Scoped honestly, because "eco mode" sounds harmless and two of these are not nothing:

• It reads and proposes edits to your Codex configuration. eco setup shows a diff of $CODEX_HOME/config.toml and writes nothing until you confirm. The shipped profiles under profiles/ are files you copy yourself; nothing installs them for you.
• The installers copy skill files into your skills directory ($HOME/.agents/skills by default, or $CODEX_SKILLS_DIR). An existing directory of the same name is moved to a timestamped backup first, never overwritten in place. --uninstall / -Uninstall removes only the directories this project installed.
• The skills instruct the agent, they do not execute anything themselves. eco and eco-max are Markdown instruction files with no bundled scripts and no network access.
• The benchmark harness runs codex exec on a throwaway fixture in a temporary git repository it creates, and writes results under bench/results/. It never touches your working tree. It uses a separate CODEX_HOME when you pass --codex-home, which is how the published studies were run.

No telemetry, no network calls of its own, no dependencies: package.json declares zero, and CI asserts that.

• It does not write to config.toml without an explicit confirmation.
• It does not install hooks. Codex hooks are configured in your own config.toml; this project documents them but ships none, because a hook runs a process on every matching tool call and that is not a cost anyone should inherit from an install.
• It does not report a dollar figure. The Codex event stream contains no cost field, and inventing one from a price table would be a fabrication.

Open a GitHub issue for anything non-sensitive. For something you would rather not post publicly, use GitHub's private vulnerability reporting on this repository.