AI agent: the core, redrawn

A new round on the model itself, from design/agent-model.md and the answers in its section 6. Three pages of the new model in the real shell with B's sidebar, the channel switches drawn both ways, and no "Draft" pill: the draft is named once, in the bar above the page. Pick the switch placement and name anything to change on the three pages; nothing is built before that.

What changes

From the apple-design review of today's six pages (shots of 27 Sep at 1440 and 390) and the model's section 5.

HighBehaviour is the prompt itself: 64 text boxes under prompt headings, 500-word blocks with [[list: …]] markup. A business change has to be written as prompt text.
HighThe channel switches sit in Settings, away from what they switch.
HighAI Settings mixes changes that wait for Publish with changes that go live at once, and looks the same for both.
HighCards inside cards ("Role" holding "Who the bot is").

MediumCopy against rulings 57 and 58: "Who the bot is", "When the bot gets stuck", dated incident counts in help text.
MediumA Global tab whose text is hidden where it acts.
MediumKnowledge as cards, with "Free text" repeated on each one.

The sidebar, one page per concept

Group	Pages	Replaces
Setup	Profile, Topics, Knowledge, Actions, Guardrails	Behaviour's four tabs and five headings, Global, the service lines in code, Triggers, the thresholds card
Jobs	Conversations, Comments, Moderation	AI Settings, the fixed replies, the eight private-message keys, the moderation tab and score card
Check	Playground, History	Playground as today (Test and Manage); the Log, now covering every concept

Rules drawn on every page

One draft holds every change to what the AI says or decides. The bar above the page is the only place it shows: "3 changes not published", Review, Publish. A changed row offers "Live wording" to compare.
On and off switches act at once, after a confirm that names who is affected. They never wait in the draft.
Every save offers "Try in Playground", opened on the job and channel of the edit.

A row shows its current value and opens a popup to edit it (ruling 18). No card inside a card.
Every notice is computed from live state (ruling 58): the WhatsApp line reads Meta's account status, the hand-over summary reads the topic's fields.
Labels follow design/copy.md: no pronouns, no questions, "AI" never "bot".

Profile: who the AI is, and where it works

The area opens here. The top of the page answers the question staff ask first: which job runs on which channel, right now. Each cell opens its job. Under it, the four things that shape every reply on every channel.

1440 · Profile, just after the voice was saved to the draft

390 · the area's first screen: every page with its current value

The grid is the page's one strong element. Everything else is quiet rows.
WhatsApp reads "On, cannot send" with Meta's reason under it, computed from the account's health, not written for today's case.

Voice is one text for every job; a channel's own note lives on that job's page.
On a phone the sidebar is the first screen: 48 px rows, each saying what the page holds.

Topics: each kind of enquiry, its owner and its hand-over

The 10 service lines, job seekers, a request for a person, a complaint and anything else, one row each. A row opens a popup. Here Mortgage Advisory is set to go to a person at once; the popup spells out what the AI will do, built from the fields above it.

1440 · Topics, Mortgage Advisory open

390 · the same topic, full screen

"Answered by" has three answers: the AI, the AI then a person, a person at once. The person and the LeadRat enquiry type are pickers, never typed.
"What happens" is computed: change the person and every line follows.

A topic never adds to what the AI offers: that list stays the website's (ruling 30).
The hand-over message has English and Arabic; the AI sends the one in the customer's language.

Comments: the switches inside the job

The whole comment journey on one page, in the order a commenter meets it: the channel switches, the reply style, the private message, the posts with their own instructions. The switches sit at the top of the job they control.

1440 · Comments, switches in the job

390 · a switch asks before it acts, and says who it affects

The public reply and the private message are one unit (ruling 13). When one runs without the other, the line under the switches says what commenters get instead.
English and Arabic sit side by side on every message.

Settings › Channels keeps the account, its credentials and its health, with one line: "AI: replies on, private message on · Open Comments".
Trade-off: turning a channel fully off means visiting three job pages.

The other way: the switches in Settings › Channels

The same switches, gathered per channel in Settings, drawn on Settings A (its own sidebar, as picked). The job pages keep the wording and show the switches' state as one read-only line that links here.

1440 · Settings › Instagram, the AI's switches with the account

390 · Comments, the state read-only, a link to Settings

One place turns a whole channel on or off, next to its connection and health.
The same confirm as in 3 before a switch acts.

Trade-off: changing a reply's wording and turning it on are two areas apart.
Trade-off: marketing needs Settings access to switch a channel.

The pick

Switches: in the job pages (3) or in Settings (4). Recommendation: 3. Apple's settings guidance (settings.md › Task-specific options) says to keep an option "in the screens they affect"; the job page is where the wording, the switch and "Try in Playground" meet, and Profile's grid already gives the one-glance view per channel that 4 offers.

Then, for pages 1 and 2: keep, or name what to change. The other pages (Knowledge, Actions, Guardrails, Conversations, Moderation, History) follow the same rows-and-popup pattern and are drawn after the pick.

Three tasks, counted from the model (read from the code, not timed)

Task	Today	New
Change the tone for Instagram comments only	13 clicks; the Instagram box can only add below text on another heading	8 clicks: Comments › Instagram note › Save › Try in Playground › Send › Review › Publish
Add a project brochure	7 clicks plus retyping; live at once, untested	11 clicks, no retyping; nothing reaches a customer before Publish
Hand every mortgage question to a person	5 clicks for a line the model may ignore; nobody is assigned	10 clicks: Topics › Mortgage Advisory › A person at once › pick the person › Save › try › Publish

What the build needs that does not exist yet

Six new settings rows (agentProfile, agentTopics, agentActions, agentGuardrails, agentJobs, agentMessages) with a one-way converter from today's rows, and the hand-over as an action (model section 5).
A knowledge_sources table for Knowledge's sources (one additive migration).