tamirs-superpowers / rules

---
description: "Objective/worker/integration worktrees — provider is metadata, not path identity. Canonical: rules/dev/git-worktree-agent-workflow.md"
alwaysApply: false
globs: ".agent-worktrees/**/*,.claude/.worktrees/**/*,.cursor/.worktrees/**/*,.codex/.worktrees/**/*"
---

<!-- GENERATED FILE — DO NOT EDIT.
     Source:    rules/dev/git-worktree-agent-workflow.md (frontmatter + summary: platforms/cursor/rules.map.json)
     Generator: scripts/build-cursor-rules.sh
     Regenerate: make cursor-rules -->

# Git Worktree Agent Workflow

Canonical rule: `rules/dev/git-worktree-agent-workflow.md`. Read it before creating any worktree — the summary below is orientation only.

One objective → zero or more worker worktrees → one integration worktree. The path and branch encode the **objective and task**, never the provider.

```text
.agent-worktrees/<objective>/{integration,task-001,task-002}
main → objective/<slug> → worker/<slug>/NNN
```

A worker ends at **commit + handoff**, never at a PR. Delivery happens once, from the integration worktree.

## Cursor-specific

Cursor has no built-in worktree tool and no `EnterWorktree` equivalent — create worktrees explicitly before editing tracked files from an agent session:

```bash
skills/dev-workflow/_shared/scripts/resolve-worktree.sh --objective <slug> --task 001
```

Legacy `.cursor/.worktrees/<name>` worktrees remain valid and are never orphaned; `resolve-worktree.sh --list` shows them and `--migrate` moves one opt-in.