Isolate edits and deliver locally

Read project AGENTS.md and CLAUDE.md first. Resolve the physical live Git checkout using Git's top-level path and symlink resolution. If launched inside a linked worktree, recover the live checkout from its common Git directory and git worktree list --porcelain.

Read-only tasks need no editing checkout. An explicit assignment to an existing task worktree or request to edit the current checkout overrides new isolation; verify the assigned checkout. If there is no Git repository, disclose the limit and do not initialize replacement history.

For ordinary editing, establish one linked worktree before detailed planning or investigation. Derive its path from the physical live root:

LIVE = physical live checkout
PARENT = LIVE's immediate parent
PROJECT = LIVE's basename
TYPE = conventional commit type
FEATURE = one descriptive task slug

Assign those values once and construct both names with the same shell variables:

WT="$PARENT/.worktrees/$PROJECT/${PROJECT}_${TYPE}-${FEATURE}"
BRANCH="$TYPE/${PROJECT}_${FEATURE}"
git -C "$LIVE" worktree add -b "$BRANCH" "$WT"

Keep the FEATURE's hyphens unchanged in the branch. The branch separates PROJECT and FEATURE with _; the directory separates TYPE and FEATURE with -. Keep these separators exactly as shown above. Resolve existing store symlinks; the store must remain under LIVE's immediate parent, outside LIVE. On collision, choose a unique FEATURE suffix and recompute both names; never delete or reuse another task. The first repository mutation is git worktree add -b BRANCH WT, not source edits or git init. Create WT/tmp/ only after that command succeeds; precreating it creates WT and makes Git reject the destination as already existing.

Retain LIVE, WT and BRANCH for the whole task. Before edits/commits and at handoff, verify WT's physical path, registered location and branch against those retained values. Use absolute task paths or an explicit tool working directory; a shell cd does not persist. Reuse this checkout and branch for every Feature, repair and documentation commit. The store holds registered worktrees only; scratch/recovery files belong inside WT/tmp/, never PARENT/.worktrees/PROJECT/tmp/. Every write and tool working directory uses the full retained task checkout path.

After each completed Feature or focused repair:

1. Commit its working code and checks in WT. Confirm HEAD advanced and source changes are committed; incidental caches may stay unstaged. A README-only commit cannot deliver uncommitted source or checks.
2. Verify the live checkout is on its actual default branch using repository metadata or project guidance. Preserve unrelated live changes and merge with git merge --no-ff BRANCH; resolve conflicts without discarding work.
3. Confirm task HEAD is an ancestor of live HEAD and inspect live status.
4. Reapply the affected consumer through its installer, selector or narrow reload, and verify installed content and relevant logs. Follow selected Linux desktop rules. For static content with no installed consumer, inspect the merged files.

Finish this delivery before the next Feature. Documentation follows the same checkout, commit and merge path. Repeat delivery after every later task commit, including ledger/doc follow-ups; a previous merge does not deliver a new HEAD. At handoff, the current task tip must be an ancestor of live HEAD. Recover only this task's misplaced drafts, without overwriting user work. If WT's location was wrong, use git worktree move to the unused correctly derived path while preserving branch and contents.

Report completion only after local default-branch delivery and consumer checks. Never push, publish, add remotes or rewrite history unless explicitly requested.