LayerFS for Allternit
Answer me with artifact · 2026-10-26 18:22

New version 0.5.10

You have 0.5.5. Found when this page was made.

What's new:

Run in a terminal:

npx skills update answer-me-with-html -y

Or tell your agent:

Update answer-me-with-html
See what's new

LayerFS for Allternit

What Ephemeral AI Lab's agent filesystem could do for Factory and cloud computers

sourcegithub.com/Ephemeral-AI-Lab/layerfs (v1.1.0, MIT)

Lead: LayerFS could replace git worktrees as the workspace each Factory agent gets. It is worth researching. It is not safe to adopt yet: it has no crash durability and reaches macOS only through DeepSeek Harness.

A

What it is

  • Each agent mounts a full writable copy of the project, including node_modules, target/ and .git, without copying files.
  • Every tool call becomes a Commit on a Branch. Agents fork, rewind, merge and publish to main.
  • One SQLite file is the Store. Content is chunked and stored once, so 157 snapshots take 100 MB, 19.9× less than Btrfs + zstd.
  • Two Commits that change the same file get Conflict, unless a .lww rule lets the newest edit win.
B

Where it helps Allternit

Pain today What LayerFS changes Fit
~127 worktrees filled the disk on 2026-09-30; cleanups have freed up to 170 GB each time Workspaces share one deduplicated Store, and unmount deletes the overlay strong
Each new worktree reinstalls pnpm deps and rebuilds the Rust target Dependencies and build output come with every mount strong
Parallel Factory agents step on each other or need manual merges One Branch per agent, with conflicts refused per file strong
No rewind of a bad agent step Every tool call is a checkpoint to go back to good
Cloud computers and the Mail VPS keep their own copies of repos Agents on different machines can share one Store needs network-share testing
Teach Replay and Agent watch show what an agent did The Commit history is a step-by-step record to show idea only
C

Blockers

callout warn Not ready for production

  • No crash or power-loss durability. Use it only for agent scratch work. Git and GitHub stay the source of truth.
  • macOS only through DSH. Linux is native, and Windows works through WSL2. Factory runs on the Mac, so the Mac path must be tested first.
  • Slower sequential I/O: 0.31× on sequential write, which affects big cargo builds.
  • 1.1.0 broke the 1.0.0 API. Expect churn.
D

Suggested path

  1. Put the link in the research queue for a baseline against Factory's worktree model.
  2. Try it on the Linux cloud computers first, where it runs natively. Run one Factory DAG with LayerFS workspaces.
  3. Compare disk use, mount time and merge conflicts with the worktree run.
E

Next

What do you want to do with it?