disk-digest

A Slack bot that scrapes the daily astro-ph new listings, uses Claude to identify papers relevant to your research group, and posts a summarized digest to the channel it is added to. Papers co-authored by channel members are highlighted with a congratulations.

Requirements

1. An LLM API key. This repo is set up for Parley (MIT's OpenAI-compatible gateway to Claude models), but any OpenAI-compatible endpoint works — set PARLEY_BASE_URL and PARLEY_API_KEY accordingly.

2. The ability to build and install an app on your Slack workspace.

3. Node.js v20.6 or later (for built-in .env loading).

Setup

There are two stages: build a Slack app that can post to your channel, then run the script daily — either via GitHub Actions (recommended) or a local cron job.

Slack App

1. Go to api.slack.com/apps and click Create New App → From scratch. Give it a name (e.g. "Disk Digest") and select your workspace.

2. In the left sidebar go to OAuth & Permissions. Under Bot Token Scopes add the following scopes:

   • chat:write — post messages
   • channels:read — list channel members for author matching
   • groups:read — same, if the digest channel is private
   • users:read — look up member names for author matching

3. Scroll to the top of the same page and click Install to Workspace, then Allow.

4. Copy the Bot OAuth Token (starts with xoxb-...) — this is your SLACK_BOT_TOKEN.

5. In Slack, right-click the channel you want the digest posted to → View channel details → scroll to the bottom to find the Channel ID (starts with C...) — this is your SLACK_CHANNEL_ID.

6. Invite the bot to the channel (/invite @Disk Digest). This is required — the bot reads the channel's member list to know who counts as "your team".

Option A: GitHub Actions (recommended)

The repo ships with .github/workflows/daily-digest.yml, which runs the digest every weekday with no server needed.

1. Fork/clone the repo to your own GitHub account.
2. In your repo go to Settings → Secrets and variables → Actions and add four repository secrets: SLACK_BOT_TOKEN, SLACK_CHANNEL_ID, PARLEY_API_KEY, PARLEY_BASE_URL.
3. That's it. The workflow runs at 12:00 UTC on weekdays; adjust the cron: line to taste. You can also trigger a run manually from the Actions tab (workflow_dispatch).

Option B: Run locally

1. Clone the repo and run npm install.
2. Copy .env.example to .env and fill in your credentials. Use KEY=value with no spaces around =:
   • SLACK_BOT_TOKEN — from your Slack app's OAuth & Permissions page
   • SLACK_CHANNEL_ID — right-click your channel in Slack → View channel details
   • PARLEY_API_KEY / PARLEY_BASE_URL — from your LLM gateway
3. Run with npm start (equivalent to node --env-file=.env disk-digest.js).

To test without posting to Slack, use npm run dry-run — the digest is printed to stdout instead.

To run daily via cron, open your crontab with crontab -e and add:

0 9 * * 1-5 cd /path/to/disk-digest && node --env-file=.env disk-digest.js >> /tmp/disk-digest.log 2>&1

arXiv announces new submissions around 00:00 UTC, so choose a time after that in your timezone.

Customizing the relevance filter

Relevance is decided by a Claude yes/no check on each paper's title and abstract. Edit the RESEARCH_TOPICS constant at the top of disk-digest.js to describe your group's interests — plain English works, e.g. "exoplanet atmospheres, JWST transmission spectroscopy, or hot Jupiter dynamics".

The model IDs used for the relevance check and the summaries are set in the RELEVANCE_MODEL and SUMMARY_MODEL constants.

Notes

• Author matching compares last name + first initial between arXiv author lists and Slack profile names (diacritics are ignored). Common surnames can occasionally produce false-positive "congrats" tags.
• Replacements are excluded: only the New submissions and Cross-lists sections of the listing page are scanned, so revised versions of old papers don't reappear.
• If the arXiv listing page hasn't been updated for today (weekends, arXiv holidays), the script exits quietly without posting.
• No duplicates across days: every paper checked is recorded in posted-ids.json (kept for 30 days) and skipped on later runs. The GitHub Actions workflow commits this file back to the repo after each run; dry runs don't write it.