Use when executing a multi-step plan and each implementation step should be delegated to subagents and independently reviewed.
Use when reviewing local changes — the working-copy diff, a branch, a commit, or a GitHub PR by number — with fresh reviewer subagents that return structured findings.
Use when a test failure, regression, exception, hang, wrong result, or unexpected behavior needs diagnosis — gathers evidence, traces the relevant path, and verifies the cause before recommending a fix.
Resolve jj (Jujutsu) conflicts. Use when jj log/status shows conflicted revisions, a rebase/squash/abandon reports 'new conflicts appeared', or files contain jj conflict markers.
Use when you want to work in an isolated jj working copy — parallel task, experimental scratch, subagent with its own tree. jj's equivalent of git worktrees: creating a workspace, working inside it from anywhere, and cleaning up without losing history.
Find things to improve in the current project
Build a mental model of code changes — guided reading order, conceptual grouping, and context.
Use when the user asks to be exhaustive, thorough, comprehensive, or rigorous, wants a finding drilled into or adversarially verified, or when missing something would be costly — a deliberate deep-work mode that trades speed for rigor.