fix(openai): trace computer use and access programs (#876)

## Summary

- Trace Agents computer-use, browser-authentication, and approval-result
items in tool spans.
- Convert screenshot data URLs from computer-use outputs into Braintrust
attachments.
- Capture Responses `access_programs` in request and response metadata,
with defensive reads for older OpenAI pins.

Fixes #869

## Live traces

- [Computer-use span with screenshot
attachment](https://www.braintrust.dev/app/Braintrust%20SDKs/p/openai-computer-use-access-programs-live-20261005/logs?r=8fa3bd730d96b150c3dcecaa729331d5&s=1befb8072daa3a4b&tvt=trace)
- [Responses `access_programs`
trace](https://www.braintrust.dev/app/Braintrust%20SDKs/p/openai-computer-use-access-programs-live-20261005/logs?r=bf79e4b24db593bd142344dff5f63208&s=ec148b3661e85412&tvt=trace)

Author: AbhiPrasad

py/src/braintrust/integrations/openai/cassettes/latest/test_openai_responses_access_programs.yaml

@@ -93,6 +93,15 @@ def _supports_agents_api() -> bool:
     return True
 
 
+def _supports_responses_access_programs() -> bool:
+    try:
+        from openai.resources.responses import Responses
+
+        return "access_programs" in inspect.signature(Responses.create).parameters
+    except (ImportError, AttributeError, ValueError):
+        return False
+
+
 @pytest.mark.vcr
 def test_openai_chat_metrics(memory_logger):
     assert not memory_logger.pop()
@@ -189,6 +198,11 @@ def test_openai_responses_metrics(memory_logger):
     assert TEST_MODEL in span["metadata"]["model"]
     assert span["metadata"]["provider"] == "openai"
     assert span["metadata"]["instructions"] == "Just the number please"
+    if hasattr(response, "access_programs"):
+        # Current OpenAI cassettes include the field with a null value. Older
+        # provider pins do not expose it; neither shape should add null metadata.
+        assert response.access_programs is None
+        assert "access_programs" not in span["metadata"]
     assert TEST_PROMPT in str(span["input"])
     assert len(span["output"]) > 0
     span_output_text = span["output"][0]["content"][0]["text"]
@@ -239,6 +253,25 @@ class NumberAnswer(BaseModel):
     assert span["output"][0]["content"][0]["parsed"]["reasoning"] == parse_response.output_parsed.reasoning
 
 
+@pytest.mark.vcr
+def test_openai_responses_access_programs(memory_logger):
+    if not _supports_responses_access_programs():
+        pytest.skip("Responses access_programs is not available in this SDK version")
+
+    access_programs = {"cyber": "standard"}
+    client = wrap_openai(openai.OpenAI())
+    response = client.responses.create(
+        model=TEST_MODEL,
+        input="Say hello in one word.",
+        access_programs=access_programs,
+    )
+
+    spans = memory_logger.pop()
+    assert len(spans) == 1
+    assert _try_to_dict(response.access_programs) == access_programs
+    assert _try_to_dict(spans[0]["metadata"]["access_programs"]) == access_programs
+
+
 @pytest.mark.vcr
 def test_openai_chat_completion_inline_moderation_metadata(memory_logger):
     if not _supports_inline_moderation():

py/src/braintrust/integrations/openai/test_openai.py

@@ -105,6 +105,7 @@ def start_span(*args, **kwargs):
     "conversation",
     "verbosity",
     "moderation",
+    "access_programs",
 )
 
 _RESPONSES_RESULT_METADATA_KEYS = (
@@ -114,6 +115,7 @@ def start_span(*args, **kwargs):
     "created_at",
     "moderation",
     "service_tier",
+    "access_programs",
     "object",
     "background",
     "incomplete_details",
@@ -1183,6 +1185,10 @@ def _log_response_tool_spans(output: Any, *, parent_export: str | None) -> None:
     "command_execution": ("command", "cwd"),
     "mcp_call": ("arguments",),
     "web_search_call": ("action",),
+    "computer_use_call": ("title",),
+    "browser_authentication_request": ("request_id",),
+    "computer_use_approval_request": ("request_id", "request"),
+    "computer_use_approval_request_result": ("request_id",),
     "create_subagent_call": ("content", "model", "reasoning_effort"),
     "send_subagent_input_call": ("content", "recipient_agent_id"),
     "resume_subagent_call": ("recipient_agent_id",),
@@ -1194,6 +1200,7 @@ def _log_response_tool_spans(output: Any, *, parent_export: str | None) -> None:
 _AGENT_TOOL_ITEM_OUTPUT_KEYS = {
     "command_execution": ("output", "exit_code", "duration_ms"),
     "mcp_call": ("output",),
+    "computer_use_approval_request_result": ("response",),
 }
 
 _AGENT_TURN_EVENTS = {
@@ -1272,7 +1279,13 @@ def _agent_tool_span_name(item: Any) -> str:
 
 
 def _agent_tool_span_data(item: Any, keys: tuple[str, ...]) -> Any:
-    values = clean_nones({key: getattr(item, key, None) for key in keys})
+    values = clean_nones(
+        {
+            key: _try_to_dict(value) if key in {"request", "response"} else value
+            for key in keys
+            if (value := getattr(item, key, None)) is not None
+        }
+    )
     if not values:
         return None
     if keys == ("arguments",):
@@ -1286,7 +1299,7 @@ def _agent_tool_span_metadata(item: Any) -> dict[str, Any]:
             "tool_type": item.type,
             "tool_id": item.id,
             "call_id": getattr(item, "call_id", None),
-            "status": item.status,
+            "status": getattr(item, "status", None),
             "turn_id": item.turn_id,
             "server_label": getattr(item, "server_label", None),
             "agent_id": getattr(item, "agent_id", None),
@@ -1299,7 +1312,7 @@ def _agent_tool_span_error(item: Any) -> Any:
     error = getattr(item, "error", None)
     if error is not None:
         return error
-    if item.status == "failed":
+    if getattr(item, "status", None) == "failed":
         return "Agent tool call failed"
     return None
 
@@ -1383,7 +1396,7 @@ def _observe_completed_item(self, item: Any) -> None:
             error = _agent_tool_span_error(item)
             if error is not None:
                 tool_span.log(error=error)
-            output = _agent_tool_span_data(item, _AGENT_TOOL_ITEM_OUTPUT_KEYS.get(item.type, ()))
+            output = _agent_tool_span_output(item)
             if output is not None:
                 tool_span.log(output=output)
 
@@ -1412,6 +1425,27 @@ def finish(self) -> None:
         self.span.end()
 
 
+def _agent_tool_span_output(item: Any) -> Any:
+    if item.type == "computer_use_call":
+        output = getattr(item, "output", None)
+        screenshot = getattr(output, "image_url", None)
+        if screenshot is None and isinstance(output, Mapping):
+            screenshot = output.get("image_url")
+        if screenshot is not None:
+            resolved = _materialize_attachment(screenshot, prefix="computer_use_screenshot")
+            if resolved is not None:
+                output_data = _try_to_dict(output)
+                if isinstance(output_data, dict):
+                    return {
+                        **output_data,
+                        "image_url": resolved.multimodal_part_payload["image_url"],
+                    }
+                return resolved.multimodal_part_payload
+        return output
+
+    return _agent_tool_span_data(item, _AGENT_TOOL_ITEM_OUTPUT_KEYS.get(item.type, ()))
+
+
 class AgentSessionWrapper:
     def __init__(self, create_fn: Callable[..., Any] | None, acreate_fn: Callable[..., Any] | None) -> None:
         self.create_fn = create_fn