epic: audio input support for agent commands

[bug-ops]

Overview

Add support for sending commands to the agent via audio — both recorded audio files (Telegram voice messages, Discord/Slack file uploads) and real-time audio streaming where supported by provider APIs.

Motivation

Voice input is a natural interaction mode for AI agents. Telegram, Discord, and Slack all support voice/audio messages natively, but Zeph currently ignores all non-text content. Adding audio support unlocks hands-free interaction and expands accessibility.

Architecture

Audio input follows a two-stage pipeline:

Audio Source → Transcription (STT) → Text → Agent Loop (unchanged)

Transcription backends (pluggable):

1. OpenAI Whisper API — cloud, high accuracy, 25 MB file limit
2. Local Whisper (candle/whisper.cpp) — offline, no API costs, GPU optional
3. Native multimodal — Claude and GPT-4o accept audio natively (no separate STT step)

Key design decisions:

• Audio is transcribed to text before entering the agent loop — minimal changes to core
• ChannelMessage gains an attachments: Vec<Attachment> field
• A new SpeechToText trait abstracts transcription backends
• Channel adapters extract audio from platform-specific message types
• Transcription happens at the channel boundary, not in the agent

Sub-issues

• multimodal: extend ChannelMessage and MessagePart with attachment support #521 Extend ChannelMessage and MessagePart with multimodal attachment support
• audio: add SpeechToText trait and OpenAI Whisper backend #522 Add SpeechToText trait and OpenAI Whisper backend
• audio: add local Whisper backend via candle (feature-gated) #523 Add local Whisper backend via candle (feature-gated)
• audio: Discord voice message and audio attachment handling #524 Telegram: handle voice messages and audio files
• audio: Slack audio file upload handling #525 Discord: handle voice messages and audio attachments
• audio: Telegram voice message and audio file handling #526 Slack: handle audio file uploads
• audio: real-time streaming STT support #527 CLI: audio file path as input argument
• audio: configuration section and documentation #528 Streaming audio input support (real-time STT)
• audio: CLI audio file input support #529 Configuration and documentation

Non-goals (v1)

• Text-to-speech (TTS) output
• Voice call support (phone, WebRTC)
• Video input

References

• OpenAI Whisper API
• OpenAI Realtime API
• Claude audio support
• Telegram voice messages
• candle-whisper