v1.3.0

GLaDOS v1.3.0

A from-scratch, ring-0 UEFI operating system in Rust, built around a language model that lives inside the kernel — no user/kernel split, no syscalls, one address space. A tool call from the model is a function call.

This is the largest stable release so far: multi-agent workflows as a real feature, a complete visual rework, and two triple-fault bugs found by fuzzing the machine's own userland and fixed.

---

Added

Multi-agent workflows (work)

A manager delegating to workers, which is a different relationship from the two multi-agent systems already in the tree (the routing council and godel's judges, where disagreement is the product).

• A manager plans once; the workers then decode nothing. agent::prompt_for re-encodes a growing prompt each step, so an episode spends O(N²) tokens over a run. work plan encodes once and writes each decided action down, so N actions cost one prefill and N decodes — and a worker handed a pre-decided action decodes nothing at all. Measured on a four-step goal with Qwen3-0.6B: planning spent 4 decodes, running the plan spent 0.
• Memory is the namespace, not a conversation. ctx_save on a 512-slot cache measures in thousands of store blocks, so a context per worker costs more than the work. The namespace is already a content-addressed Merkle tree, so a run has an address and two runs of one plan can be compared exactly, for free.
• Autonomy takes a declaration and a grant. A workflow declares it may run unattended; that grants nothing on its own, because a plan is a file and a file can edit itself. The operator grants it by typing back part of the plan's address. The grant names the plan without its statuses, so taking a step doesn't revoke the grant that allowed it.
• The honest half: role adapters were built, trained and judged — and they don't beat the base, for a structural reason (a harvested label is the base model's own argmax). That negative result ships documented rather than hidden.

Frutiger Aero visual rework

The desktop went from Windows-98/2000 flat-and-bevel to a full Frutiger Aero look, in the project's own Aperture orange rather than aqua.

• Four new span-based primitives (vgrad, tint, glass, shade) — a gradient interface costs about what a flat one did, because a vertical ramp is one memset per row.
• Warm for the machine, cool for the room: captions, selection and focus stay amber; wallpaper, taskbar glass and fields are blue-green. The wall is a sunset horizon where they meet.
• The aperture logo is now the sun on that horizon — eight solid blades with the sky showing through the gaps, over an atompunk atom-orbit graticule with glass bubbles.
• Rounded window tops, gloss, drop shadows (the back-to-front repaint makes them nearly free), a glass taskbar, and white XP-style menus.

---

Fixed — security (found by fuzzing the userland)

• Aiksi parser could triple-fault the machine. A few hundred nested brackets overflowed the recursive-descent parser; with no guard page in ring 0 that's an instant silent reboot. Reachable past the operator: the parser runs on any Aiksi source before any capability sandbox, so a skill shared by a stranger rebooted the machine the moment it was run. Two distinct overflows are now bounded — deep nesting (parser stack) and a flat 1+1+1+… chain (a tall AST that faults later in eval or drop). The boot selftest exercises all five attack shapes at depth 400.
• Workflow run name could path-traverse and corrupt the namespace. work new ../agent/policy wrote through an existing blob and turned it into a directory, destroying a file godel reads. Run and role names are now validated as single path components.

Fixed — correctness, and a few measurements that were wrong

• -smp was quietly costing the renderer ~30%. The multi-core default (added for the diag mt/migrate suites) was benchmarked on the decode path, where it's free; on the graphics path — span fills and a memcmp, i.e. memory bandwidth — it is not. video bench now runs at -smp 1.
• A 1-pixel bevel highlight made Minesweeper read as thin dark boxes. The warmer XP face left the soft highlight indistinguishable from the surface (5–6–10 per channel vs white's 19–22–39); a screenshot caught it and it's white again.
• A bare diag lists suites and runs nothing, printing 0 passed, 0 failed — which reads like a clean sweep and is the opposite. diag all is the verb. Suite count is now 29.
• The SHA256SUMS download link resolves for the first time — this is the first release with the checksums job.
• Per-core storage is now armed even on single-core / no-ACPI boots (it wasn't, which silently disabled fault recovery and allocation billing there).

---

Install media

Three bootable ISOs (El Torito), pick one:

ISO	Model	Context	Size
glados-1.3.0-qwen3-0.6b.iso	Qwen3-0.6B, int8	512	~604 MB
glados-1.3.0-qwen35-2b-8k.iso	Qwen3.5-2B hybrid (Apache-2.0)	8k	~1.90 GB
glados-1.3.0-qwen35-2b-32k.iso	Qwen3.5-2B hybrid (Apache-2.0)	32k	~1.90 GB

Write to a USB stick and boot it (hold F11 for the boot menu). A live ISO cannot update itself; ISOs are for a fresh install. The .efi + .sig + manifest are for the in-field signed updater. Verify downloads against SHA256SUMS.

Hardware note: this is built for one specific laptop (MSI Thin GF63 12UC). It boots under QEMU for development, but the wireless, ACPI battery and RTL8168 paths are written against that machine.

---

What's next

• Role adapters need a teacher that isn't themselves — a labelled outcome from teach, a judge, or a larger model — since a transcript's label is what the base already answered.
• The XPU (the compute pool that unifies the CPU, the parked application cores and the discrete GPU, formerly "QuantumGPU") lands on the experimental channel, alongside the miner. Experimental builds are a token-gated tier.

diag all 29/29, boot clean on both checkpoints.