ai-config-openai-codex-integration

Background

• Talbot (2026-09-29, in Market-timing-Red-Team): “For very important work (like Red Team stress testing), I want to use the strongest SOTA models. Currently, that is NOT Gemini. It is OpenAI (and its latest models).” · “I want to avoid API billing unless critical… I would prefer a known-cost solution, like adding a $20/mn OpenAI subscription” · “we should /task-prepare to integrate OpenAI/ChatGPT into our ai-config and KB-OS systems, as a complement to Claude Code.” · Approved: “yes, prep this task”.
• Gemini (CLI or AntiGravity) is reserved for mid-level tasks, to be wired up later (same response). The Gemini CLI trigger in ~/ai-config/UPGRADES.md (“when a Gemini CLI install appears, add it to the fan-out”) has fired, but stays deferred.

Context (Auto-Prepared)

Verified 2026-09-29 unless marked otherwise.

• Codex CLI is installed: pnpm add -g @openai/codex, v0.159.0, binary ~/.local/share/pnpm/codex. It is signed in with ChatGPT (device code) for the Red Team pass. Credentials are in ~/.codex/auth.json, which OpenAI says to treat “like a password” (Codex auth).
• Plan: ChatGPT Plus, $20/month. Per Codex pricing:
  • about 15–150 messages per 5 hours on GPT-6 Sol, plus weekly limits;
  • models listed: GPT-6 Sol/Luna, GPT-5.6 Sol/Terra/Luna, GPT-5.5, GPT-5.4 and 5.4 mini;
  • at the limit, “purchase additional credits” is optional, so there is no metered billing.
• Headless use (CLI reference): codex exec -m <model> -c model_reasoning_effort=high -s read-only|workspace-write -C <dir> -o <answer-file> "<prompt>".
• ai-config today:
  • deploy.sh copies claude/commands/* (the SMTM /task-* skills) to Claude Code only.
  • It copies skills/ (the tool-agnostic SSOT) to Claude Code and AntiGravity.
  • It sends nothing to Codex or Gemini CLI.
  • AGENTS.md is the rules SSOT, and claude/CLAUDE.md and GEMINI.md are thin wrappers.
• Unverified, check first: whether Codex reads a global ~/.codex/AGENTS.md and per-repo AGENTS.md files natively, and whether it supports custom commands or skills in a format the SMTM skills could be ported to.
• Model-choice policy lives in Core/AI/Subscriptions.md (Active Subscriptions). Add ChatGPT Plus there once it is subscribed.

Tasks

1. Verify Codex’s instruction-file and custom-command/skill support at its own docs or release notes, not blog summaries.
2. Add a Codex target to deploy.sh:
   • deploy AGENTS.md (plus a thin codex/ wrapper, if one is needed) to wherever Codex reads global instructions;
   • never commit or copy ~/.codex/auth.json;
   • follow the ai-config rule of diffing before deploying.
3. Decide which skills Codex gets:
   • the tool-agnostic skills/, by extending the fan-out;
   • whether the SMTM /task-* commands should be ported, or whether Codex should instead be used as a reviewer, called headless by Claude (the Red Team pattern) and not as a task runner.
4. Write the model-routing rule (Claude vs OpenAI vs Gemini: which work goes to which, and why) into AGENTS.md or Core/AI/. Include a reusable “second-opinion / Red Team” recipe (codex exec on a throwaway copy, with a self-contained brief).
5. Record the ChatGPT Plus subscription in Core/AI/Subscriptions.md, and register Codex in Core/IT/Utils/External/Utilities.md.

Success Criteria

• A fresh codex session in ~/projects/<any> shows it loaded the AGENTS.md rules (for example, it can state the “NEVER GUESS” rule without being told).
• deploy.sh covers Codex and was tested with a deploy run.
• The model-routing rule and the second-opinion recipe are written down, and Claude can run the recipe end to end without Talbot pasting anything.