Feat/add litellm provider

[RheagalFire]

Summary

• Adds LiteLLM as a first-class AI gateway provider, giving Mobilerun users access to 100+ LLM providers (OpenAI, Anthropic, Azure, AWS Bedrock, Google Vertex, Groq, Mistral, etc.) through a single unified interface.
• LiteLLM is installed as an optional extra (pip install mobilerun[litellm]), keeping the base install unaffected.
• Uses litellm SDK directly (import litellm / litellm.completion()) with a custom MobilerunLiteLLM wrapper extending llama-index's CustomLLM.

Changes

• pyproject.toml - Added litellm optional dependency group with litellm>=1.83.0,<2.0.0
• mobilerun/agent/providers/registry.py - Added LiteLLM provider family with API key auth mode and env key slot
• mobilerun/agent/utils/llm_picker.py - Added MobilerunLiteLLM class (extends CustomLLM, calls litellm.completion() / litellm.acompletion() directly with drop_params=True), registered in SUPPORTED_PROVIDERS, PROVIDER_ALIASES, and load_llm() dispatch
• mobilerun/agent/usage.py - Added LiteLLM to supported providers list, alias map, and usage extraction (prompt_tokens, completion_tokens, total_tokens)
• mobilerun/cli/main.py - Updated CLI --provider help text to include LiteLLM
• README.md - Added LiteLLM to feature list and installation instructions
• tests/test_llm_picker.py - 9 tests: alias resolution, class loading, base_url translation, provider/model format, bare model name, API key forwarding, temperature forwarding, max_tokens forwarding, unsupported provider error
• tests/test_provider_registry.py - 1 test: provider family variant resolution
• tests/test_usage.py - 6 tests: usage extraction, tracker support, null usage fallback, dict raw response, total computation from parts, no-raw-response error

Tests

1. Unit tests (71/71 passing):

tests/test_llm_picker.py::test_litellm_loads_correct_class PASSED
tests/test_llm_picker.py::test_litellm_alias_resolves PASSED
tests/test_llm_picker.py::test_litellm_base_url_translates_to_api_base PASSED
tests/test_llm_picker.py::test_litellm_provider_slash_model_format PASSED
tests/test_llm_picker.py::test_litellm_bare_model_name_preserved PASSED
tests/test_llm_picker.py::test_litellm_api_key_forwarded PASSED
tests/test_llm_picker.py::test_litellm_temperature_forwarded PASSED
tests/test_llm_picker.py::test_litellm_max_tokens_forwarded PASSED
tests/test_llm_picker.py::test_litellm_unsupported_provider_raises PASSED
tests/test_provider_registry.py::test_litellm_family_resolves_api_key_variant PASSED
tests/test_usage.py::test_litellm_extracts_usage_from_response PASSED
tests/test_usage.py::test_track_usage_supports_litellm PASSED
tests/test_usage.py::test_litellm_usage_with_null_usage_returns_zeros PASSED
tests/test_usage.py::test_litellm_usage_with_dict_raw_response PASSED
tests/test_usage.py::test_litellm_usage_computes_total_when_missing PASSED
tests/test_usage.py::test_litellm_no_raw_response_raises PASSED
============================== 71 passed in 1.08s ==============================

2. Lint: ruff check mobilerun/ tests/ -> All checks passed.

3. Live E2E against Anthropic (via Azure Foundry):

from mobilerun.agent.utils.llm_picker import load_llm
from mobilerun.agent.usage import get_usage_from_response
from llama_index.core.base.llms.types import ChatMessage

llm = load_llm('LiteLLM', model='anthropic/claude-sonnet-4-6',
               api_base='https://amanrai-test-resource.services.ai.azure.com/anthropic')
messages = [ChatMessage(role='user', content='What is 2+2? Reply with just the number.')]
response = llm.chat(messages)
usage = get_usage_from_response('LiteLLM', response)

Provider class: MobilerunLiteLLM
Model: anthropic/claude-sonnet-4-6
Response: 4
Usage: prompt_tokens=20, completion_tokens=5, total_tokens=25

4. Pin resolution verified: pip install -e '.[litellm]' resolves cleanly against litellm>=1.83.0,<2.0.0.

Risk / Compatibility

• Additive only. No existing providers touched.
• litellm is an optional extra - base install and all existing providers remain unaffected.
• Uses litellm SDK directly (import litellm) - no additional wrapper packages needed.
• LiteLLM appears in mobilerun configure wizard automatically via the provider family registry.

Example usage

CLI:

# Install with LiteLLM support
pip install "mobilerun[litellm]"

# Use with any LiteLLM-supported provider
mobilerun "Open Settings" --provider LiteLLM --model anthropic/claude-sonnet-4-6

# Or with Azure OpenAI
AZURE_API_KEY=... AZURE_API_BASE=... \
mobilerun "Take a screenshot" --provider LiteLLM --model azure/gpt-4o

# Or with AWS Bedrock
AWS_ACCESS_KEY_ID=... AWS_SECRET_ACCESS_KEY=... \
mobilerun "Open Chrome" --provider LiteLLM --model bedrock/anthropic.claude-v2

Python SDK:

from mobilerun.agent.utils.llm_picker import load_llm

llm = load_llm("LiteLLM", model="anthropic/claude-sonnet-4-6")
# Use with MobileAgent as usual