Serve qwen3.8-flash from desktop:8080 and enable it for model cycling. The endpoint needs no compat scaffolding: pi's URL auto-detection picks openai reasoning_effort, max_completion_tokens, developer role, and streamed usage, all of which gufo honors. Only supportsStrictMode is stated, since gufo compiles strict tool schemas and pi omits strict by default. All seven thinking levels work, so the opencode-go gating does not apply. Keep samplingParams clear: gufo rejects logit_bias, typical_p, tfs_z, mirostat*, dynatemp_*, xtc_*, dry_*, samplers, and draft fields with unsupported_sampling.