opencode-adversarial-review

Why?

Provides an adversarial code review agent that challenges implementation approach and design choices, not just finding bugs. Uses a clean-context subagent so the review is unbiased by conversation history.

Philosophy: Extending OpenCode

OpenCode is designed to be highly extensible. This plugin hooks into the OpenCode lifecycle to provide additional functionality seamlessly into your workflows. This plugin is modeled after the Codex CLI's adversarial review system, focusing on "breaking confidence" rather than validating changes.

Architecture

graph TB
    subgraph Host [OpenCode Host]
        User([User]) -->|/adversarial-review| Command[Command Template]
    end

    subgraph Injection [Context Injection]
        direction LR
        Command -->|!git status| S[Status]
        Command -->|!git log| L[Commits]
        Command -->|!git diff| D[Diff Summary]
        Command -->|!cat| U[Untracked]
    end

    subgraph Adversary [Adversarial Subagent]
        direction TB
        Reviewer{Adversarial Reviewer}
        S & L & D & U --> Reviewer
        
        Reviewer -->|Self-Collect| Tools[Explorer Tools]
        Tools -->|Source Code| Reviewer
    end

    Reviewer -->|Structured JSON| User

    classDef host fill:#bbf,stroke:#333,stroke-width:2px;
    classDef collection fill:#dfd,stroke:#333,stroke-width:1px,stroke-dasharray: 5 5;
    classDef adversary fill:#f9f,stroke:#333,stroke-width:2px;
    
    class Host host;
    class Collection collection;
    class Adversary adversary;

The plugin uses a stateless subagent architecture. When the command is invoked, a deterministic shell script collects git metadata (diff stat, untracked files, recent commits) and injects it into a fresh subagent session. This ensures the review is unbiased by the primary agent's conversation history.

Features

• Adversarial Persona: A specialized subagent prompted to find reasons not to ship, prioritizing auth gaps, race conditions, and data loss.
• Dynamic Context Collection: Automatically switches between full-inline diff (for 1-2 files) and high-level stat (for larger changes).
• Subagent Self-Collection: The subagent uses whitelisted git and file tools to "pull" the specific file contents it needs based on the initial diff summary.
• Structured JSON Output: Returns findings with severity, file locations, confidence scores, and concrete recommendations.
• Configurable Scope: Support for --base <ref>, --scope branch, and --scope working-tree.

Install

Add the plugin to opencode.json or opencode.jsonc:

{
  "plugin": ["@capybearista/opencode-adversarial-review"]
}

You can also install it through the CLI:

opencode plugin -g @capybearista/opencode-adversarial-review    # global install
opencode plugin @capybearista/opencode-adversarial-review    # project-local install

Usage

Run a review on your current working tree changes:

/adversarial-review

Arguments

| Argument | Values | Description | | :--- | :--- | :--- | | --scope | auto, working-tree, branch | The range of changes to review. Defaults to auto. | | --base | <git-ref> | The base reference (branch or commit) to compare against when using branch scope. |

• auto: Reviews using the broadest context the provided information supports.
• working-tree: Reviews staged and unstaged changes against HEAD.
• branch: Reviews all changes on the current branch since it diverged from the upstream or main branch.
• focus ...: Any trailing text is treated as a focus area for the review.

Examples

Force review of only working tree changes (ignoring branch history):

/adversarial-review --scope working-tree

Review all changes on the current branch (automatically finds the fork point):

/adversarial-review --scope branch

Review a specific branch against its fork point from main:

/adversarial-review --base main

Review with a specific focus area:

/adversarial-review --base main focus on race conditions in the auth middleware

Configuration

| Property | Type | Description | | :--- | :--- | :--- | | agent.adversarial-review.model | string | The model to use for the subagent (defaults to openai/gpt-5.4). | | agent.adversarial-review.temperature | number | Sampling temperature (defaults to 0.1 for deterministic review). | | command.adversarial-review.template | string | The shell-injected template used to gather git context. |

This plugin requires no manual configuration out of the box, but you can override the subagent model in your local opencode.json.

Permissions & Security

This plugin is designed with a "least privilege" security model. The adversarial subagent is strictly sandboxed to prevent accidental or malicious modifications to your codebase:

• Edit: Explicitly denied (edit: deny).
• Bash: Restricted to a read-only whitelist of 12 Git patterns (e.g., git diff, git log, git status). All other shell commands are blocked.
• Read/Grep/Glob: Allowed (read-only) to enable code inspection.
• Network: Web fetch and search are disabled to ensure the review remains focused on the files at hand.

Troubleshooting

• "No changes to review": Ensure you have staged or unstaged changes, or use --base to review a committed branch.
• Model timeouts: Large diffs may require a model with a larger context window or more time. The subagent is whitelisted to read files if the diff stat is too long to include.
• Git errors: Ensure you are running within a git repository. The command relies on git being available in your PATH.

Contributing

This package lives in the opencode-plugins monorepo.

• Run bun run build, bun run typecheck, bun run lint, and bun test before opening a PR.
• Keep the plugin focused on the adversarial code review function.
• Prefer small, direct changes.

Please open an issue or check for existing ones before creating a pull request.

License

MPL-2.0