operations-worktree

Operations Worktree

Purpose

Create and manage worktrees safely and consistently across projects while avoiding stale branch bases.

When to use

• You need a new task branch in a new worktree.
• You are juggling many concurrent worktrees.
• You need to avoid branching from stale local main or local HEAD.

Workflow

1. Identify the producer: shell GWT, a repository-native wrapper, or an application that owns sessions or sandbox projections. Resolve the repository identity, canonical owning checkout, Git common directory, and existing worktree registrations. A .git entry does not make an application-owned workspace a personal GWT task; preserve its native allocation, relocation and retention lifecycle. Preserve dirty owner state. Read the repository's worktree and dependency rules.
2. Inspect existing worktrees before creating another checkout. Reuse an existing task-owned worktree only when its identity, branch, and ownership match. Do not adopt another session's checkout.
3. Use a healthy owning checkout accepted by the installed wrapper. An unsafe, shallow, or promisor source is not permission to create an ad hoc clone. Do not bypass worktree creation with raw Git, copied repositories, or temporary clones. A refusal stops creation, not diagnosis. Repair that same verified owner within existing task authorization, then let the wrapper recheck it. Report the concrete blocker if no qualified owner is available.
4. Use gwt help for discovery and gwt root to resolve the configured managed root. Keep new worktrees under that root or a repository-native managed root. Do not place new checkouts in arbitrary sibling or temporary directories. For server placement policy, use the existing account-specific DOTFILES_WORKTREES_ROOT setting; preserve unrelated workstation defaults. Check the host policy, actual account homes, physical path, mount and permissions separately: gwt root reports a value, not an access boundary. Changing this setting neither relocates existing work nor controls application allocators. Use the native owner to update registrations and persisted references; do not substitute shell GWT for an application's workspace lifecycle.
5. Verify the download policy before any fetch. Task-required additive fetch or unshallow is not GC, repacking, pruning, or owner consolidation. Recheck current state; do not repeat completed repair or ask again for an already authorized step. Respect contention, suppress automatic maintenance/pruning, and preserve local branches, HEAD/index, patches, and registrations. Distinguish a verified current remote base from a cached local ref. Cached refs do not prove the latest head. When freshness cannot be verified, report that gap rather than silently substituting a stale base.
6. Before creation, choose the repository-prescribed, documented closeout route supported on this host.
   • Use gwt new <branch> <start-point> --finish-managed when required, requested, or supported by a qualified closeout route. An advertised flag alone does not qualify that route. The start-point is optional. Keep the runtime CODEX_THREAD_ID. Outside Codex, set a stable task-specific GWT_OWNER_ID before creation.
   • Where repository policy permits, use ordinary gwt new <branch> <start-point> from the verified owner and configured root. Preserve its normal ownership, recovery, and non-force removal guards.
   • Repository-native wrappers use their documented creation and closeout commands. Finish enrollment is optional and creation-only. It records ownership and does not authorize removal. It is not a prerequisite for development, maintainer work, or ordinary authorized cleanup; keep repository-native lifecycles independent. Never unenroll, recreate, relocate, or adopt an existing managed tree to bypass its record. Stop creation if the wrapper is unavailable or still refuses the repaired owner. Do not substitute an unrestricted raw-Git creation route.
7. Verify the returned path, managed root, registration, owner, branch, and HEAD. For personal finish, confirm enrollment with gwt finish-status before starting work.
8. Verify dependency reuse under Dependency Ownership.
9. Read references/task-artifacts.md before work that retains evidence, publishes expensive artifacts, or needs a resumable phase/blocker receipt.
10. For OpenClaw, use its current AGENTS.md and native review/release lifecycle helpers. Do not override their exact-head or evidence rules here.
11. Finish with the explicit Closeout result.