Review Team

Review Team — Index

Compose a multi-agent review team: pick agents, create the team, coordinate, synthesize. Each agent's full brief — scope, focus areas, how it reports — lives in its definition file at agents/<name>.md. Spawn teammates with task-specific context from there.

Agents are generic and stack-agnostic; every one inherits the session's tools, so any agent can both review and do the work in its domain.

Agents

Recipes

Starting points — drop agents that don't apply, add any the work touches.

devils-advocate is not a domain reviewer — add it to any recipe for a high-stakes, hard-to-reverse, or gate decision (ship / merge / architecture sign-off). Run it after the domain reviewers report and their findings are consolidated, not in parallel — it needs a settled direction to attack. Skip it on routine PRs.

Running a team

Requires CLAUDE_CODE_EXPERIMENTAL_AGENT_TEAMS=1. Create the team with one task per review dimension; spawn each teammate with task-specific context (they don't inherit conversation history) and ask for severity-rated findings with file:line references. The lead coordinates and synthesizes findings — it does not implement.

Synthesize the teammates' reports into one severity-ordered summary: dedupe overlapping findings (note which agent raised each), surface Critical/High first, and flag any disagreements between agents rather than silently resolving them.

When not to use a team: teams cost several times more than a single agent. For small, sequential, or token-sensitive work — or a single-dimension check one agent finishes in under ~5 minutes — use one agent or a plain subagent instead.

Source: jedwards1230/claude-plugins — distributed by TomeVault.