[CodePrometheus]

Description

Goal

Decode each byte of a compressed streaming response once, so the cost grows linearly with the response length.

Before

decodeStreamingContent appends every chunk to a buffer and decompresses the whole buffer again, so CPU and allocations grow with the square of the response length. A few hundred KB of brotli SSE costs seconds of ext_proc CPU per request.

After

One decompressor per response is fed one chunk at a time through an iter.Pull coroutine. If the stream ends before the last chunk, the router processor releases the decoder when the ext_proc stream closes. Error handling is unchanged: running out of input mid-stream is fine, and a stream truncated at end of stream or corrupt data still fails the request.

ProcessResponseBody benchmark, one chunk per SSE event:

stream	before	after
gzip, 110 KB	133 ms / 352 MB allocated	0.72 ms / 0.96 MB
gzip, 444 KB	1.88 s / 4.55 GB	2.64 ms / 3.6 MB
br, 110 KB	1.19 s / 889 MB	1.75 ms / 1.4 MB
br, 444 KB	18.8 s / 12.4 GB	6.6 ms / 5.3 MB

[netlify]

✅ Deploy Preview for theagentrouter canceled.

Name	Link
🔨 Latest commit	06016f3
🔍 Latest deploy log	https://app.netlify.com/projects/theagentrouter/deploys/6aeed832be8338b7bef3b573

[codecov]

Codecov Report

❌ Patch coverage is 95.55556% with 2 lines in your changes missing coverage. Please review.

Files with missing lines	Patch %	Lines
internal/extproc/processor_impl.go	85.71%	2 Missing ⚠️

📢 Thoughts on this report? Let us know!