Skip to content

AGENTS.md

Latest commit

 

History

History
42 lines (29 loc) · 3.21 KB

AGENTS.md

File metadata and controls

42 lines (29 loc) · 3.21 KB
Raw

Codex Instructions

This is the Codex boundary for repositories using the AI Operating System starter. Claude Code keeps using CLAUDE.md and .claude/ unchanged.

Defaults

  • Use Caveman Ultra for prose from the first reply, without asking or requiring $caveman. Keep code, commands, identifiers, errors, commits, PR text, and files normal.
  • Use plain prose for security warnings, irreversible confirmations, and ambiguous multi-step decisions, then resume Ultra. A new session restores Ultra after the user temporarily disables it.
  • Read only CLAUDE.md's Verification and Environment & Deploy Target sections for configured project facts. Use .claude/reference/ for architecture, commands, deployment, and pitfalls. Other CLAUDE.md workflow rules are not Codex instructions.
  • Never execute .claude/hooks/session-start.sh in Codex.

Capabilities

  • Inspect tools exposed in the current session before using subagents, browser control, connectors, or interactive input. Config flags alone are not proof.
  • Serial fallback is valid only when independence is not part of the deliverable. impartial-review, advocate, and why require fresh independent context; if unavailable, report the gap.
  • Claude Workflow programs are not Codex programs. Recreate their intent with exposed Codex agents or flag them blocked.
  • Keep critical Codex rules here rather than in project hooks, which require separate trust and can be disabled.

RTK

  • When installed, prefer rtk for noisy supported reads: rtk git status, rtk git diff, rtk git log, rtk git show, rtk rg, and rtk read.
  • Use rtk test <command> for failure-focused output; preserve its exit code and rerun natively when full success output is required as evidence.
  • Use native commands for mutations, interactivity, unsupported syntax, exact-output parsing, and diagnosis when filtering hides detail.
  • Codex has no Claude RTK rewrite hook here. Invoke rtk explicitly.

Safety

  • Caveman Ultra is a communication default, not side-effect authorization. Auto-merge and other persistent side-effect modes require explicit current-session intent.
  • Do not inherit Claude auto-commit, push, PR, or merge. Ship only when the request includes shipping.
  • Never push to main, force-push, merge, delete branches/worktrees, migrate, deploy, install runtime dependencies, or modify external checkouts without explicit approval.
  • Stage explicit paths, preserve unrelated changes, and verify before claiming completion.

Shared Assets

  • .claude/skills/ is canonical; .agents/skills/ contains Codex adapters. Treat $ARGUMENTS as invocation input.
  • Read relevant .claude/reference/ material before unfamiliar work and .agents/CODEX-SKILL-COMPATIBILITY.md before adapted, gated, or dangerous skills.
  • After canonical skill changes run node .claude/scripts/sync-codex-skills.mjs --write.
  • Tool mapping: .claude/skills/using-superpowers/references/codex-tools.md.

Starter Maintenance

Only in the canonical claude-starter template repository: keep bootstrap/ PowerShell 5.1 files ASCII-only and shell scripts LF. Claude gets its hooks and slash skills; Codex gets AGENTS.md and .agents/skills/. Keep private paths, tokens, and maintainer-only workflow out of defaults.