Skip to content

Installation

Adversarial Review supports three installation paths with different security guarantees.

The full multi-agent architecture with enforced isolation requires Claude Code's Agent tool.

Option A: Plugin commands

From inside a Claude Code session:

# One-time marketplace registration
/plugin marketplace add ugiordan/adversarial-reviewing

# Install globally (works in every project)
/plugin install adversarial-reviewing@ugiordan-adversarial-reviewing

After installation, start a new session. The skill activates automatically when relevant, or invoke directly via /adversarial-reviewing.

To update later:

/plugin update adversarial-reviewing

Option B: Manual setup

If /plugin commands are unavailable:

  1. Clone the marketplace repo:

    git clone https://github.com/ugiordan/adversarial-reviewing.git \
  $HOME/.claude/plugins/marketplaces/ugiordan-adversarial-reviewing

  2. Copy the plugin to the cache:

    mkdir -p $HOME/.claude/plugins/cache/ugiordan-adversarial-reviewing/adversarial-review/1.0.0
rsync -a $HOME/.claude/plugins/marketplaces/ugiordan-adversarial-reviewing/adversarial-review/ \
  $HOME/.claude/plugins/cache/ugiordan-adversarial-reviewing/adversarial-review/1.0.0/
cp $HOME/.claude/plugins/marketplaces/ugiordan-adversarial-reviewing/.claude-plugin/marketplace.json \
  $HOME/.claude/plugins/cache/ugiordan-adversarial-reviewing/adversarial-review/1.0.0/.claude-plugin/

  3. Add to ~/.claude/settings.json:

    {
  "enabledPlugins": {
    "adversarial-reviewing@ugiordan-adversarial-reviewing": true
  },
  "extraKnownMarketplaces": {
    "ugiordan-adversarial-reviewing": {
      "source": {
        "source": "git",
        "url": "https://github.com/ugiordan/adversarial-reviewing.git"
      }
    }
  }
}

  4. Add to ~/.claude/plugins/installed_plugins.json (inside the "plugins" object):

    "adversarial-reviewing@ugiordan-adversarial-reviewing": [
  {
    "scope": "user",
    "installPath": "<HOME>/.claude/plugins/cache/ugiordan-adversarial-reviewing/adversarial-review/1.0.0",
    "version": "1.0.0",
    "installedAt": "<ISO-8601-timestamp>",
    "lastUpdated": "<ISO-8601-timestamp>",
    "gitCommitSha": "<current-commit-sha>"
  }
]

Cursor (degraded single-agent mode)

Cursor cannot spawn isolated sub-agents. The plugin adapts to a sequential persona mode where the agent role-plays each specialist one at a time. Both code and strategy profiles work, but without enforced isolation between specialists.

# Clone the repo
git clone https://github.com/ugiordan/adversarial-reviewing.git $HOME/.adversarial-review

# Copy rules to your project
mkdir -p .cursor/rules
cp $HOME/.adversarial-review/.cursor/rules/adversarial-reviewing.mdc .cursor/rules/

Limitations in Cursor mode

  • No enforcement boundary between specialist personas (isolation is advisory)
  • Mediated communication is advisory only (no programmatic sanitization)
  • Output validation depends entirely on agent compliance
  • Both profiles work, but findings may be less independent than in Claude Code's isolated multi-agent mode

AGENTS.md (universal)

For any AI tool that supports AGENTS.md or similar context injection:

# Clone the repo
git clone https://github.com/ugiordan/adversarial-reviewing.git $HOME/.adversarial-review

Reference or inline AGENTS.md in your AI tool's context. Feature set depends on the tool's capabilities.

Dependencies

  • bash 4.0+
  • python3 (JSON serialization and unicode normalization)
  • git (for --diff change-impact analysis)
  • GitHub MCP tools (optional, for --triage pr:<N>)
  • Claude Code Agent tool (for full multi-agent feature set)
  • No npm or pip packages required