The model has a job.
Understand the request, ask useful questions, compare options, and propose typed actions. Model output is input to the system, not an instruction the kernel must trust.
Explore how the prompt, local intelligence, system services, and Core kernel fit together—and where authority lives.
Recommended design: a capability-based microkernel with system services in user space. Pioneer supplies the local coding-agent engine behind NIX’s intent broker. The AI proposes and assembles experiences; deterministic services enforce permissions and execute approved plans. Generated code never rewrites the running kernel.
One prompt by default. No desktop, icons, or terminal unless requested. A compositor presents the task interface when needed. Author, Wallet, and Play remain named capabilities accessible through the prompt. Asking for a desktop reveals optional icons; asking for a terminal opens an explicitly requested command environment. Deterministic input and recovery remain available without inference.
Intent understanding, questions, research synthesis, planning, code generation, and local retrieval run on the machine. Models execute in unprivileged processes. CPU inference is the baseline; GPU acceleration requires supported compute drivers.
Convert model output into typed, validated plans. Separate proposal, user decision, and execution. Limit tools, resources, network access, and Memory access per task. Retrieved text has no authority to grant permissions.
Memories, package research, builds, Pod lifecycle, networking, compositor, audio, device discovery, secrets, and signed updates communicate through versioned IPC protocols.
Boot handoff, physical and virtual memory, threads, scheduling, interrupts, IPC, capabilities, clocks, and isolation. Use Rust by default; audit unsafe code and reserve C for required interfaces or selected low-level components.
A narrow, certified hardware profile. UEFI loader, ACPI discovery, PCIe enumeration, IOMMU isolation, USB input, storage, networking, and display drivers form the foundation. Hardware selection precedes broad compatibility promises.
| Stage | System behavior | Reviewable output |
|---|---|---|
| 1. Understand | Infer the task locally. Ask only for missing details that affect the outcome: existing project, title, account, storage, or desired software. | Intent record, scope, constraints |
| 2. Research | Search the local package catalog and installed capabilities first. With network permission, retrieve documentation and package metadata. All model inference remains local. | Sources, licenses, compatibility, maintenance status |
| 3. Choose | Prefer a vetted existing application. Compare native, runtime, and guest paths. Generate a new tool only when the existing options do not meet the task. | Reuse / adapt / build decision and dependencies |
| 4. Build & verify | Compile in a restricted build Pod with pinned dependencies. Run functional checks and permission review. Separate compiler/download access from personal Memories and secrets. | Artifact hash, build recipe, checks, permission manifest |
| 5. Deploy | Install the approved artifact atomically into an isolated Pod. Register its task name for prompt access. Show its task interface on request; add an icon only when the user enables a desktop. Preserve a rollback point. | Author, Wallet, or Play available by intent |
| 6. Continue | Offer context-aware suggestions within the active task. Ask before expanding access or installing dependencies. Let the user mute suggestions. | Task history and versioned Memories |
Understand the request, ask useful questions, compare options, and propose typed actions. Model output is input to the system, not an instruction the kernel must trust.
A trusted service checks scope, permissions, artifact identity, and task state before performing an operation. Cancelling a task must prevent late execution.
The kernel provides privileged memory, scheduling, interrupt, IPC, and isolation mechanisms. Updating this foundation uses a separate build and activation path.
Ship the local model weights and inference runtime as required system components. The base image includes Core, essential drivers, services, prompt shell, compositor, recovery tools, and AI. Consumer applications are installed on demand.
Use a local model registry with hashes, versioned prompts, context limits, and a resource governor. Preempt or pause heavy inference during gaming and XR. Model weight size, RAM/VRAM, disk requirements, and latency targets must be measured on the selected hardware before a minimum specification is published.
NIX can ask “Do you have an existing manuscript?” or suggest a chapter outline. It can surface a missing dependency and explain the proposed install. Suggestions stay tied to the current intent and can be dismissed or disabled.
Retain deterministic launch, save, close, recovery, and settings paths if the model fails. Kernel updates follow a signed build/review pipeline and reboot activation; the AI does not hot-patch the trusted foundation.
The layers and interfaces above are proposed NIX design decisions. These primary sources ground the low-level constraints; they do not establish that NIX has implemented them.
Rust: the no_std environment ↗ — freestanding Rust foundations.
UEFI: architecture overview ↗ — firmware handoff and driver lifecycle.
QEMU: system emulation and acceleration ↗ — guest machine models and accelerator boundaries.