---
type: cluster-offload
coverage: 2
source: COMPLAINT_REGISTRY.jsonl via taxonomy.json _meta (2026-08-08)
absorbs: [tool-reliability-fatigue (buildbox/k3s/offload subset)]
split_ids: [C084, C166]
---

# KB — cluster-offload

audience: AI coding agents first. Type-level patterns; per-incident entries land separately.

## Coverage

2 records (explicit split from 18 `tool-reliability-fatigue` rows). Low frequency, high severity when offload path wrong.

## Recurring failure patterns

1. **Hardcoded buildbox hostname** — spill order ignored; primary down, agent retries dead host.
2. **Contact unreachable host** — registry marks `reachable: false`; SSH hammering anyway.
3. **e2e-remote mis-route** — browser+dev-server split across laptop/box; artifacts outside repo tree lost on pull-back.
4. **Remote build without registry read** — `build-remote.json` / `buildbox-hosts.json` drift.
5. **Failover not attempted** — single-host assumption when cluster has debian1/debian2 spill.

## Fix vectors that worked

- Read `~/.claude/buildbox-hosts.json` every dispatch; follow spill order.
- `e2e-remote` with `--host 127.0.0.1` server bind; artifacts under repo root.
- Box reboot as acceptable fallback; laptop unaffected.

## Related skill

`od-cluster`