fix(sandbox): keep programmatic tool calling items in memory rollouts

src/agents/sandbox/memory/rollouts.py

@@ -38,6 +38,8 @@
         "mcp_approval_request",
         "mcp_approval_response",
         "mcp_call",
+        "program",
+        "program_output",
         "shell_call",
         "shell_call_output",
         "tool_search_call",

tests/sandbox/test_memory.py

@@ -11,6 +11,8 @@
 
 import pytest
 from openai.types.responses import ResponseCustomToolCall, ResponseFunctionToolCall
+from openai.types.responses.response_function_tool_call import CallerProgram
+from openai.types.responses.response_output_item import Program, ProgramOutput
 from openai.types.responses.response_output_message import ResponseOutputMessage
 from openai.types.responses.response_reasoning_item import ResponseReasoningItem
 
@@ -33,6 +35,8 @@
     CompactionItem,
     MessageOutputItem,
     ToolApprovalItem,
+    ToolCallItem,
+    ToolCallOutputItem,
     TResponseOutputItem,
 )
 from agents.result import RunResult, RunResultStreaming
@@ -279,6 +283,120 @@ def test_build_rollout_payload_filters_developer_and_noisy_items() -> None:
     assert payload["final_output"] == "done"
 
 
+def test_build_rollout_payload_keeps_programmatic_tool_calling_items() -> None:
+    agent = Agent(name="test")
+    program = Program(
+        id="program_item",
+        call_id="call_prog_1",
+        code='lookup_inventory(sku="A-1")',
+        fingerprint="fingerprint",
+        type="program",
+    )
+    function_call = ResponseFunctionToolCall(
+        id="function_item",
+        call_id="call_fn_1",
+        name="lookup_inventory",
+        arguments='{"sku":"A-1"}',
+        caller=CallerProgram(type="program", caller_id="call_prog_1"),
+        type="function_call",
+    )
+    function_call_output = cast(
+        TResponseInputItem,
+        {
+            "type": "function_call_output",
+            "call_id": "call_fn_1",
+            "output": '{"available_units":42}',
+        },
+    )
+    program_output = ProgramOutput(
+        id="program_output_item",
+        call_id="call_prog_1",
+        result='{"sku":"A-1","available_units":42}',
+        status="completed",
+        type="program_output",
+    )
+
+    payload = build_rollout_payload(
+        input="what is in stock?",
+        new_items=[
+            ToolCallItem(agent=agent, raw_item=program),
+            ToolCallItem(agent=agent, raw_item=function_call),
+            ToolCallOutputItem(agent=agent, raw_item=function_call_output, output="42"),
+            ToolCallOutputItem(agent=agent, raw_item=program_output, output="42"),
+        ],
+        final_output="done",
+        interruptions=[],
+        terminal_metadata=RolloutTerminalMetadata(
+            terminal_state="completed",
+            has_final_output=True,
+        ),
+    )
+
+    generated_items = payload["generated_items"]
+    assert [item["type"] for item in generated_items] == [
+        "program",
+        "function_call",
+        "function_call_output",
+        "program_output",
+    ]
+    # The retained function call points back at the program that issued it, so the program
+    # it names has to survive alongside it.
+    assert generated_items[1]["caller"] == {"type": "program", "caller_id": "call_prog_1"}
+    assert generated_items[0]["call_id"] == "call_prog_1"
+    assert generated_items[0]["code"] == 'lookup_inventory(sku="A-1")'
+    assert generated_items[3]["call_id"] == "call_prog_1"
+    assert generated_items[3]["result"] == '{"sku":"A-1","available_units":42}'
+
+
+def test_build_rollout_payload_keeps_program_items_from_input() -> None:
+    payload = build_rollout_payload(
+        input=[
+            cast(
+                TResponseInputItem,
+                {
+                    "type": "program",
+                    "call_id": "call_prog_1",
+                    "code": 'lookup_inventory(sku="A-1")',
+                    "fingerprint": "fingerprint",
+                },
+            ),
+            cast(
+                TResponseInputItem,
+                {
+                    "type": "program_output",
+                    "call_id": "call_prog_1",
+                    "result": '{"available_units":42}',
+                    "status": "completed",
+                },
+            ),
+        ],
+        new_items=[],
+        final_output=None,
+        interruptions=[],
+        terminal_metadata=RolloutTerminalMetadata(terminal_state="completed"),
+    )
+
+    assert [item["type"] for item in payload["input"]] == ["program", "program_output"]
+
+
+def test_build_rollout_payload_still_drops_hosted_items_outside_the_included_set() -> None:
+    """Program items are included because every other call/output pair is; hosted tool calls
+    with no output half stay out."""
+    payload = build_rollout_payload(
+        input=[
+            cast(TResponseInputItem, {"type": "file_search_call", "id": "fs_1", "queries": []}),
+            cast(TResponseInputItem, {"type": "image_generation_call", "id": "ig_1"}),
+            cast(TResponseInputItem, {"type": "program", "call_id": "call_prog_1", "code": "x()"}),
+        ],
+        new_items=[],
+        final_output=None,
+        interruptions=[],
+        terminal_metadata=RolloutTerminalMetadata(terminal_state="completed"),
+    )
+
+    assert [item["type"] for item in payload["input"]] == ["program"]
+
+
 def test_build_rollout_payload_serializes_model_interruptions_as_dicts() -> None:
     agent = Agent(name="test")
     raw = ResponseFunctionToolCall(