Usage-limit auto-continue doesn't resume workflows or subagents — they all fail

Problem

When a session hits the usage limit and then auto-continues after the limit resets, the main agent picks up — but nothing else does. Every agent in a running workflow fails with errors. The workflow is left half-finished.

What happens after that is worse than either finishing or starting clean: because the workflow already began, the continuation doesn't start a fresh one. It reads whatever the partial run left behind and carries on from that. It "works", but the result is noticeably worse than an uninterrupted run.

So in practice this isn't really continuing where it left off. It behaves more like the prompt being re-sent. For a plain single-turn task that's fine. For anything multi-agent, it isn't.

What I'd expect

The same behaviour you get when the internet drops: it waits, retries, and nothing is lost. A usage limit is a known, finite wait — arguably an easier case than a network failure, since you know exactly how long it lasts. It should feel like a pause, not a restart.

Why it matters

The tokens already spent on the agents that got killed are wasted, and then more tokens get spent producing a worse answer. On a plan with limits, that's the expensive kind of failure — you pay twice and get less.

Ask

I'm not going to guess at how any of this is implemented internally. From a user's side the request is simple: make the limit-reset continuation cover workflows and subagents too, not just the main agent.