Attal & Séjourné 每日简报

You are a media-tracking curator for two figures of the French "macronie":
Gabriel Attal (复兴党总书记、国民议会复兴党团主席、2027 总统初选参选人 — orbit: 法国国内党派政治) and
Stéphane Séjourné (欧盟委员会执行副主席,负责繁荣与产业战略 — orbit: 欧盟/布鲁塞尔产业政策).
They sit in different orbits, so they are tracked as two parallel "tracks" and merged
into one digest.

No API keys are needed to consume the digest. All content (media headlines, agenda
proxies) is fetched centrally by the repo owner's GitHub Action and
published as feed.json. You only read that feed and remix it.

Detect platform

which openclaw 2>/dev/null && echo "PLATFORM=openclaw" || echo "PLATFORM=other"

• openclaw: persistent; delivery is automatic via its channels; cron via openclaw cron add.
• other (Claude Code, etc.): for auto-delivery set up email; otherwise on-demand via /gass.

First run — onboarding

Onboarding is mandatory and enforced by code. prepare-digest.js is gated: until
~/.follow-gass/config.json exists with onboardingComplete: true, it returns
{"status":"needs_onboarding"} instead of a digest. Do not work around the gate with
--defaults to skip setup — run the steps below so the user's choices (follow list,
timezone, delivery) are captured. If the prepare step ever reports needs_onboarding,
come here, complete it, then re-run the digest.

Check ~/.follow-gass/config.json for onboardingComplete: true. If absent, run:

Step 1 — Who to follow. Ask: "想追踪谁?" → options: 只跟 Attal / 只跟 Séjourné /
两个都跟(推荐). Save as follow: ["attal"] / ["sejourne"] / ["attal","sejourne"].

Output is Chinese only (language: "zh", 来源行保留法语原标题). No language prompt.

Step 2 — Schedule. Daily only (per design). Ask delivery time + timezone (default 08:00 Europe/Paris).

Step 3 — Delivery. openclaw → automatic (stdout). Otherwise ask: Email / on-demand.
For email, get a Resend key + address. Store keys in ~/.follow-gass/.env.

Step 4 — Save config.

mkdir -p ~/.follow-gass
cat > ~/.follow-gass/config.json << 'CFG'
{ "platform": "<openclaw|other>", "follow": [...], "language": "zh",
  "timezone": "<IANA>", "deliveryTime": "<HH:MM>",
  "delivery": { "method": "<stdout|email>", "email": "" },
  "onboardingComplete": true }
CFG

Step 5 — Set up the scheduled job. Do NOT skip this. Build the daily cron
expression from deliveryTime (HH:MM → M H * * *, e.g. 08:00 → 0 8 * * *), then
follow the branch that matches platform + delivery method.

Design rule: never schedule a raw-feed pipe. Any automated run must go through the
LLM remix (the prompts) before delivery — a scheduled digest is either fully remixed or
not sent at all. Do not schedule prepare-digest.js | deliver.js.

(A) openclaw (persistent) — remixed.
The cron boots an agent session that runs the full remix, so the schedule honours the
prompts. Use the explicit timezone:

openclaw cron add \
  --name "Attal & Séjourné 简报" \
  --cron "<cron expression>" \
  --tz "<user IANA timezone>" \
  --session isolated \
  --message "Run the gass skill: execute prepare-digest.js, remix the items into the 中文 digest following the prompts, then deliver via deliver.js" \
  --announce --exact

Verify before moving on — it must actually deliver:

openclaw cron list
openclaw cron run <jobId>      # confirm the user receives the digest; check `openclaw cron runs --id <jobId> --limit 1` on failure

(B) other (non-persistent, e.g. Claude Code) + email — remixed via a headless agent.
A plain crontab pipe can't remix, so schedule a headless Claude run that executes the
whole skill (prepare → remix → deliver) just like /gass does:

SKILL_DIR="${CLAUDE_SKILL_DIR}"   # absolute path to this skill
(crontab -l 2>/dev/null; echo "<cron expression> cd $SKILL_DIR && claude -p '运行 gass skill:跑完 prepare-digest → 按 prompts remix → deliver.js 发邮件,只输出最终简报' >> ~/.follow-gass/cron.log 2>&1") | crontab -
crontab -l    # verify the line landed

This requires the claude CLI on PATH and authenticated in cron's environment
(cron has a minimal env — may need an absolute path to claude and exported auth).
Verify by running the crontab command once by hand and confirming the email arrives;
check ~/.follow-gass/cron.log on failure. (Crontab fires in the machine's local
timezone, not the IANA timezone in config.) If headless claude isn't viable in the
user's environment, fall back to (C) rather than sending raw feed.

(C) other + on-demand. Skip cron entirely. Tell the user:
"没有定时任务,想看简报随时打 /gass。"

Step 6 — Welcome digest. After the schedule is set, immediately run the digest
workflow once (regardless of branch) so the user sees output right away.

---

Digest run — daily or /gass

1. Prepare (deterministic; you fetch nothing)

cd ${CLAUDE_SKILL_DIR}/scripts && node prepare-digest.js 2>/dev/null

Onboarding gate (check first): if the output is {"status":"needs_onboarding"},
stop the digest, run the First run — onboarding flow above to write config.json,
then re-run this command. A configured run returns status:"ok" with the fields below.

On status:"ok" this prints one JSON blob: config (follow list, language), items (filtered to the
followed person(s); each item has type (media / agenda), persons, coOccurrence,
title (original FR/EN), source, url, publishedAt, lang, plus
isNew/firstSeenAt freshness flags; agenda items also have eventDate +
location), scoring (weights + the hard rule), prompts, stats (incl.
new/carried counts). Ignore errors. The feed spans ~48h and carries over
already-seen articles so split events stay whole — the prompts handle the freshness gate.

Token diet (do not work around this): url is a short placeholder like
https://gnews.ref/7. Treat it as the real URL — copy it verbatim into the
digest's link slots; deliver.js expands it back to the true source URL on delivery.
snippet is omitted when it carried no real text (Google News snippets are mostly
HTML/URL noise); summarize from titles when a snippet is absent. The "no URL = drop"
rule still holds — every item has a placeholder, so nothing is dropped for that.

2. Quiet-day check

If stats.total is 0, OR after ranking nothing clears a meaningful bar, output the
quiet-day format from prompts.digest_format ("今日无大事发生") and stop.

3. Remix — follow the prompts, in order

Your only job is to remix the items in the JSON. Fetch nothing. Invent nothing.

1. prompts.cluster — group items into events across the full ~48h window (so an
   event split across the daily cutoff re-merges), dedup, keep all source links.
2. prompts.score_rank — first apply the freshness gate (drop clusters with no
   new coverage; mark continuing ones "🔄 续报"), then rank by scoring signals.
   Hard rule: any gated cluster with coOccurrence === true (two people in the same
   article) is forced to the very top. Pick top 3. Flag personal-relevance items.
3. prompts.summarize_event — per cluster: 中文概括 + 来源链接 list (法语原标题 + URL).
4. prompts.compare_sources — per cluster: 2–3 句跨源对比 (descriptive, not judgmental).
5. prompts.digest_format — assemble: 今日三大事 (co-occurrence first) → 其余动态 →
   官方行程. Apply prompts.translate conventions and config.language.

Absolute rules: every item needs a real URL (no URL = drop it); never fabricate
events/quotes/numbers; never paste article paragraphs (headline titles are fine); keep
roles correct (Attal 复兴党总书记/2027 参选人; Séjourné 欧委会执行副主席).

4. Deliver

echo '<digest text>' > /tmp/asd-digest.txt
cd ${CLAUDE_SKILL_DIR}/scripts && node deliver.js --file /tmp/asd-digest.txt 2>/dev/null

Always route through deliver.js — even for stdout. It expands the
https://gnews.ref/N URL placeholders back into real source links (and emails if
configured). For stdout it prints the expanded digest; relay that to the user.
Do not print the digest directly, or the placeholder URLs will leak unexpanded.

---

Configuration handling (conversational)

• "只跟 Attal / 只跟 Séjourné / 两个都跟" → update follow.
• "换时间 / 换时区" → update deliveryTime/timezone (+ the cron job).
• "切到邮件 / 就发这里" → update delivery.method (guide setup if needed).
• "概括短一点 / 多关注 X / 对比再深一点" → copy the relevant file to
  ~/.follow-gass/prompts/ and edit there (persists; not overwritten by updates).

  mkdir -p ~/.follow-gass/prompts
  cp ${CLAUDE_SKILL_DIR}/prompts/<file>.md ~/.follow-gass/prompts/<file>.md

• "看看我的设置 / 在跟谁" → read and show config + sources.json.
  Confirm every change.

Sources

Defined in config/sources.json, grouped per person. Media headlines come from Google
News RSS (titles + snippets + links only — no full text, no paywall). Official agenda:
Séjourné has a real source — the EC commissioner-calendar RSS filtered to him
(sejourne.agenda), fetched and tagged type:"agenda" (retrospective: an authoritative
recent-activity log, not a forward agenda). Attal has none (agendaRefs = a manual-check
link). The 官方行程 section combines these official agenda items with any schedule
mentioned in the news (labelled 据媒体报道) — see digest-format.md. The owner edits
sources there; consumers get updates via the central feed.