Parallel code review

I run every reviewer at once over the local diff and merge what they find into one ranked list. Fast, and no reviewer sees another one work, so the findings stay independent.

Workflow

Step 1: Determine Scope

• When the user specifies files, I review those files.
• Otherwise I review all unstaged and staged changes through git diff HEAD.
• I pin the comparison base before dispatch. When the caller supplies a merge-base SHA, as done does, I use it. Otherwise I resolve git merge-base HEAD origin/<default branch>, reading the default from gh repo view --json defaultBranchRef. Reviewers need it as context to tell base-branch behavior from behavior that earlier commits on this branch introduced. It never widens the review target, which stays the Step 1 scope.
• When a prior convergence artifact exists, I invoke converge-reviews with the current baseline, diff hash, paths, request, and planned roster and lenses before I dispatch. I reuse an unchanged result. When it returns continue, I dispatch only the invalidated coverage it names. I apply any other result without starting another review round.

Step 2: Build the roster, then dispatch

I name the roster first, then launch every member in parallel with the Agent tool.

Members 1-3 below are subagent_type values, and I pass them to the Agent tool as-is. Member 4 is a skill-running agent. /web-interface-guidelines, /ui-skills, and /rams are skills, not agent types, so they cannot go as subagent_type. I dispatch a general-purpose agent that invokes them instead.