Skip to main content
Nika’s capability system answers one question: given a provider + model, what does it accept? Which parameter carries the token limit? Is it a reasoning model? Vision input? JSON schema mode? A temperature? The answer is a ModelCapabilities value, computed at runtime from rules and a defaults block. The first rule that matches wins. Zero regex, zero allocation, zero runtime dependencies.
Canonical source: crates/nika-catalog/data/model-capabilities.toml at v0.123.0. Schema: nika/model-capabilities@1.0. The build checks every rule against llm-providers.toml (an unknown provider fails the build, so a typo cannot ship) and requires every any rule to come last within its provider scope.

Resolution algorithm

1

Scan [[rules]] in file order

A rule whose scope.providers list does not name the provider is skipped, and so is one whose scope.api_dialect does not match the provider’s dialect.
2

The first match wins

The first rule whose match clause matches the model name supplies its caps, and the scan stops. Order is load-bearing: a narrow rule must come before a broader one that would also match (o4 before o4-, gpt-4o before gpt-4).
3

Fill the rest from [defaults]

Every field the winning rule does not set comes from the defaults block:

The four match kinds

A provider-wide fallback. It must be the last rule of its scope.
A per-model override that deviates from its family.
An explicit list when the ids share no prefix.
The most common kind: one rule covers a model family. Prefix match only, no regex.

Capability fields

What a rule can set:

Supported parameter vocabulary (13)

The vocabulary is closed: adding a new parameter is a nika-catalog schema change.

Modality vocabulary (8)

text · image · audio · video · pdf · embedding · speech · image-gen
input_modalities must contain text (build invariant). A pure-image input model can’t be wired as an infer: target without text support.

Example resolution

Why not regex

L0 crates ship zero runtime dependencies beyond std + alloc. Regex would pull in regex-automata (~50k LOC) and force a runtime initialization cost. prefix_any covers 95% of actual model-family matching needs; the remaining 5% is handled by exact_any.
This is Q1 (no proc macros) applied at the data-layer. See L0 foundation decisions.

How to query capabilities

Two shipped subcommands project the embedded catalogs — nika catalog (providers · models · capabilities · env vars) and nika catalog --tools (the nika:* builtin schemas), each with a versioned --json machine form:

Drift protection

Build-time FK check: every scope.providers[i] must be a canonical provider id in llm-providers.toml. Typos fail the build.Build-time enum check: every supported_parameters entry must be in the 13-value vocabulary. New additions require a schema bump.Build-time ordering check: tags arrays in llm-providers.toml must be sorted + deduplicated. Same applies to caps.* arrays by convention.

See also

Providers catalog

The canonical providers these rules target.

Concepts · Providers

Model and access, the wires, structured output.

Schema

.nika envelope + workflow schema reference.

L0 decisions

Q1-Q13: why no regex, why manual impl, why prefix-only matching.