Implement the LLM provider contract

[Hell1213]

Description

The src/oss_dev/providers/llm/ directory is currently empty. The LLMProvider abstract base class is defined in src/oss_dev/core/contracts/provider.py with methods: chat_completion, count_tokens, model_name, context_window.

Implement a concrete LLM provider that wraps the existing client/llm_client.py logic behind this contract interface.

Tasks

• Create src/oss_dev/providers/llm/client.py with GeminiLLMProvider implementation
• Implement all 4 abstract methods
• Wire it into the ProviderRegistry
• Write tests

Skills Required

• Python
• Abstract base classes
• LLM API integration

Difficulty

Intermediate-level2,gssoc-ext

[Unsoph]

Hi @Hell1213! 👋 I'd love to work on this issue as part of GSSoC 2026.

src/oss_dev/providers/llm/ exists but is empty — there's no concrete LLMProvider implementation to bridge the LLMProvider ABC (in src/oss_dev/core/contracts/provider.py) with the existing LLMClient in client/llm_client.py.

My implementation plan:

Create src/oss_dev/providers/llm/client.py with GeminiLLMProvider(LLMProvider): Implement all 4 abstract members — chat_completion() will delegate to LLMClient.stream_chat() and yield StreamEvent objects; count_tokens() will use the client.models.count_tokens call already available via the AsyncOpenAI-compatible client; model_name and context_window will return config-driven values (e.g. gemini-2.0-flash, 1_048_576). This keeps all Gemini-specific logic in one place while the rest of the app talks only to the LLMProvider interface.

Register in ProviderRegistry (src/oss_dev/providers/llm/init.py): Call registry.register("gemini", GeminiLLMProvider, ProviderCategory.LLM) and registry.set_default(ProviderCategory.LLM, "gemini") on module import — matching the pattern the contract's ProviderRegistry API expects.

Write tests (tests/): Mock LLMClient and assert chat_completion yields correct StreamEvent types, count_tokens returns an int, and the provider is retrievable via registry.get_default(ProviderCategory.LLM).

Edge case: context_window should be read from config rather than hardcoded, so swapping models doesn't require a code change.

I'm ready to implement this and can raise a clean PR shortly. Could a maintainer please assign this to me if it's still available?

[Hell1213]

@Unsoph yeah you can work on this issue. Looking forward to your contribution! Please don't forget to star our repo! 🌟

[github-actions]

Welcome back @Hell1213! Great to see you contributing again. 🌟

I've assigned this issue to you. Happy coding! If you have any questions, feel free to ask here.

---

Note: If you find this project helpful, please consider giving us a star! ⭐ Give a Star