[BUG] SubagentHandback from a teammate-spawned subagent reports "spawning agent no longer running" and the report never reaches the caller

Environment

• Claude Code 2.1.289, Windows 11, model opus, permission mode auto.
• Agent teams in use: a lead session spawns named teammate lanes; each lane

dispatches its own subagents (code reviewers, a UI verifier) with the Agent tool.

• Plugins installed, but none registers a SubagentStop/handback hook that touches

delivery (the only SubagentStop hook blocks a stop while the subagent's own
background shell task still runs; it never fires on an Agent child).

Summary

When the spawner of a subagent is itself a teammate (not the lead session), the
subagent's SubagentHandback call fails with "the agent that spawned you is no
longer running". The harness keeps re-prompting the subagent to call
SubagentHandback (handback-send-enforce reminder), each retry fails the same way,
and the subagent ends with the report only in its final text. The report is then
delivered to the lead (or to a different, already-finished lane) or nowhere. The
lane that dispatched the subagent sits idle waiting for a report that never comes.

Steps to reproduce

1. Start a lead session with agent teams enabled.
2. Spawn a named teammate (lane) with a task that includes "dispatch a

code-reviewer subagent, wait for its report, then act on it".

1. In the lane, dispatch a subagent with the Agent tool (default, so it runs in the

background and the lane is notified on completion).

1. The lane ends its turn waiting for the completion notification (or keeps

working; both were observed).

1. The subagent finishes and calls SubagentHandback({message}).

Expected

The report is delivered to the lane that dispatched the subagent. If that lane
truly has ended, the handback either wakes the lane (as a completion
notification does for the lead) or fails once with a clear error and the lane is
told its subagent ended without delivering.

Actual

• SubagentHandback fails 3-4 times with "the spawning agent was no longer

running", while the enforce reminder keeps demanding the call (it does not check
caller liveness).

• In at least one case (2026-10-03, lane-1389, PR 1403) the spawning lane was

alive and waiting the whole time.

• SendMessage replies from the subagent to the lane by name were also lost in

one case (2026-10-02, lane-1330).

• Reports surface in the lead's context or not at all; the lane stalls until the

lead relays the report or someone reads the subagent transcript jsonl.

Evidence

Machine-wide incident log, recurrence key handback-caller-gone (plus
subagent-handback-parent-not-running, subagent-no-handback): 10 entries
2026-10-01 to 2026-10-03, across lanes 1329, 1330, 1331, 1333, 1389 and the
reviewer side, PRs 1357, 1358, 1361, 1362, 1364, 1366, 1396, 1403. Every entry has
a teammate lane as the spawner; none involves a subagent spawned by the lead
session directly.

Representative entries:

• 2026-10-01 code-reviewer: "SubagentHandback failed 3 times: the spawning agent

was no longer running, while the harness kept re-prompting for handback."

• 2026-10-02 lane-1330: "Three foreground subagents ... ended without delivering

their reports. SubagentHandback reported that the caller was gone, and their
SendMessage replies never reached the lane."

• 2026-10-03 lane-1389: "SubagentHandback failed 4 times each with 'the agent that

spawned you is no longer running', while the spawning lane was alive and
waiting."

Workarounds in use

• The lead relays the report to the lane by SendMessage.
• The lane reads the subagent's final message from its transcript jsonl.

Suspected cause

Handback routing resolves the caller from the lead session's agent registry, so a
subagent whose parent is a teammate either finds no live caller (teammate turn
ended, or teammate not registered as a handback target) or falls back to the lead.
The enforce reminder does not check caller liveness, so the subagent loops.