fix(billing): correct cache token double-counting in Claude relay to non-Claude providers

[147API]

Problem

When a Claude-format request (/v1/messages) is relayed to a non-Claude provider (e.g. Gemini, OpenAI), PostClaudeConsumeQuota incorrectly assumes Claude usage semantics for the returned usage data.

Claude API returns input_tokens that already excludes cached tokens — so no subtraction is needed.

Gemini / OpenAI / etc. return prompt_tokens that includes cached tokens — so cacheTokens must be subtracted from promptTokens before adding cacheTokens * cacheRatio, otherwise cached tokens get billed twice.

Billing breakdown (before fix)

Item	Tokens	How billed
promptTokens	16202 (includes 15802 cached)	× modelRatio
cacheTokens	15802	× cacheRatio

The 15802 cached tokens are charged twice: once at full input price, once at cache price.

Billing breakdown (after fix)

Item	Tokens	How billed
promptTokens	400 (16202 − 15802)	× modelRatio
cacheTokens	15802	× cacheRatio

Root Cause

PostClaudeConsumeQuota only subtracted cache tokens for ChannelTypeOpenRouter. The existing postConsumeQuota in relay/compatible_handler.go already handles this correctly using GetFinalRequestRelayFormat() == RelayFormatClaude to distinguish usage semantics — PostClaudeConsumeQuota was missing the same guard.

Fix

Lift the cache-token subtraction out of the OpenRouter-only block and apply it to all non-Claude-semantic usage, consistent with compatible_handler.go:

isClaudeUsageSemantic := relayInfo.GetFinalRequestRelayFormat() == types.RelayFormatClaude
if !isClaudeUsageSemantic {
    promptTokens -= cacheTokens
    promptTokens -= cacheCreationTokens
}

The OpenRouter-specific cache-creation estimation logic is preserved as-is (it runs before the subtraction).

Affected Scenarios

Any request that:

1. Enters via the Claude Messages endpoint (/v1/messages)
2. Is relayed to a non-Claude provider (Gemini, OpenAI, Azure, DeepSeek, etc.)
3. Has cached tokens returned in the response usage

Made with Cursor

Summary by CodeRabbit

• Bug Fixes
  • Cached tokens are no longer deducted from prompt token totals for Claude-format final requests.
  • Other formats continue to deduct cached tokens from prompt totals.
  • Cache-creation tokens still count toward quota as before.

[coderabbitai]

No actionable comments were generated in the recent review. 🎉

ℹ️ Recent review info

⚙️ Run configuration

Configuration used: Organization UI

Review profile: CHILL

Plan: Pro

Run ID: e9eadeae-7aa4-466c-ab73-86ec52af7ad1

📥 Commits

Reviewing files that changed from the base of the PR and between ecd7203 and 2370be6.

📒 Files selected for processing (1)

• service/quota.go

🚧 Files skipped from review as they are similar to previous changes (1)

• service/quota.go

---

Walkthrough

The PR modifies quota consumption in service/quota.go: subtraction of cacheTokens from promptTokens is now skipped when the final relay format is Claude (isClaudeUsageSemantic). cacheCreationTokens subtraction remains unchanged; non-Claude paths still deduct cached tokens. Small control-flow reorder and clarifying comment added.

Changes

Cohort / File(s)	Summary
Claude Quota Caching Logic service/quota.go	Move cache-token subtraction out of OpenRouter-specific block and guard it with isClaudeUsageSemantic; skip subtracting cacheTokens for Claude-format final requests while retaining cacheCreationTokens deduction. Add explanatory comment and minor control-flow reordering.

Sequence Diagram(s)

(Skipped — change is a small internal control-flow tweak, not a multi-component feature requiring a sequence diagram.)

Estimated code review effort

🎯 2 (Simple) | ⏱️ ~10 minutes

Possibly related PRs

• fix: claude affinity cache counter #2980: Modifies PostClaudeConsumeQuota and observes channel-affinity usage; touches Claude quota/cache handling in the same area.
• fix: log info input_token (chat -> claude model ) #2670: Adds/uses isClaudeUsageSemantic to influence cached-token subtraction for Claude paths.
• fix: prevent OpenRouter cache calculation with custom model ratios #1590: Adjusts timing/guards for cache-related token deductions in service/quota.go.

Suggested reviewers

• seefs001
• Calcium-Ion

Poem

🐰 I hopped through quotas, nibbling each line,
Claude gets a pause where cached crumbs align.
Creation crumbs stay counted, neat and bright,
While cached ones rest softly out of sight.
Tokens tiptoe onward into the night.

🚥 Pre-merge checks | ✅ 2 | ❌ 1

❌ Failed checks (1 warning)

Check name	Status	Explanation	Resolution
Docstring Coverage	⚠️ Warning	Docstring coverage is 0.00% which is insufficient. The required threshold is 80.00%.	Write docstrings for the functions missing them to satisfy the coverage threshold.

✅ Passed checks (2 passed)

Check name	Status	Explanation
Description Check	✅ Passed	Check skipped - CodeRabbit’s high-level summary is enabled.
Title check	✅ Passed	The title accurately describes the primary fix: correcting a billing issue where cached tokens were double-counted when Claude-format requests were relayed to non-Claude providers.

_{✏️ Tip: You can configure your own custom pre-merge checks in the settings.}

✨ Finishing Touches

🧪 Generate unit tests (beta)

• Create PR with unit tests
• Post copyable unit tests in a comment

📝 Coding Plan

• Generate coding plan for human review comments

---

Thanks for using CodeRabbit! It's free for OSS, and your support helps us grow. If you like it, consider giving us a shout-out.

❤️ Share

• X
• Mastodon
• Reddit
• LinkedIn

_{Comment @coderabbitai help to get the list of available commands and usage tips.}

[147API]

@seefs001 @Calcium-Ion Could you please take a look at this fix when you have time? Thank you!

[147API]

问题场景是：claude code中调用gemini时出现的，缓存token与输入token重复计费