@maxceem/agw 0.6.4 → 0.7.1

package/dist/agw.mjs

@@ -6,7 +6,7 @@ var __export = (target, all) => {
 };
 
 // <define:__AGW_MANIFEST__>
-var define_AGW_MANIFEST_default = { version: "0.6.4", wrangler: "4.96.0", 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"], release: { url: "https://github.com/maxceem/app-ai-gateway/releases/download/v0.6.4/gateway-0.6.4.tar.gz", sha256: "cd40eaa3809e18174d57091191363f14cf5ca5c00850e826ccb0d5710f7ce80c", manifest: "bdb9f85fc16464e81f03c0f14d3127b804e06ad97f4d31f0d50f8d205d1d1e7c" } };
+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" } };
 
 // 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 = {
@@ -14844,6 +14876,11 @@ var PROVIDER_DESCRIPTORS = {
     // 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: {
@@ -15015,9 +15052,6 @@ 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 +15067,6 @@ 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"];
@@ -17849,7 +17879,7 @@ 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 +17889,25 @@ 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(),
@@ -19059,7 +19104,7 @@ var CATALOG = {
     path: "/v1/cli/bootstrap",
     tags: ["CLI"],
     summary: "Initialize a recoverable CLI account",
-    description: "Persist the random token before sending: its digest is the operation's id, and an identical retry returns the same account and protected credential during its exchange window. Cloud initialization is public and rate limited. Self-hosted initialization is public too and creates the deployment's single initial account, so whoever initializes an empty deployment first owns it, exactly as its first console registration does; it is not rate limited, because the only thing a limit could refuse on a deployment nobody owns yet is its own installer retrying, and every caller after the first is refused permanently anyway. All responses are no-store.",
+    description: "Persist the random token before sending: its digest is the operation's id, and an identical retry returns the same account and protected credential during its exchange window. Cloud initialization is public and rate limited. Self-hosted initialization creates the deployment's single unclaimed account, and only for the token whose SHA-256 the deploying CLI stored as the Worker secret `CLI_BOOTSTRAP_TOKEN_DIGEST`; any other token, or a deployment without that secret, is refused with 403. It is not rate limited, because only the deploying CLI can pass it, and a limit would refuse that CLI retrying a deployment whose storage has not come up yet. All responses are no-store.",
     security: "public",
     request: CliBootstrapRequestSchema,
     response: CliOperationSchema,
@@ -20308,7 +20353,7 @@ async function openBrowser(url2) {
 }
 
 // src/deployment.ts
-import { randomUUID } from "node:crypto";
+import { createHash as createHash4, randomUUID } from "node:crypto";
 import { writeFile as writeFile3, unlink as unlink2 } from "node:fs/promises";
 import { join as join4 } from "node:path";
 
@@ -20850,7 +20895,7 @@ async function configFile(ctx, journal, artifact) {
     ],
     vars: {
       ...artifact.config.vars,
-      ALLOW_ADDITIONAL_REGISTRATIONS: "false",
+      ALLOWED_REGISTRATION_EMAILS: "",
       ...journal.vars,
       DEPLOYMENT_ID: journal.id
     },
@@ -21201,7 +21246,10 @@ async function install(ctx, cf, journal, artifact, url2, domain2) {
   });
   const secrets = {
     ...journal.secrets,
-    SECRET_VAULT_LOCAL_KEK_V1: await ctx.store.vaultKey(journal.id)
+    SECRET_VAULT_LOCAL_KEK_V1: await ctx.store.vaultKey(journal.id),
+    // The Worker creates its account only for the token this digest is of, so
+    // the account is this CLI's however soon anyone else finds the address.
+    CLI_BOOTSTRAP_TOKEN_DIGEST: createHash4("sha256").update(journal.bootstrap.token).digest("hex")
   };
   const secretPath = join4(file2.directory, "secrets.json");
   await writeFile3(secretPath, JSON.stringify(secrets), {

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@maxceem/agw",
-  "version": "0.6.4",
+  "version": "0.7.1",
   "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 +25,18 @@
     "directory": "cli"
   },
   "bugs": "https://github.com/maxceem/app-ai-gateway/issues",
-  "upgradeFrom": ["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": [
+    "0.7.0",
+    "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"