feat(transcribe): add OpenRouter as an STT/transcription provider

[HeMuling]

Summary

Would you be open to adding OpenRouter as a transcription provider for Agent Reach?

Today the transcription path appears to support only Groq and OpenAI:

• PR feat(transcribe): Whisper transcription module with Groq→OpenAI fallback #277 added the current Groq -> OpenAI fallback:
  feat(transcribe): Whisper transcription module with Groq→OpenAI fallback #277
• agent_reach/transcribe.py currently defines only groq and openai providers
• agent_reach/config.py currently exposes groq_api_key and openai_api_key, but no openrouter_api_key

At the same time, OpenRouter now provides a dedicated STT endpoint:

• STT guide: https://openrouter.ai/docs/guides/overview/multimodal/stt
• API reference: https://openrouter.ai/docs/api/api-reference/transcriptions/create-audio-transcriptions

I searched the existing issues/PRs and did not find an existing OpenRouter feature request.

Why this would help

This would be useful for users who:

• already route most model usage through OpenRouter
• want a single key / billing surface for transcription too
• want access to OpenRouter-hosted STT models without adding a new provider integration each time
• want to use OpenRouter routing / BYOK capabilities for speech-to-text

Proposed scope

A minimal version would be enough:

• add openrouter as a transcription provider
• add openrouter_api_key config support
• allow provider="openrouter" in the transcription path / CLI
• call OpenRouter's /api/v1/audio/transcriptions endpoint with its required JSON payload
• use a reasonable default model slug (for example openai/whisper-1), with optional model configurability later if desired

Implementation note

This looks like a small but real provider-specific change, not just a new endpoint string.

Current code posts multipart form data like:

• files={"file": ...}
• data={"model": ..., "response_format": "text"}

OpenRouter's STT API expects JSON with base64-encoded audio under input_audio, for example:

• input_audio.data
• input_audio.format
• model

So the clean path seems to be adding a provider-specific request branch for OpenRouter while keeping the current Groq/OpenAI flow unchanged.

Offer

If this direction makes sense, I’m happy to implement it and open the PR.