AI Coding Agent Control Plane V2.2 Released

AI Workflow Control Plane V2.2 has been released to improve the non-deterministic behavior of AI coding agents by separating task routing, retrieval, context assembly, execution, and verification. Instead of applying maximum context and the strongest model to all tasks, this tool dynamically allocates resources based on risk and task type.

Key Features

• Retrieval and Context Management: Uses a heterogeneous ranking fusion approach combining BM25 RRF and MMR, selecting optimal contexts via submodular facility location selection within a hard context budget.
• Structural Analysis: Analyzes call relationships, dependencies, test coverage, and Blast Radius through the Code Review Graph (CRG). Simple searches utilize indexes, while structural questions leverage graphs.
• Execution and Verification: The 'Superpowers' module handles planning, TDD, sub-agent execution, code review, and verification, while the control plane generates bounded orchestration contracts for these tasks.
• Safety: Supports an 'abstain' state to prevent agents from attempting modifications when evidence is insufficient, and verifies context freshness by grounding Git HEAD and file hashes.

Benchmarks and Features

A benchmark suite comprising 32 test cases is provided, covering accuracy, semantic, structural, and high-risk queries. It measures evidence state accuracy alongside standard metrics such as P@k, Recall@k, MRR, and nDCG. The system supports Python AST-aware indexing and multi-repository search, offering deterministic fallback mechanisms when external services are unavailable.