SpecAt reset, credits equal DEPTH (16), the receive buffer’s capacity, since nothing is in flight. A full-depth burst is one 16-flit transfer: the receiver returns no credit until all 16 flits land, then returns all 16. (Everyday traffic acks per flit and never nears the limit, so it isn’t modeled.)
link_model is what a model writes with confidence from that spec: credits = DEPTH - 1 reads like the index idiom used everywhere else (buf[DEPTH-1] is the last slot). But credits is a count, not an index. Right now the burst never completes.
Fix link_model so a full-depth burst completes.
One working answer: brokenbench's own fix for this exercise. Try a few runs first.
What changes, against the code you started with:
docker run --pull always --rm -p 127.0.0.1:8787:8787 ghcr.io/adibis/brokenbench-runnerLooking for a runner on this computer…
The first start downloads about 250 MB; after that it starts in seconds and only downloads again when there's an update. Your code goes to the runner on your computer, not to chipdv.io.