Debugging Trap: Reasoning Model Consumed Entire Token Budget Before Writing Output

A developer building a multi-agent content pipeline repeatedly encountered a misleading 'model returned empty content' error that pointed to a network or transport fault. After extensive logging at the HTTP and streaming layers, the root cause turned out to be unrelated to connectivity — the reasoning model (deepseek-v4 via an OpenAI-compatible API) was exhausting its entire token budget on internal chain-of-thought before generating any visible output. The telltale fingerprint was a finish_reason of 'length' paired with a populated reasoning_content field but an empty content field. The fix involved two steps: raising the token budget to account for reasoning model behavior, and updating the error message to explicitly flag when reasoning has consumed the full budget with no output produced. A new reasoningOnly diagnostic flag was added to the codebase so the fallback logic retries with a larger budget rather than switching models unnecessarily.