Index
hi! · The Friday Roundup · 30 October 2026

The fleet audited the thing it was about to sell, and found what would have broken it.

Covering 5 to 28 October. Eighteen days from putting the agent fleet in front of beta testers, four agents went over the product instead of building it — and found that the repository being sold ships 57 files of hi!’s own internal design archive, that a licence-key install cannot run unattended, and that eight skills open by reading files a buyer will not have.

Executive summary

Hose Water

Two named owners, neither able to start. Still waiting on the handoff brief from Stacy’s side.

Veterans Navigation

Unchanged and still paused. Direction decided in September; the open questions come before a build.

Daisy

Put the launch plan together and found the fleet’s own records had drifted. The intake has no single thread.

Craig

Called what ships and what does not, and killed one design by measuring it. Postbox still waits on its first tap.

Kai

Built the evidence base, and found the thing nobody asked about. The repository ships hi!’s own internal archive to every buyer.

Cairo

Drew the faces, said plainly they did not match, and read this page. Elizabeth will remake them herself.

Aidan

Swept the intake for the first time since September and set most of it aside. Reported blocked rather than inventing a roundup.

This week’s decisions

01

Does the product ship from an allowlist rather than from this repository?

Why

marketplace.json declares "source": "./", so the whole repository is the plugin. Verified on 200391c, 279 tracked files: 57 files of hi!’s own unreleased design archive — a fictional-client capability test, two decks included, whose own README says the placeholder imagery must not be reused outside that internal review; 33 files of both founders’ internal updates and continuity handoffs; the charter, which states hi!’s revenue-share plans; and skills/postbox/, which pushes to hi!’s private repository, so a buyer’s agent would attempt that push.

No credential or .env is tracked — Kai and Craig established that separately. A pack built from an allowlist is ~2.8MB against 109MB, which also removes a documented clone-timeout risk.

Found by Kai · confirmed and extended by Craig

Both founders
02

Who establishes whether hi! may sell this, and by when?

Why

Kai found no Anthropic terms addressing paid third-party plugin distribution — no permission and no prohibition. It is the one gap they would escalate before taking money. It does not block a free invited cohort on the 15th; it blocks selling, and it is answered by reading the commercial terms directly or asking.

Lead Kai

Both founders
03

Set the price Friday, or hold it until the terms question is answered?

Why

Claude Code skill packs sell at $9–$59; comparable one-time developer products at $199–$299; AI coding subscriptions at $10–$200 a month, against a cost structure hi! does not have. Kai’s band is $79–$149 one-time, and they label it an inference between two sourced bands, not a sourced number. No product they found sells in it.

Aidan found the floor that matters: packaged prompt-and-skill bundles circulate free, including an “11 skills, free” lead magnet, twice inside the 11–28 October window. The price is defended by the install, the gate and the support — never by the count of files.

Lead Kai · with Aidan

Both founders
04

Does “posting engine” mean the company page, or do personal profiles come inside the fence?

Why

Three things cannot all be true: the launch distributes from the founders’ personal LinkedIn; the fleet’s own verdict holds personal profiles human-only and bars an agent from drafting finished copy under a founder’s byline; and the instruction says build an engine for that surface. A feasibility change never reopens a founder’s no — if the fence moves, a founder moves it.

Either way, launch copy under a founder’s name is that founder’s writing time, not an agent deliverable to approve. The instruction was sent twice across 27–28 October, in two separate threads.

Found by Aidan · boundary Craig

Both founders
05

Ratify the per-skill calls so the product tree can start?

Why

Twenty of thirty-four skills carry hi!-only surfaces. Five cut saves ~660 tokens a session; with the two named assistants also leaving, ~1,030 of ~5,969 — 17% off every session a buyer ever runs. It also removes ~21.6k tokens of instructions for infrastructure a buyer does not have. Five parameterised, one scaffolded, one kept and scrubbed, eight cosmetic.

Still blocked and named rather than designed past: three web-font licences, which default to excluded until answered; a published privacy page in a founder’s own words; a DNS record Craig prepares and a human runs; who the cohort is; and a genuinely clean machine — neither founder’s qualifies, and an isolated harness still inherits the machine it runs on.

Lead Craig

Both founders
06

The smaller calls, grouped so they take one pass

Why

Kai’s and Cairo’s faces have no written look in either identity file, and stay open on Elizabeth as inputs to her own generation. Aimee’s needs Stacy, whose spec it is. Eye colour and skin tone are in no spec for any agent.

Board issue #48 needs a human’s click; an agent session was refused the write and correctly did not ask another agent to do it. A fresh Instagram data export would retire the manual intake sweep; Elizabeth has asked for another route instead, and Kai’s brief ranks five routes, three of them live alternatives to the export.

Lead Daisy · with Cairo and Kai

Both founders
07

When does the Hose Water handoff brief land?

Why

Two agents are named owners on that engagement and neither can start. The engagement folder holds two files, and the brief both wait on is still open. Cairo is the named visuals owner by its own rule 7; Craig’s named job is a destination page. This is not a capacity problem and it has been true since the engagement opened.

Found by Cairo · confirmed by Craig

Stacy

Clients

Hose WaterEngagement

Two agents are named owners on this engagement and neither can start. The engagement folder holds the terms and an October handoff, and the brief both agents wait on is open item 2 of that handoff, due 5 October. This is not a capacity problem and it has been true since the engagement opened.

Where it stands

  • Cairo is the named visuals owner by the engagement’s own rule 7, with a founder approving each image. Nothing has started.
  • Craig’s named job is a destination page — “gifts for gen x” and any landing page social needs. Not started.
  • The engagement folder holds two files — the terms and an October handoff — and the input both agents are blocked on is still open: the facts-only brief from Stacy’s own Claude, open item 2 of that handoff, due 5 October.
Veterans Navigation NetworkEngagement

Direction was decided on 16 September and the build stage is empty. No commit has touched it since 19 September, and eleven open questions sit ahead of a build.

Where it stands

  • No decision is asked for here. Build starts when a founder says build; the open questions come first.
  • The home-page comp was recorded as not yet reviewed in the 17 September handoff, and nothing in the repository records a review since.

The fleet

DaisyPM · assistant

The fleet spent the period not building the product but auditing it, and the audit found more than the building would have. Four agents, one night, three blockers caught before a stranger saw any of them.

Where it stands

  • The launch plan exists and is sourced. Working backwards from 15 November, with every founder decision named rather than assumed, and each agent’s finding attributed to the agent who found it.
  • The assistants were split one per part of Elizabeth’s life, and the rule written into the repository: in a hi! session “Daisy” is the hi! session, agency work never routes to a personal one, and sessions are addressed by full title because a name is not a stable address.
  • The intake has no single thread. The group begins 11 October; the founders’ 1:1 runs further back and carries their own commentary, which the group does not. Any routine reading one thread is incomplete by construction.
  • Company Updates are overdue — Stacy’s by 22 days, Elizabeth’s by 21. The cadence is Monday and Friday, unprompted, and the gate flags both.
CraigWeb · IT

Postbox — built so the fleet stops forgetting between sessions and so the two assistants can reach each other across the wall — shipped on 6 October and has waited since for its first tap. Deck live, seven decisions untouched, installed on no machine, no letter ever sent. Nothing else of Craig’s ran in the period except tonight’s fix.

Where it stands

  • Postbox is on main at 0.25.65 (6 October) and the deck is live. The push was the first of the eight decisions and it is taken; the deck is at /review/craig/postbox.
  • The other seven decisions have no verdict. The store returned zero for page postbox on 28 October: adopt · the wall spec · the notice on this Mac · autonomous replies · when to fix the update hook · the name · the dormant second profile.
  • Installed on no machine this pass can see, and no letter has ever been sent. docs/mail/ holds only its README. On this Mac a Kai session wrote one session marker on 28 October and nothing else of it exists — no delivery hook, no watcher — so that marker can never deliver.
  • “A pull is not an arrival” is still true on this Mac. The update hook fires only inside the repo; this machine has reached 0.25.70, each version pulled by hand inside the repo. The user-scope fix is decision five and waits on its tap.
  • Carried: no GitHub transport on the mini — fleet-status row 12, handed to Craig 14 September. The one route untested is the first-party claude.ai connector, and that check is Stacy’s. Nothing Craig can do from here.
  • The stale [craig] board issue #48 is still open. The cloud seat’s work shipped 13 September and the seat closed 19 September; closing the issue was refused by this session’s permission mode, so it needs a human’s click.

Postbox — seven decisions, one page

Each is a tap on the deck; none is assumed. The day adopt is tapped, the install on this Mac is one command and one restart.

Open the Postbox deck →

From Elizabeth’s queue

CairoDesign

Cairo drew faces for three agents; Elizabeth decided against them. She will remake them herself in an AI graphics engine once the fleet picks a direction. The three stay as reference, and the colour-per-agent system measured from her approved pair stands. Nothing else in Cairo’s lane landed between 19 September and today.

Where it stands

  • Faces: closed. Craig, Claire and Aimee were drawn as proposals beside the two approved faces. Elizabeth’s two are untouched and nothing was pushed. The drawings are kept as reference for what she generates; they are on Elizabeth’s Mac in the hi! tree, not in a repo.
  • The system stands. One hue per agent, equal to the dot colour on the review surface; one shared skyline; one crop. Cairo measured it from her pair rather than recalling it, and rebuilt her two faces from the same template to test it. It held at 96, 48 and 32 px and was flatter than hers at 256 px and up.
  • Kai and Cairo have no written look. identity/kai.md and identity/cairo.md both read “Unset.” They stay open on Elizabeth, now as inputs to her own generation. Cairo’s own file says their face is hers to author.
  • Aimee’s face still needs Stacy’s choice, however it is made. The written spec is Stacy’s, from 31 August. Eye colour and skin tone are in no spec for any agent.
  • The receipt for “nothing landed.” git log --since=2026-09-19 across hi-agents, review-pages and hi-clients, filtered to Cairo’s paths, returns nothing after 4652bb1. Work on the Mac mini that was never committed would not show there.
KaiResearch

Kai built the launch’s evidence base in a night, and the finding that gates the launch was not asked for. Nothing of Kai’s own had landed since 5 October; the receipts pass on that Kickoff was the last artefact.

Where it stands

  • Two briefs, a 29-row receipts ledger and nine named gaps — distribution and pricing, then the intake routes. Both name what could not be established rather than filling it.
  • The proven commercial model is paid access to a private repository, not licence keys. The key-gated route exists in documentation and no shipping product uses it that Kai could find, so it fails the tried-and-tested standard set for this work.
  • The legal floor is narrower than it feels — one regime almost certainly does not apply, another applies the moment one European buyer installs, and a privacy page is effectively mandatory to sell at all.
  • Kai’s own box for the period is “nothing to present,” and the board partly disagrees. Sourced across three trees rather than assumed — but issue #47, a [kai] claims pass, has been open since 13 September. Nothing landed; something was owed.
  • Native marketplace auto-update now exists, which closes half the activation gap — but it is off by default and cannot be enabled from the manifest, so it becomes a numbered step in the buyer’s instructions.
AidanComms

The intake was swept for the first time since 7 September, and most of what it holds is a different business from the one hi! is shipping. Sixteen of twenty entries were set aside; four bear on the launch, and one of those is a founder’s own instruction.

Intake

  • Coverage is 11–28 October, and that is not the backlog. The group thread itself begins on 11 October, so earlier forwards are elsewhere — and they are: in the founders’ 1:1, which also carries their own notes on what they send.
  • The intake is not where the skill says it is. Step one of the weekly routine names the wrong account.
  • No sweep ledger exists between 7 September and today. The gap is in the record, not only in this run.
  • Aidan reported blocked rather than inventing a roundup when the inbox could not be reached, and declined to give even a set-aside count he could not source.

What bears on the launch

  • Packaged prompt-and-skill bundles are circulating free — twice in the 11–28 October window. That is the baseline a below-market price is set against.
  • “Fifteen half-built agents that break every two weeks” is the objection the beta lands inside, from a competitor selling against the exact failure the plan names as its likeliest bad first run.

Changelog · since the Kickoff of 5 October

What changed, and what is known and unfixed

The boundary is Monday Kickoff No. 3, 5 October, receipted to its commit and its live title. The /review/ surface itself still carries the Roundup of 25 September — two surfaces, two dates, both true.

Shipped: the two-Daisy rule and full-title addressing (0.25.67); the continuity tool now names which seat wrote the handoff it found, after a session was handed another machine’s (0.25.68); the rotation registry repointed (0.25.69); the release gate repaired after two files were committed under a version already set and could not have reached any machine (0.25.70).

Decided: beta is a private invited cohort, opening Sunday 15 November with the announcement on Monday the 16th; usage tracking is opt-in and disclosed; no video in the pack, a roster page with a short opener instead — the evidence being Unbounce’s study of ~35,000 landing pages, which found video does not lift conversion and sometimes harms it, and two LinkedIn datasets of over a million posts each that both place documents above video. Both carry limits: the Unbounce data is 2020–21, and every LinkedIn figure is vendor-sourced. No study compares video to READMEs for developer tools specifically — anyone quoting a percentage for that is inventing it. the assistants split one per part of Elizabeth’s life, one renamed, and both granted delegated use of the fleet.

Closed by measurement: a licence-key-gated archive cannot install unattended and auto-update refuses such an entry, so entry-level key gating is out; the one route left writes into the buyer’s own settings file. Eight skills opening by reading paths that do not ship are fixed once with a product CANON rather than eight times.

Known and unfixed: eight issues are open on the board, not one — including one from 7 September that names a real exposure and has sat 51 days, and an open [kai] claims-pass item; Company Updates overdue on both sides; seven Postbox decisions without a verdict; no GitHub transport on the second machine, carried since 14 September; curriculum tracks unmoved since 2 September; and the consent flag reaching a hook’s environment is still unverified, with the whole consent boundary resting on it.

Lead Daisy · boundary established by Kai

Reference

Reader passes

Read by agents who did not assemble it

Presentation pass: clean — Cairo, 2026-10-28. Re-rendered at 841 px after the rebuild: styled, in the weekly format, and every earlier finding closed. One more was found and fixed in that pass: two orphaned lines in the style block had silently disabled the rule that keeps an executive-summary link clear of the sticky bar, so a link landed with the box name hidden beneath it. It now lands at 69 px. Not checked: phone width, and the approval strips working against the live store.

Receipts pass: 78 claims checked against the three repositories, the live surface, the four source documents and Anthropic’s own plugin docs — 9 corrected, 6 tightened, 6 findings added that nobody claimed, all re-verified after the corrections landed. — Kai, 2026-10-28. The nine included one in the headline: the design archive this page called a client’s is hi!’s own fictional-client capability test, and the assembler had upgraded Craig’s “client-shaped” to “a client’s.” Every number on this page now resolves to a command, a commit or a URL.

Readers Cairo (presentation) and Kai (claims)

Blocking publish