Jonathon's AI Wiki

Codex Inside Claude Code — Fable 5 Orchestrates, GPT-5.5 Executes

5 min read

Source: raw/x-bookmarks-recent-digest-2026-07-09.md (@cjzafir item), raw/x-bookmark-2074638582418231495.md (@steipete)

A community delegation workflow that runs OpenAI Codex as the heavy-implementation executor under Claude Fable 5, inside Claude Code, via OpenAI’s official Codex plugin. Fable 5 stays the orchestrator — planning, repo understanding, architecture, decomposition, and final review — and hands the token-hungry coding (implementation, debugging, refactoring, multi-file edits) to Codex through the /codex:rescue command, billed against your ChatGPT/Codex subscription instead of your Fable 5 budget. The plugin, its commands, and the rescue sub-agent are real and confirmed installed in this environment;^[inferred] the headline cost-savings figure is an unverified influencer claim.

Key Takeaways

  • The shape. Fable 5 = orchestrator; OpenAI Codex = executor. It is the inverse of the advisor pattern (cheap model executes, strong model advises) — here the strong Anthropic model plans and a third-party model does the heavy lifting.
  • One-time install (from the @cjzafir post, verified against the installed plugin):
    • /plugin marketplace add openai/codex-plugin-cc
    • /plugin install codex@openai-codex
    • /reload-plugins
  • Fable 5 finishes setup. Have Fable 5 run /codex:setup — it installs the Codex CLI (npm install -g @openai/codex) if missing, prompts you to authenticate your ChatGPT/Codex account once (codex login), and verifies the codex:codex-rescue sub-agent is available. It changes no project code.
  • Delegation contract. Fable 5 plans and reviews; delegate heavy implementation / debugging / test-fixing / refactoring / multi-file edits to Codex with /codex:rescue. The author prefers GPT-5.5 at xhigh effort as the go-to Codex model.
  • Always inspect before accepting. Both the source and the plugin’s own command doc stress it: read Codex’s output yourself before trusting it. Do not blindly accept.
  • Claimed ~60% Fable-5 token savings ^[inferred] — this is @cjzafir’s standalone assertion with no linked measurement; treat as directional, not benchmarked.
  • Corroboration. @steipete (6,344 likes) independently tells people to “ask Fable to make codex the workhorse,” pointing to a published codex-first skill — so the “Codex-as-workhorse-under-Fable” pattern is circulating among multiple practitioners, not a one-off.

The Delegation Prompt (verbatim from @cjzafir)

Paste this to Fable 5 after setup to establish the contract:

You are the orchestrator.

  • Use Fable 5 for planning, repo understanding, architecture decisions, task decomposition, and final review.
  • Use codex-rescue as the executor when a task needs heavy implementation, debugging, test fixing, refactoring, or multi-file code edits.
  • When delegating to Codex, use /codex:rescue.
  • Prefer GPT-5.5 (xhigh) as the go-to Codex model.
  • Keep Codex tasks focused and specific.
  • After Codex finishes, inspect the result yourself before accepting it. Do not blindly trust Codex output.

What’s Verified vs. Claimed

The vault schema flags whether tweet claims are backed by evidence. Here is the split:

  • Verified in this environment — the OpenAI codex plugin (marketplace openai-codex, v1.0.6) is installed with commands/setup.md, commands/rescue.md, and agents/codex-rescue.md. /codex:setup does exactly what the post describes (checks/installs the Codex CLI, guides auth). /codex:rescue forwards the request to the codex:codex-rescue sub-agent and exposes --model <model|spark> and --effort <none|minimal|low|medium|high|xhigh> flags — so “GPT-5.5 at xhigh” is a real, selectable configuration. The plugin’s own doc instructs the caller to “inspect the result yourself before accepting,” matching the post’s guidance.
  • Claimed, unverified — the ~60% token-savings figure ^[inferred] (no benchmark, single-author assertion), and that GPT-5.5 specifically is the best executor model (author preference, not measured here). The @steipete post is a standalone recommendation but does link a real codex-first SKILL.md as backing.

Pro Tips (from the source)

  • Wrap it as a reusable skill. @cjzafir names his “Fable-GPT” and invokes it at the start of each session so the contract is always loaded.
  • Skill + goal for long-horizon work. Pair the skill with a goal/loop; goals suit long-horizon tasks best.
  • Subagents on the Codex 20x pro plan. The author runs 5–7 Codex subagents concurrently and reports never hitting the 5-hour limit.
  • Manage context rot. Clear the conversation after ~4 compactions; use a /handoff skill to preserve context across the reset.

Implementation

  • Tool/Service: OpenAI Codex (via the official codex plugin for Claude Code) driven by Claude Fable 5 as orchestrator.
  • Setup: /plugin marketplace add openai/codex-plugin-cc → /plugin install codex@openai-codex → /reload-plugins, then have Fable 5 run /codex:setup and authenticate your ChatGPT/Codex account once.
  • Cost: Codex work bills against your existing ChatGPT/Codex subscription (e.g. Codex 20x pro), which is what shifts token load off the Fable 5 budget. The specific ~60% saving is unverified ^[inferred].
  • Integration notes: /codex:rescue runs inline and invokes the codex:codex-rescue sub-agent; it is a command/sub-agent, not a Skill(). Leave --model/--effort unset for defaults, or pass --model gpt-5.5 --effort xhigh per the author’s preference. Always review Codex output before accepting.

Open Questions

  • The 60% token-savings claim has no published measurement; a rigor-matched comparison (same task, same verification bar) would be needed to confirm it — see the coordinator-cost cookbook’s methodology for how that should look.
  • Whether GPT-5.5 at xhigh is genuinely the best executor model, or just the author’s habit, is untested.

Try It

  • Run the three install commands, then paste the setup prompt to Fable 5 and let it run /codex:setup and walk you through auth.
  • Establish the contract by pasting the delegation prompt above, then give Fable 5 a task and watch it decompose, delegate the heavy edits to /codex:rescue, and review the diff itself.
  • Save the whole thing as a “Fable-GPT” skill and load it at session start.
  • Measure your own token split (Fable 5 vs. Codex) across a few real tasks before trusting the ~60% figure — treat it as a hypothesis to verify, not a spec.

Graph View

Backlinks