Skip to content

vcs-collapse.adoc

Latest commit

 

History

History
68 lines (54 loc) · 3.07 KB

vcs-collapse.adoc

File metadata and controls

68 lines (54 loc) · 3.07 KB
Raw

VCS collapse over editions

Table of Contents

When the filesystem natively supports editions, version-control tools stop owning working trees, object stores that duplicate file bytes, and destructive checkouts. They shrink to a graph orchestrator plus semantic merge.

Two phases

Phase A — userspace bridge (today → near term)

Until editions exist in the VFS, concurrency still needs a layered userspace stack. These tiers are complementary, not competing products:

  1. Process-scoped mount / path virtualization — each IDE, agent, or build runner sees a stable logical root (e.g. /workspace) mapped to its revision, without changing absolute paths apps hard-code.

  2. Working-copy-as-revision graph (operation-log engines in the spirit of Jujutsu) — uncommitted edits are anonymous revisions; no fragile single index; conflicts can be first-class graph state.

  3. Copy-on-write storage and content-addressable caches — reflinks / OverlayFS / APFS clones plus shared dependency stores so concurrent workspaces are cheap.

  4. Headless agent sandboxes — ephemeral namespaces attached to distinct heads; success reparents; failure drops without touching the human workspace.

Git worktrees address only “more than one checkout.” GitButler-style virtual branches address hunk assignment in one working directory. Neither fully solves path stability, agent isolation, and cheap multi-workspace storage together. Phase A is the honest interim architecture.

Phase B — filesystem editions

With editions:

  • Creating a “branch” requests a new edition id on a directory.

  • “Commit” / tag marks a point-in-time edition state in the connectome graph.

  • Rebase- and cherry-pick-like remapping lean on CoW extent operations rather than userspace read/diff/write pipelines over a single live tree.

  • The VCS no longer needs a shadow copy of every blob in .git/objects for the working tree itself (history metadata and remote sync remain product concerns).

What must stay above the filesystem

Filesystems operate on extents, blocks, and bytes. They do not understand language ASTs.

  • Semantic three-way merge — conflict presentation can surface conflicting editions; resolving code meaning stays in merge engines.

  • Non-filesystem state leaks — daemons, sockets, global lockfiles, /tmp, and shared compiler services bypass folder isolation. Process boundaries remain necessary regardless of how rich the tree model becomes.