Two Claudes at once - git is the refereeCapability

You routinely run more than one Claude at once — today two sessions logged a completed job in the same second. Parallel sessions share one working tree, and git is the only referee keeping them from clobbering each other.

A Claude Code session doesn't own your files — it edits the one working tree on disk. Two sessions pointed at the same folder are two cooks in one kitchen: if both touch CLAUDE.md, whoever saves last wins and the other's edit is silently gone. Git is what makes this survivable, because every commit is a snapshot you can diff and recover — but only if each change gets committed before the next session overwrites it.

This is the actual reason CLAUDE.md's commit rule splits by repo. In a dedicated project repo (PDB, JobHunt, an app repo) a session commits and pushes as it finishes — small, frequent snapshots are exactly what let parallel work coexist. In the home repo (/Users/tom) the rule is the opposite: always ask first, and stage explicit paths, never `git add -A`. Two reasons stack there — the home tree holds secrets (.claude.json carries tokens) and it's the tree two sessions are most likely to be fighting over, so a blind bulk commit could snapshot half of another session's work plus a credential.

The clean fix, when you genuinely want two sessions running hard, is to stop sharing the tree at all: a git worktree (Thursday's lesson) gives each session its own checkout on its own branch, same history underneath, zero file collisions. Short of that, the discipline is: commit small, commit often, and never let a session run `git add -A` in a tree another session is editing.

Worked example

See exactly what a second session could clobber right now — read-only, safe.

git -C /Users/tom status --short
git -C /Users/tom/Claude/PDB status --short
tail -2 "/Users/tom/Claude/PDB/data/jarbus-events/$(date +%F).jsonl"

• The first two show uncommitted edits in the home repo and the PDB repo — every line is a file a concurrent session could overwrite before it's snapshotted.
• The tail shows the live event tape: if two lines share a timestamp, two sessions finished work at the same instant — proof this isn't hypothetical.

Gotchas

• `git add -A` in the home repo is the trap the rule exists to prevent — it can sweep in a token-bearing config or another session's half-done edit in one commit.
• 'It's just one file' is how clobbers happen — two sessions each editing 'just' CLAUDE.md still lose one copy; the tree doesn't merge, it overwrites.
• A second session that pulls or checks out under the first can yank files mid-edit; if you're running two hard, give them separate worktrees instead of hoping they miss each other.