Client or integration
Direct HTTP/API client
Area
Proxy and routing
Summary
On official v2.63.0, the first combo target can still dispatch after the shared request-send budget refuses its reservation. Expected: denied first-target reservation makes zero upstream inference requests and returns the existing typed local 429 request_send_budget_exhausted. For later targets, denial should preserve the last real upstream failure without dispatching the denied target.
Reproduction
- Configure a failover combo.
- Arrange for the shared request send budget to be exhausted before the first combo dispatch.
- Enter
executeComboResponses.
- Observe that a denied first reservation can fall through into physical dispatch.
- Local regression coverage asserts first-target refusal => upstream hit count 0 and typed 429; later-target refusal => backup hit count 0 and prior upstream response preserved.
Version
2.63.0 / commit 96b1406
Operating system
Ubuntu Server Linux x86_64
Provider and model
Not provider-specific
Logs or error output
Expected local error code: request_send_budget_exhausted
Screenshots and supporting files
Related to PR #4609, which established the shared request-send budget. This report is the remaining combo first-target admission gap.
Redacted configuration
{"combo":{"strategy":"failover","targets":["provider-a/model-a","provider-b/model-b"]}}
Checks
Client or integration
Direct HTTP/API client
Area
Proxy and routing
Summary
On official v2.63.0, the first combo target can still dispatch after the shared request-send budget refuses its reservation. Expected: denied first-target reservation makes zero upstream inference requests and returns the existing typed local 429
request_send_budget_exhausted. For later targets, denial should preserve the last real upstream failure without dispatching the denied target.Reproduction
executeComboResponses.Version
2.63.0 / commit 96b1406
Operating system
Ubuntu Server Linux x86_64
Provider and model
Not provider-specific
Logs or error output
Expected local error code: request_send_budget_exhaustedScreenshots and supporting files
Related to PR #4609, which established the shared request-send budget. This report is the remaining combo first-target admission gap.
Redacted configuration
{"combo":{"strategy":"failover","targets":["provider-a/model-a","provider-b/model-b"]}}Checks