Copilot Plugin for Claude Code

Use GitHub Copilot models from inside Claude Code — cross-model code reviews, multi-model council reviews, and task delegation.

This plugin lets Claude Code users tap into Copilot's model lineup (GPT-5.4, GPT-5.3-Codex, Claude Opus 4.6, etc.) without leaving their workflow. The headline feature is council reviews: the same diff reviewed by multiple models in parallel, with Claude synthesizing the results.

What You Get

Command	What it does
/copilot:review	Code review using any Copilot model (default GPT-5.4)
/copilot:adversarial-review	Challenge review — questions design, architecture, and assumptions
/copilot:council	Multi-model council review — parallel reviews from 3 models, synthesized by Claude
/copilot:delegate	Hand a task to Copilot (bug investigation, fixes, research)
/copilot:status	Check running and recent jobs
/copilot:result	Get the output of a completed job
/copilot:cancel	Cancel an active background job
/copilot:setup	Check installation, auth, and settings

Council Reviews

The unique feature. /copilot:council sends your diff to multiple models simultaneously and synthesizes the results:

/copilot:council

Default models: GPT-5.4, Claude Opus 4.6, GPT-5.3-Codex (configurable).

The synthesis identifies:

• Agreements — issues flagged by multiple models (high confidence)
• Disagreements — conflicting assessments (needs human judgment)
• Unique catches — issues only one model spotted (edge case signals)
• Overall assessment — combined verdict on code quality

Custom model selection:

/copilot:council --models gpt-5.4,gpt-5.4-mini,gpt-5.3-codex

Adversarial Reviews

/copilot:adversarial-review is not a stricter bug hunt — it's a design challenge. Instead of finding implementation defects, it questions whether the approach itself is right:

/copilot:adversarial-review

Focus on a specific concern:

/copilot:adversarial-review is this abstraction worth the complexity?

The review covers:

• Design choices — right abstraction? right layer?
• Architecture tradeoffs — what does this sacrifice? what breaks if requirements change?
• Assumptions — what is the code assuming? are those assumptions safe?
• Missing alternatives — was a simpler approach considered?
• Failure modes — what happens when this fails at 3am?

Review Gate

Optionally block commits that have critical issues. Enable in .claude/copilot.local.md:

- review-gate: on

When enabled, every git commit runs a fast Copilot review first. It only blocks for critical issues:

• Security vulnerabilities (hardcoded secrets, injection, auth bypass)
• Data loss risks (destructive operations without safeguards)
• Obvious production-crashing bugs

Style issues, missing tests, and subjective preferences are always allowed through. The gate uses gpt-5.4-mini by default for speed (configurable via review-gate-model).

The gate fails open — if Copilot is unavailable, times out, or errors, the commit proceeds normally.

Requirements

• GitHub Copilot subscription — usage counts toward your Copilot limits. Learn more.
• Copilot CLI installed — via npm, VS Code extension, or manual install.
• Node.js 18+

Install

Add the marketplace and install:

/plugin marketplace add mongo-ai/copilot-claude-code-plugin
/plugin install copilot

Reload plugins:

/reload-plugins

Then check everything is working:

/copilot:setup

If Copilot CLI isn't installed yet:

npm install -g @github/copilot-cli

If Copilot isn't authenticated:

!copilot login

Usage

Code Review

Review your working tree changes:

/copilot:review

Review against a base branch:

/copilot:review --base main

Use a specific model:

/copilot:review --model gpt-5.3-codex

Run in the background:

/copilot:review --background
/copilot:status
/copilot:result

Task Delegation

Hand off work to Copilot:

/copilot:delegate investigate why the auth middleware is returning 403 for valid tokens

Use a specific model:

/copilot:delegate --model gpt-5.4 refactor the database connection pool to use async initialization

Available Models

Any model available through the Copilot CLI can be used:

Model	Best for
gpt-5.4	Complex reasoning, thorough reviews (default for review)
gpt-5.3-codex	Code-specialized tasks (default for delegate)
claude-opus-4.6	Nuanced analysis, long context
gpt-5.4-mini	Quick, low-cost tasks

Configuration

Create .claude/copilot.local.md in your repo to customize defaults:

- default-model: gpt-5.4
- default-effort: high
- council-models: gpt-5.4, claude-opus-4.6, gpt-5.3-codex
- council-effort: high
- copilot-binary: auto
- review-gate: off
- review-gate-model: gpt-5.4-mini

Setting	Default	Description
default-model	gpt-5.4	Model for /copilot:review
default-effort	high	Reasoning effort level
council-models	gpt-5.4, claude-opus-4.6, gpt-5.3-codex	Models for council review
council-effort	high	Effort level for council reviews
copilot-binary	auto	Path to Copilot CLI binary, or auto to detect
review-gate	off	Set to on to gate commits on Copilot review
review-gate-model	gpt-5.4-mini	Model for review gate (fast + cheap)

How It Works

The plugin spawns the Copilot CLI directly — no SDK dependency, no extra npm install. It finds the binary automatically (PATH, VS Code extension paths, or manual config).

For council reviews, all models run in parallel via Promise.allSettled. If rate limiting is detected, it falls back to sequential execution with delays.

Background jobs are tracked as JSON files in .copilot-jobs/ (gitignored) with PID-based lifecycle management.

Architecture

commands/                Slash command definitions (8 commands)
  council.md             Multi-model council review (unique feature)
  adversarial-review.md  Challenge review — design & architecture
  review.md              Single-model code review
  delegate.md            Task delegation
  setup.md               Installation check
  status.md              Job status
  result.md              Job output
  cancel.md              Job cancellation
agents/
  copilot-delegate.md    Subagent for task delegation
hooks/
  hooks.json             SessionStart + review gate hooks
scripts/
  copilot-companion.mjs  Main entry point
  review-gate-hook.mjs   PreToolUse hook for commit gating
  lib/
    copilot.mjs          CLI binary discovery + process spawning
    git.mjs              Diff collection + review size estimation
    state.mjs            Background job tracking (JSON file-based)
    settings.mjs         Configuration reader (.claude/copilot.local.md)
tests/                   83 tests across 4 modules

License

Apache 2.0 — see LICENSE.

Architecture inspired by openai/codex-plugin-cc.