Hallucinating Canary (github.com)

A new lightweight plugin called Hallucinating Canary has been introduced for Claude Code, designed to provide early warnings about the degradation of response quality during extended sessions. As AI models like Claude Code run longer, they tend to lose track of context, leading to issues like forgotten instructions and hallucinations—where the model generates incorrect or nonsensical outputs. The Hallucinating Canary addresses this by implanting known marker tokens, or "canaries," into the conversation context at the session's start. If these canaries are found missing after compaction, it signals a breakdown in context integrity, allowing users to anticipate potential errors before they occur. This development is significant for the AI/ML community, as it tackles the challenge of context loss, a leading cause of degradation in language models. The plugin operates without the need for external dependencies, ensuring ease of use while providing a deterministic approach to monitoring response quality. By alerting users to the conditions that can lead to hallucinations, it serves not as a correctness oracle, but as a preventative tool, making it easier for developers to maintain high-quality outputs in their applications. In essence, the Hallucinating Canary enhances user control and awareness, paving the way for more robust interactions with AI models.