Merge c2a5ce669c into 976835194d

strix/core/execution.py

@@ -11,7 +11,7 @@ from functools import cache
 from typing import TYPE_CHECKING, Any, cast
 
 from agents import RunConfig, Runner
-from agents.exceptions import AgentsException, MaxTurnsExceeded, UserError
+from agents.exceptions import AgentsException, MaxTurnsExceeded, ModelRefusalError, UserError
 from agents.sandbox.errors import ExecTransportError
 from openai import (
     APIConnectionError,
@@ -72,6 +72,20 @@ class ProviderRefusalError(AgentsException):
     """Raised when a provider returns a structured refusal instead of an exception."""
 
 
+def _refusal_text(exc: BaseException) -> str | None:
+    """The refusal text of a provider refusal, however the SDK surfaced it.
+
+    A refusal can reach here as this module's own `ProviderRefusalError`, built
+    from a structured `refusal` content item, or as the SDK's own
+    `ModelRefusalError`, which some providers (observed with Gemini through
+    LiteLLM) raise directly instead of returning refusal content. Both carry
+    the same information and deserve the same one-line, non-retryable outcome.
+    """
+    if isinstance(exc, ProviderRefusalError | ModelRefusalError):
+        return str(exc.refusal if isinstance(exc, ModelRefusalError) else exc)
+    return None
+
+
 def _structured_provider_refusal(result: Any) -> str | None:
     for item in getattr(result, "new_items", ()) or ():
         raw_item = getattr(item, "raw_item", None)
@@ -130,7 +144,9 @@ def _model_error_status_code(exc: BaseException) -> int | None:
 
 
 def _is_transient_model_error(exc: BaseException) -> bool:
-    if codex.is_content_guardrail_error(exc):
+    if codex.is_content_guardrail_error(exc) or isinstance(
+        exc, ProviderRefusalError | ModelRefusalError
+    ):
         return False
     if isinstance(
         exc, APITimeoutError | APIConnectionError | TimeoutError | ConnectionError | OSError
@@ -790,9 +806,9 @@ async def _run_cycle(  # noqa: PLR0912, PLR0915
                 continue
             if session is not None:
                 await _salvage_stream_to_session(session, pre_run_items, stream, agent_id)
-            if isinstance(exc, ProviderRefusalError):
-                logger.warning("agent %s refused by the model provider: %s", agent_id, exc)
-                await coordinator.set_status(agent_id, "failed", error=str(exc))
+            if (refusal := _refusal_text(exc)) is not None:
+                logger.warning("agent %s refused by the model provider: %s", agent_id, refusal)
+                await coordinator.set_status(agent_id, "failed", error=refusal)
                 await notify_parent_on_terminal(coordinator, agent_id, "failed")
                 return None
             if isinstance(exc, MaxTurnsExceeded):

tests/test_execution.py

@@ -9,7 +9,7 @@ from typing import Any, cast
 from unittest.mock import MagicMock
 
 import pytest
-from agents.exceptions import MaxTurnsExceeded
+from agents.exceptions import MaxTurnsExceeded, ModelRefusalError
 from agents.items import MessageOutputItem
 from agents.memory import SQLiteSession
 from agents.tool_context import ToolContext
@@ -29,6 +29,26 @@ from strix.tools.finish.tool import finish_scan
 _NO_STREAM_EVENTS: list[Any] = []
 
 
+class _SdkRefusalStream:
+    """A stream where the SDK itself raises ModelRefusalError, not a content item.
+
+    Some providers (observed with Gemini through LiteLLM) surface a content-
+    filter refusal this way instead of the structured `refusal` content item
+    that OpenAI-compatible providers use.
+    """
+
+    def __init__(self, refusal: str) -> None:
+        self.run_loop_exception: BaseException | None = ModelRefusalError(refusal)
+        self.new_items: list[Any] = []
+
+    async def stream_events(self) -> Any:
+        for event in _NO_STREAM_EVENTS:
+            yield event
+
+    def cancel(self, mode: str = "immediate") -> None:  # noqa: ARG002
+        return
+
+
 class _StructuredRefusalStream:
     def __init__(self, refusal: str) -> None:
         self.run_loop_exception: BaseException | None = None
@@ -898,6 +918,48 @@ async def test_structured_provider_refusal_fails_noninteractive_child(
     session.close()
 
 
+@pytest.mark.asyncio
+async def test_sdk_model_refusal_fails_cleanly_without_a_traceback(
+    monkeypatch: pytest.MonkeyPatch,
+    caplog: pytest.LogCaptureFixture,
+) -> None:
+    """A ModelRefusalError raised by the SDK gets the same clear, non-retried
+    outcome as this module's own ProviderRefusalError, instead of falling
+    through to the generic crash path with a full traceback.
+    """
+    refusal = "Response withheld by the provider's content filter."
+    stream = _SdkRefusalStream(refusal)
+    monkeypatch.setattr(
+        "strix.core.execution.Runner.run_streamed", lambda *_args, **_kwargs: stream
+    )
+    coordinator = AgentCoordinator()
+    await coordinator.register("root", "strix", parent_id=None)
+
+    with caplog.at_level("WARNING", logger="strix.core.execution"):
+        result = await execution._run_cycle(
+            MagicMock(),
+            coordinator,
+            "root",
+            input_data="task",
+            run_config=MagicMock(),
+            context={},
+            max_turns=5,
+            session=None,
+            interactive=True,
+            event_sink=None,
+            hooks=None,
+        )
+
+    assert result is None
+    assert coordinator.statuses["root"] == "failed"
+    assert coordinator.errors["root"] == refusal
+    assert any(
+        record.levelname == "WARNING" and refusal in record.getMessage()
+        for record in caplog.records
+    )
+    assert not any(record.exc_info for record in caplog.records)
+
+
 @pytest.mark.asyncio
 async def test_crashing_noninteractive_child_settles_and_wakes_its_parent(
     monkeypatch: pytest.MonkeyPatch,