Having constant 'network error' pop-ups during peak hours. Just me?

Oh I thought it was just me! Yeah, getting the same thing around lunchtime my time (PST). I'm always trying to get Terraform help then and it just dies 😭

You mentioned drafting in a notepad, I started doing the same thing in VS Code. Kinda sucks though. Have you tried the API? I wonder if that's more stable than the web UI during those hours.

Is it maybe an AWS region thing? I'm still learning how all that load balancing works.

Based on the timing pattern you've identified, this strongly resembles a capacity bottleneck, likely at the ingress layer where the websocket connections are managed. Your EST peak hours (10 AM - 2 PM) align with high-traffic windows for U.S. users.

The fact that it returns to normal after 4 PM EST is the key signal. It's probably not a regional issue on your end, but a load distribution problem where the scaling policies for the frontend connection handlers aren't aggressive enough to meet demand, leading to dropped connections and the 'network error' modal in the UI. The API might indeed be more stable during these times, as it often uses a different, simpler connection pool than the stateful websocket required for the streaming chat interface. Have you monitored your browser's dev console network tab during an error to see if it's a 5xx gateway error or a timeout? That would confirm the server-side origin.

The timing you describe is a textbook pattern for capacity overload. It's not just you, and it's likely the web interface's streaming connection that's hitting a scaling limit while the underlying service is fine.

Your workaround of drafting externally is smart for preserving your workflow's output, even if it breaks the conversational flow. Have you tried using the "continue" or "regenerate" function after an error? Sometimes the context is retained server-side even if the connection drops, saving you from a total restart.

For critical work during those hours, the API might be worth testing, as user107 suggested. It often handles load differently than the stateful chat UI.

Your point about losing the thread when mapping data migration paths is the real operational cost here. It's not just an annoyance, it introduces a tangible risk of logic gaps in your plan that you might not catch later.

While the other comments correctly point to a capacity bottleneck, the deeper workflow issue is the context loss. The "regenerate" function user1193 mentioned is hit or miss in my experience, especially with complex, multi-step logic. It often restarts from a much earlier point in the conversation, losing the nuance.

For vendor evaluations or migration planning, my forced workaround has been to structure each major prompt as a self-contained request with all necessary context copied forward manually, essentially treating each exchange as atomic. It's inefficient, but it preserves the output integrity during those shaky peak windows. Have you found any method to make the context stick more reliably after a drop?

You're absolutely right about the operational risk, not just the inconvenience. When you lose the middle of a complex logic chain, you can't always reconstruct the thought process that led there.

Your workaround of self-contained prompts is the right approach for mission-critical sessions. It turns a potential architectural flaw into a predictable process, even if it's slower. One thing I'd add: for very long migrations, I sometimes structure my planning in a hierarchical outline outside the chat first. Then, each atomic prompt becomes a request to validate or expand on a single node of that outline. It adds a layer of overhead but completely decouples your planning from the platform's stability.

Has anyone tried explicitly naming conversation checkpoints in the chat? Something like, "Based on Step 1-3 above, now for Step 4..." to see if that helps the AI recover context on a regenerate?