refactor(llm): isolate native runtime preview

This commit is contained in:
Kit Langton
2026-05-12 21:59:30 -04:00
parent 8e132a2b82
commit 326b79a6b0
3 changed files with 165 additions and 50 deletions

View File

@@ -0,0 +1,115 @@
import type { Auth } from "@/auth"
import type { Provider } from "@/provider/provider"
import { ProviderTransform } from "@/provider/transform"
import { errorMessage } from "@/util/error"
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 "./llm-native"
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") return { type: "unsupported", reason: "provider is not openai" }
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: "OpenAI OAuth 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: input.headers,
}),
tools: nativeTools(input.tools, input),
}),
}
}
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: 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) }),
}),
}),
]),
)
}
export * as LLMNativeRuntime from "./llm-native-runtime"

View File

@@ -2,8 +2,8 @@ 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 as aiTool, jsonSchema, asSchema } from "ai"
import { tool as nativeTool, ToolFailure, type JsonSchema, type LLMEvent } from "@opencode-ai/llm"
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"
@@ -28,7 +28,7 @@ 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"
import { LLMNativeRuntime } from "./llm-native-runtime"
const log = Log.create({ service: "llm" })
export const OUTPUT_TOKEN_MAX = ProviderTransform.OUTPUT_TOKEN_MAX
@@ -358,31 +358,31 @@ const live: Layer.Layer<
}
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({
const native = LLMNativeRuntime.stream({
model: input.model,
apiKey,
baseURL,
system: isOpenaiOauth ? system : [],
messages: ProviderTransform.message(messages, input.model, options),
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: ProviderTransform.providerOptions(input.model, params.options),
providerOptions: params.options,
headers: requestHeaders,
abort: input.abort,
})
return {
type: "native" as const,
stream: llmClient.stream({ request, tools: nativeTools(sortedTools, input) }),
if (native.type === "supported") {
return {
type: "native" as const,
stream: native.stream,
}
}
l.info("native runtime unavailable; falling back to ai-sdk", { reason: native.reason })
}
return {
@@ -502,37 +502,6 @@ 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 {

View File

@@ -3,6 +3,7 @@ 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"
import { LLMNativeRuntime } from "@/session/llm-native-runtime"
import type { Provider } from "@/provider/provider"
import { ModelID, ProviderID } from "@/provider/schema"
@@ -57,6 +58,15 @@ const baseModel: Provider.Model = {
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[] = [
@@ -234,6 +244,27 @@ describe("session.llm-native.request", () => {
).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("anthropic") },
provider: { ...providerInfo, id: ProviderID.make("anthropic") },
auth: undefined,
}),
).toEqual({ type: "unsupported", reason: "provider is not openai" })
expect(
LLMNativeRuntime.status({
model: baseModel,
provider: providerInfo,
auth: { type: "oauth", refresh: "refresh", access: "access", expires: 1 },
}),
).toEqual({ type: "unsupported", reason: "OpenAI OAuth is not supported" })
})
test("compiles through the native OpenAI Responses route", async () => {
const prepared = await Effect.runPromise(
LLMClient.prepare(