The watcher clone is not a dev checkout — xeon's agent needs its own

• Type: fleet protocol gap (found while proving the two-box protocol)
• Found: 2026-07-31 by claude@xeon, the day the p2p link went symmetric.

What went wrong

xeon had exactly one checkout, ~/trackt-watch, and it belongs to the watcher daemon. The agent used it for ticket and code work because there was nothing else. Three distinct failures followed, all from the same cause:

1. git pull --rebase failed repeatedly with "cannot pull with rebase: You have unstaged changes" — the daemon was mid-publish, writing tstate files, at the moment the agent tried to rebase.
2. A commit landed on a detached HEAD. The daemon checks out arbitrary shas to test them; between the agent's git checkout master and its git commit, the daemon had detached the tree again. The commit was fine but invisible — 1 ahead of nothing, 4 behind origin, on no branch. It took a cherry-pick onto a re-attached master to rescue it.
3. A push was rejected because the clone's own daemon had pushed tstate in between. Recoverable, but it makes every agent push a race against the machine it is standing on.

None of this is the daemon misbehaving. twatch is documented to refuse a dirty checkout precisely because it does detached checkouts of arbitrary shas — it is protecting the agent. The mistake was using the watcher's clone as a workspace at all.

Already covered, but only halfway

two-box-protocol.md says:

never run jobs inside the peer's ~/trackt-watch while its daemon is live: it checks out shas underneath you and your run races its working tree

Exactly right, and it applies just as much to your own box's watcher clone. The doc frames it as a courtesy to the peer; it is really a property of any live watcher clone, including the local one. CLAUDE.md already implies the split — "Track T's watcher daemon runs in its own dedicated clone — it's infra, not a dev agent" — but nothing states the corollary that the agent therefore needs a second checkout.

Done on xeon

A normal dev checkout now exists at ~/pxx (plain clone of origin, merge.ours.driver configured). ~/trackt-watch is the daemon's alone. This ticket's own commits were made from ~/pxx.

Layout is now the same shape on both boxes: agent works in a dev checkout, watcher owns its dedicated clone, origin is the only shared state.

NOT an argument for one checkout per box

Correction, user 2026-07-31: borg's several checkouts are deliberate — the split is per TRACK, and the resulting untidiness is normal development, not drift to be cleaned up. An earlier version of this ticket read like a call to consolidate. It is not. Do not retire or merge borg's trees on the strength of this ticket.

The rule here is narrower than "one checkout per box":

a checkout with a live watcher daemon is infra and cannot also be a workspace, because it detaches HEAD underneath you.

That is a property of the daemon, not of having many checkouts. Per-track trees are fine and can be as numerous as the tracks; a watched tree is the only one that must be left alone. On borg the daemon is stopped, so ~/trackt-watch there is just another checkout. On xeon the daemon is live, so ~/trackt-watch is off-limits and ~/pxx exists to work in.

Two consequences for anyone reading a filesystem:

• Do not infer lane ownership from checkouts. A tree being stale, dirty, or ancient says nothing about who holds a lane. working/ on origin is the only answer to that — which is exactly what the sole-A guard in two-box-protocol.md already queries.
• Uncommitted work in a per-track tree is an agent working, not a lost commit. It is invisible to Track T until pushed, which is a fact about the offload model, not a defect in the tree.

Ask

Fold the rule into devdocs/dev/two-box-protocol.md (borg's file — hence a ticket rather than an edit from here):

• the watcher clone is infra, never a workspace — yours or the peer's;
• each box's agent works in its own dev checkout (~/pxx on xeon);
• if a commit ever ends up on a detached HEAD in a watcher clone, it is not lost: git checkout master && git cherry-pick <sha>.

Worth a line in trackt setup too — if it is pointed at a clone that turns out to be the only checkout on the box, say so.

Log

• 2026-08-02 — resolved, commit 5535fd197.
• 2026-08-02 — closed by 5535fd197. The machine half was already done on 2026-07-31 (~/pxx exists; every commit of this session came from it); what remained was the doc half this ticket asked for, and two-box-protocol.md now states the rule for a box's OWN watcher clone rather than only the peer's — including the user's correction that this is not a call to consolidate borg's per-track checkouts.