netclaw-dev/netclaw/run-reminder

Run a Reminder Now

A scheduled reminder runs unattended. Nobody can answer an approval prompt in that run, so a call that needs approval and has no saved grant is denied. This skill runs the reminder's exact prompt in this chat, with the user present. The user answers each prompt. An "Always" answer saves a grant that the scheduled run reads.

The user message gives the reminder ID. If it is missing, call list_reminders and ask which reminder to test.

Procedure

1. Warn the user first. Tell the user that the test does the real work.

It is not a dry run. For example, a disk-cleanup reminder really deletes files, and a reminder that posts to a channel really posts there. Ask the user to confirm before you continue.

1. Get the prompt. Call run_reminder with the reminder ID.

• If the result starts with Error:, show the error to the user and stop.

A reminder must be tested in a chat with the same audience as the reminder. The error tells the user where to run it.

1. Follow the returned prompt exactly. Do the steps that the prompt asks

for, in this chat, as the scheduled run would. Do not change commands, folders, or tools to avoid a prompt. The test must use the same calls.

1. Watch each approval result. A tool result can end with a line such as

[approval: once] or [approval: always in this folder].

• [approval: once] and [approval: this chat only] do NOT carry over to

the scheduled run. Each scheduled run is a new session. Tell the user that this step will be denied when the reminder runs on its schedule. Offer to run the step again so the user can choose an "Always" answer.

• [approval: always in this folder], [approval: always in this repo],

and [approval: always anywhere] save a grant for this audience. The scheduled run reads it.

1. Summarize at the end. List each step that needed approval:

• the steps that now have an "Always" grant;
• the steps that the scheduled run will still deny, and why (a one-time

or chat-only answer, a denial, or a call that no grant can cover).

Pitfalls

• A one-shot reminder deletes itself after it runs or after it is cancelled.

Test a one-shot before its fire time. After that, run_reminder reports that the reminder ended.

• Do not use set_reminder to make a copy of the reminder for the test. The

test must use the real reminder's prompt and audience.

• The test does not change the schedule or the reminder history.
• A reminder reaches only the child runs of the session where it fires. If the

reminder prompt names a child run, that run must belong to that session. A test in another chat does not prove that the scheduled run can reach the run.

• A model can choose other commands on a later run. A grant matches the

command and the folder that the user approved. A repository or "anywhere" grant accepts more change than a folder grant.