Retry ModelBehaviorError so a truncated stream no longer kills the run

engine/agents/llm_retry.py

@@ -7,6 +7,7 @@
 from typing import TypeVar
 
 import httpx
+from agents import ModelBehaviorError
 from openai import (
     APIConnectionError,
     APIError,
@@ -54,7 +55,16 @@ def is_retriable_llm_error(exc: BaseException) -> bool:
         "The model produced invalid content")
       - 400s except the terminal codes in ``_TERMINAL_400_CODES`` — a clean
         rerun from local history fixes stale server-side state (INF-3308)
+      - the Agents SDK's ``ModelBehaviorError`` ("Model did not produce a
+        final response!" when a stream truncates without a terminal event,
+        malformed tool-call JSON, calls to nonexistent tools): every variant
+        is nondeterministic sampled-output misbehavior, so a rerun from local
+        history can produce a well-formed turn. Other ``AgentsException``
+        subclasses stay non-retriable — ``MaxTurnsExceeded`` must terminate
+        the run and ``UserError`` / guardrail tripwires are deterministic.
     """
+    if isinstance(exc, ModelBehaviorError):
+        return True
     if isinstance(exc, (APIConnectionError, APITimeoutError, RateLimitError)):
         return True
     if isinstance(exc, APIStatusError):

engine/agents/openai_agent_runner.py

@@ -78,7 +78,10 @@ async def run(
 
         Mid-stream failures (events already processed when the stream dies —
         dropped connections, incomplete chunked reads, stale ``rs_*`` 400s from
-        the SDK re-sending Responses-API reasoning items) are recovered by
+        the SDK re-sending Responses-API reasoning items, streams that end
+        without a terminal event and surface as the SDK's
+        ``ModelBehaviorError: Model did not produce a final response!``) are
+        recovered by
         rerunning from the LOCAL conversation history (INF-3504 / INF-3308):
         completed turns stay in ``agent_context``; only a trailing incomplete
         tool turn is trimmed (see ``AgentContext.trim_incomplete_tool_turn``)

tests/unit/agents/test_llm_retry.py

@@ -2,6 +2,7 @@
 
 import httpx
 import pytest
+from agents import MaxTurnsExceeded, ModelBehaviorError, UserError
 from openai import (
     APIConnectionError,
     APIError,
@@ -71,6 +72,30 @@ def test_non_400_4xx_is_not_retriable() -> None:
     assert not is_retriable_llm_error(_status_error(404, "not found"))
 
 
+@pytest.mark.parametrize(
+    "message",
+    [
+        # Stream truncated without a terminal event (agents run_loop).
+        "Model did not produce a final response!",
+        # Other nondeterministic model-output misbehavior the SDK wraps.
+        'Invalid JSON input for tool query_traces: {"q": ',
+        "Tool nonexistent_tool not found in agent halo",
+    ],
+)
+def test_model_behavior_error_is_retriable(message: str) -> None:
+    # A rerun from local history re-samples the turn, so every
+    # ``ModelBehaviorError`` variant is worth retrying (bounded by the
+    # runner circuit breaker).
+    assert is_retriable_llm_error(ModelBehaviorError(message))
+
+
+def test_other_agents_exceptions_are_not_retriable() -> None:
+    # ``MaxTurnsExceeded`` must terminate the run; ``UserError`` is a
+    # deterministic configuration bug. Only ``ModelBehaviorError`` retries.
+    assert not is_retriable_llm_error(MaxTurnsExceeded("Max turns (30) exceeded"))
+    assert not is_retriable_llm_error(UserError("bad SDK usage"))
+
+
 def test_unrelated_exception_is_not_retriable() -> None:
     assert not is_retriable_llm_error(ValueError("nope"))
 

tests/unit/agents/test_openai_agent_runner.py

@@ -2,6 +2,7 @@
 
 import httpx
 import pytest
+from agents import ModelBehaviorError
 from openai import APIConnectionError, APIError, AsyncOpenAI, BadRequestError
 
 from engine.agents.agent_context import AgentContext
@@ -748,6 +749,54 @@ async def stream_hits_stale_reasoning_item(*, agent, input, context):
     assert [item.content for item in ctx.items] == ["worked a bit", "answer"]
 
 
+@pytest.mark.asyncio
+async def test_runner_retries_model_behavior_error_when_stream_has_no_final_response() -> None:
+    """The SDK raises ``ModelBehaviorError: Model did not produce a final
+    response!`` when a stream truncates without a terminal event. The runner
+    reruns from local history instead of failing the run — previously this
+    escaped as non-retriable and killed the whole run on the root turn."""
+    bus = EngineOutputBus()
+    ctx = _context()
+    execution = AgentExecution(
+        agent_id="root",
+        agent_name="root",
+        depth=0,
+        parent_agent_id=None,
+        parent_tool_call_id=None,
+    )
+
+    calls: list[list[dict]] = []
+
+    async def stream_truncates_without_final_response(*, agent, input, context):
+        calls.append(input)
+        if len(calls) == 1:
+            return _StreamYieldsThenRaises(
+                [_assistant_event("partial answer")],
+                ModelBehaviorError("Model did not produce a final response!"),
+            )
+        return _FakeStream([_assistant_event("answer\n<final/>")])
+
+    runner = OpenAiAgentRunner(
+        run_streamed=stream_truncates_without_final_response,
+        client=_DUMMY_CLIENT,
+        retry_backoff_base=0.0,
+    )
+
+    await runner.run(
+        sdk_agent=object(),
+        agent_context=ctx,
+        agent_execution=execution,
+        output_bus=bus,
+        is_root=True,
+    )
+
+    assert len(calls) == 2
+    # The retry replays the completed assistant message from local history.
+    assert calls[1] == [{"role": "assistant", "content": "partial answer"}]
+    assert [item.content for item in ctx.items] == ["partial answer", "answer"]
+    assert execution.consecutive_llm_failures == 0
+
+
 @pytest.mark.asyncio
 async def test_runner_propagates_terminal_400_mid_stream() -> None:
     """Terminal-code 400s mid-stream stay non-retriable — no clean rerun fixes