branching-workflow

# Branching workflow `/rewind` (Claude Code v2.0.0+): fork from a prior assistant turn, take a different forward path. The abandoned branch's tokens leave context entirely. The cleanest zoom-out primitive available natively in the CLI. This skill: WHEN to branch, WHAT to preserve first, HOW to recover. Pairs with the Critic skill (a BLOCK verdict often triggers a branch). ## When to branch Branch (use `/rewind`) when ALL three apply: 1. **The current path is going off the rails.** Output is wrong, the agent is confidently asserting unverified claims, the classifier flagged high-stakes but the Critic says BLOCK, or a path-of-least-resistance failure (closure-bias / search-truncation / sycophancy) has visibly compromised the conversation. 2. **Continuing-with-corrections would be more polluting than rewinding.** "Let me reconsider" + retry-emit will leave the bad branch in context as residue. Rewinding leaves it cleaner. 3. **The bad path started recently enough.** If the divergence is 1-3 turns back, rewind. If it's 10+ turns back, /save the insights + /compact + start fresh with prepended context. Do NOT branch when: - The output is acceptable but imperfect (continue-with-revision is cheaper). - You're in the middle of a long-running task with persistent state (database writes, sent messages, scheduled tasks fired). Branch does NOT undo external side effects. - You're under 50 turns into the session (the cost of pollution is low; rewind overhead doesn't pay back yet). ...