Addressing a slot as a model

[system]

hal0’s dispatcher resolves the model field of an OpenAI-compatible request primarily
by model id, not by slot name — see the Model → slot resolution
section of OpenAI-compatible API
for the full 5-step resolution order. Direct slot-name addressing exists, but only as
the last-resort fallback in that chain (SLOT_ALIASES direct lookup) — described in
source as “operator muscle memory,” for requesting a slot literally by name.

Chat-slot aliases

For /v1/chat/completions specifically, the route layer recognizes a small set of
aliases in the model field and rewrites them to the aliased slot’s configured model
id before dispatch even reaches the general resolver:

Alias	Resolves to
agent	The configured default chat slot.
utility	The configured utility slot.
any enabled LLM slot name	That slot’s configured model.

hal0/*-prefixed virtual model names are resolved separately, against live slot state,
and set a “lane pin” recording which specific slot the caller meant — relevant when
more than one slot serves the same underlying model id.

Omitting model entirely

If a request omits model, the request path decides the default, not a global
fallback model:

Path	Default target
/embeddings	embed capability slot
/rerank, /rerankings	rerank capability slot
/audio/speech	tts capability slot
/images/*	img capability slot
everything else	agent

Note

The default for chat-shaped requests changed from chat to agent (ADR-0023) —
if you’re carrying forward assumptions from pre-1.0 docs or muscle memory, update any
reference to a chat default.

Capability-based routing fallback

When neither an alias nor an exact model-id match resolves, the dispatcher falls back
to capability inference from the request path and model-name hints, in this order:

1. Path fragment → capability slot type (/embeddings → embed, /rerankings//rerank
   → rerank, /audio/speech → tts, /images/generations//images/edits/
   /images/variations → img — note the last two paths aren’t actually implemented
   endpoints, but their strings are still recognized here as routing hints).
2. Model-name substring hints (embed, rerank).
3. Direct SLOT_ALIASES lookup — the slot-name-as-model fallback.

See Providers, profiles & devices for
how a slot’s device/provider fields relate to which capability it can serve, and
Slot lifecycle for the states a slot must be in
(READY, SERVING, or IDLE) to be dispatchable at all.