"Why do you write like you're running out of time?"
Go to sleep. Wake up to finished work.
Install · How it works · Compared to
Your overnight Claude session, except it actually finishes. Type /nonstop before you walk away. Claude asks every question now, gets your sign-off on anything dangerous, then works through the night: solving blockers, routing around problems, and knowing when to stop trying.
You hand Claude a 4-hour task and go to bed. You wake up and it stopped 3 minutes in:
"Should I use the existing auth middleware or write a new one?"
Or it hit a permission error on step 2 of 47 and just... sat there. Waiting. For 8 hours.
nonstop makes Claude do what a good engineer does when the boss leaves: figure it out, work around it, or write it down and move on to the next thing.
Paste this into Claude Code:
Fetch and follow the instructions at https://raw.githubusercontent.com/andylizf/nonstop/main/INSTALL.md
/plugin marketplace add andylizf/nonstop
/plugin install nonstop@nonstop
curl -fsSL https://raw.githubusercontent.com/andylizf/nonstop/main/install.sh | bash/nonstop triggers a 4-phase sequence. Claude won't start working until all of this is done.
1. Mental simulation. Claude walks through every step of your task in its head and flags anything that would make it stop:
Here's my plan:
1. Refactor auth module to JWT
2. Update 12 test files
3. Run full test suite
4. Push to feature branch
Potential blockers:
- Test suite needs Redis. Is one running?
- I'll modify .github/workflows. OK?
- Ambiguous: should old session tokens be migrated or dropped?
Confirm before you go?
2. Dangerous ops manifest. Anything that touches the outside world gets explicit approval:
| Examples | |
|---|---|
| Kill processes | GPU jobs, services |
| Delete | rm -rf, git clean, drop tables |
| Force push | git push --force |
| Social | Post PRs, comments, Slack |
| Paid APIs | External calls, webhooks |
| Resources | Claim GPU, large disk |
You can approve all, deny all, or scope it ("delete in /tmp, not /data"). Your call.
3. Permissions check. Claude checks your permission mode. If you're not on bypass or auto, it'll warn you that permission prompts will block it while you're gone.
4. Confirm and go. Scope, workaround policy, ground rules. Then:
"Nonstop mode ON. Go rest, I've got this."
A stop hook blocks Claude from halting. When something goes wrong:
Can I solve it? → Solve it. Keep going.
Can I work around it? → Only if the outcome stays the same. Note what I did.
Truly stuck? → Mark it blocked. Move to next task. Don't spin.
What Claude will never do: brute-force retry, disable safety checks, guess credentials, or take unapproved destructive actions.
Long-running ops (builds, tests) get delegated to background agents. Claude doesn't sit and wait.
/nonstop off, or just check the task list:
- Completed: done
- Workaround: done, but verify the approach
- Blocked: needs you
Two files.
| File | Role |
|---|---|
SKILL.md |
Pre-flight protocol, dangerous ops manifest, blocker decision framework |
nonstop.sh |
Stop hook. Blocks premature stops, tracks nudge count, prevents infinite loops via stop_hook_active |
Session-scoped. Flag file per session ID. Auto-deactivates after NONSTOP_MAX nudges (default 5, export NONSTOP_MAX=0 for unlimited).
| Approach | What's different | |
|---|---|---|
| taskmaster | Must emit TASKMASTER_DONE:: token to stop |
No risk assessment, no blocker handling. Just a completion token. |
| Ralph loop | External orchestrator wrapping the CLI | Full-featured (~2k lines bash), prompt-based status protocol, circuit breakers. Nonstop is two files you drop in. |
| Raw stop hook | {"decision": "block"} |
Blocks blindly. No intelligence. |
| nonstop | Pre-flight + smart blocking + decision framework | Thinks ahead. Handles blockers. Knows when to actually stop. |
MIT