OpenAI-compatible runtimes differ in how they handle `response_format`:
cloud providers (OpenAI, Qwen/Dashscope, Ollama) accept
`{"type": "json_object"}`, while local runtimes like LM Studio and
llama.cpp server reject it with HTTP 400, only accepting `json_schema`
or `text`. This prevented MiroFish from running against fully-local
stacks.
Introduce `LLM_JSON_MODE` (default `json_object`) so users can opt out
of strict JSON response mode by setting `LLM_JSON_MODE=none`. The
existing prompt-based JSON + markdown-tolerant parsing already handles
the unstructured response path robustly, so `none` is viable for any
OpenAI-compatible endpoint.
Applied at all three call sites that send `response_format`:
- utils/llm_client.py (chat_json helper)
- services/oasis_profile_generator.py (persona synthesis)
- services/simulation_config_generator.py (time/event/agent config)
Documented in .env.example with guidance on when to pick each value.
Pedro Renan
666ghj