[groq-common-errors skill]: structural, fix the request: 401 invalid or missing key, 400 deprecated model or bad parameters, 413 request over the context window (Llama models cap at 128K tokens, trim the prompt). Transient, retry with backoff and honor retry-after: 429 RPM/TPM/RPD limits, 500/503 Groq-side outage or overload with a model fallback. The 413-on-TPM variant means one request exceeded the per-minute token budget: shrink the request, not the retry interval.

Context: Retry the wrong Groq error and you burn quota for nothing.