source: .forge/design/prototypes/README.md · generated — edit the markdown, not this file

Prototypes

Browsable, self-contained builds of the two apps this engagement designs against. Both are references, not contracts — where a prototype and a spec disagree, the spec wins (.claude/rules/plans.md § UI / Design Adherence, .claude/rules/source-trust.md).

Each prototype is a vendored bundle: referenced as a whole, never file-by-file. The harness reference graph excludes this folder for that reason (.claude/lib/harness-graph.py EXCLUDE_PREFIXES) — an asset in here having no inbound link is correct, not a dangling reference to go and fix.

Prototype Entry file What it is Trust
crm/ crm-prototype.dc.html The original CRM prototype, frozen: the design the client approved, with the team’s additions up to 2026-11-04. Never edited. Inventoried in ../../discovery/inventories/prototype-inventory.md High for UX/flows/screens as approved; Medium for domain facts (grade lists, stage names and thresholds in it are our design, not client-confirmed — risk R-03)
crm-current/ crm-prototype.dc.html The current CRM prototype: the same design drawn in the built CRM’s look. Built screens match forge-et-fe; screens not built yet are drawn in the same look and marked Phase 2 · not built yet. Kept in step with every UI change High for UX/flows/screens; Medium for domain facts. The application code wins on anything built
pwa/ pwa-prototype.html A frozen snapshot of the trade-show capture-app POC (../forge-et-pwa-poc), with demo data baked in. The porting reference for forge-et-pwa (PRD F36). Never edited Medium — a freeze. The POC repo wins on anything current
pwa-current/ pwa-prototype.html The current capture-app prototype. Began as a copy of pwa/ (2026-11-05), except that on 2026-11-10 its 20 pre-show exhibitors, which were real companies including Empire’s competitors, were replaced with fictional ones whose web domains were checked as unregistered, and on 2026-11-11 it was moved onto the CRM’s theme as version 2 (§ The current prototype leads). Since 2026-11-12 it is built from an editable copy of the app’s source in pwa-current/source/, with the screens and flows agreed that day (§ The capture-app screens and flows agreed on 2026-11-12). Earlier versions are kept, frozen, in pwa-current/versions/ (§ Versions of the capture-app prototype). forge-et-pwa has no capture screens yet (its one route is a placeholder); F36’s screens are to be designed here first and forge-et-pwa built to them Medium-High for capture UX/flows. The application code wins on anything built. Files in versions/ are a record of earlier versions, not a design target

The originals stay frozen; the current copies stay in step with the built apps

Decided 2026-11-05 at the lead’s request. In a client demo the team shows the built app first and, when the client asks about something not built yet, opens the prototype to show the proposed design. A prototype that looks like a different product undermines that, so each app has two:

  • The original (crm/, pwa/) is kept exactly as it is, as the record of the design the client approved. Nothing edits it; its content is unchanged since 2026-10-06.
  • The current copy (crm-current/, pwa-current/) is the one to demo and the visual target for new screens. Its built screens look like the app, and its not-built screens use the same look. There is one current copy per app, updated in place; git history holds its earlier states, so no version is ever copied to a new folder. One exception, for the capture app: a version of pwa-current/ that has been shown to Empire or accepted by the lead is kept as a frozen file in pwa-current/versions/ when the next version replaces it (§ Versions of the capture-app prototype). Added 2026-11-11 by Nanndish, to keep the POC-themed version beside the CRM-themed one; it changes the lead’s 2026-11-05 rule, so it waits on Tharinda’s agreement (.forge/owed.md).

Every UI change keeps the current copy in step. A forge-et-fe or forge-et-pwa change that alters how a screen looks — a new screen or state, a changed layout, control, label, column or colour, a theme change — updates that screen in crm-current/ or pwa-current/ in the same delivery, committed to the ticket’s harness branch. The plan lists it (.claude/rules/plans.md § Required Sections; plan-reviewer DC-11), the implementer draws it (frontend-implementer § Design reference), and the reviews check it (frontend-reviewer dimension 28, /forge-pre-pr-review). A change that cannot be drawn there in that delivery gets an .forge/owed.md row naming the screen and what differs. A design for a screen not built yet is drawn in the current copy too.

The current prototype leads; the app follows it

Decided 2026-11-10 by Tharinda (lead). The current copy is where the intended design is settled, and forge-et-fe and forge-et-pwa are built to match it — for both apps, the CRM and the capture app:

  • Design first. A new screen, or a change to how an existing screen looks, is drawn in crm-current/ or pwa-current/ before the work item that builds it is planned, and the plan’s ## UI / Design Adherence cites that screen (.claude/rules/plans.md § Required Sections). A screen not drawn yet is drawn as the first step of the feature, never left for the implementer to invent.
  • Then build to it. The implementer copies what the current prototype shows; the spec and the design system still win where they disagree with it (.claude/rules/source-trust.md).
  • When and by whom. The drawing is made at spec time, at the latest before the plan is approved, and is committed to the ticket’s harness branch with the spec or plan. Whoever approves the spec or plan accepts the drawing with it, and the plan’s ## UI / Design Adherence names the screen. A capture-app screen that is new to the product, and any change to the frame or a shared pattern, also needs the lead’s acceptance. A small change inside an existing pattern (a label, a column, a state) may be drawn in the same delivery as the build instead. A plan-less fix/BUG-* PR follows the after-the-build rule above (draw it in the same delivery, or an .forge/owed.md row).
  • Both ways. When the build adds to or changes a screen — a state the drawing missed, a fix found in review, a constraint of the real data — the current prototype is updated to match in the same delivery, as the rule above says, so the two never drift apart.

For the capture app, nothing is built yet, so pwa-current/ is brought to the intended design before F36 is specified (.forge/owed.md).

pwa-current/ took the CRM’s theme on 2026-11-11 (Nanndish, with the UI/UX designer), the first step of that work. The screens, flows, labels and demo data were then still the POC’s; only the look changed:

  • Colours. The POC’s own colour variables, which every compiled class reads, take the values of forge-et-fe src/ui/styles/theme-revamp.css: the gold primary (#8A5A12, white labels) for buttons, links, the selected tab and the POC’s brand tints; the warm neutrals for the canvas (#FCFAF4), cards (white), borders (#E7E3D7), control edges (#8C7B66) and muted text (#6E5F4E). The lead score uses the CRM’s status tones: a score of 85 or more, its avatar and the worth your time count are success green, warm is warning amber and hot is danger red. The capture shutter is the gold primary, the camera-off stripes the darkest warm neutrals.
  • Type and shape. Inter replaces Manrope (the same woff2 the apps self-host, inlined), and the radii follow the CRM’s (12, 10 and 8px).
  • Light mode only (design-system § Theming mechanics): the POC’s dark mode, its header toggle, the side menu’s Dark mode button and the Appearance setting are gone.
  • Brand. The Empire Teas logo replaces the green L mark, and Empire Teas over a gold Lead capture eyebrow replaces LeadScout in the header, the side menu, the sign-in page, the install card and the settings footer, as the CRM’s sidebar shows Empire Teas over Sales CRM. The browser icon is the same logo. Tharinda (lead) accepted the brand on 2026-11-11, as the change to the frame that § The current prototype leads asks the lead to accept.
  • How it was done. On 2026-11-11 the compiled bundle was patched in place. Since 2026-11-12 the theme lives in pwa-current/source/index.html’s <style id="crm-theme"> block, which every colour in the source reads (§ Editing the capture-app prototype).

forge-et-pwa takes the same colours from its own src/ui/styles/theme-crm.css, which points the token names at the CRM’s revamped colours over theme-empire.css (F36-a FR-SC3).

The client demo kit shows the same theme: demo-kit/crm-theme.mjs reads the crm-theme block, the Inter face and the logo out of pwa-current/pwa-prototype.html and applies them to the POC’s web app as it is served, so a colour changed here changes the demo too (demo-kit/README.md). The demo kit runs the POC’s own screens, so it shows the theme but not the screens and flows below. It is not a design reference: a change to this prototype needs no change to the demo kit (demo-kit/README.md § Not a design reference).

The capture-app screens and flows agreed on 2026-11-12

Decided in the team’s discussion of 2026-11-12 (Nanndish, with the UI/UX designer); drawn the same day. The changes to the PRD and the F36 specs these need are listed in .forge/owed.md. These are new capture-app screens; Tharinda (lead) accepted them on 2026-11-12, with the later changes the same day below (acceptance reported by Nanndish).

  • Sign In. The CRM’s Microsoft sign-in (AD #28), drawn as the CRM’s sign-in page stacked for a phone: no email or password form. A user gets in only when their role in the CRM holds the new capture-app permission, Use the capture app (capture.use), which gives full use of the app; without it they see No Access to the Capture App and can only sign out. The CRM’s role editor (crm-current/) shows the tick, in a Capture app group, and the demo Sales Representative and Sales Manager roles hold it. Demo links under the sign-in card show the refusal and the sign-in notices.
  • Leads (the home screen). Two sections. To Review holds every captured lead nobody has decided on, researching or ready. Completed holds the decided ones. A person reviews each lead and either converts it, which creates a CRM lead owned by whoever converted it (AD #35: one owner; the capturing rep is kept as captured by), or rejects it, with an optional reason from a common list (Not a fit for our products, Competitor, Supplier or service provider, Student, visitor or press, Duplicate of another lead, Other) and an optional note; one rejected without a reason shows No reason given. A rejected lead stays in the capture app only and can be moved back to review. Nothing reaches the CRM without that review. A lead has no initials avatar, in the list or on its page: there is no photo to show, and the score ring and tier badge already carry the fit, so the name takes the space. Each section has its own search (name, company or title) and own filter, kept while the other section is open: To Review filters by tier (All, Hot, Warm, Cold), Completed by outcome (All, In CRM, Rejected), each chip with its count after the search. The third tier is labelled Cold (the POC called it Filed).
  • Capture. One business card or one badge QR at a time: Scan a pile (several cards in one photo), choosing several photos at once, and the captured today count are gone. Taking the photo, or reading the badge, starts nothing: what was captured replaces the camera view with Retake (or Scan Again), the booth note sits below it, and Submit for Research sends it. Badge QR reads one badge and takes the same booth note.
  • Plan. Campaigns are archived and unarchived from the campaign picker; archived ones sit in a collapsed Archived list. Re-rank is gone. At each booth the rep writes visit notes and marks it visited. Each exhibitor is an Option B card (chosen from three in the Plan Card Options design canvas): no left column; the booth’s stand code (always in full) and hall (cut with an ellipsis), the tier and the score on the top line; then the company. The card has no AI line (Nanndish, 2026-11-12): booth, tier, score and company are enough to choose the next stop, and the reason to visit is on the booth page; a line shows only while research is not finished (researching, not researched, research failed). A visited booth has a green tick before the booth, a green border, and a strip with who visited and when (Visited by you · 2h ago) and the first line of the note (Your note, Ruwan’s note). A line under the list’s title says AI researches and ranks each booth, and the rep decides. The booth page shows the full hall and the full reason. The Plan Card Options canvas (https://claude.ai/artifact/DH4qnSPyaT6udEeHYWe7w9, shared only with the people its owner chose) also draws Option B on an iPad; those layouts are not needed for now (Nanndish, 2026-11-12), so on a tablet the prototype keeps design-system §9’s phone layout in a centred column.
  • The Plan list’s search, filters and sort, and adding, editing and archiving exhibitors (Option A of the Capture App Changes design canvas, https://claude.ai/artifact/LUvmKT7mcQ1mCC8Yqn9pa1, chosen by Nanndish on 2026-11-12). Above the list: a search (company, booth or hall), a Filter button with the number of filters on, and the status chips All · To Visit · Visited · Archived, each with its count. Filter opens Sort and Filter: sort by Best fit · highest score first (the default), Booth number · A to Z (a booth with no number goes last), Company · A to Z or Recently added · newest first; filter by tier and by hall. Below the chips one line always shows the sort and, when filters are on, them as one summary cut with an ellipsis (Hot, Warm · Hall 3, Hall 4) and Clear. Add Exhibitors opens a sheet with Upload Exhibitor List and Add by Hand; it is no longer a button that shows and hides a card. Add Exhibitor takes the company name (required), website, booth number, hall, country and notes (all optional); AI researches and ranks it once added. Several can be added in a row (Nanndish, 2026-11-12): Add Another saves the one typed, clears the form and keeps it open, a line counts and names those added, and Done closes it. The booth page’s More menu has Edit Details (the same form; a new company name or website researches the exhibitor again, the other fields save without new research) and Archive Exhibitor (an optional reason from a list and an optional note). Archived exhibitors are under the Archived chip, with who archived them, when and why, and Restore.
  • Uploading an exhibitor list uses no AI (Nanndish, 2026-11-12). The app reads the CSV or Excel file and matches each column by its heading against a fixed list of names per field (source/src/lib/exhibitor-columns.js: Exhibitor Name is Company, Stand No. is Booth, Web is Website, and so on). Check the Exhibitor List shows each field with the column it was matched to (Matched from the heading), and the rep changes any that are wrong or unmatched; the rows found update as they do, rows with no company name are skipped, and Company is required. After the import, AI researches and ranks every exhibitor on the list (Nanndish, 2026-11-12): the Research the top N by fit setting and the rough ranking pass it needed are gone, which is simpler but costs more per show and takes longer to finish (.forge/owed.md, the capture-app decisions row, item 5). The POC reads a list the same way (header aliases, its docs/plans/epic-13-preshow-exhibitor-targeting.md).
  • Plan puts the campaign first (Option D of the Plan Header Options design canvas, https://claude.ai/artifact/HBHQKCHrS5YECbUFKK9BZt, chosen by Nanndish on 2026-11-12). A campaign takes its show’s first and last day (optional; one with no end date stays current until it is archived), and where it stands follows from them. Running or upcoming: a campaign card heads the page with its name, Running · day 2 of 5 or Upcoming · starts 12 Dec, its dates, the booths visited and Switch Campaign and New; adding exhibitors is an outlined Add Exhibitors button above the list, except when the campaign has no exhibitors yet, when Upload Exhibitor List is the filled button with Add by Hand below it. Ended (its last day has passed): the page opens on Plan Your Next Show with Create Campaign as the filled button, the ended campaign as one View Plan row, and Archive; View Plan opens its plan under the same campaign card, labelled Ended, with no Add Exhibitors. No campaign chosen: Plan a Show (or Plan Your First Show when there is none) with Create Campaign. The demo has one campaign in each state: Gulfood (running), London (ended), Hong Kong (upcoming) and Colombo (archived).
  • Settings. Installing the app and the signed-in account (with Sign Out) only; the theme choice and the email alerts are gone. The prototype adds a Prototype Demo card with a No Connection switch, which the app does not have.
  • The states F36-a, F36-b and F36-d need, drawn after the 2026-11-12 review (Nanndish, 2026-11-12; they wait on Tharinda’s acceptance, .forge/owed.md):
    • No connection (F36-a, PRD Rev 64). A banner on every tab says new captures are kept on the phone. Submit for Research with no connection keeps the capture: Capture lists it under Not Sent (waiting for a connection, when it was kept, its booth note, and Remove, which asks first), and Leads shows how many are not sent with View. When the app is online again it sends them, oldest first, shows Sending…, and says how many were sent. There is no background sync: nothing is sent while the app is closed. The prototype turns the connection off from the link on Capture or the switch in Settings.
    • Correcting what was read (F36-b). A lead in To Review says it was read from the card by AI (or from the badge) and offers Edit Details: a name or a company, at least one; job title, email, phone, website and country optional. A new company or website researches the lead again; every save checks the CRM again. Once saved, the card-read check says who edited the details (Edited by you). The check’s other states — a card being read, reading off, a card not read, with Read Again where the server allows it — are reached with ?cardRead=<state> (source/src/demo/card-read-scenarios.js).
    • The existing-client check (F36-d, PRD Rev 23). A badge is checked as soon as it is read, under the badge on Capture; a card once the reader has its company, on the lead. It says Already in the CRM (the account and its owner, whether the person is already a contact on it, emails in the past year and the last contact; Phase 1’s CRM has no deals or orders to count), New to the CRM, or Customer check pending while there is no connection or the card is still being read. A lead row shows Existing account, and Convert to Lead names the account. Whether converting such a lead links it to the account is F36-d’s spec to decide (.forge/owed.md).
    • No Open in CRM (Nanndish, 2026-11-12): the app is used on a phone, where the CRM is not opened, so a converted lead says to find it in the CRM’s Leads, and nothing in the app links to the CRM.
    • The lead’s page is the booth page’s drawer (Nanndish, with Tharinda’s acceptance, 2026-11-16; F36-a WI-4.2 deviations 1, 2 and 5): a lead opens over the Leads list in the frame the booth page uses — full width on the phone, a right-hand panel from md — at every width, with Lead and ✕ in its header. There is no bottom sheet with ‹ Leads on the phone, no pane beside the list on desktop, and no Delete Lead (deleting a captured lead is out of scope, OQ-64). Leads has no running-campaign bar: every holder of Use the capture app sees every captured lead, and each row and lead page shows its campaign.
  • In-app notifications (added 2026-11-12, Nanndish). The Leads tab carries a badge with the number of leads ready in To Review. When a lead finishes its research, a Ready to review notice shows on whichever tab is open, with View, which opens the lead, and a dismiss button; it closes by itself after 8 seconds. After Submit for Research the rep now stays on Capture, ready for the next card, and the notice says when the lead is ready. A notifications bell (in the header, at every width — the capture app has no side menu, design-system §9) keeps the record: each lead whose research finished, and each lead a teammate converted or rejected, newest first, with a count of unread ones; choosing one opens the lead and marks it read, and Mark All as Read clears the count. These work only while the app is open. The capture app sends no email (AD #42: proposed by Nanndish, confirmed by Tharinda on 2026-11-12 — in-app notifications are enough), so a rep sees what happened when the app is next opened; push notifications are left for later.
  • Casing. The screens drawn here follow design-system §9’s casing for the capture app (Rev 105): titles, buttons, tabs and field labels in Title Case, messages and values in sentence case.
  • Emails, as checked on 2026-11-12. The POC sends no email: its notification step was designed (email the rep the findings once a lead is researched and ranked) and never built, and its mailer is a mock. F36-c planned one SES email when a captured lead was researched and ranked; AD #42 replaced it with the in-app notifications above. The prototype draws no email.
  • The campaign focus, drawn on 2026-11-18 for F36-c wave 3 (spec Rev 6: FR-FC2, FR-FC6, FR-LD5, FR-PL1, FR-PL3, D-29, D-30, D-33), before the plans that build it; it awaits Dinil’s review. The campaign form gains an optional Focus field after the show dates, with its hint, a of 300 count from 250 characters and the over-limit error on blur and on save; after Save Changes the campaign list shows Changes saved. and, only when the save’s answer says rankingAgain, the line Exhibitors and leads to review are being ranked again for the new focus. A To Review row ranked with a focus carries the Focus marker after the tier (an outlined pill with the focus icon, a 44px button of its own outside the row’s button, the fixed line as its name and its tooltip on hover, keyboard focus or a tap; it does not open the lead); the Plan list carries none. The lead’s page and the booth page show Ranked for the focus of : “” after the provenance, and Being ranked again for the campaign’s new focus. while a new ranking is due. The demo gives Gulfood the focus Supermarket private-label buyers in the Gulf, edited from Private-label buyers: most open leads and booths are ranked for the new focus, Aigerim Seitkali and Westmoor Tea Importers are still being ranked again, and the converted and rejected leads and the archived booths keep the old one; a focus edited in the demo ranks again 12 seconds later. Screenshots 10d, 20f, 20g, 62a, 62b and 78a–78e (phone) and 06–15 (desktop) in pwa-current/screenshots/. The marker’s look as a shared pattern of both apps is left for the design-system revision (spec D-29, FR-DC2).

Editing the capture-app prototype

pwa-current/pwa-prototype.html is built: edit the source in pwa-current/source/, never the built file. The source is a copy of the POC’s web app (../forge-et-pwa-poc apps/web at 3bd7dac), with its server calls answered by an in-memory demo store (source/src/lib/api.js, data in source/src/demo/), so the built page needs no server and no network.

cd .forge/design/prototypes/pwa-current/source
npm ci
npm run build    # writes ../pwa-prototype.html; npm run dev serves the source live
  • build.mjs runs vite build and puts the script, the stylesheet, Inter and the Empire Teas logo inside the page. Commit the source and the built file together; node_modules/ and dist/ are ignored.
  • The colours are the <style id="crm-theme"> block in source/index.html. demo-kit/crm-theme.mjs reads that block, the Inter @font-face and the window.__ET_LOGO__ script out of the built page, so keep their shape; node demo-kit/crm-theme.mjs --check <forge-et-pwa-poc> checks the demo still applies.
  • The demo data is fictional: the captured leads come from demo-kit/demo_data.py, the Gulfood campaign and its exhibitors from version 1’s snapshot; the other campaigns, London’s four booths, the halls, the two archived Gulfood booths and the sample list every upload reads (SAMPLE_EXHIBITOR_LIST) are added in source/src/demo/store.js. source/src/demo/kit.json is a copy of demo_data.py’s cards and research, made by hand. A change to the prototype’s leads is made in kit.json; demo_data.py changes with it only when the demo kit is to show the same leads.
  • Check a change in the browser (the pwa-current launch entry) at phone and desktop width before committing: no requests leave the page, and it renders its data from the URL people will open it at.
  • Screenshots of every screen and state are in pwa-current/screenshots/, at phone and desktop size, each with how to reach it, so F36’s specs, plans and reviews can compare the built app with the design screen by screen, as well as with the prototype. Whenever the prototype changes, its screenshots are re-taken in the same change: run node .forge/design/prototypes/pwa-current/screenshots/capture.mjs and commit the pictures with the change; a step it can no longer click through is reported, and needs the script updated. The pwa-screenshots-freshness.sh Stop hook asks, at the end of a turn on a branch that changed the prototype, whether to re-take them when the prototype is newer than its screenshots, and on a yes Claude builds the prototype and runs the script (.claude/hooks/README.md).

Versions of the capture-app prototype

pwa-current/pwa-prototype.html is always the latest version, the one to demo and design against, and its badge names its version. An ordinary update — keeping the prototype in step with a forge-et-pwa change, or drawing a new screen — edits source/ and rebuilds pwa-prototype.html in place, and is not a new version. A new version starts only when the lead accepts a redesign of the frame or the look, or when a version shown to Empire is about to be replaced. Only then is the file it replaces first copied, unchanged, to pwa-current/versions/v<N>-<what it was>.html and never edited again, the badge and <title> name the new version, and this table gains its row. Each file is self-contained and opens the same way (double-click, or the pwa-current launch entry, then /versions/<file>).

Version File What it is Date
2 pwa-prototype.html The CRM’s theme: gold and warm neutrals, Inter, light mode only, the Empire Teas brand. Since 2026-11-12, built from source/, with the screens and flows agreed that day (Microsoft sign-in, review before the CRM with search and filters, one capture at a time with Submit for Research, in-app notifications, a campaign-first Plan with search, filters, sort, exhibitors added, edited and archived, the list upload without AI, booth visit notes), accepted by Tharinda on 2026-11-12 2026-11-11, updated 2026-11-12
1 versions/v1-poc-theme.html The POC’s own look (LeadScout, dark and light, green): pwa/ with the fictional exhibitor list of 2026-11-10. Byte-identical to pwa-current/pwa-prototype.html on main before version 2 2026-11-10

crm-current/ was created on 2026-11-05 by redrawing the original to match forge-et-fe as deployed on dev, checked screen by screen against dev.crm.empire.lk as of that date:

  • The frame. The sidebar is flat and lists the real destinations in navConfig.ts order — Accounts, Contacts, Leads, Activities, Email, Pricing, Settings — with the user and the demo’s role switcher at its foot. There is no top bar: the global search, the New menu and the bell are gone, and each create action sits on its own screen (New lead from inquiry on Leads, New deal on Pipeline, New quote on Quotes, Pipedrive migration on the Settings page). Inter is the only font; the colours are the revamp tokens (primary #8A5A12 with white labels, muted text #6E5F4E, border #E7E3D7, control edge #8C7B66). The toast and the sign-in page match the app.

  • Screens drawn as the built CRM has them. Accounts (a list page, then a record page with Overview, Buying standards and Notes and timeline tabs); Contacts (new); Leads (a list page with the status pills and AI labels, then a record page whose cards hold the AI summary, research and the follow-up draft); Activities; Email (folders, message list and message in three panes); Pricing (Standards, Weekly prices and Auction sales tabs); and Settings — the landing page (new), Users, a user’s page, Roles, the three tag pages (new) and Email templates (new).

  • Phase 2 screens keep their content. Dashboard, Pipeline, Samples and Quotes take the same look, sit in a Phase 2 · not built yet sidebar group, and carry a banner saying so. Phase 2 entry points inside built screens stay too, marked Phase 2: the Activities Calendar view (F21), Convert to opportunity on a lead (it opens the New deal dialog and the Pipeline), Send to all clients on Pricing and Newsletter on Email. A current copy is the whole design — what is built, as built, and what is not, in the same look — never only the built app.

  • Checked a second time, with fresh eyes, on 2026-11-05. A reviewer with no part in the redraw compared every screen with dev and the forge-et-fe source, and its findings were fixed: list tables at dev’s 12.5px type and 43px rows; absolute dates (Tue 3 Jun) and money as USD 310,000.00; the account briefing’s Prepare briefing flow, the account’s standards table, notes with Add a note and a day-grouped timeline; a contact page; the lead page’s AI summary, research, tag and draft cards; one message per email with From and To; Pricing’s weekly round with dev’s columns and states; Activities’ Open and Completed views; a user’s page and a role’s page laid out as dev’s, the role editor behind Edit; dialogs with Cancel then the action, right-aligned; and the demo roles gating controls and records as forge-et-fe does — each sidebar entry, create and edit control needs its permission, and a list shows only the records the role reaches (AD #37). Switching role lands on the first entry the role holds.

  • The role editor follows AD #34 / AD #41 (2026-11-05): a role’s Edit opens forge-et-fe’s Edit role dialog — Name and Description, None / View / Manage per area with All records (AD #37), Allowed ticks, and one Settings row with its six pages and the warning that Users or Roles reaches every page. Saving updates the role page. The access-model note below now describes crm/ only.

  • Phase 2 screens redrawn on the product’s patterns (2026-11-05): Quotes as a list page and a record page, Samples as a list page with the board in product cards, a deal as a record page beside the sidebar instead of a drawer, and the Pipeline with the product’s page header.

  • Checked against the PRD and the features list (2026-11-05), by a third reviewer with no part in the redraw. Removed, because the project dropped them: Similar past quotes (F7-e) and the AI product-brief assistant (F16), both PRD Rev 61; priority P4 and the separate nurture bucket (three levels, PRD Rev 55); the Past orders panel and order-history wording (no order data in V1); the invitation wording (AD #34: an administrator creates an active user); the shared sales@ inbox (won’t build); and the “catalog” wording (a deal’s product is a buying standard). Added, because a feature has them and the copy did not: lead priority — P1–P3 with its confidence, Why this priority, Recommended next action, hot or cold, a Priority column and filter (F11-g, since built and drawn as built); email threads with Earlier in this conversation (F20-i FR-10); the record a message is linked to, with Change link (F20-f, marked Not built yet); a sent message’s tracking state (F20-g, Phase 2); and the user page’s Edit user dialog. Corrected to later decisions: the New follow-up dialog (F3-a’s six types and fields) and the New lead dialog (F11-b’s fields, no priority) as forge-et-fe draws them; the F37 dialogs as committed Phase 2 work; the share dialog’s wording (AD #35, AD #37: sharing is transitive and reach comes from the reporting line); the last-administrator rule (AD #41); and sub-floor approval asking for the permission, not a role (AD #26). Left as dev has them: no Leads count in the sidebar and no Pricing summary strip, because the built screens have none; and no notification bell, which no feature delivers (an owed.md row asks whether one is wanted).

  • Brought up to forge-et-fe main on 2026-11-06 (commit 4a91ed9c), screen by screen against its source, for everything built since the redraw. Drawn as built:

    • Leads: sortable columns, row menus, the Owner, Tag and Priority filters, the AI markers and tags in one Labels column, the archived, not-found and error states, Approve & send, the account link, the AI summary’s spam, duplicate and existing-customer states, the read-only follow-up draft, the tags card, New lead from inquiry with its failure path, and Edit lead.
    • Accounts and Contacts: the list filters, archived and dormant accounts, row menus, the contacts card with archived contacts, the contact dialogs, note editing, the timeline’s filter chips, the briefing card, New account and Edit account, and Send this week’s prices with its marker.
    • Email: All mail, the category filter, the Focused / Other tabs, one row per conversation with its count, Earlier messages, the New mail notice, the collapsed folder rail, the reading pane’s actions, the mailbox connection states and the Mailbox settings page, the compose dialog with templates and attachments, and the email template page and editor.
    • Pricing and Activities: the standards book’s filters, menus and dialogs, the weekly price round (add, edit, publish, withdraw, correct, history), new auction sales, the workbook import page, the page warnings, and the Edit follow-up dialog, Mark done with Undo, and New follow-up as built.
    • Settings and the frame: offboarding, Create user, deactivate and reactivate, the refusal panels, Edit user, New role and Delete role, the tag pages’ edit and retire, the list states, the sign-in notices, the No access, Not found and route-error pages, the landing with nothing held, and the Activities badge counting overdue follow-ups.

    The record edit dialogs (Edit account, Edit contact, Edit lead, Edit user, Edit role, the template editor and the Access dialog) use the Rev 90 frame, as built; the other dialogs keep the older footer band, as built. Replaced, because the built app redrew a feature the copy already had: the share dialog by the Access dialog, used for a lead and an account, and the older New and Edit lead forms.

    Kept and marked Not built yet:

    • Leads: the older processed-lead card, editing the follow-up draft in place, and the older extraction preview.
    • Accounts: the standards row that opens a price history, and the notes and timeline counts.
    • Email: a message row’s linked-record chip, the compose footnote on tracking, and AI draft with nothing to draft from.
    • Pricing: the standard detail panel, the Update this week’s price dialog, the older standard form, and the full price-history drawer.
    • Activities: the activity drawer, opened from the Phase 2 calendar.
    • Settings: the Pipedrive migration card.

    Not drawn: the loading skeletons, the stale-data bands, layouts below laptop width (except the Email page’s, since 2026-11-11), and offboarding’s notices for a caller who cannot reach every record (.forge/owed.md).

  • Redrawn for design-system Rev 110 on 2026-11-11 (UI polish round 2, forge-et-fe branch feature/ui-polish-round-2): Cancel without an icon and no required-fields sentence in the forms and dialogs; the account and lead headers with Add follow-up and the primary action last; the lead page’s AI cards first (AI summary and Priority in the main column, Research first in the side column), the triage tag’s ✕ inside its badge, and lead notes with Edit and Delete; the Leads data bar with More Filters; the Email page in one card with Compose in the header, and, below laptop width, one pane with the Folders: drawer; 20px sidebar icons; the draft prices icon and Add standard with no client chosen; banners in the Alert tints; and the search picker and Select look for the dialogs’ dropdowns. Not drawn: the lead page’s Follow-ups card (.forge/owed.md), the import warnings beyond No warnings, the publish dialog’s notice (drawn as a plain hint), and the Phase 2 dialogs’ dropdowns. The prototype’s own labels stay in sentence case.

  • Not changed from the original: the demo data apart from the corrections above, the role switcher, and what the Phase 2 dialogs ask for. The record pages’ spacing and the dialogs’ footer band are close to dev’s, not measured to the pixel.

  • Descriptions of the original. The discovery inventory, the screen descriptions in feature-detail.md (the Lead Triage queue, the Inbox, the top bar) and the dated notes below describe crm/.

  • Editing it. The shared styles are one UI object at the top of the script (page frame, buttons, cards, tables, chips, filters); a new screen binds those rather than writing literal styles. Tables are CSS-grid rows (gridTable), because the template’s loops cannot sit inside a <table>, which the HTML parser would move them out of. A bound style cannot take extra inline CSS (style="{{ ui.card }}; padding:0" drops the whole style): add a variant to MX instead. Check a change by rendering it (the crm-current entry in .claude/launch.json) before committing. Screens read the data state from this.state.demoData ('normal', 'empty', 'error'), which the sidebar user menu’s Demo data choice sets; the same menu’s Frame pages opens No access, Not found and the route error. The Email and Mailbox pages have a Mailbox (demo) select for the connection states, the Awaiting access demo role shows the landing with nothing held, and links under the sign-in card show its notices. Code added on 2026-11-06 sits in blocks marked // ---- <area> catch-up (2026-11-06) ----.

  • F20-g drawn on 2026-11-09, from the F20-g spec and its frozen wave-1 contract, with its frontend work items: the tracking indicator (Sent, Opened, Clicked — dot and word, design-system Rev 111) in the message view, the Sent Items rows and a timeline email entry, replacing the earlier Delivered / Link clicked chip marked Phase 2; the compose dialog’s Track Opens and Clicks checkbox, Send Later with its date and time and Schedule Send, and the warning on Edit and Send for an email that may have gone; and the email folder pane’s Scheduled entry and its list with Cancel, Edit and Send and Dismiss. The shared code sits in the block marked // ---- F20-g tracking and scheduled send (2026-11-09) ----, and each edit made in place elsewhere carries a // F20-g comment.

  • F36-c wave 3 drawn on 2026-11-18 (F36-c spec Rev 6 FR-CV5 and D-29; F11-g spec Rev 7 FR-R1, FR-V1, FR-V3, FR-V4, FR-V7), before the plans that build it; it awaits Dinil’s review. A lead’s Priority panel names a ranking carried from the capture app in its source (capture app — inquiry and research, or capture app — inquiry only, ) and, for a ranking made with a campaign focus, shows Ranked for the focus of : “” after the provenance; the Leads list’s priority cell carries the Focus marker after the priority (an outlined pill with the focus icon, a 24px button of its own outside the row’s link, named by the fixed line, whose tooltip opens on hover, keyboard focus or a click and closes on Escape, a second click or leaving; it does not open the lead), and the Priority column is wider to hold it; and the filter bar gains Campaign Focus (All, Ranked for a focus, No focus) after Hot / Cold, cleared by Clear Filters. Two demo leads carry a capture ranking with a focus: Sakura Wellness KK and Emirates Hospitality Group. Screenshots are in .forge/plans/F36-c/wave-3-screens/. The shared styles are the f36c… entries in MX, with rankSourceOf and focusLineOf beside rankFields, and each edit made in place carries an F36-c Rev 6 comment.

  • F37-a: the Pricing status moved below the header buttons on 2026-11-17 (spec Rev 5, FR-U5), as the built header has it (forge-et-fe 74172406), so its late arrival never moves a button.

  • F37-a drawn on 2026-11-15, from the F37-a spec (FR-U1–FR-U6) and its frozen wave-1 contract, before its frontend work items were planned (F37-a Decomposition Plan § Notes item 3): on the Pricing header, Send Prices to Customers before Import Weekly Prices, for a role holding Open a mailbox (the demo’s email.manage; the stretch-era Send a bulk email key now gates only F37-b’s Newsletter), with the batch’s status beside it; the review dialog Send This Week’s Prices to Customers — the week, From, the three counts, Copy Each Customer’s Other Contacts and Track Opens and Clicks, the customer list with its filter, Select All, Select None and each row’s Preview, Not Sent To with its reasons, and the confirmation — with a Review (demo) select for its empty and error states; and the results dialog in its sending and finished states, with Stop Sending and an Outcome (demo) select. These replace the Send to all clients · Phase 2 button and its dialog; F37-b’s Newsletter · Phase 2 button, chip and banner stay. The demo sends nothing: a confirmed batch moves its customers on every 5 seconds, and each customer Sent shows Prices sent on its account page. The Tea Room demo role reaches no account, so the batch is shown as the Administrator or the Sales Manager. The shared code sits in the block marked // ---- F37-a price broadcast (2026-11-15) ----, and each edit made in place elsewhere carries a // F37-a comment. Compared with the built screens on 2026-11-16 and corrected where they differed, so the screens are drawn as built.

  • F20-k drawn on 2026-11-21, from the F20-k spec, before its frontend work items were planned: Blocked Addresses on the Email › Mailbox page (between Receiving mail and Disconnect, or after Connection without a usable connection), the Block an Address dialog with its answers, Remove’s question, Block Sender in the inbox message view, and the Email Block List card and page in Settings (design-system Rev 151). The shared code sits in the block marked // ---- F20-k block lists (2026-11-21) ----, and each edit made in place elsewhere carries a // F20-k comment.

The original CRM prototype was maintained past its approval date, by team decision, until it was frozen on 2026-11-04. Client-confirmed decisions that postdated the approval were folded in rather than left to contradict it — the confirmed bulk-pricing model (Buying Standards, OQ-40/42/33/38) landed 2026-08-31, and the Microsoft-only login screen (AD #28) landed 2026-09-04. The approved baseline as of any earlier date is git history (git log --follow -- crm/crm-prototype.dc.html).

The access and sharing model is BEHIND the decisions — 2026-09-24. AD #34 (the access model) and AD #35 (ownership and sharing) supersede what the Users & roles screens and the lead share affordance show, and the prototype has not been updated. Superseded: the View / Manage / Act catalog and its outward-acting warnings, the non-deletable system Administrator role, the last-administrator check that covers role assignment only, navigation decided by role name, the “a real invite creates the account” toast, the share picker that offers users marked “No lead access”, the “cannot pass it on further” footer, sharing limited to the owner or a manager above, and a share-removal button with no permission check. The “Record sharing — 2026-09-16” and “roles are simplified — 2026-09-14” notes below describe the prototype as drawn, not the decided model. The drift rows, verified against the source, are in feature-detail.md (F5); the decided sharing design is ../data/record-sharing-model.md.

Bulk email is drawn in, and it is the one thing here that is AHEAD of the committed scope — 2026-09-16. Every other note on this page, apart from the access-model note above, is the prototype catching up to a decision already taken. This one is different and is marked as such on the screens themselves, because a prototype that shows more than the build commits to is how a demo becomes a promise.

  • What landed. A Send to all clients action on the Pricing screen (the week’s prices to every client in one action) and a Newsletter composer on the Inbox screen. Both carry a visible STRETCH chip, a dashed rather than solid control, and a Not committed to V1 banner inside the modal; both are demo-only and send nothing.
  • Why it is here at all. Empire confirmed bulk sending as an important requirement and named the two uses (PRD Rev 28, 2026-09-16, reversing the bulk-email half of the 2026-09-11 exclusion — WhatsApp stays out). It was F37, a V1 stretch candidate, and since 2026-10-09 is committed to Phase 2 as F37-a (the price broadcast) and F37-b (one message to a saved contact list, with a per-send summary) — PRD Rev 57. The on-screen Stretch — not committed labels predate that and are out of date in crm/; in crm-current/ F37-b’s button, its dialog’s chip and its banner read Phase 2. The screens show the intended shape, not the requirements: F37-a’s and F37-b’s specs decide those, and F20 stays specced without them. F37-a’s spec (approved 2026-11-15) specifies the price broadcast — a review dialog, one confirmation and a results dialog, with the button named Send prices to customers — and its screens are drawn in crm-current/ as built, without the Phase 2 labels (F37-a drawn on 2026-11-15, above, in the crm-current/ notes). F37-a was started early in Phase 1 (from 2026-11-06, PRD Rev 84; back in Phase 2 from 2026-11-11 and in Phase 1 again from 2026-11-15), as F20-g was, whose screens have been drawn as built since 2026-11-09; both are done and accepted with Phase 1 since 2026-11-18 (PRD Rev 100). F37-b is started early in Phase 1 since 2026-11-19; its Phase 2 labels in crm-current/ stay until it ships, and its delivery updates crm-current/ as for any UI change.
  • The two halves are drawn differently on purpose. The price broadcast is a batch of per-account emails — each client sees only their own standards — so its modal is a review of recipients, listing who is included, how many emails that is, and naming who would be skipped and why. The newsletter is one message to many, so its modal states the obligations that brings (unsubscribe, consent, sending address, bounce suppression — OQ-62) instead of designing around them; those are Empire’s answers to give, not ours.
  • A new permission, deliberately its own. Send a bulk email is a separate Act permission from Send email from a record (AD #26): a Sales Rep holds the second and still cannot write to every customer at once. In the demo it is held by Sales Manager, Tea Room and Administrator. The role editor’s Act-permission warning and its permission total both pick it up automatically.

Record sharing — 2026-09-16. The CRM prototype now carries the sharing path Empire asked for on 9 September (PRD Rev 29, OQ-58 closed; decisions in ../data/record-sharing-model.md). On lead detail: the owner, the people the record is shared with (each removable, with who shared it and when on hover), and a Share… control shown only to whoever may manage that record — its owner or a manager above them. The picker is where the model is legible: anyone who already reaches the record through their own role (the owner’s manager, an Administrator) is marked rather than offered, because sharing with them would change nothing; and anyone whose role holds no lead permission is marked too, because a share widens reach and their role grants no authority to use it — the Tea Room and R&D roles show this today. The queue marks a lead SHARED WITH YOU so a shared record is findable in the list rather than only by direct navigation. The sidebar identity now follows the role switcher (it named one fixed person before, which was harmless until “who am I” started deciding what is shared with me).

The products catalog is gone, the market axis is gone, and roles are simplified — 2026-09-14. Six changes landed together; each is the prototype catching up to a decision already taken, except where marked. The pricing half of the 2026-09-11 note below is now superseded by this one — read the two together: that note explains why the price moved onto the account, this one removes the market axis the note still described and the catalog it still assumed.

  • One price per standard — the market axis is removed (PRD Rev 23; forge-et-be/docs/data-model.md §2.2 Rev 6). The per-market price board, the per-market history grouping, the “unpriced markets” list and the quote builder’s destination-market dropdown are all gone. A standard now carries one price, one validUntil, one status and one weekly history series; the account row shows a Price / kg column directly, and the expanded row shows this week’s price, the week-on-week movement and validity — no cost or profit: the account section is a sales view, and since price plus profit % reconstructs cost exactly, both stay on the Pricing screen and the quote only. The quote’s price basis is the standard number, not a market. Destination is still recorded on the quote — it is just not a price key.
  • Standards can be added, edited and archived from both the account section and the Pricing screen, gated on the manage-standards permission. Archiving keeps the price history and drops the standard out of quoting and out of the weekly round; numbers are never reused, so create always mints a fresh one. Last-admin protection is enforced on role assignment for the same reason.
  • The products catalog is removed outright — nav entry, screen, detail panel and “add product” modal. The Buying Standard is the sellable record, for private label as well as bulk, so the seed now carries two private-label standards alongside the bulk ones and every surface that read the catalog reads standards instead: the quote picker (scoped to that account’s standards), the F9 certification check (deals carry standardNos, not productIds), global search, and the deal form’s product type-ahead. Reconciled in the harness on 2026-09-15 — PRD Rev 27: the Product entity is struck, F26 is retitled Buying standards & weekly pricing, OQ-37 is revised and OQ-61 raised (are private-label standards weekly-priced?); the backend’s product table, catalog package and product.* keys were removed by FS-15 on 2026-09-16 (forge-et-be#29; V11__drop_product.sql).
  • Pricing is reached by permission, not by role. The gate was role === Tea Room || Administrator; it is now role holds the pricing permission (AD #26), resolved from the same role records the role editor edits. A Sales Manager or Sales Representative holding View published weekly prices now gets the Pricing screen read-only, without costs — which is the point: ticking a permission in the role editor changes what the screens do.
  • Grade and Spec were the same string, so the duplicate column is gone from the account section and the Pricing screen. Grade is the surviving name everywhere.
  • The role model is simplified to AD #32, closing the three superseded items flagged below. No templates (only the system Administrator ships; every other role is shown as one an Administrator created); one role per user (roleId, a picker that replaces rather than adds — the two-role person and the union rule are gone); and no AI-assist kind — the catalog is View / Manage / Act, where Manage carries the ordinary actions of a record including the AI help that assists them, and only three permissions are held out as Act (send email, approve a below-floor quote, publish the week’s prices). The catalog is also coarsened to view/manage per entity per the 2026-09-13 team call: 41 permissions became 23, each carrying a one-line note saying what Manage covers. Grants, reach (own/team/everyone) per permission, the reporting line and the effective-permissions panel are unchanged in shape.

Bulk pricing was re-homed onto the account and a Tea Room Pricing screen on 2026-09-11. The prototype had every bulk price hanging off the catalog product — a price board, history graph and weekly-publish button on the Products screen, with no account dimension anywhere. That contradicted the settled model: the priced unit is a client-specific Buying Standard, N:1 Account (forge-et-be/docs/data-model.md §2.1), and the Tea Room publishes a round across the whole book, not a walk through the catalog. Nothing about the pricing data changed — same standards, same numbers, same (standard, market, validity) grain — only where it is shown. Read the rest of this note as history: the market axis and the Products screen it refers to were both removed on 2026-09-14 (see the note at the top of this file). What survives from it is the principle — the priced unit belongs to the account, and the Tea Room works the whole book:

  • The account gained a “Buying standards & prices” section (standard no · grade/product · spec · this week’s price per market with the existing Current/Expiring/Expired chip · valid until), each row expanding to the price board, the history graph and the full-history drawer.
  • A new Pricing screen carries the Tea Room’s workspace — every client’s standards at once, searchable, with this week’s validity window, published/draft counts, the unpriced-market list and the weekly publish entry point.
  • Products became the catalog it says it is: the price board, history and publish button are gone, replaced by one summary line (“N buying standards across M clients · from USD x/kg”) linking to the Pricing screen. Finished and private-label goods are untouched.
  • Cost, profit floor and profit % now render only for the Tea Room (client/roles-and-permissions.md A-10). This tightens the previous behaviour, which showed cost to every role on the Products screen. Publishing is Tea Room only (A-20); the Administrator sees the Pricing screen read-only (A-22’s default — no fallback publisher); every other role has no Pricing nav entry at all, and the route is gated, not just the nav button.
  • One nuance worth stating because the two documents can look like they disagree: the account section shows Tea Room columns, while A-21’s default is that the Tea Room sees only the pricing screens. The section demonstrates the price-visibility layering; whether the Tea Room reaches the Accounts screen at all is A-21’s question, still open with Empire.

ERP integration claims were removed 2026-09-04, leaving the handoff untouched. The prototype had claimed a live link the project does not build: a “Last synced with ERP · 6 min ago” note on the dashboard, “ERP sync” and “& ERP” in the two AI-briefing provenance captions, a “— via ERP” subtitle on the account’s Past orders, and a seeded timeline entry reading “Reorder shipped via ERP”. The issued proposal is explicit that automated CRM-to-ERP integration is not in V1 (§ERP handoff; it is a priced follow-on), and AD #10 ships the manual path behind an ErpHandoffService seam — so those five claims contradicted the contract rather than illustrating it. The Won-stage ERP handoff panel is deliberately untouched and still works — the readiness rows, Generate handoff document, the “pending confirmation of ERP API availability” caption, the resulting “ready for manual entry into the ERP” confirmation, and the Won → ERP handoff stage name are all the manual path the proposal does commit to.

The login change is a team decision already stated to the client, not a proposal awaiting their answer: the sign-in card previously offered Continue with Google and a Work email + Password fallback beside the Microsoft button, and AD #28 dropped both — so the card now carries one provider button and no password form, matching PRD Rev 20 and client-facing-prd.md. The Microsoft mark keeps its exact brand colour, which design/ui/design-system.md §1 records as the one surviving literal-colour exception (the Google mark retired with the button).

Three changes in it are team proposals, not client-confirmed — treat them as designs we are putting forward for the F5/F8/F9/F26 specs, not as settled behaviour:

  • F9’s certification check now reads a different held-cert source per pipeline (2026-09-02). The required side is unchanged — always the account’s certs. What it checks against is now: the product brief’s certsRequired for a private-label deal (the item has no catalog entry yet, so the brief is the only statement of what it will be certified to), and the certs held by the catalog products linked to the deal (productIds) for a bulk deal. It is deliberately not the linked quote’s line items — F9 has to flag the gap before a quote exists, and the quote is created later in the deal’s life. Either side falls back to the certs recorded on the deal while no brief / no linked product is on file, which is the state a brand-new inquiry arrives in. The panel names the basis it used (“Checked against …”), so the mechanism is visible in a demo rather than implicit. (Since 2026-09-15 the bulk side reads the linked buying standards — deals carry standardNos, not productIds — with the same mechanism; the rest of this bullet describes the 2026-09-02 state.)

    • Bulk deals previously had no product link at all, only a free-text product string, so productIds was added to the seed deals. Where the pricing seed already pairs an account with a product via a Buying Standard, the link follows it (Hamburg→P-01, Gulf Beverage→P-02, Emirates→P-03) rather than being invented.
    • One demo verdict changes: O-517 Gulf Beverage “Herbal teabag range” goes green → red. That is the new rule working — the buyer requires Halal and the linked herbal products hold only USDA Organic + FDA — but it is a visible difference if you have demoed that deal before.
    • The deal drawer’s cert chips were relabelled from “This deal’s product holds” to “Certifications recorded on the deal”. They still read the deal record, unchanged; the old label would otherwise contradict the panel directly beneath it on bulk deals.
    • The export-document checklist and the ERP handoff readiness rows deliberately still read the deal record’s certs, not F9’s per-pipeline basis — verified by a rendered before/after diff showing no change to either. Don’t merge the two variables.
  • A role-management area was added, for the configurable role model (2026-09-07). An Administrator-only Users & roles nav entry with four screens: a users list (name, email, roles, manager, team size, status); a user detail carrying the roles held, a manager picker and an effective-permissions panel answering “why can this person see this?” — which role granted each permission and the reach it resolved at; a roles list (the six confirmed templates plus the Managing Director profile, each marked built-in template, alongside one Empire-created example); and a role editor rendering the permission catalog grouped by area, each permission carrying a kind chip (View / Edit / AI-assist / Act) and, where it touches owned records, a reach control (own / team / everyone). The sidebar role switcher grows from four demo roles to the six templates plus Managing Director.

    • All three superseded items are now applied (2026-09-14) — see the note at the top of this file. AD #32 / PRD Rev 24 removed templates, multi-role users and the AI-assist kind; the screens described above have been rebuilt accordingly, and the catalog was coarsened to view/manage at the same time. The description in this bullet is the pre-2026-09-14 state, kept because the rest of the note (what the four screens are for, what they illustrate, what is mocked) still holds.
    • It illustrates ../../client/roles-and-permissions.md, which is itself DRAFT awaiting Empire’s sign-off — so every name, template content, reach and default on these screens is a proposal (AD #30 / PRD Rev 21), not settled behaviour. The Sales Manager template defaults to team reach (its A-13); Open a mailbox is pinned to own on every role and cannot be widened (A-6, §7).
    • The permission names are plain-language stand-ins, not the catalog. The real code-declared catalog — its keys and its full contents — is FS-6 spec Rev 3’s work and is deliberately not reproduced here: the editor shows enough permissions to make the grouping legible, and no <resource>.<action> key appears anywhere in the UI.
    • Nothing persists and several controls are mocks. Toggling a permission, moving a reach or picking a manager updates the in-memory demo state and toasts; Add user, New role and Save role only toast. A reload returns the seed. The seed is ten people across the seven templates, with one Sales Manager carrying three representatives so the reporting line and team reach are visible, and one person holding two roles so the union-to-the-most-permissive rule is.
    • No other screen’s behaviour changes. Read-only rendering follows design/ui/design-system.md §9 — a control the role can never use is hidden, not greyed — and it is reached through a single canEdit flag that is true for all four pre-existing switcher roles, so hidden controls appear only under the new Executive / Management role (the New menu, Add product — now Add standard — New activity and the lead action row). The existing isManager quote gate is untouched — verified reading “Approve pricing & submit” as Sales Manager and “Submit for approval — Manager approves pricing” as Sales Representative.
    • OQ-56’s four reporting-hierarchy questions are made visible without being answered. The users screen states the reporting line is supplied at go-live and Administrator-maintained, and that deactivating someone with reports waits for those reports to be reassigned; the reach controls show the defaults we would build to. The screens are a prompt for those answers, not a claim to have them.
  • The Pricing screen’s shape, the Excel import affordance and the email-prices action (2026-09-11). Their source is the September 2026 follow-up meeting, for which no debrief is filed — so these are statements we heard, not recorded client decisions, and they are built as proposals to put to Empire rather than as settled requirements:

    • That a Pricing screen is the right home for the weekly round at all. design/ui/design-system.md §8 had listed the Tea Room pricing screens as to-design; this is our first cut at them, and its layout, its status vocabulary (Published / Draft pending / Gap) and its header counters are all ours to defend or change.
    • “Import this week’s sheet” renders the meeting’s statement that the weekly list arrives as an Excel sheet by email with columns standard number · unit · price. It is a button and an explainer only — there is deliberately no parser and no file input, because the shape of the real sheet is unconfirmed. It toasts a draft-row count; publishing stays a separate step.
    • “Email this week’s prices to contacts” renders the statement that published prices are emailed to the account’s contacts. It toasts the contact names; there is no compose flow, and nothing about who sends, on what trigger, or in what format is settled.
    • Also heard and deliberately not built into the UI: that standard numbers are never reused (already true of the model), that about 35,000 have been issued (why the Pricing screen is a searchable list and never a dropdown), and that one client’s product has exactly one standard (the seed still shows a client holding several, because the data model permits it and the August meeting said four or five).
    • The per-market price board is gone as of 2026-09-14 — see the note at the top of this file. Empire said on 9 September that a standard’s price does not vary between customers or regions, and the weekly price workbook they shared afterwards carries one row per standard per week; PRD Rev 23 records one price per standard and revises OQ-38. The account section, the Pricing screen and the quote builder now all show one price per standard, and the market dropdown is removed.

Opening them

pwa/pwa-prototype.html — double-click it. One self-contained file, no server, no network: the app bundle, styles, Manrope, the icon and the whole dataset are inlined, and a fetch shim answers every /api call from the baked-in store — verified with nothing listening on the POC’s API port.

If it ever comes up without data, serving it works too: cd .forge/design/prototypes/pwa && python3 -m http.server 8765, then http://localhost:8765/pwa-prototype.html.

crm/crm-prototype.dc.html — double-click it. No server, no network. This corrects an earlier instruction here that said a local HTTP server plus network access were both required; the second half was true until 2026-09-02, the first half was never true.

  • file:// was never the problem. support.js is loaded by a plain <script src="./support.js">, and a classic script tag resolves fine from a file:// page — it is fetch/XHR and ES modules that file:// blocks. The runtime’s one fetch(location.href) (a re-read of the raw template) does fail there, but it is wrapped in .catch(() => {}) and the page has already been adopted from the inline document by then, so nothing breaks. Verified: booted from file://, <x-dc> consumed, zero {{ … }} placeholders left.
  • Network was the real dependency, and is now removed. support.js pulled React and ReactDOM from unpkg.com at runtime, so with no network loadReactUmd() rejected, init() never ran, and because hideRawTemplate() had already hidden <x-dc> the page came up blank — the symptom people actually hit. Both bundles are now vendored in crm/vendor/ and loaded by two tags placed before support.js; it skips its own fetch when window.React already exists, so support.js itself is unmodified (it is generated — do not edit it). The vendored files are byte-identical to the unpkg builds, checked against the SRI hashes inside support.js. Keep the two tags ahead of support.js if you ever reorder that <head>.
  • Offline, the webfonts still fail (Fraunces and Inter in crm/, Inter only in crm-current/) and the page falls back to Georgia and the system sans-serif. Cosmetic only — every rule declares a fallback. Open it online for a pixel-faithful look.
  • Babel is not involved: it is only fetched for x-imported JSX modules, and this prototype has none.

Serving it still works and is worth it when you want the real fonts or a stable URL — and it is how the Claude Code browser pane runs it, via the crm-prototype (original) and crm-current entries in .claude/launch.json (a file:// page loaded there becomes a data: snapshot, which breaks the relative vendor/ script tags and comes up blank):

cd .forge/design/prototypes/crm && python3 -m http.server 8099

then open http://localhost:8099/crm-prototype.dc.html.

The .dc.html extension is deliberate — it marks the file as a Claude Design canvas document, which is how it can still be reopened and edited in the canvas. Do not rename it to plain .html.

What the PWA snapshot does and does not do

It is the real POC frontend — a production Vite build of forge-et-pwa-poc/apps/web, not a mockup — so the screens, states and interactions are the POC’s own. Against the baked-in store:

  • Reads are exactly what the mock API returned when the snapshot was taken: one campaign (Gulfood 2026 — Dubai), three captured leads (one scored hot/95, two correctly flagged as duplicates — the mock enrichment always returns the same contact), and twenty imported exhibitors of which the top five are researched and ranked.
  • Capturing works: it adds a lead as processing, then flips it to ready, which is what the app polls for. The in-browser camera needs a secure context; the “choose from camera roll” fallback works anywhere.
  • Deleting a lead works, in memory.
  • Everything else that writes (editing campaigns, importing or re-ranking exhibitors) answers with a plain “Static snapshot — … is not available here”, surfaced by the app’s own error path rather than by a crash.
  • Nothing persists. Reload and the snapshot is back to its captured state.

A “Static snapshot” badge sits in the corner so it is never mistaken for the live app; clicking it dismisses it.

One screen in it is a mock, not the POC — “Badge QR”

The capture screen offers a third mode, Badge QR, beside the POC’s own Single card and Scan a pile. It is a demo mock and not POC behaviour. The POC has no QR scanner at all — no QR/barcode library, no BarcodeDetector — which project-prd-signals.md records as a verified gap in OQ-55, with badge QR as net-new build inside F36’s estimate. The mode opens no camera and decodes nothing; the attendee it “reads” is fixed sample data.

It lives in a hand-written <script> appended after the bundle — the same way the snapshot badge above does — so the minified build stays byte-identical and the mock can be removed in one cut. Its own header comment carries the detail. Anything else added the same way must follow the rule in that comment: the bundle’s Tailwind is tree-shaken, so only utilities the POC itself used exist, and a missing one fails silently as a 0×0 box rather than an error — reuse class strings copied verbatim from the POC’s markup, and style anything new against the app’s CSS variables.

Refreshing pwa-current/

pwa/ is frozen and never retaken. pwa-current/ leads forge-et-pwa: F36’s screens are drawn there first, and once F36 builds them, a UI change in the app redraws the matching screen in pwa-current/ (the rules above). Since 2026-11-12 pwa-current/ is rebuilt from its own source (§ Editing the capture-app prototype), not from a POC build: the snapshot-and-patch method that made pwa/, and its appended Badge QR mock, no longer apply to it — Badge QR is a screen of the source now. Before committing a rebuild, confirm the page makes zero external requests and renders its data from the URL people will actually open it at: those are two different claims, and once only the first was checked, the gap surfaced when a second person opened it.

Generated from .forge/design/prototypes/README.md by .claude/lib/render-docs-html.js. Markdown is canonical.