mirror of
https://github.com/block/buzz.git
synced 2026-11-08 10:21:46 +00:00
ad13ed8411
Shared compute pinned BUZZ_AGENT_THINKING_EFFORT. That is not a sampling hint: mesh translates `reasoning_effort` into the chat template's `enable_thinking` flag, so any value Buzz sends overrides the default the model authors shipped — and the right value is model-specific. Measured against a request captured from the real agent (full ~43k-token system prompt, real dev-mcp toolset, tool_choice=auto): effort gemma-4-E4B Qwen3-8B Qwen3-8B output tokens none 0/8 8/8 ~121 low 4/4 7/8 ~470 unset 6/6 8/8 ~500 `none` (the previous default, chosen when shared compute was Qwen3-only) stops gemma tool-calling outright, so its replies were never published. Enabling thinking instead makes Qwen3 spend ~4x the output budget, which risks the 4096 cap this preset sets. Omitting the field leaves `enable_thinking` as None, so mesh passes no override and each model uses its own template default: 3/3 delivery on gemma-4-E4B, gemma-4-26B-A4B, gemma-4-31B and Qwen3-8B, and 3/3 replies through the real buzz-acp + buzz-agent path against a live relay. Thinking levels are unaffected: an explicit agent/persona/global value is still preserved and applied. Signed-off-by: Michael Neale <michael.neale@gmail.com>