mock-verification

Smoke-walk a `web-application` in a real browser once it builds clean — stand it up in mock mode, walk the flows of its prototype, fix each failure where you find it, post progress item by item. Requi

SKILL.md

GitHub

Mock verification

A clean build says the code is well formed and nothing about what the screens do: a screen the prototype shows and the router never registered, a button wired to nothing, a form that flips a row and sends no request. So you open the app and use it. Mock mode stands it up on this machine with no cluster, no sibling service and no IDP. You verify and repair in one pass.

This is a smoke walk: breadth over depth. You are looking for a screen that is missing, an arrow that goes nowhere, a control that does nothing, a request that never leaves the page. Data, layout and wording are not your questions. Your prompt names the flows in scope: the flows of its item and of earlier items. A cycle's regressions land in shared chrome and on pages nobody meant to touch, so walk every flow in scope, whichever issue built it.

What you verify

The map

specs/design/components/<component>/prototype.json, under the project root: your working directory, one level above the App Path you edit in. A flow is one role (roleId) and the screens that role walks (screenIds), in order. prototype.tsx beside the manifest shows what each screen holds: the component for a screen id in defineApp({ screens }) has the screen's controls and, as to or useNav().go(…), where each one leads.

Each flow in scope is one item, by its id. Walk it under its role (?role=<id>) by clicking from its first screen through its screenIds, and do what its name says. Two flows on one screen are two items: the screen is not the unit.

A flow starts where its control is. Its start state is the state in which its first screen shows the control that its name needs. prototype.tsx gives that state as the branch around the control (status === "draft" around Publish menu). When the seed data does not start there, put the data in that state before you walk the flow:

  1. Through the app's own controls, by clicking (a close, a delete, a create).
  2. If no control of the app gets there, set that record of the seed in mock/handlers.ts to the state, as the contract allows it, and restart. After the flow, put the seed back as it was and restart. Change only the seed, never what a handler answers.

Open a screen in no flow at its route. Reach the shell screens (Account, Settings when the prototype has one, Signed out) through the user menu.

The manifest lists the display states (states); prototype.tsx shows which screen renders each. Walk the states that mock mode can show: the empty state, and the validation error through the empty-submit probe. Mock mode cannot show a failed or a delayed state (a request failure, a slow dependency). Such a state is its own item, after its screen. Check it ONCE in the page source: name the condition and the element that renders it. Record it outside, with "mock mode cannot show it". Never try to force it: a forced failure costs more tool calls than the rest of the screen.

The manifest is your only map: it sets what you walk. You open a source file to repair, to find the route of a screen in no flow, or to check a state that mock mode cannot show. A component with no prototype.json gives you its registered routes instead, one flow per role. The prototype is the review's copy, not the app's code: never edit it.

Per screen

On each screen that a flow reaches, and on each screen in no flow, ask three questions of the controls that the role sees there. A screen that an earlier item answered under the same role and state needs only the controls that this flow uses. Each question is one action and one snapshot with the first plausible input. A screen is answered in a handful of actions; past that you are past the smoke.

Question Evidence
Reach Did the step before this screen in its flow bring you here, and does every link it shows land on the screen it names? the target's snapshot
Act Does every control the prototype shows change something visible when used? A create is in the next list as that list opens, with no filter changed by hand; a filter narrows; a toggle flips a row. the snapshot after the action
Request Did a change leave the page as the request the contract declares, with the status it declares? A row that flips and sends nothing is the one defect a build cannot see. agent-browser network requests

Once per app

  • Roles (mock/authz/roles.gen.ts exists): each flow's entry screen under its own role (?role=<id>), and once under a role whose roleIds leave it out. Both directions are defects. Then three more, which the scope model makes walkable and which no build can see:

    • Every role, and the no-role visitor. Walk ?role=<name> for each role in mock/authz/roles.gen.ts, then ?role= (empty — signed in, holding nothing). The empty one must give NoAccess replacing the shell, naming the groups to ask to be added to. A shell with an empty nav around it is a defect; so is being let into a screen.
    • Forbidden, by hand. Open one screen a role must NOT reach, by URL, under that role. Expect that screen's Forbidden state INSIDE the app shell — the rail the role can use still there — naming the role that would unlock it. Never a redirect to sign-in, never a blank page, never the screen itself.
    • The mock must have refused something. None of this means anything unless the mock's gateway layer is actually enforcing. It reads the handles out of openapi.yaml, so they cannot be wrong — but it can be OFF, and an app where it is off is green everywhere and proves nothing. Two checks, and both are cheap: the dev server prints how many operations it enforces and from which contracts at startup, and one refusal shows in agent-browser network requests as a bare 401 with the reason on the console ([mock gateway] 401 GET /claims — …). See one before you trust the rest.

    A screen a role is meant to reach but cannot, because the design does not grant it the handle the operation declares, is open, naming the role, the screen and the handle. It is a design defect, not a code one: do not widen a mock handler or a route guard to make it pass.

  • Session (an auth dependency): ?auth=out on one entry screen runs the app's own guard and signIn() brings you back; then Sign out where the prototype puts it (the user menu) leaves the screen through signOut(). The mock signs the next load in again, so the session being gone is outside; the click leaving the page is not.

  • Probes: submit one form empty; open one detail route with an id that does not exist. These are two states the happy path implies. The empty submit also walks state.validation-error where the manifest lists it.

  • Short window: reload ONE screen that has a primary action low on the page at agent-browser set viewport 1280 600, and click that action. A layout that only works on a tall window is a layout most laptops do not have, and the way it fails is silent: a sticky footer sits over the button, the content pane does not scroll, and the click lands on the footer. Set the viewport back afterwards.

  • Deep entry: load one nested route as a fresh page load in the address bar, not by clicking into it. Clicking never leaves the first document, so only this exercises the app booting at that URL — the path a bookmark, a shared link and F5 all take. A blank page is the finding; agent-browser errors names it.

  • Console (agent-browser console, read whole, once, before you stop): a page that renders and throws is broken for whoever touches it next, and the error text is the finding.

Outcomes

Every item ends in exactly one:

  • done — what you did and what the page did back.
  • fixed — what was wrong; what you changed; what it does now, re-walked.
  • open — what happens, after three attempts.
  • outside — the truth lives outside the app, or mock mode cannot show it. Outside the app: a computed total, a generated checklist, what the real IdP grants a real account. The mock answers to openapi.yaml, so it proves the request went out, never that the number is right; validation-task judges that against the deployed system. Mock mode cannot show a failed or a delayed state: name the condition in the source and the element that renders it. A refusal is not outside — the mock's gateway layer answers it from the contract's own handle, so Forbidden is something you walk and see.

An unreachable screen is open, naming the navigation that failed, never done read off the source. A flow whose start state the walk did not reach is open, naming that state: a flow whose control the screen did not show was not walked, and it is never done. A fix does not remove or hide what the prototype shows. If the app cannot fill an element the prototype shows, the item is open.

Progress

Your prompt says where progress goes. Post there, in these three shapes and no other: the plan once, one line per item as it settles, the close once.

Mock verification: <component> — <N> items
1. <flow id> · <flow name> (<role>): <Screen> -> <Screen>; starts at <the state its control needs>; states: <its states>
2. <Screen> · <state>: source check
3. <flow id> · ...
<k>. <Screen> (no flow): <its controls>; -> <the screens its `to`s name>
<N-4>. Roles
<N-3>. Session
<N-2>. Probes
<N-1>. Short window
<N>. Console

<Item> below is the flow id, the screen in no flow, the screen's state, or the once-per-app item.

<n>/<N> <Item>: done — <what you did and what the page did back>
<n>/<N> <Item>: fixed — <what was wrong>; <what you changed>; re-walked, <what it does now>
<n>/<N> <Item>: open — <what happens, or the start state not reached>, 3 attempts
<n>/<N> <Item>: outside — <the truth that lives outside the app>
<n>/<N> <Screen> · <state>: outside — mock mode cannot show it; <the condition> renders <the element>
Mock verification done: <component> — <N> items · <d> done, <f> fixed, <o> open (<n> <Item>, ...), <t> outside

The newest line is where the walk is, and that is what the person watching reads. An item with no line is a flow or a screen you never reached, and it stays visible as one.

1 · Stand it up

cd <App Path> && bash "$AEP_SKILLS_DIR/mock-verification/scripts/walk.sh" up

The cd is part of the command, not advice around it: the script keys its state off the working directory, so run from anywhere else it starts a second, unrelated server and hangs waiting for one that was never yours.

It starts npm run dev:mock on a free port, reaps a stale server from an earlier attempt first, and prints READY <url>. The url is what you open; ?role= and ?auth=out go on it. If it prints the log instead, that is your first finding: the harness is react-webapp's references/mock-mode.md; fix it and run up again.

Done when: READY.

2 · Plan

Before the browser opens, post the plan from the map alone: one numbered item per flow in scope, by its id in the manifest's order, with its role, its screens in walking order and the state it starts in, and one per state that mock mode cannot show, after the flow that reaches its screen. Then screens in no flow, the shell screens, Roles (with mock/authz/roles.gen.ts), Session (with an auth dependency), Probes, Short window, and Console. Posting it first puts your coverage in front of the person watching while there is still time to say a screen is missing.

Done when: every flow id in scope has a number, every screen in scope is in a flow item or has its own, and the plan is posted.

3 · Walk

Load agent-browser and follow it: open, snapshot, act on what the snapshot shows.

An item ends green, and posted. Walk it; if it fails, fix it now, walk that same item again, and see it pass; then post its line, and only then open the next. A line posted at the end for an item settled an hour ago is a history, not progress. Every failure is yours, whichever issue put it there. Repair the app, never the plan.

Three attempts on an item, then post it open and walk on. The screens behind it are still unopened. A fix is the wiring a screen is missing; a defect that wants a redesign is open with its cause named.

Click between screens. Mock state lives in the page, so open, reload and back restore the seed data: a record you created a moment ago is gone, and that is the mock telling the truth, not the app. Spend full loads at the start of a block, where there is no state to lose — a role switch, ?auth=out, an unknown id. The once-per-app items are walked off the plan like any other.

restart after a change to vite.config.ts, mock/ or a dependency; everything else hot-reloads.

Done when: every number has its line, each posted before the next item was opened.

4 · Close

Post the closing line, then:

bash "$AEP_SKILLS_DIR/mock-verification/scripts/walk.sh" down

down stops the server, closes the browser and confirms the port let go. Hand back to whoever dispatched you the closing line and the numbered list with each item's outcome: the closing line is what the pull request quotes, and the open items are what it carries.

Done when: the close is posted, down printed STOPPED, and the list is in your reply.

Never

  • Make the mock agree with the app. mock/handlers.ts answers to openapi.yaml, the same document src/generated/ came from, and to nothing else. A handler bent until a screen passes hides the defect from the deployed system too. A 501 is a handler you never wrote: write it against the contract.
  • Widen a scope to make a walk pass. Not in mock/authz/roles.gen.ts, not in a route guard, and not by adding a check to a handler that overrides the gateway layer. Scope comparison is a whole-string match at the gateway, so a mock bent to accept a sibling handle passes a walk the deployed system 401s. A role that cannot reach its own screen is a design defect: post it open, naming the role, the screen and the handle.
  • Run git, commit, or open a pull request. The record belongs to the agent that dispatched you. Progress, where your prompt says it goes, is the only writing you do outside the App Path.