dev404ai commented on issue #13340:
URL: https://github.com/apache/apisix/issues/13340#issuecomment-5253066666

   While investigating this issue, I ran a small deterministic probe on Apple 
arm64 against unchanged APISIX source at `master` 39b9e43. It exposed a 
concrete fixture-fidelity gap on the request side — the same side the quoted 
concern is about.
   
   The existing `t/plugin/ai-proxy-fixture.t` baseline passes all 50 
assertions. In a separate local probe against the unchanged fake server, `GET 
/v1/messages` with only `X-AI-Fixture: anthropic/messages-basic.json` returned 
200 and the recorded response despite having no Anthropic auth headers or 
request body. As a control, the existing Bedrock Converse handler rejected the 
same invalid method with 400.
   
   This shows that the fake server can accept an invalid provider request shape 
and still return a plausible fixture response. Response-focused fixture tests 
could therefore miss adapter regressions in the method or required fields.
   
   To be clear about scope: this is a test-fidelity observation, not a security 
report. `t/` and `t/lib/` are out of the project's threat model per §4.2 and 
§4.3 point 10 of `docs/en/latest/security-threat-model.md`.
   
   This evidence does not by itself justify a separate LLM fake-server 
container. The current server already enforces a detailed request contract for 
Bedrock, so provider-specific validation in the existing server appears to be 
the smallest change that closes the demonstrated gap.
   
   A separate container may still be useful if the intended scope includes 
behavior that the current server cannot represent reliably, such as stateful 
interactions, provider SDK behavior, transport-level differences, or more 
realistic streaming. Is there a container-specific requirement beyond 
request-contract validation that should define the boundary of #13340?


-- 
This is an automated message from the Apache Git Service.
To respond to the message, please log on to GitHub and use the
URL above to go to the specific comment.

To unsubscribe, e-mail: [email protected]

For queries about this service, please contact Infrastructure at:
[email protected]

Reply via email to