OpenAI Provider

The OpenAI provider supports GPT models, reasoning models (o1/o3/o4), and Azure OpenAI deployments.

Configuration

openAiApiKey

string

required

Your OpenAI API key from OpenAI PlatformNot required if using Azure Identity authentication

openAiBaseUrl

string

Custom base URL for API requests

OpenAI: https://api.openai.com/v1 (default)
Azure: https://<resource>.openai.azure.com

openAiModelId

string

required

Model identifier. Examples:

gpt-4.5-turbo
gpt-4o
o1-preview
o3-mini
deepseek-reasoner

openAiModelInfo

OpenAiCompatibleModelInfo

Model metadata including:

maxTokens - Maximum output tokens
temperature - Sampling temperature (0-2)
isR1FormatRequired - Enable R1 message format

reasoningEffort

string

For reasoning models (o1/o3/o4). Options:

low - Faster, less thorough
medium - Balanced
high - Slower, more thorough

azureApiVersion

string

Azure OpenAI API version. Defaults to 2024-10-01-preview

azureIdentity

boolean

default:"false"

Use Azure Active Directory authentication instead of API key

openAiHeaders

Record<string, string>

Additional HTTP headers for requests

Basic Setup

OpenAI Standard

import { OpenAiHandler } from "./providers/openai"

const handler = new OpenAiHandler({
  openAiApiKey: process.env.OPENAI_API_KEY,
  openAiModelId: "gpt-4.5-turbo",
  openAiModelInfo: {
    maxTokens: 16384,
    temperature: 0
  }
})

Azure OpenAI

const handler = new OpenAiHandler({
  openAiApiKey: process.env.AZURE_OPENAI_API_KEY,
  openAiBaseUrl: "https://my-resource.openai.azure.com",
  openAiModelId: "gpt-4o",
  azureApiVersion: "2024-10-01-preview"
})

DeepSeek Reasoner on Azure

const handler = new OpenAiHandler({
  openAiApiKey: process.env.AZURE_OPENAI_API_KEY,
  openAiBaseUrl: "https://my-resource.openai.azure.com",
  openAiModelId: "deepseek-reasoner", // Note: contains "deepseek"
  // Handler detects "deepseek" and skips Azure client,
  // uses standard OpenAI client instead
})

Reasoning Models

OpenAI’s o1, o3, and o4 models support reasoning effort control:

const handler = new OpenAiHandler({
  openAiApiKey: process.env.OPENAI_API_KEY,
  openAiModelId: "o3-mini",
  reasoningEffort: "high" // low | medium | high
})

for await (const chunk of handler.createMessage(
  systemPrompt,
  messages,
  tools
)) {
  if (chunk.type === "reasoning") {
    // Reasoning models output thinking traces
    console.log("Reasoning:", chunk.reasoning)
  }
  
  if (chunk.type === "text") {
    console.log("Response:", chunk.text)
  }
}

Reasoning Model Behavior

From openai.ts:129-134:

Reasoning models use developer roleWhen using o1/o3/o4 models:

System prompt sent as developer role instead of system
Temperature is automatically set to undefined (not supported)
Messages use special reasoning format

// Standard models
openAiMessages = [
  { role: "system", content: systemPrompt },
  ...convertToOpenAiMessages(messages)
]

// Reasoning models (o1/o3/o4)
openAiMessages = [
  { role: "developer", content: systemPrompt },
  ...convertToOpenAiMessages(messages)
]

R1 Format Models

Some models (like DeepSeek Reasoner) require R1 message format:

const handler = new OpenAiHandler({
  openAiApiKey: process.env.OPENAI_API_KEY,
  openAiModelId: "deepseek-reasoner",
  openAiModelInfo: {
    isR1FormatRequired: true // Enable R1 format
  }
})

// Messages automatically converted using convertToR1Format()

Temperature Control

const handler = new OpenAiHandler({
  openAiApiKey: process.env.OPENAI_API_KEY,
  openAiModelId: "gpt-4o",
  openAiModelInfo: {
    temperature: 0.7 // 0 = deterministic, 2 = very creative
  }
})

Temperature defaults to 0 for deterministic output unless specified.Reasoning models (o1/o3/o4) ignore temperature - it’s automatically set to undefined.

Token Limits

const handler = new OpenAiHandler({
  openAiApiKey: process.env.OPENAI_API_KEY,
  openAiModelId: "gpt-4o",
  openAiModelInfo: {
    maxTokens: 8192 // Maximum output tokens
  }
})

If not specified, maxTokens is undefined (uses model’s default).

Tool Calling

import type { ChatCompletionTool } from "openai/resources/chat/completions"

const tools: ChatCompletionTool[] = [
  {
    type: "function",
    function: {
      name: "read_file",
      description: "Read file contents",
      parameters: {
        type: "object",
        properties: {
          path: {
            type: "string",
            description: "File path"
          }
        },
        required: ["path"]
      }
    }
  }
]

const handler = new OpenAiHandler({
  openAiApiKey: process.env.OPENAI_API_KEY,
  openAiModelId: "gpt-4o"
})

for await (const chunk of handler.createMessage(
  systemPrompt,
  messages,
  tools
)) {
  if (chunk.type === "tool_calls") {
    const toolCall = chunk.tool_call
    console.log("Function:", toolCall.function.name)
    console.log("Arguments:", JSON.parse(toolCall.function.arguments))
  }
}

Streaming Response

The handler processes streaming chunks:

Text Delta

{
  type: "text",
  text: string // Incremental text
}

Reasoning Content

{
  type: "reasoning",
  reasoning: string // For o1/o3/o4 models
}

Tool Calls

{
  type: "tool_calls",
  tool_call: {
    function: {
      name: string,
      arguments: string // Incremental JSON
    }
  }
}

Usage Statistics

{
  type: "usage",
  inputTokens: number,
  outputTokens: number,
  cacheReadTokens: number,
  cacheWriteTokens: number
}

Prompt Caching

OpenAI supports prompt caching (reported in usage stats):

for await (const chunk of handler.createMessage(...)) {
  if (chunk.type === "usage") {
    console.log("Cached tokens:", chunk.cacheReadTokens)
    console.log("Cache writes:", chunk.cacheWriteTokens)
    console.log("New input tokens:", chunk.inputTokens)
    console.log("Output tokens:", chunk.outputTokens)
  }
}

Custom Headers

const handler = new OpenAiHandler({
  openAiApiKey: process.env.OPENAI_API_KEY,
  openAiModelId: "gpt-4o",
  openAiHeaders: {
    "OpenAI-Organization": "org-123456",
    "OpenAI-Project": "proj-abc123"
  }
})

Azure Regional Endpoints

The handler auto-detects Azure regions for authentication:

// Azure US Government
const handler = new OpenAiHandler({
  openAiBaseUrl: "https://my-resource.openai.azure.us",
  azureIdentity: true
  // Uses scope: "https://cognitiveservices.azure.us/.default"
})

// Azure Commercial (default)
const handler2 = new OpenAiHandler({
  openAiBaseUrl: "https://my-resource.openai.azure.com",
  azureIdentity: true
  // Uses scope: "https://cognitiveservices.azure.com/.default"
})

Error Handling

const handler = new OpenAiHandler({
  openAiApiKey: process.env.OPENAI_API_KEY,
  openAiModelId: "gpt-4o",
  onRetryAttempt: (error, attempt, maxRetries) => {
    console.log(`Retry ${attempt}/${maxRetries}:`, error.message)
  }
})

try {
  for await (const chunk of handler.createMessage(...)) {
    // Process chunks
  }
} catch (error) {
  console.error("OpenAI API error:", error.message)
}

Model Information

const { id, info } = handler.getModel()

console.log("Model ID:", id)
console.log("Max tokens:", info.maxTokens)
console.log("Temperature:", info.temperature)
console.log("Context window:", info.contextWindow)

Implementation Reference

Source: ~/workspace/source/src/core/api/providers/openai.ts Key features:

Auto-detects Azure vs standard OpenAI
Handles reasoning models (o1/o3/o4) with developer role
Supports R1 format for DeepSeek Reasoner
Azure Identity authentication
Regional Azure endpoints
Streaming with tool calls

Common Patterns

Switch Between Models

const models = {
  fast: "gpt-4o-mini",
  balanced: "gpt-4o",
  reasoning: "o3-mini"
}

const handler = new OpenAiHandler({
  openAiApiKey: process.env.OPENAI_API_KEY,
  openAiModelId: models.reasoning,
  reasoningEffort: "high"
})

Cost Tracking

let totalInputTokens = 0
let totalOutputTokens = 0
let totalCachedTokens = 0

for await (const chunk of handler.createMessage(...)) {
  if (chunk.type === "usage") {
    totalInputTokens += chunk.inputTokens
    totalOutputTokens += chunk.outputTokens
    totalCachedTokens += chunk.cacheReadTokens || 0
  }
}

console.log("Input tokens:", totalInputTokens)
console.log("Output tokens:", totalOutputTokens)
console.log("Cached tokens:", totalCachedTokens)

Next Steps

Anthropic Provider

Configure Claude models

OpenRouter Provider

Access multiple providers

Provider Overview

View all providers

Custom Provider

Build your own

Extension API

CLI Reference

LLM Providers

Configuration

Basic Setup

OpenAI Standard

Azure OpenAI

DeepSeek Reasoner on Azure

Reasoning Models

Reasoning Model Behavior

R1 Format Models

Temperature Control

Token Limits

Tool Calling

Streaming Response

Text Delta

Reasoning Content

Tool Calls

Usage Statistics

Prompt Caching

Custom Headers

Azure Regional Endpoints

Error Handling

Model Information

Implementation Reference

Common Patterns

Switch Between Models

Cost Tracking

Next Steps

Anthropic Provider

OpenRouter Provider

Provider Overview

Custom Provider

Build docs developers (and LLMs) love

Extension API

CLI Reference

LLM Providers

​Configuration

​Basic Setup

​OpenAI Standard

​Azure OpenAI

​DeepSeek Reasoner on Azure

​Reasoning Models

​Reasoning Model Behavior

​R1 Format Models

​Temperature Control

​Token Limits

​Tool Calling

​Streaming Response

​Text Delta

​Reasoning Content

​Tool Calls

​Usage Statistics

​Prompt Caching

​Custom Headers

​Azure Regional Endpoints

​Error Handling

​Model Information

​Implementation Reference

​Common Patterns

​Switch Between Models

​Cost Tracking

​Next Steps

Anthropic Provider

OpenRouter Provider

Provider Overview

Custom Provider

Build docs developers (and LLMs) love

Configuration

Basic Setup

OpenAI Standard

Azure OpenAI

DeepSeek Reasoner on Azure

Reasoning Models

Reasoning Model Behavior

R1 Format Models

Temperature Control

Token Limits

Tool Calling

Streaming Response

Text Delta

Reasoning Content

Tool Calls

Usage Statistics

Prompt Caching

Custom Headers

Azure Regional Endpoints

Error Handling

Model Information

Implementation Reference

Common Patterns

Switch Between Models

Cost Tracking

Next Steps