[BUG] Worktree isolation clobbers the remote-tracking ref

Preflight Checklist

• [x] I have searched existing issues and this hasn't been reported yet
• [x] This is a single bug report (please file separate reports for different bugs)
• [x] I am using the latest version of Claude Code

What's Wrong?

Agent worktree isolation cannot base on local branches. Workaround forces clobbering refs/remotes/origin/<default>, destroying the owner's pushed-state signal.

What Should Happen?

When multiple agents work on one box, they should base on each other's landed local work. What gets pushed to the server is the owner's decision alone, and refs/remotes/* must remain truthful so the owner can see what was pushed.

Requested fix (any of, in preference order)

1. Let worktree.baseRef accept an arbitrary local ref

(e.g. "main" / "refs/heads/main"), or add a
"default-branch" mode meaning the LOCAL default branch.

1. Add a per-call base override to the Agent tool's isolation option

(e.g. isolation: {mode: "worktree", base: "main"}).

1. At minimum, document that fresh is unsuitable for offline/

sandboxed multi-agent boxes and that the update-ref workaround
destroys the remote-tracking signal.

Error Messages/Logs

Steps to Reproduce

Setup

• Headless/background Claude Code session orchestrating subagents via

the Agent tool with isolation: "worktree", all on one shared dev
box (multi-day, multi-agent project).

• The repository's remote is pushed by the OWNER on their own

schedule; the sandbox itself has no push credentials and never
fetches.

Problem

worktree.baseRef supports exactly two values, and in this topology
both are wrong:

• fresh (default) bases every new agent worktree on

refs/remotes/origin/<default-branch>. On a box that never
fetches, that ref goes stale the moment work lands locally; agents
silently build on old code and re-inherit already-fixed bugs
(observed in the field on 2026-07-14: agents reintroduced bugs that
were fixed on local main days earlier).

• head bases on the orchestrating session's current HEAD, which is

routinely parked on a docs/orchestrator branch or a mid-edit
worktree. It requires a fragile rebase-before-every-launch
discipline, and its failure mode is silent (stale or dirty base).

The workaround that keeps agents current — running
git update-ref refs/remotes/origin/main main after every landing —
REPURPOSES the remote-tracking ref and falsifies its meaning. The
owner of this repository used origin/main to know what is actually
pushed to the server and deployed; that signal was silently
destroyed, and deployment-relevant decisions were being made against
a falsified ref. This is corruption of an operational signal, not a
cosmetic inconvenience.

Claude Model

Other

Is this a regression?

No, this never worked

Last Working Version

_No response_

Claude Code Version

2.1.198

Platform

Anthropic API

Operating System

Other Linux

Terminal/Shell

Other

Additional Information

Model: Fable 5