Changelog

All notable changes to adk-smart-router are documented in this file.

The format is based on Keep a Changelog, and this project adheres to Semantic Versioning.

[1.0.1] - 2026-10-13

Added

• Environment-driven model governance (load_config() and MIN_CACHE_TOKENS):
  • load_config() now reads SMART_ROUTER_EMBEDDING_MODEL, SMART_ROUTER_EMBEDDING_DIMS, and SMART_ROUTER_LLM_JUDGE_MODEL from environment variables so deployments in model-governed Gemini Enterprise environments (constraints/vertexai.allowedModels) can pin approved embedding and judge model IDs without code changes.
  • Added "gemini-flash-latest" and "gemini-pro-latest" (4096 tokens) to MIN_CACHE_TOKENS.
• Documentation & ADK ecosystem metadata:
  • Updated README.md (## Using it) to demonstrate wiring SmartRouterPlugin alongside BigQueryAgentAnalyticsPlugin(..., custom_metadata_allowlist=["smart_router"]) and ContextCacheConfig.
  • Updated platform terminology across documentation and skills/adk-smart-router/SKILL.md to Gemini Enterprise and aligned plugin.json ownership metadata (cloud-gtm/agentic-data-science).
  • Added Google Contributor License Agreement and Community Guidelines sections to CONTRIBUTING.md.

[1.0.0] - 2026-10-12

Added

• Runner-level SmartRouterPlugin (adk_smart_router):
  • Turn-one prompt embedding (gemini-embedding-2 on Vertex AI global) and session-pinned tool declaration pruning (PrefixLock) that preserves Gemini implicit context caching across multi-turn conversations.
  • Three composable judge tiers (RouterConfig.judges / SMART_ROUTER_JUDGES):
    • "prototype" (PrototypeJudge): Zero-training-data cold-start semantic router using tool docstring + JSON schema embeddings, catalog mean-centering, 1-hop upstream Returns: $\to$ Args: prerequisite graph propagation, and optional zero-data canonical menu synthesis (cold_start_menu_count).
    • "xgboost" (Predictor): In-memory XGBoost booster exported from BigQuery ML (model.ubj + metadata.json) with greedy max-coverage canonical menu snapping.
    • ("xgboost", "prototype"): Hybrid combiner sharing a single Turn-1 prompt embedding across both judges to score trained classes via XGBoost and rescue/prune newly deployed (classless) tools via PrototypeJudge.
  • Context-cache floor protection (MIN_CACHE_TOKENS + CACHE_FLOOR_HEADROOM) and live CacheWatch automatic circuit breaker comparing routed vs. holdout billed input tokens.
  • Principal-scoped in-memory ToolResultCache (cacheable_tools) keyed by SHA-256 over (app_name, user_id, agent_name, tool_name, tool_version, canonical_args), with automatic read-your-own-writes invalidation and runtime EventActions mutation blocklisting.
  • Startup pre-warming (await plugin.warmup(app)) for embedder connection pooling and tool-catalog prototype embeddings.
• smart-router CLI (adk_smart_router.cli):
  • smart-router qualify: Pre-provisioning offline prompt-floor (PAYS, THIN, UNCACHED, TRAP) and Returns: $\to$ Args: argument-reachability audit.
  • smart-router readiness: BigQuery Agent Analytics corpus readiness check before training.
  • smart-router doctor: Static wiring verification for plugin registration order, custom_metadata_allowlist, ContextCacheConfig, and active judge configuration.
  • smart-router check-arguments: Empirical audit of --arguments markings against logged first-turn user messages.
  • smart-router status: Live model artifact, 7-day billed-token delta vs. holdout, feature drift, and retrain ledger inspection.
• BigQuery ML Training & Operations Suite (sql/ and scripts/):
  • 5-stage BigQuery ML training stored procedure (sql/train_router.sql), dashboard views (sql/dashboard.sql), and autonomous daily retraining/withdrawal cycle (scripts/run_cycle.sh, scripts/publish_router.py, scripts/tune_settings.py, scripts/watch_generations.py).
• Agent Plugin & Skill (plugin.json and skills/adk-smart-router/SKILL.md):
  • Google Agent Plugins v1.0.0 manifest and 5-step agentic integration playbook for AI coding assistants.