Second opinion

A model from another family reviews the change after you have reviewed it yourself. It catches what the authoring model is blind to. It is a pragmatic second opinion, not an opponent: a clean result is a good result, and the reviewer is told it does not have to find anything. The reviewer follows references/brief.md.

Configure

One variable, in the repository's .env or the environment: reviewers as harness:model[@effort], in order of preference. A linked worktree reads the main checkout's .env, since the file stays untracked there. The script picks the first one whose model family differs from yours, so one list serves every harness you work in.

# Grok reviews Claude and GPT; Sol, at low effort, reviews Grok.
APLUS_SECOND_OPINION_REVIEWERS=cursor:grok-4.7-high,codex:gpt-6.1-sol@low

# From Claude Code, Sol reviews; from Codex, Opus does.
APLUS_SECOND_OPINION_REVIEWERS=codex:gpt-6.1-sol,claude:opus

# OpenCode reaches several providers: Kimi reviews Claude and GPT; GPT reviews Kimi.
APLUS_SECOND_OPINION_REVIEWERS=opencode:opencode-go/kimi-k3,opencode:github-copilot/gpt-6.1-sol

• Family, not harness. Cursor and OpenCode also run Claude and GPT models, and the review is only worth it from a model the author's blind spots do not share. The family comes from the id, after any provider/: claude for Claude ids and opus, sonnet, fable and haiku; gpt for gpt, codex and o<n>; otherwise the first word, such as grok.
• Keep the pair. Run a configured or user-selected harness and model together. If it fails, report the failure and ask the user which harness and model to use next. Ask before using a local subagent too; its allowance can also be limited.
• Effort is a suffix, not a model. Write gpt-6.1-sol@low, not an invented gpt-6.1-sol-light. The suffix is one of minimal, low, medium, high, xhigh or max, and becomes the harness's own option in the table below.
• Exact ids. A listed reviewer runs as set, including a *-fast id. A model you pick must not be a fast variant: a bare alias can resolve to one (in Cursor, grok-4.7 means the account's default, such as grok-4.7-high-fast). On --model, the script refuses an id containing fast, except [fast=false]. It always switches off Claude Code's fast mode, which is a setting rather than an id.
• Ids are not checked. They change often and differ per account, so a wrong id fails in the reviewer's CLI. List the current ones with the command in the table.
• Login, with a key as fallback. Each CLI uses its own login. When the agent cannot reach it, as in CI or a sandbox, set APLUS_SECOND_OPINION_CLAUDE_API_KEY, APLUS_SECOND_OPINION_CODEX_API_KEY or APLUS_SECOND_OPINION_CURSOR_API_KEY. The script passes it to the reviewer only, as ANTHROPIC_API_KEY, CODEX_API_KEY or CURSOR_API_KEY, so the coding agent's own key is left alone. OpenCode keeps a key per provider, set with opencode auth login.

Script

node --experimental-strip-types <this skill's folder>/scripts/run-review.mts --base <ref> --notes <file> --self <your model id> [--round 1|2|3]

The coding agent runs this itself; the user never has to. Run it from inside the repository, in the background: a review takes minutes. <ref> is the base the change targets, such as origin/<integration branch>, or, for a stacked PR, the branch below it. --self is the model you run on. The script sends the brief, the round, your notes, the path of a file holding the diff from the merge base to the working tree without lockfiles, and the paths of untracked files, then prints the reviewer's findings. On success it deletes the diff; on failure it saves the prompt beside the diff and prints its path.

First run with --self to use the first eligible reviewer in APLUS_SECOND_OPINION_REVIEWERS, from the environment or .env. When the user explicitly names a harness and model for this review, pass --harness <harness> --model <id> instead of --self.

• Exit code 2 means no listed reviewer is from another family. Ask the user which harness and model to use, proposing one that is not the model you run on, not the harness's most expensive frontier model — Fable in Claude Code, Astra in Codex — and never a fast variant. When the user offers only models of your own family, ask which one instead of choosing. In Claude Code, Opus and Sonnet review each other; in Codex, Sol and Luna do. Run the script again with the chosen --harness and --model in place of --self.
• If the user chooses a local subagent instead, give the saved prompt file to a fresh, read-only subagent on the chosen model. Skip the review if the user prefers.
• These rules are only for a model you pick. A listed reviewer runs as set, whatever its tier.
• Invoke this skill by name rather than only reading this file. Its allowed-tools pre-approves the command above, so a harness that blocks one agent from starting another lets it run. A harness that ignores allowed-tools needs the same rule in its own settings.

Flow

There are at most three rounds per change, and the third is optional. The script refuses a fourth.

1. Review it yourself first, with final-review or the project's own gate. The checks must be green.
2. Round 1. Write notes to a scratch file outside the repository: what your review found, what you fixed, and what you deliberately left, with a reason for each. The reviewer treats those points as settled. Run --round 1.
3. Show the review. As soon as the review returns, show the user the reviewer's output in full, unedited, before acting on it. Show every round, including one that finds nothing.
4. Verify each finding against the code: reproduce the scenario or trace it through. The other model is a reviewer, not an authority. Fix the findings that hold and are easy, then run the checks again. Reject the ones that do not hold, with a one-line reason.
5. Round 2, only if round 1 led to a code change. Write new notes that list each round 1 finding with what you did: fixed at file:line, or rejected and why. Run --round 2. The reviewer checks only those fixes and what they touch. It does not look for new issues.
6. Round 3, optional, only if round 2 led to a code change. Run it as round 2, with notes on each round 2 finding, and --round 3.
7. Stop. Do not ask again because you disagree with a result or because the last round raised something. Verify what it reports, fix what is easy and holds, and list the rest as open findings for the user to decide.
8. Clean up. Delete your notes and any folder holding a saved prompt and diff: nothing else removes it.