Each request uses only the working code doc (optional) + your current prompt. No full chat memory.
How it works
We do not resend prior turns. Only the system rule, optional working code doc, and your current prompt are included.
Use a single-file "working doc." If you need multi-file context, paste the relevant parts or switch to a short summary of the non-active files to save tokens.
Lower temperature (e.g., 0.2–0.5) for more deterministic code.
Adjust max_tokens to cap output length (your real usage prints above).