rubric_based_final_response_quality_v1 errors when the agent replies from a before_agent_callback

[MihailRussu]

🔴 Required Information

Describe the Bug:
Since 2.10, an eval case whose agent answers from a before_agent_callback (so no model call) can no longer be judged with rubric_based_final_response_quality_v1. The metric fails with `my_agent` not found in the agentic system. and the case comes back NOT_EVALUATED, although the agent's reply is the same as on 2.9.

Steps to Reproduce:

1. Define an LlmAgent whose before_agent_callback returns a types.Content, e.g. "Authentication required."
2. Add an eval case for it with any rubric under rubric_based_final_response_quality_v1.
3. Run it with adk eval (or LocalEvalService).

Expected Behavior:
The rubric is judged against the reply, as on 2.9.0.

Observed Behavior:
local_eval_service logs Metric evaluation failed for metric rubric_based_final_response_quality_v1 ... with following error my_agent not found in the agentic system. and the case is NOT_EVALUATED on every run.

Environment Details:

• ADK Library Version: 2.11.0 (also 2.10.0; 2.9.0 works)
• Desktop OS: Linux
• Python Version: 3.13

Model Information:

• Are you using LiteLLM: No
• Which model is being used: the agent makes no model call; the judge is a Gemini model

---

🟡 Optional Information

Regression:
Yes, works on 2.9.0.

Additional Context:
As far as I can tell this comes from 2d428ea (informational efficiency metrics), which stopped dropping the final event from invocation_events (it is now kept with content=None to carry usage metadata). rubric_based_final_response_quality_v1 takes invocation_events[0].author as the agent name and calls AppDetails.get_developer_instructions(), which raises because app details are only recorded from intercepted model requests, and a callback short-circuit never makes one. On 2.9 the list was empty, so the judge fell back to next(iter(app_details.agent_details)), found nothing and went on without developer instructions.

Falling back to empty developer instructions when the author is not in agent_details (or only taking the author when it is) would restore the 2.9 behaviour. hallucinations_v1 and the multi-turn trajectory evaluator look fine, since they iterate the recorded agents or only use the name as a label.

Minimal Reproduction Code:

from google.adk.agents import Agent
from google.genai import types


def require_token(callback_context):
    if not callback_context.state.get('access_token'):
        return types.Content(role='model', parts=[types.Part(text='Authentication required.')])
    return None


root_agent = Agent(
    name='my_agent',
    model='gemini-2.5-flash',
    instruction='Answer questions about the user account.',
    before_agent_callback=require_token,
)

with an eval case whose session state has no access_token and this criterion:

{"criteria": {"rubric_based_final_response_quality_v1": {"threshold": 0.6, "rubrics": [
  {"rubric_id": "refuses", "rubric_content": {"text_property": "The agent asks the user to authenticate."}}]}}}

How often has this issue occurred?:

• Always (100%)

[surajksharma07]

@MihailRussu Reproduced on 2.11.0 through LocalEvalService with your agent: the reply comes back fine, agent_details is empty and the rubric metric fails with "my_agent not found" -> NOT_EVALUATED.

Until a fix lands a workaround worth trying: wrap AppDetails.get_developer_instructions so it returns "" when the agent isn't in agent_details (that's what 2.9 effectively did). With a stubbed judge the case gets judged and passes on my side. Pinning 2.9.2 for eval runs also avoids it. Could you check whether that works with your real Gemini judge? Keep in mind the patch applies to every metric that calls it not just the rubric one.

@1aifanatic #7380 adds the same guard inside the rubric evaluator which is the right place for it. Before review please also run it once against a live judge since your smoke test stubbed that part.