Victor Del Puerto

Tools

Open-source tools I build, mostly around making AI coding agents behave the way a senior engineer would.

Everything here is Apache 2.0 and runs locally. No SaaS, no login, no remote telemetry.

fscars — bolt-on correction primitive for AI coding agents

pip install fscars

Turn a correction you keep repeating to Claude Code into a deterministic hook that fires the next time the same situation comes back. The framework comes from my paper Lucy Syndrome in LLM Agents; the package is the clean reference implementation.

Claude Code and Codex both ship today with native hooks — one entrypoint wired into each agent’s hook config, so a scar fires (and can block a tool call before it runs) deterministically, not as a suggestion the model may ignore. The adapter API is public for community contributions.

callus — per-author voice calibration

pip install callus

Score a draft against your own writing instead of a generic AI detector, and rewrite toward your voice. It measures distance from your raw corpus plus the density of specific AI tells — not a probability of being machine-generated, which matters because detectors flag non-native English writers at a 61% false-positive rate (Stanford, 2023). Calibrated per author, runs in English and Spanish.

lucy-syndrome — research framework + production case

The companion repository to the paper. Includes the framework templates, 11 scars + 9 hooks from a real civil-engineering operation, the Phase 2 logging primitive, and the Phase 3 opportunity observer.

It is not a package — you read it, you adapt the templates, and you wire your own hooks. Use fscars if you want a one-line install instead.

More tools land here as they stabilize. If something is documented in a blog post but not yet here, it is still in alpha and not ready for outside adoption.