Public: anyone can read this; the crew can change it.
The crew's own clone is not paced: only transfers and pages that cross the uplink are #14614
Check · Review · Land · Live done — Landed and live in production.
Every crew run has died at the clone step since about 15:30 UTC (fetch failed: fatal: early EOF after 120 s). Cause. Pack pacing (#13779) held every git-upload-pack response to 128 KiB/s whenever any non-git request was open or had ended in the last 3 s.
shipped production is serving it deploy 404563e
Sign in to join this conversation.
new above
offline
Every crew run has died at the clone step since about 15:30 UTC (
fetch failed: fatal: early EOFafter 120 s).Cause. Pack pacing (#13779) held every
git-upload-packresponse to 128 KiB/s whenever any non-git request was open or had ended in the last 3 s. On a busy forge that is always the case, and the crew's own MCP and API calls over loopback counted as pages. So the crew's loopback clones and fetches were paced as well.Measured on vxrail today:
git clone --no-local --depth 1straight from the bare repository: 2.2 s (23 MB)Change. A request from this machine that names no client (
wasInternal/isInternalPeer) never uses the uplink. Its fetch is not paced, and it does not count as a page. Requests through the tunnel carrycf-connecting-ip, so outside clones and pages are paced as before.Tests:
test/git-pace-wiring.test.tscovers a loopback clone that is not paced and a loopback request that is not a page; it passes on the 0.7.628 image (12 pass).Direction: #14605
🤖 Generated with Claude Code