[Bug] Generated catalog contains models where maxTokens exceeds contextWindow (Inkling-Small, openrouter gpt-3.5-turbo-0613)

[aniruddhaadak80]

Area

AI providers and models

Description

Two entries in the generated model catalog (packages/ai/src/models.generated.ts, current main) declare maxTokens larger than their own contextWindow, which is impossible for any single request:

provider	id	contextWindow	maxTokens	line
huggingface	thinkingmachines/Inkling-Small	524288	1048576	~7078
openrouter	openai/gpt-3.5-turbo-0613	4095	4096	~13169

For comparison, the sibling entry thinkingmachines/Inkling (huggingface) declares contextWindow: 1048576 with maxTokens: 1048576, so Inkling-Small's output limit looks copy-pasted from its larger sibling while its window is half the size. The openrouter entry is off by exactly one token (4096 > 4095), suggesting an upstream rounding/data quirk that the generator copied verbatim.

Both values flow into request construction: packages/ai/src/providers/simple-options.ts:6 derives the default output budget from model.maxTokens (clamped to 32000, which masks the problem in the default path), and packages/ai/src/providers/openai-completions.ts:540 sends options.maxTokens verbatim when a caller sets it explicitly. So today the practical blast radius is limited to SDK/extension callers passing explicit budgets and to anything consuming these fields as metadata - but the invariant itself is broken, and nothing currently catches it when upstream data regenerates.

Steps to reproduce

1. Open packages/ai/src/models.generated.ts on main.
2. Inspect the "thinkingmachines/Inkling-Small" (huggingface) and "openai/gpt-3.5-turbo-0613" (openrouter) entries.
3. Compare contextWindow vs maxTokens.

Or mechanically: scan all } satisfies Model<...> entries and assert maxTokens <= contextWindow.

Expected behavior

Every catalog entry satisfies 0 < maxTokens <= contextWindow. Values that cannot be satisfied from upstream data should be clamped or overridden in packages/ai/scripts/generate-models.ts with a comment explaining why.

Actual behavior

maxTokens > contextWindow on the two entries listed above; requests built from explicit maxTokens in that range are guaranteed to be rejected by the provider.

Prime Agent version

main @ e319a66

Environment

Any (generated data; platform-independent)

Additional context

I first filed this as issue #1677; it was auto-closed by the contribution vouch gate within seconds, so I am re-filing it here per CONTRIBUTING.md's discussion-first process.

Suggested fix: add an assertion (or normalize-and-warn step) in the generator so regeneration fails or corrects instead of silently emitting impossible limits, then override/clamp the two known offenders.