@maxceem/agw 0.7.1 → 0.7.2

package/dist/agw.mjs

@@ -6,7 +6,7 @@ var __export = (target, all) => {
 };
 
 // <define:__AGW_MANIFEST__>
-var define_AGW_MANIFEST_default = { version: "0.7.1", wrangler: "4.96.0", 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"], release: { url: "https://github.com/maxceem/app-ai-gateway/releases/download/v0.7.1/gateway-0.7.1.tar.gz", sha256: "1b3ac4141c993b30bf1e90404dee7a21785a38b686c3f2562ec7b0c07fd4534d", manifest: "8455d531a2dc7df63af0fcbe1d8a181c3b6c74796e5f5a69cacf7890ca2dc42a" } };
+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";
@@ -14841,6 +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
@@ -14850,6 +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: {
@@ -14872,6 +14882,14 @@ 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`.
@@ -14888,6 +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.
@@ -14899,6 +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"
@@ -14914,6 +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",
@@ -14926,6 +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"
@@ -14935,6 +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"
   },
@@ -14945,6 +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: {
@@ -14954,6 +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.
@@ -14964,6 +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: {
@@ -14973,6 +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"
   },
@@ -15010,6 +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.
@@ -16706,6 +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 = {

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@maxceem/agw",
-  "version": "0.7.1",
+  "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": [
@@ -26,6 +26,7 @@
   },
   "bugs": "https://github.com/maxceem/app-ai-gateway/issues",
   "upgradeFrom": [
+    "0.7.1",
     "0.7.0",
     "0.6.4",
     "0.6.3",