Add native OpenAI runtime opt-in

packages/llm/src/tool-runtime.ts

@@ -200,17 +200,17 @@ const dispatch = (tools: Tools, call: ToolCallPart): Effect.Effect<ToolResultVal
   if (!tool.execute)
     return Effect.succeed({ type: "error" as const, value: `Tool has no execute handler: ${call.name}` })
 
-  return decodeAndExecute(tool, call.input).pipe(
+  return decodeAndExecute(tool, call).pipe(
     Effect.catchTag("LLM.ToolFailure", (failure) =>
       Effect.succeed({ type: "error" as const, value: failure.message } satisfies ToolResultValue),
     ),
   )
 }
 
-const decodeAndExecute = (tool: AnyTool, input: unknown): Effect.Effect<ToolResultValue, ToolFailure> =>
-  tool._decode(input).pipe(
+const decodeAndExecute = (tool: AnyTool, call: ToolCallPart): Effect.Effect<ToolResultValue, ToolFailure> =>
+  tool._decode(call.input).pipe(
     Effect.mapError((error) => new ToolFailure({ message: `Invalid tool input: ${error.message}` })),
-    Effect.flatMap((decoded) => tool.execute!(decoded)),
+    Effect.flatMap((decoded) => tool.execute!(decoded, { id: call.id, name: call.name })),
     Effect.flatMap((value) =>
       tool._encode(value).pipe(
         Effect.mapError(

packages/llm/src/tool.ts

@@ -11,6 +11,7 @@ export type ToolSchema<T> = Schema.Codec<T, any, never, never>
 
 export type ToolExecute<Parameters extends ToolSchema<any>, Success extends ToolSchema<any>> = (
   params: Schema.Schema.Type<Parameters>,
+  context?: { readonly id: string; readonly name: string },
 ) => Effect.Effect<Schema.Schema.Type<Success>, ToolFailure>
 
 /**
@@ -61,7 +62,10 @@ type TypedToolConfig = {
 type DynamicToolConfig = {
   readonly description: string
   readonly jsonSchema: JsonSchema.JsonSchema
-  readonly execute?: (params: unknown) => Effect.Effect<unknown, ToolFailure>
+  readonly execute?: (
+    params: unknown,
+    context?: { readonly id: string; readonly name: string },
+  ) => Effect.Effect<unknown, ToolFailure>
 }
 
 /**
@@ -110,7 +114,10 @@ export function make<Parameters extends ToolSchema<any>, Success extends ToolSch
 export function make(config: {
   readonly description: string
   readonly jsonSchema: JsonSchema.JsonSchema
-  readonly execute: (params: unknown) => Effect.Effect<unknown, ToolFailure>
+  readonly execute: (
+    params: unknown,
+    context?: { readonly id: string; readonly name: string },
+  ) => Effect.Effect<unknown, ToolFailure>
 }): AnyExecutableTool
 export function make(config: {
   readonly description: string

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

@@ -11,6 +11,8 @@ type ToolInput = {
 
 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>
@@ -154,14 +156,16 @@ const baseURL = (model: Provider.Model) => {
   throw new Error(`Native LLM request adapter requires a base URL for ${model.providerID}/${model.id}`)
 }
 
-export const model = (model: Provider.Model, headers?: Record<string, string>) => {
+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: baseURL(model),
+    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,
@@ -173,7 +177,7 @@ export const model = (model: Provider.Model, headers?: Record<string, string>) =
 export const request = (input: RequestInput) => {
   const converted = messages(input.messages)
   return LLM.request({
-    model: model(input.model, input.headers),
+    model: model(input, input.headers),
     system: [...(input.system ?? []).map(SystemPart.make), ...converted.system],
     messages: converted.messages,
     tools: tools(input.tools),

packages/opencode/src/session/llm.ts

@@ -2,8 +2,9 @@ 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 type { LLMEvent } from "@opencode-ai/llm"
+import { streamText, wrapLanguageModel, type ModelMessage, type Tool, tool as aiTool, jsonSchema, asSchema } from "ai"
+import { tool as nativeTool, ToolFailure, type JsonSchema, type LLMEvent } from "@opencode-ai/llm"
+import { LLMClient, RequestExecutor } from "@opencode-ai/llm/route"
 import { mergeDeep } from "remeda"
 import { GitLabWorkflowLanguageModel } from "gitlab-ai-provider"
 import { ProviderTransform } from "@/provider/transform"
@@ -17,15 +18,16 @@ import { Flag } from "@opencode-ai/core/flag/flag"
 import { Permission } from "@/permission"
 import { PermissionID } from "@/permission/schema"
 import { Bus } from "@/bus"
+import { errorMessage } from "@/util/error"
 import { Wildcard } from "@/util/wildcard"
 import { SessionID } from "@/session/schema"
 import { Auth } from "@/auth"
-import { Installation } from "@/installation"
 import { InstallationVersion } from "@opencode-ai/core/installation/version"
 import { EffectBridge } from "@/effect/bridge"
 import * as Option from "effect/Option"
 import * as OtelTracer from "@effect/opentelemetry/Tracer"
 import { LLMAISDK } from "./llm-ai-sdk"
+import { LLMNative } from "./llm-native"
 
 const log = Log.create({ service: "llm" })
 export const OUTPUT_TOKEN_MAX = ProviderTransform.OUTPUT_TOKEN_MAX
@@ -34,6 +36,8 @@ export const OUTPUT_TOKEN_MAX = ProviderTransform.OUTPUT_TOKEN_MAX
 const mergeOptions = (target: Record<string, any>, source: Record<string, any> | undefined): Record<string, any> =>
   mergeDeep(target, source ?? {}) as Record<string, any>
 
+const runtime = () => (process.env.OPENCODE_LLM_RUNTIME === "native" ? "native" : "ai-sdk")
+
 export type StreamInput = {
   user: MessageV2.User
   sessionID: string
@@ -213,7 +217,7 @@ 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",
@@ -333,86 +337,123 @@ 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": Flag.OPENCODE_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",
-          }
-        },
-        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": Flag.OPENCODE_CLIENT,
-                "User-Agent": `opencode/${InstallationVersion}`,
+        ...input.model.headers,
+        ...headers,
+      }
+
+      if (runtime() === "native") {
+        if (input.model.providerID !== "openai" || input.model.api.npm !== "@ai-sdk/openai") {
+          return yield* Effect.fail(new Error("Native LLM runtime currently only supports OpenAI models"))
+        }
+        const apiKey =
+          info?.type === "api" ? info.key : typeof item.options.apiKey === "string" ? item.options.apiKey : undefined
+        if (!apiKey) return yield* Effect.fail(new Error("Native LLM runtime requires API key auth for OpenAI"))
+        const baseURL = typeof item.options.baseURL === "string" ? item.options.baseURL : undefined
+        const request = LLMNative.request({
+          model: input.model,
+          apiKey,
+          baseURL,
+          system: isOpenaiOauth ? system : [],
+          messages: ProviderTransform.message(messages, input.model, options),
+          tools: sortedTools,
+          toolChoice: input.toolChoice,
+          temperature: params.temperature,
+          topP: params.topP,
+          topK: params.topK,
+          maxOutputTokens: params.maxOutputTokens,
+          providerOptions: ProviderTransform.providerOptions(input.model, params.options),
+          headers: requestHeaders,
+        })
+        return {
+          type: "native" as const,
+          stream: LLMClient.stream({ request, tools: nativeTools(sortedTools, input) }).pipe(
+            Stream.provide(LLMClient.layer),
+            Stream.provide(RequestExecutor.defaultLayer),
+          ),
+        }
+      }
+
+      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) =>
@@ -426,8 +467,12 @@ const live: Layer.Layer<
 
             const result = yield* run({ ...input, abort: ctrl.signal })
 
+            if (result.type === "native") return result.stream
+
             const state = LLMAISDK.adapterState()
-            return Stream.fromAsyncIterable(result.fullStream, (e) => (e instanceof Error ? e : new Error(String(e)))).pipe(
+            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)),
             )
@@ -458,6 +503,37 @@ function resolveTools(input: Pick<StreamInput, "tools" | "agent" | "permission"
   return Record.filter(input.tools, (_, k) => input.user.tools?.[k] !== false && !disabled.has(k))
 }
 
+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
+}
+
+function nativeTools(tools: Record<string, Tool>, input: StreamRequest) {
+  return Object.fromEntries(
+    Object.entries(tools).map(([name, item]) => [
+      name,
+      nativeTool({
+        description: item.description ?? "",
+        jsonSchema: nativeSchema(item.inputSchema),
+        execute: (args: unknown, ctx?: { readonly id: string; readonly name: string }) =>
+          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) }),
+          }),
+      }),
+    ]),
+  )
+}
+
 // Check if messages contain any tool-call content
 // Used to determine if a dummy tool should be added for LiteLLM proxy compatibility
 export function hasToolCalls(messages: ModelMessage[]): boolean {