From dual-review
Run parallel code review using Antigravity CLI (agy) and Codex CLI as two independent reviewers, then synthesize their findings into one unified review. Use whenever the user asks for code review with multiple AI reviewers, says "dual review", "review with agy", "agy and codex review", "agy+codex", or wants a broader review perspective beyond Claude alone. Also trigger when the user is doing /code-review and mentions wanting external reviewer opinions.
How this skill is triggered — by the user, by Claude, or both
Slash command
/dual-review:dual-reviewThe summary Claude sees in its skill listing — used to decide when to auto-load this skill
Runs `agy` (Antigravity CLI) and `codex review` as independent reviewers in parallel, then merges their findings into one unified review.
Runs agy (Antigravity CLI) and codex review as independent reviewers in parallel, then merges their findings into one unified review.
Determine what to review from user context or ask:
| Scope | agy prompt | codex flag |
|---|---|---|
| Uncommitted changes | "review the uncommitted changes in this project" | --uncommitted |
| Branch vs base | "review the changes on this branch compared to <branch>" | --base <branch> |
| Specific commit | "review commit <sha>" | --commit <sha> |
Default to uncommitted changes if not specified.
Run both in the same message turn as two independent Bash calls so they execute concurrently.
agy reviewer:
agy --add-dir "$(pwd)" \
--dangerously-skip-permissions \
--print-timeout 5m \
--print "You are a strict code reviewer. Review the uncommitted changes in this project.
For each issue found, use severity:
- P0: blocking — data loss, security hole, crash, broken API contract
- P1: correctness bug, breaking change, violates project conventions
- P2: maintainability defect, design concern, potential future breakage
- P3: optional improvement, does not affect correctness
Format each issue exactly as:
[Pn] <title>
Location: <file>:<line>
Issue: <specific description of what is wrong>
Root cause: <why this is wrong — spec, language rules, or actual behavior>
Impact: <consequence of not fixing>
Fix: <only include if non-obvious>
Rules: No praise. No encouragement. No summary. Issues only. If no issues found, output only: No issues found."
codex reviewer:
codex review --uncommitted \
"Strict review. For each issue use severity P0/P1/P2/P3 where P0=blocking/security/data-loss, P1=correctness/breaking, P2=maintainability/design, P3=optional.
Format each issue as:
[Pn] <title>
Location: <file>:<line>
Issue: <what is wrong>
Root cause: <why it is wrong>
Impact: <consequence>
Fix: <only if non-obvious>
No praise. No encouragement. Issues only."
Adjust the scope flags to match the chosen review target.
After both complete, merge into one review:
[agy+codex]. Being flagged by both reviewers adds confidence — note it concisely.[agy] or [codex] source tag. Do not drop unique findings — one reviewer catching what the other missed is the value of running both.Match the user's review standard (CLAUDE.md):
## Dual Review: agy + Codex
[P0] <title> [<source>]
位置: <file>:<line>
問題: <what is wrong>
技術根據: <why it is wrong — spec, actual behavior, or convention>
影響: <consequence of not fixing>
修正方向: <fix direction, only if non-obvious>
[P1] ...
總計: P0 x個 / P1 x個 / P2 x個 / P3 x個
<source> is [agy], [codex], or [agy+codex].
Creates, edits, and optimizes skills for Claude Code, including drafting, evaluating with test prompts, iterating on performance, and improving skill descriptions for better triggering accuracy.
npx claudepluginhub maxjkfc/claude-plugins --plugin dual-review