Merge main into model-registry

packages/ai/src/api/anthropic-messages.ts

@@ -5,6 +5,7 @@ import type {
 	MessageCreateParamsStreaming,
 	MessageParam,
 	RawMessageStreamEvent,
+	RefusalStopDetails,
 } from "@anthropic-ai/sdk/resources/messages.js";
 import { calculateCost } from "../models.ts";
 import type {
@@ -16,6 +17,7 @@ import type {
 	ImageContent,
 	Message,
 	Model,
+	ProviderEnv,
 	SimpleStreamOptions,
 	StopReason,
 	StreamFunction,
@@ -29,6 +31,7 @@ import type {
 import { AssistantMessageEventStream } from "../utils/event-stream.ts";
 import { headersToRecord } from "../utils/headers.ts";
 import { parseJsonWithRepair, parseStreamingJson } from "../utils/json-parse.ts";
+import { getProviderEnvValue } from "../utils/provider-env.ts";
 import { sanitizeSurrogates } from "../utils/sanitize-unicode.ts";
 
 import { resolveCloudflareBaseUrl } from "./cloudflare.ts";
@@ -40,11 +43,11 @@ import { transformMessages } from "./transform-messages.ts";
  * Resolve cache retention preference.
  * Defaults to "short" and uses PI_CACHE_RETENTION for backward compatibility.
  */
-function resolveCacheRetention(cacheRetention?: CacheRetention): CacheRetention {
+function resolveCacheRetention(cacheRetention?: CacheRetention, env?: ProviderEnv): CacheRetention {
 	if (cacheRetention) {
 		return cacheRetention;
 	}
-	if (typeof process !== "undefined" && process.env.PI_CACHE_RETENTION === "long") {
+	if (getProviderEnvValue("PI_CACHE_RETENTION", env) === "long") {
 		return "long";
 	}
 	return "short";
@@ -53,8 +56,9 @@ function resolveCacheRetention(cacheRetention?: CacheRetention): CacheRetention
 function getCacheControl(
 	model: Model<"anthropic-messages">,
 	cacheRetention?: CacheRetention,
+	env?: ProviderEnv,
 ): { retention: CacheRetention; cacheControl?: CacheControlEphemeral } {
-	const retention = resolveCacheRetention(cacheRetention);
+	const retention = resolveCacheRetention(cacheRetention, env);
 	if (retention === "none") {
 		return { retention };
 	}
@@ -493,7 +497,7 @@ export const stream: StreamFunction<"anthropic-messages", AnthropicOptions> = (
 					});
 				}
 
-				const cacheRetention = options?.cacheRetention ?? resolveCacheRetention();
+				const cacheRetention = resolveCacheRetention(options?.cacheRetention, options?.env);
 				const cacheSessionId = cacheRetention === "none" ? undefined : options?.sessionId;
 
 				const created = createClient(
@@ -504,6 +508,7 @@ export const stream: StreamFunction<"anthropic-messages", AnthropicOptions> = (
 					options?.headers,
 					copilotDynamicHeaders,
 					cacheSessionId,
+					options?.env,
 				);
 				client = created.client;
 				isOAuth = created.isOAuthToken;
@@ -534,6 +539,7 @@ export const stream: StreamFunction<"anthropic-messages", AnthropicOptions> = (
 					output.usage.output = event.message.usage.output_tokens || 0;
 					output.usage.cacheRead = event.message.usage.cache_read_input_tokens || 0;
 					output.usage.cacheWrite = event.message.usage.cache_creation_input_tokens || 0;
+					output.usage.cacheWrite1h = event.message.usage.cache_creation?.ephemeral_1h_input_tokens || 0;
 					// Anthropic doesn't provide total_tokens, compute from components
 					output.usage.totalTokens =
 						output.usage.input + output.usage.output + output.usage.cacheRead + output.usage.cacheWrite;
@@ -660,7 +666,11 @@ export const stream: StreamFunction<"anthropic-messages", AnthropicOptions> = (
 					}
 				} else if (event.type === "message_delta") {
 					if (event.delta.stop_reason) {
-						output.stopReason = mapStopReason(event.delta.stop_reason);
+						const stopReasonResult = mapStopReason(event.delta.stop_reason, event.delta.stop_details);
+						output.stopReason = stopReasonResult.stopReason;
+						if (stopReasonResult.errorMessage) {
+							output.errorMessage = stopReasonResult.errorMessage;
+						}
 					}
 					// Only update usage fields if present (not null).
 					// Preserves input_tokens from message_start when proxies omit it in message_delta.
@@ -688,7 +698,7 @@ export const stream: StreamFunction<"anthropic-messages", AnthropicOptions> = (
 			}
 
 			if (output.stopReason === "aborted" || output.stopReason === "error") {
-				throw new Error("An unknown error occurred");
+				throw new Error(output.errorMessage || "An unknown error occurred");
 			}
 
 			stream.push({ type: "done", reason: output.stopReason, message: output });
@@ -788,6 +798,7 @@ function createClient(
 	optionsHeaders?: Record<string, string>,
 	dynamicHeaders?: Record<string, string>,
 	sessionId?: string,
+	env?: ProviderEnv,
 ): { client: Anthropic; isOAuthToken: boolean } {
 	// Adaptive thinking models have interleaved thinking built in, so skip the beta header.
 	const needsInterleavedBeta = interleavedThinking && model.compat?.forceAdaptiveThinking !== true;
@@ -803,7 +814,7 @@ function createClient(
 		const client = new Anthropic({
 			apiKey: null,
 			authToken: null,
-			baseURL: resolveCloudflareBaseUrl(model),
+			baseURL: resolveCloudflareBaseUrl(model, env),
 			dangerouslyAllowBrowser: true,
 			defaultHeaders: mergeHeaders(
 				{
@@ -896,7 +907,7 @@ function buildParams(
 	isOAuthToken: boolean,
 	options?: AnthropicOptions,
 ): MessageCreateParamsStreaming {
-	const { cacheControl } = getCacheControl(model, options?.cacheRetention);
+	const { cacheControl } = getCacheControl(model, options?.cacheRetention, options?.env);
 	const compat = getAnthropicCompat(model);
 	const params: MessageCreateParamsStreaming = {
 		model: model.id,
@@ -1202,22 +1213,28 @@ function convertTools(
 	});
 }
 
-function mapStopReason(reason: Anthropic.Messages.StopReason | string): StopReason {
+function mapStopReason(
+	reason: Anthropic.Messages.StopReason | string,
+	stopDetails?: RefusalStopDetails | null,
+): { stopReason: StopReason; errorMessage?: string } {
 	switch (reason) {
 		case "end_turn":
-			return "stop";
+			return { stopReason: "stop" };
 		case "max_tokens":
-			return "length";
+			return { stopReason: "length" };
 		case "tool_use":
-			return "toolUse";
+			return { stopReason: "toolUse" };
 		case "refusal":
-			return "error";
+			return {
+				stopReason: "error",
+				errorMessage: stopDetails?.explanation || `The model refused to complete the request`,
+			};
 		case "pause_turn": // Stop is good enough -> resubmit
-			return "stop";
+			return { stopReason: "stop" };
 		case "stop_sequence":
-			return "stop"; // We don't supply stop sequences, so this should never happen
+			return { stopReason: "stop" }; // We don't supply stop sequences, so this should never happen
 		case "sensitive": // Content flagged by safety filters (not yet in SDK types)
-			return "error";
+			return { stopReason: "error" };
 		default:
 			// Handle unknown stop reasons gracefully (API may add new values)
 			throw new Error(`Unhandled stop reason: ${reason}`);