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.

Dependency Ownership

Inspect the source pin and actual installed package-manager metadata separately. Verify lockfile compatibility, workspace dependency graph, patches, platform, linker layout, and resolved dependency paths before reuse. A matching store path or package-manager major version does not prove compatibility. Do not assume a root dependency symlink supplies every workspace package.

APFS copy-on-write package imports can share storage without sharing a Git checkout. They are not Git snapshots, worktree creation, or cleanup. Do not infer physical reclaim from logical dependency sizes. Inspect the installed package manager's contract before applying an APFS optimization.

Follow the repository and user dependency rules, existing authorization, and shared-symlink guards. If the install is incompatible or missing, use the repository's approved proof route. Do not add a new approval requirement for dependency work already authorized by that route.

Closeout

Report the exact task-owned path, branch, HEAD, owner, and remaining work. Give each task checkout an explicit outcome:

• retained: keep the checkout, with its reason and next owner or action.
• blocked: name the missing proof or permission and the exact unblock action.
• removed: report only after authorized removal verifies path and registration absence.
• unknown: preserve an incomplete removal result; reconcile its exact intent read-only before recovery.

Normal owner finalization includes disposing of routine task output and closing its exclusive checkout when no unfinished work, required evidence or shared consumer needs it, unless the user asked to keep it. Do not request a new cleanup approval for that closeout. Use the native owner release/removal route; task completion does not permit bypassing its guards or touching another owner's workspace. Checkout retention does not keep a completed task active. Record the retained path and its concrete remaining condition once, then finish the requested work. Optional proof or profiling must not become a new closeout requirement. After required validation and delivery, or explicit cancellation/supersession, routine task logs, receipts, test captures, proof archives, build output, and checkpoints are disposable by default. Include them in authorized task cleanup; do not require an archive, external publication, retained copy, or another discard approval. Explicit user discard overrides earlier task-local retention. Required validation and deliverable contracts remain unchanged.

For a known finite task, one execution owner checks current path/registration, Git state, owners/holders/locks, and shared dependencies, uses the installed native closeout route, then verifies the outcome. Use chat for the result; do not impose fleet delegation, a typed host audit, or a new cleanup manifest. Use the installed wrapper or an explicitly reviewed coherent source bundle. Source availability alone does not qualify removal or change the installed tool. Keep unfinished or unknown source, live owners, credentials, other-owner data, shared dependencies, and actual release/customer deliverables protected. Known obsolete amend/rebase/reflog versions of finalized work need no archival refs; this does not authorize shared-store expiry or pruning.

For superseded, cancelled or already-applied work without a matching PR, use gwt cancel --reason <text> only when the installed helper advertises it. Cancellation records the disposition; it need not invent PR proof. Continue through the installed finalized-task closeout route when available. If that route is unavailable, report the concrete native limitation once rather than repeating finish or creating another recovery archive. Ordinary and repository-native checkouts keep their existing authorized lifecycle: removal requires current ownership and live-use checks, protection of unfinished work, and the repository's approved non-force procedure.