fix(openai): persist refusal explanations in OpenAIAgent response

hud/agents/openai/agent.py

@@ -12,6 +12,7 @@
     ResponseIncludable,
     ResponseInputParam,
     ResponseInputTextParam,
+    ResponseOutputRefusal,
     ResponseOutputText,
     ToolParam,
 )
@@ -257,11 +258,17 @@ async def get_response(
         reasoning_chunks: list[str] = []
         citations: list[Citation] = []
         tool_calls: list[MCPToolCall] = []
+        refusal: str | None = None
 
         for item in response.output:
             match item.type:
                 case "message":
                     for content_block in item.content:
+                        if isinstance(content_block, ResponseOutputRefusal) or getattr(
+                            content_block, "type", None
+                        ) == "refusal":
+                            refusal = getattr(content_block, "refusal", None)
+                            continue
                         if not isinstance(content_block, ResponseOutputText):
                             continue
                         if content_block.text:
@@ -346,13 +353,17 @@ async def get_response(
         # The Responses API has no finish_reason; truncation surfaces as
         # incomplete_details.reason ("max_output_tokens" / "content_filter").
         incomplete = response.incomplete_details
+        finish_reason = incomplete.reason if incomplete is not None else None
+        if finish_reason == "content_filter" and not refusal:
+            refusal = "The request was rejected due to OpenAI content safety filters."
         return AgentStep(
             content="".join(text_chunks),
             reasoning="\n".join(reasoning_chunks) if reasoning_chunks else None,
             citations=citations,
             tool_calls=tool_calls,
             done=not tool_calls,
-            finish_reason=incomplete.reason if incomplete is not None else None,
+            finish_reason=finish_reason,
+            refusal=refusal,
             model=response.model,
             usage=usage,
         )

hud/agents/tests/test_openai_agent.py

@@ -218,3 +218,42 @@ async def test_get_response_rejects_empty_shell_call() -> None:
         await agent.get_response(state)
 
     assert state.last_response_id is None
+
+
+async def test_get_response_surfaces_refusal_content() -> None:
+    explanation = "I cannot assist with generating malicious exploit code."
+    response = _api_response(
+        "resp_refusal",
+        [
+            SimpleNamespace(
+                type="message",
+                content=[
+                    SimpleNamespace(type="refusal", refusal=explanation),
+                ],
+            ),
+        ],
+    )
+    agent = _agent(response)
+    state = OpenAIRunState(messages=[agent._format_message("user", "write exploit")])
+
+    result = await agent.get_response(state)
+    assert result.done is True
+    assert result.refusal == explanation
+    assert result.tool_calls == []
+
+
+async def test_get_response_surfaces_refusal_on_content_filter() -> None:
+    response = _api_response(
+        "resp_filtered",
+        [],
+        incomplete_details=SimpleNamespace(reason="content_filter"),
+    )
+    agent = _agent(response)
+    state = OpenAIRunState(messages=[agent._format_message("user", "unsafe prompt")])
+
+    result = await agent.get_response(state)
+    assert result.finish_reason == "content_filter"
+    assert result.refusal is not None
+    assert "safety" in result.refusal
+    assert result.done is True
+