fix(pricing): stop billing reasoning tokens twice

[yidecode]

calculateCostFromTokens charges reasoning tokens on top of the completion tokens that already contain them.

const outputTokens = tokens.completion_tokens || tokens.output_tokens || 0;
cost += outputTokens * (pricing.output / 1000000);

const reasoningTokens = tokens.reasoning_tokens || 0;
if (reasoningTokens > 0) {
  cost += reasoningTokens * ((pricing.reasoning || pricing.output) / 1000000);   // added, not substituted
}

Reasoning tokens are a subset of completion tokens everywhere they are consumed:

• OpenAI counts reasoning_tokens inside completion_tokens (spec), and extractUsage's OpenAI branch reads it straight out of completion_tokens_details.
• toOpenAIUsage's gemini extractor returns completionTokens: candidates + thoughts and reasoningTokens: thoughts — already folded.

So every reasoning request was billed for its thinking twice. Because MODEL_PRICING sets reasoning equal to output for most entries (claude-sonnet-4-6: output 15.00, reasoning 15.00, gpt-5: output 10.00, reasoning 10.00, …), pricing.reasoning || pricing.output resolves to the output rate and the second line is a straight 2x on the thinking portion. On a thinking-heavy request that is a ~30% overstatement of the whole call.

The fix

Bill only the differential, and only when a model actually prices reasoning apart from output:

if (reasoningTokens > 0 && pricing.reasoning) {
  cost += reasoningTokens * ((pricing.reasoning - pricing.output) / 1000000);
}

This is the same contract the function already applies one block up, where cached and cache_creation are subtracted from prompt_tokens precisely because that field is cache-inclusive. The comment there already spells it out: "prompt_tokens is cache-inclusive (see canonicalizeUsage): cached + cache_creation are subsets, so subtract both to avoid charging them at the full input rate." Completion tokens deserve the same treatment.

Why two gemini call sites change too

extractUsage (usageTracking.js) and extractUsageFromResponse (requestDetail.js) both reported gemini usage with thoughtsTokenCount outside candidatesTokenCount. Left alone, the pricing change would have turned into an under-charge for gemini. Both now fold thoughts in, matching what toOpenAIUsage has always done — so the reasoning-inclusive invariant holds on every path that reaches the cost function.

Total gemini cost is unchanged by the fold: candidates*output + thoughts*reasoning either way. Only the field split changes. There is a test pinning that equivalence explicitly.

Heads-up for reviewers

Reported cost for reasoning models goes down after this change. That is the intended outcome — the previous figures over-counted — but it will be visible in the dashboard, so it should not come as a surprise.

Verification

• 9 new tests: the reasoning-inclusive convention (gemini fold, OpenAI passthrough, agreement with toOpenAIUsage), and the cost math for reasoning priced at / above / absent from the output rate, plus the gemini cost-equivalence guard.
• Full suite: failure set identical to the master baseline, zero regression.

本 PR 由 Claude Code 辅助完成