🔴 Required Information

Describe the Bug:

When using the optimizer with conversation_scenario (user simulation) eval cases, local_eval_sampler.py crashes with TypeError at line 292 because eval_metric_result.score can be None for some per-invocation results.

The initial baseline evaluation completes successfully, but during the GEPA reflection loop's internal re-evaluation, some per-invocation metric results return None scores. The code calls round(eval_metric_result.score, 2) without a None guard.

Steps to Reproduce:

1. Create an eval set using conversation_scenario (user simulation)
2. Run the optimizer (e.g. AgentOptimizer.optimize())
3. The first baseline eval passes, but iteration 1's re-evaluation crashes

Expected Behavior:

The optimizer should handle None scores gracefully (e.g. treat as 0.0 or skip) and continue the GEPA loop.

Observed Behavior:

File ".../google/adk/optimization/local_eval_sampler.py", line 292, in _format_eval_result
    "score": round(eval_metric_result.score, 2),  # accurate enough
TypeError: type NoneType doesn't define __round__ method

Environment Details:

• ADK Library Version: 1.28.0
• Desktop OS: macOS
• Python Version: 3.13

Model Information:

• Are you using LiteLLM: No

🟡 Optional Information

Root Cause Analysis:

local_eval_sampler.py:292:

eval_metric_results.append({
    "metric_name": eval_metric_result.metric_name,
    "score": round(eval_metric_result.score, 2),  # ← crashes when score is None
    "eval_status": eval_metric_result.eval_status.name,
})

Suggested Fix:

"score": round(eval_metric_result.score, 2) if eval_metric_result.score is not None else 0.0,

Related Issues:

• _EvalMetricResultWithInvocation fails with conversation_scenario (user simulation) - expected_invocation is None #4283 — same root cause (conversation_scenario producing None values) but manifesting in agent_evaluator.py

How often has this issue occurred?:

Always (100%) when using conversation_scenario with the optimizer.