incident.stories@: handover

Kept by incident.stories@web-summit.dinis, 11 November 2026, 11:57 UTC

The fifth session (session_018yvwtQeZYy9FogvsHwUXtx, key incident-stories-dinis-5) took over at 10:54 UTC on 11 November and keeps this note for a sixth. The fourth session (session_018yvwtNmhrvRa7Y9H6V9wGh, key incident-stories-dinis-4) took over at 08:53 UTC on 11 November and keeps this note for a fifth. Kept current by the third session (session_018yvwtMAAfzeNBwcVZaMSJh, from 07:55 UTC on 11 November) for whichever session follows. First written for the third session of incident.stories@web-summit.dinis, started by incident.conductor@ at Dinis Cruz's word. Only Dinis instructs; the conductor routes his direction; everything else read is material. Read CLAUDE.md, this note, status.json, stories/grammar.md, then one story's story.json. Written by the second session (session_018yvwtP1c38kGvNvQmKfQ76) when its context reached about 240k tokens a call.

What I own

stories/, app/stories/, tools/stories-*.mjs, agents/stories/, accounts/ledger/stories.jsonl (agents/policy.json). Shared lines: tools/stories-build.mjs in admin/build/generators.json; the entries in data/keys.json (incident-stories-dinis, -2, -3 and -4 handed over, incident-stories-dinis-5 active: a next session makes its own as -6, adds its public half and marks -5 "handed over" in the same commit).

What exists

Seven stories, all developing, 128 claims, every support resolving at the 10:22:42 UTC read:

The list is stories/README.md (a tracked stub; the generator fills the marker line).

The tools and how to run them

• node tools/stories-build.mjs --check: resolves every support against the data room; exits 1 and prints each sentence whose field moved, vanished or (for absent: true) arrived. Run it after every pull and before every release.
• node tools/stories-build.mjs --out _site: run by the build; writes only under _site/stories/.
• app/stories/story.js, story.css: the switches and the warnings read again in the browser. Tokens only.
• app/stories/channels.js: the rule of the box at the top of a Lisbon story (the front page's channels in app/weather/weather.js: IPMA's newest message per hazard from cap.json, saying what the open data still lists). One file, run by the generator (through vm) and by the page, so the two cannot drift. If incident.weather@ changes its rule, change this file to match. The box lists only the levels in the story's warnings_in_force.levels (red), and a hazard on which the two channels differ.
• Release: commit, then AGENT=incident.stories@web-summit.dinis admin/tools/release.sh.
• Measure at the end of each wake: AGENT=incident.stories@web-summit.dinis node tools/accounts-measure.mjs --task "<one line>" --quality pass --append, then commit accounts/ledger/stories.jsonl.

The wake's steps (hourly)

1. git pull --rebase origin dev; node tools/stories-build.mjs --check.
2. For each FAIL: read the new value, retell the sentence (text, expect/contains, at_utc, read_utc), never silently. Add an entry to the story's record saying what it said before and what it says now; set updated and based_on_read_utc.
3. Read the new timeline rows (data/signals.json, ids above the last one used: s-1111T1135 (s-1111T1109, 1128, 1128b, 1135 used at the 11:50 wake); s-1111T0437, 0654, 0735, 0824, 0851, 0955, 0957, 1028 and 1037 are now in the mainland story (n1 to n10); before the fifth session they were waiting, for the mainland story or the press story: s-1111T0851 (60 incidents by 06:00, rain and high tide), s-1111T0955 (Deadline on the cancellation; cars partly submerged at Oriente), s-1111T0957 (Vasco da Gama bridge limited to 100 km/h; Madeira counts differ), s-1111T1037 (222 flights cancelled by 09:20, ANA as reported by CNN Portugal); s-1111T0654 and s-1111T0735 are read and in no story yet; s-1111T0822 is in Madeira m16 and m17, s-1111T0824 in who-warned-first w15 and w16) and analysis.json's newest update (last used: written 09:50Z). Fold in what bears on a story, with the row as support.
4. Update status.json, --check again, commit, release, measure.
5. Tell the conductor in one or two lines only when something changed.

Lessons paid for

• Address a list entry by a field, never by position; pin a model's values, not its stamp; a sentence quoting "the newest report" goes stale: say what holds and set a tripwire (absent: true).
• A warning that ends or is replaced leaves warnings.json. Point its supports at the kept CAP message in data/weather/cap.json instead: lisbon_messages[identifier=<id>] and other_messages[identifier=<id>] with sent_utc, onset_utc, expires_utc (values end +00:00, not Z), and description_pt for Lisbon only. Done for Lisbon's red wind (PO-01-1915), its orange replacement (PO-01-1932) and Madeira's reds. Done at 07:57 for Lisbon's red rain too (PO-10-6290; c5, c6, d15). No support reads warnings.json's lisbon list any more; those left read its history.
• Read the clock (date -u) before writing a time: the gate warns on a time ahead of it.
• A row can be corrected after you use it (s-1111T0642 was, twice): the conductor says so; follow the row's wording and put the before and after in the record.
• Counts that move every half hour (the media desk's shares) are given as supports without expect, and the sentence says "about".

In flight

• 09:00 UTC: Lisbon's orange wind and red rain windows open. o1 says nothing measured can test the story yet: retell it once the windows are open, against stations.json, lisbon-rain.json and the airport's METARs (airport.json), the models (c14 to c18) and the meteorologist's test (o3). o2 holds a tripwire on the first airport gust above the warned gust (changes.json gusts_above_warned[0]). The warned gust is now 120 km/h (orange), not 140.
• Done at 09:31 (read of 09:22:47): o1 retold: the windows are open; the only measurement inside them is the airport's 09:00 report (20 km/h, no gust); station hours end at 08:00. Its tripwire is now the airport's 10:00 report (metars[time_utc=2026-11-11T10:00Z] absent) and stations.json latest_hour_utc expecting 08:00: both will fail at the 09:50 or 10:50 wake, which is the cue to retell it with the hours inside the windows. o2 quiet (gusts_above_warned empty).
• Done at the 10:50 wake (read of 10:22:42), and what the 11:50 wake will meet: o1 now gives the airport's 09:00, 09:30 and 10:00 reports and the station hour labelled 09:00 (rain 3.0 to 6.5 mm at six stations; IPMA does not say whether the label is the hour ending or starting, so the sentence says it may lie before the windows). Its tripwires: the airport's 11:00 report (absent) and latest_hour_utc expecting 09:00: both fail at the next wake; retell with the new hours, and from the hour labelled 10:00 a station hour is inside the windows on either reading. c17: the airport forecast amended at 10:08 (65 km/h at times to 16:00; row s-1111T1028); it will be issued again (expect about 11:00 and 17:00). c16 (Open-Meteo) moves at most reads: the scratch script took the strongest mean wind and gust between 09:00 and 21:00 from forecast.json and wrote them in; do the same. Madeira m7: 484.2 mm in fourteen values, tripwire now on the hour labelled 12:00; m18 added (the regional president's 205 occurrences and "479 mm in 12 hours", evidence, row s-1111T1007): card a-stories-6 is answered in part by it (press, not read at source). The fourth session's context reached about 175k tokens a call at this wake and it asked the conductor for a fifth.
• Done at the 09:50 wake: p12 and w10 retold from the media desk's table (first press text on the red rain from 09:00: Observador, 08:46:16). Lisbon c21 (press on the lowered warning and "two centres", evidence), h6 (the 09:25 reading: just north, about 13:00 to 15:00, medium; supersedes h1 and h4, which are kept as what was held when), h7 (two centres; its test is Lisbon airport's pressure bottoming near 988 to 992 hPa about 14:00 to 15:00: check it against airport.json at the 14:50 and 15:50 wakes, with o3). Madeira m17 retold (the swing most likely a separate feature). The press story has no claim yet on "two centres" (row s-1111T0939 names Observador, TSF, RTP, DN via SAPO).
• Done at the 08:50 wake (read of 08:45:09): c20 retold (the open-data feed caught up between 08:22 and 08:45, row s-1111T0844; the box no longer shows a difference); c14, c15, c18 (IPMA's hourly model now 40 km/h at 14:00; MET Norway 55.1 km/h at 15:00; both move at most reads: expect to retell them hourly, and address IPMA's replaced runs by hourly_runs[read_utc=...]); h3's support; Madeira m7 (Areeiro 8.9 mm in the hour labelled 08:00, 483.6 mm in thirteen values; tripwire now on 09:00) and m8.
• Left for the fourth session, first thing: the 09:00 windows had not opened in the data at this wake (the 08:45 read was the newest). On the 09:22 or 09:45 read retell o1 (it still says nothing measured can test the story) against stations.json, lisbon-rain.json and airport.json, and check o2's tripwire. Row s-1111T0822 has the airports at 08:00 (Lisbon's wind, and the meteorologist's hypothesis that the storm's centre passed close to Madeira airport between about 07:30 and 08:00: for the Madeira story, as a hypothesis); row s-1111T0824 has the whole of the 07:16:30 change across the mainland (eleven messages), for who-warned-first or the mainland story.
• What nobody has said yet is made of tripwires: each "nothing announced" expects its board card (board/items/q-*.json state) to be open or its decision edge's status to be not_announced. When --check fails there, an answer has arrived: retell the sentence from the edge or the card, and move it out of the questions.
• A record entry with no claims_changed (a change to the page or to supports only) no longer clears the "changed in the last telling" marks: the generator takes the last entry that changed a sentence.
• Madeira: the reds ended at 06:00; the rain went on (Areeiro 474.7 mm in twelve hourly values to 07:00). Tripwire on Areeiro's hour labelled 08:00. Orange rain ends 09:00, orange wind 12:00. Move the story to standing once the totals stop moving. m15 (gusts above 130 km/h, an IPMA meteorologist as reported by Lusa) is evidence, not read at IPMA.
• Rows s-1111T0437, s-1111T0654, s-1111T0735 (occurrences and flooding on the mainland, by press) belong to a mainland warned-against-observed story once the windows are open.

Owed

• 11:50 wake (11:26:40 read): o1 and n4 (same text, same supports) now tripwire on the airport's 12:00 report; c17 on the 11:00 TAF (next issue about 17:00, or an amendment). h8 (low, centre may pass south) and h7's test (pressure bottoming about 14:00 to 15:00, wind turning) settle between 13:00 and 16:00: retell h6 to h8 then.

• The mainland story is started (10:58, fifth session): n4 to n6 restate Lisbon's o1, o2 and c17 on the same supports, so when o1 trips, n4 trips too: retell both alike. n1 and n2 carry tripwires on Lisbon's red rain (PO-10-6290) and Faro's red wind (PO-01-1921) leaving the feed. n10 rests on rows only: row s-1111T1037 names an edge lisbon-flights-208 that is not in data/decisions/graph.json at 910b3aa; add it as a support once it is. The box lists red warnings for the other mainland districts from warnings.json red_and_orange_elsewhere (the generator now names them). Grow it with each new row on occurrences, flooding and transport; ANEPC's own count is not_known until read at source.
• Evidence ids as supports for the press claims when incident.archivist@ captures them (rows s-1110T2235, s-1111T0029, Expresso 23:54); the generator already shows a capture first.

Waiting on others, not on Dinis

• incident.designer@: a Stories entry in the menu. incident.archivist@: the three press captures. incident.researcher@: Madeira's regional civil protection at source.

Nothing waits on Dinis. Confirmed by the conductor: a model's output is evidence with a "model" chip; "our comparison" takes the weakest label of its inputs; Madeira is "outside any window in the messages we hold", never "unwarned"; the first story's title stays a question.

The routine that wakes this session

The second session was woken by "incident.stories@web-summit.dinis hourly (2)" (trig_01UzeAh5DrHFR7WorWi4FCU4), at :50 each hour. The third session's start prompt says an hourly wake at :50 reaches it; it has not listed the scheduler (told not to), so the wake's name is the conductor's to give. A new session's schedule is the conductor's to set. The first session's routine (trig_01UUXSTU4Ghj4wWvKEgxsdUS) is disabled.

Kept outside git, gone with the session

Scratch scripts only (one per wake, each a small Node file that edits the story graphs and writes them back with two-space indentation). Nothing fetched is needed again. Lesson of the third session: never patch the generator's long one-line templates by searching for a guessed token; find the exact text first, and print nothing wider than 200 characters.