https://gitlab.synchro.net/main/sbbs/-/commit/93048896d5586e3c154b5b58
Modified Files:
ctrl/chat_llm.ini
Log Message:
chat_llm.ini: default to Ollama's native /api/chat provider
The stock config pointed provider=openai at Ollama's OpenAI-compatible /v1/chat/completions endpoint, which silently ignores think:false. With
a "thinking" model such as qwen3.5:9b, the model spends all of
max_tokens on hidden reasoning and returns an empty reply; the guru
chat then ends with "chat_session JS call failed" after a long,
GPU-busy pause. Reported by a sysop (Keyop) in #synchronet.
Measured against the same Ollama server with qwen3.5:9b and "hi": provider=openai returned an empty reply after 12.6s; provider=ollama
with /api/chat replied in 1.7s.
Default to provider=ollama /
http://localhost:11434/api/chat, explain
the thinking-model trap next to the setting, list the endpoint form
each provider expects, and correct the keep_alive comment, which
claimed Ollama's per-request keep_alive was unsupported (the native
provider sends it).
Co-Authored-By: Claude Opus 5.5 <
noreply@anthropic.com>
--- SBBSecho 3.38-Linux
* Origin: Vertrauen - [vert/cvs/bbs].synchro.net (1:103/705)