Development Update — August 18

Two campaigns reach their point today. Route ranking stops guessing at transport quality by type and starts using measured throughput and end-to-end latency, in a three-phase build that goes from free passive observation to an active packet-pair probe to ranking that treats the type cost as a mere prior. And the warm-standby mux primitive — a leg that stays alive but parked — lands and is wired through the policy ABI, so rotating-bw rotates by promoting a warm standby instead of tearing a live leg down mid-stream. Alongside them, an app can finally override the visor-global mux defaults, and the deployment HTTP and CXO surfaces are gzipped.

Skywire: Measured Route Ranking

Ranking judged transports by RTT, which is throughput-blind — a WebRTC SCTP datachannel has low RTT but poor goodput, so it ranked well and sneaked onto mux legs. Rather than hardcode a per-type penalty (a guess), each transport gets a measured throughput estimate. 3987 is phase 1, the free half: passive observation on the existing transport-maintenance loop folds each interval’s observed goodput into a per-transport high-watermark, reading cumulative counters only, so it can never disrupt a live route or cost bandwidth. 3988 is phase 2, for idle transports a passive read never exercises: a non-obtrusive packet-pair probe sends a short back-to-back train (~11 KB) and computes the bottleneck bandwidth from the inter-arrival dispersion on a single receiver clock — no clock sync, no saturating burst. 3989 closes the loop: when a transport has a measured estimate, ranking uses it directly — a link measured fast scores zero regardless of type, a slow one is penalized regardless of type — and the per-type cost (3983, deprioritizing slow types) degrades to a prior used only until a real measurement exists.

Skywire: End-to-End Leg Latency

3985 fixes a mis-attribution: the leg-liveness pong already measured each leg’s full round-trip time, then threw it into an aggregate and credited it to the first hop, so the policy’s “evict the slowest leg” and the LatencyMs telemetry were judging the first-hop transport RTT, not the leg’s whole-route latency — very different on a multi-hop leg with a fast first hop and a slow second. A per-leg end-to-end EWMA keyed by transport ID (surviving leg re-indexing) is now folded from the pong and reported as the leg latency, so the adaptive presets decide on the real route latency. 3984 retransmits SACK gaps on the fastest leg rather than the mode-driven pick. 3990 ranks disjoint route candidates by transport cost so a mux stops taking WebRTC legs, and 3986 plans the aux legs over the global TPD graph (route-finder first) instead of local-only transports.

Skywire: Warm-Standby Legs and No-Dip Rotation

3973 is the RFC and 3974 the dormant primitive: a per-leg standby[] flag parallel to ready[], folded into legReadyAt so a standby leg is uniformly skipped by every send path while its rules stay installed and its keepalive/liveness loops keep it alive — setLegStandby demotes or promotes instantly with no route setup, and nothing calls it in a live path yet. 3977 wires demote/promote through the on_tick ABI and 3979 has the router observe and emit the gate state so the adaptive preset can hot-swap on it. 3995 puts it to work: rotating-bw had dropped the oldest live leg every interval, tearing down its transport mid-stream (a dip at mux=2, a fatal stream EOF at mux=4, measured live), and now provisions one warm standby beyond the target width and rotates by promoting the standby and demoting the oldest active leg — both pure flag-flips, no teardown, so the demoted leg’s in-flight bytes still drain and the active set rotates with no dip. 3996 takes it further: the mux is managed by standby and transport type, anchoring the active set on reliable types (stcpr/sudph/squicr/stcp) and parking fragile ones (webrtc/ws/wt/dmsg) on warm standby, since an all-webrtc group collapses under load (measured: EOF at ~14s) while a reliable-anchored group runs stable for minutes.

Skywire: The Adaptive Default and Its Harness

3978 adds a composite adaptive default preset (size + membership + explore) and 3981 makes it a fast default with a lean start that inherits the operator’s min-hops. 3980 is the per-leg mux telemetry harness — NDJSON with lifecycle events and a chart — and 3982 documents a reproducible measurement rig script and controlled far-end guide, the instrumentation the rotation work is measured against.

Skywire: Per-App Overrides, Interstitial Copy, gzip

3994 lets an app override the visor-global min_hops/mux_routes: a visor configured with aggressive defaults forced every skynet dial into multi-hop mux with no opt-out, which breaks a 1:1 forward to a single-route port, and now an explicit --routes 1 reaches the networker instead of being silently dropped. 3993 adds a --routing-policy flag to skynet start for parity with proxy start. 3991 gives the proxy interstitial mechanism-specific copy and Skywire cloud branding, and 3992 adds a proxy loadtest rig — a steady controlled sink with an exact goodput/gap recorder — the far end the mux fixes are measured against. On the deployment side, 3975 gzips the uptime-tracker, service-discovery and route-finder HTTP responses and 3976 gzips the all-transports CXO feed for bandwidth parity with the HTTP path.