@planmo/ai

The model-provider abstraction: one interface (ModelProvider) with a deterministic mock and an OpenAI-compatible HTTP provider, so no PlanMo package depends on a vendor SDK.

Publishing

Published from CI via OIDC trusted publishing (.github/workflows/release.yml). There is no token anywhere in this repository, and no npm secret configured.

npm trusted publishers are configured per package, and a package must already exist on npm before one can be configured for it — so the first version of any new package has to be published manually, once:

1. Publish it a single time with a user-level token (~/.npmrc, never a token in this repository).
2. Add its trusted publisher on npmjs.com: owner doofx0071, repository planmo, workflow filename release.yml, no environment, and tick npm publish as well as npm stage publish — changeset publish uses plain npm publish.
3. From then on CI publishes it with every release.

This matters beyond the one package: adding an unpublished package to a release aborts the entire publish run with ENEEDAUTH, so every other package in that release stays unpublished too. Bootstrap a new package on its own.

Using it

import { createMockProvider, createHttpProvider } from '@planmo/ai';

The mock is deterministic — identical input yields byte-identical output — and always reports simulated: true, so a caller can never mistake simulated output for a real model's. The default test suite uses only the mock: no network, no credentials. The HTTP provider retries transport, rate-limit and 5xx failures with bounded backoff, and never retries an auth error.

Providers

PROVIDER_PRESETS is a declared table of gateways and vendors (Token Harbor, OpenAI, Anthropic, DeepSeek, Moonshot/Kimi, xAI/Grok, OpenRouter, and an overridable custom). Each row records only the credential env var name and the wire protocol — never a key and never a model list — so a gateway like Token Harbor can serve provider/model slugs from a single key. resolveProvider selects one from an injected env (or process.env); describeProviders(env) returns render-ready rows. createHttpProvider speaks either wire format:

import { resolveProvider, createHttpProvider } from '@planmo/ai';

const resolved = resolveProvider({ env: process.env, model: 'deepseek/deepseek-v4-flash' });
if (resolved.configured) {
  const provider = createHttpProvider({
    baseUrl: resolved.baseUrl,
    apiKey: process.env[resolved.preset.credentialEnv]!,
    model: resolved.model ?? '',
    wireFormat: resolved.wireFormat, // 'openai' | 'anthropic'
  });
}