LM Studio and Ollama run one model on one GPU — concurrent requests
cause crashes. Two fixes:
1. Per-upstream semaphore (concurrency=1) in _route_agent_chat for
lm-studio/ollama providers. All agent-routed calls to the same
base URL queue instead of hitting the GPU simultaneously.
2. skip_discovery=True when routing to a local model. Context discovery
would fire a second LM Studio call alongside the main inference.
Novel words are still registered in SOAS (low saliency) but the
LLM confirmation step waits. Configure write_model_id or a separate
agent model pointing at a cloud/remote model to re-enable live
context discovery.
3. _LLM_CONCURRENCY 2 → 1 in write_queue for the same reason.
Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
gitprov
c8e8067993