fix(ai): set bedrock claude default max tokens

closes #4848

packages/ai/CHANGELOG.md

@@ -2,6 +2,10 @@
 
 ## [Unreleased]
 
+### Fixed
+
+- Fixed Amazon Bedrock Claude requests to send the model output token cap by default, matching Anthropic requests and avoiding Bedrock's 4096-token default truncation ([#4848](https://github.com/earendil-works/pi/issues/4848)).
+
 ## [0.75.4] - 2026-05-20
 
 ### Changed

packages/ai/src/providers/amazon-bedrock.ts

@@ -183,12 +183,13 @@ export const streamBedrock: StreamFunction<"bedrock-converse-stream", BedrockOpt
 		try {
 			const client = new BedrockRuntimeClient(config);
 			const cacheRetention = resolveCacheRetention(options.cacheRetention);
+			const inferenceMaxTokens = options.maxTokens ?? (isAnthropicClaudeModel(model) ? model.maxTokens : undefined);
 			let commandInput = {
 				modelId: model.id,
 				messages: convertMessages(context, model, cacheRetention),
 				system: buildSystemPrompt(context.systemPrompt, model, cacheRetention),
 				inferenceConfig: {
-					...(options.maxTokens !== undefined && { maxTokens: options.maxTokens }),
+					...(inferenceMaxTokens !== undefined && { maxTokens: inferenceMaxTokens }),
 					...(options.temperature !== undefined && { temperature: options.temperature }),
 				},
 				toolConfig: convertToolConfig(context.tools, options.toolChoice),

packages/ai/test/bedrock-thinking-payload.test.ts

@@ -2,6 +2,7 @@ import { describe, expect, it } from "vitest";
 import { getModel } from "../src/models.ts";
 import { type BedrockOptions, streamBedrock } from "../src/providers/amazon-bedrock.ts";
 import type { Context, Model } from "../src/types.ts";
+import { hasBedrockCredentials } from "./bedrock-utils.ts";
 
 interface BedrockThinkingPayload {
 	additionalModelRequestFields?: {
@@ -112,6 +113,39 @@ describe("Bedrock thinking payload", () => {
 	});
 });
 
+describe.skipIf(!hasBedrockCredentials())("Bedrock Claude max tokens E2E", () => {
+	it(
+		"uses the model maxTokens cap instead of Bedrock's 4096-token default for adaptive Claude models",
+		{ retry: 2, timeout: 180000 },
+		async () => {
+			const baseModel = getModel("amazon-bedrock", "global.anthropic.claude-sonnet-4-6");
+			const model: Model<"bedrock-converse-stream"> = {
+				...baseModel,
+				maxTokens: 6000,
+			};
+
+			const response = await streamBedrock(
+				model,
+				{
+					systemPrompt: "You are a deterministic text generator. Follow the requested output format exactly.",
+					messages: [
+						{
+							role: "user",
+							content:
+								"Output exactly 5200 repetitions of the token alpha, separated by single spaces. Do not number them. Do not use markdown. Do not add any other text.",
+							timestamp: Date.now(),
+						},
+					],
+				},
+				{ reasoning: "low" },
+			).result();
+
+			expect(response.stopReason, response.errorMessage).not.toBe("error");
+			expect(response.usage.output).toBeGreaterThan(4096);
+		},
+	);
+});
+
 describe("Application inference profile support", () => {
 	it("uses adaptive thinking when model.name contains the model name but ARN does not", async () => {
 		const baseModel = getModel("amazon-bedrock", "global.anthropic.claude-opus-4-6-v1");