World Digest — a self-evolving daily news digest

A proof-of-concept, always free, self-steering system. Every day it fetches RSS feeds from outlets across the world, deduplicates the stories, summarizes them with an LLM, and emails you a "State of the World" digest. Every night, a separate workflow asks an LLM to propose one improvement to its own pipeline code — which is merged only if the frozen test suite passes, and auto-reverted otherwise.

Architecture

/app
  core.py          <- the evolvable pipeline (fetch -> cluster -> summarize -> email)
  config.json      <- optional config override (created by evolution over time)
/tests
  test_core.py     <- FROZEN fitness definition; the system can never edit this
/evolution
  evolve.py        <- mutation engine (one change per night)
  prompts.md       <- the goals the mutation LLM is given
  metrics.json     <- performance history (the system's memory)
  attempts.log     <- record of merged/reverted/rejected mutations
/.github/workflows
  digest.yml       <- daily digest run (07:00 UTC)
  evolve.yml       <- nightly evolution run (03:00 UTC)

Why it's built this way

• Tests are sacred. tests/ encodes your intent. CI fails any mutation that touches tests/ or .github/. To change the system's goals, you edit the tests.
• One mutation per cycle. Failures stay diagnosable, reverts stay clean.
• Git history is memory. Every accepted/reverted mutation is a commit; the mutation LLM sees recent failures so it doesn't repeat them.
• Hard caps everywhere. Workflow timeouts, token limits, and a kill switch.

Setup

1. Create a new GitHub repo and push these files.
2. In repo Settings → Secrets and variables → Actions, add:
   • HF_TOKEN — a Hugging Face token with "Inference Providers" permission, used for summarization + evolution (optional for digest: without it the digest runs in free extractive mode)
   • SMTP_USER — your Gmail address
   • SMTP_PASS — a Gmail App Password (Google Account → Security → 2-Step Verification → App passwords), not your real password
   • DIGEST_TO — where the digest is sent
3. In repo Settings → Actions → General → Workflow permissions, enable Read and write permissions (so workflows can commit metrics/mutations).
4. Test manually: Actions → Daily Digest → Run workflow.

Kill switch

Create an empty file named EVOLUTION_PAUSED in the repo root. Evolution stops immediately; the daily digest keeps running. Delete the file to resume.

Your weekly role (~5 minutes)

• Skim the commit log: what did it change, what got reverted?
• Check evolution/metrics.json trends: items_fetched, regions_covered, digest_words.
• Occasionally add a new test — that's how you apply new evolutionary pressure.

Optional: close the feedback loop

Reply to the digest email with a rating like RATING: 4. A small script (your first manual extension!) can read replies via IMAP and append user_rating to metrics.json. The evolution prompt already tells the mutation engine to weight that signal heavily — it turns 10 seconds of your day into directed evolution.

Local test

pip install -r requirements.txt
pytest tests/ -q          # fitness check
python app/core.py        # runs the pipeline; prints digest if SMTP unset

Known limits (by design, for a PoC)

• Token-overlap clustering is crude; an obvious evolution target.
• Free-tier API rate limits cap evolution speed to one mutation/night — fine.
• The system will plateau. When it does, study why — that's the lesson.