fix(verifiers): judge with the configured judge model in a rubrics verifier built in code

src/agent_env/task_step/task_steps/verifiers/rubrics_verifier.py

@@ -434,7 +434,7 @@ def __init__(
         version: Optional[int],
         criteria: list[dict],
         prompt_id: str,
-        default_model: str = "claude-sonnet-4-6",
+        default_model: Optional[str] = None,
         default_model_api_base: Optional[str] = None,
         score_aggregator: Optional[ScoreAggregator] = None,
         verifier_id: Optional[str] = None,
@@ -489,7 +489,9 @@ def __init__(
         self.verifier_id = verifier_id or uuid.uuid4().hex
         self.criteria = criteria
         self.prompt_id = prompt_id
-        self.default_model = default_model
+        # A stored judge always holds a model: its author's, or the one the config names for judges, as for a step
+        # built in code, which the A2A validator's are.
+        self.default_model = default_model or get_config().get_model_for_role("judge") or "claude-sonnet-4-6"
         self.default_model_api_base = default_model_api_base
         if isinstance(score_aggregator, str):
             score_aggregator = ScoreAggregator(score_aggregator)
@@ -593,15 +595,13 @@ def to_dict(self) -> dict:
 
     @classmethod
     def from_dict(cls, data: dict) -> RubricsVerifierTaskStep:
-        from agent_env.config import get_config
-
         raw_agg = data.get("score_aggregator")
         raw_tf = data.get("trajectory_filter")
         return cls(
             **cls._base_from_dict(data),
             criteria=data["criteria"],
             prompt_id=data["prompt_id"],
-            default_model=data.get("default_model") or get_config().get_model_for_role("judge") or "claude-sonnet-4-6",
+            default_model=data.get("default_model"),
             default_model_api_base=data.get("default_model_api_base"),
             score_aggregator=ScoreAggregator(raw_agg) if raw_agg else None,
             verifier_id=data.get("verifier_id"),

tst/unit/task_step/verifiers/test_rubrics_verifier_judge_model.py

@@ -0,0 +1,59 @@
+"""A rubrics verifier built in code, as the A2A validator builds its own, judges with the configured ``judge`` model,
+as one read from a task does."""
+
+import pytest
+
+from agent_env.config import reset_config
+from agent_env.task_step.task_steps.verifiers.rubrics_verifier import RubricsVerifierTaskStep
+
+CRITERIA = [{"id": "c1", "description": "Response says hello"}]
+
+
+@pytest.fixture
+def model_config(tmp_path, monkeypatch):
+    """Install a config file holding ``model`` as its whole [model] table, so no developer config is read."""
+
+    def install(model: str) -> None:
+        config = tmp_path / "config.toml"
+        config.write_text(f"[model]\n{model}")
+        monkeypatch.setenv("AGENT_ENV_CONFIG", str(config))
+        reset_config()
+
+    yield install
+    reset_config()
+
+
+@pytest.fixture
+def judge_role(model_config):
+    model_config('default = "agent-model"\n\n[model.roles]\njudge = "judge-model"\n')
+
+
+def _built(**kwargs) -> RubricsVerifierTaskStep:
+    return RubricsVerifierTaskStep(id="verify", version=None, criteria=CRITERIA, prompt_id="p1", **kwargs)
+
+
+def test_a_verifier_built_in_code_judges_with_the_configured_judge_model(judge_role):
+    assert _built().default_model == "judge-model"
+
+
+def test_a_verifier_read_from_a_task_judges_with_the_configured_judge_model(judge_role):
+    data = _built().to_dict()
+    del data["default_model"]
+
+    assert RubricsVerifierTaskStep.from_dict(data).default_model == "judge-model"
+
+
+def test_a_model_the_step_names_wins_over_the_judge_role(judge_role):
+    assert _built(default_model="named-model").default_model == "named-model"
+
+
+def test_without_a_judge_role_the_judge_is_the_configured_default_model(model_config):
+    model_config('default = "agent-model"\n')
+
+    assert _built().default_model == "agent-model"
+
+
+def test_with_no_model_configured_the_judge_is_claude_sonnet(model_config):
+    model_config("")
+
+    assert _built().default_model == "claude-sonnet-4-6"

tst/unit/task_step/verifiers/test_rubrics_verifier_judge_trajectory.py

@@ -91,7 +91,8 @@ def issue_write_grant(self, object_url, *, media_type, max_bytes, expires_in=Non
 
 def _use_store(monkeypatch, store, *, key_prefix=""):
     config = type(
-        "Config", (), {"get_object_store": lambda self: store, "get_artifact_key_prefix": lambda self: key_prefix}
+        "Config", (), {"get_object_store": lambda self: store, "get_artifact_key_prefix": lambda self: key_prefix,
+                       "get_model_for_role": lambda self, role: None}
     )()
     monkeypatch.setattr(f"{RV}.get_config", lambda: config)
     monkeypatch.setattr(