Implement OpenAI provider with multi-modal output (images, audio)

[djthorpe]

Summary

Implement an OpenAI provider (pkg/provider/openai) supporting chat completions, embeddings, image generation, and audio output. This will likely require extending the response content model beyond text to support multi-modal outputs.

Requirements

Core Provider

• Implement llm.Client interface for OpenAI API (chat completions, model listing)
• API key via OPENAI_API_KEY environment variable
• Support streaming and non-streaming chat completions
• Support tool/function calling
• Support thinking/reasoning (o1, o3 models)

Image Output

• Support DALL-E and GPT-image models for image generation
• Responses may contain image data (base64 or URLs) alongside text
• Extend the content block model to represent image outputs (not just text)
• Images should be renderable in Telegram (send as photo) and CLI (save to file or display URL)

Audio Output

• Support audio output from GPT-4o-audio and similar models
• Responses may contain audio data (base64 WAV/MP3)
• Extend the content block model to represent audio outputs
• Audio should be sendable in Telegram (as voice/audio message) and CLI (save to file)

Content Model Changes

• Current schema.Content may need to support typed content blocks: text, image, audio, etc.
• Each block should carry MIME type and either inline data or a URL
• Downstream consumers (Telegram bot, CLI, API responses) need to handle multi-modal content blocks
• Consider how this interacts with session storage (storing large binary blobs vs references)

Models

• GPT-4o, GPT-4o-mini, GPT-4.1, o1, o3, o4-mini (chat)
• DALL-E 3, GPT-image (image generation)
• GPT-4o-audio (audio output)
• Embedding models (text-embedding-3-small, text-embedding-3-large)

Motivation

OpenAI is a major LLM provider and its multi-modal output capabilities (images, audio) will drive the content model to support rich responses across all providers, improving the overall architecture.

[djthorpe]

No audio/image generation interfaces

Old OpenAI code (etc/_old/openai/audio.go, image.go) shows this was considered
Add optional AudioProcessor and ImageGenerator interfaces to the root package