WeCom Voice Message Assistant

Problem

In private WeCom chats, voice messages often carry office requests such as checking a schedule, creating a todo, or sending a message. Text-only bots struggle with ambiguous, multi-step speech. This skill connects WeCom bot voice callbacks to a local agent that treats the transcribed text as structured intent, while keeping raw audio away from third-party processing.

How It Works

• Message intake: watches msgtype: voice, reads voice.content, then deduplicates, checks message freshness, and maintains session context.
• Intent parsing: uses intent_registry with YAML-declared intents, keyword matching, and confidence scoring; low-confidence cases ask clarifying questions instead of guessing.
• Task execution: supports schedule lookup, todo creation, weather lookup, and message sending; message sending requires confirmation, and weather queries only call wttr.in.
• Call management: v2.0 adds outbound calling, IVR handling, and multi-turn voice sessions through a state_machine covering IDLE → DIALING → LISTENING → ENDING, with automatic timeout after 30 seconds of silence.
• Compliant storage: recordings and transcripts stay in local ~/.wecom_voice/records/ and SQLite; outbound calls require admin authorization and user consent.

Boundaries

It is best suited for private single-chat voice messages in a quiet environment, with each clip under 60 seconds and one task per message. Group chat, long recordings, speaker identification, and multimodal messages are outside the current scope; avoid transmitting sensitive information by voice.