[FEATURE] Automatic pre-flight checkpoint before unattended/autonomous runs

Summary

Before starting an unattended/overnight autonomous run (background subagents, Routines, /loop, cron-driven work), there's no automatic snapshot/checkpoint taken so that a bad autonomous change can be wholesale-reverted the next morning. Recovery today means manually diagnosing what an agent did wrong after the fact. Please add an automatic pre-flight checkpoint before unattended runs.

Current behavior

Claude Code has general undo/checkpoint capabilities in interactive use (e.g. /rewind), but there's no equivalent automatic safety net specifically tied to starting an unattended autonomous run. If an overnight Routine or background-agent run goes wrong, recovering means manually figuring out what changed and reverting by hand — there's no single "restore to right before this run started" action.

Motivation

The desire for genuinely unattended, overnight agent operation is a widely reported theme across the market (e.g. commentary on Replit: "imagine if your agents really could run 24×7"), but it's consistently paired with fear of unrecoverable damage — Replit users specifically report agents causing regressions or catastrophic actions (one case: an agent deleted a production database against an explicit freeze instruction) with no clean way to roll back to before the run started. The autonomy people want and the safety net that would make it trustworthy are two different asks, and only the safety net is currently missing.

Proposed behavior

1. Automatically snapshot repo/workspace state immediately before an unattended run begins (Routine firing, background subagent dispatch, /loop, scheduled cloud agent).
2. Provide a single "revert to pre-run state" action scoped to that specific run, distinct from general-purpose undo, so recovering from a bad overnight run doesn't require reconstructing what changed by hand.
3. Should be automatic/default for unattended contexts specifically (not requiring the user to remember to set it up each time), since the whole point of unattended operation is not needing to babysit it.

Why it matters

• Directly enables trustworthy autonomous operation — the single biggest blocker to "let this run overnight" reported across the market is fear of undoable damage, not lack of autonomy itself.
• Turns "diagnose what a bad overnight run did, by hand" into "one action, revert the run" — much lower recovery cost, which is exactly what makes people willing to actually leave things unattended.

Related issues (adjacent, not duplicates)

• #64615 [BUG] /rewind (Esc Esc) silently reverts/loses code — destructive "Restore code and conversation" is the default with no confirmation — existing general undo mechanism has its own reliability problems; this request is for an automatic pre-flight snapshot specific to unattended runs, not a fix to interactive /rewind itself.
• #353 Undo/Checkpoint Feature for Correcting AI-Generated Code (closed) — general checkpoint request, presumably substantially addressed by existing undo features; this request is narrower — specifically an automatic snapshot gating unattended/autonomous run starts.
• #53610 [Feature] Multi-agent runtime needs mechanical enforcement: 9 gaps that defeat unattended overnight operation (closed/stale) — broader unattended-reliability meta-issue; this is one specific, narrower mechanism within that space.

Notes

• Reviewed for duplicates before filing (multiple search passes: "rollback checkpoint safety net before unattended overnight autonomous run", "snapshot before autonomous run revert bad changes", "checkpoint restore point unattended agent session") — nothing found proposing an automatic pre-flight snapshot specifically tied to unattended-run starts, as of 2026-07-19.
• Sourced from market research (SaaStr commentary on Replit 24×7 autonomy desire; tryorbye.com Replit incident reports of destructive actions during unattended runs), cross-referenced against Claude Code's existing (but separate, interactive-only) undo mechanism.