Development Update — June 6

A day spent unsticking the things that quietly leak, wedge, or fall back to plain HTTP. The headline is a dmsg-only fix for the blank hypervisor network tab — a casualty of the new default config — alongside a setup-node connection pool that stopped handing out dead connections and a CXO goroutine leak that was dragging whole hosts down. Around them: the rewards calc sourcing uptime from the transport-gated tracker, and the got CLI learning to actually use its SOCKS5 proxy.

Skywire: Fix the Blank Hypervisor Network Tab on dmsg-Only Visors

3034 fix(visor): fetch service data over dmsg-http first (fixes blank hypervisor network tab) repairs a casualty of the new dmsg-only default config. The hypervisor network view aggregates service-discovery and uptime-tracker data via FetchServiceData, which was HTTP-only — it read the plain-HTTP service URL and did an http.Client.Get. On a dmsg-only visor those HTTP fields are empty, so the fetch returned “service sd not configured” and the network tab rendered blank. Service data is now fetched dmsg-HTTP first whenever a *_dmsg URL is configured, with plain HTTP used only as the dual-config fallback or the sole http-only path; the fetch order is encoded and unit-tested, covering SD, UT, TPD, AR and RF.

Skywire: Stop the Setup-Node Pool From Handing Out Dead Connections

3032 fix(router): stop the setup-node pool from handing out dead connections chases down circuit breakers that were wedging open against perfectly reachable visors, blocking all multihop setup to them — the skysocks proxy server among them. Live RSN /stats showed thousands of id-reservation failures with “connection is shut down” against demonstrably reachable keys, and a pprof dump showed dozens of RPC client goroutines parked on dead dmsg streams.

The root cause was a timeout mismatch in the connection pool. After each call the stream read deadline is armed at 2 minutes so the RPC input goroutine eventually unblocks and frees its ephemeral port — but the pool’s TTL was 5 minutes, so between the 2-minute and 5-minute marks a pooled connection was already shut down yet still poolable. Get returned it unprobed, the next reservation failed with ErrShutdown, and three such failures tripped the destination breaker for 5–30 minutes. Get now reuses a pooled connection only while it has idled below the read-deadline horizon and dials fresh past it, with the pool TTL pinned to that horizon so eviction can never outlive liveness.

Skywire: Unblock a Stuck Transport Read on Close

3033 fix(cxo/node): unblock stuck transport read on Close to stop goroutine leak addresses a transport-discovery CXO node that leaked goroutines without bound — observed live at ~3000 stable, spiking to 37k under reconnection load, driving the process to ~2 GB RSS and ~300% CPU and starving its co-located route-finder and address-resolver. The read loop parks in io.ReadFull on the dmsg stream; when a remote peer vanishes without a FIN/RST, that read never errors, and Close() alone does not reliably interrupt an in-flight read — only a read deadline does. So the idle watchdog’s Close() never reached the stuck reader, and the whole connection leaked. Close now arms a past read deadline before closing, forcing any parked read to return so the loop exits and the connection tears down cleanly. A regression test drives Close() against a conn whose read blocks and whose Close() does not interrupt it — the production failure mode.

Skywire: Rewards Source Uptime From the Transport-Gated Tracker

3031 feat(rewards): source uptime from TPD-integrated tracker; drop manual tp gate removes a redundant, race-prone bookkeeping file from the reward calc. “Eligible visor” used to be the conjunction of two snapshots: heartbeats from the standalone uptime tracker, and a separate hourly tp-collect job recording which keys held ≥2 transports. The TPD-integrated uptime endpoint already publishes exactly that conjunction — it only credits heartbeats while a visor holds ≥2 transports — so the second file was duplicate work vulnerable to races between the two snapshots. The calc now hits the transport-discovery’s /uptimes?v=v2, drops the transports file, and flips --require-tp to default false; tp-collect survives as a diagnostic but is no longer load-bearing.

Skywire: The got CLI Actually Uses Its Proxy

Two fixes so skywire-cli got can talk through the resolving SOCKS5 proxy. 3029 feat(cli/got): accept socks5:// and socks5h:// proxy URL schemes makes the -x flag take curl-style scheme prefixes: socks5h:// resolves the destination at the proxy (required for dmsg web’s synthetic <pk>.dmsg hostnames, which a local resolver can’t look up), socks5:// resolves locally then sends the IP, and a bare host:port is treated as socks5h:// for backward compatibility. 3030 fix(got): propagate Got.Client and Got.ctx into hand-built Download then closes the gap that left the proxy silently bypassed anyway: a hand-built *Download with a nil client fell through to the default client, discarding the SOCKS5 dialer that NewWithProxy had configured. The receiver’s client and context are now injected before init.

And a CLI cache fix: 3028 fix(cli/ut): write JSON cache file when –cdu/–cdt is set restores the documented <dir>/uptimes.json file that the bbolt cache refactor stopped writing — the embedded reward.sh jq-slices that file, and with it missing the per-day uptime data came out empty and the version-history chart froze on the last good day.