Start isolated work from a known base
A worktree session uses a real git checkout under ~/.grok/worktrees. By default it starts from current HEAD and includes uncommitted changes; use --ref when you need a clean, named base.
A named worktree keeps discovery and cleanup understandable. Each checkout is detached at its base, so landing the result still uses normal git review, commits, and integration.
grok -w
grok --worktree=feat "refactor module X"
grok -w --ref main "fix the flaky test"Delegate bounded questions to subagents
Subagents are independent child sessions with their own context and return a summary to the parent. The built-in general-purpose type has full capabilities; explore can read, list, and search without shell or edits; plan drafts without shell or edits.
Split work by independent ownership: a package, hypothesis, test surface, or review concern. Give each child a concrete output contract and ask the parent to verify evidence rather than concatenating summaries.
- Use explore for repository mapping.
- Use plan for independent implementation proposals.
- Use a worktree when a child may edit files.
Fork, inspect, and clean up deliberately
Use /fork --worktree to branch the current conversation into an isolated checkout. Worktrees persist after a session ends, and garbage collection runs only when invoked, so housekeeping is part of the workflow.
List and inspect before removal. The rm subcommand supports --dry-run; use it to resolve exact targets, then remove only worktrees whose changes have been integrated or intentionally abandoned.
/fork --worktree
grok worktree list
grok worktree show <id>
grok worktree rm --dry-run <id>
grok worktree gc --max-age 7d