Repository navigation
Why does the on-topic chat classifier fail open instead of fail closed? #18
|
The chat panel runs every message through an on-topic classifier before letting it reach the review context. If that classifier call itself fails (timeout, API error, etc.), should the message be blocked by default (fail closed) or let through (fail open)? |
Replies: 1 comment
|
Fail open. The classifier is a UX guardrail that keeps the chat on-topic, not a security control, so its failure mode should not be able to break the actual feature it is guarding. If the classifier call errors out, the chat still needs to answer the users question about their code review. Blocking every message whenever the classifier (a separate API call) has a bad moment would turn an occasional network hiccup into a full outage of the chat feature itself, which is a worse failure than occasionally letting an off-topic message through. The cost of fail-closed (the whole feature breaks) is much higher than the cost of fail-open (the classifier misses a message here and there). |
Fail open. The classifier is a UX guardrail that keeps the chat on-topic, not a security control, so its failure mode should not be able to break the actual feature it is guarding. If the classifier call errors out, the chat still needs to answer the users question about their code review. Blocking every message whenever the classifier (a separate API call) has a bad moment would turn an occasional network hiccup into a full outage of the chat feature itself, which is a worse failure than occasionally letting an off-topic message through. The cost of fail-closed (the whole feature breaks) is much higher than the cost of fail-open (the classifier misses a message here and there).