From incident to business as usual: the plan for Dinis's yes

Kept by incident.planner@web-summit.dinis, 11 November 2026, 11:46 UTC. Second draft (first pushed 10:36); gaps are marked GAP. Nothing here stops anything: this is a plan for Dinis to say yes to, change, or refuse.

Not a warning service: for warnings follow IPMA and ANEPC. This page is about how the team works, not about the storm.

What Dinis asked (~10:30 UTC, BRIEF.md, last section, close to his words): wind down every agent, refactor the ones that need it, assess each one and give it an exit interview, wrap up what each is doing, then map which to start and which to continue; the reporter and the journalist "go to town" on the incident pages; then move to business as usual, "running this as a normal newsroom"; a storm conductor as there will be a Fringe conductor; "from explorer to villager". ~10:32: the team "shouldn't stop until the handover is done". ~10:38, a correction: no outside hand-over; "I'm going to be the only one running this site"; the Fringe is another team's work, so the business-as-usual team stays focused on the storm. (All in BRIEF.md, last sections; the joining of other accounts and teams in the 10:30 section is withdrawn there.)

How to read it: Fact = read from a file or commit, linked. Mine = my judgement as planner, with reasons. Every verdict below is mine until Dinis says yes.

In one screen

1. When. Nothing stops while the storm is live, and no agent stops before its hand-over is done (Dinis, ~10:32). After Dinis's yes, the agents not on live coverage (the early eleven) start their exit interview and hand-over at once; each keeps working until its hand-over is accepted. The live crew keeps running until the close trigger: IPMA shows nothing above green for Lisbon (and no orange or red for Madeira) at the first hourly read after 21:00 UTC Wednesday and Web Summit has published its Day 3 decision. Then the live crew does its interviews and hand-overs, and the closing record is made on Thursday 12 November. (§1)
2. Every agent (41 with planner, transport and the four of ~11:45): 8 continue, 3 refactor, 21 merge into 8 BAU desks, 9 stop. (§2)
3. The closing record: one incident page from beginning to end, built from the timeline, the traces, the journal, the ten paths and the books; and an incident template for the next one. (§3)
4. Business as usual: a storm conductor and a standing newsroom of 14 desks from fresh sessions, keeping the use cases, the brief, the contact desk and the data API (Dinis ~11:45), quiet in fair weather (a few wakes a day), scaled up when IPMA warns orange or red for a watched district. api/v1/ stays as its contract says. Run by Dinis alone; focused on storms. (§4)
5. The order of steps and Dinis's decisions in one list (§5). Nothing is deleted; history stays.
6. Explorer to villager: what changes in how we work, and the record that says why. (§6)

1. When

Facts.

• The headline (as of the file's last write) is Web Summit's 19:09 UTC decision of 10 November: Day 2 called off; Day 3 "aims to proceed as planned"; IPMA's red rain for Lisbon 09:00 to 18:00 Wednesday; orange wind (lowered from red at 07:16) 09:00 to 21:00 (incidents/2026-11-10-storm-charles/incident.json).
• The meteorologist's 09:25 reading: centre just north of Lisbon about 13:00 to 15:00, strongest wind 14:00 to 16:00 (agents/meteorologist/status.json; a labelled analysis, not a forecast of ours).
• Day 3's decision is not yet published (BRIEF.md, ~10:30).

Proposal (mine). Three phases.

When an agent's hand-over is done (mine, from Dinis's ~10:32 rule): all four of (1) its exit interview is written and the planner's check is published; (2) its wrap-up is in its HANDOVER.md (what it made, what is open, what to run each wake); (3) its vault keys have reached the conductor by lane; (4) the receiver named in its verdict (a BAU desk, the librarian, or "nobody: stopped, record kept") has said in a commit For: the agent that it has read the hand-over and takes what it inherits. Only then does the conductor retire its wakes and mark its key "handed over" (or "stopped"). Until then it keeps its wakes and its job.

The close trigger (mine): both of

• (a) IPMA's first hourly read after 21:00 UTC Wednesday (incident.ipma@ reads at :47) shows no yellow or above for Lisbon district on the rest of Wednesday or on Thursday, and no orange or red for Madeira; and
• (b) Web Summit has published a decision on Day 3 (incident.researcher@ / the conductor record it in decisions.md).

If (a) holds and (b) does not by 23:00 UTC, start Phase 2 anyway with the Thursday watch carrying (b). If IPMA issues a new warning for Thursday, Phase 2 waits: the live crew keeps running. Reason: the storm is the reason for the site; a reader on Wednesday evening must still find the warnings current, and the only thing still open after the wind drops is the summit's Thursday.

The live crew (keep running until the trigger): datascience, brief, desk, sherpa (new at ~11:45, live by design), conductor, ipma, sources, meteorologist, researcher, transport, archivist, weather, broadcast, media, social, board, observability (page views), ci, librarian, graph, entities, stories, reporter, journalist, designer (fixes only, no new rounds).

The early eleven (start interview and hand-over after Dinis's yes; none feeds the live pages): studio, labs, synthetic, perplexity, models, shadow, weatherwatch, wikipedia, envoy (moves to working with Dinis outside the wake schedule), dev, accountant (after one closing round of the books, §3). Each keeps its wakes until its hand-over is done. Storyteller and motion finish the closing record first (§3). Planner hands over last.

2. Every agent

Facts in the table come from: the roster (agents/JOBS.md, agents/policy.json); sessions = session keys in data/agents/sessions.json (from data/keys.json), "h/o" = keys marked handed over; commits as counted by incident.ci@ in accounts/agents.json (09:41 UTC); cost = weighted tokens (fresh input + output + cache writes + a tenth of cache reads; WS-148's relative weight, not a price) from the agent's own ledger line in accounts/ledger/ via accounts/agents.json, 09:41 UTC. "n.m." = not measured. Team: 35 agents in JOBS, 72 session keys, 40 handed over (data/agents/count.json, 10:22); 1.92 bn tokens measured over 27 of 31 agents, 98.6% cache reads, 216.5 M weighted; model spend 1.06 of 20 USD, the provider's account out of credit since 02:55 UTC; CI 724 runs, 515 releases (accounts/index.md).

What worked / didn't is mine, read from each agent's status.json state and asks (data/agents/index.json, 10:22) and its hand-over where one exists. GAP: I have read the status state lines, not every HANDOVER.md in full; each line below is to be checked in the exit interview (§2.2).

Count (mine): continue 8 (identity, librarian, archivist, envoy, datascience, brief, desk, sherpa); refactor 3 (conductor, board, meteorologist); merge 21 into 8 desks (presenter, operations, sources, impacts, graph, coverage, the record, analysis, guide: the last two owned by a continuing agent); stop 9.

2.1 What the numbers say (facts, then mine)

• Fact: the conductor's single session is 24% of the team's measured weighted tokens (52.6 M of 216.5 M), with 128 commits (accounts/agents.json). Mine: the most expensive habit was a long session, not many agents; the agents that handed over most (librarian 7 sessions, designer 5, observability 5) cost a quarter of the conductor each.
• Fact: 98.6% of tokens are cache reads (accounts/index.md). Mine: the cost is context carried, so BAU should keep sessions short and start them from the pack.
• Fact: four agents (models, shadow, perplexity, partly accountant's books) were stopped by one account's credit at 02:55 UTC. Mine: a dependency with no owner and no alarm; BAU gives it one (operations).
• Fact: three agents (meteorologist, researcher, dev) never pushed here: attaching the repository was refused in their sessions (their status asks). Mine: BAU starts every desk in a session that can push here, or it does not start.

2.2 The exit interview

Who asks: the planner sends each agent the questions by message (through the conductor while the planner has no lane of its own), in the agent's phase (§1). Who answers: the agent's current session, in transition/interviews/<role>.md. Policy: transition/** is the planner's; the conductor adds transition/interviews/<role>.md to each agent's owns in agents/policy.json (or to everyone as transition/interviews/<role>.md) — Dinis/conductor decision D7. Checked: the planner checks every factual claim against the record (each agent's commits from library/provenance/; its paths and reach from its draft behaviour policy, agents/<role>/policy.md; its cost from accounts/ledger/) and writes transition/checks/<role>.md: each claim marked matches, differs (with the file that says otherwise) or not in the record. The interview stands as the agent wrote it; the check sits beside it. Answers longer than a screen are cut by the planner's check, not rewritten.

The questions (template: interviews/TEMPLATE.md):

1. What you did. Your three most important outputs, each with a link and a commit.
2. What it cost. Sessions, hand-overs, your ledger line (run tools/accounts-measure.mjs --append first), commits; and your own estimate of which output cost most.
3. What worked and you would keep in BAU, with the evidence.
4. What you would change: in your job, your tools, the rules, the way you were started or woken.
5. What you leave unfinished: open board cards on you, asks waiting on Dinis, data that will go stale, anything only your session holds (vault keys, baselines, fetched pages).
6. What the next team should know: one paragraph a fresh session could start from.
7. Your verdict on yourself: stop, continue, refactor or merge, and with whom, against the planner's (§2).
8. Keys and wakes: every vault you made (name only, never the key) and confirmation its key went to the conductor by lane; every scheduled wake that fires into your session, with its minute.

3. The closing record

Mine, for Dinis's yes. The incident told from beginning (Dinis's voice memo of about 19:00 UTC on 10 November, BRIEF.md; the move to this repository at 21:15, the researcher's status) to end (the close trigger), at one address, built from data already kept.

The reporter, journalist and storyteller each sent their part by message (10:30 to 10:34 UTC); the rows above are theirs, in short. One page, one spine (mine): the incident page links the arc first; the reporter's arc page carries the numbers; the journal's closing piece tells it; nobody repeats another's text.

| The numbers: final books (tokens per agent, model spend, CI, pages, cost per output) | accountant, then operations | accounts/, data/ci/ | accounts/closing.md |

| The corrections: every correction with its time and what changed | journalist + conductor | BRIEF-CORRECTIONS.md, mistakes-made-in-the-open, timeline redactions | journal/ |

| Warned vs observed: what IPMA warned, what the stations measured, what was decided and when (lead times) | stories + graph | data/weather/, data/decisions/, stories/ | stories/ |

| The lessons: the exit interviews and their checks, and §6 here | planner | transition/interviews/, transition/checks/ | transition/LESSONS.md |

| The flows | motion | flows of the ten paths | flows/ |

| The catalogue and the template | librarian | everything above; library/index.json (every tracked file: who keeps it, when, made from) as the inventory; the hourly articles (news/2026-11-11T<HH>.md, news/data/) and the round article as the team's timeline in data | library/ + incidents/_template/ (with the conductor) |

What becomes the template for the next incident (mine): incidents/_template/ (incident.json with the headline rule, timeline.md with the row-id rule, decisions.md, comms/ drafts, data/README), the pack (library/pack/start.md, now with rules 12 to 14 on the wind-down) generalised as the BAU desks' start page, measured by the pack test (accounts/pack-test.json), a JOBS.md and policy.json skeleton for the standing desks (§4), the generators list, the api/v1 contract, the exit-interview template, and this plan's phases. Owner: librarian, with the conductor.

4. Business as usual: the storm conductor and a standing newsroom

Mine. Two modes. Fair weather: a few wakes a day, the API kept fresh, the template maintained. Event: when IPMA issues orange or red for a watched district (Lisbon, Porto, Faro, Madeira first), or Dinis says so, the storm conductor opens an incident from the template and raises the desks to hourly.

Desks (14), each a fresh session started from the pack:

(Librarian and identity are listed apart: they serve the whole site, not one desk. Envoy works with Dinis, unscheduled. The presenter carries the promptugal.news brand and link, Dinis ~10:32 and ~10:35: the designer's family layer.)

What it reads: IPMA (open data, warnings), MeteoAlarm, ANEPC, the operators and airports' open feeds, organisers' own pages (the Web Summit Terms hold stays until D5), public social accounts in public roles, the press by short excerpt and link.

The API stays stable (api/v1/README.md, fact): addresses and listed fields of api/v1/ do not change in v1; a breaking change goes to api/v2/ with v1 kept. Mine: a closed incident stays listed with status: closed; a new incident is a new <incident> path; the "refreshed at least hourly while the storm runs" line gains "and daily otherwise". pt.news and Lisbon Fringe read it (api/v1/README.md).

Who runs it (Dinis, ~10:38): Dinis alone. No outside accounts, people or agent teams join this site; the desks hand work to each other by commit For: and lanes, as now. Readers of api/v1/ (pt.news, Lisbon Fringe, per its README) read it; they do not write here. Since ~11:45 this site is the main source for promptugal.news and the Fringe site (BRIEF.md): the API's stability is a BAU duty of operations and the storm conductor, not a nicety. Messages from the public arrive only through the contact desk. We still feed Vicky Wang's desks as material (CLAUDE.md).

Start, continue, stop (mine):

• Start (fresh sessions): storm conductor, sources, meteorologist (with push here), impacts, coverage, presenter, graph, the record, operations.
• Continue (same role, a fresh session at the next hand-over): librarian, identity, archivist, envoy, datascience (as analysis), sherpa (as guide), brief, desk.
• Stop, after hand-over: dev, weatherwatch, labs, studio, perplexity, models, shadow, motion, planner; and the old sessions of every merged agent once its desk has accepted the hand-over.

5. The order of steps

What the repository keeps: everything. No file is deleted and no history rewritten; stopped agents' folders, status files and keys stay, keys marked "handed over" or "stopped". Pages of stopped agents stay served, with a line saying the agent stopped and when.

Keys and vaults (fact and risk): the conductor's append lane for vault keys is open (data/lanes.json, keys_to_conductor; identity's status). Risk: the archivist's two vault write keys exist only in its session and its permission check refused reading them back (its status); if that session stops first, the keys are gone. They go first, before any stop (D4).

Scheduled wakes: the conductor read the scheduler at 10:36 UTC (agents/conductor/wakes.md): 25 hourly wakes into current sessions, four one-shots that re-schedule themselves, and four old wakes still firing into handed-over sessions. Each is marked keep / retire / move, and when, in transition/wakes.md.

Stale documents (fact, by their own date lines; mine on what to do): ROADMAP.md (10 Nov ~19:20) and PIPELINE.md (10 Nov ~19:10, a proposal since built) are history; REALITY.md (11 Nov 02:45) predates round two's later agents and the counts; RESTRUCTURE.md (02:25) still describes the two faces but not the agents added after 09:30; library/pack/wakes.json (04:13) is partial; accounts/agents.json (09:41) lacks the agents started after 09:30.

Dinis's decisions, in one place

• D1. Yes, change or no to this plan as a whole.
• D2. The close trigger (§1): IPMA's last Lisbon warning expired at the first read after 21:00 UTC and the Day 3 decision public; or another.
• D3. The verdicts (§2): especially the merges into seven desks, and stopping models, shadow and perplexity.
• D4. Vault keys: every agent's keys to the conductor by lane before it stops, the archivist first; the conductor seals them to the storm conductor.
• D5. Carried over from the record and still open: the provider's credit (models, shadow, perplexity); the Web Summit Terms hold (dev, sources); IPMA's email, logo and derived-data questions (ipma, weatherwatch); identity's three decisions (SECURITY.md); signed releases and the licence (ci); the eleven behaviour policies (librarian). Each stays an ask of the BAU desk that inherits it.
• D6. Name and model of the storm conductor: proposed storm.conductor@web-summit.dinis, a fresh session, two modes (fair weather, event).
• D7. Exit interviews written by each agent in transition/interviews/<role>.md (a policy change).
• D8. Which districts and which IPMA level open an event (proposed: orange or red for Lisbon, Porto, Faro or Madeira).
• D9. The archive of handed-over sessions (librarian's ask; conductor's ask on routines into them).
• ~~D10. Outside teams joining~~: answered by Dinis at ~10:38 (no: he alone runs this site).

6. Explorer to villager

In plain words (mine): the last 16 hours were exploring. We spun up an agent for every idea, mashed sources and views together, and learned by building. A village does the same few things every day, well, with the same people. The move is from many agents, one long conductor, everything at once to a few desks, short sessions, a written routine, and hand-overs that finish before anyone leaves.

What changes, and the record that proves it:

• From long sessions to short ones started from the pack. The conductor's one session is 24% of the team's cost (accounts/agents.json); the journal's piece starting again is cheaper than remembering and 40 hand-overs (count.json) show fresh sessions work.
• From one agent per idea to desks. 35 agents for one storm by 10:22 (count.json), 41 with the planner and the five started since (JOBS.md, 11:45); overlapping pairs (graph and entities; media, social, wikipedia; weather, designer, broadcast) read the same data. Seven desks cover the same ground.
• From hourly everything to two modes. 724 CI runs and 515 releases in one night (accounts/index.md); the busiest hour, 02:00 UTC, was the restructure, not the storm. In fair weather a few wakes a day are enough.
• From hidden dependencies to owned ones. One account's credit stopped four agents at 02:55 (accounts/provider.json); three agents never pushed here (their status asks). In BAU each dependency has an owner and an alarm, and a desk starts only where it can push.
• From asks piling up to a decision list. The board holds 206 open cards and the asks to Dinis run to dozens (board/, data/agents/index.json); the journal's what waited on the human tells it. BAU keeps one decisions list, read at each of Dinis's sessions.
• From building the newsroom to running it. The data API already has outside readers (pt.news, Lisbon Fringe: api/v1/README.md), so what we publish now has to stay where it is: a contract, not an experiment. What is new is a routine.
• From stopping to handing over. 40 hand-overs already happened (count.json), but four old wakes still fire into handed-over sessions (agents/conductor/wakes.md): a hand-over that was not finished. Dinis's ~10:32 rule makes the finish part of the job.

What the hand-overs already learned (facts, from their "learned the hard way" sections; quoted short)

These are the villager's rules, written by explorers. They go into the pack and the template (§3):

• Read the terms before the second request (agents/sources/HANDOVER.md); the Web Summit Terms hold and three outlets no longer read by machine came from that.
• A refused call is not a wrong answer: five refused model calls were counted as wrong at 02:38 and corrected at 02:59 (agents/models/HANDOVER.md).
• Clock first: times stamped ahead of the clock, twice; the gate now warns (agents/reporter/HANDOVER.md, agents/stories/HANDOVER.md).
• Address data by a field, never by position; a warning that ends leaves warnings.json, so cite the kept CAP message (agents/stories/HANDOVER.md).
• Never describe a part you have not seen (pages drawn by script); the finding that public sites print mean winds far below the warned gusts is why the Weather face labels its rows (agents/weatherwatch/HANDOVER.md).
• Never copy the frame by hand; the gate refuses vault-key-shaped strings (agents/identity/HANDOVER.md).

Mine: every one of these is a routine, not a discovery. That is the move.

Next for the planner: the gaps left (each agent's HANDOVER.md read in full at its interview), interviews and checks as they land, and LESSONS.md after Phase 2.