feat: Add intra-invocation compaction and token compaction pre-request

Compact session events before LLM calls when token threshold is exceeded

Co-authored-by: George Weale <gweale@google.com>
PiperOrigin-RevId: 873095899

src/google/adk/agents/invocation_context.py

@@ -24,6 +24,7 @@ from pydantic import ConfigDict
 from pydantic import Field
 from pydantic import PrivateAttr
 
+from ..apps.app import EventsCompactionConfig
 from ..apps.app import ResumabilityConfig
 from ..artifacts.base_artifact_service import BaseArtifactService
 from ..auth.credential_service.base_credential_service import BaseCredentialService
@@ -200,6 +201,12 @@ class InvocationContext(BaseModel):
   resumability_config: Optional[ResumabilityConfig] = None
   """The resumability config that applies to all agents under this invocation."""
 
+  events_compaction_config: Optional[EventsCompactionConfig] = None
+  """The compaction config for this invocation."""
+
+  token_compaction_checked: bool = False
+  """Whether token-threshold compaction ran during this invocation."""
+
   plugin_manager: PluginManager = Field(default_factory=PluginManager)
   """The manager for keeping track of plugins in this invocation."""
 

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

@@ -0,0 +1,58 @@
+# Copyright 2026 Google LLC
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+"""Request processor that runs token-threshold event compaction."""
+
+from __future__ import annotations
+
+from typing import AsyncGenerator
+from typing import TYPE_CHECKING
+
+from ...apps.compaction import _has_token_threshold_config
+from ...apps.compaction import _run_compaction_for_token_threshold_config
+from ...events.event import Event
+from ._base_llm_processor import BaseLlmRequestProcessor
+
+if TYPE_CHECKING:
+  from ...agents.invocation_context import InvocationContext
+  from ...models.llm_request import LlmRequest
+
+
+class CompactionRequestProcessor(BaseLlmRequestProcessor):
+  """Compacts session events before contents are prepared for model calls."""
+
+  async def run_async(
+      self, invocation_context: InvocationContext, llm_request: LlmRequest
+  ) -> AsyncGenerator[Event, None]:
+    del llm_request
+    config = invocation_context.events_compaction_config
+    if not _has_token_threshold_config(config):
+      return
+      yield  # Required for AsyncGenerator.
+
+    token_compacted = await _run_compaction_for_token_threshold_config(
+        config=config,
+        session=invocation_context.session,
+        session_service=invocation_context.session_service,
+        agent=invocation_context.agent,
+        agent_name=invocation_context.agent.name,
+        current_branch=invocation_context.branch,
+    )
+    if token_compacted:
+      invocation_context.token_compaction_checked = True
+    return
+    yield  # Required for AsyncGenerator.
+
+
+request_processor = CompactionRequestProcessor()

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

@@ -22,6 +22,7 @@ from . import _code_execution
 from . import _nl_planning
 from . import _output_schema_processor
 from . import basic
+from . import compaction
 from . import contents
 from . import context_cache_processor
 from . import identity
@@ -42,6 +43,9 @@ def _create_request_processors():
       request_confirmation.request_processor,
       instructions.request_processor,
       identity.request_processor,
+      # Compaction should run before contents so compacted events are reflected
+      # in the model request context.
+      compaction.request_processor,
       contents.request_processor,
       # Context cache processor sets up cache config and finds
       # existing cache metadata.

src/google/adk/runners.py

@@ -553,7 +553,10 @@ class Runner:
         if self.app and self.app.events_compaction_config:
           logger.debug('Running event compactor.')
           await _run_compaction_for_sliding_window(
-              self.app, session, self.session_service
+              self.app,
+              session,
+              self.session_service,
+              skip_token_compaction=invocation_context.token_compaction_checked,
           )
 
     async with Aclosing(_run_with_trace(new_message, invocation_id)) as agen:
@@ -1362,6 +1365,9 @@ class Runner:
         credential_service=self.credential_service,
         plugin_manager=self.plugin_manager,
         context_cache_config=self.context_cache_config,
+        events_compaction_config=(
+            self.app.events_compaction_config if self.app else None
+        ),
         invocation_id=invocation_id,
         agent=self.agent,
         session=session,