[Release] google/adk-python v2.2.0: v2.2.0

ADK Python v2.2.0 introduces breaking changes including a default model migration from gemini-2.5-flash to gemini-3-flash-preview and terminology updates from "turns" to "steps" for GenAI SDK v2.0.0 compatibility. The release adds significant features like AutoTracingPlugin for OpenTelemetry, RubricBasedMultiTurnTrajectoryEvaluator, improved agent analytics, and enhanced tool capabilities. Multiple bug fixes address security vulnerabilities (CVE-2026-48710, path traversal), session management, Live protocol support for Gemini 3.1, and MCP initialization issues.

Key Points

• •Default LlmAgent model changed from gemini-2.5-flash to gemini-3-flash-preview (preview model) ahead of 2026-10-16 shutdown; set model="gemini-2.5-flash" explicitly to retain prior behavior
• •GenAI SDK v2.0.0 support renames turn-based helpers (e.g., convert_contents_to_turns → convert_contents_to_steps); update direct callers to use new "steps" terminology
• •Added AutoTracingPlugin for OpenTelemetry auto-instrumentation and native OTel gen_ai.client.* metrics emission
• •New RubricBasedMultiTurnTrajectoryEvaluator for trajectory evaluation and improved agent analytics with BigQuery reliability fixes
• •Enhanced tool capabilities: expose httpx_client_factory on RestApiTool/OpenAPIToolset, standardize request_input tool for proactive LLM clarification, support additional scopes in Google API Tools
• •Security fixes: address CVE-2026-48710 (Starlette/FastAPI), block path traversal in Agent Builder file tools, prevent Zip Slip in GCS skill extraction
• •Live protocol improvements: resolve 1007 error and support Gemini 3.1 Flash Live protocol with turn_complete_reason for safety info capture
• •Session management enhancements: add get_user_state(app_name, user_id) to BaseSessionService, exclude temp state keys from Firestore writes, preserve transparent config on live reconnect
• •API server improvements: abort runs on client drops to avoid leaks, forward custom_metadata from run requests into run config, include thoughts/tool calls in compaction summaries
• •MCP and deployment enhancements: prevent initialization hangs and task group leaks, add --trigger_sources and ADK service options to cli_deploy_agent_engine

Workflow Diagram

Concepts