`Statewright`: visual state machines to constrain coding agents

Reliability improved by shrinking the agent's action space instead of scaling model size. Statewright enforces phase-by-phase tools and transitions, which makes smaller coding models more usable and cuts wasted tokens.

• —Tool access is split by state: planning gets read-only tools, implementation gets scoped edits, testing gets test-only bash. That blocks step-skipping at the protocol level.
• —Reported gains show up across qwen-coder, gpt-oss, and gemma4, with a useful threshold around 13B+ params. Below that, context retention breaks edits.
• —Frontier models also benefit: Haiku, Sonnet, and Opus reportedly solve more reliably with fewer tokens. Better context discipline beats dumping everything into one prompt.
• —The stack is a Rust state-machine engine plus MCP plugins for Claude Code, with Codex and Cursor planned. This is an orchestration-layer play, not a new model.