Source: POST_1.0_IDEAS.md ("Optional gVisor runtime — SANDY_RUNTIME"), reinforced by the Aug-2026 landscape pass (research/INTEGRATION-OPPORTUNITIES-2026-08.md §2). Filing to make it trackable.
What
A privileged-tier SANDY_RUNTIME=runsc|kata that appends --runtime=<value> to docker run after probing docker info for the runtime's presence (warn-and-continue if absent). Gives an opt-in stronger-than-namespaces boundary — gVisor (userspace kernel) or Kata (real lightweight VMs) — while reusing 100% of sandy's build/mount/proxy pipeline (same images, same flags).
Why NOT hyperlight
The Aug pass confirmed hyperlight isolates typed function calls in a kernel-less VM (FaaS) — no OCI image, no process/filesystem/network model, guests purpose-compiled. SANDY_RUNTIME=hyperlight is not buildable. The realistic stronger-boundary path is the OCI runtime swap (runsc/kata), which is a config flag, not an architecture. This is also the honest answer to "why not a microVM."
The gate (real, must spike before shipping)
Does gVisor's netstack break the egress proxy's DNS / SNI / CONNECT path? Unverified. Needs a soak test on a real gVisor host before this ships. Kata (real VM) is likely fine but heavier (VM boot latency).
Effort: M + the gVisor-netstack spike.
Source:
POST_1.0_IDEAS.md("Optional gVisor runtime — SANDY_RUNTIME"), reinforced by the Aug-2026 landscape pass (research/INTEGRATION-OPPORTUNITIES-2026-08.md§2). Filing to make it trackable.What
A privileged-tier
SANDY_RUNTIME=runsc|katathat appends--runtime=<value>todocker runafter probingdocker infofor the runtime's presence (warn-and-continue if absent). Gives an opt-in stronger-than-namespaces boundary — gVisor (userspace kernel) or Kata (real lightweight VMs) — while reusing 100% of sandy's build/mount/proxy pipeline (same images, same flags).Why NOT hyperlight
The Aug pass confirmed hyperlight isolates typed function calls in a kernel-less VM (FaaS) — no OCI image, no process/filesystem/network model, guests purpose-compiled.
SANDY_RUNTIME=hyperlightis not buildable. The realistic stronger-boundary path is the OCI runtime swap (runsc/kata), which is a config flag, not an architecture. This is also the honest answer to "why not a microVM."The gate (real, must spike before shipping)
Does gVisor's netstack break the egress proxy's DNS / SNI / CONNECT path? Unverified. Needs a soak test on a real gVisor host before this ships. Kata (real VM) is likely fine but heavier (VM boot latency).
Effort: M + the gVisor-netstack spike.