HEXIS

What it does

• compile / update / run / guide CLI: a model drafts the machine; static checks (schema, guards, reachability, termination, clause coverage) bounce drafts back until they pass
• Trace-driven evolution: one decision per step, and a candidate is accepted only if it replays every previously accepted trace — a wrong decision cannot break the machine
• efsm-v1 machines are plain JSON with guarded transitions, bounded loops and a fallback state; guards run in a whitelisted expression language, never eval
• Every build ships GUIDE.md and PROMPT.md so any tool-using agent (e.g. Claude Code) can execute the machine state by state; 350+ hermetic tests

When to use it

Skill authors who want deterministic ordering, bounded loops and verifiable termination for multi-step procedures, and researchers studying agent reliability — the repo accompanies a paper on compiling skills into EFSMs.

Getting started

pip install hexis-agent          # imported as `hexis`; command is `hexis-agent`; Python >= 3.11
export API_KEY=...               # any OpenAI-compatible endpoint
M="--model <model> --base-url https://your-endpoint/v1"

hexis-agent compile --skill path/to/skill --out build/ $M
hexis-agent update --build build/ --traces traces/ $M
hexis-agent run --machine build/ --input request="..." --workdir work/ --executor local $M

Things to know

MIT, alpha status. Executes tool calls whose arguments — including shell commands — are generated by language models: run it in an isolated environment and only with machines and task files you trust (SECURITY.md). update sends trace steps and the skill's clauses to the configured model endpoint, and build directories keep copies of traces and every model question. Native tool execution uses OpenCode; --executor local runs bash in a subprocess instead. Compiling and updating consume model calls at roughly one call per trace step.