File size: 54,690 Bytes
14edff4 | 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280 281 282 283 284 285 286 287 288 289 290 291 292 293 294 295 296 297 298 299 300 301 302 303 304 305 306 307 308 309 310 311 312 313 314 315 316 317 318 319 320 321 322 323 324 325 326 327 328 329 330 331 332 333 334 335 336 337 338 339 340 341 342 343 344 345 346 347 348 349 350 351 352 353 354 355 356 357 358 359 360 361 362 363 364 365 366 367 368 369 370 371 372 373 374 375 376 377 378 379 380 381 382 383 384 385 386 387 388 389 390 391 392 393 394 395 396 397 398 399 400 401 402 403 404 405 406 407 408 409 410 411 412 413 414 415 416 417 418 419 420 421 422 423 424 425 426 427 428 429 430 431 432 433 434 435 436 437 438 439 440 441 442 443 444 445 446 447 448 449 450 451 452 453 454 455 456 457 458 459 460 461 462 463 464 465 466 467 468 469 470 471 472 473 474 475 476 477 478 479 480 481 482 483 484 485 486 487 488 489 490 491 492 493 494 495 496 497 498 499 500 501 502 503 504 505 506 507 508 509 510 511 512 513 514 515 516 517 518 519 520 521 522 523 524 525 526 527 528 529 530 531 532 533 534 535 536 537 538 539 540 541 542 543 544 545 546 547 548 549 550 551 552 553 554 555 556 557 558 559 560 561 562 563 564 565 566 567 568 569 570 571 572 573 574 575 576 577 578 579 580 581 582 583 584 585 586 587 588 589 590 591 592 593 594 595 596 597 598 599 600 601 602 603 604 605 606 607 608 609 610 611 612 613 614 615 616 617 618 619 620 621 622 623 624 625 626 627 628 629 630 631 632 633 634 635 636 637 638 639 640 641 642 643 644 645 646 647 648 649 650 651 652 653 654 655 656 657 658 659 660 661 662 663 664 665 666 667 668 669 670 671 672 673 674 675 676 677 678 679 680 681 682 683 684 685 686 687 688 689 690 691 692 693 694 695 696 697 698 699 700 701 702 703 704 705 706 707 708 709 710 711 712 713 714 715 716 717 718 719 720 721 722 723 724 725 726 727 728 729 730 731 732 733 734 735 736 737 738 739 740 741 742 743 744 745 746 747 748 749 750 751 752 753 754 755 756 757 758 759 760 761 762 763 764 765 766 767 768 769 770 771 772 773 774 775 776 777 778 779 780 781 782 783 784 785 786 787 788 789 790 791 792 793 794 795 796 797 798 799 800 801 802 803 804 805 806 807 808 809 810 811 812 813 814 815 816 817 818 819 820 821 822 823 824 825 826 827 828 829 830 831 832 833 834 835 836 837 838 839 840 841 842 843 844 845 846 847 848 849 850 851 852 853 854 855 856 857 858 859 860 861 862 863 864 865 866 867 868 869 870 871 872 873 874 875 876 877 878 879 880 881 882 883 884 885 886 887 888 889 890 891 892 893 894 895 896 897 898 899 900 901 902 903 904 905 906 907 908 909 910 911 912 913 914 915 916 917 918 919 920 921 922 923 924 925 926 927 928 929 930 931 932 933 934 935 936 937 938 939 940 941 942 943 944 945 946 947 948 949 950 951 952 953 954 955 956 957 958 959 960 961 962 963 964 965 966 967 968 969 970 971 972 973 974 975 976 977 978 979 980 981 982 983 984 985 986 987 988 989 990 991 992 993 994 995 996 997 998 999 1000 1001 1002 1003 1004 1005 1006 1007 1008 1009 1010 1011 1012 1013 1014 1015 1016 1017 1018 1019 1020 1021 1022 1023 1024 1025 1026 1027 1028 1029 1030 1031 1032 1033 1034 1035 1036 1037 1038 1039 1040 1041 1042 1043 1044 1045 1046 1047 1048 1049 1050 1051 1052 1053 1054 1055 1056 1057 1058 1059 1060 1061 1062 1063 1064 1065 1066 1067 1068 1069 1070 1071 1072 1073 1074 1075 1076 1077 1078 1079 1080 1081 1082 1083 1084 1085 1086 1087 1088 1089 1090 1091 1092 1093 1094 1095 1096 1097 1098 1099 1100 1101 1102 1103 1104 1105 1106 1107 1108 1109 1110 1111 1112 1113 1114 1115 1116 1117 1118 1119 1120 1121 1122 1123 1124 1125 1126 1127 1128 1129 1130 1131 1132 1133 1134 1135 1136 1137 1138 1139 1140 1141 1142 1143 1144 1145 1146 1147 1148 1149 1150 1151 1152 1153 1154 1155 1156 1157 1158 1159 1160 1161 1162 1163 1164 1165 1166 1167 1168 1169 1170 1171 1172 1173 1174 1175 1176 1177 1178 1179 1180 1181 1182 1183 1184 1185 1186 1187 1188 1189 1190 1191 1192 1193 1194 1195 1196 1197 1198 1199 1200 1201 1202 1203 1204 1205 1206 1207 1208 1209 1210 1211 1212 1213 1214 1215 1216 1217 1218 1219 1220 1221 1222 1223 1224 1225 1226 1227 1228 1229 1230 1231 1232 1233 1234 1235 1236 1237 1238 1239 1240 1241 1242 1243 1244 1245 1246 1247 1248 1249 1250 1251 1252 1253 1254 1255 1256 1257 1258 1259 1260 1261 1262 1263 1264 1265 1266 1267 1268 1269 1270 1271 1272 1273 1274 1275 1276 1277 1278 1279 1280 1281 1282 1283 1284 1285 1286 1287 1288 1289 1290 1291 1292 1293 1294 1295 1296 1297 1298 1299 1300 1301 1302 1303 1304 1305 1306 1307 1308 1309 1310 1311 1312 1313 1314 1315 1316 1317 1318 1319 1320 1321 1322 1323 1324 1325 1326 1327 1328 1329 1330 1331 1332 1333 1334 1335 1336 1337 1338 1339 1340 1341 1342 1343 1344 1345 1346 1347 1348 1349 1350 1351 1352 1353 1354 1355 1356 1357 1358 1359 1360 1361 1362 1363 1364 1365 1366 1367 1368 1369 1370 1371 1372 1373 1374 1375 1376 1377 1378 1379 1380 1381 1382 1383 1384 1385 1386 1387 1388 1389 1390 1391 1392 1393 1394 1395 1396 1397 1398 1399 1400 1401 1402 1403 1404 1405 1406 1407 1408 1409 1410 1411 1412 1413 1414 1415 1416 1417 1418 1419 1420 1421 1422 1423 1424 1425 1426 1427 1428 1429 1430 1431 1432 1433 1434 1435 1436 1437 1438 1439 1440 1441 1442 1443 | from __future__ import annotations
import asyncio
import dataclasses
import inspect
from collections.abc import Awaitable
from dataclasses import dataclass, field
from typing import TYPE_CHECKING, Any, cast
from openai.types.responses import (
ResponseComputerToolCall,
ResponseFileSearchToolCall,
ResponseFunctionToolCall,
ResponseFunctionWebSearch,
ResponseOutputMessage,
)
from openai.types.responses.response_code_interpreter_tool_call import (
ResponseCodeInterpreterToolCall,
)
from openai.types.responses.response_computer_tool_call import (
ActionClick,
ActionDoubleClick,
ActionDrag,
ActionKeypress,
ActionMove,
ActionScreenshot,
ActionScroll,
ActionType,
ActionWait,
)
from openai.types.responses.response_input_item_param import (
ComputerCallOutputAcknowledgedSafetyCheck,
)
from openai.types.responses.response_input_param import ComputerCallOutput, McpApprovalResponse
from openai.types.responses.response_output_item import (
ImageGenerationCall,
LocalShellCall,
McpApprovalRequest,
McpCall,
McpListTools,
)
from openai.types.responses.response_reasoning_item import ResponseReasoningItem
from .agent import Agent, ToolsToFinalOutputResult
from .agent_output import AgentOutputSchemaBase
from .computer import AsyncComputer, Computer
from .exceptions import (
AgentsException,
ModelBehaviorError,
ToolInputGuardrailTripwireTriggered,
ToolOutputGuardrailTripwireTriggered,
UserError,
)
from .guardrail import InputGuardrail, InputGuardrailResult, OutputGuardrail, OutputGuardrailResult
from .handoffs import Handoff, HandoffInputData
from .items import (
HandoffCallItem,
HandoffOutputItem,
ItemHelpers,
MCPApprovalRequestItem,
MCPApprovalResponseItem,
MCPListToolsItem,
MessageOutputItem,
ModelResponse,
ReasoningItem,
RunItem,
ToolCallItem,
ToolCallOutputItem,
TResponseInputItem,
)
from .lifecycle import RunHooks
from .logger import logger
from .model_settings import ModelSettings
from .models.interface import ModelTracing
from .run_context import RunContextWrapper, TContext
from .stream_events import RunItemStreamEvent, StreamEvent
from .tool import (
ComputerTool,
ComputerToolSafetyCheckData,
FunctionTool,
FunctionToolResult,
HostedMCPTool,
LocalShellCommandRequest,
LocalShellTool,
MCPToolApprovalRequest,
Tool,
)
from .tool_context import ToolContext
from .tool_guardrails import (
ToolInputGuardrailData,
ToolInputGuardrailResult,
ToolOutputGuardrailData,
ToolOutputGuardrailResult,
)
from .tracing import (
SpanError,
Trace,
function_span,
get_current_trace,
guardrail_span,
handoff_span,
trace,
)
from .util import _coro, _error_tracing
if TYPE_CHECKING:
from .run import RunConfig
class QueueCompleteSentinel:
pass
QUEUE_COMPLETE_SENTINEL = QueueCompleteSentinel()
_NOT_FINAL_OUTPUT = ToolsToFinalOutputResult(is_final_output=False, final_output=None)
@dataclass
class AgentToolUseTracker:
agent_to_tools: list[tuple[Agent, list[str]]] = field(default_factory=list)
"""Tuple of (agent, list of tools used). Can't use a dict because agents aren't hashable."""
def add_tool_use(self, agent: Agent[Any], tool_names: list[str]) -> None:
existing_data = next((item for item in self.agent_to_tools if item[0] == agent), None)
if existing_data:
existing_data[1].extend(tool_names)
else:
self.agent_to_tools.append((agent, tool_names))
def has_used_tools(self, agent: Agent[Any]) -> bool:
existing_data = next((item for item in self.agent_to_tools if item[0] == agent), None)
return existing_data is not None and len(existing_data[1]) > 0
@dataclass
class ToolRunHandoff:
handoff: Handoff
tool_call: ResponseFunctionToolCall
@dataclass
class ToolRunFunction:
tool_call: ResponseFunctionToolCall
function_tool: FunctionTool
@dataclass
class ToolRunComputerAction:
tool_call: ResponseComputerToolCall
computer_tool: ComputerTool
@dataclass
class ToolRunMCPApprovalRequest:
request_item: McpApprovalRequest
mcp_tool: HostedMCPTool
@dataclass
class ToolRunLocalShellCall:
tool_call: LocalShellCall
local_shell_tool: LocalShellTool
@dataclass
class ProcessedResponse:
new_items: list[RunItem]
handoffs: list[ToolRunHandoff]
functions: list[ToolRunFunction]
computer_actions: list[ToolRunComputerAction]
local_shell_calls: list[ToolRunLocalShellCall]
tools_used: list[str] # Names of all tools used, including hosted tools
mcp_approval_requests: list[ToolRunMCPApprovalRequest] # Only requests with callbacks
def has_tools_or_approvals_to_run(self) -> bool:
# Handoffs, functions and computer actions need local processing
# Hosted tools have already run, so there's nothing to do.
return any(
[
self.handoffs,
self.functions,
self.computer_actions,
self.local_shell_calls,
self.mcp_approval_requests,
]
)
@dataclass
class NextStepHandoff:
new_agent: Agent[Any]
@dataclass
class NextStepFinalOutput:
output: Any
@dataclass
class NextStepRunAgain:
pass
@dataclass
class SingleStepResult:
original_input: str | list[TResponseInputItem]
"""The input items i.e. the items before run() was called. May be mutated by handoff input
filters."""
model_response: ModelResponse
"""The model response for the current step."""
pre_step_items: list[RunItem]
"""Items generated before the current step."""
new_step_items: list[RunItem]
"""Items generated during this current step."""
next_step: NextStepHandoff | NextStepFinalOutput | NextStepRunAgain
"""The next step to take."""
tool_input_guardrail_results: list[ToolInputGuardrailResult]
"""Tool input guardrail results from this step."""
tool_output_guardrail_results: list[ToolOutputGuardrailResult]
"""Tool output guardrail results from this step."""
@property
def generated_items(self) -> list[RunItem]:
"""Items generated during the agent run (i.e. everything generated after
`original_input`)."""
return self.pre_step_items + self.new_step_items
def get_model_tracing_impl(
tracing_disabled: bool, trace_include_sensitive_data: bool
) -> ModelTracing:
if tracing_disabled:
return ModelTracing.DISABLED
elif trace_include_sensitive_data:
return ModelTracing.ENABLED
else:
return ModelTracing.ENABLED_WITHOUT_DATA
class RunImpl:
@classmethod
async def execute_tools_and_side_effects(
cls,
*,
agent: Agent[TContext],
# The original input to the Runner
original_input: str | list[TResponseInputItem],
# Everything generated by Runner since the original input, but before the current step
pre_step_items: list[RunItem],
new_response: ModelResponse,
processed_response: ProcessedResponse,
output_schema: AgentOutputSchemaBase | None,
hooks: RunHooks[TContext],
context_wrapper: RunContextWrapper[TContext],
run_config: RunConfig,
) -> SingleStepResult:
# Make a copy of the generated items
pre_step_items = list(pre_step_items)
new_step_items: list[RunItem] = []
new_step_items.extend(processed_response.new_items)
# First, lets run the tool calls - function tools and computer actions
(
(function_results, tool_input_guardrail_results, tool_output_guardrail_results),
computer_results,
) = await asyncio.gather(
cls.execute_function_tool_calls(
agent=agent,
tool_runs=processed_response.functions,
hooks=hooks,
context_wrapper=context_wrapper,
config=run_config,
),
cls.execute_computer_actions(
agent=agent,
actions=processed_response.computer_actions,
hooks=hooks,
context_wrapper=context_wrapper,
config=run_config,
),
)
new_step_items.extend([result.run_item for result in function_results])
new_step_items.extend(computer_results)
# Next, run the MCP approval requests
if processed_response.mcp_approval_requests:
approval_results = await cls.execute_mcp_approval_requests(
agent=agent,
approval_requests=processed_response.mcp_approval_requests,
context_wrapper=context_wrapper,
)
new_step_items.extend(approval_results)
# Next, check if there are any handoffs
if run_handoffs := processed_response.handoffs:
return await cls.execute_handoffs(
agent=agent,
original_input=original_input,
pre_step_items=pre_step_items,
new_step_items=new_step_items,
new_response=new_response,
run_handoffs=run_handoffs,
hooks=hooks,
context_wrapper=context_wrapper,
run_config=run_config,
)
# Next, we'll check if the tool use should result in a final output
check_tool_use = await cls._check_for_final_output_from_tools(
agent=agent,
tool_results=function_results,
context_wrapper=context_wrapper,
config=run_config,
)
if check_tool_use.is_final_output:
# If the output type is str, then let's just stringify it
if not agent.output_type or agent.output_type is str:
check_tool_use.final_output = str(check_tool_use.final_output)
if check_tool_use.final_output is None:
logger.error(
"Model returned a final output of None. Not raising an error because we assume"
"you know what you're doing."
)
return await cls.execute_final_output(
agent=agent,
original_input=original_input,
new_response=new_response,
pre_step_items=pre_step_items,
new_step_items=new_step_items,
final_output=check_tool_use.final_output,
hooks=hooks,
context_wrapper=context_wrapper,
tool_input_guardrail_results=tool_input_guardrail_results,
tool_output_guardrail_results=tool_output_guardrail_results,
)
# Now we can check if the model also produced a final output
message_items = [item for item in new_step_items if isinstance(item, MessageOutputItem)]
# We'll use the last content output as the final output
potential_final_output_text = (
ItemHelpers.extract_last_text(message_items[-1].raw_item) if message_items else None
)
# Generate final output only when there are no pending tool calls or approval requests.
if not processed_response.has_tools_or_approvals_to_run():
if output_schema and not output_schema.is_plain_text() and potential_final_output_text:
final_output = output_schema.validate_json(potential_final_output_text)
return await cls.execute_final_output(
agent=agent,
original_input=original_input,
new_response=new_response,
pre_step_items=pre_step_items,
new_step_items=new_step_items,
final_output=final_output,
hooks=hooks,
context_wrapper=context_wrapper,
tool_input_guardrail_results=tool_input_guardrail_results,
tool_output_guardrail_results=tool_output_guardrail_results,
)
elif not output_schema or output_schema.is_plain_text():
return await cls.execute_final_output(
agent=agent,
original_input=original_input,
new_response=new_response,
pre_step_items=pre_step_items,
new_step_items=new_step_items,
final_output=potential_final_output_text or "",
hooks=hooks,
context_wrapper=context_wrapper,
tool_input_guardrail_results=tool_input_guardrail_results,
tool_output_guardrail_results=tool_output_guardrail_results,
)
# If there's no final output, we can just run again
return SingleStepResult(
original_input=original_input,
model_response=new_response,
pre_step_items=pre_step_items,
new_step_items=new_step_items,
next_step=NextStepRunAgain(),
tool_input_guardrail_results=tool_input_guardrail_results,
tool_output_guardrail_results=tool_output_guardrail_results,
)
@classmethod
def maybe_reset_tool_choice(
cls, agent: Agent[Any], tool_use_tracker: AgentToolUseTracker, model_settings: ModelSettings
) -> ModelSettings:
"""Resets tool choice to None if the agent has used tools and the agent's reset_tool_choice
flag is True."""
if agent.reset_tool_choice is True and tool_use_tracker.has_used_tools(agent):
return dataclasses.replace(model_settings, tool_choice=None)
return model_settings
@classmethod
def process_model_response(
cls,
*,
agent: Agent[Any],
all_tools: list[Tool],
response: ModelResponse,
output_schema: AgentOutputSchemaBase | None,
handoffs: list[Handoff],
) -> ProcessedResponse:
items: list[RunItem] = []
run_handoffs = []
functions = []
computer_actions = []
local_shell_calls = []
mcp_approval_requests = []
tools_used: list[str] = []
handoff_map = {handoff.tool_name: handoff for handoff in handoffs}
function_map = {tool.name: tool for tool in all_tools if isinstance(tool, FunctionTool)}
computer_tool = next((tool for tool in all_tools if isinstance(tool, ComputerTool)), None)
local_shell_tool = next(
(tool for tool in all_tools if isinstance(tool, LocalShellTool)), None
)
hosted_mcp_server_map = {
tool.tool_config["server_label"]: tool
for tool in all_tools
if isinstance(tool, HostedMCPTool)
}
for output in response.output:
if isinstance(output, ResponseOutputMessage):
items.append(MessageOutputItem(raw_item=output, agent=agent))
elif isinstance(output, ResponseFileSearchToolCall):
items.append(ToolCallItem(raw_item=output, agent=agent))
tools_used.append("file_search")
elif isinstance(output, ResponseFunctionWebSearch):
items.append(ToolCallItem(raw_item=output, agent=agent))
tools_used.append("web_search")
elif isinstance(output, ResponseReasoningItem):
items.append(ReasoningItem(raw_item=output, agent=agent))
elif isinstance(output, ResponseComputerToolCall):
items.append(ToolCallItem(raw_item=output, agent=agent))
tools_used.append("computer_use")
if not computer_tool:
_error_tracing.attach_error_to_current_span(
SpanError(
message="Computer tool not found",
data={},
)
)
raise ModelBehaviorError(
"Model produced computer action without a computer tool."
)
computer_actions.append(
ToolRunComputerAction(tool_call=output, computer_tool=computer_tool)
)
elif isinstance(output, McpApprovalRequest):
items.append(MCPApprovalRequestItem(raw_item=output, agent=agent))
if output.server_label not in hosted_mcp_server_map:
_error_tracing.attach_error_to_current_span(
SpanError(
message="MCP server label not found",
data={"server_label": output.server_label},
)
)
raise ModelBehaviorError(f"MCP server label {output.server_label} not found")
else:
server = hosted_mcp_server_map[output.server_label]
if server.on_approval_request:
mcp_approval_requests.append(
ToolRunMCPApprovalRequest(
request_item=output,
mcp_tool=server,
)
)
else:
logger.warning(
f"MCP server {output.server_label} has no on_approval_request hook"
)
elif isinstance(output, McpListTools):
items.append(MCPListToolsItem(raw_item=output, agent=agent))
elif isinstance(output, McpCall):
items.append(ToolCallItem(raw_item=output, agent=agent))
tools_used.append("mcp")
elif isinstance(output, ImageGenerationCall):
items.append(ToolCallItem(raw_item=output, agent=agent))
tools_used.append("image_generation")
elif isinstance(output, ResponseCodeInterpreterToolCall):
items.append(ToolCallItem(raw_item=output, agent=agent))
tools_used.append("code_interpreter")
elif isinstance(output, LocalShellCall):
items.append(ToolCallItem(raw_item=output, agent=agent))
tools_used.append("local_shell")
if not local_shell_tool:
_error_tracing.attach_error_to_current_span(
SpanError(
message="Local shell tool not found",
data={},
)
)
raise ModelBehaviorError(
"Model produced local shell call without a local shell tool."
)
local_shell_calls.append(
ToolRunLocalShellCall(tool_call=output, local_shell_tool=local_shell_tool)
)
elif not isinstance(output, ResponseFunctionToolCall):
logger.warning(f"Unexpected output type, ignoring: {type(output)}")
continue
# At this point we know it's a function tool call
if not isinstance(output, ResponseFunctionToolCall):
continue
tools_used.append(output.name)
# Handoffs
if output.name in handoff_map:
items.append(HandoffCallItem(raw_item=output, agent=agent))
handoff = ToolRunHandoff(
tool_call=output,
handoff=handoff_map[output.name],
)
run_handoffs.append(handoff)
# Regular function tool call
else:
if output.name not in function_map:
if output_schema is not None and output.name == "json_tool_call":
# LiteLLM could generate non-existent tool calls for structured outputs
items.append(ToolCallItem(raw_item=output, agent=agent))
functions.append(
ToolRunFunction(
tool_call=output,
# this tool does not exist in function_map, so generate ad-hoc one,
# which just parses the input if it's a string, and returns the
# value otherwise
function_tool=_build_litellm_json_tool_call(output),
)
)
continue
else:
_error_tracing.attach_error_to_current_span(
SpanError(
message="Tool not found",
data={"tool_name": output.name},
)
)
error = f"Tool {output.name} not found in agent {agent.name}"
raise ModelBehaviorError(error)
items.append(ToolCallItem(raw_item=output, agent=agent))
functions.append(
ToolRunFunction(
tool_call=output,
function_tool=function_map[output.name],
)
)
return ProcessedResponse(
new_items=items,
handoffs=run_handoffs,
functions=functions,
computer_actions=computer_actions,
local_shell_calls=local_shell_calls,
tools_used=tools_used,
mcp_approval_requests=mcp_approval_requests,
)
@classmethod
async def _execute_input_guardrails(
cls,
*,
func_tool: FunctionTool,
tool_context: ToolContext[TContext],
agent: Agent[TContext],
tool_input_guardrail_results: list[ToolInputGuardrailResult],
) -> str | None:
"""Execute input guardrails for a tool.
Args:
func_tool: The function tool being executed.
tool_context: The tool execution context.
agent: The agent executing the tool.
tool_input_guardrail_results: List to append guardrail results to.
Returns:
None if tool execution should proceed, or a message string if execution should be
skipped.
Raises:
ToolInputGuardrailTripwireTriggered: If a guardrail triggers an exception.
"""
if not func_tool.tool_input_guardrails:
return None
for guardrail in func_tool.tool_input_guardrails:
gr_out = await guardrail.run(
ToolInputGuardrailData(
context=tool_context,
agent=agent,
)
)
# Store the guardrail result
tool_input_guardrail_results.append(
ToolInputGuardrailResult(
guardrail=guardrail,
output=gr_out,
)
)
# Handle different behavior types
if gr_out.behavior["type"] == "raise_exception":
raise ToolInputGuardrailTripwireTriggered(guardrail=guardrail, output=gr_out)
elif gr_out.behavior["type"] == "reject_content":
# Set final_result to the message and skip tool execution
return gr_out.behavior["message"]
elif gr_out.behavior["type"] == "allow":
# Continue to next guardrail or tool execution
continue
return None
@classmethod
async def _execute_output_guardrails(
cls,
*,
func_tool: FunctionTool,
tool_context: ToolContext[TContext],
agent: Agent[TContext],
real_result: Any,
tool_output_guardrail_results: list[ToolOutputGuardrailResult],
) -> Any:
"""Execute output guardrails for a tool.
Args:
func_tool: The function tool being executed.
tool_context: The tool execution context.
agent: The agent executing the tool.
real_result: The actual result from the tool execution.
tool_output_guardrail_results: List to append guardrail results to.
Returns:
The final result after guardrail processing (may be modified).
Raises:
ToolOutputGuardrailTripwireTriggered: If a guardrail triggers an exception.
"""
if not func_tool.tool_output_guardrails:
return real_result
final_result = real_result
for output_guardrail in func_tool.tool_output_guardrails:
gr_out = await output_guardrail.run(
ToolOutputGuardrailData(
context=tool_context,
agent=agent,
output=real_result,
)
)
# Store the guardrail result
tool_output_guardrail_results.append(
ToolOutputGuardrailResult(
guardrail=output_guardrail,
output=gr_out,
)
)
# Handle different behavior types
if gr_out.behavior["type"] == "raise_exception":
raise ToolOutputGuardrailTripwireTriggered(
guardrail=output_guardrail, output=gr_out
)
elif gr_out.behavior["type"] == "reject_content":
# Override the result with the guardrail message
final_result = gr_out.behavior["message"]
break
elif gr_out.behavior["type"] == "allow":
# Continue to next guardrail
continue
return final_result
@classmethod
async def _execute_tool_with_hooks(
cls,
*,
func_tool: FunctionTool,
tool_context: ToolContext[TContext],
agent: Agent[TContext],
hooks: RunHooks[TContext],
tool_call: ResponseFunctionToolCall,
) -> Any:
"""Execute the core tool function with before/after hooks.
Args:
func_tool: The function tool being executed.
tool_context: The tool execution context.
agent: The agent executing the tool.
hooks: The run hooks to execute.
tool_call: The tool call details.
Returns:
The result from the tool execution.
"""
await asyncio.gather(
hooks.on_tool_start(tool_context, agent, func_tool),
(
agent.hooks.on_tool_start(tool_context, agent, func_tool)
if agent.hooks
else _coro.noop_coroutine()
),
)
return await func_tool.on_invoke_tool(tool_context, tool_call.arguments)
@classmethod
async def execute_function_tool_calls(
cls,
*,
agent: Agent[TContext],
tool_runs: list[ToolRunFunction],
hooks: RunHooks[TContext],
context_wrapper: RunContextWrapper[TContext],
config: RunConfig,
) -> tuple[
list[FunctionToolResult], list[ToolInputGuardrailResult], list[ToolOutputGuardrailResult]
]:
# Collect guardrail results
tool_input_guardrail_results: list[ToolInputGuardrailResult] = []
tool_output_guardrail_results: list[ToolOutputGuardrailResult] = []
async def run_single_tool(
func_tool: FunctionTool, tool_call: ResponseFunctionToolCall
) -> Any:
with function_span(func_tool.name) as span_fn:
tool_context = ToolContext.from_agent_context(
context_wrapper,
tool_call.call_id,
tool_call=tool_call,
)
if config.trace_include_sensitive_data:
span_fn.span_data.input = tool_call.arguments
try:
# 1) Run input tool guardrails, if any
rejected_message = await cls._execute_input_guardrails(
func_tool=func_tool,
tool_context=tool_context,
agent=agent,
tool_input_guardrail_results=tool_input_guardrail_results,
)
if rejected_message is not None:
# Input guardrail rejected the tool call
final_result = rejected_message
else:
# 2) Actually run the tool
real_result = await cls._execute_tool_with_hooks(
func_tool=func_tool,
tool_context=tool_context,
agent=agent,
hooks=hooks,
tool_call=tool_call,
)
# 3) Run output tool guardrails, if any
final_result = await cls._execute_output_guardrails(
func_tool=func_tool,
tool_context=tool_context,
agent=agent,
real_result=real_result,
tool_output_guardrail_results=tool_output_guardrail_results,
)
# 4) Tool end hooks (with final result, which may have been overridden)
await asyncio.gather(
hooks.on_tool_end(tool_context, agent, func_tool, final_result),
(
agent.hooks.on_tool_end(
tool_context, agent, func_tool, final_result
)
if agent.hooks
else _coro.noop_coroutine()
),
)
result = final_result
except Exception as e:
_error_tracing.attach_error_to_current_span(
SpanError(
message="Error running tool",
data={"tool_name": func_tool.name, "error": str(e)},
)
)
if isinstance(e, AgentsException):
raise e
raise UserError(f"Error running tool {func_tool.name}: {e}") from e
if config.trace_include_sensitive_data:
span_fn.span_data.output = result
return result
tasks = []
for tool_run in tool_runs:
function_tool = tool_run.function_tool
tasks.append(run_single_tool(function_tool, tool_run.tool_call))
results = await asyncio.gather(*tasks)
function_tool_results = [
FunctionToolResult(
tool=tool_run.function_tool,
output=result,
run_item=ToolCallOutputItem(
output=result,
raw_item=ItemHelpers.tool_call_output_item(tool_run.tool_call, str(result)),
agent=agent,
),
)
for tool_run, result in zip(tool_runs, results)
]
return function_tool_results, tool_input_guardrail_results, tool_output_guardrail_results
@classmethod
async def execute_local_shell_calls(
cls,
*,
agent: Agent[TContext],
calls: list[ToolRunLocalShellCall],
context_wrapper: RunContextWrapper[TContext],
hooks: RunHooks[TContext],
config: RunConfig,
) -> list[RunItem]:
results: list[RunItem] = []
# Need to run these serially, because each call can affect the local shell state
for call in calls:
results.append(
await LocalShellAction.execute(
agent=agent,
call=call,
hooks=hooks,
context_wrapper=context_wrapper,
config=config,
)
)
return results
@classmethod
async def execute_computer_actions(
cls,
*,
agent: Agent[TContext],
actions: list[ToolRunComputerAction],
hooks: RunHooks[TContext],
context_wrapper: RunContextWrapper[TContext],
config: RunConfig,
) -> list[RunItem]:
results: list[RunItem] = []
# Need to run these serially, because each action can affect the computer state
for action in actions:
acknowledged: list[ComputerCallOutputAcknowledgedSafetyCheck] | None = None
if action.tool_call.pending_safety_checks and action.computer_tool.on_safety_check:
acknowledged = []
for check in action.tool_call.pending_safety_checks:
data = ComputerToolSafetyCheckData(
ctx_wrapper=context_wrapper,
agent=agent,
tool_call=action.tool_call,
safety_check=check,
)
maybe = action.computer_tool.on_safety_check(data)
ack = await maybe if inspect.isawaitable(maybe) else maybe
if ack:
acknowledged.append(
ComputerCallOutputAcknowledgedSafetyCheck(
id=check.id,
code=check.code,
message=check.message,
)
)
else:
raise UserError("Computer tool safety check was not acknowledged")
results.append(
await ComputerAction.execute(
agent=agent,
action=action,
hooks=hooks,
context_wrapper=context_wrapper,
config=config,
acknowledged_safety_checks=acknowledged,
)
)
return results
@classmethod
async def execute_handoffs(
cls,
*,
agent: Agent[TContext],
original_input: str | list[TResponseInputItem],
pre_step_items: list[RunItem],
new_step_items: list[RunItem],
new_response: ModelResponse,
run_handoffs: list[ToolRunHandoff],
hooks: RunHooks[TContext],
context_wrapper: RunContextWrapper[TContext],
run_config: RunConfig,
) -> SingleStepResult:
# If there is more than one handoff, add tool responses that reject those handoffs
multiple_handoffs = len(run_handoffs) > 1
if multiple_handoffs:
output_message = "Multiple handoffs detected, ignoring this one."
new_step_items.extend(
[
ToolCallOutputItem(
output=output_message,
raw_item=ItemHelpers.tool_call_output_item(
handoff.tool_call, output_message
),
agent=agent,
)
for handoff in run_handoffs[1:]
]
)
actual_handoff = run_handoffs[0]
with handoff_span(from_agent=agent.name) as span_handoff:
handoff = actual_handoff.handoff
new_agent: Agent[Any] = await handoff.on_invoke_handoff(
context_wrapper, actual_handoff.tool_call.arguments
)
span_handoff.span_data.to_agent = new_agent.name
if multiple_handoffs:
requested_agents = [handoff.handoff.agent_name for handoff in run_handoffs]
span_handoff.set_error(
SpanError(
message="Multiple handoffs requested",
data={
"requested_agents": requested_agents,
},
)
)
# Append a tool output item for the handoff
new_step_items.append(
HandoffOutputItem(
agent=agent,
raw_item=ItemHelpers.tool_call_output_item(
actual_handoff.tool_call,
handoff.get_transfer_message(new_agent),
),
source_agent=agent,
target_agent=new_agent,
)
)
# Execute handoff hooks
await asyncio.gather(
hooks.on_handoff(
context=context_wrapper,
from_agent=agent,
to_agent=new_agent,
),
(
agent.hooks.on_handoff(
context_wrapper,
agent=new_agent,
source=agent,
)
if agent.hooks
else _coro.noop_coroutine()
),
)
# If there's an input filter, filter the input for the next agent
input_filter = handoff.input_filter or (
run_config.handoff_input_filter if run_config else None
)
if input_filter:
logger.debug("Filtering inputs for handoff")
handoff_input_data = HandoffInputData(
input_history=tuple(original_input)
if isinstance(original_input, list)
else original_input,
pre_handoff_items=tuple(pre_step_items),
new_items=tuple(new_step_items),
run_context=context_wrapper,
)
if not callable(input_filter):
_error_tracing.attach_error_to_span(
span_handoff,
SpanError(
message="Invalid input filter",
data={"details": "not callable()"},
),
)
raise UserError(f"Invalid input filter: {input_filter}")
filtered = input_filter(handoff_input_data)
if inspect.isawaitable(filtered):
filtered = await filtered
if not isinstance(filtered, HandoffInputData):
_error_tracing.attach_error_to_span(
span_handoff,
SpanError(
message="Invalid input filter result",
data={"details": "not a HandoffInputData"},
),
)
raise UserError(f"Invalid input filter result: {filtered}")
original_input = (
filtered.input_history
if isinstance(filtered.input_history, str)
else list(filtered.input_history)
)
pre_step_items = list(filtered.pre_handoff_items)
new_step_items = list(filtered.new_items)
return SingleStepResult(
original_input=original_input,
model_response=new_response,
pre_step_items=pre_step_items,
new_step_items=new_step_items,
next_step=NextStepHandoff(new_agent),
tool_input_guardrail_results=[],
tool_output_guardrail_results=[],
)
@classmethod
async def execute_mcp_approval_requests(
cls,
*,
agent: Agent[TContext],
approval_requests: list[ToolRunMCPApprovalRequest],
context_wrapper: RunContextWrapper[TContext],
) -> list[RunItem]:
async def run_single_approval(approval_request: ToolRunMCPApprovalRequest) -> RunItem:
callback = approval_request.mcp_tool.on_approval_request
assert callback is not None, "Callback is required for MCP approval requests"
maybe_awaitable_result = callback(
MCPToolApprovalRequest(context_wrapper, approval_request.request_item)
)
if inspect.isawaitable(maybe_awaitable_result):
result = await maybe_awaitable_result
else:
result = maybe_awaitable_result
reason = result.get("reason", None)
raw_item: McpApprovalResponse = {
"approval_request_id": approval_request.request_item.id,
"approve": result["approve"],
"type": "mcp_approval_response",
}
if not result["approve"] and reason:
raw_item["reason"] = reason
return MCPApprovalResponseItem(
raw_item=raw_item,
agent=agent,
)
tasks = [run_single_approval(approval_request) for approval_request in approval_requests]
return await asyncio.gather(*tasks)
@classmethod
async def execute_final_output(
cls,
*,
agent: Agent[TContext],
original_input: str | list[TResponseInputItem],
new_response: ModelResponse,
pre_step_items: list[RunItem],
new_step_items: list[RunItem],
final_output: Any,
hooks: RunHooks[TContext],
context_wrapper: RunContextWrapper[TContext],
tool_input_guardrail_results: list[ToolInputGuardrailResult],
tool_output_guardrail_results: list[ToolOutputGuardrailResult],
) -> SingleStepResult:
# Run the on_end hooks
await cls.run_final_output_hooks(agent, hooks, context_wrapper, final_output)
return SingleStepResult(
original_input=original_input,
model_response=new_response,
pre_step_items=pre_step_items,
new_step_items=new_step_items,
next_step=NextStepFinalOutput(final_output),
tool_input_guardrail_results=tool_input_guardrail_results,
tool_output_guardrail_results=tool_output_guardrail_results,
)
@classmethod
async def run_final_output_hooks(
cls,
agent: Agent[TContext],
hooks: RunHooks[TContext],
context_wrapper: RunContextWrapper[TContext],
final_output: Any,
):
await asyncio.gather(
hooks.on_agent_end(context_wrapper, agent, final_output),
agent.hooks.on_end(context_wrapper, agent, final_output)
if agent.hooks
else _coro.noop_coroutine(),
)
@classmethod
async def run_single_input_guardrail(
cls,
agent: Agent[Any],
guardrail: InputGuardrail[TContext],
input: str | list[TResponseInputItem],
context: RunContextWrapper[TContext],
) -> InputGuardrailResult:
with guardrail_span(guardrail.get_name()) as span_guardrail:
result = await guardrail.run(agent, input, context)
span_guardrail.span_data.triggered = result.output.tripwire_triggered
return result
@classmethod
async def run_single_output_guardrail(
cls,
guardrail: OutputGuardrail[TContext],
agent: Agent[Any],
agent_output: Any,
context: RunContextWrapper[TContext],
) -> OutputGuardrailResult:
with guardrail_span(guardrail.get_name()) as span_guardrail:
result = await guardrail.run(agent=agent, agent_output=agent_output, context=context)
span_guardrail.span_data.triggered = result.output.tripwire_triggered
return result
@classmethod
def stream_step_items_to_queue(
cls,
new_step_items: list[RunItem],
queue: asyncio.Queue[StreamEvent | QueueCompleteSentinel],
):
for item in new_step_items:
if isinstance(item, MessageOutputItem):
event = RunItemStreamEvent(item=item, name="message_output_created")
elif isinstance(item, HandoffCallItem):
event = RunItemStreamEvent(item=item, name="handoff_requested")
elif isinstance(item, HandoffOutputItem):
event = RunItemStreamEvent(item=item, name="handoff_occured")
elif isinstance(item, ToolCallItem):
event = RunItemStreamEvent(item=item, name="tool_called")
elif isinstance(item, ToolCallOutputItem):
event = RunItemStreamEvent(item=item, name="tool_output")
elif isinstance(item, ReasoningItem):
event = RunItemStreamEvent(item=item, name="reasoning_item_created")
elif isinstance(item, MCPApprovalRequestItem):
event = RunItemStreamEvent(item=item, name="mcp_approval_requested")
elif isinstance(item, MCPListToolsItem):
event = RunItemStreamEvent(item=item, name="mcp_list_tools")
else:
logger.warning(f"Unexpected item type: {type(item)}")
event = None
if event:
queue.put_nowait(event)
@classmethod
def stream_step_result_to_queue(
cls,
step_result: SingleStepResult,
queue: asyncio.Queue[StreamEvent | QueueCompleteSentinel],
):
cls.stream_step_items_to_queue(step_result.new_step_items, queue)
@classmethod
async def _check_for_final_output_from_tools(
cls,
*,
agent: Agent[TContext],
tool_results: list[FunctionToolResult],
context_wrapper: RunContextWrapper[TContext],
config: RunConfig,
) -> ToolsToFinalOutputResult:
"""Determine if tool results should produce a final output.
Returns:
ToolsToFinalOutputResult: Indicates whether final output is ready, and the output value.
"""
if not tool_results:
return _NOT_FINAL_OUTPUT
if agent.tool_use_behavior == "run_llm_again":
return _NOT_FINAL_OUTPUT
elif agent.tool_use_behavior == "stop_on_first_tool":
return ToolsToFinalOutputResult(
is_final_output=True, final_output=tool_results[0].output
)
elif isinstance(agent.tool_use_behavior, dict):
names = agent.tool_use_behavior.get("stop_at_tool_names", [])
for tool_result in tool_results:
if tool_result.tool.name in names:
return ToolsToFinalOutputResult(
is_final_output=True, final_output=tool_result.output
)
return ToolsToFinalOutputResult(is_final_output=False, final_output=None)
elif callable(agent.tool_use_behavior):
if inspect.iscoroutinefunction(agent.tool_use_behavior):
return await cast(
Awaitable[ToolsToFinalOutputResult],
agent.tool_use_behavior(context_wrapper, tool_results),
)
else:
return cast(
ToolsToFinalOutputResult, agent.tool_use_behavior(context_wrapper, tool_results)
)
logger.error(f"Invalid tool_use_behavior: {agent.tool_use_behavior}")
raise UserError(f"Invalid tool_use_behavior: {agent.tool_use_behavior}")
class TraceCtxManager:
"""Creates a trace only if there is no current trace, and manages the trace lifecycle."""
def __init__(
self,
workflow_name: str,
trace_id: str | None,
group_id: str | None,
metadata: dict[str, Any] | None,
disabled: bool,
):
self.trace: Trace | None = None
self.workflow_name = workflow_name
self.trace_id = trace_id
self.group_id = group_id
self.metadata = metadata
self.disabled = disabled
def __enter__(self) -> TraceCtxManager:
current_trace = get_current_trace()
if not current_trace:
self.trace = trace(
workflow_name=self.workflow_name,
trace_id=self.trace_id,
group_id=self.group_id,
metadata=self.metadata,
disabled=self.disabled,
)
self.trace.start(mark_as_current=True)
return self
def __exit__(self, exc_type, exc_val, exc_tb):
if self.trace:
self.trace.finish(reset_current=True)
class ComputerAction:
@classmethod
async def execute(
cls,
*,
agent: Agent[TContext],
action: ToolRunComputerAction,
hooks: RunHooks[TContext],
context_wrapper: RunContextWrapper[TContext],
config: RunConfig,
acknowledged_safety_checks: list[ComputerCallOutputAcknowledgedSafetyCheck] | None = None,
) -> RunItem:
output_func = (
cls._get_screenshot_async(action.computer_tool.computer, action.tool_call)
if isinstance(action.computer_tool.computer, AsyncComputer)
else cls._get_screenshot_sync(action.computer_tool.computer, action.tool_call)
)
_, _, output = await asyncio.gather(
hooks.on_tool_start(context_wrapper, agent, action.computer_tool),
(
agent.hooks.on_tool_start(context_wrapper, agent, action.computer_tool)
if agent.hooks
else _coro.noop_coroutine()
),
output_func,
)
await asyncio.gather(
hooks.on_tool_end(context_wrapper, agent, action.computer_tool, output),
(
agent.hooks.on_tool_end(context_wrapper, agent, action.computer_tool, output)
if agent.hooks
else _coro.noop_coroutine()
),
)
# TODO: don't send a screenshot every single time, use references
image_url = f"data:image/png;base64,{output}"
return ToolCallOutputItem(
agent=agent,
output=image_url,
raw_item=ComputerCallOutput(
call_id=action.tool_call.call_id,
output={
"type": "computer_screenshot",
"image_url": image_url,
},
type="computer_call_output",
acknowledged_safety_checks=acknowledged_safety_checks,
),
)
@classmethod
async def _get_screenshot_sync(
cls,
computer: Computer,
tool_call: ResponseComputerToolCall,
) -> str:
action = tool_call.action
if isinstance(action, ActionClick):
computer.click(action.x, action.y, action.button)
elif isinstance(action, ActionDoubleClick):
computer.double_click(action.x, action.y)
elif isinstance(action, ActionDrag):
computer.drag([(p.x, p.y) for p in action.path])
elif isinstance(action, ActionKeypress):
computer.keypress(action.keys)
elif isinstance(action, ActionMove):
computer.move(action.x, action.y)
elif isinstance(action, ActionScreenshot):
computer.screenshot()
elif isinstance(action, ActionScroll):
computer.scroll(action.x, action.y, action.scroll_x, action.scroll_y)
elif isinstance(action, ActionType):
computer.type(action.text)
elif isinstance(action, ActionWait):
computer.wait()
return computer.screenshot()
@classmethod
async def _get_screenshot_async(
cls,
computer: AsyncComputer,
tool_call: ResponseComputerToolCall,
) -> str:
action = tool_call.action
if isinstance(action, ActionClick):
await computer.click(action.x, action.y, action.button)
elif isinstance(action, ActionDoubleClick):
await computer.double_click(action.x, action.y)
elif isinstance(action, ActionDrag):
await computer.drag([(p.x, p.y) for p in action.path])
elif isinstance(action, ActionKeypress):
await computer.keypress(action.keys)
elif isinstance(action, ActionMove):
await computer.move(action.x, action.y)
elif isinstance(action, ActionScreenshot):
await computer.screenshot()
elif isinstance(action, ActionScroll):
await computer.scroll(action.x, action.y, action.scroll_x, action.scroll_y)
elif isinstance(action, ActionType):
await computer.type(action.text)
elif isinstance(action, ActionWait):
await computer.wait()
return await computer.screenshot()
class LocalShellAction:
@classmethod
async def execute(
cls,
*,
agent: Agent[TContext],
call: ToolRunLocalShellCall,
hooks: RunHooks[TContext],
context_wrapper: RunContextWrapper[TContext],
config: RunConfig,
) -> RunItem:
await asyncio.gather(
hooks.on_tool_start(context_wrapper, agent, call.local_shell_tool),
(
agent.hooks.on_tool_start(context_wrapper, agent, call.local_shell_tool)
if agent.hooks
else _coro.noop_coroutine()
),
)
request = LocalShellCommandRequest(
ctx_wrapper=context_wrapper,
data=call.tool_call,
)
output = call.local_shell_tool.executor(request)
if inspect.isawaitable(output):
result = await output
else:
result = output
await asyncio.gather(
hooks.on_tool_end(context_wrapper, agent, call.local_shell_tool, result),
(
agent.hooks.on_tool_end(context_wrapper, agent, call.local_shell_tool, result)
if agent.hooks
else _coro.noop_coroutine()
),
)
return ToolCallOutputItem(
agent=agent,
output=output,
raw_item={
"type": "local_shell_call_output",
"id": call.tool_call.call_id,
"output": result,
# "id": "out" + call.tool_call.id, # TODO remove this, it should be optional
},
)
def _build_litellm_json_tool_call(output: ResponseFunctionToolCall) -> FunctionTool:
async def on_invoke_tool(_ctx: ToolContext[Any], value: Any) -> Any:
if isinstance(value, str):
import json
return json.loads(value)
return value
return FunctionTool(
name=output.name,
description=output.name,
params_json_schema={},
on_invoke_tool=on_invoke_tool,
strict_json_schema=True,
is_enabled=True,
)
|