deep-review

Pass the PR number or URL and the skill fetches the diff with gh, creates a detached git worktree, and launches five parallel review subagents covering implementation, design, correctness, runtime, and security. Findings are verified by the parent agent and posted as one PR review with inline comments.

deep-review uses a single model with five different review perspectives, while fusion-review compares multiple models on the same perspective. They are complementary and can be used together depending on whether you need viewpoint diversity or model diversity.

No. Subagents run read-only on a disposable git worktree and may only write one intermediate review file each. All repository fixes, PR comments, and final reports are produced by the parent agent after reading and verifying those files.

Checks like DRY violations, convention consistency, and caller impact cannot be judged from diff text alone. The worktree gives each reviewer the full codebase so it can grep for call sites and compare against existing patterns.

Avoid it for quick checks, since launching five subagents is costly. For lightweight review of uncommitted changes use a local-reviewer agent, and for triaging many waiting PRs use a dedicated PR-processing skill instead.