@maxceem/agw 0.7.0 → 0.7.2

package/dist/agw.mjs

@@ -6,7 +6,7 @@ var __export = (target, all) => {
 };
 
 // <define:__AGW_MANIFEST__>
-var define_AGW_MANIFEST_default = { version: "0.7.0", wrangler: "4.96.0", upgradeFrom: ["0.7.0", "0.6.4", "0.6.3", "0.6.2", "0.6.1", "0.6.0", "0.5.2", "0.5.1", "0.5.0", "0.4.0"], release: { url: "https://github.com/maxceem/app-ai-gateway/releases/download/v0.7.0/gateway-0.7.0.tar.gz", sha256: "9cbcb05a1b611470c03e522af7e34d786a4d7e9889b589bac7603695187575bd", manifest: "707e02868d2133307ff4d71a3fbcb7abe6d0b62cbc9498fab5b7d3f6d2d96fb0" } };
+var define_AGW_MANIFEST_default = { version: "0.7.2", wrangler: "4.96.0", upgradeFrom: ["0.7.2", "0.7.1", "0.7.0", "0.6.4", "0.6.3", "0.6.2", "0.6.1", "0.6.0", "0.5.2", "0.5.1", "0.5.0", "0.4.0"], release: { url: "https://github.com/maxceem/app-ai-gateway/releases/download/v0.7.2/gateway-0.7.2.tar.gz", sha256: "6673bdde6e77fcdbe5da4c13013d084e99354b67cc4ba43704b9dc81b40e1ff7", manifest: "ff987daf1a4740041f3410ac7f852a3b0bce4d98a8fc9abe2cf835f35b2d4d3a" } };
 
 // src/main.ts
 import { Writable as Writable2 } from "node:stream";
@@ -14699,7 +14699,9 @@ var GATEWAY_DESCRIPTORS = {
       { key: "accountId", label: "Cloudflare Account ID" },
       { key: "gatewayId", label: "Cloudflare Gateway ID" }
     ],
-    routes: CF_AIG_ROUTES
+    routes: CF_AIG_ROUTES,
+    // A proxy in front of the provider's own API: the body is the provider's.
+    relaysProviderResponses: true
   },
   vercel: {
     label: "Vercel AI Gateway",
@@ -14715,7 +14717,11 @@ var GATEWAY_DESCRIPTORS = {
       description: "Vercel's origin is fixed in adapter code, so it has no configuration of its own."
     }),
     connectionFields: [],
-    routes: VERCEL_ROUTES
+    routes: VERCEL_ROUTES,
+    // Vercel answers through its own client APIs, translating between them and
+    // the serving provider's protocol, so a provider's own fields are not part
+    // of the contract.
+    relaysProviderResponses: false
   }
 };
 var GATEWAY_TYPES = Object.keys(GATEWAY_DESCRIPTORS);
@@ -14767,9 +14773,18 @@ function reportedByok(root, usage2) {
   return upstreamInferenceCost(usage2) !== null;
 }
 var OPENROUTER_COST_REPORT = {
+  // Only chat completions carry `usage.cost`, which is why it is the only
+  // style the provider type offers at all.
+  styles: ["chat_completions"],
   // Which host actually served a request is opt-in per request; without this
   // header the response names no serving provider at all.
   requestHeaders: { "x-openrouter-metadata": "enabled" },
+  // `credentialSource` is deliberately not required: it is a claim that is
+  // simply absent on a non-BYOK request, so waiting for it would mean walking
+  // every chunk of every stream. The response that carries it carries the cost
+  // too.
+  settled: (report) => report.costUsd !== null && report.servedProvider !== null && report.servedModel !== null,
+  readsCredentialSource: true,
   read(root, report) {
     let said = false;
     const usage2 = asRecord(root.usage);
@@ -14801,6 +14816,23 @@ var OPENROUTER_COST_REPORT = {
     return said;
   }
 };
+var XAI_TICKS_PER_USD = 1e10;
+var XAI_COST_REPORT = {
+  // Images and videos are provider-native operations, `other` to this gateway.
+  // Its Anthropic-compatible Messages API and `v1/stt` are not documented to
+  // report a cost, so they keep billing on the catalog.
+  styles: ["chat_completions", "responses", "other"],
+  settled: (report) => report.costUsd !== null,
+  readsCredentialSource: false,
+  read(root, report) {
+    if (report.costUsd !== null) return false;
+    const usage2 = asRecord(root.usage) ?? asRecord(asRecord(root.response)?.usage);
+    const ticks = usage2 ? finiteNumber(usage2.cost_in_usd_ticks) : null;
+    if (ticks === null || ticks < 0) return false;
+    report.costUsd = ticks / XAI_TICKS_PER_USD;
+    return true;
+  }
+};
 
 // ../src/shared/providers.ts
 var PROVIDER_DESCRIPTORS = {
@@ -14809,6 +14841,13 @@ var PROVIDER_DESCRIPTORS = {
     directBaseUrl: "https://api.openai.com/",
     auth: { header: "authorization", scheme: "Bearer " },
     nativeClampStyle: "responses",
+    // Images, which report token usage, embeddings and speech.
+    inferencePaths: [
+      "v1/images/generations",
+      "v1/images/edits",
+      "v1/embeddings",
+      "v1/audio/speech"
+    ],
     probePath: "v1/models",
     examplePath: "v1/responses",
     // OpenAI's chat-completions surface deprecated `max_tokens` in favour of
@@ -14818,6 +14857,9 @@ var PROVIDER_DESCRIPTORS = {
     // The two types whose Responses and transcription request shapes the
     // gateway composes itself for custom endpoints.
     endpointPaths: { responses: "v1/responses", audio_transcription: "v1/audio/transcriptions" },
+    // Every JSON answer reports `usage`: tokens for the GPT transcription
+    // models, `{ type: "duration", seconds }` for whisper-1.
+    transcriptionFormats: { metered: ["json", "verbose_json", "diarized_json"], default: "json" },
     modelAuthor: "OpenAI"
   },
   anthropic: {
@@ -14840,10 +14882,23 @@ var PROVIDER_DESCRIPTORS = {
     directBaseUrl: "https://api.x.ai/",
     auth: { header: "authorization", scheme: "Bearer " },
     nativeClampStyle: "responses",
+    // Images and their edits report their cost like chat does; embeddings and
+    // legacy completions answer in tokens, the latter capped in `max_tokens`.
+    inferencePaths: [
+      "v1/images/generations",
+      "v1/images/edits",
+      "v1/embeddings",
+      { path: "v1/completions", clamp: "chat_completions" }
+    ],
     probePath: "v1/models",
     // Native provider paths: xAI transcribes at `v1/stt`, where OpenAI serves
     // `v1/audio/transcriptions`.
     endpointPaths: { responses: "v1/responses", audio_transcription: "v1/stt" },
+    // Every Chat Completions, Responses, image and video answer carries
+    // `usage.cost_in_usd_ticks`, what xAI actually billed for it. An image
+    // reports nothing else, so this is the only way to bill one; on the token
+    // APIs it beats the catalog's estimate, which knows no discounts.
+    costReport: XAI_COST_REPORT,
     modelAuthor: "xAI"
   },
   gemini: {
@@ -14851,6 +14906,13 @@ var PROVIDER_DESCRIPTORS = {
     directBaseUrl: "https://generativelanguage.googleapis.com/",
     auth: { header: "x-goog-api-key" },
     nativeClampStyle: "gemini_native",
+    // Embeddings. `generateContent` already carries text, images, speech and
+    // transcription; `batchEmbedContents` is left out because each item may name
+    // a model the one in the URL does not.
+    inferencePaths: [
+      "v1beta/models/{model}:embedContent",
+      "v1/models/{model}:embedContent"
+    ],
     probePath: "v1beta/models",
     // Native Gemini generation requests carry the model in the URL rather than
     // in the JSON body, so there is no example path without a model.
@@ -14862,6 +14924,15 @@ var PROVIDER_DESCRIPTORS = {
     directBaseUrl: "https://api.perplexity.ai/",
     auth: { header: "authorization", scheme: "Bearer " },
     nativeClampStyle: "responses",
+    // The Agent API, its model router's three formats, and embeddings.
+    inferencePaths: [
+      { path: "v1/agent", clamp: "responses" },
+      { path: "router/v1/chat/completions", clamp: "chat_completions" },
+      { path: "router/v1/responses", clamp: "responses" },
+      { path: "router/v1/messages", clamp: "anthropic" },
+      "v1/embeddings",
+      "v1/contextualizedembeddings"
+    ],
     // No `probePath`: Perplexity has no unmetered authenticated endpoint.
     examplePath: "chat/completions",
     modelAuthor: "Perplexity"
@@ -14877,6 +14948,13 @@ var PROVIDER_DESCRIPTORS = {
     directBaseUrl: "https://api.deepseek.com/",
     auth: { header: "authorization", scheme: "Bearer " },
     nativeClampStyle: "chat_completions",
+    // Responses, fill-in-the-middle completions and the Anthropic-format API,
+    // which like its chat path carry no `v1/` prefix of their own.
+    inferencePaths: [
+      { path: "responses", clamp: "responses" },
+      { path: "beta/completions", clamp: "chat_completions" },
+      { path: "anthropic/v1/messages", clamp: "anthropic" }
+    ],
     // DeepSeek's OpenAI base URL carries no `v1` segment.
     probePath: "models",
     examplePath: "chat/completions",
@@ -14889,6 +14967,15 @@ var PROVIDER_DESCRIPTORS = {
     directBaseUrl: "https://api.groq.com/",
     auth: { header: "authorization", scheme: "Bearer " },
     nativeClampStyle: "chat_completions",
+    // Its default `json` answer is the transcript alone; `verbose_json` adds the
+    // duration it bills by, and the same `text` a client reads.
+    transcriptionFormats: {
+      metered: ["verbose_json"],
+      default: "json",
+      upgrades: { json: "verbose_json" }
+    },
+    // Speech; transcriptions are a default style already.
+    inferencePaths: ["openai/v1/audio/speech"],
     // Groq's OpenAI-compatible surface is namespaced under `openai/`.
     probePath: "openai/v1/models",
     examplePath: "openai/v1/chat/completions"
@@ -14898,6 +14985,12 @@ var PROVIDER_DESCRIPTORS = {
     directBaseUrl: "https://api.mistral.ai/",
     auth: { header: "authorization", scheme: "Bearer " },
     nativeClampStyle: "chat_completions",
+    // Embeddings, fill-in-the-middle completions and speech.
+    inferencePaths: [
+      "v1/embeddings",
+      { path: "v1/fim/completions", clamp: "chat_completions" },
+      "v1/audio/speech"
+    ],
     probePath: "v1/models",
     modelAuthor: "Mistral"
   },
@@ -14908,6 +15001,21 @@ var PROVIDER_DESCRIPTORS = {
     directBaseUrl: "https://api.together.ai/",
     auth: { header: "authorization", scheme: "Bearer " },
     nativeClampStyle: "chat_completions",
+    // Its default `json` answer is the transcript alone; `verbose_json` adds the
+    // duration it bills by, and the same `text` a client reads.
+    transcriptionFormats: {
+      metered: ["verbose_json"],
+      default: "json",
+      upgrades: { json: "verbose_json" }
+    },
+    // Legacy completions, images, speech and embeddings. Video is a job whose
+    // result is polled for, so it is not here.
+    inferencePaths: [
+      { path: "v1/completions", clamp: "chat_completions" },
+      "v1/images/generations",
+      "v1/audio/speech",
+      "v1/embeddings"
+    ],
     probePath: "v1/models"
   },
   fireworks: {
@@ -14917,6 +15025,13 @@ var PROVIDER_DESCRIPTORS = {
     directBaseUrl: "https://api.fireworks.ai/",
     auth: { header: "authorization", scheme: "Bearer " },
     nativeClampStyle: "chat_completions",
+    // Every format its inference plane speaks, and embeddings.
+    inferencePaths: [
+      { path: "inference/v1/completions", clamp: "chat_completions" },
+      { path: "inference/v1/messages", clamp: "anthropic" },
+      { path: "inference/v1/responses", clamp: "responses" },
+      "inference/v1/embeddings"
+    ],
     // No `probePath`: its list-models call is `v1/accounts/{account}/models`,
     // and the account id cannot be derived from the key. Nothing under
     // `inference/v1/` is documented as a GET.
@@ -14927,6 +15042,8 @@ var PROVIDER_DESCRIPTORS = {
     directBaseUrl: "https://api.cerebras.ai/",
     auth: { header: "authorization", scheme: "Bearer " },
     nativeClampStyle: "chat_completions",
+    // Legacy completions.
+    inferencePaths: [{ path: "v1/completions", clamp: "chat_completions" }],
     probePath: "v1/models"
   },
   moonshot: {
@@ -14936,6 +15053,8 @@ var PROVIDER_DESCRIPTORS = {
     directBaseUrl: "https://api.moonshot.ai/",
     auth: { header: "authorization", scheme: "Bearer " },
     nativeClampStyle: "chat_completions",
+    // The Anthropic-format API.
+    inferencePaths: [{ path: "anthropic/v1/messages", clamp: "anthropic" }],
     probePath: "v1/models",
     modelAuthor: "Moonshot AI"
   },
@@ -14973,6 +15092,13 @@ var PROVIDER_DESCRIPTORS = {
     directBaseUrl: "https://ark.ap-southeast.bytepluses.com/api/v3/",
     auth: { header: "authorization", scheme: "Bearer " },
     nativeClampStyle: "chat_completions",
+    // Responses, Seedream images and multimodal embeddings, all under the base
+    // URL's own `api/v3/`. Seedance video is a polled job, so it is not here.
+    inferencePaths: [
+      { path: "responses", clamp: "responses" },
+      "images/generations",
+      "embeddings/multimodal"
+    ],
     // No `probePath`: ModelArk publishes no list-models call, and its own SDK
     // has no models resource. It also authenticates before it routes, so every
     // path answers the same 401 and a probe would prove nothing about the key.
@@ -15015,9 +15141,6 @@ var PROVIDER_TYPES = Object.keys(PROVIDER_DESCRIPTORS);
 function providerDescriptor(type) {
   return PROVIDER_DESCRIPTORS[type];
 }
-function isProviderType(value) {
-  return typeof value === "string" && Object.hasOwn(PROVIDER_DESCRIPTORS, value);
-}
 var PROVIDER_CREDENTIAL_HEADERS = [
   ...new Set(
     PROVIDER_TYPES.map((type) => providerDescriptor(type).auth.header.toLowerCase())
@@ -15033,10 +15156,6 @@ function providerCapability(provider) {
 var ENDPOINT_PROVIDER_TYPES = PROVIDER_TYPES.filter(
   (type) => providerCapability(type).endpointStyles.length > 0
 );
-function reportsCost(type) {
-  return isProviderType(type) && providerDescriptor(type).costReport !== void 0;
-}
-var COST_REPORTING_PROVIDER_TYPES = PROVIDER_TYPES.filter(reportsCost);
 
 // ../src/shared/app-status.ts
 var APP_STATUSES = ["active", "disabled"];
@@ -16676,6 +16795,12 @@ var StateStore = class {
 function selectedProviderPolicies(routing) {
   return routing.providers.mode === "selected" ? routing.providers.selected : {};
 }
+var DEFAULT_INFERENCE_PATHS = new Map(
+  PROVIDER_TYPES.map((type) => [
+    type,
+    (providerDescriptor(type).inferencePaths ?? []).map((entry) => typeof entry === "string" ? { path: entry, clamp: "none" } : { ...entry })
+  ])
+);
 
 // ../src/shared/protocols.ts
 var PROTOCOLS = {
@@ -17849,7 +17974,7 @@ var UsageRepriceResponseSchema = external_exports.object({
     description: "Matched events that carried no readable usage, and so were repriced to the zero their zero counts imply while keeping whatever cost_source they had. A non-zero number means the month contains spend nothing could meter, which repricing cannot fix and must not appear to have fixed."
   }),
   unpriced_events: external_exports.number().int().meta({
-    description: "Dry-run only: matched events whose serving instance can no longer price them."
+    description: "Dry-run only: matched events whose serving instance can no longer price them, or that predate the modality breakdown their model's price depends on."
   }),
   unpriced_cost_usd: external_exports.number(),
   previous_cost_usd: external_exports.number(),
@@ -17859,10 +17984,25 @@ var UsageRepriceResponseSchema = external_exports.object({
 var ModelPriceSchema = external_exports.object({
   input: external_exports.number().optional(),
   output: external_exports.number().optional(),
+  image_input: external_exports.number().optional().meta({
+    description: "USD per 1M image input tokens, where they are priced apart from text input."
+  }),
+  audio_input: external_exports.number().optional().meta({
+    description: "USD per 1M audio input tokens, where they are priced apart from text input."
+  }),
+  video_input: external_exports.number().optional().meta({
+    description: "USD per 1M video input tokens, where they are priced apart from text input."
+  }),
+  image_output: external_exports.number().optional().meta({
+    description: "USD per 1M generated image tokens, where they cost more than text output."
+  }),
   cached_input: external_exports.number().optional(),
   cache_write: external_exports.number().optional(),
   per_minute: external_exports.number().optional(),
   per_hour: external_exports.number().optional(),
+  per_image: external_exports.number().optional().meta({
+    description: "USD per generated image, for a model billed per image rather than per token."
+  }),
   long_context_threshold: external_exports.number().optional(),
   long_input: external_exports.number().optional(),
   long_output: external_exports.number().optional(),

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@maxceem/agw",
-  "version": "0.7.0",
+  "version": "0.7.2",
   "type": "module",
   "description": "Manage App AI Gateway from a terminal or coding agent: providers, iOS and server apps, usage, and self-hosted deployment on Cloudflare",
   "keywords": [
@@ -25,7 +25,19 @@
     "directory": "cli"
   },
   "bugs": "https://github.com/maxceem/app-ai-gateway/issues",
-  "upgradeFrom": ["0.6.4", "0.6.3", "0.6.2", "0.6.1", "0.6.0", "0.5.2", "0.5.1", "0.5.0", "0.4.0"],
+  "upgradeFrom": [
+    "0.7.1",
+    "0.7.0",
+    "0.6.4",
+    "0.6.3",
+    "0.6.2",
+    "0.6.1",
+    "0.6.0",
+    "0.5.2",
+    "0.5.1",
+    "0.5.0",
+    "0.4.0"
+  ],
   "license": "Apache-2.0",
   "bin": {
     "agw": "dist/agw.mjs"