Improve subagent batch cancellation and partial-result recovery

[puri-adityakumar]

Summary

fx can run multiple subagents in parallel, and recent work improves how their statuses are refreshed as they finish. However, a batch still needs clearer lifecycle behavior when children finish, fail, hang, or are cancelled.

Problem

If a batch contains several subagents:

• One child may fail while others have already completed
• The user may want to cancel the remaining children
• A stuck child may keep the batch open indefinitely
• Completed child results should not be lost because another child failed
• It may be unclear whether the batch can be resumed safely

Proposal

Improve subagent batch lifecycle handling by:

• Showing clear states such as running, completed, failed, cancelled, and timed out
• Preserving completed child results when another child fails
• Supporting cancellation of the remaining batch
• Ensuring child shutdown has a bounded timeout
• Showing a final batch summary
• Making interrupted batch state recoverable where practical

Scope suggestion

This should be designed separately from ACP prompt queueing. The first step could be a design and test change for deterministic child states, cancellation, and partial results before adding a larger resume workflow.

I did not find an existing fx issue specifically covering this complete subagent batch lifecycle.