AGENTS.md — ChatGPT / CodeX

This folder targets CodeX, the coding agent that powers ChatGPT Work, ChatGPT desktop, and the CodeX CLI. CodeX reads AGENTS.md files for rules and scans .agents/skills/ for skills (project root down to the working directory), so no .chatgpt folder is needed — rules live here in AGENTS.md and skills in .agents/skills/. For global use, copy the skills to ~/.agents/skills/ and this file to ~/.codex/AGENTS.md.

Language (Always English)

Always answer in English — even when the user writes or asks in another language, or explicitly requests a reply in that language. Work and information gathering always happen in English.

Safety Rules

Git safety

Never commit, push, or create a PR without explicit user approval.

• Explicit means the user literally says "commit this", "push it", or "submit a PR".
• Not explicit: "go ahead", "do it", "apply it", "fix it", "sounds good", "yes".
• Showing diffs, staging, and suggesting commit messages are always OK. When uncertain, ask.

Irreversible operations

Before any operation that can cause irreversible loss of data, state, configuration, or system behavior (migration, schema change, data manipulation, rm/drop/delete, config edits), ensure a backup exists or the change is reversible. Prefer additive, reversible changes.

Agent Behavior

• Inspect before editing — never modify a file you have not read.
• Do not assume requirements — when ambiguity affects architecture or behavior, ask before coding.
• Smallest change that satisfies the requirement; preserve existing behavior; avoid unrelated refactoring.
• Match existing patterns over inventing new ones; prefer fewer dependencies.
• Verify before uncertain changes — check official documentation or source instead of guessing.
• Report honestly — never present unverified work as done; say plainly when something is partial or failed.

Engineering Preferences

• Commits: short, meaningful Conventional Commits — <type>: <short description> with feat, fix, revert, chore, docs. Keep commits small, focused, and atomic.
• Naming: names describe actual behavior — avoid handle(), process(), doSomething(); prefer fetchUserProfile(), validatePaymentToken().
• Explicit over implicit: make contracts visible, name things clearly, define data shapes, avoid hidden behavior.
• Simple over clever: simple solutions over flexible ones; maintainability over cleverness; fewer dependencies.
• No speculative features (YAGNI) — build only what is needed now.
• File size: keep files under ~1000 lines; split into focused modules when approaching the limit.

Available Skills

Skills live in .agents/skills/ (29 portable skills, standard SKILL.md format; the ZCode-only context-compact and session-handoff ship only in zcode/). See README.md for the "what it does / why it was created" table.