esh-pfi-infrastructure/stacks/news-digest

news-digest

Twice-daily LLM-curated briefing across Reddit (via JSON API) and Miniflux's Tech aggregators category. Output is a single static HTML page styled in the Australis design system with editorial-serif headlines (Fraunces) — operations-center chrome wrapping news content.

Why this stack exists

After a few subreddits + HN + Lobste.rs, the Miniflux inbox gets noisy. This stack:

1. Pulls top-of-day posts per subreddit from Reddit's public JSON API (gives us scores + upvote ratios — RSS doesn't).
2. Filters by score >= 50 and upvote_ratio >= 0.85 (configurable) to drop flame-bait and low-effort posts.
3. Pulls non-Reddit recent items from Miniflux's Tech aggregators category (HN, Lobste.rs).
4. Sends each source through qwen3.5-35-a3b on llama-swap (one batched call per source — efficient) for a one-sentence tl;dr + single-word tag (news / tutorial / release / discussion / question / showcase / drama / meme / other).
5. Renders an HTML page styled with Australis tokens + Fraunces serif headlines.
6. Static page is served by a tiny nginx:alpine. Cron writes /output/index.html atomically; nginx serves whichever copy is there.

Two editions per day: 8am and 8pm local. Plus per-edition archives at /edition-YYYY-MM-DD-{am,pm}.html.

Architecture

Two containers, both on traefik-net, sharing a bind-mounted output dir:

news-digest-worker  (python:3.12-alpine + cron)
  ├── busybox crond fires at 0 8,20 * * *
  ├── digest.py:
  │     ├── miniflux /v1/feeds → discover subreddits
  │     ├── reddit JSON top/.json?t=day per sub (gentle 1.5s sleep)
  │     ├── miniflux /v1/entries → tech aggregators
  │     ├── llama-swap /v1/chat/completions → batched per source
  │     └── jinja2 render → /output/index.html (atomic .tmp + rename)
  │                       → /output/edition-2026-04-26-pm.html
  └── style.css served from /output (copied at deploy)

news-digest-web     (nginx:alpine)
  └── serves /output as / on host port 8181
      └── homepage card via container labels (group=News)

Deploy

scripts/elway ana-docker --playbook playbooks/deploy-news-digest.yaml

After first deploy, fill in MINIFLUX_PASSWORD on the host:

ssh ana-docker '
  cd /opt/docker/compose/news-digest
  sed -i "s|^MINIFLUX_PASSWORD=.*|MINIFLUX_PASSWORD=<your-miniflux-admin-password>|" .env
  docker compose up -d
'

The container runs the first digest immediately if /output/index.html doesn't exist, so the page populates within a minute or two of bringing the stack up with real credentials.

Visit http://10.250.50.70:8181 to read.

Tuning the noise floor

Defaults in .env.example:

Knob	Default	Effect
DIGEST_REDDIT_HOURS	12	Look-back window (matches twice-daily cadence)
DIGEST_MIN_SCORE	50	Reddit minimum upvotes to consider
DIGEST_MIN_RATIO	0.85	Reddit minimum upvote ratio (skips flamebait)
DIGEST_MAX_PER_SUB	8	Cap per subreddit, post-filter
DIGEST_MINIFLUX_HOURS	12	Look-back window for HN/Lobste.rs
DIGEST_MINIFLUX_MAX	8	Cap per non-Reddit feed

For a busier day, lower DIGEST_MIN_SCORE. For a quieter morning edition, raise it. Edit .env, no rebuild needed — the worker reads env on each cron fire.

Adding more subreddits

The digest picks up subreddit feeds from Miniflux automatically — any feed whose URL starts with https://www.reddit.com/r/<sub>/ gets queried. To add a sub, just subscribe in Miniflux (UI or API). The next digest run includes it.

Updating the LLM model

ssh ana-docker '
  cd /opt/docker/compose/news-digest
  sed -i "s|^LLAMA_SWAP_MODEL=.*|LLAMA_SWAP_MODEL=<new-model>|" .env
  docker compose up -d
'

The model must be loaded in llama-swap's config.yaml. Check http://10.250.50.54:9292/v1/models for what's available. Models with tool/JSON-mode support give better summarization quality; qwen3.5-35-a3b is the current default.

Forcing a fresh digest now

ssh ana-docker 'docker exec news-digest-worker python3 /app/digest.py'

Runs the full pipeline once, ignoring cron. Useful after changing filtering knobs or adding feeds.

Logs

ssh ana-docker 'docker logs --tail 100 news-digest-worker'

Worker logs each phase (subreddit discovery / fetching / summarizing / rendering) with timestamps. Per-source LLM filter results show how many items were kept vs skipped.

License + attribution

Reddit content surfaced here is owned by its authors and Reddit. The digest is a derived index pointing at original sources — every item links back to the Reddit thread (and to the external link if the post linked out). Same for HN / Lobste.rs.