Torsionfield / Linux

An experimental AI metadistribution

Your next Linux installer is a conversation.

Boot into your browser ChatGPT. Connect your personal MCP tools. Give intelligence a system it can reconstruct, explain and evolve.

Research prototype · Inventory tooling exists. Installer adapters and live USB are the next build.

01 / The approach

Start with a distro. Make its construction understandable.

The ambition is to recreate any Linux distribution through the ChatGPT route, then keep the knowledge of how it was built.

  1. Account for the system you already have.

    Load package manifests together: one package explains many files. For everything else, identify a repository, binary archive, build recipe, regeneration rule or preserved personal data. Record exact versions and unavailable inputs.

  2. Bring intelligence into the installer itself.

    Take the open-source installer, adapt or rewrite its orchestration, and expose each step to tools and the system record. ChatGPT participates from the first operation, rather than trying to infer every installation decision afterwards.

  3. Rebuild, observe, verify.

    Use explicit target identities and preconditions. Check the actual result after each operation. A successful rebuild establishes reproducibility; an inventory alone is the starting evidence.

02 / The live environment

A recovery drive that opens your way back in.

When the installed desktop fails, the installer should still be usable.

Full Chrome. Your ChatGPT subscription.

The proposed live USB starts an independent Linux desktop with the full Chrome browser and Proton Pass. You sign into your own browser ChatGPT account; this workflow is designed around that conversation.

All of your local MCP services.

Your browser automation server and browser, Desktop Commander, VNC and every other required personal server start with their runtimes, dependencies and tunnels. The tools must reach the live system and the intended installation target.

A personal environment you can test.

Keep account sessions and tunnel credentials separate from a public image. Verify a real boot, network access and each tool from ChatGPT. Account MCP support and authentication must be checked; offline terminal recovery remains available.

03 / The system record

The conversation can change. The evidence stays.

SQLite gives frontier intelligence persistent, queryable context about this machine. It separates what exists, what you want, and what actually happened.

Where does this file come from?

Package ownership leads to a versioned source. Residual files lead to a recipe or remain explicitly unresolved.

SELECT path, package_name, package_version,
       recipe_id, provenance_status
FROM artifacts
WHERE system_id = 'example-machine';
What should this machine become?

Desired state is distinct from an observation. An intention does not establish that a configuration was applied.

SELECT subject, state_json, reason
FROM desired_state
WHERE system_id = 'example-machine';
What ran, and how was it checked?

Record operations and attach result evidence. Interrupted operations require reconciliation before being treated as complete.

SELECT a.id, a.status, e.event, e.evidence_uri
FROM actions AS a
JOIN action_events AS e ON e.action_id = a.id
WHERE a.system_id = 'example-machine';

Illustrative queries for the draft schema. SQLite transactions cover database writes; package installation and disk changes require their own journal, recovery strategy and verification.

04 / Beyond one distro

An Ubuntu foundation. A wider Linux vocabulary.

The metadistro is the orchestration layer: a common way to reason about systems whose underlying distributions differ.

Ubuntu

A known foundation

Begin with native packages, source versions and the working kernel. Preserve what makes the machine boot and let the record explain its configuration.

Arch

More explicit construction

Bring selected ideas about deliberate system assembly and configuration into the orchestration layer. Use a distro adapter for Arch targets and compatible implementation choices on Ubuntu.

Gentoo

Source-level choices

Explore pinned source builds and recorded build options, with dependency and ABI checks. Isolate builds when needed and retain a way back to the working system.

Future adapter work. Cross-distribution composition requires compatibility checks, package ownership, update policy and tested recovery. Arbitrary package mixing does not supply those guarantees.

06 / The extended concept

Compose the experience. Account for its whole lifecycle.

Start with a usable profile. Let ChatGPT adapt its construction, and keep the evidence needed to rebuild or recover it.

  1. A profile and a locked component manifest.

    Describe the desktop, tools, services and build choices independently of the target distro. Pin source revisions, artifact digests and configuration recipes. Distro adapters resolve compatible native packages or isolated builds; SQLite records availability, deviations and observed results.

  2. Explicit roles and ownership.

    Define rescue USB, workstation, headless or VM, and connected-tool-device profiles. Browser ChatGPT remains the intended primary interface. Give each managed component one update owner and each target one active orchestrator; record dependencies, permissions and endpoint identity. A VM reaches a host only through explicitly connected tools.

  3. Recovery units with their own proof.

    Track the base OS, personal files, agent/MCP state and orchestration database separately. Keep credentials in a separate provisioning process. Check boot, data consistency, authenticated tools and the target identity after recovery. A restored snapshot or a completed database transaction alone cannot establish that the whole environment works.

Design work, not implemented integration. The first demonstration should reproduce one pinned profile in a disposable VM, interrupt and recover an operation, then verify every required MCP from the independent browser environment.

07 / Why build it?

Less lost knowledge between installation and repair.

Recovery with context

An independent boot environment can inspect the failed installation while retaining its intended configuration and known inputs.

Recreation you can examine

Versioned sources and recipes make a rebuild explainable. Missing inputs become visible work instead of hidden assumptions.

Continuity across models

A new conversation or model can query the same observations, decisions and evidence instead of depending on one chat's memory.

Controlled experimentation

Linux power users, rescue-tool builders and system researchers can test changes against explicit goals and recorded results. Repeated installation and recovery trials must establish the practical value.

08 / Project status

The foundations exist. The installer comes next.

Implemented locally

Inventory and recovery workflow

A working Ubuntu collector records installed packages, ownership and root filesystem observations in SQLite. A reusable skill describes the recovery and personal MCP workflow. Residual provenance still needs investigation.

Published here

Architecture and draft schema

This project brief and executable SQL schema define the proposed orchestration record. The current inventory database uses a separate schema; an import adapter is future work.

Next build

One observed installation

Define one pinned profile and component manifest, adapt an upstream installer, journal every step and verify a complete disposable-VM installation. Test interruption and recovery. Preserve upstream licensing and retain the native package manager's responsibilities.

Then

Personal live USB and distro adapters

Prove boot, browser sign-in, every personal MCP server and target recovery. Then test selected cross-distribution capabilities. A boot-tested ISO and universal installer are not available yet.

09 / Questions

Practical questions.

Does this require switching to Omarchy or running LINA?

The proposed orchestration layer can target different distributions. Omarchy is a potential curated target; LINA OS contributes composition and recovery ideas. Our intended entry point remains the owner's browser ChatGPT and personal MCPs. Neither an Omarchy adapter nor a LINA integration has been implemented.

Does the USB still boot Linux?

Yes. A real live Linux system provides the kernel, network, display and local tools. Browser ChatGPT becomes the intended primary interface for installing and repairing the target OS.

Can every distro really be recreated?

Every distro is a potential adapter target. A specific system is reconstructible when its required packages, sources, recipes and preserved data are available and the rebuild is verified. The project aims to make that process explicit across distributions.

Does any ChatGPT subscription support this?

MCP and developer-mode availability depend on the account and current product support. Check the actual account before building a personal image. The prototype does not promise that every subscription provides the same tools.

Is the SQLite database the operating system?

It is the structured system record used by the orchestrator. The kernel, package manager and service manager still do their jobs. The model reads and updates records through tools; database state must be reconciled with the real machine.

Where does installer adaptation begin?

Ubuntu's Subiquity autoinstall interfaces and Arch's archinstall library provide concrete starting points. The project proposes adapting their orchestration, retaining applicable licenses and adding evidence for each operation. No rewritten installer has shipped yet.