Cross model review

Skill Neznakometz/StackForge/core/skills/cross-model-review

Thin core + domain sets for a fast, stack-aware project start. Cross-agent (Claude Code, Cursor, Codex, Gemini) via the SKILL.md standard. MIT.

Install

npx -y skills add Neznakometz/StackForge --skill cross-model-review

Assembled from the repository path, not quoted from the project. Check it against their README if it does not work.

One thing to look at

0 stars0 stars. Stars are a popularity signal and not a quality one, but at this level it is likely that nobody has read this closely except its author, and you would be relying on your own review.

What its author says it does

Copied from the file, not written here

Mandatory re-check of the code by an external model (Codex/ChatGPT and/or Gemini) acting as an independent reviewer. Apply on every task after the internal reviewers and on the diff at the end of a phase.

SKILL.md

2.7 KB, as published. Nobody here has run it

Cross-model review (external "second brain")

Claude writes → an external model attacks → Claude arbitrates. An asymmetry of roles, not "merge N answers". Cheap (one extra call per task), catches what a single model doesn't see.

Mechanics

The orchestrator calls the external CLI directly via Bash (authentication is active beforehand). The plugins' own slash commands (/codex:*) are triggered only by the user.

Codex (ChatGPT):

codex review --commit <SHA> — a single commit (a task);
codex review --base <ref> — diff from ref to HEAD (a whole phase / several commits);
codex review --uncommitted — the working tree.
IMPORTANT: --commit/--base do NOT accept a text prompt. Custom instructions — only without a scope flag: codex review "<instructions>" (over the working tree).

Gemini (optional, second model): the gemini CLI in the same manner — pass it the diff and the review instruction. Enable it when a second independent point of view is needed (high-risk changes).

Fallback — review still happens with zero external setup

If NEITHER codex nor gemini CLI is available (check once: command -v codex || command -v gemini), do not skip review. Fall back to a fresh Claude subagent as the external reviewer — a different model from the implementer where possible (e.g. implementer on sonnet → reviewer on opus), with a clean, narrow context (diff + task text only, no dialog history). This is "fusion-lite": the diversity is by role and model size, not vendor, but the adversarial second-pass is preserved. External CLIs are an upgrade, not a prerequisite — review is never skipped.

Arbitration (the key part)

The external model's output is the recommendations of a THIRD-PARTY reviewer, not the truth:

a real finding → return it to the implementer as a finding (file:line);
noise/false positive → discard with a note (just like with code-reviewer);
artifacts of the external tool's sandbox (for example, a build failure due to write permissions) — not a code defect.

Don't forward the external model's whole output into the main context — only the selected findings.

When mandatory

Every task — after spec-reviewer + code-reviewer, before test-runner.
End of phase — adversarial review of the whole phase diff (--base <phase-start>).
Technical/destructive changes (auth, money, migrations, concurrency) — mandatory, both models allowed.