I’ve open-sourced Agent Lint (GitHub repo name: zhupanov/agent-lint), a linter for coding-agent configuration:
It currently validates configuration for:
- Codex:
.codex/,.codex-plugin/,AGENTS.md - Agent skills:
.agents/skills/ - MCP
- Claude Code
- Cursor
It has roughly 300 lint rules, all backed by a documented rule registry.
A few other features:
- configurable rule severity and suppression
- run individual rules with
--only - GitHub Actions integration
- prebuilt Linux and macOS binaries
- implemented in Rust
The motivation is that agent configuration is starting to become a real software surface of its own. As these configurations grow, it gets easier to accumulate invalid fields, inconsistencies, stale assumptions, or subtle mistakes that the underlying agent may simply interpret differently than intended.
I’m particularly interested in feedback from people using Codex in real projects:
- What Codex / AGENTS.md mistakes have you run into?
- Are there useful checks that Agent Lint doesn’t cover yet?
- Are any existing rules too strict or likely to produce false positives?
- Are there configuration patterns that work in practice but aren’t well documented?
Feedback, issues, and contributions are very welcome.