feat(openai-agents): initial instrumentation; collect OpenAI agent traces and metrics

packages/opentelemetry-instrumentation-openai_agents/.flake8

@@ -0,0 +1,11 @@
+[flake8]
+exclude =
+	.git,
+	__pycache__,
+	build,
+	dist,
+	.tox,
+	venv,
+	.venv,
+	.pytest_cache
+max-line-length = 120

packages/opentelemetry-instrumentation-openai_agents/.python-version

@@ -0,0 +1 @@
+3.10

packages/opentelemetry-instrumentation-openai_agents/README.md

@@ -0,0 +1,33 @@
+# OpenTelemetry OpenAI Agents Instrumentation
+
+<a href="https://pypi.org/project/opentelemetry-instrumentation-openai_agents/">
+    <img src="https://badge.fury.io/py/opentelemetry-instrumentation-openai_agents.svg">
+</a>
+
+This library enables tracing of agentic workflows implemented using the [OpenAI Agents framework](https://github.com/openai/openai-agents-python), allowing visibility into agent reasoning, tool usage, and decision-making steps.
+
+## Installation
+
+```bash
+pip install opentelemetry-instrumentation-openai_agents
+```
+
+## Example usage
+
+```python
+from opentelemetry.instrumentation.openai_agents import OpenAIAgentsInstrumentor
+
+OpenAIAgentsInstrumentor().instrument()
+```
+
+## Privacy
+
+**By default, this instrumentation logs prompts, completions, and embeddings to span attributes**. This gives you a clear visibility into how your LLM application is working, and can make it easy to debug and evaluate the quality of the outputs.
+
+However, you may want to disable this logging for privacy reasons, as they may contain highly sensitive data from your users. You may also simply want to reduce the size of your traces.
+
+To disable logging, set the `TRACELOOP_TRACE_CONTENT` environment variable to `false`.
+
+```bash
+TRACELOOP_TRACE_CONTENT=false
+```

packages/opentelemetry-instrumentation-openai_agents/opentelemetry/instrumentation/openai_agents/__init__.py

@@ -0,0 +1,8 @@
+"""OpenTelemetry OpenAI Agents instrumentation"""
+
+from opentelemetry.instrumentation.openai_agents.version import __version__
+from opentelemetry.instrumentation.openai_agents.instrumentation import (
+    OpenAIAgentsInstrumentor,
+)
+
+__all__ = ["OpenAIAgentsInstrumentor", "__version__"]

packages/opentelemetry-instrumentation-openai_agents/opentelemetry/instrumentation/openai_agents/instrumentation.py

@@ -0,0 +1,87 @@
+from typing import Collection
+
+from wrapt import wrap_function_wrapper
+from opentelemetry.trace import SpanKind, get_tracer, Tracer
+from opentelemetry.trace.status import Status, StatusCode
+from opentelemetry.instrumentation.utils import unwrap
+from opentelemetry.instrumentation.instrumentor import BaseInstrumentor
+from opentelemetry.instrumentation.openai_agents.version import __version__
+from opentelemetry.semconv_ai import SpanAttributes
+from .patch import (
+    extract_agent_details,
+    set_model_settings_span_attributes,
+    extract_run_config_details,
+    set_prompt_attributes,
+    set_response_content_span_attribute,
+    set_token_usage_span_attributes,
+)
+
+
+_instruments = ("openai_agents >= 0.0.2",)
+
+
+class OpenAIAgentsInstrumentor(BaseInstrumentor):
+    """An instrumentor for OpenAI Agents SDK."""
+
+    def instrumentation_dependencies(self) -> Collection[str]:
+        return _instruments
+
+    def _instrument(self, **kwargs):
+        tracer_provider = kwargs.get("tracer_provider")
+        tracer = get_tracer(__name__, __version__, tracer_provider)
+
+        wrap_function_wrapper(
+            "agents.run", "Runner._get_new_response", _wrap_agent_run(tracer)
+        )
+
+    def _uninstrument(self, **kwargs):
+        unwrap("agents.run.Runner", "_get_new_response")
+
+
+def with_tracer_wrapper(func):
+
+    def _with_tracer(tracer):
+        def wrapper(wrapped, instance, args, kwargs):
+            return func(tracer, wrapped, instance, args, kwargs)
+
+        return wrapper
+
+    return _with_tracer
+
+
+@with_tracer_wrapper
+async def _wrap_agent_run(tracer: Tracer, wrapped, instance, args, kwargs):
+    agent = args[0]
+    agent_name = getattr(agent, "name", "agent")
+    model_name = getattr(getattr(agent, "model", None), "model",
+                         "unknown_model")
+
+    with tracer.start_as_current_span(
+        f"openai_agents.{agent_name}",
+        kind=SpanKind.INTERNAL,
+        attributes={
+            SpanAttributes.LLM_SYSTEM: "openai",
+            SpanAttributes.LLM_REQUEST_MODEL: model_name,
+        },
+    ) as span:
+        try:
+            extract_agent_details(agent, span)
+            set_model_settings_span_attributes(agent, span)
+
+            if len(args) > 7:
+                extract_run_config_details(args[7], span)
+
+            if len(args) > 2:
+                set_prompt_attributes(span, args[2])
+
+            response = await wrapped(*args, **kwargs)
+
+            set_response_content_span_attribute(response, span)
+            set_token_usage_span_attributes(response, span)
+
+            span.set_status(Status(StatusCode.OK))
+            return response
+
+        except Exception as e:
+            span.set_status(Status(StatusCode.ERROR, str(e)))
+            raise

packages/opentelemetry-instrumentation-openai_agents/opentelemetry/instrumentation/openai_agents/patch.py

@@ -0,0 +1,119 @@
+from opentelemetry.semconv_ai import SpanAttributes
+
+
+def extract_agent_details(test_agent, span):
+    if test_agent is None:
+        return
+    agent = getattr(test_agent, "agent", test_agent)
+    if agent is None:
+        return
+
+    attributes = {}
+    agent_dict = vars(agent)
+
+    for key, value in agent_dict.items():
+        if value is not None and isinstance(value, (str, int, float, bool)):
+            attributes[f"openai.agent.{key}"] = value
+        # Optional: handle known short lists
+        elif isinstance(value, list) and len(value) != 0:
+            attributes[f"openai.agent.{key}_count"] = len(value)
+
+    if attributes:
+        span.set_attributes(attributes)
+
+
+def set_model_settings_span_attributes(agent, span):
+
+    if not hasattr(agent, "model_settings") or agent.model_settings is None:
+        return
+
+    model_settings = agent.model_settings
+    settings_dict = vars(model_settings)
+
+    key_to_span_attr = {
+        "max_tokens": SpanAttributes.LLM_REQUEST_MAX_TOKENS,
+        "temperature": SpanAttributes.LLM_REQUEST_TEMPERATURE,
+        "top_p": SpanAttributes.LLM_REQUEST_TOP_P,
+    }
+
+    for key, value in settings_dict.items():
+        if value is not None:
+            span_attr = key_to_span_attr.get(key, f"openai.agent.model.{key}")
+            span.set_attribute(span_attr, value)
+
+
+def extract_run_config_details(run_config, span):
+    if run_config is None:
+        return
+
+    config_dict = vars(run_config)
+    attributes = {}
+
+    for key, value in config_dict.items():
+
+        if value is not None and isinstance(value, (str, int, float, bool)):
+            attributes[f"openai.agent.{key}"] = value
+        elif isinstance(value, list) and len(value) != 0:
+            attributes[f"openai.agent.{key}_count"] = len(value)
+
+    if attributes:
+        span.set_attributes(attributes)
+
+
+def set_prompt_attributes(span, message_history):
+    if not message_history:
+        return
+
+    for msg in message_history:
+        if isinstance(msg, dict):
+
+            role = msg.get("role", "user")
+            content = msg.get("content", "")
+
+    span.set_attribute(f"{SpanAttributes.LLM_PROMPTS}.role", role)
+    span.set_attribute(f"{SpanAttributes.LLM_PROMPTS}.content", content)
+
+
+def set_response_content_span_attribute(response, span):
+
+    if hasattr(response, "output") and isinstance(response.output, list):
+        for output_message in response.output:
+            # Extract role and type from output_message
+            role = getattr(output_message, "role", None)
+            msg_type = getattr(output_message, "type", None)
+
+            if role:
+                span.set_attribute(
+                    f"{SpanAttributes.LLM_COMPLETIONS}.role", role)
+            if msg_type:
+                span.set_attribute(
+                    f"{SpanAttributes.LLM_COMPLETIONS}.type", msg_type)
+
+            if hasattr(output_message, "content") and isinstance(
+                output_message.content, list
+            ):
+                for content_item in output_message.content:
+                    if hasattr(content_item, "text"):
+                        span.set_attribute(
+                            f"{SpanAttributes.LLM_COMPLETIONS}.content",
+                            content_item.text,
+                        )
+
+
+def set_token_usage_span_attributes(response, span):
+    if hasattr(response, "usage"):
+        usage = response.usage
+        input_tokens = getattr(usage, "input_tokens", None)
+        output_tokens = getattr(usage, "output_tokens", None)
+        total_tokens = getattr(usage, "total_tokens", None)
+
+        if input_tokens is not None:
+            span.set_attribute(
+                SpanAttributes.LLM_USAGE_PROMPT_TOKENS, input_tokens)
+        if output_tokens is not None:
+            span.set_attribute(
+                SpanAttributes.LLM_USAGE_COMPLETION_TOKENS, output_tokens
+            )
+        if total_tokens is not None:
+            span.set_attribute(
+                SpanAttributes.LLM_USAGE_TOTAL_TOKENS, total_tokens)

packages/opentelemetry-instrumentation-openai_agents/opentelemetry/instrumentation/openai_agents/version.py

@@ -0,0 +1 @@
+__version__ = "0.40.7"