Support OpenAI-compatible endpoints via OPENAI_BASE_URL

[abliteration-ai]

What type of PR is this? (check all applicable)

• Feature

Related issue: N/A

What changed?

Detection/openai_config.py — the single place where OpenAI clients are constructed — now reads the optional OPENAI_BASE_URL environment variable:

• get_openai_client() passes it to the SDK as base_url when set, so every consumer (create_chat_completion, create_reasoning_completion, the ADR baseline agent) can be pointed at any OpenAI-compatible API.
• get_openai_config() includes a base_url key when the variable is set.
• When OPENAI_BASE_URL is unset, behavior is byte-for-byte identical to before (default api.openai.com endpoint).

Also adds unit tests for both helpers (Detection/tests/test_openai_config.py) and a short README section documenting the option with a provider example (Abliteration.ai).

Why?

The triage model is configurable in config_detector.yaml, but the client was hardcoded to the hosted OpenAI API. Teams running OpenAI-compatible gateways or alternate providers (self-hosted vLLM, LiteLLM proxies, etc.) had no way to route detector traffic there. OPENAI_BASE_URL follows the OpenAI SDK's own convention, so it's the least surprising knob.

How did you test it?

• uv run pytest tests/test_openai_config.py — 8 passed (4 existing + 4 new covering default endpoint, custom base URL, and the config dict with/without OPENAI_BASE_URL).
• black --check clean on the touched test file; openai_config.py keeps the file's existing style.
• Verified no new flake8 findings in the touched code (the file's pre-existing warnings are untouched).

Potential risks

None expected: the new code path only activates when OPENAI_BASE_URL is set; existing deployments without it are unaffected. The Anthropic/Claude path is untouched.

[CLAassistant]

All committers have signed the CLA.

[pengyuzhang]

Thanks for the PR, and especially for the tests — the coverage added in test_openai_config.py is useful regardless of the outcome here.

We're going to close this one, for two reasons.

1. The behavior already works without a code change. The OpenAI Python SDK reads OPENAI_BASE_URL from the environment itself. Verified against the version pinned in this repo (openai 1.100.2):

>>> os.environ["OPENAI_BASE_URL"] = "https://example.test/v1"
>>> OpenAI(api_key="sk-test").base_url     # no base_url argument passed
https://example.test/v1/

So setting the env var already points the triage LLM at any OpenAI-compatible endpoint on main today, and the added branch in get_openai_client() duplicates SDK behavior. The get_openai_config() change has no effect either — that function currently has no callers in the repo, so the new base_url key isn't consumed anywhere.

2. The README section. We'd prefer to keep third-party endpoint documentation vendor-neutral rather than build the example around one provider's URL, key name, and model. Separately, we'd rather not recommend running triage on an uncensored model: triage is the high-recall gate in front of the reasoning agent, and the paper's Table 2 numbers assume gpt-4o.

We'll document the SDK's existing OPENAI_BASE_URL support with a short neutral note instead.

Thanks again for taking the time to look at the project.