mirror of
https://github.com/chenhg5/cc-connect.git
synced 2026-10-25 17:36:43 +08:00
build: release v1.1.0-beta.3 — voice messages, image support, CLI send
This commit is contained in:
1 parent
8098e85c1d
commit
35e2d7ec91
2 files changed
+42
-1
No files matched your search
@@ -1,5 +1,46 @@
|
||||
# Changelog
|
||||
|
||||
## v1.1.0-beta.3 (2026-03-01)
|
||||
|
||||
### New Features
|
||||
|
||||
- **Voice Message Support (Speech-to-Text)**: Send voice messages on any supported platform — cc-connect transcribes them to text via configurable STT provider, then forwards to the agent.
|
||||
- Supported platforms: Feishu, WeChat Work, Telegram, LINE, Discord, Slack
|
||||
- STT providers: OpenAI Whisper, Groq Whisper, SiliconFlow SenseVoice, or any OpenAI-compatible endpoint
|
||||
- Audio format conversion via `ffmpeg` (AMR/OGG → MP3)
|
||||
- Configurable language hint for better recognition accuracy
|
||||
- **Image Message Support**: Send images directly from messaging platforms to AI agents.
|
||||
- Supported platforms: Feishu, WeChat Work, Telegram, LINE, Discord, Slack
|
||||
- Images are saved locally and sent to Claude Code as multimodal content (base64 + file path fallback)
|
||||
- **CLI Send Command**: `cc-connect send` allows external processes to send messages to active sessions via Unix socket API, enabling scheduled tasks (e.g., "daily summary at 6 AM")
|
||||
|
||||
### Bug Fixes
|
||||
|
||||
- **WeChat Work duplicate messages**: Added `MsgId`-based deduplication to prevent WeChat Work callback retries from causing duplicate processing
|
||||
- **WeChat Work media download timeout**: Moved image/voice download to async goroutines to prevent blocking the callback handler and triggering retries
|
||||
|
||||
### Improvements
|
||||
|
||||
- **Bilingual config template**: `config.example.toml` now includes both English and Chinese comments for all settings
|
||||
- **Updated README**: Added Voice Messages and Image Messages to support matrix and documentation
|
||||
|
||||
### Configuration Changes
|
||||
|
||||
```toml
|
||||
# New: Speech-to-Text configuration
|
||||
[speech]
|
||||
enabled = true
|
||||
provider = "openai" # "openai" or "groq"
|
||||
language = "zh" # empty = auto-detect
|
||||
|
||||
[speech.openai]
|
||||
api_key = "sk-xxx"
|
||||
base_url = "" # custom endpoint (SiliconFlow, etc.)
|
||||
model = "whisper-1"
|
||||
```
|
||||
|
||||
---
|
||||
|
||||
## v1.1.0-beta.2 (2026-03-01)
|
||||
|
||||
### Bug Fixes
|
||||
|
||||
Reference in new issue
Block a user