fix(anthropic): surface cache_read/write input tokens in metadata chunk

[jehoon-shin-mathflat]

Description of changes

Anthropic returns input_tokens as the non-cached portion only when prompt caching is in use. The metadata chunk formatter in models/anthropic.py reads input_tokens / output_tokens from the API response but drops the cache_read_input_tokens and cache_creation_input_tokens fields. As a result, anything that derives cost from Agent.metrics.accumulated_usage under-reports real usage — sometimes by an order of magnitude on image+text workloads where the cached prefix carries most of the prompt.

Repro

A vision classifier I run sends [{"text": ...}, {"cachePoint": {}}, {"image": ...}] so that the text+image prefix is cached. After warmup, the Anthropic API returns roughly:

{
  "input_tokens": 3,
  "cache_read_input_tokens": 1500,
  "cache_creation_input_tokens": 0,
  "output_tokens": 120
}

Strands currently surfaces only inputTokens=3, so accumulated cost looks ~10× lower than what Anthropic actually bills. The Usage TypedDict in types/event_loop.py already declares the cache fields as optional members — they just weren't being populated by the Anthropic adapter.

Change

models/anthropic.py — in the metadata case of format_chunk:

• Pull cache_read_input_tokens and cache_creation_input_tokens from the usage dict (defaulting to 0 when absent).
• Emit them as cacheReadInputTokens / cacheWriteInputTokens on the usage chunk, matching the existing field names already defined in types.event_loop.Usage.
• Recompute totalTokens as uncached_input + cache_read + cache_write + output_tokens so it reflects the full billed input.
• Omit the cache fields when both are zero so the chunk shape for non-cached responses is unchanged (no consumer needs to update).

Tests

tests/strands/models/test_anthropic.py:

• test_format_chunk_metadata_with_cache_tokens — both cache fields present.
• test_format_chunk_metadata_omits_zero_cache_tokens — fields absent/zero, legacy shape preserved.

All 59 tests in tests/strands/models/test_anthropic.py pass.

Related Issues

n/a — surfaced while debugging cost accounting in a downstream POC.

Documentation PR

n/a — no public API surface change, only adds optional fields that Usage already declares.

Type of change

• Bug fix (non-breaking change which fixes an issue)

Testing

pytest tests/strands/models/test_anthropic.py
# 59 passed

Checklist

• I have read the CONTRIBUTING document
• I have added tests that prove my fix is effective or that my feature works
• New and existing unit tests pass locally with my changes
• I have added/updated necessary documentation (if applicable)

By submitting this pull request, I confirm that you can use, modify, copy, and redistribute this contribution, under the terms of your choice.