Arivaran / AI Developer Toolkit

AI worktrees without repo sprawl

Give every coding agent an isolated branch workspace while unchanged files stay shared for disk and page-cache efficiency.

View release status Explore developer resources

Unavailable

Public activation is not available yet.

Public activation is unavailable. No install command is published until the signed distribution, deterministic verification, and safe rollback contract passes.

One mechanism, then the measured proof.

Release-gated

Public activation is unavailable. No install command is published until the signed distribution, deterministic verification, and safe rollback contract passes.

Step 1: Install the Linux development build
On Linux x86_64 with dnf, install the signed development package. It does not enable break-on-write safety.
curl -fsSL https://get.arivaran.ai/install.sh | sudo sh -s -- aagit
Step 2: MOWI mechanism
MOWI keeps unchanged paths shared and materializes a private inode before a write.
Step 3: Measured proof
All numbers in this table come from #10862: incremental disk is the df delta, the source tree was warmed before cache measurement, and aagit linked 16,195 of 16,197 files with 3 checkout fallbacks.

No success sequence is running.

Review the blocked release gates below. No install or rollback command is published.

Arivaran dogfood

We learned the copy tax the hard way.

Isolation solved the shared-checkout race. It also exposed the usual cost: conventional worktrees duplicate unchanged files and their file-backed cache.

“One honest note carried into memory for next time: I merged the P0 batch with agents but forgot isolation:"worktree", so they raced in the shared worktree (no damage — main is correct and my state intact, but it's the trap my own memory already flags; I'll pass isolation on parallel coding agents going forward).”

Internal Arivaran engineering-session note, July 20, 2026

The safe choice had a cost.

Plain Git worktrees isolate each agent, but each checkout materializes another copy of unchanged files.

Keep the isolation. Lose the copy tax.

Arivaran Git keeps separate worktrees and shares unchanged files. A write gets a private inode before it can change another branch.

Give every coding agent a private worktree.

Put this rule in Claude, Codex, or another coding agent's memory. Isolation is a correctness rule, not a performance setting.

When I delegate coding work to parallel agents:
- Require every coding agent to use isolation: "worktree".
- Never let two coding agents share a checkout, Git index, or uncommitted working tree.
- When aagit is available, use it for the worktree lifecycle. It shares unchanged files while each worktree keeps writes private.

The documented incident was contained with main intact. If agents have already raced in one checkout, stop before merging and rebuild the work in isolated worktrees. Recovering a trustworthy state can mean rolling back both agents' work.

See the measured trade-offs.

The benchmark below reports the actual trade-off: lower incremental disk and shared page cache, with wall time shown by filesystem. Agent runtimes still use their own memory.

One mechanism, then the measured proof.

MOWI keeps unchanged paths shared and materializes a private inode before a write.

MOWI mechanism

Placeholders first
Clone and worktree creation lay a correct directory shape immediately instead of copying every tracked byte before an agent can start.
Hydrate on access
Reads fetch byte-correct content from the git object store while background convergence links unchanged files back to the shared copy.
Break on write
Write intent moves that path to a private inode before mutation, so one agent can edit without changing another agent's worktree.

Measured proof

Open benchmark methodology

Arivaran Git worktree benchmark table
Filesystem	CoW class	Plain git	Arivaran Git	Page cache after warming source
XFS on Linux	Reflink-capable filesystem	270 MB2.81 s wall time	6.8 MB3.27 s wall time	255 MB already hot through shared inodes
ext4 on Linux	Non-CoW filesystem	274 MB1.96 s wall time	18.6 MB1.79 s wall time	255 MB already hot through shared inodes

All numbers in this table come from #10862: incremental disk is the df delta, the source tree was warmed before cache measurement, and aagit linked 16,195 of 16,197 files with 3 checkout fallbacks.

Fit, platform scope, and release gates.

Linux is the intended public release target. macOS and Windows remain outside this activation contract.

Platform scope

Linux
Release-gated
ext4 non-CoW, XFS reflink-capable
Source-build path with fanotify break-on-write and the measured #10862 benchmark surface; the public release gates still apply.
macOS
Not in public activation
APFS CoW
APFS copy-on-write engine work exists, with deeper EndpointSecurity coverage gated separately; no public activation is contracted.
Windows
Not in public activation
NTFS non-CoW
User-mode write-coverage work exists, with an enterprise minifilter track behind signing gates; no public activation is contracted.

Release gates

Public activation is unavailable. No install command is published until the signed distribution, deterministic verification, and safe rollback contract passes.

Public, immutable release channelBlocked
Bootstrap authenticated before elevationBlocked
Complete Linux safety bundleBlocked
Deterministic activation verificationBlocked
Interrupted-upgrade and rollback coverageBlocked

Documentation and animations use the same contract and cannot publish a command early.

AI worktrees without repo sprawl

One mechanism, then the measured proof.

We learned the copy tax the hard way.

The safe choice had a cost.

Keep the isolation. Lose the copy tax.

One mechanism, then the measured proof.

MOWI mechanism

Placeholders first

Hydrate on access

Break on write

Measured proof

Fit, platform scope, and release gates.

Platform scope

Linux

macOS

Windows

Release gates