|
| 1 | +"""Function tool output policies at the Realtime session dispatch boundary.""" |
| 2 | + |
| 3 | +import asyncio |
| 4 | +import json |
| 5 | +from typing import Any, Literal |
| 6 | + |
| 7 | +import pytest |
| 8 | + |
| 9 | +from agents.decorators import tool, tool_output_guardrail |
| 10 | +from agents.exceptions import ModelBehaviorError, ToolOutputGuardrailTripwireTriggered |
| 11 | +from agents.realtime import RealtimeAgent, RealtimeSession |
| 12 | +from agents.realtime.events import RealtimeSessionEvent, RealtimeToolEnd |
| 13 | +from agents.realtime.model_events import RealtimeModelToolCallEvent |
| 14 | +from agents.realtime.model_inputs import RealtimeModelSendEvent, RealtimeModelSendToolOutput |
| 15 | +from agents.realtime.testing import ScriptedRealtimeModel |
| 16 | +from agents.tool_context import ToolContext |
| 17 | +from agents.tool_guardrails import ToolGuardrailFunctionOutput, ToolOutputGuardrailData |
| 18 | + |
| 19 | +from ..mcp.helpers import FakeMCPServer |
| 20 | + |
| 21 | + |
| 22 | +def tool_call() -> RealtimeModelToolCallEvent: |
| 23 | + return RealtimeModelToolCallEvent(name="lookup", call_id="call_lookup", arguments="{}") |
| 24 | + |
| 25 | + |
| 26 | +def outputs(model: ScriptedRealtimeModel) -> list[str]: |
| 27 | + return [ |
| 28 | + event.output |
| 29 | + for event in model.sent_events |
| 30 | + if isinstance(event, RealtimeModelSendToolOutput) |
| 31 | + ] |
| 32 | + |
| 33 | + |
| 34 | +async def collect_events(session: RealtimeSession, events: list[RealtimeSessionEvent]) -> None: |
| 35 | + async for event in session: |
| 36 | + events.append(event) |
| 37 | + |
| 38 | + |
| 39 | +async def next_event(session: RealtimeSession, event_type: str) -> RealtimeSessionEvent: |
| 40 | + async def wait() -> RealtimeSessionEvent: |
| 41 | + async for event in session: |
| 42 | + if event.type == event_type: |
| 43 | + return event |
| 44 | + raise AssertionError(f"Session closed before {event_type}") |
| 45 | + |
| 46 | + return await asyncio.wait_for(wait(), timeout=2) |
| 47 | + |
| 48 | + |
| 49 | +@pytest.mark.asyncio |
| 50 | +@pytest.mark.parametrize("reject", [False, True]) |
| 51 | +async def test_tool_output_policies_check_original_result_in_order(reject: bool) -> None: |
| 52 | + original = {"value": "synthetic tool result"} |
| 53 | + seen: list[tuple[str, ToolOutputGuardrailData]] = [] |
| 54 | + invocation_contexts: list[ToolContext[Any]] = [] |
| 55 | + |
| 56 | + @tool_output_guardrail |
| 57 | + def first(data: ToolOutputGuardrailData) -> ToolGuardrailFunctionOutput: |
| 58 | + seen.append(("first", data)) |
| 59 | + return ToolGuardrailFunctionOutput.allow() |
| 60 | + |
| 61 | + @tool_output_guardrail |
| 62 | + async def second(data: ToolOutputGuardrailData) -> ToolGuardrailFunctionOutput: |
| 63 | + seen.append(("second", data)) |
| 64 | + if reject: |
| 65 | + return ToolGuardrailFunctionOutput.reject_content("replacement output") |
| 66 | + return ToolGuardrailFunctionOutput.allow() |
| 67 | + |
| 68 | + @tool_output_guardrail |
| 69 | + def third(data: ToolOutputGuardrailData) -> ToolGuardrailFunctionOutput: |
| 70 | + seen.append(("third", data)) |
| 71 | + return ToolGuardrailFunctionOutput.allow() |
| 72 | + |
| 73 | + @tool(tool_output_guardrails=[first, second, third]) |
| 74 | + async def lookup(context: ToolContext[Any]) -> dict[str, str]: |
| 75 | + invocation_contexts.append(context) |
| 76 | + return original |
| 77 | + |
| 78 | + agent = RealtimeAgent(name="source", tools=[lookup]) |
| 79 | + model = ScriptedRealtimeModel(strict=False) |
| 80 | + async with RealtimeSession( |
| 81 | + model, agent, {"user": "synthetic"}, run_config={"async_tool_calls": False} |
| 82 | + ) as session: |
| 83 | + await model.emit(tool_call()) |
| 84 | + end = await next_event(session, "tool_end") |
| 85 | + assert isinstance(end, RealtimeToolEnd) |
| 86 | + assert end.output == ("replacement output" if reject else original) |
| 87 | + assert end.agent is agent |
| 88 | + assert outputs(model) == ["replacement output" if reject else json.dumps(original)] |
| 89 | + assert [name for name, _ in seen] == ( |
| 90 | + ["first", "second"] if reject else ["first", "second", "third"] |
| 91 | + ) |
| 92 | + for _, data in seen: |
| 93 | + assert data.output is original |
| 94 | + assert data.agent is agent |
| 95 | + assert data.context is invocation_contexts[0] |
| 96 | + assert data.context.tool_call_id == "call_lookup" |
| 97 | + assert data.context.tool_arguments == "{}" |
| 98 | + await model.emit(tool_call()) |
| 99 | + assert len(invocation_contexts) == 1 |
| 100 | + assert len(outputs(model)) == 1 |
| 101 | + |
| 102 | + |
| 103 | +@pytest.mark.asyncio |
| 104 | +@pytest.mark.parametrize("async_tools", [False, True]) |
| 105 | +@pytest.mark.parametrize("behavior", ["tripwire", "error"]) |
| 106 | +async def test_output_policy_failure_never_publishes_or_reexecutes( |
| 107 | + async_tools: bool, behavior: str |
| 108 | +) -> None: |
| 109 | + calls = 0 |
| 110 | + checks = 0 |
| 111 | + events: list[RealtimeSessionEvent] = [] |
| 112 | + |
| 113 | + @tool_output_guardrail |
| 114 | + async def policy(data: ToolOutputGuardrailData) -> ToolGuardrailFunctionOutput: |
| 115 | + nonlocal checks |
| 116 | + checks += 1 |
| 117 | + assert data.output == "unchecked result" |
| 118 | + if behavior == "error": |
| 119 | + raise ValueError("policy unavailable") |
| 120 | + return ToolGuardrailFunctionOutput.raise_exception() |
| 121 | + |
| 122 | + @tool(tool_output_guardrails=[policy]) |
| 123 | + async def lookup() -> str: |
| 124 | + nonlocal calls |
| 125 | + calls += 1 |
| 126 | + return "unchecked result" |
| 127 | + |
| 128 | + model = ScriptedRealtimeModel(strict=False) |
| 129 | + session = RealtimeSession( |
| 130 | + model, |
| 131 | + RealtimeAgent(name="source", tools=[lookup]), |
| 132 | + None, |
| 133 | + run_config={"async_tool_calls": async_tools}, |
| 134 | + ) |
| 135 | + expected_error = ValueError if behavior == "error" else ToolOutputGuardrailTripwireTriggered |
| 136 | + async with session: |
| 137 | + if async_tools: |
| 138 | + consumer = asyncio.create_task(collect_events(session, events)) |
| 139 | + await model.emit(tool_call()) |
| 140 | + with pytest.raises(expected_error): |
| 141 | + await asyncio.wait_for(consumer, timeout=2) |
| 142 | + else: |
| 143 | + with pytest.raises(expected_error): |
| 144 | + await model.emit(tool_call()) |
| 145 | + with pytest.raises(ModelBehaviorError, match="already executed"): |
| 146 | + await model.emit(tool_call()) |
| 147 | + if not async_tools: |
| 148 | + await collect_events(session, events) |
| 149 | + assert calls == checks == 1 |
| 150 | + assert outputs(model) == [] |
| 151 | + assert not any(isinstance(event, RealtimeToolEnd) for event in events) |
| 152 | + assert session._pending_tool_outputs == {} |
| 153 | + |
| 154 | + |
| 155 | +@pytest.mark.asyncio |
| 156 | +async def test_mcp_output_policy_is_enforced_after_conversion() -> None: |
| 157 | + seen: list[Any] = [] |
| 158 | + |
| 159 | + @tool_output_guardrail |
| 160 | + def policy(data: ToolOutputGuardrailData) -> ToolGuardrailFunctionOutput: |
| 161 | + seen.append(data.output) |
| 162 | + return ToolGuardrailFunctionOutput.reject_content("MCP replacement") |
| 163 | + |
| 164 | + server = FakeMCPServer(tool_output_guardrails=[policy]) |
| 165 | + server.add_tool("lookup", {"type": "object", "properties": {}}) |
| 166 | + agent = RealtimeAgent(name="source", mcp_servers=[server]) |
| 167 | + model = ScriptedRealtimeModel(strict=False) |
| 168 | + async with RealtimeSession( |
| 169 | + model, agent, None, run_config={"async_tool_calls": False} |
| 170 | + ) as session: |
| 171 | + await model.emit(tool_call()) |
| 172 | + end = await next_event(session, "tool_end") |
| 173 | + assert isinstance(end, RealtimeToolEnd) |
| 174 | + assert end.output == "MCP replacement" |
| 175 | + assert outputs(model) == ["MCP replacement"] |
| 176 | + assert server.tool_calls == ["lookup"] |
| 177 | + assert seen == [{"type": "text", "text": server.tool_results[0]}] |
| 178 | + |
| 179 | + |
| 180 | +@pytest.mark.asyncio |
| 181 | +async def test_approved_async_tool_output_policy_keeps_dispatch_agent() -> None: |
| 182 | + started = asyncio.Event() |
| 183 | + release = asyncio.Event() |
| 184 | + checks: list[ToolOutputGuardrailData] = [] |
| 185 | + |
| 186 | + @tool_output_guardrail |
| 187 | + async def policy(data: ToolOutputGuardrailData) -> ToolGuardrailFunctionOutput: |
| 188 | + checks.append(data) |
| 189 | + started.set() |
| 190 | + await release.wait() |
| 191 | + return ToolGuardrailFunctionOutput.reject_content("approved replacement") |
| 192 | + |
| 193 | + @tool(needs_approval=True, tool_output_guardrails=[policy]) |
| 194 | + async def lookup() -> str: |
| 195 | + return "unchecked result" |
| 196 | + |
| 197 | + agent = RealtimeAgent(name="source", tools=[lookup]) |
| 198 | + model = ScriptedRealtimeModel(strict=False) |
| 199 | + async with RealtimeSession(model, agent, None) as session: |
| 200 | + await model.emit(tool_call()) |
| 201 | + await next_event(session, "tool_approval_required") |
| 202 | + assert checks == [] |
| 203 | + await session.approve_tool_call("call_lookup") |
| 204 | + await asyncio.wait_for(started.wait(), timeout=2) |
| 205 | + await session.update_agent(RealtimeAgent(name="replacement")) |
| 206 | + assert outputs(model) == [] |
| 207 | + release.set() |
| 208 | + end = await next_event(session, "tool_end") |
| 209 | + assert isinstance(end, RealtimeToolEnd) |
| 210 | + assert end.agent is agent |
| 211 | + assert end.output == "approved replacement" |
| 212 | + assert checks[0].agent is agent |
| 213 | + assert checks[0].context.agent is agent |
| 214 | + assert outputs(model) == ["approved replacement"] |
| 215 | + |
| 216 | + |
| 217 | +class FailOnceModel(ScriptedRealtimeModel): |
| 218 | + def __init__(self) -> None: |
| 219 | + super().__init__(strict=False) |
| 220 | + self.attempts: list[str] = [] |
| 221 | + |
| 222 | + async def send_event(self, event: RealtimeModelSendEvent) -> None: |
| 223 | + if isinstance(event, RealtimeModelSendToolOutput): |
| 224 | + self.attempts.append(event.output) |
| 225 | + if len(self.attempts) == 1: |
| 226 | + raise RuntimeError("transport unavailable") |
| 227 | + await super().send_event(event) |
| 228 | + |
| 229 | + |
| 230 | +@pytest.mark.asyncio |
| 231 | +async def test_transport_retry_reuses_only_checked_output() -> None: |
| 232 | + calls = 0 |
| 233 | + checks = 0 |
| 234 | + |
| 235 | + @tool_output_guardrail |
| 236 | + def policy(data: ToolOutputGuardrailData) -> ToolGuardrailFunctionOutput: |
| 237 | + nonlocal checks |
| 238 | + checks += 1 |
| 239 | + return ToolGuardrailFunctionOutput.reject_content("retry replacement") |
| 240 | + |
| 241 | + @tool(tool_output_guardrails=[policy]) |
| 242 | + async def lookup() -> str: |
| 243 | + nonlocal calls |
| 244 | + calls += 1 |
| 245 | + return "unchecked result" |
| 246 | + |
| 247 | + model = FailOnceModel() |
| 248 | + async with RealtimeSession( |
| 249 | + model, RealtimeAgent(name="source", tools=[lookup]), None |
| 250 | + ) as session: |
| 251 | + await model.emit(tool_call()) |
| 252 | + await next_event(session, "error") |
| 253 | + assert outputs(model) == [] |
| 254 | + pending = session._pending_tool_outputs["call_lookup"] |
| 255 | + assert pending.output == "retry replacement" |
| 256 | + assert pending.tool_end_event is not None |
| 257 | + assert pending.tool_end_event.output == "retry replacement" |
| 258 | + await model.emit(tool_call()) |
| 259 | + end = await next_event(session, "tool_end") |
| 260 | + assert isinstance(end, RealtimeToolEnd) |
| 261 | + assert end.output == "retry replacement" |
| 262 | + assert outputs(model) == ["retry replacement"] |
| 263 | + assert model.attempts == ["retry replacement", "retry replacement"] |
| 264 | + assert calls == checks == 1 |
| 265 | + |
| 266 | + |
| 267 | +@pytest.mark.asyncio |
| 268 | +@pytest.mark.parametrize("shutdown", ["cancel", "close", "close_suppress_cancel"]) |
| 269 | +async def test_output_policy_shutdown_prevents_late_output( |
| 270 | + shutdown: Literal["cancel", "close", "close_suppress_cancel"], |
| 271 | +) -> None: |
| 272 | + started = asyncio.Event() |
| 273 | + cancelled = asyncio.Event() |
| 274 | + calls = 0 |
| 275 | + events: list[RealtimeSessionEvent] = [] |
| 276 | + |
| 277 | + @tool_output_guardrail |
| 278 | + async def policy(data: ToolOutputGuardrailData) -> ToolGuardrailFunctionOutput: |
| 279 | + started.set() |
| 280 | + try: |
| 281 | + await asyncio.Event().wait() |
| 282 | + except asyncio.CancelledError: |
| 283 | + cancelled.set() |
| 284 | + if shutdown != "close_suppress_cancel": |
| 285 | + raise |
| 286 | + return ToolGuardrailFunctionOutput.allow() |
| 287 | + |
| 288 | + @tool(tool_output_guardrails=[policy]) |
| 289 | + async def lookup() -> str: |
| 290 | + nonlocal calls |
| 291 | + calls += 1 |
| 292 | + return "unchecked result" |
| 293 | + |
| 294 | + model = ScriptedRealtimeModel(strict=False) |
| 295 | + async with RealtimeSession( |
| 296 | + model, |
| 297 | + RealtimeAgent(name="source", tools=[lookup]), |
| 298 | + None, |
| 299 | + run_config={"async_tool_calls": shutdown != "cancel"}, |
| 300 | + ) as session: |
| 301 | + if shutdown == "cancel": |
| 302 | + dispatch = asyncio.create_task(session.on_event(tool_call())) |
| 303 | + await asyncio.wait_for(started.wait(), timeout=2) |
| 304 | + dispatch.cancel() |
| 305 | + with pytest.raises(asyncio.CancelledError): |
| 306 | + await dispatch |
| 307 | + with pytest.raises(ModelBehaviorError, match="already executed"): |
| 308 | + await session.on_event(tool_call()) |
| 309 | + else: |
| 310 | + await model.emit(tool_call()) |
| 311 | + await asyncio.wait_for(started.wait(), timeout=2) |
| 312 | + await session.close() |
| 313 | + assert cancelled.is_set() |
| 314 | + await collect_events(session, events) |
| 315 | + assert calls == 1 |
| 316 | + assert outputs(model) == [] |
| 317 | + assert not any(isinstance(event, RealtimeToolEnd) for event in events) |
0 commit comments