codex-auditlisted

# codex-audit - outside eyes on code Claude wrote Claude wrote this codebase, so Claude is the wrong auditor for it. This skill buys a fresh, adversarial reading from a different model family, then puts every claim it makes through a gate before acting on any of it. The division: | | Codex red team | Claude | |---|---|---| | Hunting | tears the project apart in a throwaway worktree | writes the brief, picks the lenses | | Evidence | must ship a runnable proof per finding | runs that proof; a claim that will not reproduce is dead | | Judgement | none - it reports candidates | decides what is real, reachable, worth fixing | | Fixing | never | does the fix, proven by the proof flipping red -> green | **Why Claude fixes and Codex does not:** finding needs fresh eyes, fixing needs context. A security fix is judgment work, which `codex-delegate` §0 puts on the architect-only surface. Claude has the codebase in its head; the red team does not, and a fix from a model that has read one lens of the project is how a patch closes one hole and opens another. **Requires the `codex-delegate` plugin** for `dispatch.py` / `doctor.py` and the lane model. This skill is the audit protocol on top of that machinery. ## 0. Two modes, and the difference is not cosmetic | | **A. Session audit** | **B. Comprehensive refactor** | |---|---|---| | Trigger | user asks to audit the work just done | user asks **explicitly**; never inferred | | Shape | a **hunt** - find what nobody thought of | a **ra