fix: prevent double counting of Anthropic streaming cache tokens

[Aias00]

Description

The message_delta event handler in anthropicStreamParser was using Add* methods for input tokens, cached tokens, and cache creation tokens. However, the Anthropic API reports cumulative (not incremental) token counts in message_delta, and message_start already sets these values correctly via ExtractTokenUsageFromExplicitCaching. Using Add on top of the already-set cumulative values caused cache tokens to be counted twice.

Bug Example

With input_tokens=9 and cache_read_input_tokens=1:

Step	Action	InputTokens	CachedInputTokens
message_start	SetInputTokens(10), SetCachedInputTokens(1)	10 ✅	1 ✅
message_delta (before fix)	AddInputTokens(1), AddCachedInputTokens(1)	11 ❌	2 ❌
message_delta (after fix)	Only SetOutputTokens(16)	10 ✅	1 ✅

Root Cause

ExtractTokenUsageFromExplicitCaching computes InputTokens = input_tokens + cache_read + cache_creation. In message_start, this is correctly set via Set* methods. In message_delta, the same cumulative values were being added via Add* methods, causing double counting.

Fix

The message_delta handler now only updates output tokens using SetOutputTokens, since message_start typically reports output_tokens=0 and message_delta provides the final count. This matches the pattern already used in anthropic_anthropic.go's reflectStreamingEvent. The MessageDeltaUsage struct uses non-pointer int64 fields that default to 0 when absent from JSON, making it impossible to distinguish "not provided" from "actually zero" — so we cannot safely Set override from delta either, as it would erase correct message_start values with zeros.

Changes

• internal/translator/anthropic_helper.go: Removed Add* calls for input/cache tokens from message_delta handler; replaced AddOutputTokens with SetOutputTokens; removed unused ExtractTokenUsageFromExplicitCaching call; added >= 0 guard for ThinkingTokens
• internal/translator/anthropic_helper_test.go: Added TestAnthropicStreamParserTokenUsage_NoDoubleCounting with 6 test cases

Test Coverage

• Cache tokens in both message_start and message_delta (core double-count scenario)
• Cache creation tokens in both events
• Both cache_read and cache_creation in both events
• No cache tokens (baseline correctness)
• Cache tokens only in message_start, not in message_delta
• Cache tokens only in message_delta — verifies they are ignored (non-pointer int64 fields cannot distinguish absent from zero)

[Copilot]

Pull request overview

Note

Copilot was unable to run its full agentic suite in this review.

This PR updates the Anthropic streaming token usage parsing to prevent double-counting cache-related tokens when both message_start and message_delta report usage, and adds a regression test to validate the behavior.

Changes:

• Adjust message_delta handling to treat usage values as cumulative and avoid accumulating cache/input tokens twice.
• Add a new table-driven test covering several cache token combinations across message_start and message_delta.
• Introduce testify/assert usage in the new test.

Reviewed changes

Copilot reviewed 2 out of 2 changed files in this pull request and generated 3 comments.

File	Description
internal/translator/anthropic_helper.go	Changes message_delta token accounting to avoid double-counting and sets output tokens as cumulative.
internal/translator/anthropic_helper_test.go	Adds a regression test ensuring cache tokens aren’t double-counted across streaming events.

---

💡 Add Copilot custom instructions for smarter, more guided reviews. Learn how to get started.

[codecov-commenter]

Codecov Report

✅ All modified and coverable lines are covered by tests.
✅ Project coverage is 84.89%. Comparing base (f5f83b2) to head (b41080e).

Additional details and impacted files

@@            Coverage Diff             @@
##             main    #2236      +/-   ##
==========================================
+ Coverage   84.87%   84.89%   +0.01%     
==========================================
  Files         144      144              
  Lines       21317    21345      +28     
==========================================
+ Hits        18093    18121      +28     
  Misses       2138     2138              
  Partials     1086     1086              

☔ View full report in Codecov by Harness.
📢 Have feedback on the report? Share it here.

🚀 New features to boost your workflow:

• ❄️ Test Analytics: Detect flaky tests, report on failures, and find test suite problems.
• 📦 JS Bundle Analysis: Save yourself from yourself by tracking and limiting bundle sizes in JS merges.

[Aias00]

close: #2279

[Aias00]

/retest

[Aias00]

/retest

[nacx]

Overall LGTM.
@aabchoo mind taking a look?

[nacx]

/retest