Development Update — May 25

The CLI’s connection to a visor gets unified around one idea: address a visor by its public key and reach it over whichever transport works. --rpc and --via now accept both dmsg://<pk> and skynet://<pk>, the dmsg and skynet bridges fold into a single RPC surface, and --via skynet:// works end to end. The control plane gets the same both-transports-are-first-class treatment the data plane already had — plus self-healing for degraded dmsg.

Skywire: Unified Visor-RPC — Address by Key, Reach by Any Transport

Until now, reaching a visor’s RPC over the network meant a patchwork: a dmsg path here, a transport-RPC subcommand there, a separate CLI keypair in some cases. Today consolidates all of it.

2811 feat(cli): unify –rpc with skynet:// + delete tp-rpc subcommand — --rpc skynet://<pk> reaches a visor’s RPC over the routed overlay, and the special-case tp-rpc subcommand is deleted — it’s just --rpc now.

2812 / 2813 feat: dmsg-direct visor-RPC listener + –rpc dmsg://, with a dmsg-bridge TCP listener — --rpc dmsg://<pk> reaches the visor over dmsg, and a dmsg-bridge TCP listener means it works without the CLI needing its own dmsg keypair — the bridge handles the dmsg side.

2814 / 2815 feat: fold the dmsg bridge into the existing RPC port + –via; unify dmsg + skynet bridges so –via skynet:// works end-to-end — the dmsg and skynet bridges collapse into one surface on the existing RPC port. The upshot: --via skynet://<pk> works end to end, the same way --via dmsg://<pk> does. One address scheme per transport, one flag, one mental model — which key, over which transport.

2810 fix(visor): transport-RPC actually works — shared mux + init order + auto-create transport — the foundation that makes the above real: a shared mux, the right initialization order, and auto-creation of the transport when one isn’t already up. 2816 fixes the CLI parsing the :port out of a --via URL (it was hardcoded to the dmsg visor-RPC port).

This is the management-plane expression of the system’s core property: a visor is its public key, and you reach it over dmsg or over the routed transport mesh, interchangeably.

Skywire: Self-Healing Control Connections

2806 fix(visor+hypervisor): self-heal orphan RPC conns on degraded dmsg — when dmsg degrades, RPC connections that were orphaned by it are detected and re-established rather than left dangling. 2807 fix(visor): skynet-preferred (not skynet-first) for hypervisor RPC — a subtle but important distinction: the hypervisor RPC prefers skynet when it’s healthy but doesn’t rigidly demand it first, so a visor without a ready overlay path still connects over dmsg instead of stalling.

Skywire: Transport + Interactivity

2818 fix(transport,dmsg): TCP_NODELAY on all interactive paths — TCP_NODELAY is set on the interactive paths, so a keystroke in a pty session (or any small interactive write) isn’t held by Nagle’s algorithm waiting for more bytes. The difference between a snappy and a laggy remote shell.

2821 fix(transport): remove recursive RLock in Manager.GetTransport — a recursive read-lock in Manager.GetTransport (a latent deadlock/footgun under the right contention) is removed.

Skywire: Metrics + Hypervisor + Release

• 2822 / 2825 fix(metrics): internalize VictoriaMetrics/metrics + gotop — the metrics and gotop dependencies are internalized, silencing a PSI (pressure-stall-information) log line and lint-cleaning the vendored trees.
• 2817 fix(hypervisor): no ghost rows on failed Summary + remove data race — a failed summary no longer leaves a ghost row in the dashboard, and a data race is removed.
• 2819 chore(release): auto-publish the draft after all artifacts are uploaded + 2820 v1.3.57 — the release flow auto-publishes the draft once every artifact is up (the first move toward the no-stuck-drafts release behavior), and v1.3.57 ships.