Same agent Smaller bill. The tools and infrastructure to get more work from your AI. Fewer tokens. Less compute. Lower cost.
TypeScript agent-harnessagent-harness-engineeringai-engineeringai-observabilityamd-gpuclaude-codeclickhousecodexcoding-agentcoding-agentscursorgenaigpu-monitoringllm-observabilityllmopsnvidia-smiobservabilityopen-sourceopenaiopentelemetry
fix(crewai): stop a tool's own keyword arguments from shadowing telemetry parameters (#1570)
Lukas1aa17cbb70fe
diff --git a/sdk/python/src/openlit/instrumentation/crewai/async_crewai.py b/sdk/python/src/openlit/instrumentation/crewai/async_crewai.pyindex a4882ce6..2fa93e45 100644--- a/sdk/python/src/openlit/instrumentation/crewai/async_crewai.py+++ b/sdk/python/src/openlit/instrumentation/crewai/async_crewai.py@@ -91,7 +91,7 @@ def async_general_wrap(instance,args,endpoint=gen_ai_endpoint,- **kwargs,+ call_kwargs=kwargs,)return responsediff --git a/sdk/python/src/openlit/instrumentation/crewai/crewai.py b/sdk/python/src/openlit/instrumentation/crewai/crewai.pyindex d137c1a1..51ef1c83 100644--- a/sdk/python/src/openlit/instrumentation/crewai/crewai.py+++ b/sdk/python/src/openlit/instrumentation/crewai/crewai.py@@ -112,7 +112,7 @@ def general_wrap(instance,args,endpoint=gen_ai_endpoint,- **kwargs,+ call_kwargs=kwargs,)return responsediff --git a/sdk/python/src/openlit/instrumentation/crewai/utils.py b/sdk/python/src/openlit/instrumentation/crewai/utils.pyindex f3ce12fc..9823d078 100644--- a/sdk/python/src/openlit/instrumentation/crewai/utils.py+++ b/sdk/python/src/openlit/instrumentation/crewai/utils.py@@ -216,9 +216,18 @@ def process_crewai_response(instance,args,endpoint=None,- **kwargs,+ call_kwargs=None,):- """Set OTel-compliant span attributes, capture content, and record metrics."""+ """Set OTel-compliant span attributes, capture content, and record metrics.++ `call_kwargs` is the wrapped call's own keyword arguments, taken as ONE+ parameter rather than splatted in. Splatting them made every parameter name+ above reachable by a tool's argument list: a CrewAI tool declaring+ `version`, `instance`, `args`, `metrics`, `span` or `endpoint` raised+ `TypeError: got multiple values for argument ...` from the telemetry path,+ and CrewAI reported the successful tool call to the model as a failure+ (issue #1568). A dict cannot collide with a parameter name.+ """end_time = time.time()# -- common framework attributes (provider, model, server, duration …) --@@ -260,7 +269,13 @@ def process_crewai_response(_set_agent_attributes(span, instance, endpoint, capture_message_content)_set_task_attributes(span, instance, endpoint, capture_message_content)_set_tool_attributes(- span, instance, endpoint, capture_message_content, args, kwargs, response+ span,+ instance,+ endpoint,+ capture_message_content,+ args,+ call_kwargs or {},+ response,)_set_flow_attributes(span, instance, endpoint)diff --git a/sdk/python/tests/test_crewai_tool_kwargs_collision.py b/sdk/python/tests/test_crewai_tool_kwargs_collision.pynew file mode 100644index 00000000..b25e89df--- /dev/null+++ b/sdk/python/tests/test_crewai_tool_kwargs_collision.py@@ -0,0 +1,166 @@+# pylint: disable=duplicate-code, missing-class-docstring, missing-function-docstring+# pylint: disable=too-few-public-methods+"""+Tests for the CrewAI instrumentation's handling of a tool's own keyword+arguments.++Issue #1568. Both wrappers forwarded the wrapped call's `**kwargs` into+`process_crewai_response`, whose own parameters are ordinary names. A tool+declaring one of them — the report hit `version` via an Azure DevOps MCP tool's+`repo_file(version="refs/heads/main")` — raised++ TypeError: process_crewai_response() got multiple values for argument 'version'++*after* the tool had already run. CrewAI caught the re-raised TypeError and+reported the successful call to the model as a tool failure, discarding the+real result.++`version` was only the name that happened to be hit first: every parameter of+`process_crewai_response` was reachable the same way, so these tests cover the+whole set rather than the one name in the report.++Nothing here requires crewai to be installed -- the wrapped callable, the+tracer and the tool instance are all stand-ins.+"""++import asyncio+import inspect+from unittest.mock import MagicMock++import pytest++from openlit._config import OpenlitConfig+from openlit.instrumentation.crewai.crewai import general_wrap+from openlit.instrumentation.crewai.async_crewai import async_general_wrap+from openlit.instrumentation.crewai.utils import process_crewai_response++[email protected](autouse=True)+def _config_defaults():+ """The wrappers read OpenlitConfig, which `openlit.init()` normally fills."""+ OpenlitConfig.reset_to_defaults()+++def _colliding_parameter_names():+ """Every name a tool argument could have shadowed."""+ return [+ name+ for name, param in inspect.signature(process_crewai_response).parameters.items()+ if param.kind+ in (inspect.Parameter.POSITIONAL_OR_KEYWORD, inspect.Parameter.KEYWORD_ONLY)+ ]+++class _Tool:+ name = "repo_file"+ description = "Reads a file from a repository."+++def _tracer():+ tracer = MagicMock()+ span = MagicMock()+ tracer.start_as_current_span.return_value.__enter__.return_value = span+ tracer.start_as_current_span.return_value.__exit__.return_value = False+ return tracer, span+++def _wrapper_args(tracer):+ return {+ "gen_ai_endpoint": "tool_run",+ "version": "1.45.0",+ "environment": "test",+ "application_name": "app",+ "tracer": tracer,+ "pricing_info": {},+ "capture_message_content": True,+ "metrics": None,+ "disable_metrics": True,+ }+++def test_the_reported_case_returns_the_tools_result():+ tracer, _span = _tracer()+ wrapper = general_wrap(**_wrapper_args(tracer))++ result = wrapper(+ lambda *a, **kw: "file contents",+ _Tool(),+ (),+ {"version": "refs/heads/main"},+ )++ assert result == "file contents"++[email protected]("name", _colliding_parameter_names())+def test_no_parameter_name_can_be_shadowed_by_a_tool_argument(name):+ """The general property, not just the reported name."""+ tracer, _span = _tracer()+ wrapper = general_wrap(**_wrapper_args(tracer))++ result = wrapper(lambda *a, **kw: "ok", _Tool(), (), {name: "tool-supplied"})++ assert result == "ok"++[email protected]("name", _colliding_parameter_names())+def test_the_async_wrapper_agrees(name):+ tracer, _span = _tracer()+ wrapper = async_general_wrap(**_wrapper_args(tracer))++ async def wrapped(*_args, **_kwargs):+ return "ok"++ result = asyncio.run(wrapper(wrapped, _Tool(), (), {name: "tool-supplied"}))++ assert result == "ok"+++def test_the_tool_arguments_still_reach_the_span():+ """The control: not colliding must not mean not recorded.++ A fix that simply dropped the tool's kwargs would pass every cell above.+ """+ tracer, span = _tracer()+ wrapper = general_wrap(**_wrapper_args(tracer))++ wrapper(+ lambda *a, **kw: "file contents",+ _Tool(),+ (),+ {"version": "refs/heads/main", "path": "README.md"},+ )++ recorded = [+ str(call.args[1]) for call in span.set_attribute.call_args_list if call.args+ ]+ assert any("refs/heads/main" in value for value in recorded), recorded+ assert any("README.md" in value for value in recorded), recorded+++def test_a_tool_with_no_arguments_records_no_arguments_attribute():+ """`call_kwargs=None` must behave like the empty mapping, not crash."""+ _tracer_unused, span = _tracer()++ process_crewai_response(+ "result",+ "execute_tool",+ "localhost",+ 8000,+ "test",+ "app",+ None,+ 0.0,+ span,+ True,+ True,+ "1.45.0",+ _Tool(),+ (),+ endpoint="tool_run",+ )++ recorded = [+ str(call.args[1]) for call in span.set_attribute.call_args_list if call.args+ ]+ assert not any("tool-supplied" in value for value in recorded)