← ClaudeAtlas

codex-validatelisted

Validate the codex, codex-review, and codex-discuss skills end-to-end in a session where the Codex CLI is available — preflight, hook unit tests, and per-skill smoke tests that prove context actually reaches Codex — then self-improve by recording each run and tightening checks when something breaks. Use when codex is installed and you want to confirm the skills still work.
y-a-v-a/skills · ★ 0 · AI & Automation · score 70
Install: claude install-skill y-a-v-a/skills
You validate that the sibling Codex skills actually work against a *live* Codex CLI, and you keep this validation flow sharp over time. The hard part of these skills isn't running a command — it's proving the **context handoff** works: `codex exec` is stateless, so every smoke test below is designed to confirm that what you put in the prompt actually reached Codex, not just that a command exited 0. ## Scope `$ARGUMENTS` - No target, or `all` → run preflight, hooks, drift, and all three skill smoke tests. - `codex` | `codex-review` | `codex-discuss` | `hooks` | `drift` → run just that suite (`hooks` includes the §1a wiring check; `codex` includes §2b). - `--fix` → when a check fails, apply the fix to the offending skill (not only this validator) instead of only reporting it. Without `--fix`, propose the fix and ask before editing a sibling skill. The skills under test live at `skills/codex/SKILL.md`, `skills/codex-review/SKILL.md`, `skills/codex-discuss/SKILL.md`. ## 0. Preflight (always run first) ```bash command -v codex && codex --version # must exist — this skill is pointless without it codex login status # must be authenticated command -v jq # the validation hook needs jq ``` If `codex` is missing, **stop** and tell the user — there is nothing to validate. Do the rest of the work in a throwaway directory so the repo stays clean. Make it a git repo so `--cd "$SANDBOX"` is a real working tree: ```bash SANDBOX="$(