Conventions Codex

Codex CLI Conventions

Tool-specific overlay for OpenAI Codex CLI artifacts. Loaded by the scorer and checker when an artifact is classified as Tier 2-Codex (per agents/scorer.md step 3). The universal floor lives in nlpm:conventions; this overlay adds Codex-specific schemas on top.

Primary authoritative sources:

• https://learn.chatgpt.com/docs (the developers.openai.com/codex/* tree now 308-redirects here)
• https://learn.chatgpt.com/docs/build-skills
• https://learn.chatgpt.com/docs/agent-configuration/agents-md
• https://learn.chatgpt.com/docs/agent-configuration/subagents
• https://learn.chatgpt.com/docs/config-file/config-reference
• https://learn.chatgpt.com/docs/hooks
• https://learn.chatgpt.com/docs/skills-and-plugins
• https://learn.chatgpt.com/docs/changelog
• https://github.com/openai/codex (authoritative for CLA.md, contributing.md, and machine-readable spec samples like plugin-json-spec.md)

1. File system layout

Codex separates the cross-tool surface (.agents/) from the Codex-private surface (.codex/). The Claude Code mental model of "everything under one tool directory" does not transfer.

Trust gate: Project hooks load only when .codex/ is trusted. Trust is enforced via /hooks and allow_managed_hooks_only in requirements.toml.

2. SKILL.md (Tier 1, open spec — name, description required only)

Codex reads SKILL.md from .agents/skills/, not .codex/skills/ (reconfirmed 2026-08-02 against build-skills: "Codex reads skills from .agents/skills directories, not .codex/skills"; it scans .agents/skills from CWD up to the repo root, plus ~/.agents/skills and /etc/codex/skills). Skills placed under .codex/skills/ — a common mistake in repos ported from Claude Code's .claude/skills/ habit — are not discovered by Codex; flag them. The required frontmatter is the agentskills.io baseline — name and description — same as every other tool.

Codex-specific extras live in a SIDECAR agents/openai.yaml, not in SKILL.md frontmatter. Treat the sidecar as additive metadata, not a deviation from the open spec.

Sidecar fields:

# <skill>/agents/openai.yaml
interface:
  display_name: "My Skill"
  short_description: "One-line summary"      # added 2026-06
  default_prompt: "Use this skill when..."
  icon_small: "icon-16.png"
  icon_large: "icon-512.png"                 # added 2026-06
  brand_color: "#FF6B00"
policy:
  allow_implicit_invocation: true            # default true; set false to disable auto-selection
dependencies:
  tools:
    - type: "bash"
      value: "jq"
      description: "JSON processor"
      transport: "stdio"

Correction (2026-06-07): dependencies.tools is an array of tool objects (type, value, description, transport), not a bare string list. The v0.1.0 - bash / - jq form was wrong.

Duplicate-name skills across scopes are NOT merged — both appear in selectors, repo-level wins for local workflows.

3. .codex-plugin/plugin.json (plugin manifest)

{
  "name": "my-codex-plugin",
  "version": "1.0.0",
  "description": "Short summary",
  "skills": "./skills/",
  "mcpServers": "./mcp/servers.json",
  "apps": "./apps/",
  "hooks": "./hooks.json",
  "interface": {
    "displayName": "My Plugin",
    "longDescription": "Detailed description for installer UI"
  }
}

Required field: name (kebab-case) — and only when a plugin.json is present at all. All other top-level fields are optional, including version, description, author, and interface (corrected 2026-08-02 against the vendor's plugin-json-spec.md; the earlier "version + description required" claim was wrong). Optional artifact paths — each a single relative-path STRING, not an array (corrected 2026-08-04 against the vendor plugin-json-spec.md): skills (string), hooks (string), apps (string), mcpServers (string or object). A bare directory string like "skills": "./skills/" is the documented sample form; these paths supplement Codex's default discovery, they do not replace it. Do NOT flag a scalar-string skills/hooks/apps/mcpServers as wrong — the array form the v0.3.0 overlay showed was incorrect. Optional identity fields (added 2026-06): author ({name, email, url}), homepage, repository, license, keywords. Optional UI block interface:

• displayName, shortDescription, longDescription, developerName, category, capabilities
• defaultPrompt — an array of starter prompts (not a single string). At most 3 entries (extras ignored); each capped at 128 characters (longer entries truncated).
• websiteURL, privacyPolicyURL, termsOfServiceURL
• brandColor, composerIcon, logo, logoDark (dark-mode logo variant), screenshots (PNG files stored under ./assets/, paths relative to plugin root)

4. .agents/plugins/marketplace.json (marketplace manifest)

Three marketplace tiers exist in Codex:

• Official Curated — OpenAI-managed. Workspace plugin publishing and additional marketplaces (Amazon Bedrock, Claude Code interop) shipped in 0.146.0 (2026-07-29); the exact self-serve mechanics are still settling — verify before asserting.
• Repository — <repo-root>/.agents/plugins/marketplace.json aggregates plugins shipped from that repo.
• Personal — ~/.agents/plugins/marketplace.json.

Schema:

{
  "name": "xiaolai",
  "interface": {
    "displayName": "xiaolai Marketplace"
  },
  "plugins": [
    {
      "name": "nlpm",
      "source": {
        "source": "github",
        "repo": "xiaolai/nlpm"
      },
      "policy": {
        "installation": "AVAILABLE",
        "authentication": "ON_USE"
      },
      "category": "developer-tools",
      "interface": {
        "displayName": "nlpm"
      }
    }
  ]
}

Per-plugin source types: "github", "git", "local".

policy values are UPPERCASE enums (corrected 2026-08-04 against nlpm's own shipping .agents/plugins/marketplace.json and observed real plugins): installation e.g. "AVAILABLE"; authentication e.g. "ON_USE" or "ON_INSTALL". The lowercase "auto"/"none" the v0.3.0 overlay showed was wrong — do NOT flag AVAILABLE/ON_USE/ON_INSTALL as invalid.

5. .codex/config.toml

TOML — NOT JSON. Top-level sections nlpm cares about:

• [features] — feature flags. Breaking change ~2026-04 (CLI 0.129+): [features].codex_hooks was renamed to [features].hooks (boolean; enables hooks from hooks.json or inline [hooks]). Old key is a deprecated alias and emits a warning. Flag config files that still use codex_hooks.
• [mcp_servers.<id>] — MCP server registrations. Fields: command, args, cwd, url, enabled, enabled_tools, disabled_tools, env, startup_timeout_sec, tool_timeout_sec (per-tool, default 60s — added 2026-06).
• [hooks.<event>] — inline hook registrations (alternative to .codex/hooks.json).
• [agents] — global subagent settings ONLY: enabled, max_concurrent_threads_per_session, default_subagent_model, default_subagent_reasoning_effort, interrupt_message. Individual subagents are not [agents.<name>] tables — they are standalone TOML files at .codex/agents/*.toml (project) or ~/.codex/agents/*.toml (personal), one file per agent, with required name + description + developer_instructions and optional model, model_reasoning_effort, sandbox_mode, mcp_servers, skills.config (corrected 2026-08-02 — the old [agents.<name>] / config_file / nickname_candidates schema is gone).
• [permissions.*] — permission policy.
• project_doc_max_bytes, project_doc_fallback_filenames — AGENTS.md controls (see §7).
• Optional model-catalog JSON path loaded at startup, overridable per profile.

REMOVED — [profiles.*] table syntax is gone as of 0.134.0 (2026-05-26). Profiles now live in dedicated per-profile files ~/.codex/<name>.config.toml, selected with --profile <name>. A config still using [profiles.foo] tables is stale — flag it.

No .mcp.json at repo root. Codex does NOT read Claude's .mcp.json — MCP servers live inside config.toml. A bridge from .mcp.json → .codex/config.toml is a common port pattern (see cc-suite:bridge-mcp skill).

6. Hook events (Codex CLI)

Codex hooks mostly mirror Claude Code's event names — easier than the Antigravity divergence.

Absent in Codex (but present in Claude): Notification, FileChanged, StopFailure.

Hook I/O contract: Same JSON-on-stdin / JSON-on-stdout shape as Claude. Stdin: session_id, cwd, hook_event_name, tool_name, tool_input, etc. Stdout fields: continue, stopReason, systemMessage, suppressOutput, hookSpecificOutput. Exit codes: 0 + JSON = success with directives; 0 + plain text = added as context; 2 = block (reason to stderr); other = warning.

Caveats (added 2026-06):

• SubagentStart / SubagentStop hook inputs now carry subagent identity, including permission_mode (0.134.0).
• For SubagentStart, continue: false does not stop the subagent.
• Async command hooks are parsed but not yet supported — a config declaring one is accepted but the async behavior is a no-op.

7. AGENTS.md (Codex's canonical memory file)

Codex reads AGENTS.md before every turn. Files are concatenated root-down, joined by blank lines; files closer to CWD override earlier ones by position (there is no separate global/project boundary marker — it is pure positional override). Within each directory the search order is AGENTS.override.md → AGENTS.md → fallback filenames, and at most one file per directory is taken; the walk stops at CWD. Global layer: ~/.codex/AGENTS.override.md → ~/.codex/AGENTS.md.

Default cap: 32 KiB per file (project_doc_max_bytes in config.toml). Fallback filenames: Configurable via project_doc_fallback_filenames — this is the official hook for AGENTS.md / CLAUDE.md / GEMINI.md interop (set the array to include all three to make Codex read whichever exists).

Body conventions (not enforced, but common):

• ## Working agreements — high-level decisions
• ## Repository expectations — invariants
• @file.md imports are NOT supported (unlike Gemini's GEMINI.md hierarchy) — use file concatenation instead.

The nlpm pattern of CLAUDE.md → one line @AGENTS.md does NOT work for Codex (no @-import). Codex authors should put content directly in AGENTS.md and configure Claude Code's CLAUDE.md to import it instead.

8. Slash commands / prompts

Codex's slash-command / prompt format lives at ~/.codex/prompts/<name>.md (project form .codex/prompts/). The "deprecated in favor of skills" framing is not confirmed in current docs (2026-06-07) — prompts are still documented. nlpm should NOT penalize their presence, and should treat any migration recommendation as advisory/soft rather than asserting deprecation.

Placeholders if scoring legacy prompts: $1..$9, $ARGUMENTS, $FILE, $TICKET_ID, $$.

9. Recent breaking / material changes (last 6 months)

Latest stable as of 2026-08-02: 0.146.0 (2026-07-29); pre-releases through ~`0.147.0-alpha` (verify the exact tag before quoting). Versions 0.138.0–0.144.x are omitted here — only non-first-party aggregators carried them this refresh; add rows once a first-party changelog source is read.

Repos relying on the removed built-in MCPs will silently regress under 0.131+. nlpm should flag MCP configs that name MCPs no longer shipped natively. Configs using [profiles.*] tables are stale under 0.134+.

10. Scope and uncertainty

This skill covers Codex CLI conventions. It does NOT cover:

• Universal SKILL.md spec → nlpm:conventions
• Penalty tables → nlpm:scoring
• Cross-artifact check → invoked by agents/checker.md

Resolved in the 2026-08-02 refresh:

• plugin.json required fields — only name is required; version/description/author/interface are optional (§3).
• Subagent config mechanism — standalone .codex/agents/*.toml files, not [agents.<name>] tables (§5).
• SessionEnd is a real Codex hook event (§6).
• .agents/skills/ (not .codex/skills/) reconfirmed as the sole filesystem skill-discovery path (§2).
• Contributor / CLA policy for openai/codex itself: external contributions are by invitation only — unsolicited PRs are closed without review; invited contributors sign a lightweight bot-verified CLA via a PR comment. This governs the openai/codex repo only, NOT third-party repos that merely target Codex as a tool (the auditor's actual PR targets). If PR-D ever contributes to openai/codex directly it needs an invite-only DENY-style gate, not the Google URL-CLA flow.

Resolved in the 2026-06-07 refresh:

• child_agents_md feature flag — not found in any current doc. Treat as removed/never-shipped; do not score against it.
• AGENTS.md merge boundary — resolved (§7): pure positional override, one file per directory, walk stops at CWD.

Still open (verify before scoring with confidence):

• .app.json schema for the plugin apps field — documented but the full schema is still unpublished.
• Sidecar agents/openai.yaml dependencies.tools[].url field — seen once in build-skills, not cleanly re-confirmed; do not hard-flag its presence or absence yet.
• Codex versions 0.138.0–0.144.x — first-party changelog rows still pending (§9).