App Builder

Plugin check: Run node "${PLUGIN_ROOT}/scripts/check-version.js" — if it outputs a message, show it to the user before proceeding.

app-builder — intent → model-driven app

Turn a natural-language intent into a deployed model-driven app. You author a reviewable App Spec (JSON) with the user across confirmed turns, then a deterministic engine (cds-maker-sdk, vendored) builds it — tables/columns/relationships, sample data, views, Choice-column charts, adaptive forms with sub-grids, generative pages for overview/dashboard surfaces, and the app module + sitemap. The same spec drives create and edit: download a deployed app back into a spec, change it, and re-run the build (it's idempotent).

CRITICAL — run the interactive flow in THIS conversation (the main loop)

You MUST run the authoring questions and the build narration yourself, in the main conversation. Do NOT dispatch a subagent (Task) for the interactive steps.

A subagent is headless — AskUserQuestion and plan mode do not reach the user from inside one (its only output is its final message). The whole point of this skill is the multi-turn, propose-then-confirm experience, so every AskUserQuestion, EnterPlanMode, and live build status line must originate here, in the main loop.

CRITICAL — the user sees your chat message, NOT tool output

Shell/tool output — the result of running preview-app.js, a dry-run plan, or a lint — is COLLAPSED BY DEFAULT in the UI. The user does NOT see it unless they manually expand the tool panel. Running the command is therefore NOT the same as showing the user. Every artifact the user must read, review, or approve — the whole-app preview wireframes, the dry-run build plan, and blocking lint findings — MUST be reproduced verbatim in your chat reply, inside a fenced ``` code block. Never say "the preview looks right" and leave the content buried in a collapsed panel: paste it into your message. This is the #1 cause of "the wireframes aren't visible" — the preview ran, but its output stayed hidden.

Capabilities — the full toolbox (pick best-fit per requirement)

You are a complete model-driven app builder, not a single-surface tool. Everything below ships in one App Spec and one build — choose what best serves the user's requirement to make a useful, prod-ready app; don't under-build (a bare table list) or over-build (surfaces nobody asked for):

• Data model — tables (give each custom table a meaningful Fluent-style SVG table icon by default; propose what the glyph will depict in words — never a Fluent token name — and record it as iconDescription before drawing the SVG — see references/authoring-flow.md → Table icons), columns (all types), relationships (1:N / N:N + junctions), sample data
• Record UI — forms (sub-grids, quick-create / quick-view), views (with enriched default columns), charts
• Custom grid rendering (preview) — entities[].columns[].visualization: render a column as a RadialDial, LineChart, HeatMap or StarRating in every grid and view that shows it, instead of plain text. Reach for it when a column is a magnitude a user scans (a score, a utilization %, a rating, a priority) rather than a value they read exactly — it makes a list scannable at a glance for one line of spec. It is per-column, so set it once on the column, not on each view. Preview: on an environment where it is not provisioned the build skips it and everything else still deploys, so it is always safe to include.
• Actions — modern command-bar buttons (incl. flyout / split menus), web resources (form JS / HTML / CSS). Two rules when writing that JS, both learned from buttons that deployed perfectly and then did nothing: a command handler is handed the record (function doThing(primaryControl)) — the build supplies the parameter, so write that signature; and never hardcode Choice values like 100000003, because they are assigned per publisher. Resolve by label via getOptions() (see references/app-spec-schema-advanced.md → webResources). Note also that command and web-resource edits do not redeploy on rebuild — the phases reuse what exists, so changing a button or a script means deleting it first.
• Form logic without code — businessRules[]: show/hide, lock/unlock, set-required and set-value, driven by a condition on the record. Reach for a business rule before form JS when the requirement is field-level and declarative — it is visible in the maker, survives solution export, and needs no web resource. Use form JS when the logic needs a real API call, cross-record work, or anything beyond the four supported actions.
• Guided processes — businessProcessFlows[]: the staged bar across the top of a record (ordered stages, each with steps bound to that table's columns — every step must bind a field; the platform rejects one without, so use a Boolean flag for a manual check-off). Reach for one when the user describes work moving through phases — "triage → investigate → resolve", "lead → qualify → close". Author it Active (the default) or the stage bar does not appear at all. Who may run a flow is securityRoles: { "personas": [...] } (Active flows only). v1 is single-entity and linear: cross-entity stages, branching and stage actions are rejected by the spec gate, so offer Maker for those rather than writing them into the spec.
• Surfaces — generative pages (modern dashboards / overviews / analytics / landing — the default), classic dashboards (opt-in), external URLs
• App shell — the app module + sitemap, with per-subarea icons. Turn on the modern shell with app.newLook: true unless the user asks for the classic one; it is opt-in and best-effort, so a tenant without the setting still gets a working app. app.headerNavigationRefresh controls the Wave 2 header/navigation refresh — a separate, independent setting whose platform default is ON, so set it to false only when the user explicitly wants the classic header.
• Security & access — one security role per persona, sized from that persona's jobs-to-be-done (the entity access each job needs, unioned into the role), so the app opens for non-admins.
• AI-first features (admin-gated) — form-fill assist, natural-language grid/view search, NL chart / AI data visualization, M365 Copilot (opt-in); per-table Copilot row summaries (Insight Cards) with tailored prompts, auto-selected for good-candidate tables

Author the smallest spec that fully satisfies the ask, then let the user refine. The Genpage-first policy below is the record-vs-dashboard rule; references/app-spec-schema.md documents every field.

Genpage-first policy (surface classification)

Every app surface is one of two kinds — enumerate the app's surfaces and classify each one, don't decide page-by-page in passing:

• Record surfaces (create/read/update/list a table's rows) → a model-driven form + view.
• Everything else — overview/landing, dashboard, KPIs, analytics, guided or wizard flow, composite or comparison screen → a generative page (pages[]), not a classic dashboard.

Rules:

• An app whose jobs include an overview, a queue, analytics or a guided flow but which proposes no pages has missed a surface. Where the app genuinely is record-CRUD only, say so explicitly rather than silently omitting pages.
• A traditional dashboards[] is emitted only on explicit request (e.g. "use a classic dashboard").
• A generative page is authored as a design intent (source: { kind: "intent" }, schemaVersion: 2) during Phase 1 — its .tsx is written in Phase 1.5 — Generate pages, after plan approval and after the data pre-build creates the tables so pac model genpage generate-types can emit RuntimeTypes.ts. See references/authoring-flow.md → Pages.
• The build's pages phase uploads each page via pac model genpage upload without --add-to-sitemap — the SDK is the single sitemap writer, so a page's nav entry comes from a page subarea in appShell (referenced by the page's key). See references/app-spec-schema.md → pages[].
• Every page in pages[] must be sitemap-placed — validation rejects any page absent from the sitemap. A "detail" page that receives a caller-supplied id is a normal sitemap page; it reads its input via pageInput?.data?.<field>. Navigation-only (headless) pages are not supported.
• A page that declares pageInput MUST declare directEntry. Because every page is sitemap-placed, a detail page is also reachable straight from the app navigation with no input — a state a user reaches by clicking. Say what happens then: { "behavior": "selector" } shows a picker and then the record, { "behavior": "emptyState" } explains and renders nothing broken. Prefer selector when the table is browsable; it is the more useful landing. Every key in pageInput.data must also be produced by some page's navigatesTo[].data, or the generated page reads a key nothing ever sets.
• Three-authority page identity (build + download + verify all follow this): (1) IDENTITY — the durable <app>_pagemanifest (key → pageId); a downloaded spec's own pages[].pageId outranks it for that rebuild. (2) EXISTENCE — env-wide pac model genpage list (crash-safe; decides create-vs-reuse). (3) MEMBERSHIP — the app's sitemap GenPageId set (placement, download enumeration, verify coverage). All matching is by id — never by display name.
• Multi-page navigation uses PAGEREF_<key> (the stable pages[].key) as the pageId placeholder; the build resolves each to the real page GUID in a run-scoped staging copy (the canonical .tsx is never mutated), then verifies every nav edge resolves to the deployed page's GenPageId.

Workflow

Phase 0 — Working directory

1. Derive a short kebab-case slug from $ARGUMENTS (e.g. "Project Tracker" → project-tracker).
2. Create the directory (mkdir -p <slug> on bash/PowerShell; mkdir <slug> on cmd, which has no -p and errors if it exists) and resolve its absolute path. It holds app-spec.json, model-app-plan.md, and workflow-log.md.

Phase 1 — Author the App Spec (interactive, main loop)

Follow references/authoring-flow.md step by step, running every prompt yourself via AskUserQuestion. In short:

Unattended runs (Copilot autopilot / automation)

Resolve the interaction mode once before the first authoring question:

node "${PLUGIN_ROOT}/scripts/resolve-interaction-mode.js" [--non-interactive]

POWER_PLATFORM_SKILLS_NONINTERACTIVE=1/true (or the equivalent --non-interactive invocation) means there is no user waiting for prompts. In that mode:

• do not call AskUserQuestion, EnterPlanMode, or ExitPlanMode;
• use explicit requirements and supplied existing specs as authoritative, and use documented defaults only where the request leaves a non-destructive choice open;
• record each skipped gate as Unattended default: <question> → <answer> (<reason>) in workflow-log.md;
• treat the completed lint, preview, rendered design document, and build dry-run as the approved plan, then continue with the normal approved path;
• halt rather than guessing an ambiguous environment, app identity, or destructive structural edit.

Suppressing interaction never authorizes destructive work. An unattended existing-app apply still requires --allow-destructive wherever the build normally requires it; POWER_PLATFORM_SKILLS_NONINTERACTIVE only suppresses prompts.

Carry the resolved mode through subprocesses: when it is unattended, append --non-interactive to every build-model-app.js invocation (dry-run, data pre-build, recovery rerun, and full apply). Do not append --allow-destructive unless destructive authority was supplied independently.

1. Prereqs — node --version, pac help (> 2.10.0, the plugin's minimum).
2. Environment (PAC) — pac auth list. If exactly one / an active profile, confirm it (FYI), don't ask. If several and none active, ask which to use. If none, ask the user to pac auth create. Capture the org URL (pac org who).
3. Detect existing — pac model list-tables --search … (exact-match) and pac model list to find tables/apps already present; build around them.
4. Levelled authoring — first read the App Spec format so you author to the exact shape (do this once; don't go spelunking through scripts): references/app-spec-schema.md and the worked sample samples/app-spec.support-desk.json. That document covers everything an app always has. Additionally read references/app-spec-schema-advanced.md once your design calls for a conditional feature — business rules, a business process flow, command-bar buttons, web resources, global choices, classic dashboards, or roleGrants[]. Its pointer table is in the main schema, and the toolbox above tells you when to reach for each; read the detail only for the ones you are actually using, and never skip a capability just to avoid the read. Phase 1 is design-only: never emit page .tsx here. Each level is confirmed via AskUserQuestion before the next begins, and app-spec.json is persisted after each — full prompts in the playbook.
   • Level (a0) — personas & jobs-to-be-done (personas[]): before proposing any tables, ask who will use the app and what each needs to get done. Jobs drive everything that follows — a table exists because a job needs its data, a surface because a job needs to act on it. Deriving the data model first reliably misses surfaces. Privileges come later; capture the jobs now.
   • Level (a) — data model: entities/columns/relationships derived from those jobs; run the early data-model lint (catches e.g. the relationship-vs-lookup collision before forms are authored on top).
   • Descriptions are part of authoring, not a cleanup pass. Every table, column, view, chart, form, dashboard and business rule takes an optional description, and you should write one as you create the artifact — never as a backfill. A name says what a thing is called; a description says what it is for, and it is the only intent an app carries that an agent can read back later when it inspects an app it did not build. Describe the purpose, not the shape: "Severity 1-5; drives the escalation rule and the SLA clock", not "The priority column". (commands[] and Customer columns accept one but the SDK cannot write it — you'll get a warning; personas[] does not take one at all. See the schema reference for why.) The app itself takes two: app.description is the short tile text, and app.aiDescription is the routing description an orchestrator reads to choose between apps — who the app is for, what it covers and excludes, and how to tell it from a sibling app over the same tables. Write it whenever such a sibling exists or is planned.
   • Level (b) — artifacts + page-intents + design: enumerate every surface each job needs and classify it per the genpage-first policy above — record CRUD → form + view; anything else (overview/landing, dashboard, KPIs, analytics, guided/wizard flow, composite or comparison screen) → a page intent (source: { kind: "intent" }). State each classification out loud; if the app needs no pages, say so and why rather than silently omitting them. Then forms + views
     • charts + sample data and the optional design contract, mapping each job to its surfaces via jobs[].surfaces[]. No page .tsx here. dashboards[] only on explicit request.
   • Level (c) — access (personas[].jobs[].privileges[]): personas and jobs already exist from (a0) and the entities now exist, so only add the privileges each job needs — don't re-ask who the users are. The builder unions them into one role per persona and grants the app to it so it opens for that persona, not just sysadmins. Render the roles + per-entity access as a table in your chat reply (the user can't approve an access model they can't see — see the CRITICAL note above). If they want no roles, tell them jobs live in personas[] so dropping the roles drops the recorded jobs too; offer read-only privileges to keep the record. (Column-level security and access teams are not yet supported — see Notes & limits.)
   • Whole-app preview (design gate for Level (b)): node "${PLUGIN_ROOT}/scripts/preview-app.js" --spec @<working-dir>/app-spec.json renders data-model + sitemap + form wireframes + page-intents + design contract. **Reproduce the ENTIRE rendered output verbatim in your chat reply, inside a fenced ``` code block — do NOT leave it in the (collapsed, invisible) tool output, and do NOT just summarize "the preview looks right" (see the CRITICAL note above).** The user must be able to SEE each form, the sitemap, and the page intents they are approving. For a single form only: node "${PLUGIN_ROOT}/scripts/preview-form.js" --spec @<working-dir>/app-spec.json.
   • Don't pre-create tables/columns — the build does it idempotently.
5. Guardrail lint (hard gate) — run the full lint on the complete spec; errors block, warnings teach. If it blocks (or warns), paste the findings into your chat reply — tool output is collapsed and invisible to the user (see the CRITICAL note above), so the user can't fix what they can't see:

   node "${PLUGIN_ROOT}/scripts/lint-app-spec.js" --spec @<working-dir>/app-spec.json --json

   This runs migration and validateAppSpec — the gates the build itself runs on load — plus the lintAppSpec authoring guardrails, which the builder does not run. It exits non-zero on errors. It validates under the plan profile (which allows not-yet-generated intent pages), matching Step 6's dry run.
6. Plan-mode approval (the single build approval) — present the plan including the build dry-run's phase-grouped plan (run build-model-app.js without --apply, using the plan profile that allows intent pages) inside EnterPlanMode, then ExitPlanMode to get the user's go-ahead. On approval, Phase 1.5 (generate-pages) runs first, then Phase 2 applies directly (no second dry-run/go-ahead). Render the design document — never hand-write it:

   node "${PLUGIN_ROOT}/scripts/write-app-spec-doc.js" --spec @<working-dir>/app-spec.json --env <envUrl>

   It writes <working-dir>/model-app-plan.md (jobs → surfaces traceability, data model, every surface, navigation, access model, sample data) and prints { ok, docPath, bytes, warnings }. Surface those warnings — they name design gaps such as a job with no covering surface or an app with no generative pages. Tell the user where the document is; it's theirs to keep, and it's regenerable after any spec edit.

Phase 1.5 — Generate pages (main loop, headless workers)

After plan-mode approval (before the full build):

1. Data pre-build — schema-only build so generate-types can resolve real column names:

   node "${PLUGIN_ROOT}/scripts/build-model-app.js" \
     --env <envUrl> --spec @<working-dir>/app-spec.json --stage data --apply

   --stage data applies solution + data-model only — no --sample-data (rows are created once in the full build). Only --stage data is apply-safe; all other --stage selectors and legacy --from/--to/--only/--skip selectors are dry-run inspection only.

2. Types — generate Dataverse type bindings for the entities the pages read:

   pac model genpage generate-types --data-sources "<entity1,entity2,…>" --output-file <working-dir>/RuntimeTypes.ts

   --data-sources is the union of every intent page's dataSources. Skip this step entirely when every intent page is mock-only. On Windows use forward slashes in the path.

3. Page plan (adapter) — the page worker's input contract is a plan document, not an App Spec, so project the spec into one. This also echoes the per-page dispatch parameters:

   node "${PLUGIN_ROOT}/scripts/write-page-plan.js" \
     --spec @<working-dir>/app-spec.json --working-dir <working-dir> --env <envUrl> \
     --app "<app name>" --languages "<languages from the environment probe>"

   It writes <working-dir>/app-builder-page-plan.md and prints { ok, planPath, pages: [{ name, key, file, dataMode, intent }] }. Pass --languages through from the environment probe — omitting it silently defaults every plan to English-only and drops the localization pattern. The command fails (before writing) if the plan would name a sample that doesn't exist, or if a page a worker will write is not a safe target in the working directory — a link at its path, a folder that links outside the directory, or a page already there under another spelling — or if a link, a hard link or a folder is already at the plan's own path. Stop and fix what it names; do not dispatch any worker.

4. Generate — for each page from step 3 with intent: true, dispatch the headless genpage-page-builder worker via Task. Use its documented input contract verbatim — a missing field is why a page silently never becomes .tsx:

   You are the genpage-page-builder agent. Generate the [name from step 3] page.

   • Target file: [file from step 3 — already includes .tsx; do NOT append another]
   • Plan document: [absolute path to the app-builder-page-plan.md written in step 3]
   • Data mode: [dataMode from step 3 — dataverse or mock]
   • Connectors: none
   • Telemetry: disabled
   • RuntimeTypes: [absolute path to RuntimeTypes.ts] ← omit this line when Data mode is mock
   • Working directory: [absolute working-dir path]
   • Plugin root: ${PLUGIN_ROOT}

   Follow the instructions in your agent file. Write [file] and return your result when done.

   The plan's ## Environment carries Mode: app-builder and every page row carries a Key, so the worker emits "PAGEREF_<key>" for cross-page navigation (never a file-derived token — a downloaded page's codeFile is a path, not its identity). Custom nav ids go in data: — never recordId. Connectors: none and Telemetry: disabled are constants here: the App Spec has no connector-binding concept (so the projected plan always says No connector bindings.), and /app-builder never runs the Phase 4.7 custom-telemetry probe. Both are stated explicitly rather than omitted — the page-builder treats a missing line as the same fail-closed value, but an explicit line is what makes the dispatch contract checkable.

5. Validate + commit the transition (transactional) — never flip source by hand, and never flip pages one at a time as workers return. Run:

   node "${PLUGIN_ROOT}/scripts/promote-intent-pages.js" \
     --spec @<working-dir>/app-spec.json --working-dir <working-dir>

   It checks every generated page (file written, structurally a module, PAGEREF_ tokens canonical and in exact parity with the spec's navigatesTo edges) and only then flips all of them intent → { kind: "tsx", codeFile } in a single atomic write. On any failure it exits 3 and leaves app-spec.json untouched, printing which page failed and why — regenerate just those pages and re-run. All-or-nothing on purpose: a half-flipped spec would claim a .tsx that was never written, and Phase 2's deploy profile fails fast on any remaining source.kind === "intent".

6. Proceed to Phase 2 (full idempotent build).

⚠️ The interactive author never runs inside a Task subagent. Only pure, headless code-gen workers are dispatched here — all user-facing prompts originate in the main loop.

Phase 2 — Build (narrated, main loop)

Always use scripts/build-model-app.js. Never hand-write a builder. It's idempotent (skips existing solution/tables/columns/relationships — so new, existing, and mixed envs all just work), so you don't pre-create anything or special-case existing tables.

The build plan was already presented and approved in plan mode (Phase 1 Step 6 shows the engine's real dry-run plan), so on approval apply directly — one build approval, no second go-ahead. Add --verify so the build self-checks after applying (see Phase 3):

node "${PLUGIN_ROOT}/scripts/build-model-app.js" \
  --env <envUrl> --spec @<working-dir>/app-spec.json --apply --verify [--sample-data] [--publish]

Each step streams its status live ([n/total] ✓ created / ⊘ skipped / ✗ failed — <error>) and a closing ✓ build complete — X created, Y skipped, Z failed summary.

Keep the build's progress visible. A full build runs for several minutes. Let its output stream — do NOT pipe it through Select-Object -First/-Last N or Select-String head-limits, which buffer and hide progress until the run ends (and can truncate a still-running pipe). To capture the log use Tee-Object -FilePath <log> (no head-limit). The build also prints a ▸ live progress: line pointing at <workspace>/.maker-workspace/build-status.json — a snapshot (state, steps, lastPhase, lastLabel) overwritten every step. Read it (or tail build-log.jsonl) any time to report where a long build is, even if stdout is buffered.

(Reaching Phase 2 without a fresh plan-mode approval — resuming a failed build, or a quick edit re-run — do a dry-run first (drop --apply), paste the phase-grouped plan verbatim into your chat reply (tool output is collapsed — see the CRITICAL note above), and get a go-ahead before applying.)

Stage selector (--stage <data|ui|app|publish>) maps to its phase range. On --apply, ONLY --stage data is accepted (solution + data-model, no rows in run 1; run 2 is a full build). All other stages and the legacy --from/--to/--only/--skip selectors are dry-run inspection only — their phase ranges are not dependency-closed and are rejected on --apply.

Language (--language-code / --languageCode <lcid>) — you normally do not pass this. Data-model labels are stamped with the organization's own base language, read once per build. Pass it only to author labels in a different provisioned language, or if the build warns that it could not determine the base language and fell back to 1033.

If you pass an LCID the organization has not provisioned, the build now stops in the data-model phase, before any label is written, and lists the ones it does have. That check exists because Dataverse handles this inconsistently: table and choice labels are accepted (HTTP 204) and silently stored under the organization's base language, while DateTime and Memo columns are rejected with "The language code N is not a valid language for this organization". Without the check a build therefore creates the table with the wrong labels and then dies partway through the columns, phases away from the flag that caused it — which reads like an environment fault. List the provisioned set with node "${PLUGIN_ROOT}/scripts/dataverse-request.js" <envUrl> GET RetrieveProvisionedLanguages. The check is best-effort: if that read fails the build proceeds unchanged.

The App Spec field languageCode pins the same value across runs. It now covers the whole build: data-model labels, and form, dashboard and sitemap labels — the SDK serializers that used to hardcode 1033 take the authoring language as an option.

Narrate progress as it runs. Transient env errors (429 customization-lock, 503 SQL-timeout, a SQL deadlock, concurrent-op guards) are auto-retried with backoff on --apply (the build is idempotent, so a retry reuses what's already created). If the build still halts (BuildHalt) on an unrecoverable error, surface it and ask the user how to proceed via AskUserQuestion (adjust the spec / cancel), then re-run. Everything is scoped to a dedicated unmanaged solution; --publish gates the final bulk publish — edit/finalize paths still publish their one targeted artifact so the change takes effect (see Notes & limits).

Recovery from a failed or halted build: run the full build again. The build is idempotent — every phase re-uses what's already created and only fills the gaps. SDK metadata is persisted under <working-dir>/.maker-workspace/ (override with --workspace). There is no apply-safe --from <phase> shortcut; a full rerun is the correct and safe recovery path. One exception: an already-exists halt recurs on a rerun until that workspace directory is deleted — follow its message.