agents 0.28.1 → 0.29.0

package/dist/models/pi-ai/index.d.ts

@@ -1,18 +1,16 @@
 import {
-  d as DynamicRouteId,
-  m as WorkersAIModelId,
-  n as GatewayOptions,
-  o as CloudflareEmbeddingModelId,
-  r as ModelOptions,
-  t as AISettings
-} from "../../settings-DnD__xuE.js";
-import {
+  _ as DynamicRouteId,
   a as CloudflareAIError,
+  b as WorkersAIModelId,
+  c as GatewayOptions,
+  f as CloudflareEmbeddingModelId,
   i as CloudflareAIAttempt,
+  l as ModelOptions,
   n as GATEWAY_PROVIDER_NAMES,
   o as CloudflareAIErrorCode,
-  r as GatewayProviderName
-} from "../../transport-DCyarnwm.js";
+  r as GatewayProviderName,
+  s as AISettings
+} from "../../transport-CPh7YGiw.js";
 import {
   Api,
   AssistantImages,
@@ -108,7 +106,7 @@ declare const CLOUDFLARE_AI_API = "cloudflare-ai";
 type CloudflareAIApi = typeof CLOUDFLARE_AI_API;
 /**
  * The pi-ai provider id. A `Models` registry resolves specifiers as
- * `cloudflare/<id>`, e.g. `cloudflare/anthropic/claude-opus-4-8`.
+ * `cloudflare/<id>`, e.g. `cloudflare/anthropic/claude-opus-4.8`.
  *
  * @beta This surface is in beta and may change.
  */
@@ -267,9 +265,10 @@ interface AI {
 /**
  * Creates a pi-ai provider over Cloudflare's AI.
  *
- * The Workers AI binding is the only way in — Workers AI models go through
- * `env.AI.run`, vendor models through `env.AI.gateway(id).run`, and there is
- * no HTTP transport and no API token. Gateway options may be given flat
+ * Workers AI and catalog models go through `env.AI.run`, other model objects
+ * through the binding's `fetch` to AI Gateway's provider-native path. With
+ * `{ accountId, apiToken }` instead of `binding`, the same requests go to the
+ * REST API and to `gateway.ai.cloudflare.com`. Gateway options may be given flat
  * (`{ binding, id: "prod", cacheTtl: 60 }`) or nested under `gateway`;
  * per-model options and per-call pi-ai options (`headers`, `metadata`,
  * `sessionId`) override them in that order.

package/dist/models/pi-ai/index.js

@@ -1,5 +1,5 @@
-import { _ as isGatewayErrorEnvelope, a as normalizeChatCompletion, c as gatewaySlugForBaseUrl, d as resolveGatewayProvider, f as createTransport, g as headersToObject, h as errorFromResponse, i as normalizeChatCompletionsStream, l as isGatewayShapedUrl, m as errorFromGatewayEnvelope, n as imageMediaTypeOf, o as prepareChatCompletionsRequest, r as sseDataStream, s as GATEWAY_PROVIDER_NAMES, t as buildImageInput, v as unwrapEnvelope, w as isRunPathLanguageModelId, x as isAbortError, y as CloudflareAIError } from "../../images-C4hiIexM.js";
-import { i as mergeModelOptions, o as resolveOptions, r as gatewayLayerOf } from "../../settings-BjAiPd6Z.js";
+import { C as isCatalogModelId, T as isRunPathLanguageModelId, _ as isGatewayErrorEnvelope, a as normalizeChatCompletion, c as gatewaySlugForBaseUrl, d as resolveGatewayProvider, f as createTransport, g as headersToObject, h as errorFromResponse, i as normalizeChatCompletionsStream, l as isGatewayShapedUrl, m as errorFromGatewayEnvelope, n as imageMediaTypeOf, o as prepareChatCompletionsRequest, r as sseDataStream, s as GATEWAY_PROVIDER_NAMES, t as buildImageInput, v as unwrapEnvelope, x as isAbortError, y as CloudflareAIError } from "../../images-xaDv2Fqe.js";
+import { i as mergeModelOptions, o as resolveOptions, r as gatewayLayerOf } from "../../settings-Dk2il2Hf.js";
 import { calculateCost, clampThinkingLevel, createAssistantMessageEventStream, createProvider, getCurrentTools, normalizeContext, parseStreamingJson } from "@earendil-works/pi-ai";
 import { CLOUDFLARE_AI_GATEWAY_MODELS } from "@earendil-works/pi-ai/providers/cloudflare-ai-gateway.models";
 import { CLOUDFLARE_WORKERS_AI_MODELS } from "@earendil-works/pi-ai/providers/cloudflare-workers-ai.models";
@@ -19,7 +19,7 @@ import { clampOpenAIPromptCacheKey } from "@earendil-works/pi-ai/api/openai-prom
 const CLOUDFLARE_AI_API = "cloudflare-ai";
 /**
 * The pi-ai provider id. A `Models` registry resolves specifiers as
-* `cloudflare/<id>`, e.g. `cloudflare/anthropic/claude-opus-4-8`.
+* `cloudflare/<id>`, e.g. `cloudflare/anthropic/claude-opus-4.8`.
 *
 * @beta This surface is in beta and may change.
 */
@@ -125,9 +125,22 @@ function buildWorkersAIModel(modelId, options) {
 function slugOf(entry) {
 	return entry.baseUrl.split("/").filter((part) => part !== "").at(-1);
 }
+/**
+* The id AI Gateway's model catalog — `env.AI.run` — knows a registry entry
+* by. pi-ai spells Claude ids as Anthropic does (`claude-opus-4-8`), because pi
+* sends them through the gateway's Anthropic passthrough, while the catalog
+* puts a dot between version numbers (`anthropic/claude-opus-4.8`) and answers
+* `Model not found` to the dashed spelling. Checked against the live catalog
+* for every entry in pi-ai 1.2: that dot is the only difference, and only
+* Anthropic ids have it.
+*/
+function gatewayCatalogId(slug, id) {
+	return `${slug}/${slug === "anthropic" ? id.replace(/(\d)-(\d)(?!\d)/g, "$1.$2") : id}`;
+}
 let gatewayRegistry;
 /**
-* pi-ai's generated Cloudflare AI Gateway registry, keyed by `<slug>/<id>`.
+* pi-ai's generated Cloudflare AI Gateway registry, keyed by the id AI
+* Gateway's catalog runs each entry under (see {@link gatewayCatalogId}).
 * `/compat` entries are skipped: they alias `@cf/` ids the Workers AI registry
 * already declares, and those belong on the run path.
 */
@@ -136,7 +149,7 @@ function buildGatewayRegistry() {
 	for (const entry of Object.values(CLOUDFLARE_AI_GATEWAY_MODELS)) {
 		const slug = slugOf(entry);
 		if (slug === void 0 || slug === "compat") continue;
-		map.set(`${slug}/${entry.id}`, entry);
+		map.set(gatewayCatalogId(slug, entry.id), entry);
 	}
 	return map;
 }
@@ -205,7 +218,7 @@ const API_PATHS = {
 	"openai-responses": "responses"
 };
 /**
-* The universal request's `endpoint` for a gateway-shaped base URL, whose path
+* The provider-native `endpoint` for a gateway-shaped base URL, whose path
 * already ends at the provider slug, so what follows is the vendor's own path.
 */
 const GATEWAY_ENDPOINTS = {
@@ -482,26 +495,61 @@ function openAIVendorHeaders() {
 	};
 }
 /**
-* Sends a vendor's own request body through AI Gateway's universal endpoint.
+* The AI Gateway catalog id a model runs under on `/run`, when it has one: the
+* `<author>/<model>` specifier it was resolved from, spelled as the gateway's
+* catalog spells it (`anthropic/claude-opus-4.8`, not pi's `claude-opus-4-8`).
+*/
+function catalogIdOf(model) {
+	const specifier = specifierOf(model);
+	return specifier !== void 0 && isCatalogModelId(specifier) ? specifier : void 0;
+}
+/**
+* Sends a vendor-format request body through AI Gateway, by the binding.
+*
+* A model from pi's Cloudflare AI Gateway catalog runs on `env.AI.run` under
+* its catalog id, in its native `format`: `/run` converts nothing, so the
+* answer — streams included — is the shape this wire parses. Any other model
+* goes to the gateway's provider-native path with the request unchanged; the
+* provider slug comes from the model's base URL and the endpoint from its
+* `api`.
 *
-* The provider slug comes from the model's base URL and the endpoint from its
-* `api`; the body is what pi-ai's converter built, `model` included, because
-* that is what the universal request carries. The gateway's *own* failures
-* (no such gateway, unified billing unpaid) are lifted into a
-* `CloudflareAIError` so fallback and retry classification behave as they do on
-* the Workers AI run path; a vendor's own error response is left alone for the
-* wire's parser to read.
+* The gateway's *own* failures (no such gateway, unified billing unpaid, an
+* unknown catalog model) become a `CloudflareAIError`, so fallback and retry
+* behave as they do on the Workers AI run path; a vendor's own error response
+* on the provider-native path is left for the wire's parser to read.
 */
-async function sendUniversal(request, transport, body, vendorHeaders, onResponse) {
+async function sendGateway(request, transport, body, vendorHeaders, format, onResponse) {
 	const { model, options, resolved } = request;
+	const catalogId = catalogIdOf(model);
+	if (catalogId !== void 0) {
+		const { model: _vendorId, ...input } = body;
+		const response = await transport.run({
+			format,
+			gateway: resolved.gateway,
+			headers: request.headers,
+			input,
+			model: catalogId,
+			signal: options.signal
+		});
+		onResponse?.(response);
+		if (!response.ok) throw await errorFromResponse(response, {
+			model: catalogId,
+			requestBodyValues: input,
+			url: transport.url
+		});
+		return response;
+	}
 	const route = routeForModel(model);
-	const response = await transport.universal({
+	const response = await transport.native({
+		body: JSON.stringify(body),
 		endpoint: route.endpoint,
-		extraHeaders: Object.keys(request.headers).length > 0 ? request.headers : void 0,
 		gateway: resolved.gateway,
-		headers: vendorHeaders,
+		headers: {
+			...vendorHeaders,
+			...request.headers
+		},
+		method: "POST",
 		provider: route.provider,
-		query: body,
 		signal: options.signal
 	});
 	onResponse?.(response);
@@ -599,7 +647,7 @@ async function assertOk(response, context) {
 //#region src/models/pi-ai/wires/chat-completions.ts
 /**
 * Chat-completions wire: Workers AI on the run path, and any pi-ai model whose
-* `api` is `openai-completions` on AI Gateway's universal endpoint.
+* `api` is `openai-completions`, through AI Gateway.
 *
 * The request is shaped with pi-ai's own message converter so replay rules
 * (tool-call id normalization, reasoning replay, role bridging) match pi's
@@ -836,7 +884,7 @@ var CompletionAssembler = class {
 		const choice = chunk.choices?.[0];
 		if (chunk.id !== void 0) this.output.responseId ??= chunk.id;
 		if (typeof chunk.model === "string" && chunk.model.length > 0 && chunk.model !== this.model.id) this.output.responseModel ??= chunk.model;
-		if (chunk.usage !== void 0) this.output.usage = usageFromOpenAI(this.model, chunk.usage);
+		if (chunk.usage !== void 0 && chunk.usage !== null) this.output.usage = usageFromOpenAI(this.model, chunk.usage);
 		const finished = this.applyFinishReason(choice?.finish_reason);
 		const delta = choice?.delta ?? choice?.message;
 		if (delta !== void 0) {
@@ -960,12 +1008,13 @@ function streamCompletions(request, transport) {
 			const overridden = await options.onPayload?.(prepared.body, model);
 			const payload = overridden === void 0 ? prepared.body : overridden;
 			response = workersAI ? await transport.run({
+				format: "chat-completions",
 				gateway: resolved.gateway,
 				headers: request.headers,
 				input: payload,
 				model: model.id,
 				signal: options.signal
-			}) : await sendUniversal(request, transport, payload, openAIVendorHeaders());
+			}) : await sendGateway(request, transport, payload, openAIVendorHeaders(), "chat-completions");
 			await options.onResponse?.(responseInfo(response), model);
 			attachCorrelation(output, correlationDetails(request, response, transport));
 			await assertOk(response, {
@@ -1257,7 +1306,7 @@ function buildRequest(model, context, options, simple, config) {
 //#region src/models/pi-ai/wires/anthropic.ts
 /**
 * Anthropic Messages wire: pi-ai's own Anthropic implementation, over AI
-* Gateway's universal endpoint.
+* Gateway — `/run` for a catalog model, the provider-native path otherwise.
 *
 * Everything about the wire — request shaping, thinking, tool use, stream
 * parsing, thinking-signature replay — is pi-ai's. It accepts a `client` whose
@@ -1317,7 +1366,7 @@ function streamAnthropic(request, transport) {
 			model: wireModelId(model)
 		};
 		try {
-			const response = await sendUniversal(request, transport, body, anthropicVendorHeaders(Array.isArray(betas) ? betas : []), (raw) => {
+			const response = await sendGateway(request, transport, body, anthropicVendorHeaders(Array.isArray(betas) ? betas : []), "messages", (raw) => {
 				correlation = correlationDetails(request, raw, transport);
 			});
 			await assertOk(response, {
@@ -1380,7 +1429,8 @@ function streamAnthropic(request, transport) {
 //#region src/models/pi-ai/wires/responses.ts
 /**
 * OpenAI Responses wire: pi-ai's own Responses converters and stream
-* processor, over AI Gateway's universal endpoint.
+* processor, over AI Gateway: `/run` for a catalog model, the provider-native
+* path otherwise.
 *
 * Only the transport differs from talking to OpenAI directly: the body goes to
 * the gateway, which forwards it to `api.openai.com/v1/responses` with the key
@@ -1445,7 +1495,7 @@ function streamResponses(request, transport) {
 			if (sampling !== void 0) Object.assign(body, sampling);
 			const overridden = await options.onPayload?.(body, model);
 			const payload = overridden === void 0 ? body : overridden;
-			response = await sendUniversal(request, transport, payload, openAIVendorHeaders());
+			response = await sendGateway(request, transport, payload, openAIVendorHeaders(), "responses");
 			await options.onResponse?.(responseInfo(response), model);
 			attachCorrelation(output, correlationDetails(request, response, transport));
 			recordWarnings(output, ignoredWorkersAIKnobs(request));
@@ -1612,7 +1662,8 @@ function streamWithFallback(legs, signal) {
 * The `api` is the model author's declaration, not a guess of ours: Workers AI
 * models carry {@link CLOUDFLARE_AI_API} and run on `env.AI.run`; every other
 * model carries whatever protocol its registry says it speaks, and goes to AI
-* Gateway's universal endpoint.
+* Gateway in that protocol: `/run` for a catalog model, the provider-native
+* path otherwise.
 */
 /** A stream that fails immediately, for a model this provider cannot route. */
 function failedStream(request, error) {
@@ -1700,8 +1751,9 @@ function createStreams(config) {
 * ```
 *
 * Workers AI runs on `env.AI.run` with Cloudflare's own compat layer. Every
-* other model goes to AI Gateway's universal endpoint, which forwards the
-* vendor's own request to the vendor with the key the gateway holds — so no
+* other model goes to AI Gateway in its own protocol — `/run` for a model
+* from the gateway's catalog, the provider-native path otherwise — and the
+* gateway calls the vendor with the key it holds, so no
 * provider key, and no vendor wire format, lives in your Worker. The model's
 * metadata (`api`, `compat`, `thinkingLevelMap`, cost, limits) is whatever its
 * own registry says, and pi-ai's converters and stream processors do the rest.
@@ -1740,9 +1792,10 @@ function buildModel(target, options) {
 /**
 * Creates a pi-ai provider over Cloudflare's AI.
 *
-* The Workers AI binding is the only way in — Workers AI models go through
-* `env.AI.run`, vendor models through `env.AI.gateway(id).run`, and there is
-* no HTTP transport and no API token. Gateway options may be given flat
+* Workers AI and catalog models go through `env.AI.run`, other model objects
+* through the binding's `fetch` to AI Gateway's provider-native path. With
+* `{ accountId, apiToken }` instead of `binding`, the same requests go to the
+* REST API and to `gateway.ai.cloudflare.com`. Gateway options may be given flat
 * (`{ binding, id: "prod", cacheTtl: 60 }`) or nested under `gateway`;
 * per-model options and per-call pi-ai options (`headers`, `metadata`,
 * `sessionId`) override them in that order.