mervexai 0.0.97

package/src/provider/provider.ts

@@ -0,0 +1,2399 @@
+import { LayerNode } from "@msrfteam/core/effect/layer-node"
+import { ProxyAgent } from "undici"
+import os from "os"
+import { ConfigV1 } from "@msrfteam/core/v1/config/config"
+import fuzzysort from "fuzzysort"
+import { Config } from "@/config/config"
+import { mapValues, mergeDeep, omit, pickBy, sortBy } from "remeda"
+import { NoSuchModelError, type Provider as SDK } from "ai"
+import { Npm } from "@msrfteam/core/npm"
+import { Hash } from "@msrfteam/core/util/hash"
+import { Plugin } from "../plugin"
+import { serviceUse } from "@msrfteam/core/effect/service-use"
+import { type LanguageModelV3 } from "@ai-sdk/provider"
+import { ModelsDev } from "@msrfteam/core/models-dev"
+import { Auth } from "../auth"
+import { Env } from "../env"
+import { InstallationVersion } from "@msrfteam/core/installation/version"
+import { iife } from "@/util/iife"
+import { Global } from "@msrfteam/core/global"
+import path from "path"
+import { pathToFileURL } from "url"
+import { Effect, Layer, Context, Schema, Types } from "effect"
+import { EffectBridge } from "@/effect/bridge"
+import { InstanceState } from "@/effect/instance-state"
+import { EffectPromise } from "@/effect/promise"
+import { FSUtil } from "@msrfteam/core/fs-util"
+import { isRecord } from "@/util/record"
+import { optional } from "@msrfteam/core/schema"
+import { ProviderTransform } from "./transform"
+import { ProviderV2 } from "@msrfteam/core/provider"
+import { ModelV2 } from "@msrfteam/core/model"
+import { ModelStatus } from "./model-status"
+import { RuntimeFlags } from "@/effect/runtime-flags"
+import { ProviderError } from "./error"
+
+const OPENAI_HEADER_TIMEOUT_DEFAULT = 300_000
+
+function wrapSSE(res: Response, ms: number, ctl: AbortController) {
+  if (typeof ms !== "number" || ms <= 0) return res
+  if (!res.body) return res
+  if (!res.headers.get("content-type")?.includes("text/event-stream")) return res
+
+  const reader = res.body.getReader()
+  const body = new ReadableStream<Uint8Array>({
+    async pull(ctrl) {
+      const part = await new Promise<Awaited<ReturnType<typeof reader.read>>>((resolve, reject) => {
+        const id = setTimeout(() => {
+          const err = new ProviderError.ResponseStreamError("SSE read timed out")
+          ctl.abort(err)
+          void reader.cancel(err)
+          reject(err)
+        }, ms)
+
+        reader.read().then(
+          (part) => {
+            clearTimeout(id)
+            resolve(part)
+          },
+          (err) => {
+            clearTimeout(id)
+            reject(err)
+          },
+        )
+      })
+
+      if (part.done) {
+        ctrl.close()
+        return
+      }
+
+      ctrl.enqueue(part.value)
+    },
+    async cancel(reason) {
+      ctl.abort(reason)
+      await reader.cancel(reason)
+    },
+  })
+
+  return new Response(body, {
+    headers: new Headers(res.headers),
+    status: res.status,
+    statusText: res.statusText,
+  })
+}
+
+function timeoutController(ms: number) {
+  const ctl = new AbortController()
+  const id = setTimeout(() => ctl.abort(new ProviderError.HeaderTimeoutError(ms)), ms)
+  return {
+    signal: ctl.signal,
+    clear: () => clearTimeout(id),
+  }
+}
+
+function googleVertexAnthropicBaseURL(project: string | undefined, location: string | undefined) {
+  if (!project) return
+  if (location !== "eu" && location !== "us") return
+  // Continental multi-regions require Regional Endpoint Platform domains.
+  return `https://aiplatform.${location}.rep.googleapis.com/v1/projects/${project}/locations/${location}/publishers/anthropic/models`
+}
+
+function googleVertexEndpoint(location: string) {
+  if (location === "global") return "aiplatform.googleapis.com"
+  if (location === "eu" || location === "us") return `aiplatform.${location}.rep.googleapis.com`
+  return `${location}-aiplatform.googleapis.com`
+}
+
+type BundledSDK = {
+  languageModel(modelId: string): LanguageModelV3
+  chat?: (modelId: string) => LanguageModelV3
+  responses?: (modelId: string) => LanguageModelV3
+}
+
+const BUNDLED_PROVIDERS: Record<string, () => Promise<(opts: any) => BundledSDK>> = {
+  "@ai-sdk/amazon-bedrock": () => import("@ai-sdk/amazon-bedrock").then((m) => m.createAmazonBedrock),
+  "@ai-sdk/amazon-bedrock/mantle": () => import("@ai-sdk/amazon-bedrock/mantle").then((m) => m.createBedrockMantle),
+  "@ai-sdk/anthropic": () => import("@ai-sdk/anthropic").then((m) => m.createAnthropic),
+  "@ai-sdk/azure": () => import("@ai-sdk/azure").then((m) => m.createAzure),
+  "@ai-sdk/google": () => import("@ai-sdk/google").then((m) => m.createGoogleGenerativeAI),
+  "@ai-sdk/google-vertex": () => import("@ai-sdk/google-vertex").then((m) => m.createVertex),
+  "@ai-sdk/google-vertex/anthropic": () =>
+    import("@ai-sdk/google-vertex/anthropic").then((m) => m.createVertexAnthropic),
+  "@ai-sdk/openai": () => import("@ai-sdk/openai").then((m) => m.createOpenAI),
+  "@ai-sdk/openai-compatible": () => import("@ai-sdk/openai-compatible").then((m) => m.createOpenAICompatible),
+  "@openrouter/ai-sdk-provider": () => import("@openrouter/ai-sdk-provider").then((m) => m.createOpenRouter),
+  "@ai-sdk/xai": () => import("@ai-sdk/xai").then((m) => m.createXai),
+  "@ai-sdk/mistral": () => import("@ai-sdk/mistral").then((m) => m.createMistral),
+  "@ai-sdk/groq": () => import("@ai-sdk/groq").then((m) => m.createGroq),
+  "@ai-sdk/deepinfra": () => import("@ai-sdk/deepinfra").then((m) => m.createDeepInfra),
+  "@ai-sdk/cerebras": () => import("@ai-sdk/cerebras").then((m) => m.createCerebras),
+  "@ai-sdk/cohere": () => import("@ai-sdk/cohere").then((m) => m.createCohere),
+  "@ai-sdk/gateway": () => import("@ai-sdk/gateway").then((m) => m.createGateway),
+  "@ai-sdk/togetherai": () => import("@ai-sdk/togetherai").then((m) => m.createTogetherAI),
+  "@ai-sdk/perplexity": () => import("@ai-sdk/perplexity").then((m) => m.createPerplexity),
+  "@ai-sdk/vercel": () => import("@ai-sdk/vercel").then((m) => m.createVercel),
+  "@ai-sdk/alibaba": () => import("@ai-sdk/alibaba").then((m) => m.createAlibaba),
+  "gitlab-ai-provider": () => import("gitlab-ai-provider").then((m) => m.createGitLab),
+  "venice-ai-sdk-provider": () => import("venice-ai-sdk-provider").then((m) => m.createVenice),
+}
+
+type CustomModelLoader = (sdk: any, modelID: string, options?: Record<string, any>, model?: Model) => Promise<any>
+type CustomVarsLoader = (options: Record<string, any>) => Record<string, string>
+type CustomDiscoverModels = () => Promise<Record<string, Model>>
+type CustomLoader = (provider: Info) => Effect.Effect<{
+  autoload: boolean
+  getModel?: CustomModelLoader
+  vars?: CustomVarsLoader
+  options?: Record<string, any>
+  discoverModels?: CustomDiscoverModels
+}>
+
+type CustomDep = {
+  auth: (id: string) => Effect.Effect<Auth.Info | undefined>
+  config: () => Effect.Effect<ConfigV1.Info>
+  env: () => Effect.Effect<Record<string, string | undefined>>
+  get: (key: string) => Effect.Effect<string | undefined>
+}
+
+function selectAzureLanguageModel(sdk: any, modelID: string, useChat: boolean) {
+  if (useChat && sdk.chat) return sdk.chat(modelID)
+  if (sdk.responses) return sdk.responses(modelID)
+  if (sdk.messages) return sdk.messages(modelID)
+  if (sdk.chat) return sdk.chat(modelID)
+  return sdk.languageModel(modelID)
+}
+
+function selectBedrockMantleLanguageModel(sdk: BundledSDK, modelID: string) {
+  if (modelID === "openai.gpt-oss-safeguard-20b" || modelID === "openai.gpt-oss-safeguard-120b")
+    return sdk.chat?.(modelID) ?? sdk.languageModel(modelID)
+  return sdk.responses?.(modelID) ?? sdk.languageModel(modelID)
+}
+
+function custom(dep: CustomDep): Record<string, CustomLoader> {
+  return {
+    anthropic: () =>
+      Effect.succeed({
+        autoload: false,
+        options: {
+          headers: {
+            "anthropic-beta": "interleaved-thinking-2025-05-14,fine-grained-tool-streaming-2025-05-14",
+          },
+        },
+      }),
+    mervex: Effect.fnUntraced(function* (input: Info) {
+      const env = yield* dep.env()
+      const hasKey = iife(() => {
+        if (input.env.some((item) => env[item])) return true
+        return false
+      })
+      const ok =
+        hasKey ||
+        Boolean(yield* dep.auth(input.id)) ||
+        Boolean((yield* dep.config()).provider?.["mervex"]?.options?.apiKey)
+
+      if (!ok) {
+        for (const [key, value] of Object.entries(input.models)) {
+          if (value.cost.input === 0) continue
+          delete input.models[key]
+        }
+      }
+
+      return {
+        autoload: Object.keys(input.models).length > 0,
+        options: ok ? { baseURL: MERVEX_BASE_URL } : { apiKey: "public", baseURL: MERVEX_BASE_URL },
+      }
+    }),
+    openai: () =>
+      Effect.succeed({
+        autoload: false,
+        async getModel(sdk: any, modelID: string, _options?: Record<string, any>) {
+          return sdk.responses(modelID)
+        },
+        options: { headerTimeout: OPENAI_HEADER_TIMEOUT_DEFAULT },
+      }),
+    meta: () =>
+      Effect.succeed({
+        autoload: false,
+        async getModel(sdk: any, modelID: string, _options?: Record<string, any>) {
+          return sdk.responses(modelID)
+        },
+      }),
+    xai: () =>
+      Effect.succeed({
+        autoload: false,
+        async getModel(sdk: any, modelID: string, _options?: Record<string, any>) {
+          return sdk.responses(modelID)
+        },
+        options: {},
+      }),
+    azure: Effect.fnUntraced(function* (provider: Info) {
+      const env = yield* dep.env()
+      const auth = yield* dep.auth(provider.id)
+      const resource = iife(() => {
+        return [
+          provider.options?.resourceName,
+          auth?.type === "api" ? auth.metadata?.resourceName : undefined,
+          env["AZURE_RESOURCE_NAME"],
+        ].find((name) => typeof name === "string" && name.trim() !== "")
+      })
+
+      if (!resource && !provider.options?.baseURL) {
+        return {
+          autoload: false,
+          async getModel() {
+            throw new Error(
+              "AZURE_RESOURCE_NAME is missing, set it using env var or reconnecting the azure provider and setting it",
+            )
+          },
+        }
+      }
+
+      return {
+        autoload: false,
+        async getModel(sdk: any, modelID: string, options?: Record<string, any>) {
+          return selectAzureLanguageModel(sdk, modelID, Boolean(options?.["useCompletionUrls"]))
+        },
+        options: {
+          resourceName: resource,
+        },
+        vars(_options): Record<string, string> {
+          if (resource) {
+            return {
+              AZURE_RESOURCE_NAME: resource,
+            }
+          }
+          return {}
+        },
+      }
+    }),
+    "azure-cognitive-services": Effect.fnUntraced(function* (provider: Info) {
+      const resourceName = yield* dep.get("AZURE_COGNITIVE_SERVICES_RESOURCE_NAME")
+      return {
+        autoload: false,
+        async getModel(sdk: any, modelID: string, options?: Record<string, any>) {
+          return selectAzureLanguageModel(sdk, modelID, Boolean(options?.["useCompletionUrls"]))
+        },
+        options: {
+          baseURL: resourceName
+            ? `https://${resourceName}.cognitiveservices.azure.com/openai${provider.options?.useDeploymentBasedUrls ? "" : "/v1"}`
+            : undefined,
+        },
+      }
+    }),
+    "amazon-bedrock": Effect.fnUntraced(function* () {
+      const providerConfig = (yield* dep.config()).provider?.["amazon-bedrock"]
+      const auth = yield* dep.auth("amazon-bedrock")
+      const env = yield* dep.env()
+
+      // Region precedence: 1) config file, 2) env var, 3) default
+      const configRegion = providerConfig?.options?.region
+      const envRegion = env["AWS_REGION"]
+      const defaultRegion = configRegion ?? envRegion ?? "us-east-1"
+
+      // Profile: config file takes precedence over env var
+      const configProfile = providerConfig?.options?.profile
+      const envProfile = env["AWS_PROFILE"]
+      const profile = configProfile ?? envProfile
+
+      const awsAccessKeyId = env["AWS_ACCESS_KEY_ID"]
+      const configApiKey = providerConfig?.options?.apiKey
+
+      // TODO: Using process.env directly because Env.set only updates a process.env shallow copy,
+      // until the scope of the Env API is clarified (test only or runtime?)
+      const awsBearerToken = iife(() => {
+        const envToken = process.env.AWS_BEARER_TOKEN_BEDROCK
+        if (envToken) return envToken
+        if (auth?.type === "api") {
+          process.env.AWS_BEARER_TOKEN_BEDROCK = auth.key
+          return auth.key
+        }
+        return undefined
+      })
+
+      const awsWebIdentityTokenFile = env["AWS_WEB_IDENTITY_TOKEN_FILE"]
+
+      const containerCreds = Boolean(
+        process.env.AWS_CONTAINER_CREDENTIALS_RELATIVE_URI || process.env.AWS_CONTAINER_CREDENTIALS_FULL_URI,
+      )
+
+      if (
+        !profile &&
+        !awsAccessKeyId &&
+        !awsBearerToken &&
+        !configApiKey &&
+        !awsWebIdentityTokenFile &&
+        !containerCreds
+      )
+        return { autoload: false }
+
+      const { fromNodeProviderChain } = yield* Effect.promise(() => import("@aws-sdk/credential-providers"))
+
+      const providerOptions: Record<string, any> = {
+        region: defaultRegion,
+      }
+
+      // Only use credential chain if no bearer token exists
+      // Bearer token takes precedence over credential chain (profiles, access keys, IAM roles, web identity tokens)
+      if (!awsBearerToken && !configApiKey) {
+        // Build credential provider options (only pass profile if specified)
+        const credentialProviderOptions = profile ? { profile } : {}
+
+        providerOptions.credentialProvider = fromNodeProviderChain(credentialProviderOptions)
+      }
+
+      // Add custom endpoint if specified (endpoint takes precedence over baseURL)
+      const endpoint = providerConfig?.options?.endpoint ?? providerConfig?.options?.baseURL
+      if (endpoint) {
+        providerOptions.baseURL = endpoint
+      }
+
+      return {
+        autoload: true,
+        options: providerOptions,
+        vars(options: Record<string, any>) {
+          return { AWS_REGION: options.region ?? defaultRegion }
+        },
+        async getModel(sdk: any, modelID: string, options?: Record<string, any>, model?: Model) {
+          if (model?.api.npm === "@ai-sdk/amazon-bedrock/mantle") return selectBedrockMantleLanguageModel(sdk, modelID)
+
+          // Skip region prefixing if model already has a cross-region inference profile prefix
+          // Models from models.dev may already include prefixes like us., eu., global., etc.
+          const crossRegionPrefixes = ["global.", "us.", "eu.", "jp.", "apac.", "au."]
+          if (crossRegionPrefixes.some((prefix) => modelID.startsWith(prefix))) {
+            return sdk.languageModel(modelID)
+          }
+
+          // Region resolution precedence (highest to lowest):
+          // 1. options.region from mervex.json provider config
+          // 2. defaultRegion from AWS_REGION environment variable
+          // 3. Default "us-east-1" (baked into defaultRegion)
+          const region = options?.region ?? defaultRegion
+
+          let regionPrefix = region.split("-")[0]
+
+          switch (regionPrefix) {
+            case "us": {
+              const modelRequiresPrefix = [
+                "nova-micro",
+                "nova-lite",
+                "nova-pro",
+                "nova-premier",
+                "nova-2",
+                "claude",
+                "deepseek",
+              ].some((m) => modelID.includes(m))
+              const isGovCloud = region.startsWith("us-gov")
+              if (modelRequiresPrefix && !isGovCloud) {
+                modelID = `${regionPrefix}.${modelID}`
+              }
+              break
+            }
+            case "eu": {
+              const regionRequiresPrefix = [
+                "eu-west-1",
+                "eu-west-2",
+                "eu-west-3",
+                "eu-north-1",
+                "eu-central-1",
+                "eu-south-1",
+                "eu-south-2",
+              ].some((r) => region.includes(r))
+              const modelRequiresPrefix = ["claude", "nova-lite", "nova-micro", "llama3", "pixtral"].some((m) =>
+                modelID.includes(m),
+              )
+              if (regionRequiresPrefix && modelRequiresPrefix) {
+                modelID = `${regionPrefix}.${modelID}`
+              }
+              break
+            }
+            case "ap": {
+              const isAustraliaRegion = ["ap-southeast-2", "ap-southeast-4"].includes(region)
+              const isTokyoRegion = region === "ap-northeast-1"
+              if (
+                isAustraliaRegion &&
+                ["anthropic.claude-sonnet-4-5", "anthropic.claude-haiku"].some((m) => modelID.includes(m))
+              ) {
+                regionPrefix = "au"
+                modelID = `${regionPrefix}.${modelID}`
+              } else if (isTokyoRegion) {
+                // Tokyo region uses jp. prefix for cross-region inference
+                const modelRequiresPrefix = ["claude", "nova-lite", "nova-micro", "nova-pro"].some((m) =>
+                  modelID.includes(m),
+                )
+                if (modelRequiresPrefix) {
+                  regionPrefix = "jp"
+                  modelID = `${regionPrefix}.${modelID}`
+                }
+              } else {
+                // Other APAC regions use apac. prefix
+                const modelRequiresPrefix = ["claude", "nova-lite", "nova-micro", "nova-pro"].some((m) =>
+                  modelID.includes(m),
+                )
+                if (modelRequiresPrefix) {
+                  regionPrefix = "apac"
+                  modelID = `${regionPrefix}.${modelID}`
+                }
+              }
+              break
+            }
+          }
+
+          return sdk.languageModel(modelID)
+        },
+      }
+    }),
+    llmgateway: () =>
+      Effect.succeed({
+        autoload: false,
+        options: {
+          headers: {
+            "HTTP-Referer": "https://mervex.ai/",
+            "X-Title": "mervex",
+            "X-Source": "mervex",
+          },
+        },
+      }),
+    openrouter: () =>
+      Effect.succeed({
+        autoload: false,
+        options: {
+          headers: {
+            "HTTP-Referer": "https://mervex.ai/",
+            "X-Title": "mervex",
+          },
+        },
+      }),
+    nvidia: (provider) =>
+      Effect.succeed({
+        autoload: provider.source === "config",
+        options: {
+          headers: {
+            "HTTP-Referer": "https://mervex.ai/",
+            "X-Title": "mervex",
+            "X-BILLING-INVOKE-ORIGIN": "Mervex",
+          },
+        },
+      }),
+    vercel: () =>
+      Effect.succeed({
+        autoload: false,
+        options: {
+          headers: {
+            "http-referer": "https://mervex.ai/",
+            "x-title": "mervex",
+          },
+        },
+      }),
+    "google-vertex": Effect.fnUntraced(function* (provider: Info) {
+      const env = yield* dep.env()
+      // models.dev advertises GOOGLE_VERTEX_PROJECT for Vertex; keep the wider
+      // Google Cloud project env names as fallbacks for existing ADC setups.
+      const project =
+        provider.options?.project ??
+        env["GOOGLE_VERTEX_PROJECT"] ??
+        env["GOOGLE_CLOUD_PROJECT"] ??
+        env["GCP_PROJECT"] ??
+        env["GCLOUD_PROJECT"]
+
+      const location = String(
+        provider.options?.location ??
+          env["GOOGLE_VERTEX_LOCATION"] ??
+          env["GOOGLE_CLOUD_LOCATION"] ??
+          env["VERTEX_LOCATION"] ??
+          "us-central1",
+      )
+
+      const autoload = Boolean(project)
+      if (!autoload) return { autoload: false }
+      return {
+        autoload: true,
+        vars(_options: Record<string, any>) {
+          return {
+            ...(project && { GOOGLE_VERTEX_PROJECT: project }),
+            GOOGLE_VERTEX_LOCATION: location,
+            GOOGLE_VERTEX_ENDPOINT: googleVertexEndpoint(location),
+          }
+        },
+        options: {
+          project,
+          location,
+          fetch: async (input: RequestInfo | URL, init?: RequestInit) => {
+            const { GoogleAuth } = await import("google-auth-library")
+            const auth = new GoogleAuth({ scopes: ["https://www.googleapis.com/auth/cloud-platform"] })
+            const client = await auth.getClient()
+            const token = await client.getAccessToken()
+
+            const headers = new Headers(init?.headers)
+            headers.set("Authorization", `Bearer ${token.token}`)
+
+            return fetch(input, { ...init, headers })
+          },
+        },
+        async getModel(sdk: any, modelID: string) {
+          const id = String(modelID).trim()
+          return sdk.languageModel(id)
+        },
+      }
+    }),
+    "google-vertex-anthropic": Effect.fnUntraced(function* () {
+      const env = yield* dep.env()
+      const project = env["GOOGLE_CLOUD_PROJECT"] ?? env["GCP_PROJECT"] ?? env["GCLOUD_PROJECT"]
+      const location = env["GOOGLE_CLOUD_LOCATION"] ?? env["VERTEX_LOCATION"] ?? "global"
+      const autoload = Boolean(project)
+      if (!autoload) return { autoload: false }
+      const baseURL = googleVertexAnthropicBaseURL(project, location)
+      return {
+        autoload: true,
+        options: {
+          project,
+          location,
+          ...(baseURL && { baseURL }),
+        },
+        async getModel(sdk: any, modelID) {
+          const id = String(modelID).trim()
+          return sdk.languageModel(id)
+        },
+      }
+    }),
+    "sap-ai-core": Effect.fnUntraced(function* () {
+      const auth = yield* dep.auth("sap-ai-core")
+      // TODO: Using process.env directly because Env.set only updates a shallow copy (not process.env),
+      // until the scope of the Env API is clarified (test only or runtime?)
+      const envServiceKey = iife(() => {
+        const envAICoreServiceKey = process.env.AICORE_SERVICE_KEY
+        if (envAICoreServiceKey) return envAICoreServiceKey
+        if (auth?.type === "api") {
+          process.env.AICORE_SERVICE_KEY = auth.key
+          return auth.key
+        }
+        return undefined
+      })
+      const deploymentId = process.env.AICORE_DEPLOYMENT_ID
+      const resourceGroup = process.env.AICORE_RESOURCE_GROUP
+
+      return {
+        autoload: !!envServiceKey,
+        options: envServiceKey ? { deploymentId, resourceGroup } : {},
+        async getModel(sdk: any, modelID: string) {
+          return sdk(modelID)
+        },
+      }
+    }),
+    zenmux: () =>
+      Effect.succeed({
+        autoload: false,
+        options: {
+          headers: {
+            "HTTP-Referer": "https://mervex.ai/",
+            "X-Title": "mervex",
+          },
+        },
+      }),
+    gitlab: Effect.fnUntraced(function* (input: Info) {
+      const {
+        VERSION: GITLAB_PROVIDER_VERSION,
+        isWorkflowModel,
+        discoverWorkflowModels,
+      } = yield* Effect.promise(() => import("gitlab-ai-provider"))
+
+      const instanceUrl = (yield* dep.get("GITLAB_INSTANCE_URL")) || "https://gitlab.com"
+
+      const auth = yield* dep.auth(input.id)
+      const apiKey = auth?.type === "oauth" ? auth.access : auth?.type === "api" ? auth.key : undefined
+      const token = apiKey ?? (yield* dep.get("GITLAB_TOKEN"))
+
+      const providerConfig = (yield* dep.config()).provider?.["gitlab"]
+      const directory = yield* InstanceState.directory
+
+      const aiGatewayHeaders = {
+        "User-Agent": `mervex/${InstallationVersion} gitlab-ai-provider/${GITLAB_PROVIDER_VERSION} (${os.platform()} ${os.release()}; ${os.arch()})`,
+        "anthropic-beta": "context-1m-2025-08-07",
+        ...providerConfig?.options?.aiGatewayHeaders,
+      }
+
+      const featureFlags = {
+        duo_agent_platform_agentic_chat: true,
+        duo_agent_platform: true,
+        ...providerConfig?.options?.featureFlags,
+      }
+
+      return {
+        autoload: !!token,
+        options: {
+          instanceUrl,
+          apiKey: token,
+          aiGatewayHeaders,
+          featureFlags,
+        },
+        async getModel(sdk: any, modelID: string, options?: Record<string, any>) {
+          if (modelID.startsWith("duo-workflow-")) {
+            const workflowRef = typeof options?.workflowRef === "string" ? options.workflowRef : undefined
+            // Use the static mapping if it exists, otherwise use duo-workflow with selectedModelRef
+            const sdkModelID = isWorkflowModel(modelID) ? modelID : "duo-workflow"
+            const workflowDefinition =
+              typeof options?.workflowDefinition === "string" ? options.workflowDefinition : undefined
+            const model = sdk.workflowChat(sdkModelID, {
+              featureFlags,
+              workflowDefinition,
+            })
+            if (workflowRef) {
+              model.selectedModelRef = workflowRef
+            }
+            return model
+          }
+          return sdk.agenticChat(modelID, {
+            aiGatewayHeaders,
+            featureFlags,
+          })
+        },
+        async discoverModels(): Promise<Record<string, Model>> {
+          if (!apiKey) {
+            return {}
+          }
+
+          try {
+            const token = apiKey
+            const getHeaders = (): Record<string, string> =>
+              auth?.type === "api" ? { "PRIVATE-TOKEN": token } : { Authorization: `Bearer ${token}` }
+
+            const result = await discoverWorkflowModels({ instanceUrl, getHeaders }, { workingDirectory: directory })
+
+            if (!result.models.length) {
+              return {}
+            }
+
+            const models: Record<string, Model> = {}
+            for (const m of result.models) {
+              if (!input.models[m.id]) {
+                models[m.id] = {
+                  id: ModelV2.ID.make(m.id),
+                  providerID: ProviderV2.ID.make("gitlab"),
+                  name: `Agent Platform (${m.name})`,
+                  family: "",
+                  api: {
+                    id: m.id,
+                    url: instanceUrl,
+                    npm: "gitlab-ai-provider",
+                  },
+                  status: "active",
+                  headers: {},
+                  options: { workflowRef: m.ref },
+                  cost: { input: 0, output: 0, cache: { read: 0, write: 0 } },
+                  limit: { context: m.context, output: m.output },
+                  capabilities: {
+                    temperature: false,
+                    reasoning: true,
+                    attachment: true,
+                    toolcall: true,
+                    input: {
+                      text: true,
+                      audio: false,
+                      image: true,
+                      video: false,
+                      pdf: true,
+                    },
+                    output: {
+                      text: true,
+                      audio: false,
+                      image: false,
+                      video: false,
+                      pdf: false,
+                    },
+                    interleaved: false,
+                  },
+                  release_date: "",
+                  variants: {},
+                }
+              }
+            }
+
+            return models
+          } catch (e) {
+            return {}
+          }
+        },
+      }
+    }),
+    "cloudflare-workers-ai": Effect.fnUntraced(function* (input: Info) {
+      // When baseURL is already configured (e.g. corporate config routing through a proxy/gateway),
+      // skip the account ID check because the URL is already fully specified.
+      if (input.options?.baseURL) return { autoload: false }
+
+      const auth = yield* dep.auth(input.id)
+      const env = yield* dep.env()
+      const accountId = env["CLOUDFLARE_ACCOUNT_ID"] || (auth?.type === "api" ? auth.metadata?.accountId : undefined)
+      if (!accountId)
+        return {
+          autoload: false,
+          async getModel() {
+            throw new Error(
+              "CLOUDFLARE_ACCOUNT_ID is missing. Set it with: export CLOUDFLARE_ACCOUNT_ID=<your-account-id>",
+            )
+          },
+        }
+
+      const apiKey = env["CLOUDFLARE_API_KEY"] || (auth?.type === "api" ? auth.key : undefined)
+
+      return {
+        autoload: !!apiKey,
+        options: {
+          apiKey,
+          headers: {
+            "User-Agent": `mervex/${InstallationVersion} cloudflare-workers-ai (${os.platform()} ${os.release()}; ${os.arch()})`,
+          },
+        },
+        async getModel(sdk: any, modelID: string) {
+          return sdk.languageModel(modelID)
+        },
+        vars(_options) {
+          return {
+            CLOUDFLARE_ACCOUNT_ID: accountId,
+          }
+        },
+      }
+    }),
+    "cloudflare-ai-gateway": Effect.fnUntraced(function* (input: Info) {
+      // When baseURL is already configured (e.g. corporate config), skip the ID checks.
+      if (input.options?.baseURL) return { autoload: false }
+
+      const auth = yield* dep.auth(input.id)
+      const env = yield* dep.env()
+      const accountId = env["CLOUDFLARE_ACCOUNT_ID"] || (auth?.type === "api" ? auth.metadata?.accountId : undefined)
+      // The Cloudflare auth prompt stores this value as gatewayId metadata.
+      const gateway = env["CLOUDFLARE_GATEWAY_ID"] || (auth?.type === "api" ? auth.metadata?.gatewayId : undefined)
+
+      if (!accountId || !gateway) {
+        const missing = [
+          !accountId ? "CLOUDFLARE_ACCOUNT_ID" : undefined,
+          !gateway ? "CLOUDFLARE_GATEWAY_ID" : undefined,
+        ].filter((x): x is string => Boolean(x))
+        return {
+          autoload: false,
+          async getModel() {
+            throw new Error(
+              `${missing.join(" and ")} missing. Set with: ${missing.map((x) => `export ${x}=<value>`).join(" && ")}`,
+            )
+          },
+        }
+      }
+
+      // Get API token from env or auth - required for authenticated gateways
+      const apiToken =
+        env["CLOUDFLARE_API_TOKEN"] || env["CF_AIG_TOKEN"] || (auth?.type === "api" ? auth.key : undefined)
+
+      if (!apiToken) {
+        throw new Error(
+          "CLOUDFLARE_API_TOKEN (or CF_AIG_TOKEN) is required for Cloudflare AI Gateway. " +
+            "Set it via environment variable or run `mervex auth cloudflare-ai-gateway`.",
+        )
+      }
+
+      const { createAiGateway } = yield* Effect.promise(() => import("ai-gateway-provider"))
+      const { createUnified } = yield* Effect.promise(() => import("ai-gateway-provider/providers/unified"))
+      const { createOpenAI } = yield* Effect.promise(() => import("ai-gateway-provider/providers/openai"))
+      const { createAnthropic } = yield* Effect.promise(() => import("ai-gateway-provider/providers/anthropic"))
+      const { createOpenAICompatible } = yield* Effect.promise(() => import("@ai-sdk/openai-compatible"))
+
+      const metadata = iife(() => {
+        if (input.options?.metadata) return input.options.metadata
+        try {
+          return JSON.parse(input.options?.headers?.["cf-aig-metadata"])
+        } catch {
+          return undefined
+        }
+      })
+      const opts = {
+        metadata,
+        cacheTtl: input.options?.cacheTtl,
+        cacheKey: input.options?.cacheKey,
+        skipCache: input.options?.skipCache,
+        collectLog: input.options?.collectLog,
+        headers: {
+          "User-Agent": `mervex/${InstallationVersion} cloudflare-ai-gateway (${os.platform()} ${os.release()}; ${os.arch()})`,
+        },
+      }
+
+      const aigateway = createAiGateway({
+        accountId,
+        gateway,
+        apiKey: apiToken,
+        ...(Object.values(opts).some((v) => v !== undefined) ? { options: opts } : {}),
+      })
+      return {
+        autoload: true,
+        async getModel(_sdk: any, modelID: string, _options?: Record<string, any>) {
+          // Model IDs use Unified API format: provider/model (e.g., "anthropic/claude-sonnet-4-5").
+          // OpenAI and Anthropic ride their native passthrough routes so agents get the Responses
+          // and Messages APIs; new OpenAI models reject tools+reasoning_effort on chat completions.
+          // The passthrough wrappers inject a CF_TEMP_TOKEN sentinel that the gateway strips before
+          // dispatch, so upstream billing stays on the gateway (Unified Billing / stored BYOK).
+          if (modelID.startsWith("openai/")) return aigateway(createOpenAI()(modelID.slice("openai/".length)))
+          // models.dev lists Anthropic ids with dotted versions (claude-haiku-4.5); Anthropic's
+          // Messages API expects dashed native slugs (claude-haiku-4-5), so translate before passing.
+          // No native Anthropic slug contains a dot, so the blanket replacement is lossless here -
+          // unlike OpenAI above, whose native ids (e.g. gpt-4.1) keep their dots and must not be touched.
+          if (modelID.startsWith("anthropic/"))
+            return aigateway(createAnthropic()(modelID.slice("anthropic/".length).replaceAll(".", "-")))
+          // Workers AI is the only first-party provider whose upstream is Cloudflare itself, so it is
+          // the only one that should receive the Cloudflare token as its upstream Authorization header.
+          // The Unified API addresses Workers AI both with the explicit "workers-ai/" prefix and as
+          // bare "@cf/..." ids. Third-party providers must not receive the token; they rely on the
+          // gateway's stored/BYOK keys instead.
+          // Workers AI is Cloudflare's own upstream, so it rides the unified compat route with the
+          // Cloudflare token as its upstream Authorization header.
+          const isWorkersAi = modelID.startsWith("workers-ai/") || modelID.startsWith("@cf/")
+          if (isWorkersAi) return aigateway(createUnified({ apiKey: apiToken })(modelID))
+
+          // Every other third-party provider (google, xai, alibaba, deepseek, moonshotai, …) is only
+          // served by Cloudflare's catalog-aware REST API. The universal/compat gateway route rejects
+          // them with "Invalid provider" (the gateway's compat endpoint doesn't front those upstreams),
+          // so point an OpenAI-compatible client at the REST endpoint and bind it to the gateway with
+          // cf-aig-gateway-id — that keeps requests gateway-routed (analytics/caching/BYOK), not a
+          // bypass. models.dev ids (provider/model, dotted) pass through unchanged.
+          return createOpenAICompatible({
+            name: "cloudflare-ai-gateway",
+            baseURL: `https://api.cloudflare.com/client/v4/accounts/${accountId}/ai/v1`,
+            apiKey: apiToken,
+            headers: { "cf-aig-gateway-id": gateway },
+          })(modelID)
+        },
+        options: {},
+      }
+    }),
+    cerebras: () =>
+      Effect.succeed({
+        autoload: false,
+        options: {
+          headers: {
+            "X-Cerebras-3rd-Party-Integration": "mervex",
+          },
+        },
+      }),
+    kilo: () =>
+      Effect.succeed({
+        autoload: false,
+        options: {
+          headers: {
+            "HTTP-Referer": "https://mervex.ai/",
+            "X-Title": "mervex",
+          },
+        },
+      }),
+    "snowflake-cortex": Effect.fnUntraced(function* (input: Info) {
+      const env = yield* dep.env()
+      const auth = yield* dep.auth(input.id)
+
+      const account =
+        env["SNOWFLAKE_ACCOUNT"] ??
+        (auth?.type === "api" ? auth.metadata?.account : undefined) ??
+        (auth?.type === "oauth" ? auth.accountId : undefined) ??
+        input.options?.account
+
+      const envToken = env["SNOWFLAKE_CORTEX_TOKEN"] ?? env["SNOWFLAKE_CORTEX_PAT"]
+      const apiKeyToken = auth?.type === "api" ? auth.key : undefined
+      const oauthToken = auth?.type === "oauth" ? auth.access : undefined
+      const configToken = input.options?.token ?? input.options?.apiKey
+
+      const token = envToken ?? apiKeyToken ?? oauthToken ?? configToken
+
+      if (!account || !token) {
+        const missing = [!account && "SNOWFLAKE_ACCOUNT", !token && "SNOWFLAKE_CORTEX_TOKEN"].filter(Boolean).join(", ")
+        return {
+          autoload: false,
+          async getModel() {
+            throw new Error(
+              `Snowflake Cortex: missing credentials (${missing}). Provide a bearer token (OAuth, JWT, or PAT) via env var, mervex auth, or provider options.`,
+            )
+          },
+        }
+      }
+
+      const baseURL = `https://${account}.snowflakecomputing.com/api/v2/cortex/v1`
+
+      const options: Record<string, any> = { baseURL, apiKey: token }
+
+      // Only skip provider-level fetch when the token is from OAuth with no override.
+      // For OAuth tokens, the plugin auth loader's combined fetch handles
+      // OAuth refresh + snowflake transformations in one place.
+      // For env/config/API-key tokens, the provider fetch applies snowflake
+      // transformations directly.
+      const useOAuthHandler =
+        oauthToken !== undefined && envToken === undefined && apiKeyToken === undefined && configToken === undefined
+      if (!useOAuthHandler) {
+        options.fetch = async (url: RequestInfo | URL, init?: RequestInit) => {
+          if (init?.body && typeof init.body === "string") {
+            try {
+              const body = JSON.parse(init.body)
+              if ("max_tokens" in body) {
+                body.max_completion_tokens = body.max_tokens
+                delete body.max_tokens
+                init = { ...init, body: JSON.stringify(body) }
+              }
+            } catch {}
+          }
+
+          const response = await fetch(url, init)
+
+          if (!response.ok && response.status === 400) {
+            try {
+              const errorData = await response.clone().json()
+              const errorMessage = String(errorData.message || errorData.error || "")
+              if (errorMessage.toLowerCase().includes("conversation complete")) {
+                return new Response(
+                  JSON.stringify({
+                    choices: [{ finish_reason: "stop", message: { content: "", role: "assistant" } }],
+                  }),
+                  { status: 200, headers: new Headers({ "content-type": "application/json" }) },
+                )
+              }
+            } catch {}
+          }
+
+          if (response.body && response.headers.get("content-type")?.includes("text/event-stream")) {
+            const reader = response.body.getReader()
+            const encoder = new TextEncoder()
+            const decoder = new TextDecoder()
+            const stream = new ReadableStream({
+              async pull(ctrl) {
+                const { done, value } = await reader.read()
+                if (done) {
+                  ctrl.close()
+                  return
+                }
+                const text = decoder.decode(value, { stream: true })
+                ctrl.enqueue(encoder.encode(text.replace(/"role"\s*:\s*""/g, '"role":"assistant"')))
+              },
+              cancel() {
+                reader.cancel()
+              },
+            })
+            return new Response(stream, { headers: response.headers, status: response.status })
+          }
+
+          return response
+        }
+      }
+
+      return {
+        autoload: input.source === "config",
+        options,
+      }
+    }),
+  }
+}
+
+const ProviderApiInfo = Schema.Struct({
+  id: Schema.String,
+  url: Schema.String,
+  npm: Schema.String,
+})
+
+const ProviderModalities = Schema.Struct({
+  text: Schema.Boolean,
+  audio: Schema.Boolean,
+  image: Schema.Boolean,
+  video: Schema.Boolean,
+  pdf: Schema.Boolean,
+})
+
+const ProviderInterleavedField = Schema.Union([
+  Schema.Literals(["reasoning", "reasoning_content", "reasoning_text"]),
+  Schema.String,
+])
+
+const ProviderInterleaved = Schema.Union([
+  Schema.Boolean,
+  Schema.Struct({
+    field: ProviderInterleavedField,
+  }),
+])
+
+const ProviderCapabilities = Schema.Struct({
+  temperature: Schema.Boolean,
+  reasoning: Schema.Boolean,
+  attachment: Schema.Boolean,
+  toolcall: Schema.Boolean,
+  input: ProviderModalities,
+  output: ProviderModalities,
+  interleaved: ProviderInterleaved,
+})
+
+const ProviderCacheCost = Schema.Struct({
+  read: Schema.Finite,
+  write: Schema.Finite,
+})
+
+const ProviderCostTier = Schema.Struct({
+  input: Schema.Finite,
+  output: Schema.Finite,
+  cache: ProviderCacheCost,
+  tier: Schema.Struct({
+    type: Schema.Literal("context"),
+    size: Schema.Finite,
+  }),
+})
+
+const ProviderCost = Schema.Struct({
+  input: Schema.Finite,
+  output: Schema.Finite,
+  cache: ProviderCacheCost,
+  tiers: optional(Schema.Array(ProviderCostTier)),
+  experimentalOver200K: optional(
+    Schema.Struct({
+      input: Schema.Finite,
+      output: Schema.Finite,
+      cache: ProviderCacheCost,
+    }),
+  ),
+})
+
+const ProviderLimit = Schema.Struct({
+  context: Schema.Finite,
+  input: optional(Schema.Finite),
+  output: Schema.Finite,
+})
+
+export const Model = Schema.Struct({
+  id: ModelV2.ID,
+  providerID: ProviderV2.ID,
+  api: ProviderApiInfo,
+  name: Schema.String,
+  family: optional(Schema.String),
+  capabilities: ProviderCapabilities,
+  cost: ProviderCost,
+  limit: ProviderLimit,
+  status: ModelStatus,
+  options: Schema.Record(Schema.String, Schema.Any),
+  headers: Schema.Record(Schema.String, Schema.String),
+  release_date: Schema.String,
+  variants: optional(Schema.Record(Schema.String, Schema.Record(Schema.String, Schema.Any))),
+}).annotate({ identifier: "Model" })
+export type Model = Types.DeepMutable<Schema.Schema.Type<typeof Model>>
+
+export const Info = Schema.Struct({
+  id: ProviderV2.ID,
+  name: Schema.String,
+  source: Schema.Literals(["env", "config", "custom", "api"]),
+  env: Schema.Array(Schema.String),
+  key: optional(Schema.String),
+  options: Schema.Record(Schema.String, Schema.Any),
+  models: Schema.Record(Schema.String, Model),
+  rateLimit: optional(
+    Schema.Struct({
+      rpm: optional(Schema.Number),
+      tpm: optional(Schema.Number),
+      rpd: optional(Schema.Number),
+      tpd: optional(Schema.Number),
+      maxInputTokens: optional(Schema.Number),
+    }),
+  ),
+}).annotate({ identifier: "Provider" })
+export type Info = Types.DeepMutable<Schema.Schema.Type<typeof Info>>
+
+const DefaultModelIDs = Schema.Record(Schema.String, Schema.String)
+
+export const ListResult = Schema.Struct({
+  all: Schema.Array(Info),
+  default: DefaultModelIDs,
+  connected: Schema.Array(Schema.String),
+})
+export type ListResult = Types.DeepMutable<Schema.Schema.Type<typeof ListResult>>
+
+export const ConfigProvidersResult = Schema.Struct({
+  providers: Schema.Array(Info),
+  default: DefaultModelIDs,
+})
+export type ConfigProvidersResult = Types.DeepMutable<Schema.Schema.Type<typeof ConfigProvidersResult>>
+
+export function toPublicInfo(provider: Info): Info {
+  return JSON.parse(
+    JSON.stringify(
+      {
+        ...provider,
+        models: Object.fromEntries(Object.entries(provider.models).filter(([, model]) => Schema.is(Model)(model))),
+      },
+      (_, value) => {
+        if (typeof value === "function" || typeof value === "symbol" || value === undefined) return undefined
+        if (typeof value === "bigint") return value.toString()
+        return value
+      },
+    ),
+  )
+}
+
+export function defaultModelIDs<T extends { models: Record<string, { id: string }> }>(providers: Record<string, T>) {
+  return mapValues(providers, (item) => sort(Object.values(item.models))[0].id)
+}
+
+export class ModelNotFoundError extends Schema.TaggedErrorClass<ModelNotFoundError>()("ProviderModelNotFoundError", {
+  providerID: ProviderV2.ID,
+  modelID: ModelV2.ID,
+  suggestions: Schema.optional(Schema.Array(Schema.String)),
+  cause: Schema.optional(Schema.Defect()),
+}) {
+  override get message() {
+    const suggestions = this.suggestions?.length ? ` Did you mean: ${this.suggestions.join(", ")}?` : ""
+    return `Model not found: ${this.providerID}/${this.modelID}.${suggestions}`
+  }
+
+  static isInstance(input: unknown): input is ModelNotFoundError {
+    return input instanceof ModelNotFoundError
+  }
+}
+
+export class InitError extends Schema.TaggedErrorClass<InitError>()("ProviderInitError", {
+  providerID: ProviderV2.ID,
+  cause: Schema.optional(Schema.Defect()),
+}) {
+  override get message() {
+    return `Failed to initialize provider: ${this.providerID}`
+  }
+
+  static isInstance(input: unknown): input is InitError {
+    return input instanceof InitError
+  }
+}
+
+export class NoProvidersError extends Schema.TaggedErrorClass<NoProvidersError>()("ProviderNoProvidersError", {}) {
+  override get message() {
+    return "No providers are available"
+  }
+
+  static isInstance(input: unknown): input is NoProvidersError {
+    return input instanceof NoProvidersError
+  }
+}
+
+export class NoModelsError extends Schema.TaggedErrorClass<NoModelsError>()("ProviderNoModelsError", {
+  providerID: ProviderV2.ID,
+}) {
+  override get message() {
+    return `No models are available for provider: ${this.providerID}`
+  }
+
+  static isInstance(input: unknown): input is NoModelsError {
+    return input instanceof NoModelsError
+  }
+}
+
+export type DefaultModelError = ModelNotFoundError | NoProvidersError | NoModelsError
+export type Error = ModelNotFoundError | InitError | NoProvidersError | NoModelsError
+
+export interface Interface {
+  readonly list: () => Effect.Effect<Record<ProviderV2.ID, Info>>
+  readonly catalog: () => Effect.Effect<Record<ProviderV2.ID, Info>>
+  readonly getProvider: (providerID: ProviderV2.ID) => Effect.Effect<Info>
+  readonly getModel: (providerID: ProviderV2.ID, modelID: ModelV2.ID) => Effect.Effect<Model, ModelNotFoundError>
+  readonly getLanguage: (model: Model) => Effect.Effect<LanguageModelV3, ModelNotFoundError>
+  readonly closest: (
+    providerID: ProviderV2.ID,
+    query: string[],
+  ) => Effect.Effect<{ providerID: ProviderV2.ID; modelID: string } | undefined>
+  readonly getSmallModel: (providerID: ProviderV2.ID) => Effect.Effect<Model | undefined>
+  readonly defaultModel: () => Effect.Effect<{ providerID: ProviderV2.ID; modelID: ModelV2.ID }, DefaultModelError>
+}
+
+interface State {
+  models: Map<string, LanguageModelV3>
+  providers: Record<ProviderV2.ID, Info>
+  catalog: Record<ProviderV2.ID, Info>
+  sdk: Map<string, BundledSDK>
+  modelLoaders: Record<string, CustomModelLoader>
+  varsLoaders: Record<string, CustomVarsLoader>
+}
+
+export class Service extends Context.Service<Service, Interface>()("@mervex/Provider") {}
+
+export const use = serviceUse(Service)
+
+function cost(c: ModelsDev.Model["cost"]): Model["cost"] {
+  const result: Model["cost"] = {
+    input: c?.input ?? 0,
+    output: c?.output ?? 0,
+    cache: {
+      read: c?.cache_read ?? 0,
+      write: c?.cache_write ?? 0,
+    },
+  }
+  if (c?.tiers) {
+    result.tiers = c.tiers.map((item) => ({
+      input: item.input,
+      output: item.output,
+      cache: {
+        read: item.cache_read ?? 0,
+        write: item.cache_write ?? 0,
+      },
+      tier: item.tier,
+    }))
+  }
+  if (c?.context_over_200k) {
+    result.experimentalOver200K = {
+      cache: {
+        read: c.context_over_200k.cache_read ?? 0,
+        write: c.context_over_200k.cache_write ?? 0,
+      },
+      input: c.context_over_200k.input,
+      output: c.context_over_200k.output,
+    }
+  }
+  return result
+}
+
+// Cloudflare AI Gateway routes OpenAI and Anthropic models through their native
+// passthrough SDKs (Responses / Messages APIs). Resolving the native npm before
+// variants are computed makes reasoning variants produce payloads the native
+// SDKs understand (e.g. anthropic `effort` instead of compat `reasoningEffort`).
+function cloudflareGatewayNpm(providerID: string, modelID: string) {
+  if (providerID !== "cloudflare-ai-gateway") return undefined
+  if (modelID.startsWith("openai/")) return "@ai-sdk/openai"
+  if (modelID.startsWith("anthropic/")) return "@ai-sdk/anthropic"
+  return undefined
+}
+
+function fromModelsDevModel(provider: ModelsDev.Provider, model: ModelsDev.Model): Model {
+  const base: Model = {
+    id: ModelV2.ID.make(model.id),
+    providerID: ProviderV2.ID.make(provider.id),
+    name: model.name,
+    family: model.family,
+    api: {
+      id: model.id,
+      url: model.provider?.api ?? provider.api ?? "",
+      npm:
+        cloudflareGatewayNpm(provider.id, model.id) ??
+        model.provider?.npm ??
+        provider.npm ??
+        "@ai-sdk/openai-compatible",
+    },
+    status: model.status ?? "active",
+    headers: {},
+    options: {},
+    cost: cost(model.cost),
+    limit: {
+      context: model.limit.context,
+      input: model.limit.input,
+      output: model.limit.output,
+    },
+    capabilities: {
+      temperature: model.temperature ?? false,
+      reasoning: model.reasoning ?? false,
+      attachment: model.attachment ?? false,
+      toolcall: model.tool_call ?? true,
+      input: {
+        text: model.modalities?.input?.includes("text") ?? false,
+        audio: model.modalities?.input?.includes("audio") ?? false,
+        image: model.modalities?.input?.includes("image") ?? false,
+        video: model.modalities?.input?.includes("video") ?? false,
+        pdf: model.modalities?.input?.includes("pdf") ?? false,
+      },
+      output: {
+        text: model.modalities?.output?.includes("text") ?? false,
+        audio: model.modalities?.output?.includes("audio") ?? false,
+        image: model.modalities?.output?.includes("image") ?? false,
+        video: model.modalities?.output?.includes("video") ?? false,
+        pdf: model.modalities?.output?.includes("pdf") ?? false,
+      },
+      interleaved: typeof model.interleaved === "string" ? { field: model.interleaved } : (model.interleaved ?? false),
+    },
+    release_date: model.release_date ?? "",
+    variants: {},
+  }
+
+  const variants = ProviderTransform.reasoningVariants(model, base) ?? ProviderTransform.variants(base)
+
+  return {
+    ...base,
+    variants: mapValues(variants, (v) => v),
+  }
+}
+
+export function fromModelsDevProvider(provider: ModelsDev.Provider): Info {
+  const models: Record<string, Model> = {}
+  for (const [key, model] of Object.entries(provider.models)) {
+    models[key] = fromModelsDevModel(provider, model)
+    for (const [mode, opts] of Object.entries(model.experimental?.modes ?? {})) {
+      const id = `${model.id}-${mode}`
+      const base = fromModelsDevModel(provider, model)
+      models[id] = {
+        ...base,
+        id: ModelV2.ID.make(id),
+        name: `${model.name} ${mode[0].toUpperCase()}${mode.slice(1)}`,
+        cost: opts.cost ? mergeDeep(base.cost, cost(opts.cost)) : base.cost,
+        options: modeOptions(base, opts.provider?.body),
+        headers: opts.provider?.headers ?? base.headers,
+      }
+    }
+  }
+  return {
+    id: ProviderV2.ID.make(provider.id),
+    source: "custom",
+    name: provider.name,
+    env: [...(provider.env ?? [])],
+    options: {},
+    models,
+  }
+}
+
+const MERVEX_BASE_URL = "https://api.mervex.cc.cd/v1"
+
+const GEEKHUB_BASE_URL = "https://api.geekhub.mx/v1"
+
+function geekhubModel(id: string, name: string, context: number, reasoning = false): Model {
+  return {
+    id: ModelV2.ID.make(id),
+    providerID: ProviderV2.ID.make("geekhub"),
+    name,
+    family: "geekhub-free",
+    api: {
+      id,
+      url: GEEKHUB_BASE_URL,
+      npm: "@ai-sdk/openai-compatible",
+    },
+    status: "active",
+    headers: {},
+    options: {},
+    cost: {
+      input: 0,
+      output: 0,
+      cache: { read: 0, write: 0 },
+    },
+    limit: {
+      context,
+      output: context > 262144 ? 16000 : 8000,
+    },
+    capabilities: {
+      temperature: true,
+      reasoning,
+      attachment: false,
+      toolcall: true,
+      input: { text: true, audio: false, image: false, video: false, pdf: false },
+      output: { text: true, audio: false, image: false, video: false, pdf: false },
+      interleaved: false,
+    },
+    release_date: "2026-09-01",
+    variants: {},
+  }
+}
+
+function geekhubProvider(): Info {
+  const defs: Array<[string, string, number, boolean?]> = [
+    ["cohere/north-mini-code:free", "North Mini Code Free", 256000],
+    ["dots/dots-3-note-preview:free", "Dots 3 Note Preview Free", 512000],
+    ["google/gemma-4-26b-a4b-it:free", "Gemma 4 26B A4B IT Free", 262144],
+    ["google/gemma-4-31b-it:free", "Gemma 4 31B IT Free", 262144],
+    ["inclusionai/ling-3.0-flash-fin:free", "Ling 3.0 Flash Fin Free", 262144],
+    ["inclusionai/ling-3.0-flash-sante:free", "Ling 3.0 Flash Sante Free", 262144],
+    ["inclusionai/ling-3.0-flash-vl:free", "Ling 3.0 Flash VL Free", 262144],
+    ["liquid/lfm-2.5-2.6b:free", "LFM 2.5 2.6B Free", 65536],
+    ["nex-agi/nex-n2.5-mini:free", "Nex N2.5 Mini Free", 262144],
+    ["nex-agi/nex-n2.5-pro:free", "Nex N2.5 Pro Free", 262144],
+    ["nvidia/nemotron-3-nano-omni-30b-a3b-reasoning:free", "Nemotron 3 Nano Omni Reasoning Free", 256000, true],
+    ["nvidia/nemotron-3-super-120b-a12b:free", "Nemotron 3 Super 120B Free", 262144],
+    ["nvidia/nemotron-3-ultra-550b-a55b:free", "Nemotron 3 Ultra 550B Free", 1000000],
+    ["nvidia/nemotron-3.5-content-safety:free", "Nemotron 3.5 Content Safety Free", 128000],
+    ["nvidia/nemotron-3.5-lightning:free", "Nemotron 3.5 Lightning Free", 1000000],
+    ["poolside/laguna-s-2.1:free", "Laguna S 2.1 Free", 262144],
+    ["poolside/laguna-xs-2.1:free", "Laguna XS 2.1 Free", 262144],
+    ["thinkingmachines/inkling-small:free", "Inkling Small Free", 1048576],
+    ["thinkingmachines/inkling:free", "Inkling Free", 1048576],
+    ["qwen/qwen3.8-27b:free", "Qwen 3.8 27B Free", 262144],
+    ["dots-studio/dots-3-note-preview:free", "Dots 3 Note Preview Free", 512000],
+    ["z-ai/glm-5.2:free", "GLM 5.2 Free", 32768],
+  ]
+  const models: Record<string, Model> = {}
+  for (const [id, name, context, reasoning] of defs) {
+    models[id] = geekhubModel(id, name, context, reasoning ?? false)
+  }
+  for (const model of Object.values(models)) {
+    const variants = ProviderTransform.variants(model)
+    model.variants = mapValues(variants, (v) => v)
+  }
+  return {
+    id: ProviderV2.ID.make("geekhub"),
+    source: "custom",
+    name: "GeekHub",
+    env: ["GEEKHUB_API_KEY"],
+    options: {},
+    models,
+  }
+}
+
+function mervexModel(id: string, name: string, opts: {
+  context: number
+  output: number
+  reasoning?: boolean
+  attachment?: boolean
+  costInput?: number
+  costOutput?: number
+  vip?: boolean
+}): Model {
+  return {
+    id: ModelV2.ID.make(id),
+    providerID: ProviderV2.ID.make("mervex"),
+    name,
+    family: opts.vip ? "mervex-vip" : "mervex",
+    api: {
+      id,
+      url: MERVEX_BASE_URL,
+      npm: "@ai-sdk/openai-compatible",
+    },
+    status: "active",
+    headers: {},
+    options: {},
+    cost: {
+      input: opts.costInput ?? 0,
+      output: opts.costOutput ?? 0,
+      cache: { read: 0, write: 0 },
+    },
+    limit: {
+      context: opts.context,
+      output: opts.output,
+    },
+    capabilities: {
+      temperature: true,
+      reasoning: opts.reasoning ?? false,
+      attachment: opts.attachment ?? false,
+      toolcall: true,
+      input: { text: true, audio: false, image: opts.attachment ?? false, video: false, pdf: opts.attachment ?? false },
+      output: { text: true, audio: false, image: false, video: false, pdf: false },
+      interleaved: false,
+    },
+    release_date: "2026-01-01",
+    variants: {},
+  }
+}
+
+type MervexRemoteModel = {
+  id: string
+  name: string
+  tier?: string
+  vipOnly: boolean
+  vision: boolean
+  contextLimit?: number
+  outputLimit?: number
+}
+
+const fetchMervexRemoteModels = Effect.fn("Provider.fetchMervexRemoteModels")(function* (baseURL: string) {
+  const response = yield* Effect.tryPromise({
+    try: () =>
+      fetch(`${baseURL.replace(/\/$/, "")}/models`, {
+        signal: AbortSignal.timeout(10000),
+        headers: { Accept: "application/json" },
+      }),
+    catch: () => new Error("mervex models fetch failed"),
+  }).pipe(Effect.catch(() => Effect.succeed(undefined)))
+  if (!response || !response.ok) return undefined
+  const json = yield* Effect.tryPromise({
+    try: () => response.json() as Promise<unknown>,
+    catch: () => new Error("mervex models parse failed"),
+  }).pipe(Effect.catch(() => Effect.succeed(undefined)))
+  if (!isRecord(json) || !Array.isArray(json.data)) return undefined
+  const models: MervexRemoteModel[] = []
+  for (const item of json.data) {
+    if (!isRecord(item) || typeof item.id !== "string") continue
+    const num = (v: unknown) => (typeof v === "number" && Number.isFinite(v) && v > 0 ? v : undefined)
+    models.push({
+      id: item.id,
+      name: typeof item.name === "string" && item.name ? item.name : item.id,
+      tier: typeof item.tier === "string" ? item.tier : undefined,
+      vipOnly: item.is_vip_only === true,
+      vision: item.supports_vision === true,
+      contextLimit: num(item.context_limit),
+      outputLimit: num(item.output_limit),
+    })
+  }
+  return models.length ? models : undefined
+})
+
+function mervexRemoteDefaults(tier: string | undefined): { context: number; output: number; reasoning: boolean } {
+  if (tier === "pro") return { context: 64000, output: 16000, reasoning: true }
+  if (tier === "max") return { context: 64000, output: 16000, reasoning: true }
+  if (tier === "medium" || tier === "chat") return { context: 32000, output: 8000, reasoning: true }
+  return { context: 16000, output: 4000, reasoning: false }
+}
+
+// Display policy: model names shown without parenthetical descriptors.
+function cleanModelName(name: string): string {
+  return name.replace(/\s*\([^()]*\)\s*/g, " ").replace(/\s+/g, " ").trim()
+}
+
+function mergeMervexRemoteModels(info: Info, remote: MervexRemoteModel[]) {
+  const remoteIDs = new Set(remote.map((item) => item.id))
+  for (const id of Object.keys(info.models)) {
+    // Server no longer lists this model: hide it via deprecated status.
+    // The existing loader drops deprecated models from listings.
+    if (!remoteIDs.has(id)) info.models[id].status = "deprecated"
+  }
+  for (const item of remote) {
+    const existing = info.models[item.id]
+    if (existing) {
+      existing.name = cleanModelName(item.name)
+      existing.family = item.vipOnly ? "mervex-vip" : "mervex"
+      existing.capabilities.attachment = existing.capabilities.attachment || item.vision
+      existing.capabilities.input.image = existing.capabilities.input.image || item.vision
+      existing.capabilities.input.pdf = existing.capabilities.input.pdf || item.vision
+      // Server is the source of truth for enforced limits; sync them so
+      // requests stay within what the backend accepts.
+      if (item.contextLimit) existing.limit.context = item.contextLimit
+      if (item.outputLimit) existing.limit.output = item.outputLimit
+      continue
+    }
+    const defaults = mervexRemoteDefaults(item.tier)
+    info.models[item.id] = mervexModel(item.id, cleanModelName(item.name), {
+      context: item.contextLimit ?? defaults.context,
+      output: item.outputLimit ?? defaults.output,
+      reasoning: defaults.reasoning,
+      attachment: item.vision,
+      vip: item.vipOnly,
+    })
+  }
+}
+
+function mervexProvider(): Info {
+  const models: Record<string, Model> = {
+    "Msrfteam/Mervex-ai-v0.1-lite": mervexModel("Msrfteam/Mervex-ai-v0.1-lite", "Mervex V0.1 Lite", {
+      context: 8000,
+      output: 2000,
+    }),
+    "Msrfteam/Mervex-ai-v0.1": mervexModel("Msrfteam/Mervex-ai-v0.1", "Mervex V0.1", {
+      context: 16000,
+      output: 4000,
+      reasoning: true,
+      attachment: true,
+    }),
+    "Msrfteam/Mervex-ai-v0.1-pro": mervexModel("Msrfteam/Mervex-ai-v0.1-pro", "Mervex V0.1 Pro", {
+      context: 16000,
+      output: 4000,
+      reasoning: true,
+    }),
+    "Msrfteam/Mervex-ai-v0.2": mervexModel("Msrfteam/Mervex-ai-v0.2", "Mervex V0.2", {
+      context: 32000,
+      output: 8000,
+      reasoning: true,
+      attachment: true,
+    }),
+    "Msrfteam/Mervex-ai-v0.2-pro": mervexModel("Msrfteam/Mervex-ai-v0.2-pro", "Mervex V0.2 Pro", {
+      context: 64000,
+      output: 16000,
+      reasoning: true,
+      attachment: true,
+      vip: true,
+    }),
+    "Msrfteam/Mervex-ai-v0.1-mg": mervexModel("Msrfteam/Mervex-ai-v0.1-mg", "Mervex V0.1 MG", {
+      context: 32000,
+      output: 8000,
+      vip: true,
+    }),
+    "Msrfteam/Mervex-ai-v0.1-sch": mervexModel("Msrfteam/Mervex-ai-v0.1-sch", "Mervex V0.1 SCH", {
+      context: 32000,
+      output: 8000,
+      vip: true,
+    }),
+    "Msrfteam/Mervex-ai-v0.1-vd": mervexModel("Msrfteam/Mervex-ai-v0.1-vd", "Mervex V0.1 VD", {
+      context: 32000,
+      output: 8000,
+      vip: true,
+    }),
+    "Msrfteam/Mervex-ai-v0.1-aud": mervexModel("Msrfteam/Mervex-ai-v0.1-aud", "Mervex V0.1 AUD", {
+      context: 32000,
+      output: 8000,
+      vip: true,
+    }),
+  }
+  for (const [id, model] of Object.entries(models)) {
+    const variants = ProviderTransform.variants(model)
+    model.variants = mapValues(variants, (v) => v)
+  }
+  return {
+    id: ProviderV2.ID.make("mervex"),
+    source: "custom",
+    name: "Msr F Team",
+    env: ["MERVEX_API_KEY"],
+    options: {},
+    models,
+    // Cautious free-tier defaults: the backend itself pays per token
+    // (OpenRouter etc.), so guard both request count AND token volume.
+    // VIP users can raise these in mervex.jsonc `provider.mervex.rateLimit`.
+    rateLimit: {
+      rpm: 10,
+      rpd: 200,
+      tpd: 1_000_000,
+      maxInputTokens: 32_768,
+    },
+  }
+}
+
+function modeOptions(model: Model, body: Record<string, unknown> | undefined) {
+  if (!body) return model.options
+  const options = Object.fromEntries(
+    Object.entries(body).map(([key, value]) => [key.replace(/_([a-z])/g, (_, char) => char.toUpperCase()), value]),
+  )
+  const reasoning = body.reasoning
+  if (model.api.npm !== "@ai-sdk/openai" || !isRecord(reasoning) || typeof reasoning.mode !== "string") return options
+  const { reasoning: _, ...rest } = options
+  return { ...rest, reasoningMode: reasoning.mode }
+}
+
+function modelSuggestions(provider: Info | undefined, modelID: ModelV2.ID, enableExperimentalModels: boolean) {
+  const available = provider
+    ? Object.keys(provider.models).filter((id) => {
+        const model = provider.models[id]
+        if (model.status === "deprecated") return false
+        if (model.status === "alpha" && !enableExperimentalModels) return false
+        return true
+      })
+    : []
+  const fuzzy = fuzzysort.go(modelID, available, { limit: 3, threshold: -10000 }).map((m) => m.target)
+  if (fuzzy.length) return fuzzy
+  const query = modelID
+    .toLowerCase()
+    .split(/[^a-z0-9]+/)
+    .filter((part) => part.length > 1)
+  return sortBy(
+    available
+      .map((id) => ({
+        id,
+        score: query.filter((part) => id.toLowerCase().includes(part)).length,
+      }))
+      .filter((item) => item.score > 0),
+    [(item) => item.score, "desc"],
+    [(item) => item.id, "asc"],
+  )
+    .slice(0, 3)
+    .map((item) => item.id)
+}
+
+const layer = Layer.effect(
+  Service,
+  Effect.gen(function* () {
+    const fs = yield* FSUtil.Service
+    const config = yield* Config.Service
+    const auth = yield* Auth.Service
+    const env = yield* Env.Service
+    const plugin = yield* Plugin.Service
+    const modelsDevSvc = yield* ModelsDev.Service
+    const runtimeFlags = yield* RuntimeFlags.Service
+
+    const state = yield* InstanceState.make<State>(() =>
+      Effect.gen(function* () {
+        const bridge = yield* EffectBridge.make()
+        const cfg = yield* config.get()
+        const modelsDev = yield* modelsDevSvc.get()
+        const catalog = mapValues(modelsDev, fromModelsDevProvider)
+        const mervexInfo = mervexProvider()
+        const mervexCfg = (cfg.provider as Record<string, { options?: Record<string, any> }> | undefined)?.["mervex"]
+        const mervexBaseURL =
+          typeof mervexCfg?.options?.baseURL === "string" && mervexCfg.options.baseURL
+            ? mervexCfg.options.baseURL
+            : MERVEX_BASE_URL
+        const remoteMervexModels = yield* fetchMervexRemoteModels(mervexBaseURL).pipe(
+          Effect.catch(() => Effect.succeed(undefined)),
+        )
+        if (remoteMervexModels) mergeMervexRemoteModels(mervexInfo, remoteMervexModels)
+        catalog[mervexInfo.id] = mervexInfo
+        const geekhubInfo = geekhubProvider()
+        catalog[geekhubInfo.id] = geekhubInfo
+        const database = mapValues(catalog, toPublicInfo)
+
+        const providers: Record<ProviderV2.ID, Info> = {} as Record<ProviderV2.ID, Info>
+        const languages = new Map<string, LanguageModelV3>()
+        const modelLoaders: {
+          [providerID: string]: CustomModelLoader
+        } = {}
+        const varsLoaders: {
+          [providerID: string]: CustomVarsLoader
+        } = {}
+        const sdk = new Map<string, BundledSDK>()
+        const discoveryLoaders: {
+          [providerID: string]: CustomDiscoverModels
+        } = {}
+        const dep = {
+          auth: (id: string) => auth.get(id).pipe(Effect.orDie),
+          config: () => config.get(),
+          env: () => env.all(),
+          get: (key: string) => env.get(key),
+        }
+
+        function mergeProvider(providerID: ProviderV2.ID, provider: Partial<Info>) {
+          const existing = providers[providerID]
+          if (existing) {
+            // @ts-expect-error
+            providers[providerID] = mergeDeep(existing, provider)
+            return
+          }
+          const match = database[providerID]
+          if (!match) return
+          // @ts-expect-error
+          providers[providerID] = mergeDeep(match, provider)
+        }
+
+        // load plugins first so config() hook runs before reading cfg.provider
+        const plugins = yield* plugin.list()
+
+        // now read config providers - includes any modifications from plugin config() hook
+        const configProviders = Object.entries(cfg.provider ?? {})
+        const disabled = new Set(cfg.disabled_providers ?? [])
+        const enabled = cfg.enabled_providers ? new Set(cfg.enabled_providers) : null
+
+        function isProviderAllowed(providerID: ProviderV2.ID): boolean {
+          if (enabled && !enabled.has(providerID)) return false
+          if (disabled.has(providerID)) return false
+          return true
+        }
+
+        for (const hook of plugins) {
+          const p = hook.provider
+          const models = p?.models
+          if (!p || !models) continue
+
+          const providerID = ProviderV2.ID.make(p.id)
+          if (disabled.has(providerID)) continue
+
+          const provider = database[providerID]
+          if (!provider) continue
+          const pluginAuth = yield* auth.get(providerID).pipe(Effect.orDie)
+
+          provider.models = yield* Effect.promise(async () => {
+            const next = await models(toPublicInfo(provider), { auth: pluginAuth })
+            return Object.fromEntries(
+              Object.entries(next).map(([id, model]) => [
+                id,
+                {
+                  ...model,
+                  id: ModelV2.ID.make(id),
+                  providerID,
+                },
+              ]),
+            )
+          })
+        }
+
+        // extend database from config
+        for (const [providerID, provider] of configProviders) {
+          const existing = database[providerID]
+          const parsed: Info = {
+            id: ProviderV2.ID.make(providerID),
+            name: provider.name ?? existing?.name ?? providerID,
+            env: provider.env ?? existing?.env ?? [],
+            options: mergeDeep(existing?.options ?? {}, provider.options ?? {}),
+            source: "config",
+            models: existing?.models ?? {},
+          }
+
+          for (const [modelID, model] of Object.entries(provider.models ?? {})) {
+            const existingModel = parsed.models[model.id ?? modelID]
+            const apiID = model.id ?? existingModel?.api.id ?? modelID
+            const apiNpm =
+              model.provider?.npm ??
+              provider.npm ??
+              existingModel?.api.npm ??
+              // Config-defined gateway models bypass fromModelsDevModel, so resolve the
+              // native passthrough npm here before falling back to the catalog default.
+              cloudflareGatewayNpm(providerID, apiID) ??
+              modelsDev[providerID]?.npm ??
+              "@ai-sdk/openai-compatible"
+            const name = iife(() => {
+              if (model.name) return model.name
+              if (model.id && model.id !== modelID) return modelID
+              return existingModel?.name ?? modelID
+            })
+            const parsedModel: Model = {
+              id: ModelV2.ID.make(modelID),
+              api: {
+                id: apiID,
+                npm: apiNpm,
+                url: model.provider?.api ?? provider?.api ?? existingModel?.api.url ?? modelsDev[providerID]?.api ?? "",
+              },
+              status: model.status ?? existingModel?.status ?? "active",
+              name,
+              providerID: ProviderV2.ID.make(providerID),
+              capabilities: {
+                temperature: model.temperature ?? existingModel?.capabilities.temperature ?? false,
+                reasoning: model.reasoning ?? existingModel?.capabilities.reasoning ?? false,
+                attachment: model.attachment ?? existingModel?.capabilities.attachment ?? false,
+                toolcall: model.tool_call ?? existingModel?.capabilities.toolcall ?? true,
+                input: {
+                  text: model.modalities?.input?.includes("text") ?? existingModel?.capabilities.input.text ?? true,
+                  audio: model.modalities?.input?.includes("audio") ?? existingModel?.capabilities.input.audio ?? false,
+                  image: model.modalities?.input?.includes("image") ?? existingModel?.capabilities.input.image ?? false,
+                  video: model.modalities?.input?.includes("video") ?? existingModel?.capabilities.input.video ?? false,
+                  pdf: model.modalities?.input?.includes("pdf") ?? existingModel?.capabilities.input.pdf ?? false,
+                },
+                output: {
+                  text: model.modalities?.output?.includes("text") ?? existingModel?.capabilities.output.text ?? true,
+                  audio:
+                    model.modalities?.output?.includes("audio") ?? existingModel?.capabilities.output.audio ?? false,
+                  image:
+                    model.modalities?.output?.includes("image") ?? existingModel?.capabilities.output.image ?? false,
+                  video:
+                    model.modalities?.output?.includes("video") ?? existingModel?.capabilities.output.video ?? false,
+                  pdf: model.modalities?.output?.includes("pdf") ?? existingModel?.capabilities.output.pdf ?? false,
+                },
+                interleaved:
+                  (typeof model.interleaved === "string" ? { field: model.interleaved } : model.interleaved) ??
+                  existingModel?.capabilities.interleaved ??
+                  (!existingModel && apiNpm === "@ai-sdk/openai-compatible" && apiID.includes("deepseek")
+                    ? { field: "reasoning_content" }
+                    : false),
+              },
+              cost: {
+                input: model?.cost?.input ?? existingModel?.cost?.input ?? 0,
+                output: model?.cost?.output ?? existingModel?.cost?.output ?? 0,
+                cache: {
+                  read: model?.cost?.cache_read ?? existingModel?.cost?.cache.read ?? 0,
+                  write: model?.cost?.cache_write ?? existingModel?.cost?.cache.write ?? 0,
+                },
+              },
+              options: mergeDeep(existingModel?.options ?? {}, model.options ?? {}),
+              limit: {
+                context: model.limit?.context ?? existingModel?.limit?.context ?? 0,
+                input: model.limit?.input ?? existingModel?.limit?.input,
+                output: model.limit?.output ?? existingModel?.limit?.output ?? 0,
+              },
+              headers: mergeDeep(existingModel?.headers ?? {}, model.headers ?? {}),
+              family: model.family ?? existingModel?.family ?? "",
+              release_date: model.release_date ?? existingModel?.release_date ?? "",
+              variants: {},
+            }
+            const variants =
+              existingModel?.api.npm === parsedModel.api.npm
+                ? (existingModel.variants ?? ProviderTransform.variants(parsedModel))
+                : ProviderTransform.variants(parsedModel)
+            const merged = mergeDeep(variants, model.variants ?? {})
+            parsedModel.variants = mapValues(
+              pickBy(merged, (v) => !v.disabled),
+              (v) => omit(v, ["disabled"]),
+            )
+            parsed.models[modelID] = parsedModel
+          }
+          database[providerID] = parsed
+        }
+
+        // load env
+        const envs = yield* env.all()
+        for (const [id, provider] of Object.entries(database)) {
+          const providerID = ProviderV2.ID.make(id)
+          if (disabled.has(providerID)) continue
+          const apiKey = provider.env.map((item) => envs[item]).find(Boolean)
+          if (!apiKey) continue
+          mergeProvider(providerID, {
+            source: "env",
+            key: provider.env.length === 1 ? apiKey : undefined,
+          })
+        }
+
+        // load apikeys
+        const auths = yield* auth.all().pipe(Effect.orDie)
+        for (const [id, provider] of Object.entries(auths)) {
+          const providerID = ProviderV2.ID.make(id)
+          if (disabled.has(providerID)) continue
+          if (provider.type === "api") {
+            mergeProvider(providerID, {
+              source: "api",
+              key: provider.key,
+            })
+          }
+        }
+
+        // plugin auth loader - database now has entries for config providers
+        for (const plugin of plugins) {
+          if (!plugin.auth) continue
+          const providerID = ProviderV2.ID.make(plugin.auth.provider)
+          if (disabled.has(providerID)) continue
+
+          const stored = yield* auth.get(providerID).pipe(Effect.orDie)
+          if (!stored) continue
+          if (!plugin.auth.loader) continue
+
+          const options = yield* Effect.promise(() =>
+            plugin.auth!.loader!(
+              () => bridge.promise(auth.get(providerID).pipe(Effect.orDie)) as any,
+              toPublicInfo(database[plugin.auth!.provider]),
+            ),
+          )
+          const opts = options ?? {}
+          const patch: Partial<Info> = providers[providerID] ? { options: opts } : { source: "custom", options: opts }
+          mergeProvider(providerID, patch)
+        }
+
+        for (const [id, fn] of Object.entries(custom(dep))) {
+          const providerID = ProviderV2.ID.make(id)
+          if (disabled.has(providerID)) continue
+          const data = database[providerID]
+          if (!data) {
+            continue
+          }
+          const result = yield* fn(data)
+          if (result && (result.autoload || providers[providerID])) {
+            if (result.getModel) modelLoaders[providerID] = result.getModel
+            if (result.vars) varsLoaders[providerID] = result.vars
+            if (result.discoverModels) discoveryLoaders[providerID] = result.discoverModels
+            const opts = result.options ?? {}
+            const patch: Partial<Info> = providers[providerID] ? { options: opts } : { source: "custom", options: opts }
+            mergeProvider(providerID, patch)
+          }
+        }
+
+        // load config - re-apply with updated data
+        for (const [id, provider] of configProviders) {
+          const providerID = ProviderV2.ID.make(id)
+          const partial: Partial<Info> = { source: "config" }
+          if (provider.env) partial.env = provider.env
+          if (provider.name) partial.name = provider.name
+          if (provider.options) partial.options = provider.options
+          if (provider.rateLimit) partial.rateLimit = provider.rateLimit
+          mergeProvider(providerID, partial)
+        }
+
+        const gitlab = ProviderV2.ID.make("gitlab")
+        if (discoveryLoaders[gitlab] && providers[gitlab] && isProviderAllowed(gitlab)) {
+          yield* Effect.promise(async () => {
+            try {
+              const discovered = await discoveryLoaders[gitlab]()
+              for (const [modelID, model] of Object.entries(discovered)) {
+                if (!providers[gitlab].models[modelID]) {
+                  providers[gitlab].models[modelID] = model
+                }
+              }
+            } catch (e) {}
+          })
+        }
+
+        for (const [id, provider] of Object.entries(providers)) {
+          const providerID = ProviderV2.ID.make(id)
+          if (!isProviderAllowed(providerID)) {
+            delete providers[providerID]
+            continue
+          }
+
+          const configProvider = cfg.provider?.[providerID]
+
+          for (const [modelID, model] of Object.entries(provider.models)) {
+            model.api.id = model.api.id ?? model.id ?? modelID
+
+            if (
+              // These chat aliases are invalid for the special handling in the
+              // built-in providers below, but custom providers may support them.
+              (modelID === "gpt-5-chat-latest" &&
+                (providerID === ProviderV2.ID.openai ||
+                  providerID === ProviderV2.ID.githubCopilot ||
+                  providerID === ProviderV2.ID.openrouter)) ||
+              (providerID === ProviderV2.ID.openrouter && modelID === "openai/gpt-5-chat")
+            )
+              delete provider.models[modelID]
+            if (model.status === "alpha" && !runtimeFlags.enableExperimentalModels) delete provider.models[modelID]
+            if (model.status === "deprecated") delete provider.models[modelID]
+            if (
+              (configProvider?.blacklist && configProvider.blacklist.includes(modelID)) ||
+              (configProvider?.whitelist && !configProvider.whitelist.includes(modelID))
+            )
+              delete provider.models[modelID]
+
+            if (model.variants === undefined) {
+              model.variants = mapValues(ProviderTransform.variants(model), (v) => v)
+            }
+
+            const configVariants = configProvider?.models?.[modelID]?.variants
+            if (configVariants && model.variants) {
+              const merged = mergeDeep(model.variants, configVariants)
+              model.variants = mapValues(
+                pickBy(merged, (v) => !v.disabled),
+                (v) => omit(v, ["disabled"]),
+              )
+            }
+          }
+
+          if (Object.keys(provider.models).length === 0) {
+            delete providers[providerID]
+            continue
+          }
+        }
+
+        return {
+          models: languages,
+          providers,
+          catalog,
+          sdk,
+          modelLoaders,
+          varsLoaders,
+        }
+      }),
+    )
+
+    const list = Effect.fn("Provider.list")(() => InstanceState.use(state, (s) => s.providers))
+
+    const catalog = Effect.fn("Provider.catalog")(() => InstanceState.use(state, (s) => s.catalog))
+
+    async function resolveSDK(
+      model: Model,
+      s: State,
+      envs: Record<string, string | undefined>,
+      globalProxy?: ConfigV1.Info["proxy"],
+      providerProxy?: ConfigV1.Info["proxy"],
+    ) {
+      try {
+        const provider = s.providers[model.providerID]
+        const options = { ...provider.options }
+
+        if (
+          model.providerID === "google-vertex" &&
+          model.api.npm === "@ai-sdk/google-vertex/anthropic" &&
+          !options.baseURL
+        ) {
+          const baseURL = googleVertexAnthropicBaseURL(
+            typeof options.project === "string" ? options.project : undefined,
+            typeof options.location === "string" ? options.location : undefined,
+          )
+          if (baseURL) options.baseURL = baseURL
+        }
+
+        if (model.providerID === "google-vertex" && !model.api.npm.includes("@ai-sdk/openai-compatible")) {
+          delete options.fetch
+        }
+
+        if (model.api.npm.includes("@ai-sdk/openai-compatible") && options["includeUsage"] !== false) {
+          options["includeUsage"] = true
+        }
+
+        const baseURL = iife(() => {
+          let url =
+            typeof options["baseURL"] === "string" && options["baseURL"] !== "" ? options["baseURL"] : model.api.url
+          if (!url) return
+
+          const loader = s.varsLoaders[model.providerID]
+          if (loader) {
+            const vars = loader(options)
+            for (const [key, value] of Object.entries(vars)) {
+              const field = "${" + key + "}"
+              url = url.replaceAll(field, value)
+            }
+          }
+
+          url = url.replace(/\$\{([^}]+)\}/g, (item, key) => {
+            const val = envs[String(key)]
+            return val ?? item
+          })
+          return url
+        })
+
+        if (baseURL !== undefined) options["baseURL"] = baseURL
+        if (options["apiKey"] === undefined && provider.key) options["apiKey"] = provider.key
+        if (model.headers)
+          options["headers"] = {
+            ...options["headers"],
+            ...model.headers,
+          }
+
+        const key = Hash.fast(
+          JSON.stringify({
+            providerID: model.providerID,
+            npm: model.api.npm,
+            options,
+          }),
+        )
+        const existing = s.sdk.get(key)
+        if (existing) return existing
+
+        const customFetch = options["fetch"]
+        const chunkTimeout = options["chunkTimeout"]
+        const headerTimeout = options["headerTimeout"]
+        delete options["chunkTimeout"]
+        delete options["headerTimeout"]
+
+        const proxyConfig = providerProxy ?? globalProxy
+        let proxyAgent: ProxyAgent | undefined
+        if (proxyConfig) {
+          const url = proxyConfig.https ?? proxyConfig.http
+          if (url) {
+            proxyAgent = new ProxyAgent({
+              uri: url,
+              requestTls: { rejectUnauthorized: true },
+            })
+          }
+        }
+
+        options["fetch"] = async (input: any, init?: BunFetchRequestInit) => {
+          const fetchFn = customFetch ?? fetch
+          const opts = init ?? {}
+          const chunkAbortCtl = typeof chunkTimeout === "number" && chunkTimeout > 0 ? new AbortController() : undefined
+          const headerTimeoutMs = headerTimeout === false ? undefined : headerTimeout
+          const headerTimeoutCtl = typeof headerTimeoutMs === "number" ? timeoutController(headerTimeoutMs) : undefined
+          const signals: AbortSignal[] = []
+
+          if (opts.signal) signals.push(opts.signal)
+          if (chunkAbortCtl) signals.push(chunkAbortCtl.signal)
+          if (headerTimeoutCtl) signals.push(headerTimeoutCtl.signal)
+          if (options["timeout"] !== undefined && options["timeout"] !== null && options["timeout"] !== false)
+            signals.push(AbortSignal.timeout(options["timeout"]))
+
+          const combined = signals.length === 0 ? null : signals.length === 1 ? signals[0] : AbortSignal.any(signals)
+          if (combined) opts.signal = combined
+
+          if (proxyAgent) {
+            // @ts-ignore undici dispatcher
+            opts.dispatcher = proxyAgent
+          }
+
+          const res = await fetchFn(input, {
+            ...opts,
+            // @ts-ignore see here: https://github.com/oven-sh/bun/issues/16682
+            timeout: false,
+          }).finally(() => headerTimeoutCtl?.clear())
+
+          if (!chunkAbortCtl) return res
+          return wrapSSE(res, chunkTimeout, chunkAbortCtl)
+        }
+
+        const bundledLoader = BUNDLED_PROVIDERS[model.api.npm]
+        if (bundledLoader) {
+          const factory = await bundledLoader()
+          const loaded = factory({
+            name: model.providerID,
+            ...options,
+          })
+          s.sdk.set(key, loaded)
+          return loaded as SDK
+        }
+
+        const installedPath = await (async () => {
+          if (model.api.npm.startsWith("file://")) {
+            return model.api.npm
+          }
+          const item = await Npm.add(model.api.npm)
+          if (!item.entrypoint) throw new Error(`Package ${model.api.npm} has no import entrypoint`)
+          return item.entrypoint
+        })()
+
+        // `installedPath` is a local entry path or an existing `file://` URL. Normalize
+        // only path inputs so Node on Windows accepts the dynamic import.
+        const importSpec = installedPath.startsWith("file://") ? installedPath : pathToFileURL(installedPath).href
+        const mod = await import(importSpec)
+
+        const fn = mod[Object.keys(mod).find((key) => key.startsWith("create"))!]
+        const loaded = fn({
+          name: model.providerID,
+          ...options,
+        })
+        s.sdk.set(key, loaded)
+        return loaded as SDK
+      } catch (e) {
+        throw new InitError({ providerID: model.providerID, cause: e })
+      }
+    }
+
+    const getProvider = Effect.fn("Provider.getProvider")((providerID: ProviderV2.ID) =>
+      InstanceState.use(state, (s) => s.providers[providerID]),
+    )
+
+    const getModel = Effect.fn("Provider.getModel")(function* (providerID: ProviderV2.ID, modelID: ModelV2.ID) {
+      const s = yield* InstanceState.get(state)
+      const provider = s.providers[providerID]
+      if (!provider) {
+        const catalogProvider = s.catalog[providerID]
+        const suggestions = catalogProvider
+          ? modelSuggestions(catalogProvider, modelID, runtimeFlags.enableExperimentalModels)
+          : fuzzysort
+              .go(providerID, Object.keys({ ...s.catalog, ...s.providers }), { limit: 3, threshold: -10000 })
+              .map((m) => m.target)
+        return yield* new ModelNotFoundError({ providerID, modelID, suggestions })
+      }
+
+      const info = provider.models[modelID]
+      if (!info) {
+        const current = modelSuggestions(provider, modelID, runtimeFlags.enableExperimentalModels)
+        const suggestions = current.length
+          ? current
+          : modelSuggestions(s.catalog[providerID], modelID, runtimeFlags.enableExperimentalModels)
+        return yield* new ModelNotFoundError({ providerID, modelID, suggestions })
+      }
+      return info
+    })
+
+    const getLanguage = Effect.fn("Provider.getLanguage")(function* (model: Model) {
+      const s = yield* InstanceState.get(state)
+      const envs = yield* env.all()
+      const cfg = yield* config.get()
+      const key = `${model.providerID}/${model.id}`
+      if (s.models.has(key)) return s.models.get(key)!
+
+      const provider = s.providers[model.providerID]
+      return yield* EffectPromise.refineRejection(
+        async () => {
+          const sdk = await resolveSDK(model, s, envs, cfg.proxy, cfg.provider?.[model.providerID]?.proxy)
+          const language = s.modelLoaders[model.providerID]
+            ? await s.modelLoaders[model.providerID](
+                sdk,
+                model.api.id,
+                {
+                  ...provider.options,
+                  ...model.options,
+                },
+                model,
+              )
+            : sdk.languageModel(model.api.id)
+          s.models.set(key, language)
+          return language
+        },
+        (cause) =>
+          cause instanceof NoSuchModelError
+            ? new ModelNotFoundError({ modelID: model.id, providerID: model.providerID, cause })
+            : undefined,
+      )
+    })
+
+    const closest = Effect.fn("Provider.closest")(function* (providerID: ProviderV2.ID, query: string[]) {
+      const s = yield* InstanceState.get(state)
+      const provider = s.providers[providerID]
+      if (!provider) return undefined
+      for (const item of query) {
+        for (const modelID of Object.keys(provider.models)) {
+          if (modelID.includes(item)) return { providerID, modelID }
+        }
+      }
+      return undefined
+    })
+
+    const getSmallModel = Effect.fn("Provider.getSmallModel")(function* (providerID: ProviderV2.ID) {
+      const cfg = yield* config.get()
+
+      if (cfg.small_model) {
+        const parsed = parseModel(cfg.small_model)
+        return yield* getModel(parsed.providerID, parsed.modelID).pipe(
+          Effect.catchTag("ProviderModelNotFoundError", () => Effect.succeed(undefined)),
+        )
+      }
+
+      const s = yield* InstanceState.get(state)
+      const provider = s.providers[providerID]
+      if (!provider) return undefined
+
+      const experimental = yield* plugin.trigger<"experimental.provider.small_model">(
+        "experimental.provider.small_model",
+        { provider: toPublicInfo(provider) },
+        { model: undefined },
+      )
+      if (experimental.model) {
+        return {
+          ...experimental.model,
+          id: ModelV2.ID.make(experimental.model.id),
+          providerID: ProviderV2.ID.make(experimental.model.providerID),
+        }
+      }
+
+      // TODO: Remove these provider-specific assumptions once model syncing reliably reports available deployments.
+      if (providerID === ProviderV2.ID.azure || providerID === ProviderV2.ID.make("azure-cognitive-services")) {
+        return undefined
+      }
+
+      const priority = providerID.startsWith("mervex")
+        ? ["gpt-nano", ...smallModelFamilyPriority]
+        : providerID.startsWith("mervex")
+          ? ["gpt-nano"]
+          : providerID.startsWith("deleted-provider")
+            ? ["gpt-mini", ...smallModelFamilyPriority]
+            : smallModelFamilyPriority
+      const models = sortBy(
+        Object.values(provider.models),
+        [(model) => model.release_date, "desc"],
+        [(model) => model.id, "desc"],
+      )
+      for (const family of priority) {
+        const candidates = models.filter((model) => model.family === family)
+        if (providerID === ProviderV2.ID.amazonBedrock) {
+          const crossRegionPrefixes = ["global.", "us.", "eu."]
+
+          const globalMatch = candidates.find((model) => model.id.startsWith("global."))
+          if (globalMatch) return globalMatch
+
+          const region = provider.options?.region
+          if (region) {
+            const regionPrefix = region.split("-")[0]
+            if (regionPrefix === "us" || regionPrefix === "eu") {
+              const regionalMatch = candidates.find((model) => model.id.startsWith(`${regionPrefix}.`))
+              if (regionalMatch) return regionalMatch
+            }
+          }
+
+          const unprefixed = candidates.find((model) => !crossRegionPrefixes.some((p) => model.id.startsWith(p)))
+          if (unprefixed) return unprefixed
+          continue
+        }
+        if (candidates[0]) return candidates[0]
+      }
+
+      return undefined
+    })
+
+    const defaultModel = Effect.fn("Provider.defaultModel")(function* () {
+      const cfg = yield* config.get()
+      if (cfg.model) return parseModel(cfg.model)
+
+      const s = yield* InstanceState.get(state)
+      const recent = yield* fs.readJson(path.join(Global.Path.state, "model.json")).pipe(
+        Effect.map((x): { providerID: ProviderV2.ID; modelID: ModelV2.ID }[] => {
+          if (!isRecord(x) || !Array.isArray(x.recent)) return []
+          return x.recent.flatMap((item) => {
+            if (!isRecord(item)) return []
+            if (typeof item.providerID !== "string") return []
+            if (typeof item.modelID !== "string") return []
+            return [{ providerID: ProviderV2.ID.make(item.providerID), modelID: ModelV2.ID.make(item.modelID) }]
+          })
+        }),
+        Effect.catch(() => Effect.succeed([] as { providerID: ProviderV2.ID; modelID: ModelV2.ID }[])),
+      )
+      for (const entry of recent) {
+        const provider = s.providers[entry.providerID]
+        if (!provider) continue
+        if (!provider.models[entry.modelID]) continue
+        return { providerID: entry.providerID, modelID: entry.modelID }
+      }
+
+      const configured = Object.keys(cfg.provider ?? {})
+      const allProviders = Object.values(s.providers).sort((a, b) => {
+        if (a.id === "mervex") return -1
+        if (b.id === "mervex") return 1
+        if (a.id === "mervex") return -1
+        if (b.id === "mervex") return 1
+        return 0
+      })
+      const provider = allProviders.find((p) => configured.length === 0 || configured.includes(p.id))
+      if (!provider) return yield* new NoProvidersError()
+      const [model] = sort(Object.values(provider.models))
+      if (!model) return yield* new NoModelsError({ providerID: provider.id })
+      return {
+        providerID: provider.id,
+        modelID: model.id,
+      }
+    })
+
+    return Service.of({ list, catalog, getProvider, getModel, getLanguage, closest, getSmallModel, defaultModel })
+  }),
+)
+
+const priority = ["mervex-ai", "gpt-5", "claude-sonnet-4", "big-pickle", "gemini-3-pro"]
+const smallModelFamilyPriority = ["gemini-flash", "gpt-nano", "claude-haiku"]
+export function sort<T extends { id: string }>(models: T[]) {
+  return sortBy(
+    models,
+    [(model) => priority.findIndex((filter) => model.id.includes(filter)), "desc"],
+    [(model) => (model.id.includes("latest") ? 0 : 1), "asc"],
+    [(model) => model.id, "desc"],
+  )
+}
+
+export function parseModel(model: string) {
+  const [providerID, ...rest] = model.split("/")
+  return {
+    providerID: ProviderV2.ID.make(providerID),
+    modelID: ModelV2.ID.make(rest.join("/")),
+  }
+}
+
+export const node = LayerNode.make({
+  service: Service,
+  layer: layer,
+  deps: [FSUtil.node, Config.node, Auth.node, Env.node, Plugin.node, ModelsDev.node, RuntimeFlags.node],
+})
+
+export * as Provider from "./provider"