I Shipped a Feature in May. Nobody Could Reach It Until Yesterday.

A feature was shipped in May, but it remained unaccessible to visitors until recently. The issue stemmed from the chat widget being available on only one out of thirty-nine pages, which was problematic. However, the real issue was that the widget failed to send necessary session_id and scope fields. Despite the server having both fields documented, tested, and deployed, the browser was never wired to send them. Consequently, the feature was never invoked.

The failure involved two components: correct individual components with an untested integration. The issue did not throw errors or log notifications, making it difficult to detect. The health endpoint was green, and a metric could indicate a problem without being visibly broken. In this case, a sessions_active counter read zero, indicating no active sessions, but it appeared plausible given the low traffic of the site.

To verify the fix, the author sent questions from different pages and observed if the answer retained context. Sending a question from the datasets page with a session id yielded the correct response, while sending the same question from the portfolio page with an empty client history resulted in the same answer, confirming the fix's success.

However, while fixing the issue, the author accidentally expanded the widget's scope to include internal research vocabulary. This inadvertently connected a buyer's licensing question to the research corpus, leading to a revision. The author learned the importance of testing a sales surface with real visitor interactions rather than internal vocabulary and the significance of monitoring specific health metrics that could indicate a feature's failure.