# Rules that will bite you

Five things that are not obvious and have each cost someone time.

1. **Public hostnames, never internal DNS.** From a workspace, `https://litellm.k8s.semanticscience.org` — not `litellm.litellm-prod.svc.cluster.local`. Internal names are blocked by policy. It looks like the service is down; it is not.

2. **GPU workspaces: keep `cpu` ≤ 6.** g1 has 32 cores for three cards, and every VM costs one more. The template does not enforce this yet; the scheduler does, by leaving an oversized request Pending with no message. If a GPU workspace will not start, lower `cpu` first; if it still will not, it is waiting on a *card* — check who holds the three.

3. **Stop GPU workspaces you are not using.** Running holds the card, idle or not.

4. **`/mnt` is ephemeral.** Only `/home/coder` survives a rebuild.

5. **No scale-to-zero on models.** A served model occupies its hardware until removed. New models are swaps.

Two smaller ones:

- The `model` field is the **gateway name**. `gemma4-g1`, not the HF repo.
- **Prefix nomic inputs** with `search_query: ` / `search_document: `.