## OpenWebUI Integration

- `frank-dragoon` (frank-large, 192.168.0.143:7770) must be available in OpenWebUI alongside the other fleet endpoints.

## Model Naming

- Official model ids are StarCraft unit names: `frank-archon` (frank-primary), `frank-dragoon` (frank-large), `frank-zealot` (frank-secondary).
- Served names on the wire match these ids after 2026-10-09; verify renames on `/v1/models`, not chat responses.
- `models.yml` provider/model blocks, `modelRoles` entries, and homepage cards must update together when renaming endpoints.
- Subagent overrides use bare model ids; uniqueness across providers is required for resolution.

## Runtime-to-Model Mapping

- `frank-primary` → radiance runtime → 192.168.0.235:7770 (`FRANK_PRIMARY_SERVED_NAME`)
- `frank-large` → halogen-flash-server runtime → 192.168.0.143:7770 (`HALOGEN_MODEL_ID`)
- `frank-secondary` → llama.cpp runtime → 192.168.0.143:7771 (`ALIAS`)
- All three endpoints accept images (`input: [text, image]`) as of 2026-10-06.
- `frank-secondary` `contextWindow` is 200704 (per-slot, not pool/native).
- Reasoning channel differs: llama.cpp/SGLang emit `reasoning_content`, vLLM emits `reasoning`.

## Endpoint Lifecycle

- Retired endpoints (`frank-npu` and prior runtimes) live in the runbook index; the full swap history is in `serverAppConfigs-docs/LLM/index.md`.
- Before retiring an endpoint: update provider blocks, model ids, context windows, max tokens, and input specs in `models.yml`, then fix every `modelRoles` entry referencing the old name.

## Fleet Topology

- **Dankserver** (192.168.0.235): media/library stack plus `frank-primary` LLM service.
- **Dankclaw** (192.168.0.143): AI/search/dev/observability stack plus `frank-large` and `frank-secondary` LLM services; Caddy reverse-proxies both hosts.
- **FrankBook-Pro** (192.168.0.112): macOS mini running the *arr stack as quadlets; egress rides Gluetun WireGuard netns; deployed via the `frankbook-pro` handler.

- **Retired** endpoints (`frank-npu` and prior runtimes) are in the runbook index; the full swap history is in `serverAppConfigs-docs/LLM/index.md`.
- Before retiring an endpoint: update provider blocks, model ids, context windows, max tokens, and input specs in `models.yml`, then fix every `modelRoles` entry referencing the old name.