Daily Telegram digest of arXiv papers with working code

Every weekday morning at 9am Eastern, send me a Telegram digest of the freshest arXiv papers that ship with working code I can actually clone and run.

Build this as an agent workflow. Trigger is cron, weekdays at 9am ET, and I should be able to change the schedule later.

Configurable inputs I want exposed when I save the workflow:

- watched_categories: arXiv subject categories to scan. Default cs.LG, cs.AI, cs.CL.

- focus_statement: a sentence or two describing what I care about reading. Default: "agent frameworks, retrieval, post-training, and evals". The agent uses this to rank papers.

- telegram_chat: where to post the digest. Default is my Saved Messages (peer "me").

- min_stars: minimum GitHub stars for the linked repo to count as alive. Default 5.

- recent_push_days: maximum days since the linked repo's last push to count as alive. Default 90.

- digest_size: how many papers to include. Default 3 to 5.

On each run, the agent should:

1. For each watched category, call arXiv's List Recent Papers in a Category to grab the last 24 to 48 hours of new submissions and updates. If I have optional keyword filters set, use arXiv's Search Papers with the keyword plus the category instead.

2. For every candidate paper, try to find an associated code repository. First scan the abstract text and the comment field for any github.com URL and parse owner/repo out of it. If at least one URL is present, prefer the one mentioned alongside words like "code", "implementation", "official", or "released", then call GitHub's Get a Repository on it. If no URL is in the metadata, fall back to GitHub's Search Repositories with the paper title (and optionally the first author's surname) as the query, take the top plausible match (good title overlap and recent activity), and call Get a Repository on that. If nothing turns up, drop the paper silently.

3. From the Get a Repository response, capture stars, last push date, and primary language. Mark the repo as alive only if stars are at least min_stars AND the last push is within recent_push_days.

4. Rank the surviving papers by: (a) fit to focus_statement, weighted highest, (b) repo liveness, (c) arXiv recency. Pick the top digest_size.

5. Post a single Telegram message to telegram_chat using Send Message. Format the body as one block per paper, separated by a blank line, with: a bold title, a one or two sentence "what's new" summary written from the abstract in plain English, a link to the arXiv abstract page, a link to the GitHub repo, and a small footer line like "⭐ {stars} · last push {humanized date} · {primary_language}".

Rules and edge cases:

- Papers without a discoverable code release are dropped silently. The whole point is "things I can clone today". Don't pad the digest with paper-only entries.

- Don't repeat papers. Remember which arXiv IDs were posted in the last 14 days and skip them on re-runs.

- Be polite to arXiv. Honor the 3 second between-requests pacing and back off on 429s.

- Keep the message under Telegram's 4096 character limit. If the formatted digest would overflow, drop the lowest-ranked paper rather than truncating mid-block.

- If zero papers survive the filters, still post a short "no runnable papers today" message so I know the workflow ran. Don't go silent on me.

This replaces my morning ritual of scrolling arXiv-sanity and Papers-With-Code looking for ML work that actually ships with code.