mirror of
https://github.com/block/buzz.git
synced 2026-08-18 06:50:31 +02:00
Michael Neale
7c3dce28ec
Shared-compute serve nodes currently run whatever reasoning behaviour the model ships with. For reasoning models (e.g. Qwen3.6-35B-A3B) that means every agent turn burns hidden thinking tokens — a 1-word reply cost 161 completion tokens in testing — which is slow and expensive under the goose / Buzz agent harness. Buzz lets mesh-llm synthesise an isolated throwaway serve config that only disables the telemetry + blobstore plugins. This points the embedded serve node at a Buzz-owned config file instead, which re-asserts those no-leak disables and adds a reasoning-off request default. Clients can still opt back into thinking per request. No mesh-llm source change and no touching the user's global mesh config. DRAFT: mechanism (config_path plumbing) verified from source and compiles; the reasoning-off *default* content is NOT yet verified end-to-end on the embedded v0.73.1 runtime. See PR description.