Agent compatibility

Purpose: Separate officially discovered behavior, shared conventions, and portable fallbacks.

Intended reader: Maintainers installing FPAT Lite for Codex, Claude Code, or another coding agent.

Consult this when: Editing agent-specific files or troubleshooting discovery.

Compatibility matrix#

Capability	Codex	Claude Code	FPAT Lite choice
Repository instructions	`AGENTS.md`; hierarchical scope	`CLAUDE.md`; hierarchical scope	Shared rules in `AGENTS.md`; `CLAUDE.md` imports it
Repository skill	`.agents/skills/<name>/SKILL.md`	`.claude/skills/<name>/SKILL.md`	One thin router in each native location
Explicit invocation	`$fpat-lite ...` or skill selector	`/fpat-lite ...`	Same mode and arguments
Reusable command files	Codex custom prompts are deprecated and user-home scoped	`.claude/commands/` works but is legacy	`commands/` are portable contracts, not native commands
Permission control	Sandbox, approvals, configuration, experimental exec rules	Permission modes, allow/ask/deny rules, hooks	No silent grants; document approval boundary
Shared detailed knowledge	Skill references and ordinary repository files	Skill references and ordinary repository files	Shared `commands/`, `rules/`, `templates/`, `knowledge-base/`
Other agents	No guaranteed native discovery	No guaranteed native discovery	Read `AGENTS.md` then a named command contract

Codex behavior used#

Current Codex guidance documents that:

Codex builds an AGENTS.md instruction chain from global guidance and the project root down to the working directory; closer files appear later and take precedence.
Repository skills are discovered under .agents/skills from the working directory through the repository root.
Skills require a SKILL.md with name and description, and support progressive disclosure.
Custom prompts are deprecated in favor of skills and live in the user's Codex home rather than the repository.

Sources:

Claude Code behavior used#

Current Claude Code documentation states that:

Project instructions may live in ./CLAUDE.md or ./.claude/CLAUDE.md.
Claude Code reads CLAUDE.md, not AGENTS.md, but CLAUDE.md can import AGENTS.md with @AGENTS.md.
Project skills live at .claude/skills/<skill-name>/SKILL.md and can be invoked as /skill-name.
.claude/commands/ continues to work, but skills are the recommended current format.
Instructions shape behavior; permission rules and hooks provide enforcement.

Sources:

Instruction precedence differences#

Both products combine broader and narrower guidance, but their details are not identical:

Codex checks AGENTS.override.md, then AGENTS.md, at each applicable directory and uses one file per directory.
Claude Code loads CLAUDE.md and CLAUDE.local.md; project instructions appear after user instructions, and subdirectory files can load on demand.
Claude imports are native Claude behavior. Codex is not told to interpret the @AGENTS.md line; it reads AGENTS.md directly.

Do not describe either product's exact precedence as a universal Agent Skills standard.

Permission guidance#

FPAT Lite deliberately omits checked-in allowed-tools, auto-approval rules, and broad platform settings. Such files can alter the user's security posture and are environment-specific.

Teams that need hard enforcement should configure the platform directly:

Codex: sandbox, approval policy, trusted project config, and organization policy.
Claude Code: permission modes, allow/ask/deny rules, or hooks.
GitHub: least-privilege tokens, branch protection, required checks, and reviewer policy.

Keep the Markdown approval rule even when enforcement exists; it explains the intent and improves handoffs.

Portable fallback#

When native skill discovery is unavailable, use:

Read AGENTS.md, rules/core-rules.md, and
rules/safety-and-approval.md. Then follow commands/<mode>.md for this task.