Upstream issue draft — live CLI session torn down at run-end kills in-flight background subagents

Repo: openclaw/openclaw
Companion to: PR #99139 (fix(claude-cli): preserve reused session on empty response) and issue #99131.
Status: draft for Josh to file.

Summary

Follow-up to #99131 (which we reported). #99139 fixes the session-binding half of #99131;
this issue is the process-teardown half that #99139 leaves open — the part that actually kills
in-flight background subagents.

When a claude-cli-backed turn ends, OpenClaw force-closes the reused "live session"
(closeClaudeLiveSessionForContext → closeLiveSession(session,"restart") →
session.managedRun.cancel()), which cancels that CLI process and its whole tree. Any
background subagents (Task/Agent tool with run_in_background) that the ending turn spawned are
driven by that same live-session process, so they are stranded/killed the instant the turn ends —
they cannot outlive the turn that created them. Elapsed wall-clock is irrelevant: with the driver
gone, a "background" subagent makes zero further progress and is later reaped.

This is the process-layer half of the same family as #99131/#99139. #99139 correctly stops the
persisted session binding from being cleared on empty_response; it does not touch the process
teardown, so the subagent kill (and the self-reinforcing notification-preemption cascade) persists.

Repro

1. In a session, spawn ≥1 background subagent (run_in_background: true).
2. End the turn (normal reply, sessions_yield, or let a resumed turn return empty via a queued
   task-notification).
3. Observe: the subagent processes stop making progress and are torn down; on the interactive path
   the gateway logs claude live session close: … reason=restart at every turn-end (even on
   successful turns), confirming cleanupCliLiveSessionOnRunEnd:true closes the live session
   unconditionally.

Root cause (from dist trace, 2026.6.11)

• cli-runner's runCliAgentInternal: run-end cleanup calls closeClaudeLiveSessionForContext
  unconditionally when params.cleanupCliLiveSessionOnRunEnd === true, with no awareness of
  runError or of any still-running background children.
• claude-live-session's session object has no in-flight/background-children counter — only a
  single currentTurn slot (scheduleIdleClose guards on !session.currentTurn, which never
  reflects a background child) and CLAUDE_LIVE_IDLE_TIMEOUT_MS = 600000 (hardcoded 10-min idle
  close, no env override).

Proposed fix (the real one)

Introduce a live-children refcount on the session (increment when a background subagent/child run
attaches to the live session, decrement on its completion) and make all three teardown paths
"drain before close":

1. run-end cleanup (cleanupCliLiveSessionOnRunEnd) — skip/defer while liveChildren > 0;
2. scheduleIdleClose — extend the guard from !session.currentTurn to
   !session.currentTurn && liveChildren === 0;
3. closeLiveSession(reason) — for "idle"/"restart" (not "abort"/error-driven), no-op while
   liveChildren > 0 and re-arm a drain check.

Secondary: make CLAUDE_LIVE_IDLE_TIMEOUT_MS env-configurable.

Interim local mitigations we run (hand-patches, reapplied after each upgrade)

• PR fix(claude-cli): preserve reused session on empty response #99139 applied locally (session binding preserved on empty_response).
• Local patch Login fails with 'WebSocket Error (socket hang up)' ECONNRESET #2: cli-runner skips the run-end force-close when runError is a FailoverError
  with reason === "empty_response" (script:
  scripts/ops/openclaw-patches/apply-cli-runner-empty-response-keep-live-session.js). This
  neutralizes the empty_response cascade but does not implement the full drain-before-close, so
  background subagents still don't survive a normal successful turn-end. Hence this issue.

Note

shouldClearReusedCliSessionAfterError returns true for AbortError, so a genuine user
interrupt still clears the session binding (fresh context). If that's undesirable for
interactive sessions, consider sparing AbortError too, or making it configurable.