Docs · Worker
Worker limits
What the contributor limits screen controls, what is enforced locally, and what is not available yet.
On this page
The Limits screen writes policy.json and restarts the worker sidecar so the
running process matches the form. The desktop core enforces local limits
before it starts the sidecar. A server assignment never overrides them: if a
job exceeds the local policy, the worker rejects it.
What the Limits screen controls
| Setting | Behavior |
|---|---|
| Enable contribution | The desktop core will not start the worker while contribution is disabled. |
| Model runner | Canned, Ollama, Subprocess, or llama.cpp, matching the installed backend. |
| Ollama model and host | The model name and server origin used when the runner is Ollama. |
| VRAM budget | The operator-declared budget used when GPU telemetry is unavailable. Eligibility also applies a 15% safety margin over the plan requirement. |
| Model-cache quota | The configured budget for downloaded model files. |
| Idle-only minutes | The core waits for the configured idle period before starting or resuming work. |
| Active hours | A local schedule with start time, end time, and optional weekdays. Hours may cross midnight. |
| Autostart | Stored as a preference. |
Saving a policy stops or restarts the sidecar so the running process uses the new limits.
Enforcement status
Enforced locally by the desktop core:
enabledidleOnlyMinutesactiveHours
Not enforced yet:
- Model-cache quota. The app does not currently evict cached models. The quota is not a guarantee about free disk space on the machine.
- Autostart. The preference is stored, but the current installer does not register the app at Windows sign-in. OS-level autostart registration is planned for the signed installer.
CPU utilization ceilings, GPU utilization ceilings, temperature limits, and network bandwidth caps are in the worker design but are not exposed by the current desktop policy. Do not assume they are active.
Pause, resume, and revoke
- Pause stops the sidecar gracefully. In-flight work drains, and the app keeps its device session. Resume starts the worker again without the enrollment token.
- Revoke asks the server to remove the worker, then deletes the local
device credentials (
device-key.jsonandrefresh-token) and stops the sidecar. If the server call fails, nothing local is removed, so you can retry.
Uninstalling removes local data from this PC but does not unenroll the worker. Choose Revoke in the app while it is running to remove the worker from the network.
Model notes
LFM3-1.7B is the only live model. Qwen3 models are planned. The desktop UI configures a single worker; advanced pipeline roles are configured outside the UI with environment variables and are documented in the repository.