fix(ai): respect model output token limits

closes #4539

packages/ai/CHANGELOG.md

@@ -16,6 +16,7 @@
 - Fixed Inception Mercury 2 tool calling on OpenRouter by marking `off` as unsupported in `thinkingLevelMap`, so the openai-completions provider omits the reasoning param instead of defaulting to `{reasoning:{effort:"none"}}` (which puts Mercury 2 in instant mode, disabling tool calls).
 - Fixed OpenAI Codex SSE retries to honor `retry-after-ms` and `retry-after` headers before falling back to exponential backoff.
 - Fixed context overflow detection for LiteLLM-wrapped OpenAI-compatible errors using `exceeds the model's maximum context length of ... tokens` wording ([#4563](https://github.com/earendil-works/pi/issues/4563)).
+- Fixed `streamSimple()` defaults to respect model output limits above 32000 tokens instead of clamping provider requests to 32000 ([#4539](https://github.com/earendil-works/pi/issues/4539)).
 
 ## [0.74.0] - 2026-05-07
 

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

@@ -3,7 +3,7 @@ import type { Api, Model, SimpleStreamOptions, StreamOptions, ThinkingBudgets, T
 export function buildBaseOptions(model: Model<Api>, options?: SimpleStreamOptions, apiKey?: string): StreamOptions {
 	return {
 		temperature: options?.temperature,
-		maxTokens: options?.maxTokens ?? (model.maxTokens > 0 ? Math.min(model.maxTokens, 32000) : undefined),
+		maxTokens: options?.maxTokens ?? (model.maxTokens > 0 ? model.maxTokens : undefined),
 		signal: options?.signal,
 		apiKey: apiKey || options?.apiKey,
 		transport: options?.transport,