GOAT0.1 (Kid) · Source preview

ADR-0020: Engine presets: pick a known engine, or point at any custom URL ​

Status: Accepted · 2026-08-30

Context ​

ADR-0017 made the client engine-agnostic: it speaks the OpenAI /v1 dialect to whatever URL it is given. But Settings still exposed only a raw endpoint field and a single oMLX-specific button ("Manage Models in oMLX…"). A user running Ollama or LM Studio had to know the right port, know GOAT was OpenAI-compatible at all, and got no help managing models. The transport was general; the UX was not.

Decision ​

Settings → Engine offers a preset picker plus a Custom option:

  • A preset is pure data (EnginePreset in GoatInference): a name, the root URL its server listens on (the client still appends /v1/...), a one-line blurb, and how models are managed. It is not a second transport. Every preset speaks the one dialect.
  • Shipped presets: oMLX (recommended), vMLX, Ollama (:11434), LM Studio (:1234), llama.cpp server (:8080), and Custom… (type any URL).
  • Choosing a preset fills the URL and re-probes; the URL stays editable, so a preset is a convenience, never a cage. The chosen preset id is remembered (engine.preset).
  • Model management is engine-aware, not removed: open the engine's app when the preset names one and it's installed (oMLX/vMLX/LM Studio); show a copyable command for CLI engines (ollama pull <model>, llama-server …); show nothing for bare/custom endpoints. The composer capsule and the empty-state banner use the same generalized affordance ("Wake <engine>").
  • Discovery also probes Ollama's :11434 alongside the MLX ports.

oMLX remains the recommended engine everywhere (picker default, discovery seed, docs).

Consequences ​

Adding an engine is a row in EnginePreset.all: no new types, transports, or views. docs/ENGINES.md carries the preset table and stays the compatibility contract. Presets are conventional defaults; an unusual port is one edit away in the same field. Remote/hosted engines remain parked (ADR-0017). Presets are local-first only.

Alternatives considered ​

Raw URL only (rejected: leaves non-oMLX users to guess ports and compatibility), presets only with no free-form field (rejected: blocks unusual ports and self-hosted setups), removing model management entirely (rejected: the oMLX "manage models" jump is genuinely useful; generalizing it serves more engines than dropping it), a per-engine adapter class (rejected: still one dialect; a data preset is enough, mirroring ADR-0017's reasoning).