@opengeni/config 0.10.1 → 0.10.6

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@opengeni/config",
-  "version": "0.10.1",
+  "version": "0.10.6",
   "description": "OpenGeni runtime configuration: settings resolution, deployment knobs, and config validation shared across the server packages.",
   "license": "Apache-2.0",
   "repository": {
@@ -33,8 +33,8 @@
     "prepublishOnly": "bash ../../scripts/prepublish-guard"
   },
   "dependencies": {
-    "@opengeni/codex": "^0.2.9",
-    "@opengeni/contracts": "^0.31.1",
+    "@opengeni/codex": "^0.2.10",
+    "@opengeni/contracts": "^0.35.0",
     "zod": "^4.2.1"
   },
   "engines": {

package/src/index.ts

@@ -108,7 +108,7 @@ export const DEFAULT_AGENT_INSTRUCTIONS = [
   "Follow the user's task and any enabled pack or skill instructions for the current role.",
   "Work inside the sandbox workspace and use filesystem and shell tools when useful.",
   "Repository resources are mounted under repos/<host>/<owner>/<repo> unless the session specifies another collision-free mount path.",
-  "File resources are mounted under files/<file-id>/ unless the session specifies another mount path.",
+  "File resources are mounted under .opengeni/files/<file-id>/ unless the session specifies another mount path.",
   "Attached files are mounted read-only; copy them before modifying.",
   "Bundled skills are under .agents/ and can include infrastructure, marketing, or other role-specific guidance.",
   "Use Checkov, Terraform, Azure CLI, git provider CLIs, and repository tools when relevant; gh, glab, and az repos are pre-authenticated when the host brokers matching git credentials.",
@@ -432,7 +432,7 @@ const SettingsSchema = z.object({
   // match the UI danger flip (UsageBar danger at pct >= 90). OPENGENI_CODEX_ROTATION_NEAR_EXHAUSTION_PCT.
   codexRotationNearExhaustionPct: z.coerce.number().int().min(1).max(100).default(90),
   openaiReasoningEffort: ReasoningEffort.default("low"),
-  openaiAllowedReasoningEfforts: z.string().default("low,medium,high,xhigh"),
+  openaiAllowedReasoningEfforts: z.string().default("low,medium,high,xhigh,max"),
   openaiResponsesTransport: z.enum(["http", "websocket"]).default("http"),
   // Provider-assigned item ids (rs_/msg_/fc_…) in Responses API input are
   // resolved against the provider's server-side response store. That store is
@@ -1176,6 +1176,8 @@ export const ModelCapabilitiesV1Schema = z
       codeExecution: CapabilityStateV1Schema,
     }),
     inputModalities: z.array(ModelModalityV1).min(1),
+    /** Exact MIME types accepted as typed `input_file`; `text/*` is allowed. */
+    inputFileMediaTypes: z.array(z.string()).default([]),
     outputModalities: z.array(ModelModalityV1).min(1),
     transports: z.object({
       sse: CapabilityStateV1Schema,
@@ -1415,7 +1417,7 @@ export interface ConfiguredModel {
   capabilities: ModelCapabilitiesV1;
   requestPolicy?: {
     gateway: {
-      only: [string];
+      only: [string, ...string[]];
       caching: "auto" | "none";
     };
   };
@@ -1430,27 +1432,74 @@ export interface ConfiguredModel {
 }
 
 export const VERCEL_AI_GATEWAY_BASE_URL = "https://ai-gateway.vercel.sh/v1" as const;
+export const VERCEL_AI_GATEWAY_AI_SDK_BASE_URL = "https://ai-gateway.vercel.sh/v4/ai" as const;
 export const OPENGENI_GATEWAY_PROVIDER_ID = "opengeni-gateway" as const;
 export const WORKSPACE_GATEWAY_PROVIDER_ID = "workspace-gateway" as const;
 export const WORKSPACE_GATEWAY_MODEL_ID_PREFIX = "workspace-gateway/" as const;
 export const VERCEL_AI_GATEWAY_CONNECTION_DOMAIN = "ai-gateway.vercel.sh" as const;
 export const VERCEL_AI_GATEWAY_CONNECTION_ROLE = "vercel_ai_gateway" as const;
 
+export const CODEX_REALTIME_MODEL_ID = "gpt-live-1-boulder-alpha" as const;
+export const OPENGENI_REALTIME_MODEL_ID_PREFIX = "opengeni-gateway/" as const;
+export const WORKSPACE_REALTIME_MODEL_ID_PREFIX = "workspace-gateway/" as const;
+
+/** Curated voice models exposed through AI Gateway's normalized realtime API. */
+export const AI_GATEWAY_REALTIME_MODELS = {
+  openaiRealtime21: {
+    upstreamModelId: "openai/gpt-realtime-2.1",
+    managedModelId: `${OPENGENI_REALTIME_MODEL_ID_PREFIX}openai/gpt-realtime-2.1`,
+    workspaceModelId: `${WORKSPACE_REALTIME_MODEL_ID_PREFIX}openai/gpt-realtime-2.1`,
+    label: "GPT Realtime 2.1",
+    description: "Best overall voice intelligence",
+  },
+  openaiRealtimeMini: {
+    upstreamModelId: "openai/gpt-realtime-mini",
+    managedModelId: `${OPENGENI_REALTIME_MODEL_ID_PREFIX}openai/gpt-realtime-mini`,
+    workspaceModelId: `${WORKSPACE_REALTIME_MODEL_ID_PREFIX}openai/gpt-realtime-mini`,
+    label: "GPT Realtime Mini",
+    description: "Faster, lighter live voice",
+  },
+  grokVoiceThinkFast20: {
+    upstreamModelId: "xai/grok-voice-think-fast-2.0",
+    managedModelId: `${OPENGENI_REALTIME_MODEL_ID_PREFIX}xai/grok-voice-think-fast-2.0`,
+    workspaceModelId: `${WORKSPACE_REALTIME_MODEL_ID_PREFIX}xai/grok-voice-think-fast-2.0`,
+    label: "Grok Voice Think Fast 2.0",
+    description: "Fast, natural xAI voice",
+  },
+} as const;
+
+export type AiGatewayRealtimeModel =
+  (typeof AI_GATEWAY_REALTIME_MODELS)[keyof typeof AI_GATEWAY_REALTIME_MODELS];
+
+export function resolveAiGatewayRealtimeModel(
+  modelId: string,
+): { source: "managed" | "workspace"; upstreamModelId: string } | null {
+  for (const model of Object.values(AI_GATEWAY_REALTIME_MODELS)) {
+    if (model.managedModelId === modelId) {
+      return { source: "managed", upstreamModelId: model.upstreamModelId };
+    }
+    if (model.workspaceModelId === modelId) {
+      return { source: "workspace", upstreamModelId: model.upstreamModelId };
+    }
+  }
+  return null;
+}
+
 export const OPENGENI_GATEWAY_MODELS = {
   deepseek: {
     productId: "deepseek-v4-flash-0731",
     workspaceProductId: `${WORKSPACE_GATEWAY_MODEL_ID_PREFIX}deepseek-v4-flash-0731`,
     upstreamModelId: "deepseek/deepseek-v4-flash-0731",
     label: "DeepSeek V4 Flash 0731",
-    provider: "deepinfra",
+    providers: ["baseten", "novita", "deepinfra"],
     implicitCaching: true,
   },
   kimi: {
-    productId: "kimi-k3-fast",
-    workspaceProductId: `${WORKSPACE_GATEWAY_MODEL_ID_PREFIX}kimi-k3-fast`,
-    upstreamModelId: "moonshotai/kimi-k3-fast",
-    label: "Kimi K3 Fast",
-    provider: "wafer",
+    productId: "kimi-k3",
+    workspaceProductId: `${WORKSPACE_GATEWAY_MODEL_ID_PREFIX}kimi-k3`,
+    upstreamModelId: "moonshotai/kimi-k3",
+    label: "Kimi K3",
+    providers: ["baseten", "fireworks"],
     implicitCaching: true,
   },
 } as const;
@@ -1529,24 +1578,24 @@ export const defaultModelPricing: Record<string, ModelPricingScheduleV1> = {
       },
     ],
   },
-  // Vercel AI Gateway endpoint prices, provider-pinned in the runtime.
-  // Snapshot: 2026-08-02. Both pinned routes returned discounted implicit
-  // cache reads in live Gateway responses. Wafer/Kimi reported $0.45/M even
-  // though the provider-discovery flag currently says otherwise; bill from
-  // the response-backed rate, not that inconsistent boolean.
+  // Conservative Vercel AI Gateway fallback prices. Normal managed Gateway
+  // billing uses the exact response Gateway `cost` / `inferenceCost` and applies
+  // the same margin. These token rates are used only if that
+  // metadata is absent. DeepSeek therefore carries the highest approved route
+  // (Novita); both approved Kimi routes have the same list price.
   [OPENGENI_GATEWAY_MODELS.deepseek.productId]: {
     default: {
-      inputMicrosPerMillionTokens: 90_000,
-      cachedInputMicrosPerMillionTokens: 18_000,
-      outputMicrosPerMillionTokens: 180_000,
+      inputMicrosPerMillionTokens: 140_000,
+      cachedInputMicrosPerMillionTokens: 28_000,
+      outputMicrosPerMillionTokens: 280_000,
       marginBps: 2_500,
     },
   },
   [OPENGENI_GATEWAY_MODELS.kimi.productId]: {
     default: {
-      inputMicrosPerMillionTokens: 4_500_000,
-      cachedInputMicrosPerMillionTokens: 450_000,
-      outputMicrosPerMillionTokens: 22_500_000,
+      inputMicrosPerMillionTokens: 3_000_000,
+      cachedInputMicrosPerMillionTokens: 300_000,
+      outputMicrosPerMillionTokens: 15_000_000,
       marginBps: 2_500,
     },
   },
@@ -2224,6 +2273,7 @@ function normalizeCapabilities(capabilities: ModelCapabilitiesV1): ModelCapabili
     inputModalities: [...parsed.inputModalities].sort(
       (left, right) => (MODALITY_ORDER.get(left) ?? 0) - (MODALITY_ORDER.get(right) ?? 0),
     ),
+    inputFileMediaTypes: [...new Set(parsed.inputFileMediaTypes)].sort(),
     outputModalities: [...parsed.outputModalities].sort(
       (left, right) => (MODALITY_ORDER.get(left) ?? 0) - (MODALITY_ORDER.get(right) ?? 0),
     ),
@@ -2236,7 +2286,7 @@ function normalizeCapabilities(capabilities: ModelCapabilitiesV1): ModelCapabili
 
 function legacyModelCapabilities(
   settings: Settings,
-  input: { reasoningEffort: boolean; hostedWebSearch: boolean },
+  input: { reasoningEffort: boolean; hostedWebSearch: boolean; vision?: boolean },
 ): ModelCapabilitiesV1 {
   const reasoningEfforts = input.reasoningEffort ? configuredAllowedReasoningEfforts(settings) : [];
   return normalizeCapabilities({
@@ -2257,7 +2307,14 @@ function legacyModelCapabilities(
       xSearch: { upstream: "unknown", runnable: false },
       codeExecution: { upstream: "unknown", runnable: false },
     },
-    inputModalities: ["text"],
+    inputModalities: input.vision ? ["text", "image"] : ["text"],
+    inputFileMediaTypes: [
+      "application/json",
+      "application/pdf",
+      "application/x-yaml",
+      "application/yaml",
+      "text/*",
+    ],
     outputModalities: ["text"],
     transports: {
       sse: { upstream: "unknown", runnable: true },
@@ -2279,7 +2336,7 @@ export function gatewayRequestPolicyForUpstreamModel(
   }
   return {
     gateway: {
-      only: [model.provider],
+      only: [...model.providers] as [string, ...string[]],
       caching: model.implicitCaching ? "auto" : "none",
     },
   };
@@ -2287,7 +2344,7 @@ export function gatewayRequestPolicyForUpstreamModel(
 
 function gatewayModelCapabilities(
   settings: Settings,
-  input: { implicitCaching: boolean; vision: boolean },
+  input: { implicitCaching: boolean; vision: boolean; inputFileMediaTypes?: string[] },
 ): ModelCapabilitiesV1 {
   const legacy = legacyModelCapabilities(settings, {
     reasoningEffort: true,
@@ -2297,6 +2354,7 @@ function gatewayModelCapabilities(
     ...legacy,
     functionCalling: { upstream: "supported", runnable: true },
     inputModalities: input.vision ? ["text", "image"] : ["text"],
+    inputFileMediaTypes: input.inputFileMediaTypes ?? [],
     transports: {
       ...legacy.transports,
       sse: { upstream: "supported", runnable: true },
@@ -2304,8 +2362,8 @@ function gatewayModelCapabilities(
     promptCaching: input.implicitCaching
       ? { upstream: "supported", runnable: true, mode: "implicit" }
       : { upstream: "unsupported", runnable: false, mode: "none" },
-    // "Fast" is part of Kimi's product name, not OpenGeni's separately billed
-    // latency mode. Both Gateway products expose only standard here.
+    // Both Gateway products expose one reviewed route policy and no separately
+    // billed latency mode.
     latencyModes: [{ id: "standard", upstream: "supported", runnable: true }],
   });
 }
@@ -2317,19 +2375,23 @@ function gatewayRegistryProvider(
     | { kind: "vercel-gateway-workspace"; apiKey?: string },
 ): RegistryProvider {
   const workspace = input.kind === "vercel-gateway-workspace";
-  const models = [OPENGENI_GATEWAY_MODELS.deepseek, OPENGENI_GATEWAY_MODELS.kimi].map((model) => ({
-    id: workspace ? model.workspaceProductId : model.productId,
-    upstreamModelId: model.upstreamModelId,
-    label: model.label,
-    capabilities: gatewayModelCapabilities(settings, {
-      implicitCaching: model.implicitCaching,
-      vision: model === OPENGENI_GATEWAY_MODELS.kimi,
-    }),
-    contextWindowTokens: 1_000_000,
-    effectiveContextWindowTokens: 900_000,
-    autoCompactTokenLimit: 850_000,
-    toolOutputTruncationTokens: settings.modelToolOutputTruncationTokens,
-  }));
+  const models = [OPENGENI_GATEWAY_MODELS.deepseek, OPENGENI_GATEWAY_MODELS.kimi].map((model) => {
+    const kimi = model === OPENGENI_GATEWAY_MODELS.kimi;
+    return {
+      id: workspace ? model.workspaceProductId : model.productId,
+      upstreamModelId: model.upstreamModelId,
+      label: model.label,
+      capabilities: gatewayModelCapabilities(settings, {
+        implicitCaching: model.implicitCaching,
+        vision: kimi,
+        inputFileMediaTypes: kimi ? ["application/pdf"] : [],
+      }),
+      contextWindowTokens: 1_000_000,
+      effectiveContextWindowTokens: 900_000,
+      autoCompactTokenLimit: 850_000,
+      toolOutputTruncationTokens: settings.modelToolOutputTruncationTokens,
+    };
+  });
   return {
     kind: input.kind,
     id: workspace ? WORKSPACE_GATEWAY_PROVIDER_ID : OPENGENI_GATEWAY_PROVIDER_ID,
@@ -2705,6 +2767,7 @@ export function withCodexCatalogProvider(settings: Settings): Settings {
         ...legacyModelCapabilities(settings, {
           reasoningEffort: true,
           hostedWebSearch: true,
+          vision: slug.startsWith("gpt-5.6-"),
         }),
         ...(builtinPromptCachingForModel(`${CODEX_MODEL_ID_PREFIX}${slug}`)
           ? {
@@ -2891,6 +2954,7 @@ export function configuredModels(settings: Settings): ConfiguredModel[] {
         ...legacyModelCapabilities(settings, {
           reasoningEffort: true,
           hostedWebSearch: settings.webSearchEnabled,
+          vision: id.startsWith("gpt-5.6-"),
         }),
         ...(builtinPromptCachingForModel(id)
           ? { promptCaching: builtinPromptCachingForModel(id)! }
@@ -3316,6 +3380,39 @@ export function calculateModelUsageCostMicros(
   return total;
 }
 
+/**
+ * Convert AI Gateway's exact USD inference cost to OpenGeni credit micros and
+ * apply the configured model margin. Decimal arithmetic is integer-only so a
+ * sub-micro provider charge cannot be lost to floating-point rounding.
+ */
+export function calculateGatewayReportedCostMicros(
+  settings: Settings,
+  model: string,
+  inferenceCostUsd: string,
+  options?: { inputTokens?: number },
+): number {
+  const schedule = configuredModelPricingSchedules(settings)[model];
+  if (!schedule) {
+    throw new Error(`Missing model pricing for ${model}`);
+  }
+  const pricing = selectModelPricing(schedule, positiveInt(options?.inputTokens));
+  const match = /^(0|[1-9]\d*)(?:\.(\d{1,18}))?$/.exec(inferenceCostUsd);
+  if (!match) {
+    throw new Error("Invalid AI Gateway inference cost");
+  }
+  const fraction = match[2] ?? "";
+  const decimalDigits = BigInt(`${match[1]}${fraction}`);
+  const decimalScale = 10n ** BigInt(fraction.length);
+  const marginBps = BigInt(10_000 + (pricing.marginBps ?? 0));
+  const numerator = decimalDigits * 1_000_000n * marginBps;
+  const denominator = decimalScale * 10_000n;
+  const micros = (numerator + denominator - 1n) / denominator;
+  if (micros > BigInt(Number.MAX_SAFE_INTEGER)) {
+    throw new Error("AI Gateway inference cost exceeds the supported billing range");
+  }
+  return Number(micros);
+}
+
 export function configuredAllowedReasoningEfforts(
   settings: Settings,
 ): Array<z.infer<typeof ReasoningEffort>> {