Skip to content

Repository files navigation

Deep Review

Deep Review is a configurable, review-only Agent Skill that coordinates specialized reviewers over repository changes and returns one result: ready, blocked, or incomplete.

Install

Paste the three verified-install commands for your operating system from the v1.1.1 installation guide. The guided installer asks for your AI client and user or project scope, then shows the destination before changing anything.

Quick start

  1. Install Deep Review for your AI client.
  2. Open the repository you want to review.
  3. Ask your client:
Use the deep-review skill to review my current repository changes.

The final status tells you what to do next:

Status Meaning
ready The configured review completed without blockers. Run the repository's normal tests and checks separately.
blocked One or more findings must be addressed before readiness. Fix them and run Deep Review again.
incomplete Required evidence was unavailable or could not be reviewed safely. Resolve the reported gap and retry.

Deep Review does not edit consumer source files or apply fixes. A ready result is review readiness, not proof that builds, tests, linters, or other project validation passed.

Choose a review scope

Natural language works in every supported client. Clients may also expose a prefix:

Client Example
Claude Code /deep-review --base main
Codex CLI and IDE $deep-review --base main
Devin @skills:deep-review --base main
Windsurf Cascade @deep-review --base main

Use the syntax supported by your client with one of these selectors:

Task Selector
Review current changes no selector
Compare with a base branch --base main
Review a Git range --range release..HEAD
Review a file or directory --path src/payments
Review a GitHub pull request --github-pr 123
Review a GitLab merge request --gitlab-mr 123
Prioritize a concern --focus "pay particular attention to retry behavior"

Remote shorthand is also supported: Use deep-review #123 infers the provider from the trusted repository remote, while Use deep-review !123 explicitly selects GitLab. Prefer --github-pr or --gitlab-mr when provider selection should be unambiguous.

Remote reviews require the corresponding authenticated GitHub or GitLab command-line client. See Installation.

Important safety boundary

A skill stored inside a repository is trusted client configuration. Do not initiate a review from an untrusted checkout that can replace that copy. To review a pull request, merge request, or fork safely, start the client in a separate trusted checkout and invoke a user- or administrator-installed Deep Review with --github-pr or --gitlab-mr. See the installation trust model.

Diffs, descriptions, paths, checklists, and reviewer focus are treated as untrusted input. Deep Review fails closed when it cannot establish the required evidence or filesystem guarantees.

What it reviews

The general roster covers correctness, security, architecture, simplification, documentation, CI, and an optional project checklist. Matching changes also receive repository-neutral language review for:

  • TypeScript
  • Python
  • Swift
  • Java
  • JavaScript
  • Groovy, including Gradle Groovy DSL and Jenkins Shared Libraries
  • Kotlin, including Gradle Kotlin DSL
  • C#, including Razor C# regions

Language reviewers own only their documented language-semantic rules. General correctness, security, architecture, and simplification remain with the corresponding general reviewers.

Deep Review safely bounds descriptions, model input, context reads, parallel work, results, and cache records. Oversized reviews are split and synthesized under package-owned limits; unavailable coverage produces incomplete, never a false ready. Exact behavior is defined by the runtime contracts.

Configuration

Configuration is optional. Defaults enable all matching reviewers and make high, medium, and checklist failures blocking. Create .deep-review/config.toml only when you need to change behavior:

[language_agents]
disabled = ["swift"]

[language_rules]
disabled = ["typescript.no-explicit-any", "python.runtime-assert"]

See Configuration for task-oriented examples, supported language-rule IDs, triggers, checklists, and project-specific reviewers.

Documentation

The files under skills/deep-review/references/ are the normative runtime contract. User guides describe tasks and observable behavior; the references own exact algorithms, schemas, and limits.

License

Deep Review is available under the MIT License.

About

Read-only mirror of the canonical GitLab repository, maintained for Anthropic plugin distribution.

Resources

Contributing

Security policy

Stars

0 stars

Watchers

0 watching

Forks

Releases

Packages

Contributors

Languages