AI agent supervisor·Developer Tools·2026

Spec-to-State

From specification to supervised execution.

Working PrototypeNEW

Currently unavailable to view: this project is being extended with new functionality, so access is temporarily paused.

Tech Stack

PythonPydanticSQLiteDAG PlanningSHA-256Pytest

About

Spec-to-State is a runtime supervisor designed to manage autonomous AI coding agents reliably. It parses software specifications and constructs a DAG of dependent tasks, then supervises agent execution using configurable policy gates that validate outputs before allowing the agent to proceed. Each completed step is logged as a SHA-256-signed evidence entry in a tamper-evident ledger, enabling full auditability. The system supports checkpoints and recovery so that failed tasks can restart from a known-good state rather than from scratch.

The Problem

Autonomous AI coding agents frequently fail silently, produce incorrect outputs, or lose track of state across long multi-step tasks. There is no standard mechanism to validate agent output before allowing execution to continue.

The Approach

Spec-to-State introduces a DAG-based task planner, independent verification oracles, cryptographic evidence logging, and policy gates that block progression when verification fails.

Key Features

  • Dependency-aware task graph from specification (DAG planner)
  • Policy gates with independent verification oracles
  • SHA-256 tamper-evident evidence ledger
  • Checkpoint and recovery for failed tasks
  • Pytest-based verification framework
  • Pydantic-typed configuration and data models

Highlights

  • DAG-based planner decomposes specs into atomic, verifiable tasks.
  • Cryptographic evidence ledger creates an auditable execution trail.