What is proactive-nudge?

This is a small background worker that sends proactive messages in chat applications. You connect it behind your existing chat app, AI companion, or agent gateway. The idea is simple: when a user has been silent for a configured amount of time, the worker injects a custom message into the normal chat flow, and the model replies as it normally would. The project exists because many proactive messaging systems just push a fixed message on a timer, like an alarm. That feels artificial: the user might have just finished talking, the message has no context, and it does not read like real conversation. This worker takes a different approach. It only creates a conversation opportunity by injecting a special user message marked with [nudge]. The actual reply still goes through the normal chat pipeline, so the model sees full context, memory, and tools, and the conversation history stays complete. Users configure everything themselves: the silence interval (fixed or random), the injected message text, and which conversation to target. The worker checks when the last user message was sent and waits only the remaining time if the user spoke recently. When the interval passes, it calls the normal chat endpoint with the nudge message. The system prompt tells the model that [nudge] messages are scheduled prompts, not freshly typed user messages, and the model should reply naturally without mentioning automation. Optionally, the worker can push a short reply summary to the user's phone. A notable design choice is that the nudge message is just a user-defined instruction. It can be a simple reminder, or it can tell the model to check specific tools or data before responding. The worker itself stays thin: it only triggers and injects. All business logic lives in the existing chat gateway and tool system. The README discusses the tradeoff between injecting data directly into the worker versus letting the model fetch it, noting that direct injection is cheaper but makes the worker dependent on business logic, while letting the model fetch is more flexible at the cost of an extra API call.