fix(ai): stop defaulting max token request caps

closes #4675

packages/ai/CHANGELOG.md

@@ -4,6 +4,7 @@
 
 ### Fixed
 
+- Fixed OpenAI-compatible `streamSimple()` requests to stop sending model-derived default output token caps, avoiding context-window reservation failures on servers such as vLLM while preserving explicit `maxTokens` and required Anthropic `max_tokens` handling ([#4675](https://github.com/earendil-works/pi/issues/4675)).
 - Fixed OpenAI prompt cache keys to clamp session-derived values to the 64-character API limit across OpenAI Responses, Chat Completions, Codex Responses, and Azure OpenAI Responses ([#4720](https://github.com/earendil-works/pi/issues/4720)).
 
 ## [0.75.3] - 2026-05-18

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

@@ -314,8 +314,10 @@ export const streamSimpleBedrock: StreamFunction<"bedrock-converse-stream", Simp
 			} satisfies BedrockOptions);
 		}
 
+		// Undefined means the caller did not request an output cap; let the helper use the model cap.
+		// Do not coerce to 0 here, or the thinking budget would become the entire maxTokens value.
 		const adjusted = adjustMaxTokensForThinking(
-			base.maxTokens || 0,
+			base.maxTokens,
 			model.maxTokens,
 			options.reasoning,
 			options.thinkingBudgets,

packages/ai/src/providers/anthropic.ts

@@ -751,8 +751,10 @@ export const streamSimpleAnthropic: StreamFunction<"anthropic-messages", SimpleS
 		} satisfies AnthropicOptions);
 	}
 
+	// Undefined means the caller did not request an output cap; let the helper use the model cap.
+	// Do not coerce to 0 here, or the thinking budget would become the entire max_tokens value.
 	const adjusted = adjustMaxTokensForThinking(
-		base.maxTokens || 0,
+		base.maxTokens,
 		model.maxTokens,
 		options.reasoning,
 		options.thinkingBudgets,
@@ -891,7 +893,7 @@ function buildParams(
 	const params: MessageCreateParamsStreaming = {
 		model: model.id,
 		messages: convertMessages(context.messages, model, isOAuthToken, cacheControl),
-		max_tokens: options?.maxTokens || (model.maxTokens / 3) | 0,
+		max_tokens: options?.maxTokens ?? model.maxTokens,
 		stream: true,
 	};
 

packages/ai/src/providers/simple-options.ts

@@ -1,19 +1,9 @@
 import type { Api, Model, SimpleStreamOptions, StreamOptions, ThinkingBudgets, ThinkingLevel } from "../types.js";
 
-const DEFAULT_MAX_OUTPUT_TOKENS = 32000;
-const CONTEXT_WINDOW_OUTPUT_TOLERANCE = 1024;
-
-export function buildBaseOptions(model: Model<Api>, options?: SimpleStreamOptions, apiKey?: string): StreamOptions {
-	const defaultMaxTokens =
-		model.maxTokens > 0
-			? model.maxTokens >= model.contextWindow - CONTEXT_WINDOW_OUTPUT_TOLERANCE
-				? Math.min(model.maxTokens, DEFAULT_MAX_OUTPUT_TOKENS)
-				: model.maxTokens
-			: undefined;
-
+export function buildBaseOptions(_model: Model<Api>, options?: SimpleStreamOptions, apiKey?: string): StreamOptions {
 	return {
 		temperature: options?.temperature,
-		maxTokens: options?.maxTokens ?? defaultMaxTokens,
+		maxTokens: options?.maxTokens,
 		signal: options?.signal,
 		apiKey: apiKey || options?.apiKey,
 		transport: options?.transport,
@@ -34,7 +24,8 @@ export function clampReasoning(effort: ThinkingLevel | undefined): Exclude<Think
 }
 
 export function adjustMaxTokensForThinking(
-	baseMaxTokens: number,
+	// Undefined means no explicit caller cap. Use the model cap and fit thinking inside it.
+	baseMaxTokens: number | undefined,
 	modelMaxTokens: number,
 	reasoningLevel: ThinkingLevel,
 	customBudgets?: ThinkingBudgets,
@@ -50,7 +41,8 @@ export function adjustMaxTokensForThinking(
 	const minOutputTokens = 1024;
 	const level = clampReasoning(reasoningLevel)!;
 	let thinkingBudget = budgets[level]!;
-	const maxTokens = Math.min(baseMaxTokens + thinkingBudget, modelMaxTokens);
+	const maxTokens =
+		baseMaxTokens === undefined ? modelMaxTokens : Math.min(baseMaxTokens + thinkingBudget, modelMaxTokens);
 
 	if (maxTokens <= thinkingBudget) {
 		thinkingBudget = Math.max(0, maxTokens - minOutputTokens);

packages/ai/test/github-copilot-anthropic.test.ts

@@ -87,7 +87,7 @@ describe("Copilot Claude via Anthropic Messages", () => {
 		const params = mockState.createParams!;
 		expect(params.model).toBe("claude-sonnet-4.6");
 		expect(params.stream).toBe(true);
-		expect(params.max_tokens).toBeGreaterThan(0);
+		expect(params.max_tokens).toBe(model.maxTokens);
 		expect(Array.isArray(params.messages)).toBe(true);
 	});
 

packages/ai/test/openai-completions-empty-tools.test.ts

@@ -93,6 +93,40 @@ describe("openai-completions empty tools handling", () => {
 		expect("tools" in (params as object)).toBe(false);
 	});
 
+	it("does not send default max token fields", async () => {
+		const { compat: _compat, ...baseModel } = getModel("openai", "gpt-4o-mini")!;
+		const model = { ...baseModel, api: "openai-completions" } as const;
+
+		await streamSimple(
+			model,
+			{
+				messages: [{ role: "user", content: "hi", timestamp: Date.now() }],
+			},
+			{ apiKey: "test" },
+		).result();
+
+		const params = mockState.lastParams as { max_tokens?: number; max_completion_tokens?: number };
+		expect(params.max_tokens).toBeUndefined();
+		expect(params.max_completion_tokens).toBeUndefined();
+	});
+
+	it("sends explicit maxTokens", async () => {
+		const { compat: _compat, ...baseModel } = getModel("openai", "gpt-4o-mini")!;
+		const model = { ...baseModel, api: "openai-completions" } as const;
+
+		await streamSimple(
+			model,
+			{
+				messages: [{ role: "user", content: "hi", timestamp: Date.now() }],
+			},
+			{ apiKey: "test", maxTokens: 1234 },
+		).result();
+
+		const params = mockState.lastParams as { max_tokens?: number; max_completion_tokens?: number };
+		expect(params.max_tokens).toBeUndefined();
+		expect(params.max_completion_tokens).toBe(1234);
+	});
+
 	it("uses conservative OpenAI-compatible fields for Cloudflare AI Gateway /compat models", async () => {
 		process.env.CLOUDFLARE_ACCOUNT_ID = "account-id";
 		process.env.CLOUDFLARE_GATEWAY_ID = "gateway-id";
@@ -104,7 +138,7 @@ describe("openai-completions empty tools handling", () => {
 				systemPrompt: "You are helpful.",
 				messages: [{ role: "user", content: "hi", timestamp: Date.now() }],
 			},
-			{ apiKey: "test", reasoning: "high" },
+			{ apiKey: "test", maxTokens: 1234, reasoning: "high" },
 		).result();
 
 		const params = mockState.lastParams as {
@@ -115,7 +149,7 @@ describe("openai-completions empty tools handling", () => {
 			store?: boolean;
 		};
 		expect(params.messages[0].role).toBe("system");
-		expect(params.max_tokens).toBeDefined();
+		expect(params.max_tokens).toBe(1234);
 		expect(params.max_completion_tokens).toBeUndefined();
 		expect(params.reasoning_effort).toBeUndefined();
 		expect(params.store).toBeUndefined();