Scheduled prompts ("send later") for agent chats

[mgaitan]

What are you proposing?

Add support for scheduled prompts ("send later") in agent chats.

When a prompt fails due to quota/rate limits, allow turning it into a scheduled action (e.g. retry at a specific time). Also expose a general “Send later” option from the prompt box.

Why does this matter?

When working with agents, hitting quota limits breaks the flow at the worst moment.

In many cases, the user already knows the next prompt (continue, resume, etc.), but has to remember to retry later. This feature removes that friction.

It also covers a different use case than queues:

• queue = run as soon as possible
• scheduled = run not before a specific time

Are there any examples or context?

Typical workflow:

• agent run fails → “retry at 3:00 AM”
• user clicks “Retry automatically at 3:01 AM”
• prompt is sent later without manual intervention

I’ve implemented a similar idea in my Telegram-ACP-bot (an acp client) and it works well for “continue later” scenarios after quota exhaustion. mgaitan/telegram-acp-bot#195

Possible approach

MVP:

• allow scheduling a prompt with a datetime
• show a list of scheduled prompts (edit/cancel)
• surface failures if execution doesn’t succeed

Nice-to-have:

• detect retry/reset time from provider errors and suggest it automatically
• quick actions from error messages (“Retry at X”, “Schedule later”)

[silvenon]

I need this really bad! I'm working on features that are very draining for Claude Agent, and continuing my effort after running out of tokens is very painful:

• having to set a timer when my tokens will be restored
• if I just send "continue" in the same chat, that won't work, I need to restart Zed
• Claude Agent lost its train of thought, I need to tell it to regain context, it does it by calling Claude Code context file with python3
• it doesn't really know where it left off, mostly it focuses on the initial prompt, repeating some things, I assume due to compacting

ACP being able to handle running out of tokens would be a killer feature! I envision it as a "Continue" button at the bottom.

But it sounds pretty complicated, not sure how or if CLI handles this already.

[malfroid]

Example related solution: claude-auto-retry

This is the only reason why I currently prefer claude code in the terminal instead of through Zed.

[mgaitan]

Maybe a simple mechanism via Tasks could already cover this, if the prompt input were exposed as a file/environment variable.

For example, that would allow scheduling something like:

echo "echo 'continue please' >> '$ZED_AGENT_SESSION_1_PROMPT'" | at now + 3 hours

This would also enable injecting other kinds of events into the loop. For instance, a script that monitors a CI job and, once it finishes, posts the result back into the agent as a new prompt.

It would also be useful to expose additional environment variables, such as something like $ZED_AGENT_SESSION_1_ERROR, so errors can be parsed and handled automatically. In my case, for example, I could extract the retry time from the error message and schedule the next attempt accordingly.

As an extra, it could be helpful if agent failures triggered a hook, so users could attach a Task to it.