tentac

Live site: https://baiddd.github.io/tentac/

A weekly digest of what happened in AI — papers, journals, lab blogs, and security research, picked and summarized by Claude. No ANTHROPIC_API_KEY anywhere: fetching is automated, the AI classification step is done locally via Claude Code, and the result is reviewed as a normal pull request before it goes live.

(The codebase and Python package are still internally named ai-weekly — that's the pipeline's name, not the site's brand. Only user-facing text — page titles, headings, the RSS feed — says "tentac".)

• config/sources.yaml — every source, grouped by section, with a tier
• pipeline/ — fetch → score → build
• data/ — one committed JSON file per week; the archive is the git history. prefiltered/<week>.jsonl reflects the prefilter logic as of whenever it was last (re)generated — if relevance_keywords gets tuned after a week ships, that week's prefiltered file won't necessarily still contain every item in its already-published <week>.json/scored/<week>.jsonl. The published issue itself is never affected by re-running prefilter.
• web/ — Astro site published to GitHub Pages
• .claude/skills/weekly-analysis/ — the local, no-API-key classification step. Gitignored (internal tooling, not part of the public repo) — you need your own local copy to run it.

How a weekly issue happens

1. Automatic (GitHub Actions, every Monday 06:00 UTC, no secrets needed): fetch.yml runs pipeline/fetch.py then pipeline/score.py --stage prefilter (dedupe + prefilter — pure code, no LLM), and opens a PR on a branch named issue/<week> with data/raw/<week>.jsonl and data/prefiltered/<week>.jsonl.

2. Local, via Claude Code (no API key — uses your own Claude Code session):

   git fetch && git checkout issue/2026-W34

   Open Claude Code in the repo and run /weekly-analysis (requires your local, gitignored copy of the skill — see above). It classifies and scores every item, checks that each source link actually resolves (drops dead links), ranks them, writes the headline, and assembles data/2026-W34.json. Review the diff, commit, and push to the same branch.

   Before finishing, optionally check the source config/sources.yaml documents as blocked with no automated fix (currently openreview — active bot challenge; see its note:). If something standout turns up, hand-add it to data/raw/<week>.jsonl as a RawItem-shaped JSON line (see pipeline/models.py) before running /weekly-analysis, so it flows through prefilter/classification/ranking like everything else.

3. Merge the PR. Merging to main triggers deploy.yml (GitHub Actions), which builds the Astro site and publishes it to GitHub Pages automatically — no manual deploy step.

Local run (manual, for testing a single stage)

python -m venv .venv && .venv/Scripts/pip install -r requirements-dev.txt   # or .venv/bin/pip on macOS/Linux
.venv/Scripts/python pipeline/fetch.py --week 2026-W34        # or --date 2026-08-24
.venv/Scripts/python pipeline/score.py --week 2026-W34 --stage prefilter
# then run the /weekly-analysis Claude Code skill, or manually:
.venv/Scripts/python pipeline/score.py --week 2026-W34 --stage rank
.venv/Scripts/python pipeline/build.py --week 2026-W34 --headline "..."

classify_and_score() (in pipeline/score.py) and write_headline()/ write_title() (in pipeline/build.py) still exist and are fully unit-tested — they call the anthropic SDK directly and need a real ANTHROPIC_API_KEY. Nothing in this repo's default workflow calls them; they're there for anyone who later wants to fully automate the classification step in CI instead of doing it locally via Claude Code.

Setup (one-time)

1. Enable Pages: Settings → Pages → source GitHub Actions.
2. Trigger Fetch weekly issue manually once (gh workflow run fetch.yml) to confirm the PR-opening flow works, before relying on the Monday cron.