[charchit7]

What

Adds Google Gemini as a first-class model provider, selectable via the
gemini/<model> (Google AI Studio) and vertex_ai/<model> (Vertex AI) id
prefixes.

Previously any gemini/� id fell through to the HuggingFace router catch-all
and was misrouted (e.g. openai/gemini/�). This wires Gemini through
LiteLLM's native Google adapters instead.

Changes

• agent/core/llm_params.py: new gemini/ + vertex_ai/ branch in
  _resolve_llm_params. Forwards reasoning_effort as a top-level kwarg,
  which LiteLLM translates into the thinking budget for Gemini 2.5+ thinking
  models. Adds _GEMINI_EFFORTS = {low, medium, high}; minimal normalizes
  to low, and xhigh/max are handled by the existing effort cascade
  (no per-model capability table, consistent with the other providers).
• agent/core/model_switcher.py: adds gemini/gemini-2.5-pro and
  gemini/gemini-2.5-flash to the /model suggestions, marks gemini/ and
  vertex_ai/ as direct prefixes (skip the HF router catalog lookup), and
  updates the /model help text and is_valid_model_id docstring.
• tests/unit/test_llm_params.py: 6 new tests â�� effort forwarding,
  vertex_ai/ route, minimal�low normalization, strict-mode rejection of
  unsupported efforts, non-strict effort drop, and that Gemini is not routed
  through the HF router.

Usage

export GEMINI_API_KEY=...            # AI Studio route
ml-intern --model gemini/gemini-2.5-pro
# or, in-session:  /model gemini/gemini-2.5-pro

@lewtun Please review whenever you are free.

[lewtun]

Thanks for contributing to ML Intern. The project, including its CLI, has been retired and this repository is being archived. We will therefore not be able to review or merge this work. For continued AI assistance, please use HuggingChat.