R

reasoning-effort

DeepSeek Harness plugin: Kimi For Coding (Kimi Code) as an ordinary model provider. Serves the K3.1 Pro catalog with low/high/max thinking efforts, Kimi's request contract (max_tokens, reasoning_content replay, no store or developer role), a 1 MiB image budget for the 2 MB request-body cap, and a per-route queue that keeps one request in flight per key. Writes the catalog into the pi-ai adapter's settings section, so the Models page and every selector see it without a restart.