cross-reviewlisted

# cross-review — second-opinion review Get an independent review of a file (plan, spec, code change, anything) from a reviewer that did **not** write it. The default path is a different model family via the Codex CLI — useful for catching blind spots that same-family review misses: race conditions in async code, silent behavior changes from edits, brittle test assumptions. On a Claude-only machine the review runs through a fresh same-model subagent instead (see **Fallback**, below) — an equally valid path, not a degraded one. **Cost note:** the Codex path makes one round-trip to Codex per invocation (~30s–2min and counts against your subscription). Use it for high-stakes artifacts (plans, designs, refactors), not trivial edits. The fresh-subagent path spends one subagent turn instead. ## Arguments `$ARGUMENTS` — the path to the file to review, optionally followed by a focus hint. If no focus hint is given, default to "correctness, missing tests, ordering / sequencing bugs, brittle assumptions, security." ## Steps ### 1. Parse arguments Split `$ARGUMENTS` into the file path (first word) and the optional focus hint (everything after). If the path doesn't exist, stop and ask for clarification. ### 2. Build the review prompt Assemble a self-contained prompt for the reviewer with four parts: 1. **Project orientation** — if a `CLAUDE.md` exists in the working directory (project root), summarize the first ~200 words: what this project is, tech stack, key conventions. If