Local Codex Review (pre-push)

Runs the exact same review Codex performs in CI (.github/workflows/codex-pr-review.yml),
but locally and scoped to work you have not pushed yet — so you catch what CI would flag
before the PR exists. Use this before git push on a non-trivial change.

Correspondence with CI — identical:

• Policy: REVIEW.md (severity triage, public-surface checklist, AGENTS.md compliance, test coverage).
• Model: gpt-5.6-sol, model_reasoning_effort="xhigh".
• Output: markdown starting with ## Codex Review, findings tagged P0 / P1 / P2 with file:line.

Differences from CI — local-only:

• Scope is the current branch vs main at the merge-base, including uncommitted changes (CI reviews a pushed PR diff).
• Sandbox is read-only (CI uses danger-full-access on an ephemeral runner). Codex reads the diff and files but cannot modify your working tree.
• Fresh context is inherent: codex exec is a separate cold process, so it does not anchor on the current chat session — the same reason local-review insists on a subagent.

Prerequisites

• codex CLI >= 0.144.1 installed and authed (codex login or OPENAI_API_KEY). Older CLIs reject gpt-5.6-sol with "requires a newer version of Codex". Upgrade with npm install --global @openai/codex@0.144.1 (may need sudo for a global install). Keep this in sync with the pin in .github/workflows/codex-pr-review.yml.
• git fetch the base ref if it's stale, so the merge-base is accurate.

Run

bash .agents/skills/local-review-codex/run.sh          # review vs main (default)
bash .agents/skills/local-review-codex/run.sh <base>   # review vs a different base ref

Invoke with bash (or run the executable directly) — the script needs Bash for
set -o pipefail; sh is Dash on Debian/Ubuntu and would fail. If main isn't a
local branch (e.g. a fresh single-branch checkout), the runner falls back to
origin/main automatically.

The script computes BASE_SHA = git merge-base HEAD <base>, feeds Codex REVIEW.md plus a
diff context pointing at git diff <BASE_SHA> (which folds in uncommitted edits), and prints
the review. It writes only temp files — nothing lands in the working tree.

Relaying the result

Print the Codex output verbatim. Do not re-summarize or filter it — the value of a cold Codex
pass is surfacing what the current session would rationalize away. Then decide with the user
whether to address findings before pushing.

For a Claude-native review instead, use local-review (branch-diff-reviewer subagent). This
skill is the Codex counterpart; run both for independent perspectives.