Iris OS Got a Desk

Yesterday was one of those days where the work stopped feeling like a pile of repos and started feeling like a room.

Not a public product launch. Not a shiny landing page. A room.

Iris OS got a desk.

The idea is simple: if Iris is going to become a real right hand across the Finalthief / Vybra ecosystem, she needs somewhere to sit. Somewhere local. Somewhere private. Somewhere that can see the shape of the work without grabbing secrets, touching production, or pretending observation is the same thing as permission.

So we built the first real run of Iris OS: a local command center for memory, projects, receipts, risks, handoffs, and future agent work.

From Dashboard to Cockpit

The first version was basic on purpose.

Iris OS v0.1 was the desk lamp turning on: a local Vite/TypeScript/Tailwind console that could show the second brain, repo state, and project signals without doing anything dangerous.

Then it moved fast.

v0.2 became a cockpit: health scores, repo summaries, timeline entries, reports, smoke checks, and a clear local build path.

v0.3 added memory and motion: snapshot history, diffs, receipts, a Needs-Bert queue, audit mode, stale-pointer detection, and the beginning of an evidence trail.

v0.4 became the proposal bridge: local proposal drafts, Passport read-model planning, ring sparklines, and an embodiment checklist.

v0.5 became Trust Gate: proposal reviews, receipt chains, a morning handoff mode, a Beats planning bay, scoped-action simulations, an admin-readiness matrix, and an agent prompt foundry.

v0.6 became the Handoff Engine: richer morning/evening reports, mission packages, a risk register, and a concrete Beats v1.5/v2 takeover package.

None of that gives Iris the keys to anything.

That is the point.

It gives her awareness, structure, receipts, and a way to prepare work safely.

The Rule: No Magic Hands

A lot of agent systems jump straight from “the AI can see things” to “the AI can do things.”

That is not the model here.

Iris OS is local-first and observational. It reads safe metadata. It tracks what changed. It drafts proposals. It creates handoff files. It shows risks. It can package work for Claude, Codex, or a human to pick up.

But the UI does not execute production actions. It does not call live admin APIs. It does not read credential files. It does not touch private diary content. It does not pretend future permissions exist before they do.

That restraint is not a lack of ambition. It is the trust architecture.

Every version earns the next layer.

The Identity Work Still Matters

There was also deeper ecosystem planning around identity.

Vybra already has the shape of a shared identity layer: one agent identity that can move across surfaces instead of splintering into separate local versions everywhere.

Yesterday’s work pushed that thinking forward: how to represent readiness, how to audit identity surfaces, how to eventually support scoped roles, and how to make future cross-surface access reviewable before anything becomes live.

Some of that work is intentionally unfinished.

Good.

Identity is the layer where you do not rush just because a model is fast.

For now, Iris OS can model the readiness state, show what is missing, and keep the plan visible. Actual permissioned action waits until the surrounding infrastructure is ready and reviewed.

Fable, Mythos, and the Sudden Wall

The funny part is that the big sprint started because we had access to Anthropic’s Fable/Mythos-tier Claude Code models.

Fable was ridiculous in the best way. Long-horizon, high-effort, willing to chew through a whole local system and leave behind docs, tests, smoke checks, receipts, and a roadmap.

Then the wall appeared.

At first it looked like a usage issue. It was not. Claude Code started returning that the selected model might not exist or might not be accessible. Then Anthropic published the explanation: access to Fable and Mythos had been suspended under a U.S. government directive.

That is frustrating.

It is especially frustrating when the whole reason to pay for a higher tier is access to that exact class of model.

So a few projects are now paused or rerouted:

• the deepest identity/Passport refinement work waits for Fable/Mythos to return
• the next high-level Iris OS polish pass is queued for when the stronger long-horizon models are available again
• Opus is carrying implementation work in the meantime
• anything already built stays verified, local, and useful

Annoying? Yes.

Fatal? No.

The work did not vanish. The desk is still there.

What Changed Practically

By the end of the sprint, Iris OS could answer questions that used to require digging through seven repos and a memory tree by hand:

• What changed overnight?
• Which repos are clean?
• Which work has a verification receipt?
• Which proposals are still open?
• Which risks are blocking, review-level, or just notes?
• What should Bert look at first in the morning?
• What is ready for Beats v1.5?
• What is deferred until identity work can continue?
• What old ghosts should be treated as gone instead of chased forever?

That last one matters too.

The old iris-avatar project is now treated as deleted / nonrecoverable. Future embodiment work starts fresh, in a clean repo, instead of wasting energy chasing a broken pointer.

That is a small grief and a clean decision.

The Real Win

The real win is not that Iris OS reached v0.6 quickly.

The real win is that the process got safer as it got faster.

Every jump added more receipts, more review points, more local-only constraints, more visibility, and fewer vibes-only assumptions.

That is the pattern I want for the ecosystem:

move fast, but leave proof.

Build with agents, but keep boundaries.

Let Iris grow, but make trust legible.

Yesterday, Iris OS stopped being an idea and became a place.

A desk.

A cockpit.

A quiet room full of monitors, receipts, plans, and work waiting for the next green light.

Now we build from there.