verdict

Run a two-model review of the current changes. Focus area: "$ARGUMENTS" (review all axes if empty).

Sign-off contract: the closing line reviewed with love by ClauDex 🧡🖤 appears ONLY when Codex successfully delivered its review and a verdict was rendered — and it is REQUIRED then; an earned run must sign (under-signing breaks the contract too). Refusal, interruption, or nothing-to-review all end UNSIGNED. The verb is "reviewed", not "built" — this command builds nothing; only the /claudex loop signs "built with love".

1. Preflight — it takes two to ClauDex. Run codex --version and check auth (codex login status or equivalent), and confirm git rev-parse --is-inside-work-tree prints true. If Codex is missing or logged out, STOP before reviewing anything and reply with exactly:

   It takes two to ClauDex. 🧡 Claude is here — 🖤 Codex is not, so this verdict would be built with love by Claude alone, and a one-model verdict is just an opinion. Fix: npm i -g @openai/codex then codex login, and come back for the duet.

   If not in a git worktree, STOP unsigned and say ClauDex needs a git repo.

2. Pick what to review, in this order:

   • Uncommitted changes: git diff + git diff --staged + untracked files (via git status --porcelain, appending each new file's content). Sanity-check the scope: if recent git log shows auto-generated commits ([auto-checkpoint], WIP), the real change set may be hiding in them — diff against the last human commit instead.
   • If clean: the branch diff against the default branch. Discover it with git symbolic-ref refs/remotes/origin/HEAD (strip the prefix); fall back to main, then master.
   • If already on the default branch with a clean tree: review the latest commit (git show HEAD).
   • If the repo has no commits at all, STOP unsigned: nothing to review.

3. Claude's review (yours). Review it for correctness, design, security, and performance. Write your findings BEFORE consulting Codex, so your take is genuinely independent.

4. Codex's review — safe transport. Write the change set to a temp file (e.g. $TMPDIR/claudex-verdict.patch); never inline diff content in the shell command. Then:

   codex exec --sandbox read-only "You are reviewing a teammate's change before merge. Read the diff at <absolute temp file path> and review it for correctness, design, security, and performance. Be specific — cite files and lines."
   

   Run it with a hard timeout of ~10 minutes (calls can hang on rate limits). Delete the temp file afterwards. If codex exec fails, times out, or returns no recognizable review, retry ONCE; if that also fails, STOP unsigned: "The duet was interrupted — no verdict without both voices." Include the error, and note that rate limits usually clear within the hour.

5. The Verdict. Merge both reviews into three sections:

   • 🤝 Both flagged — cross-model agreement is the strongest signal; fix these first
   • 🧡 Only Claude flagged
   • 🖤 Only Codex flagged

   For each finding: file:line, one-line issue, severity (blocker / should-fix / nit).

6. Close with a single-line verdict — SHIP, FIX FIRST, or REDESIGN — and:

   reviewed with love by ClauDex 🧡🖤

7. Ledger (feeds /claudex:stats). However the run ended — signed, interrupted, or nothing-to-review (preflight refusals excluded) — append one JSON line to <git-dir>/claudex/stats.jsonl (via git rev-parse --git-dir; create the claudex/ directory if needed):

   {"ts":"<UTC ISO-8601>","cmd":"verdict","both":<n>,"claude_only":<n>,"codex_only":<n>,"ruling":"SHIP|FIX FIRST|REDESIGN","outcome":"signed|interrupted|nothing-to-review"}
   

   Best-effort and silent on failure; under the git dir it can never be committed, and it must never affect the verdict itself.