[BOT ISSUE] Anthropic prompt caching token metrics (cache_creation_input_tokens, cache_read_input_tokens) not captured

[braintrust-bot]

Summary

The Anthropic instrumentation extracts only input_tokens and output_tokens from the usage object, silently dropping the prompt caching fields cache_creation_input_tokens and cache_read_input_tokens. These fields are present in every Anthropic API response (including this repo's own test cassettes) and are critical for tracking prompt caching cost savings.

What is missing

In InstrumentationSemConv.tagAnthropicResponse() (lines 216-226), only two usage fields are extracted:

if (usage.has("input_tokens")) metrics.put("prompt_tokens", usage.get("input_tokens"));
if (usage.has("output_tokens")) metrics.put("completion_tokens", usage.get("output_tokens"));

The following fields from the Anthropic usage object are never extracted:

• cache_creation_input_tokens — tokens written to the cache (billed at 25% premium)
• cache_read_input_tokens — tokens read from the cache (billed at 90% discount)
• cache_creation.ephemeral_5m_input_tokens / cache_creation.ephemeral_1h_input_tokens — breakdown by TTL

For comparison, the OpenAI handler in the same file extracts output_tokens_details.reasoning_tokens (line 145-150), showing that nested usage detail extraction is an established pattern here.

The repo's own Anthropic test cassettes contain these fields — e.g., every response in test-harness/src/testFixtures/resources/cassettes/anthropic/__files/ includes "cache_creation_input_tokens":0,"cache_read_input_tokens":0 — but no test asserts they are captured in metrics.

Braintrust docs status

• Braintrust's Anthropic integration docs at https://braintrust.dev/docs/integrations/ai-providers/anthropic.md explicitly document Prompt Caching support with "ephemeral" cache control to "reduce costs and latency."
• Status: supported (documented as a feature, but cache token metrics are not captured in this Java SDK)

Upstream sources

• Anthropic prompt caching docs: https://docs.anthropic.com/en/docs/build-with-claude/prompt-caching
• Anthropic API usage object: https://docs.anthropic.com/en/api/messages — the usage object includes cache_creation_input_tokens and cache_read_input_tokens
• Anthropic Java SDK Usage class includes cacheCreationInputTokens() and cacheReadInputTokens() fields

Local files inspected

• braintrust-sdk/src/main/java/dev/braintrust/instrumentation/InstrumentationSemConv.java — lines 203-231 (tagAnthropicResponse)
• braintrust-sdk/instrumentation/anthropic_2_2_0/src/test/java/dev/braintrust/instrumentation/anthropic/v2_2_0/BraintrustAnthropicTest.java
• test-harness/src/testFixtures/resources/cassettes/anthropic/__files/v1_messages-759ba9d9-fbff-4177-8666-aac06350c678.json — example response with cache_creation_input_tokens and cache_read_input_tokens present but ignored