Lectures

In this lecture we break down the MCP event model on top of JSON-RPC 2.0: how notifications differ from request/response, how to design methods like notifications/job/*, which fields must live in params (for example, jobId, eventId, timestamp), how to agree on a taxonomy ( job.*, resource.*, system events), and why handlers must be idempotent (the “at-least-once” delivery model, de-duplication by eventId, validation of state machine transitions).

We break down where “streams” appear in the ChatGPT App architecture and why “plain” HTTP quickly stops being enough for long-running operations. We compare SSE (event subscription via EventSource) and HTTP-stream (streaming the response to a single request via fetch + ReadableStream), look at minimal implementations in Next.js 16, and discuss reconnect, timeouts, heartbeat, and common production pitfalls.

In this lecture we design a “live” UX for long-running operations in ChatGPT Apps: we introduce a clear status model ( idle, pending, in_progress, partial_ready, completed…), wire up an event stream ( job.progress, job.partial_result, job.canceled) and map it onto React state via useState/ useEffect. We also cover honest progress (determinate vs. indeterminate), careful rendering of partial results without UI “jitter,” and proper job cancelation on both the frontend and backend.

In this lecture, we’ll break down how to design long-running operations in a ChatGPT App via an async-job pipeline: quickly launching a tool and returning a jobId, enqueueing a task, executing it in the background with a worker, delivering progress via MCP events ( job.progress, job.partial, job.completed, job.failed), and ensuring reliability through retry with exponential backoff, attempt limits, and idempotent handlers.

In this lecture we examine why long‑lived streams ( SSE/ HTTP-stream) and MCP events are especially vulnerable under production load, how to introduce limits on event frequency and volume (including throttling job.progress), how to design backpressure via bounded buffering and conflation, and which metrics and logs you need to see growth of the event-rate, delivery lag, and a reconnection storm. With GiftGenius examples, we show practical TypeScript snippets: limiting progress frequency, capping parallel streams (response 429), a basic event counter, and approaches to UX degradation.