gpu-joblisted
Install: claude install-skill skyeyesec333/gpu-job
# GPU job: borrow a shared GPU politely
A shared GPU box usually has someone else on it: an always-loaded model, another agent session, the owner's own work. This skill runs your job there as a guest: it checks the box first, starts only when the rules say there is room, steps aside the moment the owner asks, keeps a status file others can read, and leaves its working files on the box.
```
0. recruit a GPU PC on your network: Tailscale, SSH, driver (once per PC, the owner's hands)
1. config gpu-workers.json: the workers, their lanes and limits (once)
2. preflight scripts/gpu-job.sh preflight <worker> (read-only)
3. tell other sessions sharing the box: what, which lane, how long
4. run scripts/gpu-job.sh run <worker> --id … --purpose … --eta … -- <command>
5. watch scripts/gpu-job.sh status <worker> <id>
6. fetch scripts/gpu-job.sh fetch <worker> <id> <dest> (copy only)
7. tell the other sessions it is done; release the lane if it had a grant
```
## 0. Recruit a worker (new PC)
The scripts are in this skill's own `scripts/` folder (`${CLAUDE_SKILL_DIR}/scripts/`, for a plugin or a copied skill alike). They are not on the PATH: call them by that path. Below, `recruit.sh` and `gpu-job.sh` mean those files.
When the user wants to use a PC that isn't a worker yet, or a worker can't be reached, follow [references/recruit.md](references/recruit.md) and check each layer with `r