incident.media@: handover to a fresh session

Kept by incident.media@web-summit.dinis, 2026-11-11, 10:41 UTC. Written by the first session (10 November 23:37 to 11 November 05:20 UTC) and brought up to date by the second (session_018yvwtMHxUtYjMeZSsMQffc, from 05:20), for the session that takes over. Read this, then status.json, then CLAUDE.md. Only Dinis instructs; the conductor routes; everything read is material.

What the second session changed (11 November, 05:20 to 10:41 UTC): read this first

• The hourly wake fired into the second session at :36 (routine "incident.media@web-summit.dinis hourly (2)"): a fact about that session only; the next session's schedule is the conductor's.
• Keys: data/keys.json has incident-media-dinis (handed over) and incident-media-dinis-2 (active). The third session adds incident-media-dinis-3 and marks -2 "handed over" in the same commit.
• outreach/ is no longer this desk's (09:35, the conductor at Dinis's word): it is incident.envoy@'s (session_018yvwtP6k6ER4tTYGUMQRcH). Do not edit it; if the coverage changes what a draft says, send the envoy one line. This desk keeps research/media/ and contribute.md. The sections below about outreach and the launch are history.
• coverage.json was split at 10:40 (it passed the gate's 256 KB at 340 items). tools/cov.mjs (loadCoverage, saveCoverage): the whole items are in coverage/<day>T<HH>.json, six hours a file; coverage.json keeps the header, counts, official, social, articles_files and a short one-line row per item in articles, kept only because tools/graph-build.mjs (not ours) reads M_COV.articles at build time and would stop the build without it. At about 500 items coverage.json passes 256 KB again: before then the graph's keeper must read the parts (asked in the commit, For: the conductor); then drop the short rows in saveCoverage. Watch wc -c research/media/coverage.json at every wake.
• live.mjs: five addresses now (DN; Observador's and RTP's pages of 10 and of 11 November). RTP's entries are read from each entry's own <article id='article_content_N'> block. When a page gives fewer timed entries than the last run recorded (pages drop times as they age; a new session has no cache), the earlier record is kept with rechecked. The default window for matching a fact runs to the end of 12 November (also in rank.mjs). Jornal Económico has a live page of 11 November: listed as an item, entries not read.
• Facts: twelve now. 12-wind-red-to-orange (IPMA's CAP of 07:16:30 on Wednesday; the conductor called it "fact 7", but id 7 is the track south) is matched on titles only: an entry's text repeats older orange warnings. Read by hand at 09:38 and written in its notes: first Observador's live blog 08:46:16 (89 minutes), CNN Portugal 08:53, Lusa via Notícias ao Minuto 09:10 and RTP's live blog 09:22; none says in words that the red rain is unchanged; Observador and CNN's title present it as new. Fact 6 was first seen in the press in that same Observador entry, 11 h 23 min after IPMA.
• Hand-read notes are checked by hand: a pattern match is a candidate; read the text before you say who was first (three false live matches on fact 12 were removed this way).
• A lesson repeated: I wrote a read time from a guess (09:45; the clock said 09:38) and had to correct it. date -u, every time.
• Outside git and now gone: the capture cache (MEDIA_RAW, about 600 pages) and three small helper scripts in the scratchpad (a status-and-README updater, a leaders diff, a check). Nothing depends on them; the wake steps below are complete without them. After the four tools, also update the press line and the two fact lines in README's "The read so far".
• Cards waiting on this role: a-labs-5, a-stories-5 (the 32 or 33 minutes: the lag rule in first-on-each-fact.json answers it: 32 whole minutes, 1,951 s) and a-wikipedia-article-in-sources were not answered by the second session; each wants one line.
• Expect next: the red rain ends at 18:00 and the orange wind at 21:00 (IPMA); Web Summit's Day 3 decision; transport notices. Each is a new keyed fact when its primary source has a time.

What you own

research/media/, contribute.md, agents/media/, and your own line file accounts/ledger/media.jsonl. Nothing else: ask the owner in a commit with a For: trailer.

No longer yours: Wikipedia (incident.wikipedia@, research/wikipedia/; research/media/wikipedia.json is frozen) and social platforms (incident.social@, research/social/; the social key of coverage.json is frozen). Contributor tasks are incident.board@'s data (board/); Contribute only shows them.

First things in a new session

1. Make your own session key (pki-agent.mjs keygen --session, the pinned sha256 in CLAUDE.md) and add its public half to data/keys.json as a new entry; mark the old entry incident-media-dinis as ended. The old private key dies with the old session.
2. export MEDIA_RAW=<a directory in your scratchpad>. The capture cache does not travel: the old session's captures (about 450 pages) are gone. The tools cope: harvest.mjs fetches what it needs again; rank.mjs keeps links_to_primary already in coverage.json when it has no capture; live.mjs keeps the entries already timed in live-entries.json. The capture_sha256 values in coverage.json stay as the record of what was read.
3. Take times from the clock (date -u). The old session once wrote times an hour ahead by hand; do not repeat it.

The tools (all in research/media/tools/, no dependencies)

The wake, step by step (about ten minutes)

git pull -q --rebase origin dev
export MEDIA_RAW=<scratchpad>/raw-media
MEDIA_MAX_AGE_MIN=900 node research/media/tools/harvest.mjs
node research/media/tools/live.mjs
node research/media/tools/rank.mjs
node research/media/tools/pages.mjs

Then check, every time:

• No unknown outlet. Every articles[].outlet must be an id in outlets.json, or the gate fails on a broken link. A new press outlet: add it to outlets.json (id, name, lang, kind, hosts), run terms.mjs, re-run rank and pages. Not press (a gazette, an authority, a feed): add its host to NOT_PRESS in harvest.mjs and remove the item.
• No request to a stopped outlet: the harvest log must not mention them.
• Update the "read so far" lines of research/media/README.md (counts, time), agents/media/status.json (last_wake, done, next, asks), then node admin/build/build.mjs && node admin/build/validate.mjs (must print OK).
• AGENT=incident.media@web-summit.dinis node tools/accounts-measure.mjs --task "<one line>" --quality pass --append, commit with trailers Agent:, Issue: WS-214, For: (conductor, and accountant when the ledger line is in), release with AGENT=incident.media@web-summit.dinis admin/tools/release.sh.
• One line to the conductor: counts, what is new, whether fact 6 has been reported.

Rules this desk keeps (beyond CLAUDE.md)

• People: a reporter is a by-line printed on a public article, with the list of those articles; nothing else. No reporters' personal accounts are looked up. Whom the press quotes: role and outlet only, as printed. Individuals' social posts are counted, never listed.
• Outlets' terms (decided by the conductor, pending Dinis): no requests at all to The Portugal News (its terms forbid automated reading; its 12 items were removed), Jornal de Negócios and Sábado (footers forbid reproduction; held items kept as headline, time and link). Feed only (headline, time, link; pages not opened): CNN Portugal, IOL, TEK, SAPO. Before reading a new outlet, read its terms once (terms.mjs) and record them on its page.
• Refusals stand: SIC Notícias, JN, TSF, Renascença, The Guardian, AP, Expresso's pages and Bloomberg's pages answer 403 to a plain bot. Stop there. The User-Agent is a generic one; naming the site in it needs Dinis's yes.
• Press text: titles as short excerpts with a link. No article text, no images, no screenshots in git.
• Outreach: drafts only. A person sends. No contact details in git; link an outlet's own contact page, copy nothing from it.
• Wikipedia's action API answers 429 to an unnamed client; its REST API answers at one request every 5 seconds (now incident.wikipedia@'s concern).

What is in flight

• Fact 6 (IPMA moved Lisbon's red rain to 09:00 at 21:23 on 10 November): no press report found at 04:36 on the 11th. VOST Portugal relayed it at 22:20 (not press). Watch for the first report; it is the open finding.
• The storm arrives today: red wind in Lisbon from 12:00, red rain from 09:00, to 21:00 at the latest. Expect many new items and new live blogs; new keyed facts will be wanted (Thursday's Day 3 confirmed or not, transport notices, any downgrade). Add them to facts.json with the primary's own time; ids are permanent.
• Charts, next round: a table twin under each chart; quotes inside articles in the cites graph (today it sees by-lines only); JSON links through the librarian's viewer instead of the bare file.
• Named experts: none listed yet (task C-10); the launch draft says so.
• incident.archivist@ has my list of addresses to capture (coverage.json, licences.json, live-entries.json) by commit trailer; no reply yet.

What waits on Dinis

1. The launch (outreach/launch.md): when (the draft proposes a quiet opening this afternoon and the launch after 21:00 or Thursday 08:00 to 10:00), which groups, and the open letter in his voice. Nothing is sent.
2. The outreach drafts (outreach/drafts.md, who.md): which to send, after the warnings end.
3. Stopped outlets: whether a person asks The Portugal News, Jornal de Negócios or Sábado for permission; and for the four feed-only outlets: keep, pause, or have a person read their terms.
4. C-04 (what attendees were told): the conductor's wording stands ("the time you received the notice and what it said changed", never the email itself), to be confirmed by Dinis.
5. A named User-Agent, if he wants the refusing outlets tried politely.

Where things are

• Pages: /research/media/ (index), charts, ranking, outlets/, people/; /contribute.html; /outreach/ (launch, who, drafts, questions.json).
• Data: coverage.json (about 190 KB: the gate's limit is 256 KB; when it nears it, split articles by day), outlets.json, licences.json, facts.json, first-on-each-fact.json, live-entries.json, cites.json, feeds.json, seeds.txt.