Skip to content
Kindgi

@kindgi/capabilities

View as Markdown

npm install @kindgi/capabilities · source

Capability declarations, a tenant-scoped model-provider registry, and a deterministic router for Kindgi. An agent declares what it needs from a model (features, context window, cost, region, latency, provider or model allow and deny lists) and what it prefers; the router picks a (provider, model) pair from the tenant's registered providers that meets every hard requirement and the tenant's policy, ranked by the preferences. The package also defines ModelProvider, the provider-neutral contract that model adapters implement.

Purpose

Section titled “Purpose”

Keep agent definitions independent of any one model vendor. Agents state requirements instead of naming a model; the binding to a concrete model happens at run time, and when nothing qualifies the router returns a structured explanation of which candidate failed which requirement. Providers are registered per tenant, so routing for one tenant never sees another tenant's providers.

Exports

Section titled “Exports”
  • Declaring requirements
    • Capability — { kind?, needs, prefer?, budget? }. kind defaults to DEFAULT_CAPABILITY_KIND ('llm-inference'); BUILT_IN_CAPABILITY_KINDS / BuiltInCapabilityKind list the well-known kinds, and CapabilityKind is an open string.
    • Requirement — hard constraints: { feature }, { contextWindow: { op, value } }, { costPerCall: { op, usd } } (compared against the model's prompt price per 1K tokens), { region }, { p95LatencyMs: { op, value } }, { providers: { allow?, deny? } }, { models: { allow?, deny? } }. Operators are ComparisonOp (>=, >, =, <=, <) and UpperBoundOp (<=, <) for latency.
    • Preference — { feature, weight }; feature is a Feature or a provider attribute such as 'lower-cost'. Weights add up per candidate and may be negative.
    • FEATURES / Feature — the closed feature set: structured-output, vision, audio-input, audio-output, tool-use, parallel-tool-use, thinking, long-context, code-execution, web-search, file-search, streaming, batch. structured-output means the model can follow a JSON schema natively (ModelCallInput.structuredOutput, where its adapter maps it); an agent's typed output doesn't use that yet, and on every provider is checked by parse and repair.
    • Budget — maxCostUsd, maxTokens, maxDurationMs.
    • defineCapability(spec) — validates a declaration against the bundled capability.schema.json (the same file as in @kindgi/specs; its $id is exported as CAPABILITY_SCHEMA_URI) and returns Result<Capability, InvalidCapabilityError>. The schema accepts needs, prefer and budget with every requirement except models, so a declaration that sets kind or uses a models requirement is rejected here even though the type and the router accept it.
  • Providers
    • ModelProvider — metadata plus invoke(input: ModelCallInput): Promise<ModelCallResult>.
    • ProviderMetadata — id, region ('unspecified' if not region-scoped), models, and optional attributes, description, capabilityKind. One provider is one connection that can expose several models.
    • ModelInfo — name, contextWindow, features, cost (promptUsdPer1kTokens, completionUsdPer1kTokens), and optional p95LatencyMs, maxOutputTokens, description.
  • Model call shape — ModelCallInput (model, messages, optional tools, structuredOutput, temperature, maxOutputTokens, abortSignal), ModelMessage (roles system, user, assistant, tool), ModelToolCall, ModelToolDefinition, StructuredOutputRequest, ModelCallResult (message, finishReason, usage, costUsd, durationMs, provider), UsageCounters. Adapters translate these to and from their vendor SDKs.
  • Registry
    • createProviderRegistry(seed?) — in-memory registry. Returns { registry, register }: register(tenantId, provider) validates the metadata and returns a Result (invalid-provider, duplicate-provider); registry is the ProviderRegistry view, whose own register throws on the same errors.
    • ProviderRegistry — register, get, list, has, all keyed by tenantId, plus optional hydrate(tenantId) and invalidate(tenantId) for implementations backed by persistent storage, and setupFailures(tenantId): the registered providers its last hydrate couldn't set up, each a ProviderSetupFailure (providerId, message). list leaves those out.
  • Routing
    • route(input: RouteInput) — returns Result<RoutingDecision, CapabilityError>. It expands each provider into one candidate per model, filters by kind, needs and the tenant policy, ranks by summed preference weight, moves candidates matching preferredProvider / preferredModel to the front, and breaks ties on (providerId, modelName). The same inputs always produce the same decision.
    • RouteInput — capability, providers (already scoped to one tenant, e.g. registry.list(tenantId)), and optional tenantPolicy, preferredProvider, preferredModel, and setupFailures (registry.setupFailures(tenantId)): when nothing routes, the error names each under the requirement provider set up, by provider id, and its message gives the first one's why.
    • RoutingDecision — the chosen provider and model, a human-readable reason, and alternates (ProviderModelPick[]) in rank order.
    • matchTuples(input) — the filtering step on its own: the passing candidates, unranked, or capability-unsatisfiable when none pass.
    • TenantPolicy — tenantId, provider and model allow / deny lists, regionAllow, maxCostPerCallUsd, maxTokensPerCall. An allow list (or regionAllow) that is present restricts even when empty: [] allows nothing.
  • Adapter factories — createAdapterFactoryRegistry(seed?, options?) returns an AdapterFactoryRegistry (register, get, has, list): a deployment-wide map from adapter id to the code that builds providers. An AdapterFactoryEntry holds adapterId, capabilityKind, a factory (AdapterFactory: (input: AdapterFactoryInput) => ModelProvider, where AdapterFactoryInput is metadata plus optional config, resolveApiKey, fetch and identities: the runtime's own cloud identities, AdapterIdentities { azure?: AzureTokenClient, aws?: AwsCredentialClient }, for a registration that signs in as the server), an optional prepare(params) that streams PrepareEvents (progress, ready, error) for downloads or warm-up, and an optional checkConfig(input): the adapter's static check of a provider registration (AdapterConfigCheckInput: metadata, config, hasSecretRef, and identities, which of the runtime's identities are there; no network, no secret read), returning **AdapterConfigProblem**s (path, a JSON pointer into the registration, and message). Each problem's message names the setting and what it takes, and factory throws it as adapterConfigError(adapterId, providerId, problem) words it; the runtime runs the check when a provider registers. A registry made with options.identities hands them to every factory and tells every check which are there, so a registration needing an identity the runtime lacks is refused when it registers. Registering the same id twice throws.
  • Errors — CapabilityError, a union of InvalidCapabilityError, InvalidProviderError, DuplicateProviderError, CapabilityUnsatisfiableError and BudgetExceededError. CapabilityUnsatisfiableError.reasons lists, for each requirement that rejected at least one candidate, the candidates that satisfied it and a structured RejectionReason for each that did not (missing-feature, context-window-mismatch, region-mismatch, tenant-denied, capability-kind-mismatch, provider-setup-failed, …).

Example

Section titled “Example”
import { createProviderRegistry, defineCapability, route } from '@kindgi/capabilities';
import type { ModelProvider } from '@kindgi/capabilities';
import type { TenantId } from '@kindgi/types';


// One connection to a model vendor, exposing two models.
const acmeLlm: ModelProvider = {
  metadata: {
    id: 'acme-llm',
    region: 'eu-west-1',
    attributes: ['lower-cost'],
    models: [
      {
        name: 'acme-large',
        contextWindow: 200_000,
        features: ['tool-use', 'structured-output', 'streaming'],
        cost: { promptUsdPer1kTokens: 0.003, completionUsdPer1kTokens: 0.015 },
        p95LatencyMs: 4_000,
      },
      {
        name: 'acme-small',
        contextWindow: 32_000,
        features: ['tool-use'],
        cost: { promptUsdPer1kTokens: 0.0002, completionUsdPer1kTokens: 0.0008 },
      },
    ],
  },
  invoke: (input) => callAcmeApi(input), // your vendor SDK call
};


const tenantId = process.env.TENANT_ID as TenantId;
const { registry, register } = createProviderRegistry();
const registered = register(tenantId, acmeLlm);
if (registered.kind === 'err') throw new Error(registered.error.message);


const capability = defineCapability({
  needs: [{ feature: 'tool-use' }, { contextWindow: { op: '>=', value: 100_000 } }],
  prefer: [{ feature: 'lower-cost', weight: 1 }],
});
if (capability.kind === 'err') throw new Error(capability.error.message);


const decision = route({
  capability: capability.value,
  providers: registry.list(tenantId),
  tenantPolicy: { tenantId, regionAllow: ['eu-west-1'] },
});
if (decision.kind === 'err') {
  if (decision.error.code === 'capability-unsatisfiable') {
    for (const r of decision.error.reasons) {
      console.error(r.requirement, r.rejectingProviders.map((p) => `${p.id}: ${p.reason.code}`));
    }
  }
  process.exit(1);
}


// acme-small fails the context-window requirement, so this is acme-llm / acme-large.
const { provider, model } = decision.value;
const reply = await provider.invoke({
  model: model.name,
  messages: [{ role: 'user', content: 'Summarize clause 4.2 in one sentence.' }],
});
console.log(reply.message.content, reply.costUsd);
Excerpt

Non-goals

Section titled “Non-goals”
  • No vendor adapters. Concrete ModelProvider implementations live in adapter packages such as those under packages/adapters.
  • No cost metering. Budget and BudgetExceededError describe limits; nothing in this package tracks spend or enforces them.
  • No retries or failover. route returns ranked alternates; retrying with the next candidate is up to the caller.
  • No persistence. Both registries are in-memory. Implementations backed by storage plug in through ProviderRegistry.hydrate and invalidate.
Section titled “Related”