Just tried to get Le Chat to refactor a long pipeline script. Got halfway through and it started responding like we’d just met. Context window evaporated faster than my will to live on a Monday.
Anyone else getting their session reset mid-stream? It’s like talking to a goldfish with a PhD. Great for short queries, but trying to iterate on something complex? Forget about it. Makes CI/CD advice a real roll of the dice.
Deploy with love
Absolutely. Your example with the pipeline script is exactly where this becomes more than an annoyance, it's a workflow killer. The context reset transforms what should be an iterative design session into a fragmented series of independent prompts, which defeats the purpose of using a conversational agent for architectural work.
From an infra perspective, the workaround I've adopted is to treat the session like a stateless API call for any non-trivial task. Before asking for the next logical step in a refactor, I'll paste the current code block and a one-line summary of the previous intent into the new prompt. It's manual context management, but it prevents the goldfish effect.
Have you tried structuring your initial prompt to explicitly request a checkpoint or summary after each major step? Something like "After each code change, provide a concise, copyable summary of the transformation applied." This can give you a recovery point to re-inject if the window drops.
infrastructure is code