Second opinion

Second Opinion

Spawn an independent Codex CLI review of the current change for a fresh perspective from a different model. Codex advises only — it runs in a read-only sandbox and cannot edit code. Complements your own review or /aetheris-review; it does not replace them.

When to use

• You want a fresh pair of eyes on a diff before merging — from a model other than the one that wrote the code.
• The user asks for "a second opinion", "fresh eyes", or "what would Codex say" about a change.
• Before a risky merge, when independent confirmation is worth a pass.

Not for: applying fixes (advisory only — acting on a finding is a separate, explicit step the user must request), or posting PR comments (use /aetheris-review).

How to run

Run the bundled review.sh from this skill's own directory. It selects scope, invokes Codex read-only, and saves the report outside the repo (~/.codex-reviews/<repo>/) so the working tree stays pristine:

bash <this-skill-dir>/review.sh [scope-flag]

Two modes — code review (a diff) and plan review (a doc):

Flag	Reviews
(none)	code: smart default (feature branch → vs default branch; else working tree)
--uncommitted	code: staged + unstaged + untracked
--base <branch>	code: diff vs a base branch (PR-equivalent)
--commit <sha>	code: a single commit's changes
--plan <file>	a plan / spec / design doc, before it's built — Codex critiques soundness, dependency ordering, risk/cost/reversibility, and gaps, grounding against the code the doc references

The helper streams Codex's review to stdout and ends with a REPORT_PATH: line. Code mode uses codex exec review on the chosen scope, running Codex's built-in review harness; plan mode uses codex exec with a baked-in plan-critique prompt. Both run from the repo so Codex can cross-reference code.

The report is the verdict, not the transcript. Both modes capture Codex's final message via --output-last-message and write it to the report; the streamed run lands beside it in <same-name>.transcript.log. Read the report; open the transcript only when the report says Codex never reached a verdict.

codex exec review does accept custom review instructions, but not alongside a scope flag — --uncommitted / --base / --commit and a custom prompt are mutually exclusive, so scope-selected reviews use the built-in harness.

Presenting the result — faithful passthrough

Codex is the advisor. Do not re-rank, soften, re-judge, or restate its findings as your own analysis.

1. Surface Codex's verdict under a clear frame, e.g. "Codex's independent review (read-only):" and quote its findings as written.
2. Give the saved report path.
3. Close with one line: "Want me to act on any of these?"

Acting on a finding (editing code) is a separate step — only after the user asks. This skill itself never edits code.

Guarantees

• No changes: --sandbox read-only makes writes impossible. Confirm with git status that the repo is untouched after a run.
• Out-of-repo report: written to ~/.codex-reviews/<repo>/, never inside the repo, alongside a .transcript.log of the full run.
• Focused & lean: runs Codex with --disable hooks so the host machine's skill/plugin session_start hooks aren't injected into the review (without it, Codex rabbit-holes reading skill files — ~100x more output). Per-run only; your global Codex config is untouched.

Troubleshooting

• codex CLI not found → install (npm i -g @openai/codex) or run the marketplace's install-codex.sh.
• codex may not be logged in → codex login.
• no changes to review → no diff for the chosen scope; pick a different scope flag.