Preview native LLM runtime stack (#27114)

bun.lock

@@ -421,6 +421,7 @@
         "@octokit/graphql": "9.0.2",
         "@octokit/rest": "catalog:",
         "@openauthjs/openauth": "catalog:",
+        "@opencode-ai/llm": "workspace:*",
         "@opencode-ai/plugin": "workspace:*",
         "@opencode-ai/script": "workspace:*",
         "@opencode-ai/sdk": "workspace:*",
@@ -489,6 +490,7 @@
         "@babel/core": "7.28.4",
         "@octokit/webhooks-types": "7.6.1",
         "@opencode-ai/core": "workspace:*",
+        "@opencode-ai/http-recorder": "workspace:*",
         "@opencode-ai/script": "workspace:*",
         "@parcel/watcher-darwin-arm64": "2.5.1",
         "@parcel/watcher-darwin-x64": "2.5.1",

packages/llm/src/schema/events.ts

@@ -91,6 +91,7 @@ export const TextDelta = Schema.Struct({
   type: Schema.tag("text-delta"),
   id: ContentBlockID,
   text: Schema.String,
+  providerMetadata: Schema.optional(ProviderMetadata),
 }).annotate({ identifier: "LLM.Event.TextDelta" })
 export type TextDelta = Schema.Schema.Type<typeof TextDelta>
 
@@ -112,6 +113,7 @@ export const ReasoningDelta = Schema.Struct({
   type: Schema.tag("reasoning-delta"),
   id: ContentBlockID,
   text: Schema.String,
+  providerMetadata: Schema.optional(ProviderMetadata),
 }).annotate({ identifier: "LLM.Event.ReasoningDelta" })
 export type ReasoningDelta = Schema.Schema.Type<typeof ReasoningDelta>
 

packages/opencode/package.json

@@ -39,8 +39,9 @@
   "devDependencies": {
     "@babel/core": "7.28.4",
     "@octokit/webhooks-types": "7.6.1",
-    "@opencode-ai/script": "workspace:*",
     "@opencode-ai/core": "workspace:*",
+    "@opencode-ai/http-recorder": "workspace:*",
+    "@opencode-ai/script": "workspace:*",
     "@parcel/watcher-darwin-arm64": "2.5.1",
     "@parcel/watcher-darwin-x64": "2.5.1",
     "@parcel/watcher-linux-arm64-glibc": "2.5.1",
@@ -101,6 +102,7 @@
     "@octokit/graphql": "9.0.2",
     "@octokit/rest": "catalog:",
     "@openauthjs/openauth": "catalog:",
+    "@opencode-ai/llm": "workspace:*",
     "@opencode-ai/plugin": "workspace:*",
     "@opencode-ai/script": "workspace:*",
     "@opencode-ai/sdk": "workspace:*",

packages/opencode/src/effect/runtime-flags.ts

@@ -50,6 +50,7 @@ export class Service extends ConfigService.Service<Service>()("@opencode/Runtime
   experimentalIconDiscovery: enabledByExperimental("OPENCODE_EXPERIMENTAL_ICON_DISCOVERY"),
   outputTokenMax: positiveInteger("OPENCODE_EXPERIMENTAL_OUTPUT_TOKEN_MAX"),
   bashDefaultTimeoutMs: positiveInteger("OPENCODE_EXPERIMENTAL_BASH_DEFAULT_TIMEOUT_MS"),
+  experimentalNativeLlm: enabledByExperimental("OPENCODE_EXPERIMENTAL_NATIVE_LLM"),
   client: Config.string("OPENCODE_CLIENT").pipe(Config.withDefault("cli")),
 }) {}
 

packages/opencode/src/session/llm.ts

@@ -2,7 +2,10 @@ import { Provider } from "@/provider/provider"
 import * as Log from "@opencode-ai/core/util/log"
 import { Context, Effect, Layer, Record } from "effect"
 import * as Stream from "effect/Stream"
-import { streamText, wrapLanguageModel, type ModelMessage, type Tool, tool, jsonSchema } from "ai"
+import { streamText, wrapLanguageModel, type ModelMessage, type Tool, tool as aiTool, jsonSchema } from "ai"
+import type { LLMEvent } from "@opencode-ai/llm"
+import { LLMClient, RequestExecutor } from "@opencode-ai/llm/route"
+import type { LLMClientService } from "@opencode-ai/llm/route"
 import { mergeDeep } from "remeda"
 import { GitLabWorkflowLanguageModel } from "gitlab-ai-provider"
 import { ProviderTransform } from "@/provider/transform"
@@ -23,10 +26,11 @@ import { EffectBridge } from "@/effect/bridge"
 import { RuntimeFlags } from "@/effect/runtime-flags"
 import * as Option from "effect/Option"
 import * as OtelTracer from "@effect/opentelemetry/Tracer"
+import { LLMAISDK } from "./llm/ai-sdk"
+import { LLMNativeRuntime } from "./llm/native-runtime"
 
 const log = Log.create({ service: "llm" })
 export const OUTPUT_TOKEN_MAX = ProviderTransform.OUTPUT_TOKEN_MAX
-type Result = Awaited<ReturnType<typeof streamText>>
 
 // Avoid re-instantiating remeda's deep merge types in this hot LLM path; the runtime behavior is still mergeDeep.
 const mergeOptions = (target: Record<string, any>, source: Record<string, any> | undefined): Record<string, any> =>
@@ -51,10 +55,8 @@ export type StreamRequest = StreamInput & {
   abort: AbortSignal
 }
 
-export type Event = Result["fullStream"] extends AsyncIterable<infer T> ? T : never
-
 export interface Interface {
-  readonly stream: (input: StreamInput) => Stream.Stream<Event, unknown>
+  readonly stream: (input: StreamInput) => Stream.Stream<LLMEvent, unknown>
 }
 
 export class Service extends Context.Service<Service, Interface>()("@opencode/LLM") {}
@@ -62,7 +64,13 @@ export class Service extends Context.Service<Service, Interface>()("@opencode/LL
 const live: Layer.Layer<
   Service,
   never,
-  Auth.Service | Config.Service | Provider.Service | Plugin.Service | Permission.Service | RuntimeFlags.Service
+  | Auth.Service
+  | Config.Service
+  | Provider.Service
+  | Plugin.Service
+  | Permission.Service
+  | LLMClientService
+  | RuntimeFlags.Service
 > = Layer.effect(
   Service,
   Effect.gen(function* () {
@@ -71,6 +79,7 @@ const live: Layer.Layer<
     const provider = yield* Provider.Service
     const plugin = yield* Plugin.Service
     const perm = yield* Permission.Service
+    const llmClient = yield* LLMClient.Service
     const flags = yield* RuntimeFlags.Service
 
     const run = Effect.fn("LLM.run")(function* (input: StreamRequest) {
@@ -202,7 +211,7 @@ const live: Layer.Layer<
         Object.keys(tools).length === 0 &&
         hasToolCalls(input.messages)
       ) {
-        tools["_noop"] = tool({
+        tools["_noop"] = aiTool({
           description: "Do not call this tool. It exists only for API compatibility and must never be invoked.",
           inputSchema: jsonSchema({
             type: "object",
@@ -322,86 +331,141 @@ const live: Layer.Layer<
         ? (yield* InstanceState.context).project.id
         : undefined
 
-      return streamText({
-        onError(error) {
-          l.error("stream error", {
-            error,
-          })
-        },
-        async experimental_repairToolCall(failed) {
-          const lower = failed.toolCall.toolName.toLowerCase()
-          if (lower !== failed.toolCall.toolName && sortedTools[lower]) {
-            l.info("repairing tool call", {
-              tool: failed.toolCall.toolName,
-              repaired: lower,
-            })
-            return {
-              ...failed.toolCall,
-              toolName: lower,
+      const requestHeaders = {
+        ...(input.model.providerID.startsWith("opencode")
+          ? {
+              ...(opencodeProjectID ? { "x-opencode-project": opencodeProjectID } : {}),
+              "x-opencode-session": input.sessionID,
+              "x-opencode-request": input.user.id,
+              "x-opencode-client": flags.client,
+              "User-Agent": `opencode/${InstallationVersion}`,
             }
-          }
-          return {
-            ...failed.toolCall,
-            input: JSON.stringify({
-              tool: failed.toolCall.toolName,
-              error: failed.error.message,
+          : {
+              "x-session-affinity": input.sessionID,
+              ...(input.parentSessionID ? { "x-parent-session-id": input.parentSessionID } : {}),
+              "User-Agent": `opencode/${InstallationVersion}`,
             }),
-            toolName: "invalid",
+        ...input.model.headers,
+        ...headers,
+      }
+
+      if (flags.experimentalNativeLlm) {
+        const native = LLMNativeRuntime.stream({
+          model: input.model,
+          provider: item,
+          auth: info,
+          llmClient,
+          isOpenaiOauth,
+          system,
+          messages,
+          tools: sortedTools,
+          toolChoice: input.toolChoice,
+          temperature: params.temperature,
+          topP: params.topP,
+          topK: params.topK,
+          maxOutputTokens: params.maxOutputTokens,
+          providerOptions: params.options,
+          headers: requestHeaders,
+          abort: input.abort,
+        })
+        if (native.type === "supported") {
+          yield* Effect.logInfo("llm runtime selected").pipe(
+            Effect.annotateLogs({
+              "llm.runtime": "native",
+              "llm.provider": input.model.providerID,
+              "llm.model": input.model.id,
+            }),
+          )
+          return {
+            type: "native" as const,
+            stream: native.stream,
           }
-        },
-        temperature: params.temperature,
-        topP: params.topP,
-        topK: params.topK,
-        providerOptions: ProviderTransform.providerOptions(input.model, params.options),
-        activeTools: Object.keys(sortedTools).filter((x) => x !== "invalid"),
-        tools: sortedTools,
-        toolChoice: input.toolChoice,
-        maxOutputTokens: params.maxOutputTokens,
-        abortSignal: input.abort,
-        headers: {
-          ...(input.model.providerID.startsWith("opencode")
-            ? {
-                "x-opencode-project": opencodeProjectID,
-                "x-opencode-session": input.sessionID,
-                "x-opencode-request": input.user.id,
-                "x-opencode-client": flags.client,
-                "User-Agent": `opencode/${InstallationVersion}`,
+        }
+        yield* Effect.logInfo("llm runtime selected").pipe(
+          Effect.annotateLogs({
+            "llm.runtime": "ai-sdk",
+            "llm.provider": input.model.providerID,
+            "llm.model": input.model.id,
+            "llm.native_unsupported_reason": native.reason,
+          }),
+        )
+        l.info("native runtime unavailable; falling back to ai-sdk", { reason: native.reason })
+      }
+
+      yield* Effect.logInfo("llm runtime selected").pipe(
+        Effect.annotateLogs({
+          "llm.runtime": "ai-sdk",
+          "llm.provider": input.model.providerID,
+          "llm.model": input.model.id,
+        }),
+      )
+      return {
+        type: "ai-sdk" as const,
+        result: streamText({
+          onError(error) {
+            l.error("stream error", {
+              error,
+            })
+          },
+          async experimental_repairToolCall(failed) {
+            const lower = failed.toolCall.toolName.toLowerCase()
+            if (lower !== failed.toolCall.toolName && sortedTools[lower]) {
+              l.info("repairing tool call", {
+                tool: failed.toolCall.toolName,
+                repaired: lower,
+              })
+              return {
+                ...failed.toolCall,
+                toolName: lower,
               }
-            : {
-                "x-session-affinity": input.sessionID,
-                ...(input.parentSessionID ? { "x-parent-session-id": input.parentSessionID } : {}),
-                "User-Agent": `opencode/${InstallationVersion}`,
+            }
+            return {
+              ...failed.toolCall,
+              input: JSON.stringify({
+                tool: failed.toolCall.toolName,
+                error: failed.error.message,
               }),
-          ...input.model.headers,
-          ...headers,
-        },
-        maxRetries: input.retries ?? 0,
-        messages,
-        model: wrapLanguageModel({
-          model: language,
-          middleware: [
-            {
-              specificationVersion: "v3" as const,
-              async transformParams(args) {
-                if (args.type === "stream") {
-                  // @ts-expect-error
-                  args.params.prompt = ProviderTransform.message(args.params.prompt, input.model, options)
-                }
-                return args.params
+              toolName: "invalid",
+            }
+          },
+          temperature: params.temperature,
+          topP: params.topP,
+          topK: params.topK,
+          providerOptions: ProviderTransform.providerOptions(input.model, params.options),
+          activeTools: Object.keys(sortedTools).filter((x) => x !== "invalid"),
+          tools: sortedTools,
+          toolChoice: input.toolChoice,
+          maxOutputTokens: params.maxOutputTokens,
+          abortSignal: input.abort,
+          headers: requestHeaders,
+          maxRetries: input.retries ?? 0,
+          messages,
+          model: wrapLanguageModel({
+            model: language,
+            middleware: [
+              {
+                specificationVersion: "v3" as const,
+                async transformParams(args) {
+                  if (args.type === "stream") {
+                    // @ts-expect-error
+                    args.params.prompt = ProviderTransform.message(args.params.prompt, input.model, options)
+                  }
+                  return args.params
+                },
               },
+            ],
+          }),
+          experimental_telemetry: {
+            isEnabled: cfg.experimental?.openTelemetry,
+            functionId: "session.llm",
+            tracer: telemetryTracer,
+            metadata: {
+              userId: cfg.username ?? "unknown",
+              sessionId: input.sessionID,
             },
-          ],
-        }),
-        experimental_telemetry: {
-          isEnabled: cfg.experimental?.openTelemetry,
-          functionId: "session.llm",
-          tracer: telemetryTracer,
-          metadata: {
-            userId: cfg.username ?? "unknown",
-            sessionId: input.sessionID,
           },
-        },
-      })
+        }),
+      }
     })
 
     const stream: Interface["stream"] = (input) =>
@@ -415,7 +479,15 @@ const live: Layer.Layer<
 
             const result = yield* run({ ...input, abort: ctrl.signal })
 
-            return Stream.fromAsyncIterable(result.fullStream, (e) => (e instanceof Error ? e : new Error(String(e))))
+            if (result.type === "native") return result.stream
+
+            const state = LLMAISDK.adapterState()
+            return Stream.fromAsyncIterable(result.result.fullStream, (e) =>
+              e instanceof Error ? e : new Error(String(e)),
+            ).pipe(
+              Stream.mapEffect((event) => LLMAISDK.toLLMEvents(state, event)),
+              Stream.flatMap((events) => Stream.fromIterable(events)),
+            )
           }),
         ),
       )
@@ -432,6 +504,7 @@ export const defaultLayer = Layer.suspend(() =>
     Layer.provide(Config.defaultLayer),
     Layer.provide(Provider.defaultLayer),
     Layer.provide(Plugin.defaultLayer),
+    Layer.provide(LLMClient.layer.pipe(Layer.provide(RequestExecutor.defaultLayer))),
     Layer.provide(RuntimeFlags.defaultLayer),
   ),
 )

packages/opencode/src/session/llm/AGENTS.md

@@ -0,0 +1,67 @@
+# Session LLM Runtime Boundaries
+
+`../llm.ts` is the opencode session LLM service. It owns opencode concerns: auth, config, model/provider resolution, plugins, permissions, telemetry headers, and runtime selection.
+
+This folder contains adapters behind that service boundary:
+
+- `ai-sdk.ts` converts AI SDK `fullStream` parts into `@opencode-ai/llm` `LLMEvent`s. This is the default runtime path.
+- `native-request.ts` converts opencode's normalized session input into a native `@opencode-ai/llm` `LLMRequest`. It does not execute requests.
+- `native-runtime.ts` is the opt-in native runtime adapter. It decides whether a selected model is supported, builds the native request, bridges opencode tools into native executable tools, and delegates transport to `LLMClient` / `RequestExecutor`.
+
+## Runtime selection
+
+Both runtimes converge on the same `LLMEvent` stream consumed by the session processor. The gate is per-request: a single session can route some calls through native and fall back for others.
+
+```txt
+                             ╭───────────────────╮
+╭───────────────────────────▶│ session processor │
+│                            ╰─────────┬─────────╯
+│                                      │
+│                                      │
+│                                      │
+│                                      ▼
+│                         ╭─────────────────────────╮
+│                         │ LLM.Service (../llm.ts) │
+│                         ╰────────────┬────────────╯
+│                                      │
+│                                      │
+│                                      │
+│                                      ▼
+│                                ╭───────────╮
+│                              ╭─╯           ╰─╮
+│                              │  native gate  │
+│                              ╰─╮           ╭─╯
+│                                ╰─────┬─────╯
+│                                      │
+│                     ╭────── no ──────┴─────── yes ────────╮
+│                     │                                     │
+│                     ▼                                     ▼
+│       ╭───────────────────────────╮             ╭───────────────────╮
+│       │          AI SDK           │             │ native-runtime.ts │
+│       │ streamText / generateText │             ╰────────┬──────────╯
+│       ╰─────────────┬─────────────╯                      │
+│                     │                                    │
+│                 ╭───╯                                    │
+│                 │                                        │
+│                 ▼                                        ▼
+│     ╭───────────────────────╮             ╭────────────────────────────╮
+│     │       ai-sdk.ts       │             │     native-request.ts      │
+│     │ fullStream → LLMEvent │             │ session input → LLMRequest │
+│     ╰──────────┬────────────╯             ╰──────────────┬─────────────╯
+│                │                                         │
+│                │                                     ╭───╯
+│                │                                     │
+│                ▼                                     ▼
+│       ╭─────────────────╮             ╭─────────────────────────────╮
+╰───────┤ LLMEvent stream │◀────────────┤ LLMClient · RequestExecutor │
+        ╰─────────────────╯             ╰─────────────────────────────╯
+```
+
+`native-runtime.ts` evaluates the gate and either bridges into `@opencode-ai/llm` or returns control so `llm.ts` can take the AI SDK path. Tool execution stays opencode-owned in both branches; only request lowering and transport differ.
+
+Safety boundary:
+
+- AI SDK remains the default.
+- `OPENCODE_EXPERIMENTAL_NATIVE_LLM=true` or the umbrella `OPENCODE_EXPERIMENTAL=true` opts in. Native is not a global replacement.
+- Native execution currently runs only for OpenAI-compatible Responses models exposed through `@ai-sdk/openai`: direct `openai` API-key auth and console-managed `opencode`/Zen API-key config.
+- Unsupported providers, OpenAI OAuth, and missing API-key cases fall back to AI SDK.

packages/opencode/src/session/llm/ai-sdk.ts

@@ -0,0 +1,254 @@
+import { FinishReason, LLMEvent, ProviderMetadata, ToolResultValue } from "@opencode-ai/llm"
+import { Effect, Schema } from "effect"
+import { type streamText } from "ai"
+import { errorMessage } from "@/util/error"
+
+type Result = Awaited<ReturnType<typeof streamText>>
+type AISDKEvent = Result["fullStream"] extends AsyncIterable<infer T> ? T : never
+
+export function adapterState() {
+  return {
+    step: 0,
+    text: 0,
+    reasoning: 0,
+    currentTextID: undefined as string | undefined,
+    currentReasoningID: undefined as string | undefined,
+    toolNames: {} as Record<string, string>,
+  }
+}
+
+function finishReason(value: string | undefined): FinishReason {
+  return Schema.is(FinishReason)(value) ? value : "unknown"
+}
+
+function providerMetadata(value: unknown): ProviderMetadata | undefined {
+  if (value == null) return undefined
+  return Schema.is(ProviderMetadata)(value) ? value : undefined
+}
+
+function usage(value: unknown) {
+  if (!value || typeof value !== "object") return undefined
+  const item = value as {
+    inputTokens?: number
+    outputTokens?: number
+    totalTokens?: number
+    reasoningTokens?: number
+    cachedInputTokens?: number
+    inputTokenDetails?: { cacheReadTokens?: number; cacheWriteTokens?: number }
+    outputTokenDetails?: { reasoningTokens?: number }
+  }
+  const entries = Object.entries({
+    inputTokens: item.inputTokens,
+    outputTokens: item.outputTokens,
+    totalTokens: item.totalTokens,
+    reasoningTokens: item.outputTokenDetails?.reasoningTokens ?? item.reasoningTokens,
+    cacheReadInputTokens: item.inputTokenDetails?.cacheReadTokens ?? item.cachedInputTokens,
+    cacheWriteInputTokens: item.inputTokenDetails?.cacheWriteTokens,
+  }).filter((entry) => entry[1] !== undefined)
+  return entries.length === 0 ? undefined : Object.fromEntries(entries)
+}
+
+function currentTextID(state: ReturnType<typeof adapterState>, id: string | undefined) {
+  state.currentTextID = id ?? state.currentTextID ?? `text-${state.text++}`
+  return state.currentTextID
+}
+
+function currentReasoningID(state: ReturnType<typeof adapterState>, id: string | undefined) {
+  state.currentReasoningID = id ?? state.currentReasoningID ?? `reasoning-${state.reasoning++}`
+  return state.currentReasoningID
+}
+
+export function toLLMEvents(
+  state: ReturnType<typeof adapterState>,
+  event: AISDKEvent,
+): Effect.Effect<ReadonlyArray<LLMEvent>, unknown> {
+  switch (event.type) {
+    case "start":
+      return Effect.succeed([])
+
+    case "start-step":
+      return Effect.succeed([LLMEvent.stepStart({ index: state.step })])
+
+    case "finish-step":
+      return Effect.sync(() => [
+        LLMEvent.stepFinish({
+          index: state.step++,
+          reason: finishReason(event.finishReason),
+          usage: usage(event.usage),
+          providerMetadata: providerMetadata(event.providerMetadata),
+        }),
+      ])
+
+    case "finish":
+      return Effect.sync(() => {
+        const events = [
+          LLMEvent.finish({
+            reason: finishReason(event.finishReason),
+            usage: usage(event.totalUsage),
+            providerMetadata: "providerMetadata" in event ? providerMetadata(event.providerMetadata) : undefined,
+          }),
+        ]
+        // Reset so the adapter can be reused for a follow-up stream without leaking
+        // counters or block IDs. adapterState() is the single source of truth for shape.
+        Object.assign(state, adapterState())
+        return events
+      })
+
+    case "text-start":
+      return Effect.sync(() => {
+        state.currentTextID = currentTextID(state, event.id)
+        return [
+          LLMEvent.textStart({
+            id: state.currentTextID,
+            providerMetadata: providerMetadata(event.providerMetadata),
+          }),
+        ]
+      })
+
+    case "text-delta":
+      return Effect.succeed([
+        LLMEvent.textDelta({
+          id: currentTextID(state, event.id),
+          text: event.text,
+          providerMetadata: providerMetadata(event.providerMetadata),
+        }),
+      ])
+
+    case "text-end":
+      return Effect.sync(() => {
+        const id = currentTextID(state, event.id)
+        state.currentTextID = undefined
+        return [
+          LLMEvent.textEnd({
+            id,
+            providerMetadata: providerMetadata(event.providerMetadata),
+          }),
+        ]
+      })
+
+    case "reasoning-start":
+      return Effect.sync(() => {
+        state.currentReasoningID = currentReasoningID(state, event.id)
+        return [
+          LLMEvent.reasoningStart({
+            id: state.currentReasoningID,
+            providerMetadata: providerMetadata(event.providerMetadata),
+          }),
+        ]
+      })
+
+    case "reasoning-delta":
+      return Effect.succeed([
+        LLMEvent.reasoningDelta({
+          id: currentReasoningID(state, event.id),
+          text: event.text,
+          providerMetadata: providerMetadata(event.providerMetadata),
+        }),
+      ])
+
+    case "reasoning-end":
+      return Effect.sync(() => {
+        const id = currentReasoningID(state, event.id)
+        state.currentReasoningID = undefined
+        return [
+          LLMEvent.reasoningEnd({
+            id,
+            providerMetadata: providerMetadata(event.providerMetadata),
+          }),
+        ]
+      })
+
+    case "tool-input-start":
+      return Effect.sync(() => {
+        state.toolNames[event.id] = event.toolName
+        return [
+          LLMEvent.toolInputStart({
+            id: event.id,
+            name: event.toolName,
+            providerMetadata: providerMetadata(event.providerMetadata),
+          }),
+        ]
+      })
+
+    case "tool-input-delta":
+      return Effect.succeed([
+        LLMEvent.toolInputDelta({
+          id: event.id,
+          name: state.toolNames[event.id] ?? "unknown",
+          text: event.delta ?? "",
+        }),
+      ])
+
+    case "tool-input-end":
+      return Effect.succeed([
+        LLMEvent.toolInputEnd({
+          id: event.id,
+          name: state.toolNames[event.id] ?? "unknown",
+          providerMetadata: providerMetadata(event.providerMetadata),
+        }),
+      ])
+
+    case "tool-call":
+      return Effect.sync(() => {
+        state.toolNames[event.toolCallId] = event.toolName
+        return [
+          LLMEvent.toolCall({
+            id: event.toolCallId,
+            name: event.toolName,
+            input: event.input,
+            providerExecuted: "providerExecuted" in event ? event.providerExecuted : undefined,
+            providerMetadata: providerMetadata(event.providerMetadata),
+          }),
+        ]
+      })
+
+    case "tool-result":
+      return Effect.sync(() => {
+        const name = state.toolNames[event.toolCallId] ?? "unknown"
+        delete state.toolNames[event.toolCallId]
+        return [
+          LLMEvent.toolResult({
+            id: event.toolCallId,
+            name,
+            result: ToolResultValue.make(event.output),
+            providerExecuted: "providerExecuted" in event ? event.providerExecuted : undefined,
+            providerMetadata: providerMetadata(event.providerMetadata),
+          }),
+        ]
+      })
+
+    case "tool-error":
+      return Effect.sync(() => {
+        const name = state.toolNames[event.toolCallId] ?? ("toolName" in event ? event.toolName : "unknown")
+        delete state.toolNames[event.toolCallId]
+        return [
+          LLMEvent.toolError({
+            id: event.toolCallId,
+            name,
+            message: errorMessage(event.error),
+            error: event.error,
+            providerMetadata: providerMetadata(event.providerMetadata),
+          }),
+        ]
+      })
+
+    case "error":
+      return Effect.fail(event.error)
+
+    case "abort":
+    case "source":
+    case "file":
+    case "raw":
+    case "tool-output-denied":
+    case "tool-approval-request":
+      return Effect.succeed([])
+
+    default: {
+      const _exhaustive: never = event
+      void _exhaustive
+      return Effect.succeed([])
+    }
+  }
+}
+
+export * as LLMAISDK from "./ai-sdk"

packages/opencode/src/session/llm/native-request.ts

@@ -0,0 +1,188 @@
+import type { JsonSchema, LLMRequest, ProviderMetadata } from "@opencode-ai/llm"
+import { LLM, Message, SystemPart, ToolCallPart, ToolDefinition, ToolResultPart } from "@opencode-ai/llm"
+import "@opencode-ai/llm/providers"
+import type { ModelMessage } from "ai"
+import type { Provider } from "@/provider/provider"
+import { isRecord } from "@/util/record"
+
+type ToolInput = {
+  readonly description?: string
+  readonly inputSchema?: unknown
+}
+
+export type RequestInput = {
+  readonly model: Provider.Model
+  readonly apiKey?: string
+  readonly baseURL?: string
+  readonly system?: readonly string[]
+  readonly messages: readonly ModelMessage[]
+  readonly tools?: Record<string, ToolInput>
+  readonly toolChoice?: "auto" | "required" | "none"
+  readonly temperature?: number
+  readonly topP?: number
+  readonly topK?: number
+  readonly maxOutputTokens?: number
+  readonly providerOptions?: LLMRequest["providerOptions"]
+  readonly headers?: Record<string, string>
+}
+
+const DEFAULT_BASE_URL: Record<string, string> = {
+  "@ai-sdk/openai": "https://api.openai.com/v1",
+  "@ai-sdk/anthropic": "https://api.anthropic.com/v1",
+  "@ai-sdk/google": "https://generativelanguage.googleapis.com/v1beta",
+  "@ai-sdk/amazon-bedrock": "https://bedrock-runtime.us-east-1.amazonaws.com",
+  "@openrouter/ai-sdk-provider": "https://openrouter.ai/api/v1",
+}
+
+const ROUTE: Record<string, string> = {
+  "@ai-sdk/openai": "openai-responses",
+  "@ai-sdk/azure": "azure-openai-responses",
+  "@ai-sdk/anthropic": "anthropic-messages",
+  "@ai-sdk/google": "gemini",
+  "@ai-sdk/amazon-bedrock": "bedrock-converse",
+  "@ai-sdk/openai-compatible": "openai-compatible-chat",
+  "@openrouter/ai-sdk-provider": "openrouter",
+}
+
+const providerMetadata = (value: unknown): ProviderMetadata | undefined => {
+  if (!isRecord(value)) return undefined
+  const result = Object.fromEntries(
+    Object.entries(value).filter((entry): entry is [string, Record<string, unknown>] => isRecord(entry[1])),
+  )
+  return Object.keys(result).length === 0 ? undefined : result
+}
+
+const textPart = (part: Record<string, unknown>) => ({
+  type: "text" as const,
+  text: typeof part.text === "string" ? part.text : "",
+  providerMetadata: providerMetadata(part.providerOptions),
+})
+
+const mediaPart = (part: Record<string, unknown>) => {
+  if (typeof part.data !== "string" && !(part.data instanceof Uint8Array))
+    throw new Error("Native LLM request adapter only supports file parts with string or Uint8Array data")
+  return {
+    type: "media" as const,
+    mediaType: typeof part.mediaType === "string" ? part.mediaType : "application/octet-stream",
+    data: part.data,
+    filename: typeof part.filename === "string" ? part.filename : undefined,
+  }
+}
+
+const toolResult = (part: Record<string, unknown>) => {
+  const output = isRecord(part.output) ? part.output : { type: "json", value: part.output }
+  const type = output.type === "text" ? "text" : output.type === "error-text" ? "error" : "json"
+  return ToolResultPart.make({
+    id: typeof part.toolCallId === "string" ? part.toolCallId : "",
+    name: typeof part.toolName === "string" ? part.toolName : "",
+    result: "value" in output ? output.value : output,
+    resultType: type,
+    providerExecuted: typeof part.providerExecuted === "boolean" ? part.providerExecuted : undefined,
+    providerMetadata: providerMetadata(part.providerOptions),
+  })
+}
+
+const contentPart = (part: unknown) => {
+  if (!isRecord(part)) throw new Error("Native LLM request adapter only supports object content parts")
+  if (part.type === "text") return textPart(part)
+  if (part.type === "file") return mediaPart(part)
+  if (part.type === "reasoning")
+    return {
+      type: "reasoning" as const,
+      text: typeof part.text === "string" ? part.text : "",
+      providerMetadata: providerMetadata(part.providerOptions),
+    }
+  if (part.type === "tool-call")
+    return ToolCallPart.make({
+      id: typeof part.toolCallId === "string" ? part.toolCallId : "",
+      name: typeof part.toolName === "string" ? part.toolName : "",
+      input: part.input,
+      providerExecuted: typeof part.providerExecuted === "boolean" ? part.providerExecuted : undefined,
+      providerMetadata: providerMetadata(part.providerOptions),
+    })
+  if (part.type === "tool-result") return toolResult(part)
+  throw new Error(`Native LLM request adapter does not support ${String(part.type)} content parts`)
+}
+
+const content = (value: ModelMessage["content"]) =>
+  typeof value === "string" ? [{ type: "text" as const, text: value }] : value.map(contentPart)
+
+const messages = (input: readonly ModelMessage[]) => {
+  const system = input.flatMap((message) => (message.role === "system" ? [SystemPart.make(message.content)] : []))
+  const messages = input.flatMap((message) => {
+    if (message.role === "system") return []
+    return [
+      Message.make({
+        role: message.role,
+        content: content(message.content),
+        native: isRecord(message.providerOptions) ? { providerOptions: message.providerOptions } : undefined,
+      }),
+    ]
+  })
+  return { system, messages }
+}
+
+const schema = (value: unknown): JsonSchema => {
+  if (!isRecord(value)) return { type: "object", properties: {} }
+  if (isRecord(value.jsonSchema)) return value.jsonSchema
+  return value
+}
+
+const tools = (input: Record<string, ToolInput> | undefined): ToolDefinition[] =>
+  Object.entries(input ?? {}).map(([name, item]) =>
+    ToolDefinition.make({
+      name,
+      description: item.description ?? "",
+      inputSchema: schema(item.inputSchema),
+    }),
+  )
+
+const generation = (input: RequestInput) => {
+  const result = {
+    temperature: input.temperature,
+    topP: input.topP,
+    topK: input.topK,
+    maxTokens: input.maxOutputTokens,
+  }
+  return Object.values(result).some((value) => value !== undefined) ? result : undefined
+}
+
+const baseURL = (model: Provider.Model) => {
+  if (model.api.url) return model.api.url
+  const fallback = DEFAULT_BASE_URL[model.api.npm]
+  if (fallback) return fallback
+  throw new Error(`Native LLM request adapter requires a base URL for ${model.providerID}/${model.id}`)
+}
+
+export const model = (input: Provider.Model | RequestInput, headers?: Record<string, string>) => {
+  const model = "model" in input ? input.model : input
+  const route = ROUTE[model.api.npm]
+  if (!route) throw new Error(`Native LLM request adapter does not support provider package ${model.api.npm}`)
+  return LLM.model({
+    id: model.api.id,
+    provider: model.providerID,
+    route,
+    baseURL: "model" in input && input.baseURL ? input.baseURL : baseURL(model),
+    apiKey: "model" in input ? input.apiKey : undefined,
+    headers: Object.keys({ ...model.headers, ...headers }).length === 0 ? undefined : { ...model.headers, ...headers },
+    limits: {
+      context: model.limit.context,
+      output: model.limit.output,
+    },
+  })
+}
+
+export const request = (input: RequestInput) => {
+  const converted = messages(input.messages)
+  return LLM.request({
+    model: model(input, input.headers),
+    system: [...(input.system ?? []).map(SystemPart.make), ...converted.system],
+    messages: converted.messages,
+    tools: tools(input.tools),
+    toolChoice: input.toolChoice,
+    generation: generation(input),
+    providerOptions: input.providerOptions,
+  })
+}
+
+export * as LLMNative from "./native-request"

packages/opencode/src/session/llm/native-runtime.ts

@@ -0,0 +1,124 @@
+import type { Auth } from "@/auth"
+import type { Provider } from "@/provider/provider"
+import { ProviderTransform } from "@/provider/transform"
+import { errorMessage } from "@/util/error"
+import { isRecord } from "@/util/record"
+import { asSchema, type ModelMessage, type Tool } from "ai"
+import { Effect } from "effect"
+import * as Stream from "effect/Stream"
+import { tool as nativeTool, ToolFailure, type JsonSchema, type LLMEvent } from "@opencode-ai/llm"
+import type { LLMClientShape } from "@opencode-ai/llm/route"
+import { LLMNative } from "./native-request"
+
+export type RuntimeStatus =
+  | { readonly type: "supported"; readonly apiKey: string; readonly baseURL?: string }
+  | { readonly type: "unsupported"; readonly reason: string }
+export type StreamResult =
+  | { readonly type: "supported"; readonly stream: Stream.Stream<LLMEvent, unknown> }
+  | { readonly type: "unsupported"; readonly reason: string }
+
+type StreamInput = {
+  readonly model: Provider.Model
+  readonly provider: Provider.Info
+  readonly auth: Auth.Info | undefined
+  readonly llmClient: LLMClientShape
+  readonly isOpenaiOauth: boolean
+  readonly system: string[]
+  readonly messages: ModelMessage[]
+  readonly tools: Record<string, Tool>
+  readonly toolChoice?: "auto" | "required" | "none"
+  readonly temperature?: number
+  readonly topP?: number
+  readonly topK?: number
+  readonly maxOutputTokens?: number
+  readonly providerOptions?: Record<string, any>
+  readonly headers: Record<string, string>
+  readonly abort: AbortSignal
+}
+
+export function status(input: Pick<StreamInput, "model" | "provider" | "auth">): RuntimeStatus {
+  if (input.model.providerID !== "openai" && !input.model.providerID.startsWith("opencode"))
+    return { type: "unsupported", reason: "provider is not openai or opencode" }
+  if (input.model.api.npm !== "@ai-sdk/openai") return { type: "unsupported", reason: "provider package is not OpenAI" }
+  if (input.auth?.type === "oauth") return { type: "unsupported", reason: "OAuth auth is not supported" }
+
+  const apiKey =
+    input.auth?.type === "api"
+      ? input.auth.key
+      : typeof input.provider.options.apiKey === "string"
+        ? input.provider.options.apiKey
+        : undefined
+  if (!apiKey) return { type: "unsupported", reason: "OpenAI API key is not configured" }
+
+  return {
+    type: "supported",
+    apiKey,
+    baseURL: typeof input.provider.options.baseURL === "string" ? input.provider.options.baseURL : undefined,
+  }
+}
+
+export function stream(input: StreamInput): StreamResult {
+  const current = status(input)
+  if (current.type === "unsupported") return current
+
+  return {
+    ...current,
+    stream: input.llmClient.stream({
+      request: LLMNative.request({
+        model: input.model,
+        apiKey: current.apiKey,
+        baseURL: current.baseURL,
+        system: input.isOpenaiOauth ? input.system : [],
+        messages: ProviderTransform.message(input.messages, input.model, input.providerOptions ?? {}),
+        toolChoice: input.toolChoice,
+        temperature: input.temperature,
+        topP: input.topP,
+        topK: input.topK,
+        maxOutputTokens: input.maxOutputTokens,
+        providerOptions: ProviderTransform.providerOptions(input.model, input.providerOptions ?? {}),
+        headers: { ...providerHeaders(input.provider.options.headers), ...input.headers },
+      }),
+      tools: nativeTools(input.tools, input),
+    }),
+  }
+}
+
+function providerHeaders(value: unknown): Record<string, string> | undefined {
+  if (!isRecord(value)) return undefined
+  return Object.fromEntries(
+    Object.entries(value).filter((entry): entry is [string, string] => typeof entry[1] === "string"),
+  )
+}
+
+function nativeSchema(value: unknown): JsonSchema {
+  if (!value || typeof value !== "object") return { type: "object", properties: {} }
+  if ("jsonSchema" in value && value.jsonSchema && typeof value.jsonSchema === "object")
+    return value.jsonSchema as JsonSchema
+  return asSchema(value as Parameters<typeof asSchema>[0]).jsonSchema as JsonSchema
+}
+
+export function nativeTools(tools: Record<string, Tool>, input: Pick<StreamInput, "messages" | "abort">) {
+  return Object.fromEntries(
+    Object.entries(tools).map(([name, item]) => [
+      name,
+      nativeTool({
+        description: item.description ?? "",
+        jsonSchema: nativeSchema(item.inputSchema),
+        execute: (args: unknown, ctx) =>
+          Effect.tryPromise({
+            try: () => {
+              if (!item.execute) throw new Error(`Tool has no execute handler: ${name}`)
+              return item.execute(args, {
+                toolCallId: ctx?.id ?? name,
+                messages: input.messages,
+                abortSignal: input.abort,
+              })
+            },
+            catch: (error) => new ToolFailure({ message: errorMessage(error), error }),
+          }),
+      }),
+    ]),
+  )
+}
+
+export * as LLMNativeRuntime from "./native-runtime"

packages/opencode/src/session/processor.ts

@@ -1,4 +1,5 @@
-import { Cause, Deferred, Effect, Exit, Layer, Context, Scope } from "effect"
+import { Image } from "@/image/image"
+import { Cause, Deferred, Effect, Exit, Layer, Context, Scope, Schema } from "effect"
 import * as Stream from "effect/Stream"
 import { Agent } from "@/agent/agent"
 import { Bus } from "@/bus"
@@ -9,7 +10,6 @@ import { Snapshot } from "@/snapshot"
 import * as Session from "./session"
 import { LLM } from "./llm"
 import { MessageV2 } from "./message-v2"
-import { Image } from "@/image/image"
 import { isOverflow } from "./overflow"
 import { PartID } from "./schema"
 import type { SessionID } from "./schema"
@@ -28,14 +28,13 @@ import { ModelV2 } from "@opencode-ai/core/model"
 import { ProviderV2 } from "@opencode-ai/core/provider"
 import * as DateTime from "effect/DateTime"
 import { RuntimeFlags } from "@/effect/runtime-flags"
+import { Usage, type LLMEvent } from "@opencode-ai/llm"
 
 const DOOM_LOOP_THRESHOLD = 3
 const log = Log.create({ service: "session.processor" })
 
 export type Result = "compact" | "stop" | "continue"
 
-export type Event = LLM.Event
-
 export interface Handle {
   readonly message: MessageV2.Assistant
   readonly updateToolCall: (
@@ -69,6 +68,7 @@ type ToolCall = {
   messageID: MessageV2.ToolPart["messageID"]
   sessionID: MessageV2.ToolPart["sessionID"]
   done: Deferred.Deferred<void>
+  inputEnded: boolean
 }
 
 interface ProcessorContext extends Input {
@@ -81,7 +81,7 @@ interface ProcessorContext extends Input {
   reasoningMap: Record<string, MessageV2.ReasoningPart>
 }
 
-type StreamEvent = Event
+type StreamEvent = LLMEvent
 
 export class Service extends Context.Service<Service, Interface>()("@opencode/SessionProcessor") {}
 
@@ -137,7 +137,7 @@ export const layer = Layer.effect(
 
       const readToolCall = Effect.fn("SessionProcessor.readToolCall")(function* (toolCallID: string) {
         const call = ctx.toolcalls[toolCallID]
-        if (!call) return
+        if (!call) return undefined
         const part = yield* session.getPart({
           partID: call.partID,
           messageID: call.messageID,
@@ -145,7 +145,7 @@ export const layer = Layer.effect(
         })
         if (!part || part.type !== "tool") {
           delete ctx.toolcalls[toolCallID]
-          return
+          return undefined
         }
         return { call, part }
       })
@@ -155,7 +155,7 @@ export const layer = Layer.effect(
         update: (part: MessageV2.ToolPart) => MessageV2.ToolPart,
       ) {
         const match = yield* readToolCall(toolCallID)
-        if (!match) return
+        if (!match) return undefined
         const part = yield* session.updatePart(update(match.part))
         ctx.toolcalls[toolCallID] = {
           ...match.call,
@@ -211,12 +211,98 @@ export const layer = Layer.effect(
         return true
       })
 
+      const finishReasoning = Effect.fn("SessionProcessor.finishReasoning")(function* (reasoningID: string) {
+        if (!(reasoningID in ctx.reasoningMap)) return
+        // TODO(v2): Temporary dual-write while migrating session messages to v2 events.
+        if (flags.experimentalEventSystem) {
+          yield* events.publish(SessionEvent.Reasoning.Ended, {
+            sessionID: ctx.sessionID,
+            reasoningID,
+            text: ctx.reasoningMap[reasoningID].text,
+            timestamp: DateTime.makeUnsafe(Date.now()),
+          })
+        }
+        // oxlint-disable-next-line no-self-assign -- reactivity trigger
+        ctx.reasoningMap[reasoningID].text = ctx.reasoningMap[reasoningID].text
+        ctx.reasoningMap[reasoningID].time = { ...ctx.reasoningMap[reasoningID].time, end: Date.now() }
+        yield* session.updatePart(ctx.reasoningMap[reasoningID])
+        delete ctx.reasoningMap[reasoningID]
+      })
+
+      const ensureToolCall = Effect.fn("SessionProcessor.ensureToolCall")(function* (input: {
+        id: string
+        name: string
+        providerExecuted?: boolean
+      }) {
+        const existing = yield* readToolCall(input.id)
+        if (existing) {
+          if (!input.providerExecuted || existing.part.metadata?.providerExecuted) return existing
+          const part = yield* session.updatePart({
+            ...existing.part,
+            metadata: { ...existing.part.metadata, providerExecuted: true },
+          })
+          ctx.toolcalls[input.id] = {
+            ...existing.call,
+            partID: part.id,
+            messageID: part.messageID,
+            sessionID: part.sessionID,
+          }
+          return { call: ctx.toolcalls[input.id], part }
+        }
+        // TODO(v2): Temporary dual-write while migrating session messages to v2 events.
+        if (flags.experimentalEventSystem) {
+          yield* events.publish(SessionEvent.Tool.Input.Started, {
+            sessionID: ctx.sessionID,
+            callID: input.id,
+            name: input.name,
+            timestamp: DateTime.makeUnsafe(Date.now()),
+          })
+        }
+        const part = yield* session.updatePart({
+          id: PartID.ascending(),
+          messageID: ctx.assistantMessage.id,
+          sessionID: ctx.assistantMessage.sessionID,
+          type: "tool",
+          tool: input.name,
+          callID: input.id,
+          state: { status: "pending", input: {}, raw: "" },
+          metadata: input.providerExecuted ? { providerExecuted: true } : undefined,
+        } satisfies MessageV2.ToolPart)
+        ctx.toolcalls[input.id] = {
+          done: yield* Deferred.make<void>(),
+          partID: part.id,
+          messageID: part.messageID,
+          sessionID: part.sessionID,
+          inputEnded: false,
+        }
+        return { call: ctx.toolcalls[input.id], part }
+      })
+
+      const isFilePart = Schema.is(MessageV2.FilePart)
+
+      const toolResultOutput = (value: Extract<StreamEvent, { type: "tool-result" }>) => {
+        if (isRecord(value.result.value) && typeof value.result.value.output === "string") {
+          return {
+            title: typeof value.result.value.title === "string" ? value.result.value.title : value.name,
+            metadata: isRecord(value.result.value.metadata) ? value.result.value.metadata : {},
+            output: value.result.value.output,
+            attachments: Array.isArray(value.result.value.attachments)
+              ? value.result.value.attachments.filter(isFilePart)
+              : undefined,
+          }
+        }
+        return {
+          title: value.name,
+          metadata: value.result.type === "json" && isRecord(value.result.value) ? value.result.value : {},
+          output:
+            typeof value.result.value === "string" ? value.result.value : (JSON.stringify(value.result.value) ?? ""),
+        }
+      }
+
+      const toolInput = (value: unknown): Record<string, any> => (isRecord(value) ? value : { value })
+
       const handleEvent = Effect.fnUntraced(function* (value: StreamEvent) {
         switch (value.type) {
-          case "start":
-            yield* status.set(ctx.sessionID, { type: "busy" })
-            return
-
           case "reasoning-start":
             if (value.id in ctx.reasoningMap) return
             // TODO(v2): Temporary dual-write while migrating session messages to v2 events.
@@ -240,6 +326,7 @@ export const layer = Layer.effect(
             return
 
           case "reasoning-delta":
+            // Match dev: silently drop orphan deltas (no preceding reasoning-start).
             if (!(value.id in ctx.reasoningMap)) return
             ctx.reasoningMap[value.id].text += value.text
             if (value.providerMetadata) ctx.reasoningMap[value.id].metadata = value.providerMetadata
@@ -253,59 +340,26 @@ export const layer = Layer.effect(
             return
 
           case "reasoning-end":
-            if (!(value.id in ctx.reasoningMap)) return
-            // TODO(v2): Temporary dual-write while migrating session messages to v2 events.
-            if (flags.experimentalEventSystem) {
-              yield* events.publish(SessionEvent.Reasoning.Ended, {
-                sessionID: ctx.sessionID,
-                reasoningID: value.id,
-                text: ctx.reasoningMap[value.id].text,
-                timestamp: DateTime.makeUnsafe(Date.now()),
-              })
+            if (value.providerMetadata && value.id in ctx.reasoningMap) {
+              ctx.reasoningMap[value.id].metadata = value.providerMetadata
             }
-            // oxlint-disable-next-line no-self-assign -- reactivity trigger
-            ctx.reasoningMap[value.id].text = ctx.reasoningMap[value.id].text
-            ctx.reasoningMap[value.id].time = { ...ctx.reasoningMap[value.id].time, end: Date.now() }
-            if (value.providerMetadata) ctx.reasoningMap[value.id].metadata = value.providerMetadata
-            yield* session.updatePart(ctx.reasoningMap[value.id])
-            delete ctx.reasoningMap[value.id]
+            yield* finishReasoning(value.id)
             return
 
           case "tool-input-start":
             if (ctx.assistantMessage.summary) {
-              throw new Error(`Tool call not allowed while generating summary: ${value.toolName}`)
-            }
-            // TODO(v2): Temporary dual-write while migrating session messages to v2 events.
-            if (flags.experimentalEventSystem) {
-              yield* events.publish(SessionEvent.Tool.Input.Started, {
-                sessionID: ctx.sessionID,
-                callID: value.id,
-                name: value.toolName,
-                timestamp: DateTime.makeUnsafe(Date.now()),
-              })
-            }
-            const part = yield* session.updatePart({
-              id: ctx.toolcalls[value.id]?.partID ?? PartID.ascending(),
-              messageID: ctx.assistantMessage.id,
-              sessionID: ctx.assistantMessage.sessionID,
-              type: "tool",
-              tool: value.toolName,
-              callID: value.id,
-              state: { status: "pending", input: {}, raw: "" },
-              metadata: value.providerExecuted ? { providerExecuted: true } : undefined,
-            } satisfies MessageV2.ToolPart)
-            ctx.toolcalls[value.id] = {
-              done: yield* Deferred.make<void>(),
-              partID: part.id,
-              messageID: part.messageID,
-              sessionID: part.sessionID,
+              throw new Error(`Tool call not allowed while generating summary: ${value.name}`)
             }
+            yield* ensureToolCall(value)
             return
 
           case "tool-input-delta":
+            // AI SDK emits a final `tool-call` with the parsed `input`; accumulating
+            // delta fragments into `state.raw` is redundant work for no current consumer.
             return
 
           case "tool-input-end": {
+            const toolCall = yield* ensureToolCall(value)
             // TODO(v2): Temporary dual-write while migrating session messages to v2 events.
             if (flags.experimentalEventSystem) {
               yield* events.publish(SessionEvent.Tool.Input.Ended, {
@@ -315,37 +369,52 @@ export const layer = Layer.effect(
                 timestamp: DateTime.makeUnsafe(Date.now()),
               })
             }
+            ctx.toolcalls[value.id] = { ...toolCall.call, inputEnded: true }
             return
           }
 
           case "tool-call": {
             if (ctx.assistantMessage.summary) {
-              throw new Error(`Tool call not allowed while generating summary: ${value.toolName}`)
+              throw new Error(`Tool call not allowed while generating summary: ${value.name}`)
+            }
+            const toolCall = yield* ensureToolCall(value)
+            const input = toolInput(value.input)
+            if (!toolCall.call.inputEnded) {
+              // TODO(v2): Temporary dual-write while migrating session messages to v2 events.
+              if (flags.experimentalEventSystem) {
+                yield* events.publish(SessionEvent.Tool.Input.Ended, {
+                  sessionID: ctx.sessionID,
+                  callID: value.id,
+                  text: "",
+                  timestamp: DateTime.makeUnsafe(Date.now()),
+                })
+              }
             }
-            const toolCall = yield* readToolCall(value.toolCallId)
             // TODO(v2): Temporary dual-write while migrating session messages to v2 events.
             if (flags.experimentalEventSystem) {
               yield* events.publish(SessionEvent.Tool.Called, {
                 sessionID: ctx.sessionID,
-                callID: value.toolCallId,
-                tool: value.toolName,
-                input: value.input,
+                callID: value.id,
+                tool: value.name,
+                input,
                 provider: {
-                  executed: toolCall?.part.metadata?.providerExecuted === true,
+                  executed: toolCall.part.metadata?.providerExecuted === true,
                   ...(value.providerMetadata ? { metadata: value.providerMetadata } : {}),
                 },
                 timestamp: DateTime.makeUnsafe(Date.now()),
               })
             }
-            yield* updateToolCall(value.toolCallId, (match) => ({
+            yield* updateToolCall(value.id, (match) => ({
               ...match,
-              tool: value.toolName,
-              state: {
-                ...match.state,
-                status: "running",
-                input: value.input,
-                time: { start: Date.now() },
-              },
+              tool: value.name,
+              state:
+                match.state.status === "running"
+                  ? { ...match.state, input }
+                  : {
+                      status: "running",
+                      input,
+                      time: { start: Date.now() },
+                    },
               metadata: match.metadata?.providerExecuted
                 ? { ...value.providerMetadata, providerExecuted: true }
                 : value.providerMetadata,
@@ -359,9 +428,9 @@ export const layer = Layer.effect(
               !recentParts.every(
                 (part) =>
                   part.type === "tool" &&
-                  part.tool === value.toolName &&
+                  part.tool === value.name &&
                   part.state.status !== "pending" &&
-                  JSON.stringify(part.state.input) === JSON.stringify(value.input),
+                  JSON.stringify(part.state.input) === JSON.stringify(input),
               )
             ) {
               return
@@ -370,27 +439,19 @@ export const layer = Layer.effect(
             const agent = yield* agents.get(ctx.assistantMessage.agent)
             yield* permission.ask({
               permission: "doom_loop",
-              patterns: [value.toolName],
+              patterns: [value.name],
               sessionID: ctx.assistantMessage.sessionID,
-              metadata: { tool: value.toolName, input: value.input },
-              always: [value.toolName],
+              metadata: { tool: value.name, input },
+              always: [value.name],
               ruleset: agent.permission,
             })
             return
           }
 
           case "tool-result": {
-            const toolCall = yield* readToolCall(value.toolCallId)
-            const toolAttachments: MessageV2.FilePart[] = (
-              Array.isArray(value.output.attachments) ? value.output.attachments : []
-            ).filter(
-              (attachment: unknown): attachment is MessageV2.FilePart =>
-                isRecord(attachment) &&
-                attachment.type === "file" &&
-                typeof attachment.mime === "string" &&
-                typeof attachment.url === "string",
-            )
-            const normalized = yield* Effect.forEach(toolAttachments, (attachment) =>
+            const toolCall = yield* readToolCall(value.id)
+            const rawOutput = toolResultOutput(value)
+            const normalized = yield* Effect.forEach(rawOutput.attachments ?? [], (attachment) =>
               attachment.mime.startsWith("image/")
                 ? image.normalize(attachment).pipe(
                     Effect.catchIf(
@@ -404,18 +465,18 @@ export const layer = Layer.effect(
             const omitted = normalized.filter(Exit.isFailure).length
             const attachments = normalized.filter(Exit.isSuccess).map((item) => item.value)
             const output = {
-              ...value.output,
+              ...rawOutput,
               output:
                 omitted === 0
-                  ? value.output.output
-                  : `${value.output.output}\n\n[${omitted} image${omitted === 1 ? "" : "s"} omitted: could not be resized below the image size limit.]`,
-              attachments: attachments?.length ? attachments : undefined,
+                  ? rawOutput.output
+                  : `${rawOutput.output}\n\n[${omitted} image${omitted === 1 ? "" : "s"} omitted: could not be resized below the image size limit.]`,
+              attachments: attachments.length ? attachments : undefined,
             }
             // TODO(v2): Temporary dual-write while migrating session messages to v2 events.
             if (flags.experimentalEventSystem) {
               yield* events.publish(SessionEvent.Tool.Success, {
                 sessionID: ctx.sessionID,
-                callID: value.toolCallId,
+                callID: value.id,
                 structured: output.metadata,
                 content: [
                   {
@@ -423,32 +484,32 @@ export const layer = Layer.effect(
                     text: output.output,
                   },
                   ...(output.attachments?.map((item: MessageV2.FilePart) => ({
-                    type: "file",
+                    type: "file" as const,
                     uri: item.url,
                     mime: item.mime,
                     name: item.filename,
                   })) ?? []),
                 ],
                 provider: {
-                  executed: toolCall?.part.metadata?.providerExecuted === true,
+                  executed: value.providerExecuted === true || toolCall?.part.metadata?.providerExecuted === true,
                 },
                 timestamp: DateTime.makeUnsafe(Date.now()),
               })
             }
-            yield* completeToolCall(value.toolCallId, output)
+            yield* completeToolCall(value.id, output)
             return
           }
 
           case "tool-error": {
-            const toolCall = yield* readToolCall(value.toolCallId)
+            const toolCall = yield* readToolCall(value.id)
             // TODO(v2): Temporary dual-write while migrating session messages to v2 events.
             if (flags.experimentalEventSystem) {
               yield* events.publish(SessionEvent.Tool.Failed, {
                 sessionID: ctx.sessionID,
-                callID: value.toolCallId,
+                callID: value.id,
                 error: {
                   type: "unknown",
-                  message: errorMessage(value.error),
+                  message: value.message,
                 },
                 provider: {
                   executed: toolCall?.part.metadata?.providerExecuted === true,
@@ -456,14 +517,14 @@ export const layer = Layer.effect(
                 timestamp: DateTime.makeUnsafe(Date.now()),
               })
             }
-            yield* failToolCall(value.toolCallId, value.error)
+            yield* failToolCall(value.id, value.error ?? new Error(value.message))
             return
           }
 
-          case "error":
-            throw value.error
+          case "provider-error":
+            throw new Error(value.message)
 
-          case "start-step":
+          case "step-start":
             if (!ctx.snapshot) ctx.snapshot = yield* snapshot.track()
             if (!ctx.assistantMessage.summary) {
               // TODO(v2): Temporary dual-write while migrating session messages to v2 events.
@@ -490,11 +551,12 @@ export const layer = Layer.effect(
             })
             return
 
-          case "finish-step": {
+          case "step-finish": {
             const completedSnapshot = yield* snapshot.track()
+            yield* Effect.forEach(Object.keys(ctx.reasoningMap), finishReasoning)
             const usage = Session.getUsage({
               model: ctx.model,
-              usage: value.usage,
+              usage: value.usage ?? new Usage({}),
               metadata: value.providerMetadata,
             })
             if (!ctx.assistantMessage.summary) {
@@ -502,7 +564,7 @@ export const layer = Layer.effect(
               if (flags.experimentalEventSystem) {
                 yield* events.publish(SessionEvent.Step.Ended, {
                   sessionID: ctx.sessionID,
-                  finish: value.finishReason,
+                  finish: value.reason,
                   cost: usage.cost,
                   tokens: usage.tokens,
                   snapshot: completedSnapshot,
@@ -510,12 +572,12 @@ export const layer = Layer.effect(
                 })
               }
             }
-            ctx.assistantMessage.finish = value.finishReason
+            ctx.assistantMessage.finish = value.reason
             ctx.assistantMessage.cost += usage.cost
             ctx.assistantMessage.tokens = usage.tokens
             yield* session.updatePart({
               id: PartID.ascending(),
-              reason: value.finishReason,
+              reason: value.reason,
               snapshot: completedSnapshot,
               messageID: ctx.assistantMessage.id,
               sessionID: ctx.assistantMessage.sessionID,
@@ -622,10 +684,6 @@ export const layer = Layer.effect(
 
           case "finish":
             return
-
-          default:
-            slog.info("unhandled", { event: value.type, value })
-            return
         }
       })
 
@@ -727,6 +785,7 @@ export const layer = Layer.effect(
           yield* Effect.gen(function* () {
             ctx.currentText = undefined
             ctx.reasoningMap = {}
+            yield* status.set(ctx.sessionID, { type: "busy" })
             const stream = llm.stream(streamInput)
 
             yield* stream.pipe(

packages/opencode/src/session/prompt.ts

@@ -63,6 +63,7 @@ import * as Database from "@/storage/db"
 import { SessionTable } from "./session.sql"
 import { referencePromptMetadata, referenceTextPart } from "./prompt/reference"
 import { SessionReminders } from "./reminders"
+import { LLMEvent } from "@opencode-ai/llm"
 
 // @ts-ignore
 globalThis.AI_SDK_LOG_WARNINGS = false
@@ -283,7 +284,7 @@ export const layer = Layer.effect(
           messages: [{ role: "user", content: "Generate a title for this conversation:\n" }, ...msgs],
         })
         .pipe(
-          Stream.filter((e): e is Extract<LLM.Event, { type: "text-delta" }> => e.type === "text-delta"),
+          Stream.filter(LLMEvent.is.textDelta),
           Stream.map((e) => e.text),
           Stream.mkString,
           Effect.orDie,

packages/opencode/src/session/session.ts

@@ -4,7 +4,8 @@ import { BackgroundJob } from "@/background/job"
 import { BusEvent } from "@/bus/bus-event"
 import { Bus } from "@/bus"
 import { Decimal } from "decimal.js"
-import { type ProviderMetadata, type LanguageModelUsage } from "ai"
+import { Flag } from "@opencode-ai/core/flag/flag"
+import type { ProviderMetadata, Usage } from "@opencode-ai/llm"
 import { InstallationVersion } from "@opencode-ai/core/installation/version"
 
 import { Database } from "@/storage/db"
@@ -374,21 +375,19 @@ export function plan(input: { slug: string; time: { created: number } }, instanc
   return path.join(base, [input.time.created, input.slug].join("-") + ".md")
 }
 
-export const getUsage = (input: { model: Provider.Model; usage: LanguageModelUsage; metadata?: ProviderMetadata }) => {
+export const getUsage = (input: { model: Provider.Model; usage: Usage; metadata?: ProviderMetadata }) => {
   const safe = (value: number) => {
     if (!Number.isFinite(value)) return 0
     return Math.max(0, value)
   }
   const inputTokens = safe(input.usage.inputTokens ?? 0)
   const outputTokens = safe(input.usage.outputTokens ?? 0)
-  const reasoningTokens = safe(input.usage.outputTokenDetails?.reasoningTokens ?? input.usage.reasoningTokens ?? 0)
+  const reasoningTokens = safe(input.usage.reasoningTokens ?? 0)
 
-  const cacheReadInputTokens = safe(
-    input.usage.inputTokenDetails?.cacheReadTokens ?? input.usage.cachedInputTokens ?? 0,
-  )
+  const cacheReadInputTokens = safe(input.usage.cacheReadInputTokens ?? 0)
   const cacheWriteInputTokens = safe(
     Number(
-      input.usage.inputTokenDetails?.cacheWriteTokens ??
+      input.usage.cacheWriteInputTokens ??
         input.metadata?.["anthropic"]?.["cacheCreationInputTokens"] ??
         // google-vertex-anthropic returns metadata under "vertex" key
         // (AnthropicMessagesLanguageModel custom provider key from 'vertex.anthropic.messages')

packages/opencode/test/effect/runtime-flags.test.ts

@@ -62,6 +62,7 @@ describe("RuntimeFlags", () => {
       expect(flags.experimentalEventSystem).toBe(true)
       expect(flags.experimentalWorkspaces).toBe(true)
       expect(flags.experimentalIconDiscovery).toBe(true)
+      expect(flags.experimentalNativeLlm).toBe(true)
       expect(flags.client).toBe("desktop")
     }),
   )
@@ -80,6 +81,17 @@ describe("RuntimeFlags", () => {
     }),
   )
 
+  it.effect("enables native LLM via dedicated or umbrella flag", () =>
+    Effect.gen(function* () {
+      const explicit = yield* readFlags.pipe(Effect.provide(fromConfig({ OPENCODE_EXPERIMENTAL_NATIVE_LLM: "true" })))
+      const umbrella = yield* readFlags.pipe(Effect.provide(fromConfig({ OPENCODE_EXPERIMENTAL: "true" })))
+
+      expect(explicit.experimentalNativeLlm).toBe(true)
+      expect(umbrella.experimentalNativeLlm).toBe(true)
+    }),
+  )
+
+
   it.effect("layer accepts partial test overrides and fills defaults from Config definitions", () =>
     Effect.gen(function* () {
       const flags = yield* readFlags.pipe(

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 }) =>