OpenAI provider sends max_completion_tokens — breaks Mistral (and likely other) OpenAI-compatible endpoints

[philippop]

Problem

When using HINDSIGHT_API_LLM_PROVIDER=openai with HINDSIGHT_API_LLM_BASE_URL pointed at Mistral's API (https://api.mistral.ai/v1), Hindsight crashes on startup during verify_connection.

The OpenAI-compatible provider sends max_completion_tokens in the chat completion request, but Mistral's API only accepts max_tokens and rejects the request with a 422:

openai.UnprocessableEntityError: Error code: 422 - {
  'object': 'error',
  'message': {'detail': [{'type': 'extra_forbidden', 'loc': ['body', 'max_completion_tokens'],
    'msg': 'Extra inputs are not permitted', 'input': 100}]},
  'type': 'invalid_request_error'
}

RuntimeError: Connection verification failed for openai/mistral-small-latest

This is a known incompatibility — Mistral (and several other providers) haven't adopted max_completion_tokens and expect max_tokens instead. See mistralai/platform-docs-public#7.

Expected behavior

The openai provider with a custom BASE_URL should work with any OpenAI-compatible API, including Mistral. Either:

1. Use max_tokens instead of max_completion_tokens when a custom base URL is set, or
2. Add mistral as a first-class provider (with correct parameter mapping), or
3. Catch the 422 and fall back to max_tokens

Workaround

Using litellm as the provider works, since LiteLLM handles the parameter translation:

HINDSIGHT_API_LLM_PROVIDER=litellm
HINDSIGHT_API_LLM_MODEL=mistral/mistral-small-latest
HINDSIGHT_API_LLM_API_KEY=<key>
MISTRAL_API_KEY=<key>

Environment

• Hindsight v0.4.22 (Docker, ghcr.io/vectorize-io/hindsight:latest)
• Mistral model: mistral-small-latest
• Base URL: https://api.mistral.ai/v1