API

Named exports only. There is no default export.

You will usually call four functions: compile (schema), check (every cloud provider), checkRequest (model + endpoint + tools + reasoning), and checkDeployment (local or self-hosted runtime).

import {
  compile,
  check,
  checkRequest,
  checkDeployment,
  analyze,
  fingerprint,
  listTargets,
  resolveTarget,
  listRequestProfiles,
  resolveRequestProfile,
  listRuntimeProfiles,
  resolveRuntimeProfile,
  listModelProfiles,
  resolveModelProfile,
} from "llm-abi";

compile(schema, target | options)

Compiles a JSON Schema, Standard JSON Schema object, or a closed TypeScript type subset into a provider-safe schema.

const result = compile(schema, "anthropic");
const same = compile(schema, { target: "anthropic/messages/structured" });

compile() is pure: no network, filesystem, clocks, or randomness.

strict: true throws SchemaCompatibilityError when compatibility is unsupported.

constraintFallback: "description" | "strip" controls whether runtime-only constraints are appended to description.

optimize: true drops titles that duplicate the property name and descriptions that duplicate the title. Each omission emits redundant-annotation-removed. Default is false, so previously emitted JSON stays the same when you leave it off.

typeName selects which type or interface to compile from TypeScript source. Default: the last exported declaration, otherwise the last declaration.

TypeScript input is a closed subset: type / interface, primitives, literals, arrays, tuples, unions, intersections, optional properties, nested objects, Array<T>, ReadonlyArray<T>, and Record<string, T>. String, number, and boolean literal unions emit enum. There is no typescript compiler dependency, no import resolution, and no generic type declarations. Unsupported syntax throws instead of being dropped silently.

check(schema, options?)

Runs compile for every built-in provider target (or options.targets). Runtime schema targets are omitted unless listed explicitly. Each row includes that target's emitted size. options.optimize is forwarded to each compile.

analyze(schema)

Returns fingerprint, node/depth/property stats, unused root $defs count, input schema byte/token size, and parse notes. Does not lower to a provider.

fingerprint(schema)

Canonical SHA-256 of the input schema. Property order does not matter.

checkRequest(request)

Checks whether a provider request combination is representable. This is not compile(). It does not call a model, rewrite a payload, or pick an SDK.

const result = checkRequest({
  provider: "openai",
  model: "gpt-5.6-terra",
  endpoint: "chat-completions",
  tools: true,
});

provider is openai, azure-openai, or azure for the shipped OpenAI-compatible family. endpoint accepts chat-completions, responses, and short aliases (chat, /v1/chat/completions). tools: true or tools: "function" means function tools are present.

checkRequest() is pure: no network, filesystem, clocks, or randomness. Same input returns the same result.

If no request profile matches the provider and model, coverage is unknown, compatibility is lossless, profile is undefined, and no default is invented. Treat unknown as unchecked, not as safe to send.

The shipped GPT-5.6 rule: Chat Completions plus function tools is compatible only when effective reasoning is none. GPT-5.6 defaults omitted reasoning_effort to medium, so omitting it is enough to fail. Responses accepts function tools with reasoning.

listTargets() / resolveTarget(id)

resolveTarget("claude") returns the same ResolvedTarget as listTargets() for the Anthropic structured profile. resolveTarget("deepseek") returns deepseek/chat/strict-tools. resolveTarget("grok") returns xai/grok/structured. resolveTarget("qwen") returns alibaba/qwen/tools (Alibaba Model Studio tools/structured output, not local Qwen / LM Studio / Ollama / llama.cpp). resolveTarget("mistral") returns mistral/chat/structured. resolveTarget("openrouter") returns openrouter/structured. resolveTarget("mcp") returns mcp/2026-06/tools. listTargets() defaults to scope: "provider". Pass { scope: "runtime" } or { scope: "all" } for local schema engines such as llamacpp/server/structured, vllm/openai/structured, and sglang/openai/structured. resolveTarget("vllm") and resolveTarget("sglang") stay runtime-scoped. Unknown ids throw LlmAbiError.

Each resolved target includes:

{
  maturity: "supported" | "partial" | "experimental",
  evidence: {
    kind: "documented" | "sdk-observed" | "empirical",
    source: "https://...",
    lastVerified: "2026-08-18",
    live: "nightly" | "not-configured",
  },
}

Maturity, documentary evidence, and live coverage are independent signals. A documentation verification date does not claim a successful live API request.

listRequestProfiles() / resolveRequestProfile(id)

resolveRequestProfile("openai/gpt-5.6") returns the GPT-5.6 request family. Unknown ids throw LlmAbiError.

checkDeployment(input)

Checks whether a schema and request contract can be represented on a runtime × engine × model deployment. This is not compile() and not a generator.

const result = checkDeployment({
  schema,
  deployment: {
    runtime: { kind: "lmstudio", apiSurface: "openai", engine: { kind: "llamacpp" } },
    model: { id: "Qwen3.8-27B-Q4_K_M", format: "gguf" },
  },
  request: { endpoint: "chat-completions", structuredOutput: true, tools: true },
});

checkDeployment() is pure. If no runtime profile matches, coverage is unknown and compatibility is lossless. Unknown features do not become unsupported. Probe results never live here.

Ollama Responses with stateful: true is unsupported (previous_response_id is documented unsupported). LM Studio without format does not guess GGUF vs MLX.

llm-abi/local

import {
  discoverLocalDeployments,
  probeDeployment,
  matrixLocalDeployments,
  createDeploymentLock,
  diffDeploymentLocks,
} from "llm-abi/local";

Node only. Default discovery is 127.0.0.1:1234, :11434, and :8080. It does not scan a LAN or follow redirects. local doctor is GET metadata and lists loaded models only. local probe is explicit generation with synthetic fixtures. matrixLocalDeployments() compares loaded models. createDeploymentLock() splits contract, deployment, and evaluation fingerprints and redacts absolute paths. An explicit model may cause the runtime to load that model. A probe pass does not upgrade static compatibility.