Issues #39886 + #47548 — still open · Claude Code 2.x
Worktree Guard — Early Access

Stop losing work to Claude Code worktree collisions.

isolation: "worktree" in Claude Code settings does not actually isolate branches. It silently falls back to the main working directory. Two agents, one directory, second write wins — prior agent work erased without error or warning. GitHub issues #39886 open and #47548 open — unresolved. No tooling exists. Worktree Guard is a drop-in shell bundle that prevents the collision before it happens.

GitHub #39886 — OPEN GitHub #47548 — OPEN
Confirmed failure mode — community + primary sources
0
error messages when the silent fallback happens
GitHub #39886 · confirmed behavior
2+
open issues documenting branch collision, unresolved in Claude Code 2.x
GitHub #39886 + #47548
$0
tooling exists today — only manual workarounds and blog posts
Community search 2026-05-02
$29
one-time, prevents the first lost-morning incident before it happens
Septim Labs — ships in 14 days
What actually happens — the silent failure sequence
01
You set isolation: "worktree" in .claude/settings.json and start two parallel subagents on the same repo. Looks correct. Claude Code accepts it without warning.
02
The setting silently falls back to the main working directory. Both agents are now writing to the same file tree with no coordination. No error is thrown. No log entry. The agents themselves are unaware.
03
Agent A writes src/api/auth.ts. Agent B, running a different task in the same session, also writes src/api/auth.ts. Second write wins. Agent A's work is gone.
04
You discover the collision at commit time — or worse, at deploy. Git history shows two conflicting writes with no branch separation. You spend the next two hours reconstructing Agent A's work from context.

What ships

Four tools. Zero server. Zero ops.

A shell script bundle and a browser-side config generator. POSIX-compatible bash. Runs locally. No telemetry, no phone-home, no SaaS account. You pay once, you keep it.

Script 01
wg-init

Generate a safe multi-agent config

Reads your existing CLAUDE.md or .claude/settings.json, detects subagent definitions, and outputs a verified per-agent worktree configuration with named branches and explicit isolation paths. One command, correct setup.

Script 02
wg-check

Pre-session collision audit

Run before any multi-agent session. Verifies each defined subagent has its own worktree path, confirms no two agents share a working directory, prints a pass/fail matrix. The free script — open-source MIT. The full bundle is the upsell.

Script 03
wg-clean

Safe session teardown

Tears down worktrees at session end, merges completed work to a staging branch, removes stale locks. Conservative by default — stages only, never auto-merges to main. Prevents worktree accumulation across long sessions.

Tool 04
config generator

Browser-side setup, no install

Paste your agent names and repo path. The generator outputs a valid .claude/settings.json worktree block, tested against the known-bad config format that triggers the silent fallback. Zero install needed to verify your current setup.

wg-check ships free and open-source — the detection script so any developer can verify their isolation status. Full bundle (init + clean + config generator) at $29 pay-once.


Common questions

Before you sign up.

Does wg-check send anything off my machine?

No. The detection script reads your local .claude/settings.json and your git working tree — nothing leaves your machine. No network calls, no analytics, no license server. The conflict log that wg-check writes goes to .wg/session.log on disk, local and append-only. You can read the full source on GitHub before you spend a dollar.

What if Anthropic fixes the underlying bug before you ship?

The waitlist is free — you pay nothing until the build is live and you choose to buy. If GitHub issues #39886 and #47548 are resolved with a native fix before the studio ships, the waitlist email will say so plainly and point you at the Anthropic patch instead. No Stripe link, no pressure.

Can I run this on a client repo?

Yes. The $29 license covers personal and client use — no per-seat fees, no "commercial license" upsell. If you are an agency developer running multi-agent Claude Code setups for clients, one purchase covers every repo you manage. The scripts are bash, POSIX-compatible, and leave no persistent process behind on the client machine.


Early access

Ships in 14 days. $29 pay-once.

No subscription. No account required. Pay once when it ships, use it indefinitely. The waitlist is email-only — the studio notifies you when the build is live, the free wg-check script is on GitHub, and the Stripe link for the full bundle is open.

$29 pay-once for full bundle  ·  wg-check free on GitHub  ·  ETA <14 days  ·  no server, no SaaS
by Septim Labs  ·  pay-once tools  ·  no subscription