Skip to content

Cold restore forks a second local claude --resume across worktrees onto a live transcript #17807

Closed

Description

@ikitkwan

Summary

Cold-restoring a sleeping agent can fork a second local claude --resume <sessionId> onto a transcript another live pane is still writing, when the two panes are in different worktrees. The existing dedupe cannot see across that boundary.

workspace-terminal-host-authority.ts already names this failure mode for the SSH case ("resuming a sleeping agent forks a second claude --resume onto a transcript the host is still writing", STA-3500/3498/3374). This is the same shape, purely local — no remote host involved.

Why the guard misses it

activeOrQueuedResumeClaimsProviderSession() in src/renderer/src/lib/resume-sleeping-agent-session.ts checks three sources for a live/queued claim on the same providerSession — agentStatusByPaneKey, pendingStartupByTabId, automaticAgentResumeClaimsByTabId — and every branch is scoped to the record's own worktree:

const worktreeTabIds = new Set((state.tabsByWorktree[record.worktreeId] ?? []).map(t => t.id))
…
worktreeTabIds.has(getAgentStatusTabId(entry) ?? '') && entry.worktreeId === record.worktreeId
worktreeTabIds.has(tabId) && …
worktreeTabIds.has(tabId) && claim.worktreeId === record.worktreeId

A Claude transcript is identified by sessionId and is global to the machine — it is not owned by a worktree. So a pane running that session in worktree A is invisible when a record for worktree B is resumed, and the resume proceeds.

Observed

macOS, Orca 1.4.192, three sessionIds duplicated at once. Two show the cross-worktree pattern plainly — original in the repo's main checkout, duplicate in an Orca workspace worktree:

sessionId 284b624a…   pid 29026  cwd …/imbee-automation          (main worktree)
                      pid 45980  cwd ~/orca/workspaces/…/T94     (different worktree)

sessionId 95c4f830…   pid 29625  cwd …/imbee-automation          (main worktree)
                      pid 45960  cwd ~/orca/workspaces/…/T126    (different worktree)

Both processes carry the same sessionId in ~/.claude/sessions/<pid>.json and resolve to one transcript. Ancestry confirms Orca spawned the second:

Orca Helper -> login -> zsh -> claude --resume <sessionId>

13 of 21 live sessions on this machine were started with --resume. Killing the extras is not durable — the same sessionIds re-duplicated within hours.

Impact

  • Two processes appending to one .jsonl. No corruption observed — 5,041 lines all parsed cleanly, and the file is not held open by either process — so this is a latent risk, not a realised one.
  • The measured cost is context: each duplicate re-sends the whole conversation per request. An 8.8 MB transcript loaded twice is the concrete waste.
  • The restored copy loads the transcript as of its resume, so subsequent writes from both interleave and the file stops being a faithful record of either.

Suggested direction

Deduping on providerSession for claude probably needs to be worktree-agnostic, since the identity it keys on is machine-global. I have not proposed a patch: the surrounding invariants span hibernation, mobile wake fanout and the SSH host-authority verdicts, and I could not tell from outside which cross-worktree resumes are legitimate.

Not verified

I did not confirm the cross-worktree path for a third duplicated session whose panes shared one cwd — a floating terminal (FLOATING_TERMINAL_WORKTREE_ID) would produce the same symptom by a different route, and I could not distinguish them from outside the app.

Activity

  1. added theissue type on Sep 19, 2026
  2. nwparker commented on Oct 15, 2026

    @nwparker
    Contributor

    Thanks for the detailed report and for identifying the cross-worktree boundary. This case is fixed on main by #19882 (e45cf438bc5adc845bc14554c174b480687b6210): a matching session with an exact live pane now prevents another automatic resume across worktrees, including after a completed turn.

    Verified on main e2f2277d51272bb36a906c87df68dd145e02b5a5 on macOS arm64 by executing the production restore sweep. Both working and done Claude peers in another worktree prevented duplicate launches; the control without a live peer PTY still resumed. All 40 targeted tests passed. This was production store-level verification, not a live two-Claude UI test.

    Closing as fixed on main. If this recurs on a build containing #19882, please let us know; a live process missing from the status/PTY inventory would need a separate investigation.

Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Metadata

Metadata

Assignees

No one assigned

    Labels

    bugSomething isn't working

    Type

    Projects

    No projects

      Milestone

      No milestone

      Relationships

      None yet

      Development

      No branches or pull requests

      Issue actions