ChatRequest reasoning cannot express max_tokens — API accepts it, SDK silently strips it

[daveycodez]

Summary

The chat-completions HTTP API accepts a per-request reasoning token budget:

{ "reasoning": { "max_tokens": 2000 } }

but the SDK cannot send it. ChatRequestReasoning (models/chatrequest.ts) only models effort and summary:

export type ChatRequestReasoning = {
    effort?: ChatRequestEffort | null | undefined;
    summary?: ChatReasoningSummaryVerbosityEnum | null | undefined;
};

and because ChatRequestReasoning$outboundSchema is a plain z.object(...), an unknown key passed by a caller (e.g. via a type cast from JS/TS) is silently stripped before the request is serialized — the request goes out with no reasoning config at all, with no error or warning. That silent-drop behavior is the sharp edge: a caller who casts around the type believes they are testing a capped configuration while actually measuring the default.

Notably the SDK already models the right shape elsewhere — ReasoningConfig (models/reasoningconfig.ts) has maxTokens with a max_tokens wire mapping — it just isn't referenced by ChatRequest.

Ask

Add maxTokens to ChatRequestReasoning (serializing to max_tokens), or reference the existing ReasoningConfig from ChatRequest, so the SDK matches the documented API surface.

Use case

Effort tiers are too coarse for cost control on budget-based thinking models (Gemini flash family, Anthropic). On a large-batch classification workload the thinking tail scales with batch size — a per-request budget like max_tokens: batchItems * K caps the tail proportionally, which no fixed effort tier can express. Today the only SDK-expressible step below our current tier is a whole effort level down, which measurably degrades task quality.

Tested on @openrouter/sdk 0.13.20.