multi-model-review

Multi-model code review for Spec-Driven Development.

Build with one model (Claude Code, Codex CLI, Gemini CLI, ...), review with a different one. This plugin turns GitHub Spec Kit artifacts (spec.md, plan.md, tasks.md) into a portable handoff surface so the two models never have to talk to each other — they just read and write the same markdown files.

The reviewer model does not need to be connected to your current session. The handoff is a single markdown file on disk. You run the other model yourself.

Why multi-model?

A single LLM reviewing its own output tends to rationalize rather than challenge. A different model reading the same spec and diff has:

• Different training data — catches issues the builder's model is blind to.
• Different calibration on severity — less confirmation bias.
• Different house style — flags idiomatic drift the builder normalizes.

Spec Kit itself is great at orchestrating one agent through specify → plan → tasks → implement. It does not prescribe a cross-model review loop. multi-model-review fills that gap:

• The builder model does implement.
• /multi-model-review:review-package bundles the spec, plan, tasks, diff, and project rules (CLAUDE.md) into a self-contained prompt for the reviewer.
• You run the reviewer model on that file.
• /multi-model-review:apply-review ingests the reviewer's report and walks you through fixes.

Because the bundle is just markdown, any LLM with a CLI (or an API) can play reviewer. Swap models freely — even run the same diff past two different reviewers for triangulation.

Supported reviewers out of the box

Reviewer	Template	Example command
Claude	templates/claude-review-prompt.md	claude -p "$(cat <pkg>/review-package.md)" > <pkg>/review-report.md
Codex	templates/codex-review-prompt.md	codex exec --file <pkg>/review-package.md > <pkg>/review-report.md
Gemini	templates/gemini-review-prompt.md	gemini --file <pkg>/review-package.md > <pkg>/review-report.md

Adding a new reviewer (Qwen, Mistral, a local llama.cpp process, ...) is a one-file change: drop templates/<reviewer>-review-prompt.md and use the new ID in config.json.

Requirements

• Claude Code 1.x — needed for the skill and slash commands.
• A Spec Kit–initialized project (optional but recommended) — specs/<slug>/spec.md is picked up automatically. Without it, the plugin falls back to diff-only mode.
• The reviewer model installed locally (any of):
  • Codex CLI
  • Gemini CLI
  • A second Claude Code instance (claude -p)
  • Any other CLI-capable LLM — write a template for it.
• git on PATH.

Install

Slash vs shell. Commands that start with / (e.g. /plugin install) are typed inside a running Claude Code session — not at your OS shell. Commands shown in bash blocks (git …, claude …, cp …) are typed at your OS shell (PowerShell, bash, zsh). Mixing them up returns '/plugin' is not recognized in PowerShell or command not found elsewhere.

As a Claude Code plugin

In a Claude Code session:

/plugin install multi-model-review@multi-model-review

Or at your OS shell (equivalent):

claude plugin install multi-model-review@multi-model-review
claude plugin list      # verify "Status: enabled"

Then run /reload-plugins inside Claude Code (or restart it). The three slash commands /multi-model-review:cross-review, /multi-model-review:review-package, and /multi-model-review:apply-review will appear in the picker.

Develop against a local clone

git clone https://github.com/formin/multi-model-review.git
cd multi-model-review
claude --plugin-dir .

After editing plugin files, run /reload-plugins inside Claude Code to pick up changes.

As a standalone skill (minimal)

If you don't want the full plugin:

mkdir -p ~/.claude/skills ~/.claude/commands
cp -r skills/multi-model-review ~/.claude/skills/
cp -r templates ~/.claude/skills/multi-model-review/
cp commands/*.md ~/.claude/commands/     # optional: /cross-review, /review-package, /apply-review (no plugin namespace)

The skill activates on phrases like "review with Codex", "review with Gemini", "cross-review", or when you invoke any of the commands.

Update / reinstall

Pull the latest version of the plugin from this repo — in a Claude Code session:

/plugin update multi-model-review@multi-model-review

Or at your OS shell (outside Claude Code), the equivalent CLI subcommand:

claude plugin update multi-model-review@multi-model-review

Uninstall the plugin (in a Claude Code session):

/plugin uninstall multi-model-review@multi-model-review

For a clean reinstall, uninstall and then install again.

In development mode (you cloned the repo and ran claude --plugin-dir .), update with git pull at your OS shell from inside the clone — no reinstall needed. If you see fatal: destination path 'multi-model-review' already exists when re-cloning, delete the existing directory first or pull in place:

# Windows PowerShell (at the OS shell, NOT inside Claude Code)
Remove-Item -Recurse -Force multi-model-review
git clone https://github.com/formin/multi-model-review.git

# macOS / Linux / Git Bash
rm -rf multi-model-review
git clone https://github.com/formin/multi-model-review.git

# or update in place
cd multi-model-review && git pull

Starting over from a corrupted state

If the install is partially broken (install didn't complete, you renamed the clone directory):

1. In Claude Code: /plugin uninstall multi-model-review@multi-model-review.
2. Exit Claude Code.
3. If ~/.claude/plugins/installed_plugins.json still has a stale entry for this plugin, hand-edit it out.
4. Delete any local clone directory if you used --plugin-dir.
5. Restart Claude Code and reinstall.

Quick start

From inside a repo with a feature branch checked out:

/multi-model-review:cross-review init

Answers: builder = claude-code, reviewer = gemini-cli (or codex-cli, etc.), base ref = main.

Build your change as you normally would (or via /speckit.implement), then:

/multi-model-review:review-package

The plugin writes .cross-review/packages/<timestamp>-<slug>/review-package.md and prints the exact command to run the reviewer. Example for Gemini:

gemini --file .cross-review/packages/20260417-1020-auth-rework/review-package.md \
  > .cross-review/packages/20260417-1020-auth-rework/review-report.md

When the reviewer finishes:

/multi-model-review:apply-review

Claude reads the report, filters findings by confidence, and walks you through each one.

Multi-reviewer pass (triangulate)

Nothing stops you from running the same package through two reviewers:

codex  exec --file <pkg>/review-package.md > <pkg>/review-report-codex.md
gemini --file <pkg>/review-package.md       > <pkg>/review-report-gemini.md

Then rename one to review-report.md and pass the other to /multi-model-review:apply-review <pkg> manually, or merge the two reports by hand. A built-in multi-reviewer merge is on the roadmap.

Commands

Claude Code namespaces plugin commands as /<plugin-name>:<command>. This plugin provides three:

Invocation	Purpose
/multi-model-review:cross-review [init|status]	Init config or show current roles, base ref, and packages.
/multi-model-review:review-package [slug] [--base <ref>]	Export the handoff bundle for the reviewer model.
/multi-model-review:apply-review [path] [--min-confidence N]	Ingest the reviewer's report and drive remediation.

Full details in docs/USAGE.md. A scenario-driven walkthrough (Claude builds → Codex reviews → Claude applies fixes) lives in docs/EXAMPLE.md.

Directory layout

multi-model-review/
├── .claude-plugin/
│   ├── plugin.json                    # plugin manifest
│   └── marketplace.json               # single-plugin marketplace catalog
├── skills/multi-model-review/SKILL.md # the skill (model-invoked)
├── commands/
│   └── multi-model-review.md          # single slash command, dispatches by flag
├── templates/                         # reviewer prompts + report schema
│   ├── codex-review-prompt.md
│   ├── claude-review-prompt.md
│   ├── gemini-review-prompt.md
│   └── review-report.md
└── docs/USAGE.md

At runtime, per-project state lives in .cross-review/ inside your target repo.

Design choices

• No side calls. The skill never shells out to the reviewer model. You run it yourself. This keeps auth, cost, and rate limits in your hands and avoids coupling the plugin to any one vendor's CLI flags.
• Markdown is the only interchange. The review package and review report are plain markdown files. Version them, diff them, commit them if you want an audit trail.
• Confidence-gated. Findings with confidence < 70 are hidden by default. The reviewer explicitly scores itself, so you don't drown in nits.
• Preserve reviewer language. During /multi-model-review:apply-review, the builder quotes the report rather than paraphrasing. Prevents one model from rewriting the other's wording into agreement.
• Template-driven, not hard-coded. Supporting a new reviewer model = dropping a new markdown template. No code change.

Relationship to Spec Kit

This plugin is complementary, not a fork. It assumes you are already using Spec Kit for /speckit.specify, /speckit.plan, /speckit.tasks, /speckit.implement. If you aren't, multi-model-review still works in ad-hoc mode — it just has less context to hand the reviewer.

Spec Kit's own /speckit.analyze, /speckit.checklist, and /speckit.verify commands are single-model checks. multi-model-review adds the cross-model loop.

License

MIT. See LICENSE.

Contributing

Issues and PRs welcome. This is an early version (0.1.0) and the schema for review-report.md may still change. Adding templates for new reviewer models is especially appreciated.