Donate Your Instance to Curation (`/donate-curation`)

This is a plain-language guide to /donate-curation — a way to lend your Claude Code instance to dismech disease curation during hours you choose (for example, every evening while you're away from your desk). You say when; it does the curating.

You do not need to understand cron, UTC, or the code to use it. But there is a short one-time setup, and the recurring version cannot run at all until you do it — so start with Set-up you must do once below.

What it actually does

Think of it as a volunteer shift you schedule for your Claude Code instance.

It works on one disease at a time. It never juggles several at once.
Each time it wakes up, it first checks GitHub to see if anything it already started still needs its attention — a reviewer asked for changes, a check failed, or a merge conflict appeared. If so, it fixes that and stops for the session.
Only when nothing it started needs work does it pick up a brand-new disease, research it, and open a new proposal (a "pull request", or PR — the unit of proposed change on GitHub).
It hands finished work off and leaves it alone. Work waiting for a reviewer, or approved and just waiting for the automatic merge, does not hold up the next disease.

What it will never do

By design, it never:

merges its own work,
approves its own work, or
dismisses a reviewer's request for changes.

A human (or the reviewer bot) always stays in the loop for those. The worst it can do is open proposals and improve them — never publish them unreviewed.

Set-up you must do once

There are three things to set up before scheduling. The easiest way to do all of them is to just type /donate-curation and let Claude walk you through what's missing — but here's what each one is and why it's needed.

The tool acts on GitHub as you, so you must be signed in. Check with:

gh auth status

If it says you're logged in, you're set. If not, run gh auth login and follow the prompts.

2. Tell it your GitHub username

This is a safety catch, and it matters most for the scheduled (cloud) version. Each run only claims a new disease when your in-progress work is all clear. If a scheduled run ever signed in as the wrong account (a robot account, say), it would see an empty to-do list and wrongly start a brand-new disease every hour. Recording your username stops that: a run that isn't signed in as you does nothing at all.

You set this once in the settings file /.claude/schedule-config.yaml, on the line:

github_login: null      # change null to your GitHub username, e.g. jsmith

It ships blank on purpose — your username is personal to you and must never be committed as a shared default. If you'd rather not edit the file by hand, ask Claude: "set my github_login in the schedule config".

3. Set up the cloud environment (required for scheduled shifts)

This is the one people miss. A recurring shift runs in Anthropic's cloud — not on your Mac — so it needs a cloud environment: a ready-to-go copy of the dismech project in the cloud, with the tools and sign-ins the curation work needs. Without one, a scheduled shift cannot start.

If you already have a Claude Code cloud environment for dismech, put its id on the environment_id: line of /.claude/schedule-config.yaml.
If you don't, create/connect one from the Claude Code cloud area at https://claude.ai/code, then copy its id into that line.
Not sure? Ask Claude: "help me set up the cloud environment for scheduling" — it will point you at the right place and, when you register a shift, prompt you for the id if it's still blank.

Like your username, this line ships blank so nobody inherits one person's environment by accident.

The run-once version needs only step 1. Typing /donate-curation with no schedule runs immediately in your current session (not the cloud), so it needs GitHub sign-in but no cloud environment. Steps 2 and 3 are what unlock the recurring, unattended shifts.

The four things you can type

Type these in Claude Code while you're in the dismech project.

You type	What happens
`/donate-curation`	Runs once, right now, in this session. Good for trying it out (needs only step 1).
`/donate-curation 8pm-12am every weekday`	Sets up a recurring cloud shift (needs steps 1–3).
`/donate-curation status`	Shows what's scheduled and what it's working on.
`/donate-curation stop`	Ends the recurring shift.

Saying when (the recurring form)

You describe three things in plain words — the daily window, how often, and (optionally) when to stop:

Window — "8pm to midnight", "9am–5pm".
How often — "every day", "every weekday", "weekends", or "once".
Stop after (optional) — "for 1 week", "for 3 days", "until 2026-10-01". If you leave this off, it keeps running until you stop it.

Examples you can type as-is:

/donate-curation 8pm-12am every weekday
/donate-curation 9am-11am every day for 2 weeks
/donate-curation 8pm-11pm weekends until 2026-12-31
/donate-curation 8pm-9pm once (just tonight — see below)

"Once" (or "tonight") does not mean a single wake-up at 8pm. It means the instance takes its shift tonight only — waking each hour across the window just like a recurring shift, then switching itself off the next day. It's the right choice for a one-evening trial.

Claude will read back the schedule it understood — including the exact clock times, adjusted for daylight-saving — and ask you to confirm before setting anything up. Always glance at that read-back; the time-zone conversion is the one thing worth double-checking.

What a scheduled shift looks like

Once set up, the shift runs in the cloud on your own subscription — your computer does not need to be awake or even switched on. Each evening in your window, it wakes up, checks GitHub, and does exactly one thing: either finish something that needs its attention, or start one new disease. Anything it finishes is left for a reviewer; you'll see the proposals pile up on GitHub, get reviewed, and merge on their own over the following days.

If it ever runs out of its usage allowance mid-task, it stops cleanly and saves its place. The next evening it simply picks up where it left off.

The very first scheduled run is a test run. Before trusting the nightly rhythm, the tool does one cloud run to confirm the environment really has everything the curation work needs. Until that test run is signed off, keep the schedule short (e.g. once, or for 1 day) and check the result.

Checking in on it

/donate-curation status

This tells you the schedule (window, how often, stop date) and the one thing currently in hand — whether it's researching a disease, waiting on a reviewer, or waiting to be merged.

Changing or stopping it

To change the hours or frequency: just type a new /donate-curation <window> ... — it replaces the old shift.
To stop entirely: type /donate-curation stop. In-progress work is allowed to finish saving; nothing is thrown away.
To remove it completely (an extra, optional step): /donate-curation stop turns the shift off; to delete it from your account outright, open https://claude.ai/code/routines in a browser and remove it there. Claude will print that link for you.

Frequently asked

Why does it need my GitHub username? As a safety catch. A scheduled run only starts a new disease when your in-progress work is clear — so if it ever ran signed in as the wrong account, it could misread an empty list and start new work every hour. Recording your username means a run that isn't you does nothing.

What is a "cloud environment" and why do I need one? It's a ready-made copy of the dismech project that lives in the cloud, with the tools and sign-ins the curation needs. Recurring shifts run there (so your Mac can be off), and they can't start without one. The run-now version doesn't need it — it uses your current session.

Will it curate a disease I don't know anything about? It picks from the shared queue of diseases waiting to be curated. Every proposal it opens goes through the normal review before anything is accepted.

Can two evenings' runs collide and do the same disease twice? No. Each run checks first, and if the previous night's run is somehow still going, the new one simply steps aside.

Does this cost the project money / use an org account? No — it runs on your personal Claude subscription, not the project's. There's no external service involved.

How many diseases at once? Exactly one, in this first version. Doing several in parallel is a planned future improvement, not something this version does.

Status of this feature

The run-once command (/donate-curation with no schedule) and the GitHub check-in logic work today. The recurring cloud shift depends on the one-time set-up above — chiefly the cloud environment — plus the first test run described earlier. Until you've done that set-up and seen a test run succeed, prefer a short once / for 1 day schedule and check the results rather than leaving a long unattended shift running.

For the technical details behind all of this — the exact scheduling API, the PR classification rules, and the frozen instructions each cloud run follows — see the donate-curation skill definition (.claude/skills/donate-curation/SKILL.md) and issue #11657.

Donate Your Instance to Curation (/donate-curation)