Developer Traces LLM Pipeline Failures to Context Overflow, Not Free Server Tier

A developer migrating a batch LLM pipeline to a free server endpoint began seeing empty completions and hanging requests after roughly 40 prompts, initially suspecting the free tier was unreliable. To isolate the cause, they wrote a 40-line async probe that sent identical prompts padded to increasing lengths and recorded status, latency, and output validity. The probe revealed that failures scaled directly with input size rather than server load, ruling out infrastructure issues. The actual bug was a context-window overflow caused by concatenating a large system prompt, few-shot examples, and user input — a flaw masked previously by a paid endpoint with a larger context window. The free model's smaller effective window exposed the long-standing harness error, with a dead keep-alive connection identified as a secondary culprit.