Proposal: deterministic tool-call expectations for conversation_scenario eval cases #8077

🔴 Required Information

Is your feature request related to a specific problem?

An eval case with a conversation_scenario (simulated user) has no deterministic check on tool use. tool_trajectory_avg_score and tool_parameter_match both need expected_invocations, which a simulated conversation does not have. They raise expected_invocations is needed by this metric., LocalEvalService records the metric as NOT_EVALUATED, and NOT_EVALUATED does not affect the case's status.

So when the agent calls a wrong or destructive tool in a simulated conversation, only a judge-backed metric can notice. The snippet below runs offline on main at ab63844 (2.14.0, Python 3.12.7). The agent calls delete_account in a "cancel my order" scenario:

final_eval_status: NOT_EVALUATED
  tool_trajectory_avg_score: NOT_EVALUATED
  tool_parameter_match: NOT_EVALUATED

Raising without expected_invocations is correct for those two metrics. This request is for a metric that does not need them.

Describe the Solution You'd Like

A built-in, deterministic metric that checks the tool calls of a whole conversation against expectations stored on the scenario (names open to discussion):

• required: calls that must appear somewhere in the conversation.
• forbidden: calls that must not appear.
• Arguments are optional. When given, they are compared with a ToolArgsMatcher, as in tool_parameter_match.

Score: 0.0 if any forbidden call appears, otherwise the fraction of required calls that were matched (each actual call matches at most one). A scenario without expectations is NOT_EVALUATED.

Impact on your work

I found this gap while reading the evaluation code for other contributions, not from a production need, so there is no deadline. A check like "the agent must call cancel_order for this order and must never call delete_account" is one a CI gate should be able to make without depending on a judge model's verdict or cost.

Willingness to contribute

Yes. I would like to agree on the design here first.

🟡 Recommended Information

Describe Alternatives You've Considered

• rubric_based_tool_use_quality_v1 / multi_turn_tool_use_quality_v1: they work without a reference, but need a judge model, so the verdict is not reproducible and costs model calls.
• A custom metric (custom_metrics): it receives the conversation_scenario, but there is nowhere to put per-case expectations. ConversationScenario forbids extra fields, and the evaluator does not receive the EvalCase.
• Expectations on the criterion in EvalConfig: they would apply to every case, while required calls differ per case.

Proposed API / Implementation

class ToolCallExpectation(EvalBaseModel):
  name: str
  args: dict[str, Any] | None = None  # None: any arguments


class ToolCallExpectations(EvalBaseModel):
  required: list[ToolCallExpectation] = []
  forbidden: list[ToolCallExpectation] = []


class ConversationScenario(EvalBaseModel):
  ...
  tool_call_expectations: ToolCallExpectations | None = None


class ToolCallExpectationsCriterion(BaseCriterion):
  args_matcher: ToolArgsMatcher | None = None
  in_order: bool = False  # required calls must appear in the listed order

• Putting the expectations on ConversationScenario needs no change to Evaluator.evaluate_invocations, which already receives the scenario.
• Result shape as recommended in Support conversation-level evaluators that don't require per-invocation results #4533: the score goes on the last invocation, earlier ones are NOT_EVALUATED.

Questions for maintainers:

1. Is a built-in metric like this welcome, or is tool use in simulated conversations meant to stay judge-only?
2. Should the expectations live on ConversationScenario or on EvalCase (which would also cover static conversations, but needs the evaluator to receive them)?
3. Is "any forbidden call scores 0.0" the right rule, or should forbidden calls be a separate metric?

Additional Context

Reproduction (no API key needed):

import asyncio
import logging

from google.adk.agents.llm_agent import LlmAgent
from google.adk.evaluation import base_eval_service as svc
from google.adk.evaluation import eval_case as ec
from google.adk.evaluation.conversation_scenarios import ConversationScenario
from google.adk.evaluation.eval_metrics import EvalMetric
from google.adk.evaluation.in_memory_eval_sets_manager import InMemoryEvalSetsManager
from google.adk.evaluation.local_eval_service import LocalEvalService
from google.genai import types

logging.disable(logging.CRITICAL)


async def main():
  manager = InMemoryEvalSetsManager()
  manager.create_eval_set("app", "set")
  scenario = ConversationScenario(
      starting_prompt="Cancel my order.",
      conversation_plan="Ask the agent to cancel order 1.",
  )
  manager.add_eval_case(
      "app", "set", ec.EvalCase(eval_id="c1", conversation_scenario=scenario)
  )
  # What a simulated conversation could produce: a wrong, destructive call.
  call = types.Part.from_function_call(name="delete_account", args={})
  actual = ec.Invocation(
      user_content=types.UserContent("Cancel my order."),
      final_response=types.ModelContent("Done."),
      intermediate_data=ec.InvocationEvents(
          invocation_events=[
              ec.InvocationEvent(
                  author="agent", content=types.ModelContent(parts=[call])
              )
          ]
      ),
  )
  service = LocalEvalService(
      root_agent=LlmAgent(name="agent", model="fake"),
      eval_sets_manager=manager,
  )
  request = svc.EvaluateRequest(
      inference_results=[
          svc.InferenceResult(
              app_name="app",
              eval_set_id="set",
              eval_case_id="c1",
              inferences=[actual],
              session_id="s",
              status=svc.InferenceStatus.SUCCESS,
          )
      ],
      evaluate_config=svc.EvaluateConfig(
          eval_metrics=[
              EvalMetric(metric_name="tool_trajectory_avg_score", threshold=1.0),
              EvalMetric(metric_name="tool_parameter_match", threshold=1.0),
          ]
      ),
  )
  async for result in service.evaluate(request):
    print("final_eval_status:", result.final_eval_status.name)
    for metric in result.overall_eval_metric_results:
      print(f"  {metric.metric_name}: {metric.eval_status.name}")


asyncio.run(main())