feat: Otel instrumentation version 3 (#3021)

Co-authored-by: Alex Hall <[email protected]>

Author: bitnahian

pydantic_ai_slim/pydantic_ai/_agent_graph.py

@@ -16,6 +16,7 @@
 from typing_extensions import TypeVar, assert_never
 
 from pydantic_ai._function_schema import _takes_ctx as is_takes_ctx  # type: ignore
+from pydantic_ai._instrumentation import DEFAULT_INSTRUMENTATION_VERSION
 from pydantic_ai._tool_manager import ToolManager
 from pydantic_ai._utils import dataclasses_no_defaults_repr, get_union_args, is_async_callable, run_in_executor
 from pydantic_ai.builtin_tools import AbstractBuiltinTool
@@ -704,6 +705,9 @@ def build_run_context(ctx: GraphRunContext[GraphAgentState, GraphAgentDeps[DepsT
         tracer=ctx.deps.tracer,
         trace_include_content=ctx.deps.instrumentation_settings is not None
         and ctx.deps.instrumentation_settings.include_content,
+        instrumentation_version=ctx.deps.instrumentation_settings.version
+        if ctx.deps.instrumentation_settings
+        else DEFAULT_INSTRUMENTATION_VERSION,
         run_step=ctx.state.run_step,
         tool_call_approved=ctx.state.run_step == 0,
     )

pydantic_ai_slim/pydantic_ai/_instrumentation.py

@@ -0,0 +1,95 @@
+from __future__ import annotations
+
+from dataclasses import dataclass
+from typing import TYPE_CHECKING
+
+if TYPE_CHECKING:
+    from typing_extensions import Self
+
+DEFAULT_INSTRUMENTATION_VERSION = 2
+"""Default instrumentation version for `InstrumentationSettings`."""
+
+
+@dataclass(frozen=True)
+class InstrumentationNames:
+    """Configuration for instrumentation span names and attributes based on version."""
+
+    # Agent run span configuration
+    agent_run_span_name: str
+    agent_name_attr: str
+
+    # Tool execution span configuration
+    tool_span_name: str
+    tool_arguments_attr: str
+    tool_result_attr: str
+
+    # Output Tool execution span configuration
+    output_tool_span_name: str
+
+    @classmethod
+    def for_version(cls, version: int) -> Self:
+        """Create instrumentation configuration for a specific version.
+
+        Args:
+            version: The instrumentation version (1, 2, or 3+)
+
+        Returns:
+            InstrumentationConfig instance with version-appropriate settings
+        """
+        if version <= 2:
+            return cls(
+                agent_run_span_name='agent run',
+                agent_name_attr='agent_name',
+                tool_span_name='running tool',
+                tool_arguments_attr='tool_arguments',
+                tool_result_attr='tool_response',
+                output_tool_span_name='running output function',
+            )
+        else:
+            return cls(
+                agent_run_span_name='invoke_agent',
+                agent_name_attr='gen_ai.agent.name',
+                tool_span_name='execute_tool',  # Will be formatted with tool name
+                tool_arguments_attr='gen_ai.tool.call.arguments',
+                tool_result_attr='gen_ai.tool.call.result',
+                output_tool_span_name='execute_tool',
+            )
+
+    def get_agent_run_span_name(self, agent_name: str) -> str:
+        """Get the formatted agent span name.
+
+        Args:
+            agent_name: Name of the agent being executed
+
+        Returns:
+            Formatted span name
+        """
+        if self.agent_run_span_name == 'invoke_agent':
+            return f'invoke_agent {agent_name}'
+        return self.agent_run_span_name
+
+    def get_tool_span_name(self, tool_name: str) -> str:
+        """Get the formatted tool span name.
+
+        Args:
+            tool_name: Name of the tool being executed
+
+        Returns:
+            Formatted span name
+        """
+        if self.tool_span_name == 'execute_tool':
+            return f'execute_tool {tool_name}'
+        return self.tool_span_name
+
+    def get_output_tool_span_name(self, tool_name: str) -> str:
+        """Get the formatted output tool span name.
+
+        Args:
+            tool_name: Name of the tool being executed
+
+        Returns:
+            Formatted span name
+        """
+        if self.output_tool_span_name == 'execute_tool':
+            return f'execute_tool {tool_name}'
+        return self.output_tool_span_name

pydantic_ai_slim/pydantic_ai/_output.py

@@ -11,6 +11,8 @@
 from pydantic_core import SchemaValidator, to_json
 from typing_extensions import Self, TypedDict, TypeVar, assert_never
 
+from pydantic_ai._instrumentation import InstrumentationNames
+
 from . import _function_schema, _utils, messages as _messages
 from ._run_context import AgentDepsT, RunContext
 from .exceptions import ModelRetry, ToolRetryError, UserError
@@ -95,6 +97,7 @@ async def execute_traced_output_function(
         ToolRetryError: When wrap_validation_errors is True and a ModelRetry is caught
         ModelRetry: When wrap_validation_errors is False and a ModelRetry occurs
     """
+    instrumentation_names = InstrumentationNames.for_version(run_context.instrumentation_version)
     # Set up span attributes
     tool_name = run_context.tool_name or getattr(function_schema.function, '__name__', 'output_function')
     attributes = {
@@ -104,18 +107,29 @@ async def execute_traced_output_function(
     if run_context.tool_call_id:
         attributes['gen_ai.tool.call.id'] = run_context.tool_call_id
     if run_context.trace_include_content:
-        attributes['tool_arguments'] = to_json(args).decode()
-        attributes['logfire.json_schema'] = json.dumps(
-            {
-                'type': 'object',
-                'properties': {
-                    'tool_arguments': {'type': 'object'},
-                    'tool_response': {'type': 'object'},
-                },
-            }
-        )
+        attributes[instrumentation_names.tool_arguments_attr] = to_json(args).decode()
+
+    attributes['logfire.json_schema'] = json.dumps(
+        {
+            'type': 'object',
+            'properties': {
+                **(
+                    {
+                        instrumentation_names.tool_arguments_attr: {'type': 'object'},
+                        instrumentation_names.tool_result_attr: {'type': 'object'},
+                    }
+                    if run_context.trace_include_content
+                    else {}
+                ),
+                'gen_ai.tool.name': {},
+                **({'gen_ai.tool.call.id': {}} if run_context.tool_call_id else {}),
+            },
+        }
+    )
 
-    with run_context.tracer.start_as_current_span('running output function', attributes=attributes) as span:
+    with run_context.tracer.start_as_current_span(
+        instrumentation_names.get_output_tool_span_name(tool_name), attributes=attributes
+    ) as span:
         try:
             output = await function_schema.call(args, run_context)
         except ModelRetry as r:
@@ -135,7 +149,7 @@ async def execute_traced_output_function(
             from .models.instrumented import InstrumentedModel
 
             span.set_attribute(
-                'tool_response',
+                instrumentation_names.tool_result_attr,
                 output if isinstance(output, str) else json.dumps(InstrumentedModel.serialize_any(output)),
             )
 

pydantic_ai_slim/pydantic_ai/_run_context.py

@@ -8,6 +8,8 @@
 from opentelemetry.trace import NoOpTracer, Tracer
 from typing_extensions import TypeVar
 
+from pydantic_ai._instrumentation import DEFAULT_INSTRUMENTATION_VERSION
+
 from . import _utils, messages as _messages
 
 if TYPE_CHECKING:
@@ -36,6 +38,8 @@ class RunContext(Generic[AgentDepsT]):
     """The tracer to use for tracing the run."""
     trace_include_content: bool = False
     """Whether to include the content of the messages in the trace."""
+    instrumentation_version: int = DEFAULT_INSTRUMENTATION_VERSION
+    """Instrumentation settings version, if instrumentation is enabled."""
     retries: dict[str, int] = field(default_factory=dict)
     """Number of retries for each tool so far."""
     tool_call_id: str | None = None

pydantic_ai_slim/pydantic_ai/_tool_manager.py

@@ -12,6 +12,7 @@
 from typing_extensions import assert_never
 
 from . import messages as _messages
+from ._instrumentation import InstrumentationNames
 from ._run_context import AgentDepsT, RunContext
 from .exceptions import ModelRetry, ToolRetryError, UnexpectedModelBehavior
 from .messages import ToolCallPart
@@ -115,6 +116,7 @@ async def handle_call(
                 wrap_validation_errors,
                 self.ctx.tracer,
                 self.ctx.trace_include_content,
+                self.ctx.instrumentation_version,
                 usage_limits,
             )
 
@@ -203,15 +205,18 @@ async def _call_tool_traced(
         allow_partial: bool,
         wrap_validation_errors: bool,
         tracer: Tracer,
-        include_content: bool = False,
+        include_content: bool,
+        instrumentation_version: int,
         usage_limits: UsageLimits | None = None,
     ) -> Any:
         """See <https://opentelemetry.io/docs/specs/semconv/gen-ai/gen-ai-spans/#execute-tool-span>."""
+        instrumentation_names = InstrumentationNames.for_version(instrumentation_version)
+
         span_attributes = {
             'gen_ai.tool.name': call.tool_name,
             # NOTE: this means `gen_ai.tool.call.id` will be included even if it was generated by pydantic-ai
             'gen_ai.tool.call.id': call.tool_call_id,
-            **({'tool_arguments': call.args_as_json_str()} if include_content else {}),
+            **({instrumentation_names.tool_arguments_attr: call.args_as_json_str()} if include_content else {}),
             'logfire.msg': f'running tool: {call.tool_name}',
             # add the JSON schema so these attributes are formatted nicely in Logfire
             'logfire.json_schema': json.dumps(
@@ -220,8 +225,8 @@ async def _call_tool_traced(
                     'properties': {
                         **(
                             {
-                                'tool_arguments': {'type': 'object'},
-                                'tool_response': {'type': 'object'},
+                                instrumentation_names.tool_arguments_attr: {'type': 'object'},
+                                instrumentation_names.tool_result_attr: {'type': 'object'},
                             }
                             if include_content
                             else {}
@@ -232,18 +237,21 @@ async def _call_tool_traced(
                 }
             ),
         }
-        with tracer.start_as_current_span('running tool', attributes=span_attributes) as span:
+        with tracer.start_as_current_span(
+            instrumentation_names.get_tool_span_name(call.tool_name),
+            attributes=span_attributes,
+        ) as span:
             try:
                 tool_result = await self._call_tool(call, allow_partial, wrap_validation_errors, usage_limits)
             except ToolRetryError as e:
                 part = e.tool_retry
                 if include_content and span.is_recording():
-                    span.set_attribute('tool_response', part.model_response())
+                    span.set_attribute(instrumentation_names.tool_result_attr, part.model_response())
                 raise e
 
             if include_content and span.is_recording():
                 span.set_attribute(
-                    'tool_response',
+                    instrumentation_names.tool_result_attr,
                     tool_result
                     if isinstance(tool_result, str)
                     else _messages.tool_return_ta.dump_json(tool_result).decode(),

pydantic_ai_slim/pydantic_ai/agent/__init__.py

@@ -14,6 +14,7 @@
 from pydantic.json_schema import GenerateJsonSchema
 from typing_extensions import Self, TypeVar, deprecated
 
+from pydantic_ai._instrumentation import DEFAULT_INSTRUMENTATION_VERSION, InstrumentationNames
 from pydantic_graph import Graph
 
 from .. import (
@@ -644,11 +645,16 @@ async def get_instructions(run_context: RunContext[AgentDepsT]) -> str | None:
         )
 
         agent_name = self.name or 'agent'
+        instrumentation_names = InstrumentationNames.for_version(
+            instrumentation_settings.version if instrumentation_settings else DEFAULT_INSTRUMENTATION_VERSION
+        )
+
         run_span = tracer.start_span(
-            'agent run',
+            instrumentation_names.get_agent_run_span_name(agent_name),
             attributes={
                 'model_name': model_used.model_name if model_used else 'no-model',
                 'agent_name': agent_name,
+                'gen_ai.agent.name': agent_name,
                 'logfire.msg': f'{agent_name} run',
             },
         )

pydantic_ai_slim/pydantic_ai/models/instrumented.py

@@ -21,6 +21,8 @@
 from opentelemetry.util.types import AttributeValue
 from pydantic import TypeAdapter
 
+from pydantic_ai._instrumentation import DEFAULT_INSTRUMENTATION_VERSION
+
 from .. import _otel_messages
 from .._run_context import RunContext
 from ..messages import (
@@ -90,7 +92,7 @@ class InstrumentationSettings:
     event_mode: Literal['attributes', 'logs'] = 'attributes'
     include_binary_content: bool = True
     include_content: bool = True
-    version: Literal[1, 2] = 1
+    version: Literal[1, 2, 3] = DEFAULT_INSTRUMENTATION_VERSION
 
     def __init__(
         self,
@@ -99,7 +101,7 @@ def __init__(
         meter_provider: MeterProvider | None = None,
         include_binary_content: bool = True,
         include_content: bool = True,
-        version: Literal[1, 2] = 2,
+        version: Literal[1, 2, 3] = DEFAULT_INSTRUMENTATION_VERSION,
         event_mode: Literal['attributes', 'logs'] = 'attributes',
         event_logger_provider: EventLoggerProvider | None = None,
     ):

tests/models/test_fallback.py

@@ -169,6 +169,7 @@ def test_first_failed_instrumented(capfire: CaptureLogfire) -> None:
                 'attributes': {
                     'model_name': 'fallback:function:failure_response:,function:success_response:',
                     'agent_name': 'agent',
+                    'gen_ai.agent.name': 'agent',
                     'logfire.msg': 'agent run',
                     'logfire.span_type': 'span',
                     'gen_ai.usage.input_tokens': 51,
@@ -268,6 +269,7 @@ async def test_first_failed_instrumented_stream(capfire: CaptureLogfire) -> None
                 'attributes': {
                     'model_name': 'fallback:function::failure_response_stream,function::success_response_stream',
                     'agent_name': 'agent',
+                    'gen_ai.agent.name': 'agent',
                     'logfire.msg': 'agent run',
                     'logfire.span_type': 'span',
                     'gen_ai.usage.input_tokens': 50,
@@ -375,6 +377,7 @@ def test_all_failed_instrumented(capfire: CaptureLogfire) -> None:
                 'attributes': {
                     'model_name': 'fallback:function:failure_response:,function:failure_response:',
                     'agent_name': 'agent',
+                    'gen_ai.agent.name': 'agent',
                     'logfire.msg': 'agent run',
                     'logfire.span_type': 'span',
                     'pydantic_ai.all_messages': [{'role': 'user', 'parts': [{'type': 'text', 'content': 'hello'}]}],