Availability and platform support
What is implemented, what has been exercised and which capabilities are still planned.
Choose a qualified compute backend
| Platform | Current scope |
|---|---|
| Linux with Incus/KVM | VM compute, native helper, templates, warm reserves and workspace lifecycle. Live qualification currently uses x86_64. |
| macOS | Operator/controller target; archive staging and supervised service updates exercised. Manage separate Linux compute nodes. |
| Windows | Native operator/controller build target. Windows service installation and managed provider sign-in require further execution qualification. |
| Docker laptop contribution | Planned. Installing an operator does not add Docker Desktop or personal-machine capacity to the fleet. |
| Windows/macOS native compute and GPUs | Planned backends. GPU detection is not qualification for GPU workloads or saved GPU state. |
Keep explicit CPU, memory and disk limits. Frozen VMs retain memory commitments; stopped VMs still consume disk. Fleet capacity does not account for every unrelated host process.
Feature availability
| Capability | Status |
|---|---|
| Official Codex, Claude Code, Cursor Agent | Native integration and account delegation implemented. Rust Codex/Claude coding, exact resume and browser MCP use passed; Cursor remains provider-quota blocked. |
| Private native homes | Persist in each workspace. Cross-workspace memory synchronization and movable tenant volumes are planned. |
| Templates and warm pools | Controller registry, node synchronization and never-reused fresh reserves implemented. |
| LayerFS checkout branches | Controller/API/MCP/dashboard implemented on explicitly provisioned, qualified Linux nodes. General installation does not provision it automatically. |
| LayerFS Incus attachment | Explicit administrator operation; qualified guest file operations and clean lifecycle. No automatic allocation attachment or coordinated RAM hibernation. |
| Persistent project core and data forks | Managed main tracking, optional dev command, sleep/wake and online publication planned. A local stopped-state Wrangler fork prototype exists. |
| Public multi-tenant service | Not implemented. Current authorization assumes one trusted controller owner. |
Storage qualification is specific
The LayerFS preview requires a local Linux store on the supported filesystem/kernel topology and a compatible local Docker runtime. The qualified integration uses Docker API 1.54. It is not a cross-host store or a Docker Desktop filesystem qualification.
Power-loss durability, bounded history retention, large Unity/Unreal builds, native-provider sessions on the mounted checkout and live Wrangler/SQLite consistency still need qualification. Keep authoritative project state and irreplaceable data on separately qualified persistent storage.
File checkpoint and VM RAM restoration are different operations. Do not combine them without a consistency protocol. See LayerFS scope and remaining gates.
Read timings as measurements of a specific path
| Measured path | Result |
|---|---|
| Rust resume from paused VM | 0.512 s median, 0.541 s p95; 20 samples |
| Rust start from stopped VM | 6.034 s median, 7.096 s p95; 10 samples |
| Rust restore from hibernation | 5.846 s median, 8.053 s p95; 5 samples |
These are readiness timings on the recorded Linux host with a 2-vCPU, 2-GiB VM. The stopped path reused a prepared disk and host caches; it was not a new-machine install. The measurements did not use a LayerFS-attached VM. They do not establish a controlled speedup over Python or a universal subsecond boot promise.
The LayerFS backend created fresh Docker workspaces in 1.437 seconds and a workspace from a prepared checkpoint in 1.549 seconds in its small-fixture qualification. These were individual observations with prepared images and caches. A cheap branch reference is not a booted VM or a completed application start. See VM timing evidence and storage measurements.
Before a wider release
The reference fleet completed its initial Rust cutover on October 20, 2026. Separate disposable tests proved Codex/Claude coding, exact conversation resume and browser MCP use; Cursor remains provider-quota blocked. A fresh revision-4 warm allocation then passed real Codex/Claude code assertions through the installed operator MCP and HTTP controller, followed by exact conversation resume after MCP pause/resume with boot identity preserved. Cursor was not retried. Inference on LayerFS attachments, a fresh physical-host install, Windows service execution and an independent installation reproduction remain acceptance work.
Cloudflare environment publication is new product work, not a missing checkbox on the Rust migration. Its acceptance test must exercise concurrent writes, interruption during capture, complete binding state, two isolated forks and restart at representative sizes.
Use the validation record for evidence and the implementation plan for delivery order.