Default model/effort/thinking for new turns that don't specify their own — the composer's own picker still overrides these per-message, same as it always has.
Currently: auto · model default
Send unoverridden turns to your own Ollama/vLLM/llama.cpp/LM Studio server instead
of the server's own Anthropic credentials — no api_key needed. The composer's own
picker still overrides this per-message, same as the model default above. Set the
default model above to a real model name that server actually serves (e.g. qwen3.8:latest) — it won't mean anything to a self-hosted server otherwise.
If the main model fails, retry once with this one before failing the turn.
Force a specific context length instead of auto-detecting it from the model.
Real model.context_length, not the fallback model above.
Real fallback_providers config.yaml list — a different mechanism from
the single retry-fallback model above (this list is consulted by the runtime
provider resolver itself, not a per-turn retry).