OpenAI-compatible providers that only support json_object (DeepSeek,
Mistral, some vLLM/llama.cpp) reject the json_schema response_format
with a 400, and the resulting error was swallowed silently: not logged
server-side and never rendered by the background-task widget, leaving
only a generic "no reservations" message.
- Retry the chat/completions request once with response_format
json_object when the json_schema attempt returns 400 (non-NuExtract
only); the system prompt already dictates the output shape
- Log swallowed llm-parse errors with an [llm-parse] tag so failures
show up in server logs
- Render task warnings under the empty-preview note in the background
tasks widget so the actual provider error reaches the user
jubnl
2195facbe5