@totemheart/provider-openai

A real OpenAI-compatible chat-completions LanguageProvider for Totemheart — same resilience contract as the core's built-in OllamaProvider: it throws on any failure (missing key, unreachable host, non-2xx response, malformed JSON), and Totemheart's own #analyze() catches that and falls back to HeuristicProvider transparently. A misconfigured or down provider never breaks the pipeline.

Works with the real OpenAI API and with any real OpenAI-compatible endpoint (baseURL is a genuine parameter, not hardcoded), since many local/hosted inference servers implement the same /chat/completions shape.

Install

npm install totemheart @totemheart/provider-openai

Usage

import { Totemheart, Personality } from 'totemheart'
import { OpenAIProvider }          from '@totemheart/provider-openai'

const provider = new OpenAIProvider( {
	apiKey : process.env.OPENAI_API_KEY, // or pass explicitly
	model  : 'gpt-4o-mini',              // default
} )

const ai = new Totemheart( { personality: new Personality(), provider } )

const result = await ai.processInput( 'hola, ¿qué tal tu día?', { userId: 'u' } )
console.log( result.text )

API

new OpenAIProvider({ apiKey, baseURL, model, timeoutMs })

| Option | Default | Notes | | --- | --- | --- | | apiKey | process.env.OPENAI_API_KEY | Required for any real call to succeed | | baseURL | https://api.openai.com/v1 | Point at any real OpenAI-compatible endpoint | | model | gpt-4o-mini | Any real chat-completions model your account/endpoint supports | | timeoutMs | 8000 | Aborts and throws past this real wall-clock limit |

analyze(task, payload)

Duck-types Totemheart's real provider contract directly — this class doesn't extend LanguageProvider from the core package (Totemheart's own orchestrator calls this.provider.analyze(task, payload) with no instanceof check), so this package has no runtime dependency on totemheart, only a peerDependency for discoverability. Supports the same 5 tasks Totemheart's pipeline calls: sentiment, appraisal, beliefConflict, mentalState, selfCritique.

Testing without a real API key

OpenAIProvider throws on a missing key or unreachable baseURL — the same pattern the core repo uses to test OllamaProvider without a live server:

const provider = new OpenAIProvider( { apiKey: 'sk-fake', baseURL: 'http://127.0.0.1:1', timeoutMs: 500 } )
await provider.analyze( 'sentiment', { text: 'hola' } ) // rejects — real, honest failure

License

MIT