I build tools that keep AI coding agents in check: the boring checks that stop an agent from touching your secrets or calling a task done while the tests fail. I also run small model experiments and publish the results even when they don't work.

Reflex Control is my main project: a Rust CLI that plugs into Claude Code, Cursor, Codex CLI, OpenCode, Kilo Code, Cline and pi as a hook. So far it has run live in a real Claude Code session; the adapters for the other agents are checked against their source or docs but not yet run in a real session.

• Blocks writes to files you protect (like .env or secrets/**).
• Asks before risky shell commands (like a force push).
• Runs your tests at the end of a turn when the agent says it is done, sending failures back to the agent a limited number of times.

curl -fsSL https://raw.githubusercontent.com/rustfuture/reflex-control/main/install.sh | sh
export PATH="$HOME/.local/bin:$PATH"
reflex demo agent

Run reflex install to launch the interactive setup wizard.

Repository · Design · Evaluation

• Rust Agent Runtime: runs coding tasks with command allowlists, timeouts and required verification, and records task state in an append-only log.
• Repository Intelligence: searches local code and answers with exact source lines instead of prose; works offline. Quotes are verbatim; relevance is not guaranteed.
• grainx: displays CPU, memory, disk, network and processes in the terminal; can share readings over a local HTTP service and export JSON/CSV.
• deltasafe: sends files between computers on the same local network and checks that each file arrived unchanged.
• RustHound: reads log files and reports lines that match rules or look unusual.

Each of them has a prebuilt CLI on its release page, with an install script in its README.

• Model Adaptation Lab: I tried LoRA fine-tuning on a small model (Qwen2.5-Coder-1.5B) to explain Rust compiler errors; the recorded run showed no gain (report).
• RLT-RSI Experiment: I tested reusing model layers in a loop on a sequence parity task; the runs stayed at chance level.
• RSI Experimental Framework: a loop that proposes and keeps keyword rules for text classification on a small synthetic dataset; in the recorded run held-out accuracy went from 3/8 to 8/8 (+0.29 on average over 3 seeds), on a toy dataset rather than real-world data.

If you run into a bug or have ideas, please open an issue on any repo.