TL;DR Open the run from History at
claude.ai/code/routines, or its link on your phone, and write to it there. It comes back with its whole conversation on a fresh machine, and your reply counts as the yes the scheduled prompt can't give. Don't archive it if you plan to come back.
Every run of a cloud scheduled task is a full Claude Code session, a new one per run. Its machine goes away when the session ends or sits idle, but the conversation doesn't close: open the run again and Claude Code provisions a new machine, loads the history, and carries on from your next message. A finished run is still a conversation.
When a run stops to ask you something or leaves the work half done, the reflex is to hit Run now. That starts from zero: it clones again, reads again, and knows nothing about what the last run found out or why it stopped.
What happens when you come back
My wmedia.es tips task ran on the morning of November 5, published its tip and finished. That evening I went back to that same run, we settled a second tip's title in the conversation, and it published that one too, with nothing new launched. An excerpt of that session's metadata:
{
"title": "⚡ CC Tips (cloud)",
"origin": "scheduled_trigger",
"created_at": "2026-11-05T09:34:46Z",
"updated_at": "2026-11-05T19:35:09Z",
"permission_mode": "auto",
"context_usage": { "used_tokens": 443646 }
}
A run's machine doesn't wait for you, so coming back hours later means a fresh one. Here is what that brings back and what it doesn't:
| Comes back | Doesn't |
|---|---|
| The whole conversation | Shell commands and subagents that were still running (Claude Code tells Claude which ones stopped) |
| Its permission mode, auto unless auto mode was unavailable | Tasks and /loop it scheduled inside the session |
| What it pushed to a branch, which lives on GitHub | What it left on disk, with no guarantee (below) |
The disk is the part nobody promises you. The run that wrote this tip stopped waiting for the yes to publish, and when it came, less than half an hour later, it started on a freshly booted machine with everything in place: the uncommitted drafts, its files in /tmp, even the packages it had installed with apt. But the docs only promise the conversation, so a run that might leave something pending commits and pushes before it stops.
How to pick a run back up
1. Open it
In the browser: claude.ai/code/routines, the task, then the run under History. It opens as a regular session. On your phone, open the run's link (claude.ai/code/session_...): it is the same session in the Code tab of the Claude app, with nothing running on your computer. That is why my task stores that link on every row it touches.
2. Answer it, or ask for what's missing
Write as you would in any session. One difference matters: the prompt that fires the task isn't a live message from you and doesn't count as approval, so a well-written run stops before it publishes, sends or deletes anything. Your message in the run is yours: "publish it" there is the yes it was missing.
If what's holding it is a permission prompt (an action that matches an ask rule, or auto mode blocking 3 actions in a row or 20 in total), it is waiting in that same session. Answer it there and it moves on.
3. If you'd rather carry on in your terminal
claude --teleport session_01AbCdEf...
Run it from a clean checkout of the same repository: the branch and the conversation come down, and if the run was compacted, only the summary of what came before. Whatever you do after that stays local and doesn't go back to the run: how that differs from Remote Control has its own tip.
What coming back costs
- The first message rereads the whole context. Unless you set it yourself, the conversation's cache expires an hour after the last request within your subscription's included usage, five minutes outside it, so hours later it's gone and the first turn reprocesses the entire context. A long run carries a lot of it: mine ended the day at 443,646 tokens. It isn't free: why a turn after a break is slow and costly.
- Archiving closes it. An archived session no longer accepts messages, and if you archive a run that succeeded, its History row turns to Cancelled. If you need it again, bring it back from the Archived tab at
claude.ai/code/sessionswith Unarchive.
Run now is still the right call when you changed the prompt or want a clean context. To finish what a run left half done, write to that run.
Official docs: View and interact with runs · Environment expired
Requirements
- Cloud scheduled tasks, on a Pro, Max, Ultra, Team or Enterprise plan. The Claude app on iOS or Android to pick it up from your phone.
