[ArnavAgrawal03]

Video frame descriptions failed on OpenAI reasoning models. VisionModelClient.get_frame_description always sent max_tokens = 300, and gpt-5 and later reject it:

litellm.BadRequestError: OpenAIException - Unsupported parameter: 'max_tokens' is not supported with this model. Use 'max_completion_tokens' instead.

Image OCR already handled this (#454). This moves that logic into core/parser/model_params.py and uses it for both.

Changes

• add_output_limit sets max_completion_tokens for OpenAI models and max_tokens for other providers, keeps a limit set on the registered model, and allows reasoning_effort through LiteLLM. Image OCR calls it instead of its inline copy, so OCR behavior is unchanged.
• Frame descriptions get a 4096-token budget instead of 300, so a reasoning model has room to think before it writes. Ollama models still get no limit, as before.
• A description that runs out of output with no text raises a ValueError that suggests reasoning_effort = "low", instead of indexing a blank frame.
• Rate limits, timeouts and outages raise ConnectionError, so the ingestion worker retries, as for OCR.
• DOCKER.md says reasoning models work for frames and recommends reasoning_effort = "low".

Testing

• test_video_transcription.py adds 4 tests: OpenAI reasoning model parameters, other providers and registered limits, running out of output, and transient errors. With test_video_parser and test_image_ocr_model: 37 passed.
• Live calls through get_frame_description on a generated frame with a red "Error 4012: Invoice total does not match" box. openai/gpt-6.1-sol returned the error above before this change. After it, openai/gpt-6.1-sol (with and without reasoning_effort = "low") and gpt-4.1-mini all described the box and quoted its text.
• ruff with its long-standing default rules (E4, E7, E9, F) passes on the changed files.

🤖 Generated with Claude Code

---

^{Need help on this PR? Tag @codesmith-bot with what you need.}