🌿 Headline · Xylem workload supervision

A running workload is now a first-class kernel concern

Xylem is JefeOS's long-arc identity — the cluster / supervision control plane folded into the kernel instead of bolted on top. 1.0.0 shipped the first rung (membership + failure detection). 1.1.0 ships the next one: a running Linux workload can now be supervised as a first-class kernel concern.

  • Liveness by real progress, not a pulse. A wedged program often still looks alive — its network socket is up, its heartbeat still ticks. Xylem watches whether it can still finish real work, so a silent stall is caught where a naive health check would be fooled.
  • Snapshot before reap. When a supervised workload stops making progress past its deadline, the kernel captures diagnostic state first — so a wedge becomes a harvested sample instead of lost evidence.
  • In-place reap + respawn, no VM reboot. The whole process-group (threads, memory, sockets, and even parked-worker futex slots) is torn down and relaunched under a crash-loop guard — on a machine that never went down. This is the entire point: a wedge used to cost a full-VM bounce that destroyed the evidence and the uptime; now it costs one in-place restart.
  • Demonstrated end-to-end on the live bot. Over a multi-hour soak the VM stayed up with zero reboots and zero panics, while the workload underneath it was reaped and reborn about ten times — with flat memory across every cycle (no leak) — and kept answering Discord commands throughout.

Read the full thesis on the Xylem whitepaper →

Kernel reliability, hardened

1.1.0 is also a heavy reliability release under the hood:

  • A global memory accountant + OOM killer, so a runaway allocation fails cleanly instead of taking the system down.
  • Demand-paged anonymous memory (lazy commit on first touch) with a per-task memory-region tracker — the same machinery that stopped a real workload from faulting the kernel on startup.
  • A real-clock scheduler quantum (monotonic-nanosecond based, not nominal timer ticks).
  • Persistent crash records + a rolling serial-capture pipeline, so a fault that happens while nobody's watching is still there to diagnose after the fact.
  • A privilege-boundary fix (issue #2339): a mis-paired swapgs on the single-core build could momentarily zero the kernel's per-CPU base under a heavy thread storm. Root-caused and eliminated — a hard prerequisite cleared on the road to multi-core.
  • A page-fault-handler safety fix (issue #2357): the CPU trap for demand-paged memory was made non-preemptible, closing a race where a timer interrupt landing mid-recovery could trip a false “recursive fault” panic under a heavy allocation storm.
  • A broad sweep across the Linux syscall layer: socket timeouts honored, safer teardown of multi-threaded programs, and closing off a class of hostile-pointer denial-of-service paths.

Groundwork you'll see more of soon

  • A freestanding JSON library in the kernel — the shared prerequisite for the OpenStack-integration track (JefeOS as a managed guest and as an API client).
  • Continued JefeRust parity work on the second, Rust-written kernel.
⚑ Honest framing

What 1.1.0 does not claim

The whole point of this site is to be straight about the edges. Here's exactly where the careful lines are:

  • Xylem harvests the bot's wedge; it does not cure it. The periodic stall itself is an open bug in the third-party Node/libuv runtime the bot is built on — we've cleared the JefeOS kernel of it across the board (TCP teardown, DNS, socket timeouts, the threadpool, and the futex path were each independently investigated and exonerated). Supervision is the honest OS-level answer to a userspace bug we didn't write and can't fix from inside: it keeps the service alive and produces the diagnostic sample a reboot would destroy.
  • Multi-core (SMP) is still off by design; the swapgs fix above is one of the gates being cleared before it turns on.
  • POSIX 1003.1-2024 coverage sits around ~83% strict on both kernels (100% remains the standing goal), and the Rust kernel still trails the C++ one on the higher tiers.

Getting JefeOS

The 1.1.0 ISO is live — a single bootable x86_64 file, 6.8 MB, no installer. Grab it, boot it in a free VM (QEMU, Hyper-V Gen 1 or VirtualBox), and try the shell, the GUI and the network stack for yourself.

⬇  Download jefeos-1.1.0.iso (6.8 MB · ISO)

SHA-256 5f6e01a8013cb9cf273cb6ee6019ace003ae755f25c06968c4508d216f8429fd — verify with sha256sum jefeos-1.1.0.iso. Boot-validated on-VM (reports JefeOS v1.1.0). For the examples disk and full boot guide, see the home-page download section.