fix(worker): cap completion budget by model context window

max_tokens=16384 caused HTTP 400 on small-context models (e.g. 16k-window
models); budget is now min(16384, context_length - 4096) with a 2048 floor.

Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>

Simon-Pierre Boucher committed 6 days ago (Aug 5, 2026) parent 6cf7329

Showing 1 changed file with +7 and −2

modified apps/worker/src/eval-runner.ts +7 −2

		`@@ -118,6 +118,11 @@ export async function runEvalBatch(opts: EvalBatchOptions): Promise<{ runId: str`
118	118	const client = new OpenRouterClient();
119	119	let failures = 0;
120	120	let completedCalls = 0;
	121	+ // Small-context models reject max_tokens near/above their window (HTTP 400):
	122	+ // cap the completion budget by the context length, keeping prompt headroom.
	123	+ const maxTokens = model.contextLength
	124	+ ? Math.min(MAX_COMPLETION_TOKENS, Math.max(2048, model.contextLength - 4096))
	125	+ : MAX_COMPLETION_TOKENS;
121	126
122	127	interface Call {
123	128	itemIdx: number;
		`@@ -150,8 +155,8 @@ export async function runEvalBatch(opts: EvalBatchOptions): Promise<{ runId: str`
150	155	// Scored sample: temperature 0. Consistency samples: model default temperature (§6).
151	156	const requestParams =
152	157	sampleIndex === 0
153		− ? { model: model.slug, messages, temperature: 0, max_tokens: MAX_COMPLETION_TOKENS }
154		− : { model: model.slug, messages, max_tokens: MAX_COMPLETION_TOKENS };
	158	+ ? { model: model.slug, messages, temperature: 0, max_tokens: maxTokens }
	159	+ : { model: model.slug, messages, max_tokens: maxTokens };
155	160	// Audit copy: keep exact params but elide the base64 image payload (the
156	161	// scene is reproducible from the item's stored SVG/perturb seed).
157	162	const auditParams = item.svg
158	163