← ClaudeAtlas

background-runnerlisted

Run any script, command, or work loop as a detached background process that survives across Claude tool-call boundaries and turns, with disk-checkpointed state so it resumes after the VM's idle teardown. Use this WHENEVER the user wants something to run in the background, keep running between turns, run asynchronously while other work happens, persist a long-running or overnight job, fan out bulk I/O work, poll/monitor something on an interval, stand up a daemon/supervisor/poller, or "kick this off and check on it later." Trigger even if they never say "daemon" or "background" explicitly — any "run X while we do other things", "keep this going", "let it churn", "set up a watcher", "collect data over time", or otherwise long-running/async/persistent job qualifies. Also use it to correctly resume such jobs after returning to a cold-rebuilt VM.
williamyates/bgrunner-skill · ★ 1 · Data & Documents · score 64
Install: claude install-skill williamyates/bgrunner-skill
# Background Runner Run things that keep going between your turns, and that come back from disk after the VM is torn down. Two tools ship with this skill; pick one. Built for Claude's hosted code-execution sandbox (claude.ai chat and Cowork). On a persistent local machine — e.g. Claude Code — none of the teardown-survival machinery is needed; plain background execution is enough there. ## First: know the substrate (one paragraph) The VM persists while warm, but **after ~5 minutes idle it is fully torn down — every process dies, including PID 1.** On return you get a **fresh VM with the disk restored**. So detaching a process (double-fork) only buys survival across tool calls and short gaps; **disk is the only thing that crosses a teardown, and `resume` is the real workhorse.** Network egress is rate-governed (~8–25 req/s regardless of worker count), and the box has 1 CPU core (concurrency helps I/O, not CPU). Full details + how to re-measure: `references/environment-model.md` — read it if anything about persistence, timing, or throughput surprises you. ## Pick your tool | You want to… | Use | Why | |---|---|---| | Run an **arbitrary command/script/service** in the background (and optionally auto-restart it) | `scripts/bgrun.py` | named jobs, status/logs/stop, **post-teardown resume**, supervisor/restart | | Write the **work yourself in Python** with checkpoint + skip-already-done | `scripts/work_daemon.py` | resume-safe unit loop; re-run same command to continue | Both