fix(ci): set LLM_PROVIDER=mock to prevent OpenAI init at import time

[steadhac]

No description provided.

[joshlochner]

Could we use the Ollama build in #497 as a way to have an instance of LLM_PROVIDER instantiated that doesn't rely on mock

[stealthwhizz]

@joshlochner Using Ollama in CI would require installing the binary, pulling a model, and starting a server on every run. That adds significant time and disk usage, and LLM responses are non-deterministic so unit tests would be flaky. LLM_PROVIDER=mock is the right fix for CI - it's instant, deterministic, and tests the code logic rather than model behaviour.

Ollama makes more sense for local dev or E2E tests.

[steadhac]

@joshlochner Thanks for the suggestion! The reason we use the mock client in CI rather than Ollama/Llama is that it keeps the test environment self-contained — no running server, no model weights, no GPU dependency. A live Ollama process isn't available in the CI runner, so tests would fail with a connection error the same way they currently fail with the missing OpenAI key.

The mock client is also intentional for unit tests specifically: it returns deterministic responses, runs instantly, and tests the code around the LLM rather than the LLM itself. If you want to validate actual model behavior, that belongs in a separate integration test suite that runs on demand, not on every PR.

Happy to discuss adding an optional Ollama integration test job that developers can trigger manually — but for the default CI pipeline, mock is the right choice.

Thanks, @stealthwhizz, for answering.