[Bug]: Sending to another Claude instance leaves the old Claude process running unseen #20343

Before submitting

• I searched existing issues and did not find a duplicate.
• I included enough detail to reproduce or investigate the problem.

Area

apps/server

Steps to reproduce

1. Configure two Claude provider instances, A and B, with different Claude config directories, so they don't share native history.
2. On instance A in Full access, ask Claude to start a background subagent (Agent tool, run_in_background) that works for a few minutes and then edits a file. Let the turn finish, so the thread is idle while the subagent runs.
3. Select instance B in the composer and send a message. The thread shows a "Provider handoff" item and the run starts on B.
4. Check instance A's claude process with ps, and wait for the subagent to finish.

Expected behavior

When the send hands the thread to B, T3 detaches A's session and stops its process. It already does this when the same switch starts from a queued run or from thread.model-selection.set. If the background subagent should survive, T3 refuses or defers the switch, as discussed in #16000. No work runs on A that the thread doesn't show.

Actual behavior

A's process keeps running and stays bound to the thread. Its session row still reads ready. When the subagent finishes, Claude Code on A starts its own wake turn. T3 offers one continuation. It runs on B for about 5 seconds with no output, because B can't replay A's buffer (#17099). After that, A keeps taking turns on its own. T3 reads them from the stream but shows none of them.

This is an unsupervised Full access session. It changes files with no UI, and it keeps spending A's quota after the user moved the thread off A.

One case, nightly build, times UTC:

• 14:57 T3 writes its last message to A. No query.close for A is logged afterwards.
• 15:47:40 the user sends with B selected. Handoff from A to B. Only provider-turn.start is queued. The thread has no provider-session.detach effect at all.
• 15:51 to 17:05 the provider event log for A's session has 7 task_notification, 6 task_started and 7 result frames. In that time A ran shell commands and made 9 git commits in the workspace.
• 17:05:45 A's account reaches its weekly usage limit.
• 17:20 the process was stopped by hand with SIGTERM.

Code on main @ e58a485:

• Confirmed: the send path ignores the release list. dispatchMessage detaches the old session only when canResumeAcrossInstances is true (Orchestrator.ts#L6492-L6519). Instances with different config dirs get create_with_handoff (ProviderSessionTransitionPolicy.ts#L48-L53), and ProviderSwitchService puts A's live session in releaseProviderSessionIds (ProviderSwitchService.ts#L180-L193). startNextQueuedRun (L1736-L1741, L1873-L1882) and dispatchThreadMutation (L3381) use that list. The send path does not. A detach would have stopped the process: Claude sets supportsMultipleProviderThreadsPerSession: false, so detach calls releaseEntry (ProviderSessionManager.ts#L2414-L2423).
• Suspected: the idle timeout can't catch it. Every event from a session calls touchActivity(..., { providerActivity: true }) (L1946), which also clears the background-work pin (L1226). A session that keeps starting its own wake turns never reaches the 30-minute idle release.
• Related, already reported: the wake is offered to the thread's current provider, once per native thread (ClaudeAdapterV2.ts#L6206-L6234). That is [Bug]: Claude background subagents stay Running forever when their wake continuation runs on another provider #17099. PR fix(server): Claude background subagents end when their wake can't run on Claude #17109 drops such wakes. That removes the empty run on B, but A's process still runs.

Impact

Major degradation or frequent failure

Version or commit

0.1.1-nightly.20261016.3190 (9b150f7). The cited code is unchanged on main @ e58a485.

Environment

macOS 26.6.2, T3 Code Nightly desktop. The message was sent from the mobile app; web sends the selection the same way through client-runtime. Claude Code 2.1.301, claude-opus-5-5 on both instances.

Workaround

Wait for background subagents to finish before switching instances. After a switch, stop the old claude process by its PID. Archiving or settling the thread also detaches every bound session.

Suggested fix

• In dispatchMessage, detach the sessions in the switch plan's releaseProviderSessionIds for every instance switch, as startNextQueuedRun does. Apply the background-work guard that [Bug]: Runtime-mode, instance and worktree changes detach the Claude session and kill live subagents without the background-work check #16000 settles on.
• Don't let a session the thread no longer uses extend its own life through provider activity, so a missed detach still ends.

Related: #16000, #17099 and #17109, #16565 and #17096 (same trigger, different symptom).