fix(a2a): forward get_session_config in _prepare_session()

[vnykdfd]

Link to Issue

Closes: #6410

Problem Description

Environment

• ADK Version: 2.4.0 (latest main branch)
• Python: 3.10-3.14 (tested on 3.13)
• Component: A2A executor (legacy path)

Bug Summary

_prepare_session() in A2aAgentExecutor ignores GetSessionConfig(num_recent_events=N) by not forwarding it to session_service.get_session(), causing the entire event history to be loaded instead of respecting the configured limit.

Reproduction

from google.adk.agents.run_config import RunConfig
from google.adk.sessions.base_session_service import GetSessionConfig

# Configure to load only 10 recent events
run_config = RunConfig(
    get_session_config=GetSessionConfig(num_recent_events=10)
)

# Bug: _prepare_session() calls get_session() WITHOUT config parameter
# Result: All events loaded (e.g., 291 events instead of 10)

Expected Behavior

Session service should load at most 10 events when num_recent_events=10 is configured.

Observed Behavior

Session service loads ALL events regardless of get_session_config setting.

Root Cause

_prepare_session() (line 320-324) calls session_service.get_session() without passing config= parameter:

# BEFORE (missing config parameter)
session = await runner.session_service.get_session(
    app_name=runner.app_name,
    user_id=user_id,
    session_id=session_id,
    # ← Missing: config=get_session_config
)

Impact

For sessions with hundreds of events:

• 10-100x more database I/O than configured
• 200-300ms added latency per request (observed in production)
• Higher memory usage (~300KB vs ~10KB)
• Higher token costs (sending excessive context to LLM)

Solution

Extract get_session_config from run_request.run_config and forward it to session_service.get_session():

# AFTER (with fix)
get_session_config = None
if run_request.run_config:
    get_session_config = run_request.run_config.get_session_config

session = await runner.session_service.get_session(
    app_name=runner.app_name,
    user_id=user_id,
    session_id=session_id,
    config=get_session_config,  # ← ADDED
)

Testing Plan

Unit Tests Added

• test_prepare_session_passes_get_session_config - Verifies config is forwarded correctly
• test_prepare_session_none_run_config - Handles run_config=None gracefully
• test_prepare_session_no_get_session_config - Handles missing get_session_config

Test Results

Local pytest (Python 3.13):

$ pytest tests/unittests/a2a/executor/test_a2a_agent_executor.py -v
==================== 29 passed, 1 skipped in 1.50s ====================

Full A2A test suite:

$ pytest tests/unittests/a2a/ -v
==================== 372 passed, 17 skipped in 2.54s ====================

Pre-commit hooks:

$ pre-commit run --all-files
All checks passed ✅

Manual E2E Verification

Reproduction script confirms the fix works (script available on request).

Changes Made

Files Modified

1. src/google/adk/a2a/executor/a2a_agent_executor.py (+4 lines)

   • Extract get_session_config from run request
   • Pass it to session_service.get_session(config=...)

2. tests/unittests/a2a/executor/test_a2a_agent_executor.py (+103 insertions, -14 deletions)

   • Added 3 new unit tests
   • Fixed existing tests: replaced Mock(spec=RunConfig) with RunConfig() instances
     (Pydantic v2 fields not accessible via class-level Mock spec)

Backward Compatibility

✅ Fully backward compatible - no breaking changes
✅ Existing behavior unchanged when get_session_config is None
✅ All 372 A2A tests pass

Checklist

• Read CONTRIBUTING.md
• Self-review performed
• Added code comments for clarity
• Added unit tests proving fix works
• Tests pass locally
• Pre-commit hooks pass
• License headers present on all files
• No dependent changes required
• CLA signed

Additional Context

This bug only affects the legacy A2aAgentExecutor path. The newer _A2aAgentExecutor implementation already handles session loading more efficiently by passing num_recent_events=0 during existence checks.

Tested locally on Python 3.13. CI will verify compatibility with Python 3.10-3.14.

[guglielmo-san]

Hi @vnykdfd, this change is targeting the legacy a2a_agent_executor which we no longer support. Can you try to use the a2a_agent_executor_impl and see if in that version this issue still exists?

[vdoifode-equinix]

Hi @guglielmo-san

Thanks for clarifying.....

Yes, I can confirm the issue does exist in a2a_agent_executor_impl as well.

As per src/google/adk/a2a/executor/a2a_agent_executor_impl.py (lines 265-279):

async def _resolve_session(
    self,
    run_request: AgentRunRequest,
    runner: Runner,
):
  session_id = run_request.session_id
  user_id = run_request.user_id
  session = await runner.session_service.get_session(
      app_name=runner.app_name,
      user_id=user_id,
      session_id=session_id,
      # Checking existence doesn't require event history.
      config=base_session_service.GetSessionConfig(num_recent_events=0),  # ← Hardcoded to 0
  )

The bug _resolve_session() never extracts or forwards run_request.run_config.get_session_config - it always hardcodes num_recent_events=0.

Impact: Sessions always load with 0 events regardless of caller configuration, causing the same performance issue (database loads all events, then filters in-memory).

Should I close this PR and submit a new one targeting _A2aAgentExecutor._resolve_session() in a2a_agent_executor_impl.py instead?

[guglielmo-san]

You're right, and there is some parallel internal work on slightly changing the way the session is retrieved and shared to the runner in the a2a integration. This work is also fixing this code

# Checking existence doesn't require event history.
      config=base_session_service.GetSessionConfig(num_recent_events=0),  # ← Hardcoded to 0

You might wait to open your PR as it should be fixed in the coming releases

[vdoifode-equinix]

Got it.
Thanks