BFE Files

Added

• Add per-protocol upstream path rewrite via AIConf.ProtocolPaths (Pull [#1373])
• Add plaintext dial option for the EPP client and formalize the EPPTLS default (Pull [#1376])

Changed

• Improve WRR/WLC load balancing policies (Pull [#1375])

Fixed

• Fix proxy_delay_time uint32 wraparound in access logs for requests that never reached a backend (401/404/redirect/close), which corrupted downstream MySQL batch inserts (Pull [#1390])
• Fix duplicate billing of Responses API cached_tokens by restoring subset semantics (input_tokens already includes cached tokens) (Pull [#1388])
• Validate API-Key allowed/blocked models against the post-routing target model, so model redirection composes with model whitelists (Pull [#1385])
• Match AI rate limit policies by post-routing target model via the new HandleAfterAITargetModel callback point (Pull [#1385])
• Fix token model whitelist check for native Gemini path models (no more false 400 rejects) (Pull [#1383])
• Fix OpenAI provider path rewrite for client entry paths without the /v1 prefix (Pull [#1380])
• Fix AI request mode detection for client entry paths without the /v1 prefix, with the endpoint table shared in bfe_basic (Pull [#1380])
• Fix provider-native /compatible-mode/v1/responses misdetected as Chat mode, causing price miss (Pull [#1380])
• Fix unrecognized Responses API response.completed events that left Codex streaming requests unbilled (Pull [#1380])
• Recognize Anthropic message_delta as final usage in the openai adapter, fixing lost usage on cross-protocol streams (Pull [#1371])