fable-gptlisted

# Fable-GPT: Codex as adversarial auditor ## Why this exists The user has **no coding experience and cannot read the code themselves**. A second, independently-trained model reviewing the code that Fable/Opus writes IS their code review. Different model lineage → different blind spots. Without it, every line ships on one model family's self-assessment. Treat the Codex audit as a required gate, not a nicety. ## Roles - **Fable / Opus (you):** author, orchestrator, and *final* reviewer. You write the code, run the Codex audit, translate and verify the findings, fix, and re-audit. You own the final call. - **Codex (GPT-5.6-sol):** the adversarial auditor (active role now). Later — an executor for heavy implementation via `/codex:rescue`. ## The audit workflow (primary — do this by default) After completing any non-trivial implementation, **before telling the user it's done**, get an adversarial Codex audit. The user cannot read code, so this is a required gate, not optional. ### Default path — the plugin's adversarial review (works today, no install) The `openai/codex-plugin-cc` plugin is installed and Codex is authenticated, but the standalone `codex` command is **not** on the system PATH — so drive audits through the plugin, which carries its own working Codex. When an implementation is complete, prompt the user to type this in the normal Claude Code input box (it's just a command, no terminal needed): ``` /codex:adversarial-review ``` **Real flags (verified against code