Daily AI Digest

A GitHub Actions workflow that builds one Markdown file per day under _daily/ from a curated list of AI RSS / arXiv sources, with each item summarised by an LLM (OpenRouter, DeepSeek primary / fallback).

Published on the site at https://runguoli.com/daily/, with an RSS feed at https://runguoli.com/daily.xml.

Files

Path	Purpose
`scripts/daily/build_digest.py`	Fetches sources, calls the LLM, writes the daily markdown.
`scripts/daily/sources.yml`	Sources and their sections. Edit to add / remove feeds.
`scripts/daily/requirements.txt`	Python dependencies pinned for CI.
`scripts/daily/state.json`	Maintained automatically — seen article GUIDs to prevent duplicates. Do not edit by hand.
`.github/workflows/daily-digest.yml`	Cron workflow (03:00 UTC+8) and manual trigger.
`_daily/YYYY-MM-DD.md`	Generated digests; committed by the bot.
`_layouts/daily.html`	Jekyll layout for a single day.
`_pages/daily.html`	Index page listing the most recent digests.
`daily.xml`	RSS feed of the last 30 digests.

First-time setup (once)

Go to repo → Settings → Secrets and variables → Actions → New repository secret.
Name: OPENROUTER_API_KEY
Value: your full OpenRouter key (starts with sk-or-v1-…).
Save.

That’s it — the nightly schedule will take over.

Trigger a run manually

GitHub → Actions tab → Daily AI Digest.
Run workflow. Optional inputs:
- dry_run = true: skip all LLM calls, emit stub summaries. Useful to verify the plumbing without spending tokens.
- digest_date: override the date (e.g. 2026-05-11).
The workflow commits _daily/<date>.md and an updated scripts/daily/state.json. GitHub Pages redeploys automatically.

Editing the sources

Open scripts/daily/sources.yml:

sections: controls the grouping headers and order on the rendered page.
sources: is a list; each entry has a kind (rss or arxiv) and a section.
For RSS sources, set fetch_full_text: true to scrape the article body with trafilatura before summarising. Skip it for feeds whose summary already has enough text (e.g. short announcements).
For arXiv sources, set the category (cs.AI, cs.LG, cs.CL, …) and a lookback_hours window.

Changes take effect at the next scheduled run.

Cost and rate limits

With ~100-150 items/day × ~500 tokens avg, the daily OpenRouter bill on deepseek/deepseek-chat stays under $0.10/day (roughly $3/month). Set a monthly spend limit on OpenRouter to be safe.

Troubleshooting

A source is marked unavailable on the digest page. The warning box at the top of the day’s page names the source and the error. Most commonly this is an RSS URL that changed or a flaky mirror — add a fallback: URL in sources.yml.
LLM calls fail with “model not found”. The script already falls back from deepseek/deepseek-chat to deepseek/deepseek-v4-flash. If both fail, check OpenRouter’s model catalogue and update OPENROUTER_MODEL / OPENROUTER_FALLBACK in the workflow.
State got corrupted / want to replay a day. Delete the offending guid entries from scripts/daily/state.json, or pass digest_date via the manual trigger.

Locally running a dry run

pip install -r scripts/daily/requirements.txt
DIGEST_DRY_RUN=1 DIGEST_DATE=2026-05-12 python scripts/daily/build_digest.py

No OpenRouter key needed in dry-run; summaries will be placeholders.

Runguo Li