[sfloess]

Summary

• Add a standard-library Python smoke-test harness for selected provider/model candidates.
• Copy the authenticated model-catalogue inventory utility into the evaluation repository.
• Tighten chat-candidate classification so missing metadata is treated as unknown, and known task-specific models are excluded from the general chat shortlist.
• Classify all responses terminated by the output limit as incomplete, including responses containing partial text.
• Add deterministic unit tests for response classification, provider request headers, and provider endpoint configuration.
• Identify the HTTP client as FlossWare-ProviderSmoke/0.1 with explicit JSON headers.
• Add DeepInfra using its OpenAI-compatible chat endpoint and a representative model candidate.
• Add NVIDIA NIM using its documented OpenAI-compatible /v1/chat/completions endpoint and openai/gpt-oss-120b candidate.
• Record requested output-token budgets, provider-native usage, normalized reasoning-token usage, visible-output length, request IDs, rate-limit headers, and finish reasons.
• Append JSONL evidence across runs; do not automatically retry 429, 402, or other failures.
• Document these as research utilities, not part of the published evaluation-ai package or production provider transport.

Safety and scope

• Reads only FlossWare-scoped environment variables and never prints credential values.
• Uses a synthetic prompt and bounded output.
• Paid/promotional-credit candidates require explicit --include-paid.
• Smoke tests are non-streaming; TTFT measurement belongs in the later capability suite.
• Never send real user data, proprietary code, or secrets.

Research notes remain in FlossWare/research/provider-api-inference.

Run deterministic tests with python3 -m unittest discover -s tests -v.

• Add --all-configured for one representative candidate per provider in a single resilient batch. Missing credentials and paid-only candidates are explicitly skipped unless --include-paid is supplied; no implicit paid fan-out.

[sonarqubecloud]

Quality Gate failed

Failed conditions
C Security Rating on New Code (required ≥ A)
C Reliability Rating on New Code (required ≥ A)

See analysis details on SonarQube Cloud

Catch issues before they fail your Quality Gate with our IDE extension SonarQube for IDE