[nicoloboschi]

Closes #4935.

Problem

On the soft (non-strict) structured-output path, the OpenAI-compatible provider decides whether to send response_format={"type": "json_object"} from a hardcoded provider-name list: lmstudio, ollama and volcano skip it, llamacpp follows HINDSIGHT_API_LLAMACPP_NO_GRAMMAR, and every other provider sends it.

provider=openai with a custom base_url pointing at a local server is a common setup, and it had no way to opt out. The reporter's local server can't constrain output, so it rewrites the prompt with its own "respond with valid JSON only" line instead. A thinking model loops on that line until the LLM timeout. The same request without json_object returned valid JSON in 77 s.

Change

• New tri-state setting HINDSIGHT_API_LLM_OPENAI_COMPATIBLE_JSON_MODE, following the HINDSIGHT_API_LLM_VISION pattern:
  • unset: the provider decides, same as today
  • false: never send json_object; the schema goes in the prompt only
  • true: always send json_object
• The provider-name list moves into one helper, OpenAICompatibleLLM._supports_json_mode().
• It is a server-level setting, not per bank, like the other structured-output flags (LLM_SUPPORTS_MAX_ITEMS, LLM_STRUCTURED_OUTPUT_FORCED_TOOL).

This doesn't make a weak model write JSON. It only drops a request that some servers mishandle.

Tests

tests/test_llm_strict_schema.py now covers the provider defaults (openai sends json_object, lmstudio doesn't), both overrides, and reading the setting from the environment. The docs, .env.example and the hindsight-embed copy of the template are updated.