Awesome Reviewers

When integrating LLM providers/adapters, treat reasoning and token budgets as provider-specific data shapes and normalize them so downstream systems don’t double-count or receive invalid fields.

Apply these rules: 1) Normalize usage so reasoning tokens are not counted twice.

• If a provider reports reasoning as a subset of outputTokens (including nested shapes), subtract those reasoning tokens from outputTokens and track reasoning separately.

// Example pattern
const reasoningTokenCount = /* sum across configured reasoning token paths */;
const outputTokens = Math.max(0, normalizedUsage.outputTokens - reasoningTokenCount);
return {
  ...normalizedUsage,
  outputTokens,
  reasoningTokenCount,
};

2) Only send reasoning fields that the target backend supports.

• For models the resolver/adapter can’t classify as having reasoning support, do not emit reasoning fields that the backend rejects.

3) Map model capability fields correctly to avoid unsafe defaults.

• Prefer parsing alternate catalog field names (e.g., vLLM’s max_model_len) so you don’t silently fall back to an incorrect context/window.

4) Make budgeting semantics explicit.

• Treat “floor vs ceiling” intentionally (e.g., defaults as floors that may be raised by fractions, then clamped by model/context caps), and document the invariants.

5) Lock the behavior with regression tests at the integration boundary.

• Add tests for nested reasoning shapes and for “catalog lags launch” scenarios where resolution must remain stable.

This prevents cost/telemetry/session inconsistencies (double-booking), avoids invalid request payloads (backend errors), and reduces the risk of silent capability regressions caused by heterogeneous provider metadata.

This site is open source. Improve this page.