Grouped pages — one event, one interruption, one signature
Three agents noticed the same thing at the same minute: git push, gh release create, terraform apply. Without grouping, that is three beeps, three reads,
three decisions — three interruptions for one question: "do you want this
release out?". Noise is the enemy n°1 of a pager, and the shared attention
budget is the constraint n°1: grouping is what makes both survivable.
POST /v1/pages { command: "git push origin main", group: "deploy-42" } ──▶ rings
POST /v1/pages { command: "gh release create v1", group: "deploy-42" } ──▶ attached
POST /v1/pages { command: "terraform apply", group: "deploy-42" } ──▶ attached
the pager shows ONE item:
310 fw PUSH main>origin 3 commits +120-14 #a1b2
GROUP 1/3: +2 linked pages in this block
GROUP 2/3: 320 fw DEPLOY gh release create v1 #9c3d
GROUP 3/3: 110 fw DELETE terraform apply #7f10
The hard part is the signature, not the screen
BIP-F-18: a decision can never apply to another page. The naive way to group is to propagate the leader's approval to its members — and that is precisely the lie the fingerprint exists to prevent: you would be answering a payload you never signed.
So the block is closed the other way round. When a member joins, the leader's fingerprint is recomputed over the whole block:
digest = canonical({ v: 1, leader: <leader fingerprint>, members: [<each member fingerprint>] })
leader.payload_sha256 = sha256(digest)
The human reads the list on the leader's screens and signs the list. One
signature covers exactly what was displayed — nothing is inherited, nothing is
implied. The members are settled by that decision, and each member's audit
record says via_group: <leader>, so the journal never pretends a member was
signed individually.
Two rules keep that honest
A leader is re-signed only while it is pending and unread. Once a human has laid eyes on it, what they saw is fixed: the incoming page opens its own group and rings. Rewriting a read page would move the ground under a decision already in flight — and the whole product rests on the line you see is the line you sign.
No member is ever lost from the list. The block is packed into screens with
the label reserved before packing (GROUP 2/3: is 13 characters taken out of
the 80, not bolted on after). A block that cannot fit in the ceiling is
refused, not truncated: it rings on its own. Signing an incomplete list
would be worse than two interruptions.
What a block costs
A member never rings, so it never spends the attention budget: the block is one slot, not three. The budget counts deliveries, and there is one delivery.
budget 6/h · 6 blocks delivered → the 7th ordinary page is HELD
↘ but a member of an existing block still joins, free
That asymmetry is deliberate: holding a new question is protection; holding a member of a question already on screen would split a block across the hour and produce exactly the two-interruptions-for-one-decision this feature removes.
The projection: what this page will cost
A page that costs 4 of the 6 remaining is a different decision from a page that costs 1 of 6 — and the human must see it before tapping, not after.
BUDGET: this page costs 1 of 6 per hour — 4 left after it
BUDGET: this page costs 1 of 6 per hour — 6 already spent, it would be HELD
The note is written into the stored screens, so the replay of BIP-F-46 shows what was on the LCD, projection included. It is added only when it changes something: a free page (a receipt, a member, a red page) always says so, and an ordinary page says so once the window is partly spent. On the first page of an empty hour, "costs 1 of 6, 5 left" is noise — and noise is what this feature exists to fight.
GET /v1/metrics carries projection for the next page, so the dashboard
draws the hour that comes and not only the hour that passed.
On the wire
group is a fact on both kinds of page, so questions group too — the most
natural case ("two design choices, one answer"):
{ "kind": "question", "question": "date lib?", "group": "choix-date", "options": [...] }
{ "kind": "question", "question": "timezone lib?", "group": "choix-date", "options": [...] }
It is capped at 64 characters and enters the fingerprint, like the diff. The queue a pager shows contains leaders only: a member is the same question, listed, never a second item.
What it is not
Not a workflow engine. A block is one human decision about pages that arrived close together and share a key the station chose — it is not a sequence, not a dependency graph, not a plan. If the pages must be decided separately, do not group them.
Not a way to sign less, either. The signature covers more than before: the whole block. Grouping removes interruptions, never scrutiny.