Status: Adopted
Category: Standards Track
Organization: fountain.coach
Applies to: All Codex-operated repositories
Version: 1.0.0

This document defines the Fountain Codex Instruction System (FCIS), a normative, organization-wide standard for governing how OpenAI Codex is instructed, constrained, and enabled across repositories.

FCIS establishes four orthogonal instruction layers:

1. AGENTS.md — behavioral law
2. PLANS.md — intent and execution reasoning
3. Skills (SKILL.md) — reusable execution techniques
4. MCP servers — external capabilities and data access

Strict separation of these layers is required to ensure predictability, auditability, and scalability of LLM-driven development.

The key words MUST, MUST NOT, REQUIRED, SHALL, SHOULD, SHOULD NOT, RECOMMENDED, MAY, and OPTIONAL in this document are to be interpreted as described in RFC 2119.

Within FCIS, orthogonality means:

Instruction layers are independent dimensions of responsibility, such that modifying one layer does not redefine, weaken, or invalidate the others.

Orthogonality is a hard requirement, not a stylistic preference.

LLM-based systems degrade rapidly when:

• responsibilities overlap
• priorities are implicit
• instructions conflict probabilistically

Orthogonality prevents these failure modes by making responsibility explicit, exclusive, and reviewable.

Each instruction layer answers exactly one question.

No layer MAY answer another layer’s question.

AGENTS.md MUST define non-negotiable behavioral constraints under which Codex operates in a repository.

AGENTS.md MUST include:

• invariants and prohibitions
• safety and review requirements
• routing rules for PLANS.md and skills

AGENTS.md MUST NOT:

• contain step-by-step procedures
• include tool configuration details
• encode execution workflows

Violation constitutes a law breach.

PLANS.md MUST capture the intent, structure, and risk management of multi-step work.

AGENTS.md MUST define when PLANS.md is required. Codex SHALL NOT execute qualifying tasks without a plan.

A plan SHOULD include:

• goal
• constraints
• phased approach
• test plan
• rollback plan
• definition of done

A skill MUST:

• exist in its own directory
• contain exactly one SKILL.md

A skill MUST define:

• when it applies
• ordered execution steps
• output guarantees

Skills MUST NOT:

• redefine behavioral law
• replace PLANS.md for complex work
• configure MCP servers

MCP servers MUST provide access to external tools or data only.

They MUST NOT:

• encode workflow logic
• enforce policy
• substitute for skills

MCP configuration MUST occur outside repositories. Repositories MUST NOT rely on MCP presence for correctness.

The following are spec violations:

• procedures in AGENTS.md
• behavioral rules in skills
• workflow logic in MCP
• unplanned long-running tasks
• duplicated responsibility across layers

Violations SHOULD block merges.

repo-root/
  AGENTS.md
  PLANS.md
  .codex/
    skills/
      <skill-name>/
        SKILL.md

User-level configuration:

~/.codex/config.toml
~/.codex/AGENTS.override.md

A repository is FCIS-compliant if:

• all applicable layers are present
• responsibilities are orthogonal
• plans exist for complex work
• skills are modular and declarative

FCIS treats Codex as a compiler-like system:

• law → intent → execution → capability This mirrors successful separation patterns from compilers, operating systems, and safety-critical automation.

Non-orthogonal instruction systems increase:

• unintended behavior
• silent policy drift
• non-reproducible execution

FCIS reduces these risks through explicit separation and review.

• RFC 0001 defines FCIS v1.0.0
• Backward-incompatible changes require a new RFC

FCIS establishes a contractual architecture for Codex:

• AGENTS.md defines law
• PLANS.md defines intent
• Skills define execution
• MCP defines capability

Adherence is mandatory for all fountain.coach Codex repositories.