[Observability] add metadata to track llm request for sessions (#7381)

Co-authored-by: Robert Brennan <accounts@rbren.io>

openhands/agenthub/codeact_agent/codeact_agent.py

@@ -119,6 +119,8 @@ class CodeActAgent(Agent):
             'messages': self.llm.format_messages_for_llm(messages),
         }
         params['tools'] = self.tools
+        # log to litellm proxy if possible
+        params['extra_body'] = {'metadata': state.to_llm_metadata(agent_name=self.name)}
         response = self.llm.completion(**params)
         actions = codeact_function_calling.response_to_actions(response)
         for action in actions:

openhands/controller/agent_controller.py

@@ -605,6 +605,7 @@ class AgentController:
         llm = LLM(config=llm_config, retry_listener=self._notify_on_llm_retry)
         delegate_agent = agent_cls(llm=llm, config=agent_config)
         state = State(
+            session_id=self.id.removesuffix('-delegate'),
             inputs=action.inputs or {},
             local_iteration=0,
             iteration=self.state.iteration,
@@ -873,6 +874,7 @@ class AgentController:
         # If state is None, we create a brand new state and still load the event stream so we can restore the history
         if state is None:
             self.state = State(
+                session_id=self.id.removesuffix('-delegate'),
                 inputs={},
                 max_iterations=max_iterations,
                 confirmation_mode=confirmation_mode,

openhands/controller/state/state.py

@@ -1,9 +1,11 @@
 import base64
+import os
 import pickle
 from dataclasses import dataclass, field
 from enum import Enum
 from typing import Any
 
+import openhands
 from openhands.controller.state.task import RootTask
 from openhands.core.logger import openhands_logger as logger
 from openhands.core.schema import AgentState
@@ -71,6 +73,7 @@ class State:
     """
 
     root_task: RootTask = field(default_factory=RootTask)
+    session_id: str = ''
     # global iteration for the current task
     iteration: int = 0
     # local iteration for the current subtask
@@ -201,3 +204,14 @@ class State:
             if isinstance(event, MessageAction) and event.source == EventSource.USER:
                 return event
         return None
+
+    def to_llm_metadata(self, agent_name: str) -> dict:
+        return {
+            'session_id': self.session_id,
+            'trace_version': openhands.__version__,
+            'tags': [
+                f'agent:{agent_name}',
+                f'web_host:{os.environ.get("WEB_HOST", "unspecified")}',
+                f'openhands_version:{openhands.__version__}',
+            ],
+        }

openhands/llm/llm.py

@@ -235,9 +235,16 @@ class LLM(RetryMixin, DebugMixin):
             # NOTE: this setting is global; unlike drop_params, it cannot be overridden in the litellm completion partial
             litellm.modify_params = self.config.modify_params
 
+            # if we're not using litellm proxy, remove the extra_body
+            if 'litellm_proxy' not in self.config.model:
+                kwargs.pop('extra_body', None)
+
             # Record start time for latency measurement
             start_time = time.time()
             # we don't support streaming here, thus we get a ModelResponse
+            logger.debug(
+                f'LLM: calling litellm completion with model: {self.config.model}, base_url: {self.config.base_url}, args: {args}, kwargs: {kwargs}'
+            )
             resp: ModelResponse = self._completion_unwrapped(*args, **kwargs)
 
             # Calculate and record latency