+
+  // Anthropic emits cache write counts in providerMetadata.anthropic.cacheCreationInputTokens
+  // rather than usage.inputTokenDetails.cacheWriteTokens. Session.getUsage falls back to the
+  // metadata path — but only if the adapter preserves providerMetadata on step-finish.
+  test("preserves providerMetadata on step-finish so Anthropic cache writes survive getUsage", async () => {
+    const events = await adapt([
+      {
+        type: "finish-step",
+        response: { id: "msg_test", timestamp: new Date(0), modelId: "claude-3-5-sonnet" },
+        finishReason: "stop",
+        rawFinishReason: "stop",
+        // Anthropic's AI SDK shape: cacheWriteTokens is NOT in usage, it arrives via providerMetadata.
+        usage: {
+          inputTokens: 1000,
+          outputTokens: 500,
+          totalTokens: 1500,
+          inputTokenDetails: { noCacheTokens: 800, cacheReadTokens: 200, cacheWriteTokens: undefined },
+          outputTokenDetails: { textTokens: 500, reasoningTokens: undefined },
+        },
+        providerMetadata: { anthropic: { cacheCreationInputTokens: 300 } },
+      },
+    ])
+
+    expect(events).toHaveLength(1)
+    const stepFinish = events[0]
+    if (stepFinish.type !== "step-finish") throw new Error("expected step-finish")
+    expect(stepFinish.providerMetadata).toEqual({ anthropic: { cacheCreationInputTokens: 300 } })
+    expect(stepFinish.usage?.cacheWriteInputTokens).toBeUndefined()
+    expect(stepFinish.usage?.cacheReadInputTokens).toBe(200)
+
+    // End-to-end: with the metadata preserved, getUsage extracts cache.write from the fallback path.
+    const result = SessionNs.getUsage({
+      model: {
+        id: "claude-3-5-sonnet",
+        providerID: "anthropic",
+        name: "Claude",
+        limit: { context: 200_000, output: 8_000 },
+        cost: { input: 0, output: 0, cache: { read: 0, write: 0 } },
+        capabilities: {
+          toolcall: true,
+          attachment: false,
+          reasoning: false,
+          temperature: true,
+          input: { text: true, image: false, audio: false, video: false },
+          output: { text: true, image: false, audio: false, video: false },
+        },
+        api: { npm: "@ai-sdk/anthropic" },
+        options: {},
+      } as never,
+      usage: stepFinish.usage!,
+      metadata: stepFinish.providerMetadata,
+    })
+    expect(result.tokens.cache.write).toBe(300)
+    expect(result.tokens.cache.read).toBe(200)
+  })
+})
+
 type Capture = {
   url: URL
   headers: Headers
@@ -608,6 +992,18 @@ describe("session.llm.stream", () => {
           service_tier: null,
         },
       },
+      {
+        type: "response.output_item.added",
+        output_index: 0,
+        item: { type: "message", id: "item-1", status: "in_progress", role: "assistant", content: [] },
+      },
+      {
+        type: "response.content_part.added",
+        item_id: "item-1",
+        output_index: 0,
+        content_index: 0,
+        part: { type: "output_text", text: "", annotations: [] },
+      },
       {
         type: "response.output_text.delta",
         item_id: "item-1",
@@ -630,32 +1026,7 @@ describe("session.llm.stream", () => {
     ]
     const request = waitRequest("/responses", createEventResponse(responseChunks, true))
 
-    await using tmp = await tmpdir({
-      init: async (dir) => {
-        await Bun.write(
-          path.join(dir, "opencode.json"),
-          JSON.stringify({
-            $schema: "https://opencode.ai/config.json",
-            enabled_providers: ["openai"],
-            provider: {
-              openai: {
-                name: "OpenAI",
-                env: ["OPENAI_API_KEY"],
-                npm: "@ai-sdk/openai",
-                api: "https://api.openai.com/v1",
-                models: {
-                  [model.id]: configModel(model),
-                },
-                options: {
-                  apiKey: "test-openai-key",
-                  baseURL: `${server.url.origin}/v1`,
-                },
-              },
-            },
-          }),
-        )
-      },
-    })
+    await using tmp = await tmpdir({ config: openAIConfig(model, `${server.url.origin}/v1`) })
 
     await withTestInstance({
       directory: tmp.path,
@@ -706,6 +1077,438 @@ describe("session.llm.stream", () => {
     })
   })
 
+  test("keeps supported OpenAI models on AI SDK path when native flag is off", async () => {
+    const server = state.server
+    if (!server) {
+      throw new Error("Server not initialized")
+    }
+
+    const source = await loadFixture("openai", "gpt-5.2")
+    const model = source.model
+    const request = waitRequest(
+      "/responses",
+      createEventResponse(
+        [
+          {
+            type: "response.created",
+            response: {
+              id: "resp-flag-off",
+              created_at: Math.floor(Date.now() / 1000),
+              model: model.id,
+              service_tier: null,
+            },
+          },
+          {
+            type: "response.output_item.added",
+            output_index: 0,
+            item: { type: "message", id: "item-flag-off", status: "in_progress", role: "assistant", content: [] },
+          },
+          {
+            type: "response.content_part.added",
+            item_id: "item-flag-off",
+            output_index: 0,
+            content_index: 0,
+            part: { type: "output_text", text: "", annotations: [] },
+          },
+          {
+            type: "response.output_text.delta",
+            item_id: "item-flag-off",
+            delta: "Flag off",
+            logprobs: null,
+          },
+          {
+            type: "response.completed",
+            response: {
+              incomplete_details: null,
+              usage: {
+                input_tokens: 1,
+                input_tokens_details: null,
+                output_tokens: 1,
+                output_tokens_details: null,
+              },
+              service_tier: null,
+            },
+          },
+        ],
+        true,
+      ),
+    )
+    const failingNativeClient = Layer.succeed(
+      LLMClient.Service,
+      LLMClient.Service.of({
+        prepare: () => Effect.die(new Error("native LLM client should not be used when the flag is off")),
+        stream: () => Stream.die(new Error("native LLM client should not be used when the flag is off")),
+        generate: () => Effect.die(new Error("native LLM client should not be used when the flag is off")),
+      }),
+    )
+
+    await using tmp = await tmpdir({ config: openAIConfig(model, `${server.url.origin}/v1`) })
+
+    await withTestInstance({
+      directory: tmp.path,
+      fn: async (ctx) => {
+        const resolved = await getModel(ProviderID.openai, ModelID.make(model.id), ctx)
+        const sessionID = SessionID.make("session-test-native-flag-off")
+        const agent = {
+          name: "test",
+          mode: "primary",
+          options: {},
+          permission: [{ permission: "*", pattern: "*", action: "allow" }],
+        } satisfies Agent.Info
+
+        await drainWith(
+          LLM.layer.pipe(
+            Layer.provide(Auth.defaultLayer),
+            Layer.provide(Config.defaultLayer),
+            Layer.provide(Provider.defaultLayer),
+            Layer.provide(Plugin.defaultLayer),
+            Layer.provide(failingNativeClient),
+            Layer.provide(RuntimeFlags.layer({ experimentalNativeLlm: false })),
+          ),
+          {
+            user: {
+              id: MessageID.make("msg_user-native-flag-off"),
+              sessionID,
+              role: "user",
+              time: { created: Date.now() },
+              agent: agent.name,
+              model: { providerID: ProviderID.make("openai"), modelID: resolved.id, variant: "high" },
+            } satisfies MessageV2.User,
+            sessionID,
+            model: resolved,
+            agent,
+            system: ["You are a helpful assistant."],
+            messages: [{ role: "user", content: "Hello" }],
+            tools: {},
+          },
+          ctx,
+        )
+
+        const capture = await request
+        expect(capture.url.pathname.endsWith("/responses")).toBe(true)
+        expect(capture.body.model).toBe(resolved.api.id)
+      },
+    })
+  })
+
+  test("streams OpenAI through native runtime when opted in", async () => {
+    const server = state.server
+    if (!server) {
+      throw new Error("Server not initialized")
+    }
+
+    const source = await loadFixture("openai", "gpt-5.2")
+    const model = source.model
+    const chunks = [
+      {
+        type: "response.created",
+        response: {
+          id: "resp-native",
+        },
+      },
+      {
+        type: "response.output_item.added",
+        item: { type: "message", id: "item-native", status: "in_progress" },
+      },
+      {
+        type: "response.output_text.delta",
+        item_id: "item-native",
+        delta: "Hello native",
+      },
+      {
+        type: "response.completed",
+        response: {
+          incomplete_details: null,
+          usage: {
+            input_tokens: 1,
+            input_tokens_details: null,
+            output_tokens: 1,
+            output_tokens_details: null,
+          },
+        },
+      },
+    ]
+    const request = waitRequest("/responses", createEventResponse(chunks, true))
+
+    await using tmp = await tmpdir({ config: openAIConfig(model, `${server.url.origin}/v1`) })
+
+    await withTestInstance({
+      directory: tmp.path,
+      fn: async (ctx) => {
+        const resolved = await getModel(ProviderID.openai, ModelID.make(model.id), ctx)
+        const sessionID = SessionID.make("session-test-native")
+        const agent = {
+          name: "test",
+          mode: "primary",
+          options: {},
+          permission: [{ permission: "*", pattern: "*", action: "allow" }],
+          temperature: 0.2,
+        } satisfies Agent.Info
+
+        await drainWith(
+          llmLayerWithExecutor(RequestExecutor.defaultLayer, { experimentalNativeLlm: true }),
+          {
+            user: {
+              id: MessageID.make("msg_user-native"),
+              sessionID,
+              role: "user",
+              time: { created: Date.now() },
+              agent: agent.name,
+              model: { providerID: ProviderID.make("openai"), modelID: resolved.id, variant: "high" },
+            } satisfies MessageV2.User,
+            sessionID,
+            model: resolved,
+            agent,
+            system: ["You are a helpful assistant."],
+            messages: [{ role: "user", content: "Hello" }],
+            tools: {},
+          },
+          ctx,
+        )
+
+        const capture = await request
+        expect(capture.url.pathname.endsWith("/responses")).toBe(true)
+        expect(capture.headers.get("Authorization")).toBe("Bearer test-openai-key")
+        expect(capture.body.model).toBe(model.id)
+        expect(capture.body.stream).toBe(true)
+        expect((capture.body.reasoning as { effort?: string } | undefined)?.effort).toBe("high")
+        expect(JSON.stringify(capture.body.input)).toContain("You are a helpful assistant.")
+        expect(capture.body.input).toContainEqual({ role: "user", content: [{ type: "input_text", text: "Hello" }] })
+      },
+    })
+  })
+
+  test("uses injected native request executor for tool calls", async () => {
+    const source = await loadFixture("openai", "gpt-5.2")
+    const model = source.model
+    const chunks = [
+      {
+        type: "response.output_item.added",
+        item: { type: "function_call", id: "item-injected-tool", call_id: "call-injected-tool", name: "lookup" },
+      },
+      {
+        type: "response.function_call_arguments.delta",
+        item_id: "item-injected-tool",
+        delta: '{"query":"weather"}',
+      },
+      {
+        type: "response.output_item.done",
+        item: {
+          type: "function_call",
+          id: "item-injected-tool",
+          call_id: "call-injected-tool",
+          name: "lookup",
+          arguments: '{"query":"weather"}',
+        },
+      },
+      {
+        type: "response.completed",
+        response: { incomplete_details: null, usage: { input_tokens: 1, output_tokens: 1 } },
+      },
+    ]
+    let captured: Record<string, unknown> | undefined
+    let executed: unknown
+    const executor = Layer.succeed(
+      RequestExecutor.Service,
+      RequestExecutor.Service.of({
+        execute: (request) =>
+          Effect.gen(function* () {
+            const web = yield* HttpClientRequest.toWeb(request).pipe(Effect.orDie)
+            captured = (yield* Effect.promise(() => web.json())) as Record<string, unknown>
+            return HttpClientResponse.fromWeb(request, createEventResponse(chunks, true))
+          }),
+      }),
+    )
+
+    await using tmp = await tmpdir({ config: openAIConfig(model, "https://injected-openai.test/v1") })
+
+    await withTestInstance({
+      directory: tmp.path,
+      fn: async (ctx) => {
+        const resolved = await getModel(ProviderID.openai, ModelID.make(model.id), ctx)
+        const sessionID = SessionID.make("session-test-native-injected-tool")
+        const agent = {
+          name: "test",
+          mode: "primary",
+          options: {},
+          permission: [{ permission: "*", pattern: "*", action: "allow" }],
+        } satisfies Agent.Info
+
+        await drainWith(
+          llmLayerWithExecutor(executor, { experimentalNativeLlm: true }),
+          {
+            user: {
+              id: MessageID.make("msg_user-native-injected-tool"),
+              sessionID,
+              role: "user",
+              time: { created: Date.now() },
+              agent: agent.name,
+              model: { providerID: ProviderID.make("openai"), modelID: resolved.id },
+            } satisfies MessageV2.User,
+            sessionID,
+            model: resolved,
+            agent,
+            system: [],
+            messages: [{ role: "user", content: "Use lookup" }],
+            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" }
+                },
+              }),
+            },
+          },
+          ctx,
+        )
+
+        expect(captured?.model).toBe(model.id)
+        expect(captured?.tools).toEqual([
+          {
+            type: "function",
+            name: "lookup",
+            description: "Lookup data",
+            parameters: {
+              type: "object",
+              properties: { query: { type: "string" } },
+              required: ["query"],
+              additionalProperties: false,
+              $schema: "http://json-schema.org/draft-07/schema#",
+            },
+          },
+        ])
+        expect(executed).toEqual({ args: { query: "weather" }, toolCallId: "call-injected-tool" })
+      },
+    })
+  })
+
+  test("executes OpenAI tool calls through native runtime", async () => {
+    const server = state.server
+    if (!server) {
+      throw new Error("Server not initialized")
+    }
+
+    const source = await loadFixture("openai", "gpt-5.2")
+    const model = source.model
+    const chunks = [
+      {
+        type: "response.output_item.added",
+        item: { type: "function_call", id: "item-native-tool", call_id: "call-native-tool", name: "lookup" },
+      },
+      {
+        type: "response.function_call_arguments.delta",
+        item_id: "item-native-tool",
+        delta: '{"query":"weather"}',
+      },
+      {
+        type: "response.output_item.done",
+        item: {
+          type: "function_call",
+          id: "item-native-tool",
+          call_id: "call-native-tool",
+          name: "lookup",
+          arguments: '{"query":"weather"}',
+        },
+      },
+      {
+        type: "response.completed",
+        response: { incomplete_details: null, usage: { input_tokens: 1, output_tokens: 1 } },
+      },
+    ]
+    const request = waitRequest("/responses", createEventResponse(chunks, true))
+    let executed: unknown
+
+    await using tmp = await tmpdir({
+      init: async (dir) => {
+        await Bun.write(
+          path.join(dir, "opencode.json"),
+          JSON.stringify({
+            $schema: "https://opencode.ai/config.json",
+            enabled_providers: ["openai"],
+            provider: {
+              openai: {
+                name: "OpenAI",
+                env: ["OPENAI_API_KEY"],
+                npm: "@ai-sdk/openai",
+                api: "https://api.openai.com/v1",
+                models: {
+                  [model.id]: model,
+                },
+                options: {
+                  apiKey: "test-openai-key",
+                  baseURL: `${server.url.origin}/v1`,
+                },
+              },
+            },
+          }),
+        )
+      },
+    })
+
+    await withTestInstance({
+      directory: tmp.path,
+      fn: async (ctx) => {
+        const resolved = await getModel(ProviderID.openai, ModelID.make(model.id), ctx)
+        const sessionID = SessionID.make("session-test-native-tool")
+        const agent = {
+          name: "test",
+          mode: "primary",
+          options: {},
+          permission: [{ permission: "*", pattern: "*", action: "allow" }],
+        } satisfies Agent.Info
+
+        await drainWith(
+          llmLayerWithExecutor(RequestExecutor.defaultLayer, { experimentalNativeLlm: true }),
+          {
+            user: {
+              id: MessageID.make("msg_user-native-tool"),
+              sessionID,
+              role: "user",
+              time: { created: Date.now() },
+              agent: agent.name,
+              model: { providerID: ProviderID.make("openai"), modelID: resolved.id },
+            } satisfies MessageV2.User,
+            sessionID,
+            model: resolved,
+            agent,
+            system: [],
+            messages: [{ role: "user", content: "Use lookup" }],
+            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" }
+                },
+              }),
+            },
+          },
+          ctx,
+        )
+
+        const capture = await request
+        expect(capture.body.tools).toEqual([
+          {
+            type: "function",
+            name: "lookup",
+            description: "Lookup data",
+            parameters: {
+              type: "object",
+              properties: { query: { type: "string" } },
+              required: ["query"],
+              additionalProperties: false,
+              $schema: "http://json-schema.org/draft-07/schema#",
+            },
+          },
+        ])
+        expect(executed).toEqual({ args: { query: "weather" }, toolCallId: "call-native-tool" })
+      },
+    })
+  })
+
   test("accepts user image attachments as data URLs for OpenAI models", async () => {
     const server = state.server
     if (!server) {
@@ -724,6 +1527,18 @@ describe("session.llm.stream", () => {
           service_tier: null,
         },
       },
+      {
+        type: "response.output_item.added",
+        output_index: 0,
+        item: { type: "message", id: "item-data-url", status: "in_progress", role: "assistant", content: [] },
+      },
+      {
+        type: "response.content_part.added",
+        item_id: "item-data-url",
+        output_index: 0,
+        content_index: 0,
+        part: { type: "output_text", text: "", annotations: [] },
+      },
       {
         type: "response.output_text.delta",
         item_id: "item-data-url",

packages/opencode/test/session/processor-effect.test.ts

@@ -1,7 +1,9 @@
 import { NodeFileSystem } from "@effect/platform-node"
 import { expect } from "bun:test"
+import { tool } from "ai"
 import { Cause, Effect, Exit, Fiber, Layer } from "effect"
 import path from "path"
+import z from "zod"
 import type { Agent } from "../../src/agent/agent"
 import { Agent as AgentSvc } from "../../src/agent/agent"
 import { Bus } from "../../src/bus"
@@ -661,6 +663,71 @@ it.live("session.processor effect tests compact on structured context overflow",
   ),
 )
 
+it.live("session.processor effect tests complete AI SDK tool calls when native flag is off", () =>
+  provideTmpdirServer(
+    ({ dir, llm }) =>
+      Effect.gen(function* () {
+        const { processors, session, provider } = yield* boot()
+
+        yield* llm.tool("lookup", { query: "weather" })
+
+        const chat = yield* session.create({})
+        const parent = yield* user(chat.id, "tool")
+        const msg = yield* assistant(chat.id, parent.id, path.resolve(dir))
+        const mdl = yield* provider.getModel(ref.providerID, ref.modelID)
+        const handle = yield* processors.create({
+          assistantMessage: msg,
+          sessionID: chat.id,
+          model: mdl,
+        })
+
+        const value = yield* handle.process({
+          user: {
+            id: parent.id,
+            sessionID: chat.id,
+            role: "user",
+            time: parent.time,
+            agent: parent.agent,
+            model: { providerID: ref.providerID, modelID: ref.modelID },
+          } satisfies MessageV2.User,
+          sessionID: chat.id,
+          model: mdl,
+          agent: agent(),
+          system: [],
+          messages: [{ role: "user", content: "tool" }],
+          tools: {
+            lookup: tool({
+              description: "Look up information",
+              inputSchema: z.object({ query: z.string() }),
+              execute: async (input) => ({
+                title: "Weather lookup",
+                output: `result:${input.query}`,
+                metadata: { source: "test" },
+              }),
+            }),
+          },
+        })
+
+        const parts = MessageV2.parts(msg.id)
+        const call = parts.find((part): part is MessageV2.ToolPart => part.type === "tool")
+
+        expect(value).toBe("continue")
+        expect(yield* llm.calls).toBe(1)
+        expect(call?.callID).toBe("call_1")
+        expect(call?.tool).toBe("lookup")
+        expect(call?.state.status).toBe("completed")
+        if (call?.state.status !== "completed") return
+        expect(call.state.input).toEqual({ query: "weather" })
+        expect(call.state.output).toBe("result:weather")
+        expect(call.state.title).toBe("Weather lookup")
+        expect(call.state.metadata).toEqual({ source: "test" })
+        expect(call.state.time.start).toBeDefined()
+        expect(call.state.time.end).toBeDefined()
+      }),
+    { git: true, config: (url) => providerCfg(url) },
+  ),
+)
+
 it.live("session.processor effect tests mark pending tools as aborted on cleanup", () =>
   provideTmpdirServer(
     ({ dir, llm }) =>