AI News Watcher

A small agent that sends one well-chosen WhatsApp message per day to keep me current on the AI industry — leaders, labs, tools, concepts. If something genuinely big breaks, that takes over the daily message instead. Replies are threaded: ask follow-up questions and Claude answers, with optional web search.

Architecture mirrors release-watcher: Railway cron → Python module → Postgres → Brave + Anthropic → Twilio. See .claude/plans/i-want-to-brainstorm-delegated-thacker.md for the full design rationale.

Status

Live in production on Railway.

• Inbound webhook: https://ainewswatcher-production.up.railway.app/sms/inbound (Twilio "When a message comes in" URL for the WhatsApp Sender)
• Daily fact crons: daily-fact-weekday (Mon–Fri 6:45 PM MST), daily-fact-weekend (Sat–Sun 1:00 PM MST). Arizona = UTC−7 year-round.
• Seeds in DB: ~80; refill via generate_seeds when low.

Tech stack

Concern	Choice
Language	Python 3.11+
HTTP	httpx
DB	Postgres (Railway in prod, local brew in dev), SQLAlchemy 2 + Alembic
Web search	Brave Search API
LLM	Anthropic SDK, Sonnet 4.6
Messaging	Twilio WhatsApp (whatsapp: prefix on FROM/TO)
Web service	FastAPI + uvicorn (inbound webhook only)
Hosting	Railway (3 services: 2 crons + 1 always-on webhook)

Quick start (local)

python -m venv .venv && source .venv/bin/activate
pip install -r requirements.txt
cp .env.example .env   # fill in secrets
alembic upgrade head
python -m watcher.jobs.sync_seed
python -m watcher.jobs.daily_fact --dry-run

Jobs

• python -m watcher.jobs.daily_fact — the one daily entrypoint. Flags: --dry-run, --force-news.
• python -m watcher.jobs.sync_seed — load config.yaml seed list into the DB.
• python -m watcher.jobs.generate_seeds --count 25 — ask Claude for new seed entries, append to config.yaml for review, then re-run sync_seed to persist.

Operational tips

Trigger a one-off run against the Railway DB from your laptop:

AI_NEWS_DATABASE_URL='postgresql+psycopg2://...kodama.proxy.rlwy.net.../railway' \
  python -m watcher.jobs.daily_fact

(Public Postgres URL is in Railway: Postgres service → Variables → DATABASE_PUBLIC_URL.)

Inspect recent activity (Postgres):

-- Last 5 sends + their text
SELECT sent_at, fact_kind, left(sms_body, 100) FROM sent_facts ORDER BY sent_at DESC LIMIT 5;

-- Cron run history (success / error)
SELECT run_at, status, fact_kind, left(detail, 100) FROM send_log ORDER BY run_at DESC LIMIT 10;

-- Current reply threads
SELECT root_sent_fact_id, jsonb_array_length(transcript) AS turns, last_message_at
FROM sms_threads ORDER BY last_message_at DESC LIMIT 5;

Add more seeds when running low:

python -m watcher.jobs.generate_seeds --count 25  # writes to config.yaml
# review the diff, then:
AI_NEWS_DATABASE_URL='...railway-public-url...' python -m watcher.jobs.sync_seed

Inbound replies (stretch goal)

Replies on WhatsApp are threaded onto your most recent sent_fact. Claude has access to a brave_search tool (max 2 calls per reply) and answers in SMS-length form. The webhook (watcher/web.py) verifies Twilio signatures and dispatches the LLM work as a BackgroundTasks so the HTTP response returns immediately.

Whitelist: only the configured SMS_TO_NUMBER is allowed to trigger replies. Everything else is silently dropped.

Schedule

Configured per-service in railway.toml / railway.weekday.toml / railway.weekend.toml. Arizona = MST year-round; cron expressions in UTC.

• Weekday cron: 45 1 * * 2-6 (Tue–Sat 01:45 UTC = Mon–Fri 6:45 PM MST)
• Weekend cron: 0 20 * * 0,6 (Sat–Sun 20:00 UTC = Sat–Sun 1:00 PM MST)