BFE Files

Added

• Support Gemini protocol with adapter-based usage extraction and a three-stage cross-protocol usage fallback chain (Pull [#1366])
• Support responses/video_generation modes, image input token and output_cost_per_video billing (Pull [#1346])
• Support length-tier pricing (input/output_cost_per_token_above_{200k,256k,272k,512k}_tokens) and 1h-TTL cache write cost (Pull [#1365])
• Add ai_cache_write_1h_tokens, ai_image_input_tokens and ai_video_count access log fields (Pull [#1365], Pull [#1346])
• Production-harden BFE/EPP ext-proc integration: inference-pool metadata, primary/backup EPP addresses with gRPC health check, failover hysteresis and circuit breaker, TLS/timeouts config, /monitor/epp_metrics observability (Pull [#1358])

Changed

• Converge AI protocol handling (auth injection, usage extraction, error normalization) into the bfe_model_protocol adapter layer (Pull [#1350])
• Keep model prices as float64 and convert per billing item to preserve 10-12 decimal price precision (Pull [#1360])
• Allow zero total_token quota plans and treat a missing Redis balance as exhausted (Pull [#1353])
• Add tzdata package to Dockerfile (Pull [#1347])

Fixed

• Fix fresh input token billing for Anthropic high cache hit (Pull [#1350])
• Fix three billing issues: cacheRead truncation, skip /count_tokens endpoint, duplicate deduction guard (Pull [#1346])
• Skip full-request estimation billing on client abort (Pull [#1351])
• Bill complete usage for cross-protocol non-streaming responses (Pull [#1362])
• Mask raw API Key in access log (authorization header, ai_route_rule_hits rule owner, ai_auth_hit_quota_plans) (Pull [#1356])
• Relocate client CA/CRL base dirs on tls_conf reload with path (Pull [#1355])
• Attach HttpResponse for internally-constructed responses to fix panic when mod_header rewrites them (Pull [#1358])
• Send :path/:method pseudo-headers in EPP ext_proc request so chat-completions payloads parse correctly (Pull [#1365])
• Make joinPath separator-aware on Windows (Pull [#1360])