[erauner12]

Closes #263.

Summary

ADK judges need to know which tools an agent had available, not just which ones it called.

This PR passes captured tool definitions into Invocation.app_details, including descriptions and parameter schemas, so judges have the context needed to evaluate tool use.

It also:

• Uses upstream's source-owned tool events.
• Groups declarations by agent and skips non-function tools.
• Falls back to observed tool names when definitions aren't available.
• Keeps exported eval-set JSON unchanged.

Testing

Added regression coverage using OTel GenAI semconv fixtures, including rubric_based_final_response_quality_v1 with tool calls.

Validation after rebase: 2,059 model-free tests passed on Python 3.13.15, including 29 integration tests; 14 skipped. E2E tests excluded. Ruff, lockfile, and diff checks passed. No live model calls.

Base

Rebased directly onto upstream main at 890dc65, which includes merged #267. The identity repair is inherited from upstream; this PR now adds only captured tool declarations and their tests. Producers use canonical GenAI semconv; no legacy producer fixture or private extraction override is added.

AI assistance: non-trivial code and test assistance.

[krisztianfekete]

Thanks, couple of small things:

1. rubric_based_final_response_quality_v1 now errors on every ADK trace with a tool call (see inline). CI misses it because no test runs that metric on a converted trace.
2. We're dropping legacy formats in favour of OTel GenAI semconv, which is also where ADK telemetry v2 is going. Please remove the parts built on OpenLLMetry gen_ai.prompt.N, ADK's deprecated call_llm span and the Gemini Schema format. Please also use the semconv fixtures for tests.

[erauner12]

Rebased this PR onto #267. Removed the legacy compatibility branch and dedicated tests, converted only type == "function" entries, dropped the OBJECT case, and switched the judge fixture to v2proj_span_only/S2.json, including a rubric_based_final_response_quality_v1 run asserting result.error is None.

@krisztianfekete Would you accept a minimal synthetic kagent 0.10 compatibility fixture, or should semconv be the only supported path?

[krisztianfekete]

@krisztianfekete Would you accept a minimal synthetic kagent 0.10 compatibility fixture, or should semconv be the only supported path?

Optimally no, semconv is the only supported path. Older producers should upgrade or normalize to OTel GenAI before export. Would that work with you?

[krisztianfekete]

Let's rebase it to main, then it LGTM!