Background forked skill completes before its subagents finish; SendMessage resume then fails with contradictory state

Environment: Claude Code 2.1.227, macOS (Darwin 25.5.0), terminal CLI.

What happened

1. I ran /code-review <PR#>, which launched as a background forked skill (<forked-skill-launch> with a named subagent code-review).
2. The skill agent spawned several finder subagents, then ended its turn with a final message saying the finders were still running and "results will arrive as notifications". The harness marked the agent completed.
3. The finder subagents' results were never collected — the review produced no final report. From the main session's view, ListAgents shows code-review · completed, and its TaskOutput is just the mid-work status message.
4. Attempting to resume it with SendMessage (which the docs say resumes a completed agent from its transcript) fails every time with a self-contradictory error:

Agent "code-review" is stopped (completed) and could not be resumed:
Agent aaef6d27daeba2157 is already running or being resumed

The agent is simultaneously reported as completed (ListAgents, the error's first clause) and "already running or being resumed" (the error's second clause). Retrying gives the identical error, so the agent is permanently wedged: not running, not resumable.

Expected

• A skill agent that spawns background subagents should not be marked completed while its children are still running — or its children's completion notifications should still be deliverable somewhere instead of dropped.
• SendMessage to a completed agent should resume it, per the tool's documented behavior. If a resume lock is stuck, it should time out or be clearable rather than wedging the agent forever.

Impact

The user-facing result of a background /code-review was silently lost, and there is no way to recover the child agents' work from the main session. The main session had to redo the review from scratch.