Llm response token counts

[umaannamalai]

This PR includes changes to send token counts pulled from LLM response objects directly from the agent.

[github-actions]

✅MegaLinter analysis: Success

Descriptor	Linter	Files	Fixed	Errors	Warnings	Elapsed time
✅ ACTION	actionlint	8		0	0	1.05s
✅ MARKDOWN	markdownlint	7	0	0	0	1.45s
✅ PYTHON	ruff	1030	0	0	0	1.11s
✅ PYTHON	ruff-format	1030	0	0	0	0.41s
✅ YAML	prettier	19	0	0	0	1.66s
✅ YAML	v8r	19		0	0	6.29s
✅ YAML	yamllint	19		0	0	0.75s

See detailed reports in MegaLinter artifacts

Show us your support by starring ⭐ the repository

[codecov-commenter]

Codecov Report

❌ Patch coverage is 90.32258% with 18 lines in your changes missing coverage. Please review.
✅ Project coverage is 81.92%. Comparing base (7343f58) to head (35c85ad).

Files with missing lines	Patch %	Lines
newrelic/hooks/mlmodel_gemini.py	67.39%	7 Missing and 8 partials ⚠️
newrelic/hooks/external_botocore.py	97.32%	0 Missing and 3 partials ⚠️

Additional details and impacted files

@@            Coverage Diff             @@
##             main    #1682      +/-   ##
==========================================
+ Coverage   81.85%   81.92%   +0.07%     
==========================================
  Files         214      214              
  Lines       25683    25813     +130     
  Branches     4076     4090      +14     
==========================================
+ Hits        21022    21147     +125     
- Misses       3264     3265       +1     
- Partials     1397     1401       +4     

☔ View full report in Codecov by Sentry.
📢 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.

[JiwaniZakir]

In create_chat_completion_message_event, the all_token_counts parameter is added to the signature and checked for truthiness, but its value is never actually used — both the request message (line ~231) and response message (line ~274) unconditionally set "token_count": 0 whenever the flag is truthy. This looks like a bug where the actual per-message token count should be extracted from all_token_counts rather than hardcoded to zero. If all_token_counts is a dict or list keyed by message index, the lookup logic is missing entirely.

Additionally, extract_bedrock_titan_embedding_model_response sets only response.usage.total_tokens (equal to inputTextTokenCount), while the text model response handler sets all three — prompt_tokens, completion_tokens, and total_tokens. The inconsistency is reasonable for embeddings (no output), but it would be worth a comment clarifying the intentional omission of prompt_tokens and completion_tokens to avoid confusion for future maintainers.

In extract_bedrock_titan_text_model_streaming_response, the accumulation pattern (bedrock_attrs.get(..., 0) + value) assumes invocation metrics may appear across multiple chunks, but amazon-bedrock-invocationMetrics is typically only present in the final chunk — so for most chunks prompt_tokens and completion_tokens will be 0, meaning the accumulation adds unnecessary overhead without risk of double-counting. A guard checking whether the metrics key exists before updating would make the intent clearer.