[Bug]: PR discovery checks hidden subagent threads every minute, so agent-heavy servers run hundreds of git lookups#26970

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. Use agents that start subagents or delegates on their own branches (e.g. delegate_task, review fan-outs). Each one becomes a child thread.
2. Keep inactivity settling at its default of 3 days.
3. Leave t3 serve running with a client connected, and count runGitCommand spans in server.trace.ndjson.

Expected behavior

Subagent and delegate threads don't appear in the sidebar, and they are finished once their parent has the result. The per-minute PR discovery sweep should skip them, or settle them with their parent, so the git and PR lookups scale with the threads a user actually watches.

Actual behavior

ThreadPullRequestServiceV2.synchronize reads getShellSnapshot({ location: "active", unsettledOnly: true }). That query filters out settled and archived threads but not child threads, so every unsettled subagent with a branch gets a git.branchPullRequest lookup each minute (ThreadPullRequestService.ts).

On my server, agents create about 300 subagent threads a day. They only settle by inactivity after 3 days, so about 600 of them are always in the sweep. Unsettled threads, read from statev2.sqlite:

In about 8 minutes of server.trace.ndjson, the server ran about 3,160 git commands. About 1,900 of them came from ThreadPullRequestServiceV2.synchronize, and most of the rest from other PR or branch status reads. t3 serve sat at about 100% of one core. Shell and PR reads took up to 40 s (ws.orchestrationV2.projectShellItems up to 39 s, PullRequestSyncReactor.syncGroup up to 97 s), and remote clients dropped with SocketCloseError: 1006 and RemoteEnvironmentAuthTimeoutError ... timed out after 6000ms.

Settling the idle subagent threads by hand helps, but they aren't visible in the sidebar, so a user can't find them.

Related: #24098 (merged/closed PR threads re-read git every minute), #18618 (PR discovery rebuilds full thread shells), #26224 (event-loop freezes drop remote clients).

Impact

Major degradation or frequent failure

Version or commit

0.1.3-nightly.20261025.3669

Environment

Linux, t3 serve headless, clients over Tailscale (web + mobile). statev2.sqlite 8.7 GB.

Logs or stack traces

Screenshots, recordings, or supporting files

No response

Workaround

asking my agent to settle just the subagents