feat(runner): add metadata parameter to Runner.run_async()

src/google/adk/agents/invocation_context.py

@@ -206,6 +206,17 @@ class InvocationContext(BaseModel):
   canonical_tools_cache: Optional[list[BaseTool]] = None
   """The cache of canonical tools for this invocation."""
 
+  metadata: Optional[dict[str, Any]] = None
+  """Per-request metadata passed from Runner entry points.
+
+  This field allows passing arbitrary metadata that can be accessed during
+  the invocation lifecycle, particularly in callbacks like before_model_callback.
+  Common use cases include passing user_id, trace_id, memory context keys, or
+  other request-specific context that needs to be available during processing.
+
+  Supported entry points: run(), run_async(), run_live(), run_debug().
+  """
+
   _invocation_cost_manager: _InvocationCostManager = PrivateAttr(
       default_factory=_InvocationCostManager
   )

src/google/adk/flows/llm_flows/base_llm_flow.py

@@ -86,7 +86,7 @@ async def run_live(
       invocation_context: InvocationContext,
   ) -> AsyncGenerator[Event, None]:
     """Runs the flow using live api."""
-    llm_request = LlmRequest()
+    llm_request = LlmRequest(metadata=invocation_context.metadata)
     event_id = Event.new_id()
 
     # Preprocess before calling the LLM.
@@ -392,7 +392,7 @@ async def _run_one_step_async(
       invocation_context: InvocationContext,
   ) -> AsyncGenerator[Event, None]:
     """One step means one LLM call."""
-    llm_request = LlmRequest()
+    llm_request = LlmRequest(metadata=invocation_context.metadata)
 
     # Preprocess before calling the LLM.
     async with Aclosing(

src/google/adk/models/llm_request.py

@@ -15,6 +15,7 @@
 from __future__ import annotations
 
 import logging
+from typing import Any
 from typing import Optional
 from typing import Union
 
@@ -99,6 +100,15 @@ class LlmRequest(BaseModel):
   the full history.
   """
 
+  metadata: Optional[dict[str, Any]] = None
+  """Per-request metadata for callbacks and custom processing.
+
+  This field allows passing arbitrary metadata from the Runner.run_async()
+  call to callbacks like before_model_callback. This is useful for passing
+  request-specific context such as user_id, trace_id, or memory context keys
+  that need to be available during model invocation.
+  """
+
   def append_instructions(
       self, instructions: Union[list[str], types.Content]
   ) -> list[types.Content]:

src/google/adk/runners.py

@@ -392,6 +392,7 @@ def run(
       session_id: str,
       new_message: types.Content,
       run_config: Optional[RunConfig] = None,
+      metadata: Optional[dict[str, Any]] = None,
   ) -> Generator[Event, None, None]:
     """Runs the agent.
 
@@ -409,6 +410,7 @@ def run(
       session_id: The session ID of the session.
       new_message: A new message to append to the session.
       run_config: The run config for the agent.
+      metadata: Optional per-request metadata that will be passed to callbacks.
 
     Yields:
       The events generated by the agent.
@@ -424,6 +426,7 @@ async def _invoke_run_async():
                 session_id=session_id,
                 new_message=new_message,
                 run_config=run_config,
+                metadata=metadata,
             )
         ) as agen:
           async for event in agen:
@@ -459,6 +462,7 @@ async def run_async(
       new_message: Optional[types.Content] = None,
       state_delta: Optional[dict[str, Any]] = None,
       run_config: Optional[RunConfig] = None,
+      metadata: Optional[dict[str, Any]] = None,
   ) -> AsyncGenerator[Event, None]:
     """Main entry method to run the agent in this runner.
 
@@ -476,6 +480,13 @@ async def run_async(
       new_message: A new message to append to the session.
       state_delta: Optional state changes to apply to the session.
       run_config: The run config for the agent.
+      metadata: Optional per-request metadata that will be passed to callbacks.
+        This allows passing request-specific context such as user_id, trace_id,
+        or memory context keys to before_model_callback and other callbacks.
+        Note: A shallow copy is made of this dictionary, so top-level changes
+        within callbacks won't affect the original. However, modifications to
+        nested mutable objects (e.g., nested dicts or lists) will affect the
+        original.
 
     Yields:
       The events generated by the agent.
@@ -485,13 +496,16 @@ async def run_async(
         new_message are None.
     """
     run_config = run_config or RunConfig()
+    # Create a shallow copy to isolate from caller's modifications
+    metadata = metadata.copy() if metadata is not None else None
 
     if new_message and not new_message.role:
       new_message.role = 'user'
 
     async def _run_with_trace(
         new_message: Optional[types.Content] = None,
         invocation_id: Optional[str] = None,
+        metadata: Optional[dict[str, Any]] = None,
     ) -> AsyncGenerator[Event, None]:
       with tracer.start_as_current_span('invocation'):
         session = await self._get_or_create_session(
@@ -519,6 +533,7 @@ async def _run_with_trace(
               invocation_id=invocation_id,
               run_config=run_config,
               state_delta=state_delta,
+              metadata=metadata,
           )
           if invocation_context.end_of_agents.get(
               invocation_context.agent.name
@@ -532,6 +547,7 @@ async def _run_with_trace(
               new_message=new_message,  # new_message is not None.
               run_config=run_config,
               state_delta=state_delta,
+              metadata=metadata,
           )
 
         async def execute(ctx: InvocationContext) -> AsyncGenerator[Event]:
@@ -558,7 +574,9 @@ async def execute(ctx: InvocationContext) -> AsyncGenerator[Event]:
               self.app, session, self.session_service
           )
 
-    async with Aclosing(_run_with_trace(new_message, invocation_id)) as agen:
+    async with Aclosing(
+        _run_with_trace(new_message, invocation_id, metadata)
+    ) as agen:
       async for event in agen:
         yield event
 
@@ -928,6 +946,7 @@ async def run_live(
       live_request_queue: LiveRequestQueue,
       run_config: Optional[RunConfig] = None,
       session: Optional[Session] = None,
+      metadata: Optional[dict[str, Any]] = None,
   ) -> AsyncGenerator[Event, None]:
     """Runs the agent in live mode (experimental feature).
 
@@ -969,6 +988,7 @@ async def run_live(
         run_config: The run config for the agent.
         session: The session to use. This parameter is deprecated, please use
           `user_id` and `session_id` instead.
+        metadata: Optional per-request metadata that will be passed to callbacks.
 
     Yields:
         AsyncGenerator[Event, None]: An asynchronous generator that yields
@@ -983,6 +1003,7 @@ async def run_live(
         Either `session` or both `user_id` and `session_id` must be provided.
     """
     run_config = run_config or RunConfig()
+    metadata = metadata.copy() if metadata is not None else None
     # Some native audio models requires the modality to be set. So we set it to
     # AUDIO by default.
     if run_config.response_modalities is None:
@@ -1008,6 +1029,7 @@ async def run_live(
         session,
         live_request_queue=live_request_queue,
         run_config=run_config,
+        metadata=metadata,
     )
 
     root_agent = self.agent
@@ -1160,6 +1182,7 @@ async def run_debug(
       run_config: RunConfig | None = None,
       quiet: bool = False,
       verbose: bool = False,
+      metadata: dict[str, Any] | None = None,
   ) -> list[Event]:
     """Debug helper for quick agent experimentation and testing.
 
@@ -1183,6 +1206,7 @@ async def run_debug(
           shown).
         verbose: If True, shows detailed tool calls and responses. Defaults to
           False for cleaner output showing only final agent responses.
+        metadata: Optional per-request metadata that will be passed to callbacks.
 
     Returns:
         list[Event]: All events from all messages.
@@ -1245,6 +1269,7 @@ async def run_debug(
           session_id=session.id,
           new_message=types.UserContent(parts=[types.Part(text=message)]),
           run_config=run_config,
+          metadata=metadata,
       ):
         if not quiet:
           print_event(event, verbose=verbose)
@@ -1260,6 +1285,7 @@ async def _setup_context_for_new_invocation(
       new_message: types.Content,
       run_config: RunConfig,
       state_delta: Optional[dict[str, Any]],
+      metadata: Optional[dict[str, Any]] = None,
   ) -> InvocationContext:
     """Sets up the context for a new invocation.
 
@@ -1268,6 +1294,7 @@ async def _setup_context_for_new_invocation(
       new_message: The new message to process and append to the session.
       run_config: The run config of the agent.
       state_delta: Optional state changes to apply to the session.
+      metadata: Optional per-request metadata to pass to callbacks.
 
     Returns:
       The invocation context for the new invocation.
@@ -1277,6 +1304,7 @@ async def _setup_context_for_new_invocation(
         session,
         new_message=new_message,
         run_config=run_config,
+        metadata=metadata,
     )
     # Step 2: Handle new message, by running callbacks and appending to
     # session.
@@ -1299,6 +1327,7 @@ async def _setup_context_for_resumed_invocation(
       invocation_id: Optional[str],
       run_config: RunConfig,
       state_delta: Optional[dict[str, Any]],
+      metadata: Optional[dict[str, Any]] = None,
   ) -> InvocationContext:
     """Sets up the context for a resumed invocation.
 
@@ -1308,6 +1337,7 @@ async def _setup_context_for_resumed_invocation(
       invocation_id: The invocation id to resume.
       run_config: The run config of the agent.
       state_delta: Optional state changes to apply to the session.
+      metadata: Optional per-request metadata to pass to callbacks.
 
     Returns:
       The invocation context for the resumed invocation.
@@ -1333,6 +1363,7 @@ async def _setup_context_for_resumed_invocation(
         new_message=user_message,
         run_config=run_config,
         invocation_id=invocation_id,
+        metadata=metadata,
     )
     # Step 3: Maybe handle new message.
     if new_message:
@@ -1377,6 +1408,7 @@ def _new_invocation_context(
       new_message: Optional[types.Content] = None,
       live_request_queue: Optional[LiveRequestQueue] = None,
       run_config: Optional[RunConfig] = None,
+      metadata: Optional[dict[str, Any]] = None,
   ) -> InvocationContext:
     """Creates a new invocation context.
 
@@ -1386,6 +1418,7 @@ def _new_invocation_context(
         new_message: The new message for the context.
         live_request_queue: The live request queue for the context.
         run_config: The run config for the context.
+        metadata: Optional per-request metadata for the context.
 
     Returns:
         The new invocation context.
@@ -1417,6 +1450,7 @@ def _new_invocation_context(
         live_request_queue=live_request_queue,
         run_config=run_config,
         resumability_config=self.resumability_config,
+        metadata=metadata,
     )
 
   def _new_invocation_context_for_live(
@@ -1425,6 +1459,7 @@ def _new_invocation_context_for_live(
       *,
       live_request_queue: LiveRequestQueue,
       run_config: Optional[RunConfig] = None,
+      metadata: Optional[dict[str, Any]] = None,
   ) -> InvocationContext:
     """Creates a new invocation context for live multi-agent."""
     run_config = run_config or RunConfig()
@@ -1443,6 +1478,7 @@ def _new_invocation_context_for_live(
         session,
         live_request_queue=live_request_queue,
         run_config=run_config,
+        metadata=metadata,
     )
 
   async def _handle_new_message(