BFE Files

Added

• Add Server.AIUpstreamAcceptEncoding (auto/identity/off, default auto): how Accept-Encoding is rewritten on an AI upstream request, so content negotiation stays inside the coding set the gateway can decode (gzip, deflate). auto keeps only the decodable codings of the client's list (q values preserved, an empty intersection is left untouched, identity;q=0 is honoured), identity strips compression for diagnostics, off keeps the previous behaviour. The rewrite only touches the outbound header: the inbound request, the access log, mod_ai_cache and mod_compress still see the client value (issue [#1468])
• Add bfe_util.NewContentDecoder / CanDecodeContentEncoding / SupportedContentEncodings, a streaming Content-Encoding decoder covering gzip/x-gzip and deflate (zlib header sniffed, raw DEFLATE accepted). DecodeContentEncoding now delegates to it, so the non-streaming parse-only copy also understands deflate (issue [#1468])
• Add the ResStreamCompressed / ResStreamDecoded / ResStreamDecodeSkipped / ResStreamRawOnly mod_body_process counters: how many responses carried a Content-Encoding, how many got the decompress-for-parsing tap, how many could not be decoded and how many degraded mid-stream. The last two mean "this request is billed by estimation", i.e. the only signal that an upstream ignored the narrowed Accept-Encoding (issue [#1468])

Changed

• An AI upstream request whose client Accept-Encoding contains codings the gateway cannot decode no longer receives them (Server.AIUpstreamAcceptEncoding = auto, the default): the upstream is offered gzip/deflate only, so a client that asked for zstd may now receive gzip instead (still inside its own Accept-Encoding, <1% size difference). Set Server.AIUpstreamAcceptEncoding = off to keep the previous negotiation (issue [#1468])

Fixed

• Parse the usage of a compressed streaming (SSE) response instead of estimating it: mod_body_process bypassed any Content-Encoding body to a byte-wise pass-through, so cache_read_tokens stayed 0, prompt_tokens fell back to the request-body estimate and completion_tokens to the compressed-bytes estimate, and the whole prompt was billed at the input price — measured on the field at a 96% cache-hit rate that overcharges by roughly 46x and writes ai_usage_source=estimated. Compressed SSE now goes through CompressedStreamDecoder, which tees the source into a decompressor used only for parsing while the original compressed bytes are forwarded verbatim together with the Content-Encoding header; a decompression failure or an unsupported coding degrades to a byte-exact pass-through plus the estimate, visibly (issue [#1468])
• Report a skipped stream_options.include_usage injection when the request body exceeds Server.AccessibleBodySize: the injection used to fail silently, which raised the probability of an upstream returning no usage at all (issue [#1468])