Repository navigation
[Bug]: Claude background subagents die when T3 restarts the session after auto-compaction or idle, then show as "stopped by the user" #29319
Description
Version: T3 Code 0.1.7-nightly.20261102.4066, Linux (Fedora, KDE Plasma on Wayland), provider Claude (claude-opus-5-5).
What happens
Background subagents started with Claude Code's Agent tool (run_in_background: true) run inside the session's claude process. T3 ends and resumes that process several times a day, and every running subagent dies with it, halfway through its work. The restarts are not only T3 updates. In the systemd user journal for 2026-11-02, a t3code-claude-<id>.scope ends:
- at 07:27 and 09:29 together with
t3up.service, which is an update; - at 11:41, 12:13, 12:45, 12:48, 12:52, 13:20, 13:25 and 14:14 without an update. These seem to follow auto-compaction or longer idle time in the thread. In one case
/compactshowed "Compaction canceled" right before.
After the resume, Claude Code reports each of these subagents as "was stopped by the user and won't be resumed". Nobody stopped them, and the agent then refuses to continue the work (SendMessage fails). The work left in the subagents' checkouts stays uncommitted.
Expected
- An auto-compaction or idle restart of the Claude process does not end subagents that are still running. For example, the process stays up, or the subagents run outside the session's scope.
- If it has to end them, they are shown as interrupted rather than stopped by the user, so they can be resumed.
Related: #28331 (Claude subagents carried over from an earlier process stay Running). This report is the other side of it: the subagents are actually killed.
Workaround we now use: long background work goes through delegate_task (T3-owned child threads) instead of Claude Code's in-process background agents.
Activity
juliusmarminge commentedon Nov 2, 2026
I tried to reproduce this on current main (
47f2c8d29e, Linux) with a real Claude Code session. I shortened the 30-minute idle release to 60 seconds in a scratch build so the release would fire during the test. In all four runs theclaudeprocess stayed up while its background subagent was working. The server loggedorchestration-v2.driver-session.idle-release-deferredeach minute and released the process only after the subagent had finished:Agentsubagent withrun_in_background: truerunning a 7-minute command, with the parent idle./compactand a follow-up turn while the subagent ran./compactwith a steered message sent during it.So an idle Claude session with a running background subagent is kept alive here. Something else ended your processes at those times. To narrow it down, could you share:
driver-session,idle-release,claude-query,interrupt, orreleaseIfStillIdle. They are in the server's journal, or in~/.t3/userdata/logs/server.trace.ndjson*for the same time.t3code-claude-<id>.scope, including the result line (for exampleDeactivated successfullyor an OOM-kill message).t3_thread_send, or a steer or Stop.