From 390cbde78326a33451eb5d7cfcf885ecf6d44db8 Mon Sep 17 00:00:00 2001
From: "Eric A. Litman" <ericlitman@studio2.tail062eee.ts.net>
Date: Tue, 4 Aug 2026 03:05:26 -0400
Subject: [PATCH] fix: isolate adversarial stage sub-agents from spine
 persistence [refs OSWE-239]

Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
---
 agent/reviewer_adversarial.py               |  2 +
 tests/reviewer/test_reviewer_adversarial.py | 72 +++++++++++++++++++++
 2 files changed, 74 insertions(+)

diff --git a/agent/reviewer_adversarial.py b/agent/reviewer_adversarial.py
index 839014da1..84f335d1d 100644
--- a/agent/reviewer_adversarial.py
+++ b/agent/reviewer_adversarial.py
@@ -352,6 +352,8 @@ def _bounded_agent(
         system_prompt=prompt,
         tools=tools or [],
         backend=backend,
+        # Stage sub-agents must not join the spine's persistence.
+        checkpointer=False,
         response_format=ToolStrategy(response_format),
         middleware=cast(list[AgentMiddleware[Any, Any, Any]], _stage_middleware(middleware)),
     )
diff --git a/tests/reviewer/test_reviewer_adversarial.py b/tests/reviewer/test_reviewer_adversarial.py
index 3946ad55c..997b998d6 100644
--- a/tests/reviewer/test_reviewer_adversarial.py
+++ b/tests/reviewer/test_reviewer_adversarial.py
@@ -14,6 +14,7 @@
 from langchain_core.language_models.chat_models import BaseChatModel
 from langchain_core.messages import AIMessage
 from langchain_core.outputs import ChatGeneration, ChatResult
+from langgraph.constants import CONFIG_KEY_CHECKPOINTER, Send
 from langgraph.graph.state import RunnableConfig
 from langgraph.runtime import Runtime
 from pydantic import Field
@@ -1553,6 +1554,77 @@ async def test_bounded_finder_agent_calls_its_model_and_returns_structured_outpu
         assert "FinderOutput" in model.bound_tool_names, str(spec["name"])
 
 
+@pytest.mark.asyncio
+async def test_stage_subagent_ignores_spine_checkpointer_and_thread(
+    caplog: pytest.LogCaptureFixture,
+) -> None:
+    from deepagents.backends import StateBackend
+    from langgraph.checkpoint.memory import InMemorySaver
+    from langgraph.graph import END, START, StateGraph
+
+    from agent.review.adversarial import AdversarialState, FinderOutput
+
+    caplog.set_level(logging.WARNING)
+    specs = _finder_specs(cast(BaseChatModel, _StubToolCallModel(response_tool="FinderOutput")))
+    spec = next(spec for spec in specs if spec["name"] == "correctness")
+    model = cast(_StubToolCallModel, spec["model"])
+    finder_graph = _bounded_agent(
+        model=cast(BaseChatModel, model),
+        response_format=FinderOutput,
+        backend=StateBackend(),
+        tools=cast(list[Any], spec.get("tools", [])),
+        middleware=[
+            *cast(list[Any], spec.get("middleware", [])),
+            ExcludeToolsMiddleware(excluded=frozenset({"task"})),
+        ],
+    )
+    saver = InMemorySaver()
+    config = cast(
+        RunnableConfig,
+        {
+            "configurable": {
+                "thread_id": "adversarial-spine-replica",
+                "checkpoint_ns": "",
+                CONFIG_KEY_CHECKPOINTER: saver,
+            }
+        },
+    )
+    observed: dict[str, Any] = {}
+
+    async def prepare(state: AdversarialState) -> dict[str, Any]:
+        del state
+        return {"finders_expected": ["correctness", "security"]}
+
+    def fanout(state: dict[str, Any]) -> list[Send]:
+        return [Send("find", {"finder_name": name}) for name in state["finders_expected"]]
+
+    async def find(state: AdversarialState) -> dict[str, Any]:
+        data = cast(dict[str, Any], state)
+        name = data["finder_name"]
+        try:
+            observed[name] = await _run_stage(finder_graph, "review the diff", config)
+        except Exception as exc:
+            observed[name] = exc
+        return {"finder_results": [{"finder": name, "candidates": [], "error": None}]}
+
+    builder = StateGraph(AdversarialState)
+    builder.add_node("prepare", prepare)
+    builder.add_node("find", find)
+    builder.add_edge(START, "prepare")
+    builder.add_conditional_edges("prepare", fanout)
+    builder.add_edge("find", END)
+    spine = builder.compile()
+
+    await spine.ainvoke({}, config=config, durability="sync")
+
+    assert set(observed) == {"correctness", "security"}
+    assert all(isinstance(result, FinderOutput) for result in observed.values()), (
+        f"observed={observed!r}; logs={caplog.text}"
+    )
+    assert "Ignoring unknown node name find in pending sends" not in caplog.text
+    assert model.invocations
+
+
 @pytest.mark.asyncio
 async def test_run_stage_logs_diagnostics_when_no_structured_response(
     caplog: pytest.LogCaptureFixture,