The parent agent's turn ends right after `agent_spawn` returns
`status: Running`, leaving children to finish in the background with no
path back into the parent's inference loop. The model has to be poked
by a human before it resumes the plan.
Wire a wakeup channel from `run_subagent_task` into the engine's turn
loop. When the model produces no more tool calls but direct children
are still running, the loop now blocks on the next completion (with
cancel and steer escape hatches), drains all pending sentinels, and
re-enters inference with the existing `<deepseek:subagent.done>`
elements injected as user messages. This fulfils the contract already
documented in `prompts/base.md` (lines 189-205): the parent is promised
it'll see the sentinel when a child finishes.
The async `agent_spawn` semantics from #239 are preserved — only direct
children fire on the channel (gated by `spawn_depth == 1`), so
grandchildren spawned recursively don't flood the parent.
y0sif
cae4ca3c0f