Problem
The Python detectors can emit typed spans, but the HTTP response model retains only start, end, text, and confidence. It drops category, subcategory, and reasoning. Detector selection is also inconsistent: configuration implies multiple methods, while the start script exposes only the transformer path, and the current LLM construction path is not compatible with the server's model_path argument.
The existing API tests are outside the default test suite and load a real model. The server also serializes all requests through one process-wide lock without documenting that concurrency policy.
Decisions and work
- Decide whether the HTTP API is supported or experimental and document that status.
- Preserve optional typed-span metadata end-to-end without breaking binary-detector clients.
- Support configured detector methods correctly, or reject unsupported methods at the CLI/config boundary before startup.
- Define and document concurrency semantics; replace the process-wide lock only if detector safety permits.
- Use an injected or stub detector for network-free server/client tests and execute them in a separate API CI job.
- Keep a real-model/real-server smoke test explicitly marked as integration or manual.
Acceptance
category, subcategory, and reasoning round-trip when present and remain optional for binary detectors.- Invalid method/config combinations fail before server startup with a clear error.
- Offline tests cover the OpenAPI response shape, sync and async clients, and the selected concurrency policy.
- README/API documentation contains one typed-span request and response.
This is a contract-design issue. Please do not begin a large implementation before the supported surface is decided here.
Problem
The Python detectors can emit typed spans, but the HTTP response model retains only
start,end,text, and confidence. It dropscategory,subcategory, and reasoning. Detector selection is also inconsistent: configuration implies multiple methods, while the start script exposes only the transformer path, and the current LLM construction path is not compatible with the server'smodel_pathargument.The existing API tests are outside the default test suite and load a real model. The server also serializes all requests through one process-wide lock without documenting that concurrency policy.
Decisions and work
Acceptance
category,subcategory, and reasoning round-trip when present and remain optional for binary detectors.This is a contract-design issue. Please do not begin a large implementation before the supported surface is decided here.