Repair a customer's streaming assistant

A customer's support assistant streams replies from a model, and their team filed three complaints this week as a single ticket called "the chat is broken". Accented characters arrive as black diamonds. Long replies lose most of their text and only the end shows up. And sending a second message while the first is still writing tangles the two replies together.

They are three separate defects with three separate mechanisms, which is the first thing to establish out loud. A rewrite typically fixes two of them and quietly reintroduces the third, so the grader checks each symptom on its own.

There is a fourth defect nobody reported, because it is invisible until the day it matters: what the reader sees when the server returns 500, or when the connection dies halfway through a reply. On most implementations the answer is a blank bubble and a stack trace in a console nobody has open.

The pack ships a hostile stream server that splits characters across byte boundaries, stalls, returns errors and drops connections mid-stream, plus a grader that stands in for React so what is scored is your state model rather than your JSX. It runs on Node's standard library with nothing to install.

The tests are the smaller half. The half that decides the round is being able to answer two questions about your own fix: which message does a chunk belong to, and which run is allowed to write. A functional state update answers neither, and cancelling a request does not stop work that was already scheduled.

What you can run

The Premium exercise download contains the broken client, browser UI, local HTTP server and both deterministic and real HTTP checks. The separate reference download includes working code and a step-by-step written walkthrough of decoding, message ownership, cancellation and failure handling. The HTTP checks can also save per-request timing observations.

This local JavaScript project uses synthetic streamed text. It does not call an LLM or deploy to a cloud account, and it has no narrated coding video. The walkthrough explains the additional requirements for hosted inference and deployment behind a proxy. This brief is the public preview; both downloads require Premium access.