Python: surface cache and reasoning token counts for the Bedrock and Gemini connectors

[he-yufeng]

Motivation & Context

Several connectors drop token counts that UsageDetails already has canonical fields for, so cache and reasoning usage silently reads as zero for those providers. This PR fixes that for the two connectors that were missing it:

• Bedrock: the Converse API reports cacheReadInputTokens (input tokens served from a cache) and cacheWriteInputTokens (input tokens written to a cache) when prompt caching is active, but _parse_usage dropped both.
• Gemini: the response usage metadata reports cached_content_token_count and thoughts_token_count, but _parse_usage dropped both.

UsageDetails already defines cache_read_input_token_count, cache_creation_input_token_count, and reasoning_output_token_count, and the OpenAI/Anthropic connectors already populate the cache fields, so these two connectors were the odd ones out. The change is the same small is not None mapping pattern in each connector's _parse_usage.

Description & Review Guide

• What are the major changes?
  • BedrockChatClient._parse_usage maps cacheReadInputTokens -> cache_read_input_token_count and cacheWriteInputTokens -> cache_creation_input_token_count.
  • GeminiChatClient._parse_usage maps cached_content_token_count -> cache_read_input_token_count and thoughts_token_count -> reasoning_output_token_count.
  • Bedrock's _parse_usage now returns details or None (matching its UsageDetails | None annotation and the Gemini connector), so a usage payload with no recognized keys no longer propagates an empty mapping.
• What is the impact of these changes? Cache and reasoning token counts are now reported for Bedrock and Gemini, consistent with the OpenAI/Anthropic connectors. Responses without caching/thinking are unchanged (the fields stay unset).
• What do you want reviewers to focus on? That the provider field names map to the right UsageDetails keys.

Added focused tests in both packages (test_parse_usage_surfaces_cache_tokens, test_parse_usage_returns_none_when_no_recognized_keys, test_get_response_usage_details_includes_cached_and_reasoning_tokens) and ran both suites locally (Bedrock: 9 passed; Gemini suite green). ruff check and ruff format --check pass on the changed files.

[Copilot]

Pull request overview

Surfaces Bedrock Converse prompt-cache token counts in the framework’s canonical usage_details so cached prompts don’t report zero cache usage, aligning Bedrock behavior with other connectors that already populate these fields.

Changes:

• Map Bedrock Converse cacheReadInputTokens → cache_read_input_token_count and cacheWriteInputTokens → cache_creation_input_token_count in BedrockChatClient._parse_usage.
• Add a Bedrock unit test asserting the cache token counts are surfaced.
• Also includes Gemini usage parsing + tests for cached/thinking token counts (note: this expands scope beyond the PR title/description).

Reviewed changes

Copilot reviewed 4 out of 4 changed files in this pull request and generated 2 comments.

File	Description
python/packages/bedrock/agent_framework_bedrock/_chat_client.py	Map Bedrock Converse cache token fields into canonical UsageDetails keys.
python/packages/bedrock/tests/test_bedrock_client.py	Add unit test covering Bedrock cache token parsing.
python/packages/gemini/agent_framework_gemini/_chat_client.py	Add parsing of Gemini cached/thinking token counts into canonical usage fields.
python/packages/gemini/tests/test_gemini_client.py	Extend test helpers and add a test for Gemini cached/reasoning usage fields.

[eavanvalkenburg]

Please fix the comment from copilot

[github-actions]

Python Test Coverage Report •

File	Stmts	Miss	Cover	Missing
packages/bedrock/agent_framework_bedrock
_chat_client.py	449	98	78%	304–305, 321–330, 336, 404, 413, 424, 426, 428, 433, 452–453, 477, 490, 502, 505, 513–514, 517–518, 520–521, 526–528, 530, 540–541, 563, 570, 579–580, 582–583, 585–587, 589, 591–592, 598–600, 603–604, 610–613, 619–629, 632, 651, 656, 706–707, 720, 746, 758, 763, 791, 795–796, 799, 817, 841, 853, 857, 871, 879–880, 884, 886–893
packages/gemini/agent_framework_gemini
_chat_client.py	380	3	99%	398, 821, 832
TOTAL	42161	4981	88%

Python Unit Test Overview

Tests	Skipped	Failures	Errors	Time
8306	37 💤	0 ❌	0 🔥	2m 10s ⏱️

[eavanvalkenburg]

@he-yufeng there are still comments form copilot that need to be addressed or closed with a reason

[he-yufeng]

@eavanvalkenburg both Copilot comments are addressed:

1. Empty-mapping return (Bedrock _parse_usage): fixed in ab5bdf5, it now returns details or None like the Gemini connector, with a regression test.
2. Scope (the PR also touches Gemini): intentional, both connectors had the same gap. I've updated the title and description to cover Bedrock and Gemini together since it's the same _parse_usage mapping fix. I can split it into two PRs if you'd rather keep one connector per PR.

Both test suites pass locally (Bedrock 9 passed, Gemini green) and ruff is clean on the changed files.