qwibitai

NanoClaw became a production-ready conversational AI this week. We fixed two critical WhatsApp bugs that were preventing direct messages from working at all - the bot was looking for its own number instead of the user's number, breaking DM registration completely. With that resolved, we added emoji reactions so the agent can signal task progress (👀 received, 💭 thinking, ✅ done) and users can react to messages naturally. The bigger infrastructure move was implementing a credential proxy that keeps API keys completely outside containers, replacing our previous stdin approach with a transparent host-side authentication layer. We also shipped sender allowlists for group chat access control, letting admins choose between trigger mode (everyone's context stored but only allowed users activate) or drop mode (denied users ignored entirely). The /compact command now lets users manually compress conversation history when sessions get unwieldy, using Claude's native compaction. A container lifecycle bug was eating 30-minute idle windows after IPC tasks, which we fixed to close within 10 seconds. The product is now handling real conversations with proper access controls and reliable message delivery.