Create Mobile App

📋 Shared instructions: shared-instructions.md — read first. Covers safety guardrails, memory bank usage, preferred-environment policy, connector-first rule, Windows CLI compat, command-failure handling.

Create Power Apps Code App (Native)

Top-level orchestrator. Owns the user-visible flow; delegates planning to the native-app-planner agent and per-domain mutation to dedicated /add-* skills.

Workflow

0. Resume check + fresh-template gate → 1. Prerequisites → 2. Gather requirements → 2b. Requirements discovery → 2c. Plan preview (rough cost + abort gate) → 3. Plan (planner agent + 4 gates) → 4. Auth & environment → 5. Prepare existing template → 6. npx power-apps init → 6.5 verify npm install → 6.5b SafeAreaProvider gate (always runs, idempotent) → 6.6 scaffold tsc smoke check → 6.7 seed memory bank → 6.85 Offline profile (always asked) → 7. Auth config → 8. Apply data model → 9. Apply native capabilities → 9b. Design system → 10. Add connectors → 10b. Wire navigation layout → 11. Build screens (parallel) → 11.4 Stylistic fix sweep → 11.5 Dev smoke test → 12. Start dev server (npx expo start) → 13. Summary

Fresh-template working-directory mode

This skill assumes the user already has a fresh pa-wrap-tools/templates/expo-app-standalone template materialized with degit in the target working directory and has already run npm install there. The skill turns that fresh template into an app; it does not clone, degit, or copy a template itself.

Fresh template required. If the working directory is not a template, or if it already looks like an app created by this skill, STOP and tell the user to materialize a fresh expo-app-standalone template with degit into a new folder, run npm install, then rerun /create-mobile-app --working-dir <fresh-template-dir>.

Use these markers:

This gate is intentionally simple: /create-mobile-app creates a new app from a fresh template. It does not adopt, repair, resume, or overwrite an already-created app.

TypeScript Gate Policy — no quality compromise

tsc is a phase gate, not a reflex after every tiny edit. The app may not advance past a gate until TypeScript is clean.

Required gates:

• Scaffold gate: Step 6.6 after existing-template preparation, npx power-apps init, and dependency verification.
• Dataverse/generated-services gate: immediately after Step 8 returns and generated services/models are refreshed.
• Navigation/skeleton gate: after Step 10b layouts and Step 10.8 shared code/skeletons are written, before Step 11 builders launch.
• Screen-wave gate: after each Step 11 screen-builder wave returns, before launching the next wave.
• Final gate: before Step 12 starts the dev server.

When a gate fails:

1. Capture the full tsc --noEmit output once.
2. Classify errors by root cause (for example: generated model names, service option shapes, invalid UI props, typed percentage values, create/update payload typing, missing imports).
3. Repair in a batch.
4. Re-run the same gate once after the batch.
5. Continue only when the gate is clean, or stop/block according to the retry policy.

Do not run full-app tsc after every microscopic local edit inside the same repair pass. That is slower and encourages line-by-line patching. Batch root-cause fixes, then re-run the gate. This is a speed improvement only; it does not lower the quality bar.

Hard stops:

• Do not launch data-source work from a broken scaffold gate.
• Do not launch screen-builders from broken generated services, layouts, shared code, or skeletons.
• Do not launch wave N+1 until wave N passes its tsc gate.
• Do not start the dev server until the final gate is clean.
• Do not hide approved capability failures behind mocks or TODOs just to satisfy tsc.

Step 0 — Resume check + fresh-template gate

If $ARGUMENTS includes a --working-dir (or the user names an existing directory), check whether <working_dir>/memory-bank.md exists.

• Bank present → read it. Identify the highest-numbered completed step. Inform the user:

  "Found existing project '' at <dir>. Steps 1– already completed (last update ). Resume from Step <N+1>?" Wait for confirmation. If the user says yes, jump to that step. Skip the wizard (Step 2) and re-use the values stored in the bank.

• Bank absent → fresh project. Continue to Step 1.
• Bank present but corrupted (missing required headings) → surface the parse error, ask the user whether to overwrite (lose history) or fix manually before proceeding.

The bank is the only resume mechanism. Do not infer resume state from package.json or node_modules/ — those can lie.

After the resume check, run the fresh-template gate from the section above. This is a create-only command:

• If memory-bank.md exists and the user confirms resume, resume as documented above.
• If any already-created-app marker exists and there is no approved resume path, STOP and tell the user to materialize a fresh template into a new folder with degit.
• If required template files are missing, STOP and tell the user to materialize pa-wrap-tools/templates/expo-app-standalone into the working directory with degit and run npm install.
• If node_modules/expo is missing, STOP and tell the user to run npm install in that template folder before rerunning this skill.

Do not silently copy a bundled template over the user's folder. A fresh pa-wrap-tools-1 template may contain placeholder power.config.json with an empty environmentId; Step 5 removes that placeholder immediately before Step 6 runs npx power-apps init.

Step 1 — Prerequisites

Run all checks first — no point gathering requirements if the toolchain isn't ready.

Important: npm auth and Power Platform app auth are separate. The account used for npm install can be different from the account used by npx power-apps:

Renewing npm feed auth does not sign the user into npx power-apps. If the Power Apps CLI prompts for browser auth later, that is expected and unrelated to the npm/ADO feed token.

Then run the checks:

node --version                                      # v22+
npm  --version                                      # v10+
az account show --query "user.name" -o tsv          # Azure CLI logged in (needed for Dataverse helper scripts)
git --version                                       # optional

Do NOT probe Xcode, Java, Android Studio, or CocoaPods here. This plugin's flow is plan → scaffold → code → local Expo dev server. Build + deploy (npm run build / npx power-apps push) is a separate user-driven step via the /deploy skill. Local native compile is the user's choice and lives outside this skill (run the platform-specific native command directly when needed). See shared/version-check.md — only the Always required tier matters here.

Template-only rule: this skill no longer provisions npm feed tokens, PAT fallbacks, vendor fallbacks, or registry rewrites. The user must run npm install in the fresh template folder before invoking /create-mobile-app.

Capture target Power Platform environment for the remaining flow.

Source of truth for env selection: the generated power.config.json first, explicit environment ID second. In the normal template-folder flow, npx power-apps init runs first and writes the selected environment ID into power.config.json; read that ID and pass it to scripts/resolve-environment.js to resolve the Dataverse URL and tenant. If power.config.json is missing or has an empty placeholder environmentId, ask for an environment ID. A Dataverse URL is useful as a resolver fallback for existing apps, but it is not enough for npx power-apps init because init needs --environment-id.

TARGET_ENV="<environment-id-or-empty>"
if [ -z "$TARGET_ENV" ] && [ -f power.config.json ]; then
  TARGET_ENV=$(node -e "try { const id=require('./power.config.json').environmentId || ''; console.log(id); } catch { console.log(''); }")
fi
test -n "$TARGET_ENV" || { echo "✗ Environment missing. Provide an environment ID."; exit 2; }
ENV_JSON=$(node "${CLAUDE_SKILL_DIR}/../../scripts/resolve-environment.js" "$TARGET_ENV")
printf '%s\n' "$ENV_JSON" > .resolved-environment.json
ACTIVE_ENV_ID=$(node -e "const j=JSON.parse(process.argv[1]); console.log(j.environmentId || '')" "$ENV_JSON")
ACTIVE_ENV_NAME=$(node -e "const j=JSON.parse(process.argv[1]); console.log(j.displayName || j.environmentUrl || '')" "$ENV_JSON")
ACTIVE_ENV_URL=$(node -e "const j=JSON.parse(process.argv[1]); console.log(j.environmentUrl || '')" "$ENV_JSON")
ACTIVE_TENANT_ID=$(node -e "const j=JSON.parse(process.argv[1]); console.log(j.tenantId || '')" "$ENV_JSON")
test -n "$ACTIVE_ENV_ID" || { echo "✗ Environment ID missing. Provide the environment ID directly."; exit 2; }
echo "✓ Target env: $ACTIVE_ENV_NAME ($ACTIVE_ENV_ID)"
echo "✓ Target env URL: $ACTIVE_ENV_URL"
echo "✓ Target tenant: ${ACTIVE_TENANT_ID:-unknown}"

Orchestrator handling for exit 2: ask the user for their environment ID directly, then re-run the capture block above. Do not run npx power-apps init here; Step 6 owns initialization after the user confirms the target environment.

Stash $ACTIVE_ENV_ID, $ACTIVE_ENV_NAME, $ACTIVE_ENV_URL, and $ACTIVE_TENANT_ID for Step 2 (env confirmation), Step 6 (npx power-apps init), and Step 7 (auth.config.json tenant/environment cache). If parsing fails, ask for an environment ID again.

If resolve-environment.js cannot get tokens, run az login --tenant <env-tenant> in the foreground. If npx power-apps init later uses the wrong account, follow shared-instructions standalone CLI auth handling and retry once.

Step 1.7 — Detect publisher prefix

Detect the publisher prefix for the env's Default solution so the planner uses the correct prefix rather than assuming cr_.

node "${CLAUDE_SKILL_DIR}/../../scripts/detect-publisher-prefix.js" "$ACTIVE_ENV_URL"

Output is one line of JSON, e.g.:

{"prefix": "cr8142a", "source": "detected"}
{"prefix": null, "reason": "no token (run `az login --tenant <env-tenant>`)"}

The script queries the Default solution's publisher via: /api/data/v9.2/solutions?$select=uniquename&$expand=publisherid($select=customizationprefix)&$filter=uniquename eq 'Default'

A second solution name can be passed as a second argument if the env uses a different solution (defaults to 'Default').

Token tenant note: the script's getAuthToken discovers the env's tenant ID from the Dataverse HTTPS auth challenge and passes --tenant <env-tenant> to az, so detection works even when the active az identity is on a DIFFERENT tenant. If the user has not run az login --tenant <env-tenant> at any point, detection may return null.

Stash the result for Step 3 (planner spawn):

Do NOT block on null detection — the user can still proceed; the Power Apps CLI normalizes prefixes when npx power-apps add-data-source runs. The detection step is purely to make the plan output accurate.

If the script exits non-zero (rare — should always exit 0 with prefix: null), treat it as the null case and continue.

Step 2 — Gather requirements

Skip questions the user already answered in $ARGUMENTS.

If the user gave no description, ask one open-ended question first:

"What would you like to build? Describe it in your own words — what it does, who uses it, and what problem it solves."

Then collect with AskUserQuestion (batch where possible):

App slug is auto-derived from the display name (slugify(displayName) — kebab-case, ASCII-only, strip non-alphanumerics). Do NOT ask the user; the derived slug is correct >95% of the time. Show the resolved slug as part of Step 2c's plan preview so the user can override via edit if needed.

Environment override branch: If the user picks "use a different environment", ask for the Power Platform environment ID via AskUserQuestion, then run scripts/resolve-environment.js again and refresh $ACTIVE_ENV_ID / $ACTIVE_ENV_URL / $ACTIVE_TENANT_ID.

App-name collision pre-flight. Once <displayName> is fixed, check the chosen env for a name collision:

npx power-apps list-codeapps --environment-id "$ACTIVE_ENV_ID" --json 2>/dev/null | grep -F "<displayName>" >/dev/null && \
  echo "COLLISION" || echo "OK"

If COLLISION, ask the user via AskUserQuestion:

"An app named <displayName> already exists in <ACTIVE_ENV_NAME>. Choose:

1. Pick a different name (recommended)
2. Delete the existing app in Maker portal — DESTRUCTIVE, asks confirmation outside this skill
3. Continue anyway (bg npx power-apps init will fail; you'll have to rename later — NOT recommended)"

Re-prompt for name if (1). If (2), send the user to Maker portal to delete the existing app, then re-run the collision check. Only proceed once collision is resolved.

If npx power-apps list-codeapps is unavailable in the installed CLI version, skip the pre-flight silently and continue.

Don't enter plan mode here — that's the planner agent's job in Step 3.

Step 2b — Requirements discovery

Goal: Turn the user's thin prompt into a confirmed feature brief before the planner runs. The planner agent receives this brief verbatim — richer input means better data model inference, accurate connector detection, and correct screen specs.

Step 2b.0 — Prompt richness scoring (decides which path to take)

Before asking anything, score the description on four signals. The score decides whether we ask a multi-select feature picker, a single confirmation, or skip the discovery question entirely.

Run this scorer mentally on <description> (the prompt the user gave with /create-mobile-app, plus any clarifying text from Step 2a). Count how many of the four trip:

Tier the result:

Print the chosen tier so the user knows which path is running:

"→ Prompt richness: 4/4 — skipping discovery questions, extracting brief and going straight to the plan-cost preview." (or 3/4 / ≤2/4 with the matching path name)

--full-discovery escape hatch: if $ARGUMENTS contains --full-discovery, force walk-through regardless of score. Use this in dogfood runs where you want to exercise the multi-select path.

--no-discovery escape hatch: if $ARGUMENTS contains --no-discovery, force auto-plan regardless of score. Use this for fully-headless runs from the wrapper templates repo.

Step 2b.1 — Walk-through path (only when tier = walk-through)

Read references/requirements-discovery.md. Infer context-aware options from the user's description, ask exactly one structured AskUserQuestion, and never use markdown checkboxes in the question text.

Wait for the user's response. Summarize their answers into a requirements brief — 4–8 bullet points covering what users can do, what data is tracked, and integrations.

Confirm once:

"Here's the brief I'll use for planning: • (bullet 1) • (bullet 2) ... Look right? (yes / adjust)"

Store the confirmed brief as <requirements_brief>. This replaces the thin $ARGUMENTS as the primary input to the planner.

Step 2b.2 — One-tap path (tier = one-tap)

Skip the multi-select question. Extract a 6–10 bullet brief directly from <description> covering: user roles, key entities, primary workflow, severity / status enums if present, integrations / connectors, native capabilities, and any explicit constraints. Show it with a single confirm:

"Your description is detailed enough to skip the feature picker. Here's the brief I extracted: • (bullet 1) • (bullet 2) ... Look right? (yes / adjust / start over)"

• yes → store as <requirements_brief>, fall through to Step 2c.
• adjust → drop to Step 2b.1 (walk-through) so the user can edit via the multi-select.
• start over → return to Step 2a and re-prompt for the description.

Step 2b.3 — Auto-plan path (tier = auto-plan)

Skip both the multi-select AND the brief confirmation. Extract the brief silently and store it as <requirements_brief>. Print it as a transparency log only:

"→ Auto-plan tier (4/4). Extracted brief from your description: • (bullet 1) • (bullet 2) ... → Going straight to the plan-cost preview (Step 2c). The brief above is locked in unless you abort there."

Do not ask for confirmation here — the user agreed to this when their prompt scored 4/4. The plan-preview gate at Step 2c remains in force as the last cheap exit before any side effects.

Step 2b.4 — Common to all paths

Auto-proceed after yes (or after auto-plan transparency log). Fall through directly to Step 2c (plan preview). Do NOT add a separate "Proceed to planning?" prompt — the brief confirmation IS the planning go-ahead. The only abort gate after this is Step 2c's proceed/edit/abort block, which is intentionally distinct because it shows the rough cost estimate.

Design decisions are deferred to Step 6.75 — /design-system (ships with this plugin) handles brand inputs, the style picker, and visual companion preference in one flow after the project is scaffolded. Do NOT ask design questions here.

Set tentative defaults (used by Step 3b before /design-system runs):

• <visual_companion> = yes — open _plan_preview.html in browser at Gate 4 by default. /design-system at Step 6.75 may downgrade this to no (path (d) in its cost picker), persisted to memory-bank for future runs.
• <design_vibe_opt_in> = deferred — Step 6.75 sets the real value. While deferred, the planner does NOT prompt for a direction; it writes a placeholder ## Design Direction: <deferred — set by /design-system> block so screen-planner can still run.

--no-design escape hatch. For headless / token-constrained runs, set --no-design in $ARGUMENTS. It forces <visual_companion> = no, skips the style-picker handoff at Step 3a entirely, and short-circuits Step 6.75 to a no-op (placeholder block stays in native-app-plan.md; screen-builders fall back to industry-inferred defaults).

Step 2c — Plan preview (rough, always shown)

Goal: Give the user a cheap exit before any mutation happens. This is the last point in the flow with zero side effects — no git clone, no npm install, no npx power-apps init, no agent tokens spent on planning. After Step 3 starts, every abort gets more expensive (half-written native-app-plan.md, partial _screens_section.md, architect tokens already burnt).

Always runs. There is no --no-preview flag in v0 — we need calibration data (~10+ runs with recorded estimate-vs-actual) before we can trust the rough estimates enough to let users skip them. Once the data shows estimates are reliably within ±50%, evaluate adding a skip flag for repeat-user workflows.

Compute the estimates from inputs already in hand (no agent spawn — pure heuristics on the confirmed brief and the wizard answers):

Print the block once, exactly in this format (substitute computed values; ranges as low-high):

─── Plan preview (rough) ─────────────────────────────────
Based on your confirmed brief, before any agent runs:

Scope (proxy estimates — actual numbers come from architects):
  Tables       ~<low>-<high>      ← from <N> nouns in brief; architect may merge/split
  Connectors    <N> inferred      ← <comma-separated names>  (confirm at Gate 3)
  Screens     ~<low>-<high>       ← from <N> features × ~2-3 screens each
  Approval gates  4               ← fixed (data model, native, connectors, screen plan)

Time (rough — agent time only, excludes your approval latency at gates):
  Planning      ~<low>-<high> min ← architects + your gate approvals add to this
  Scaffolding   ~1-2 min          ← validates prepared template + runs power-apps init
  Screen build  ~<low>-<high> min ← parallel, capped at 5 concurrent

Token tier: Opus everywhere in v0 (model routing not yet shipped).

⚠ These are proxies, not measurements:
  • Table count is "noun count in brief" — architect may collapse or split
  • Time excludes your approval latency at the 4 gates
  • If industry inference is low-confidence, +1 picker prompt
  • If you opted into the design vibe picker, +1 prompt + planner re-spawn
  • If any gate is rejected, that section regenerates (~2-3 min each)

Proceed, edit brief, or abort? [proceed/edit/abort]
─────────────────────────────────────────────────────────

Three-option exit:

Why "always show" is correct in v0 (do not skip without explicit user request):

• Cost when user proceeds: ~30s (read + decide). Token cost ~500/run = ~$0.008.
• Cost when user aborts late (after Step 3 starts): 5-10 min + dirty working dir + frustration.
• Asymmetry: bounded 30s vs unbounded 30 min. Always show the bounded cost.
• Forced calibration: every run produces the <estimate, actual> data we need for v0.x model routing decisions. Skipping drops calibration data.

Set expectations before handing off to the planner:

"Brief locked in. Planning surfaces 4 approval prompts (data model → native capabilities → connectors → screens). Total time is 5–12 minutes: • Gate 1 (data model) — ~60–90s • Gate 2 (native capabilities) — ~10s (quick) • Gate 3 (connectors) — ~30–60s • Gate 4 (screens + design) — 3–8 minutes (this is the heavy one: design vibe picker if opted in, then per-screen specs and HTML preview generation)

Between gates the spinner will sit quiet — sub-agents can't stream their progress to this view. If a gate hasn't surfaced after 12 minutes, something is stuck and you should interrupt."

Step 2d — Template-only mode

No background scaffold pipeline is used. The template is already present in <working_dir> and dependencies are expected to be installed before this skill starts (npm install). Continue directly to Step 3.

Step 3 — Plan (planner agent + 4 approval gates)

First, create the empty working directory so the planner has a place to write:

mkdir -p <working_dir>

Hard rule — planner writes are restricted during Step 3. The planner (and any sub-agents it spawns) is permitted to write to only:

• <working_dir>/native-app-plan.md
• <working_dir>/_screens_section.md
• <working_dir>/.tmp/*

All other paths in <working_dir>/ (notably app/, src/, package.json, power.config.json, tamagui.config.ts, tsconfig.json, node_modules/, memory-bank.md) are owned by the foreground setup phases. Do not mutate them during planning.

If the planner needs to record a DONE_WITH_CONCERNS from a sub-agent (data-model architect, screen-planner), add it to an in-memory queue DEFERRED_CONCERNS[] during Step 3. Do not write memory-bank.md yet. Step 6.7 must always flush DEFERRED_CONCERNS[] into memory-bank.md ## Concerns immediately after the file is created.

Resume-from-draft check. Before spawning, check if <working_dir>/native-app-plan.md already exists with content. If yes, a previous planner run (possibly in a degraded context with no Task/gate tools) already drafted sections. Read it. If it has populated ## Data Model / ## Native Capabilities / ## Connectors but the gates were never run (no ## Approvals block, or the file was authored by an agent that returned BLOCKED: tool surface missing), pass resume_from_draft: true and the existing path to the planner so it loads the draft as baseline instead of regenerating from scratch.

Planner preflight (silent). Before the full Task spawn, do a no-op Task probe for mobile-app:native-app-planner (same pattern as Step 11.0). If the probe fails with Agent type … not found, tool unavailable, or the host clearly cannot route nested agents, fall through to inline-gate mode (described below) without prompting. The orchestrator has the full tool surface itself — it can run the gates directly. Do not retry, do not ask the user.

Announce the handoff before the Task call (so the user isn't staring at a blank screen while the planner spins up):

"→ Spawning planner agent. First prompt (data model) appears in ~60–90 seconds while the data-model architect analyzes your requirements. Later gates take longer — see the timing breakdown above."

Then spawn the mobile-app:native-app-planner agent via Task (the plugin name mobile-app: prefix is required — without it Task returns Agent type not found):

Spawn agent: mobile-app:native-app-planner

Prompt:
  Plan a Power Apps mobile app.

  Requirements brief (confirmed with user):
  <requirements_brief — bullet points from Step 2b>

  Design vibe opt-in: <design_vibe_opt_in — always "deferred" unless `--no-design` is in $ARGUMENTS, in which case use "skip". Never invent yes/no/other values.>
  Visual companion: <visual_companion — "yes" or "no">

  Original prompt: <full $ARGUMENTS verbatim>
  Wizard answers: <Step 2 answers>
  Working directory: <absolute path of <working_dir>>
  Plugin root: ${CLAUDE_SKILL_DIR}/../../
  Publisher prefix (detected from env): <DETECTED_PUBLISHER_PREFIX from Step 1.7, e.g. "cr8142a" — use literally as `<prefix>_<entity>` in all logical names. If empty/NOT DETECTED, fall back to `cr` placeholder and surface a `DONE_WITH_CONCERNS` note that Dataverse will normalize at create time.>

  Follow native-app-planner.md. Run all 4 approval gates. On terminal return, emit one of `DONE` / `DONE_WITH_CONCERNS:` / `NEEDS_CONTEXT:` / `BLOCKED:` as the literal first line per AGENTS.md rule #10.

The planner runs gates internally for data model → native capabilities → connectors → screen plan, and writes <working_dir>/native-app-plan.md. Wait for it to return before continuing — do not proceed on a partially-approved plan.

3.0a — Inline-gate fallback (planner unavailable OR returned BLOCKED: tool surface missing)

When the preflight fails OR the planner returns BLOCKED: tool surface missing <…>, the orchestrator runs the four gates inline. Do NOT re-spawn the planner — it cannot succeed in this host. Print once:

"→ Planner agent unavailable in this host — running approval gates inline. (No action needed; this is automatic.)"

Then execute, in order, using your own EnterPlanMode + AskUserQuestion:

1. If a draft native-app-plan.md exists: read it as baseline. Surface each populated section (## Data Model, ## Native Capabilities, ## Connectors) one at a time via EnterPlanMode, take user feedback inline, edit the file in place. Skip generating sections that are already populated and approved.

2. If no draft exists: spawn mobile-app:data-model-architect directly via Task (single architect, not the orchestrator agent) to draft ## Data Model; then build ## Native Capabilities + ## Connectors inline from the brief; then spawn mobile-app:screen-planner with phase: graph and phase: specs per the two-phase Gate 4 split.

   Before each screen-planner spawn, print a one-line ETA so the user knows the agent is live and roughly how long to wait (the agent's own Bash echo progress markers — see agents/screen-planner.md "Progress streaming" — surface every milestone, but the orchestrator's pre-spawn line gives the wall-clock budget):

   • Before phase: graph: > "→ [Gate 4a] Spawning screen-planner phase=graph (~2 min for ${N} screens)…"
   • Before phase: specs: > "→ [Gate 4b] Spawning screen-planner phase=specs (~1 min/screen, ~${N} min for ${N} screens). Progress markers will appear inline."

MUST forward $DETECTED_PUBLISHER_PREFIX from Step 1.7 in the architect prompt — same line as the planner prompt at Step 3 line 1034: "Publisher prefix (detected from env): <DETECTED_PUBLISHER_PREFIX> — use literally as <prefix>_<entity> in all logical names. If empty/NOT DETECTED, fall back to cr placeholder and surface a DONE_WITH_CONCERNS note that Dataverse will normalize at create time." Without this, the architect defaults to cr_ and the whole plan needs a post-hoc sweep when the real prefix is something else (e.g. cr3e9).

Why this works even though the planner just returned BLOCKED for tool surface: the orchestrator (this skill, running in the user's slash-command session) always has the full tool surface — Task, EnterPlanMode, ExitPlanMode, AskUserQuestion, Read, Write, Bash. What's missing is the surface inside nested agent contexts (the native-app-planner agent runs in a sandbox without EnterPlanMode/AskUserQuestion, which is why its Step 0 preflight returned BLOCKED). The leaf agents data-model-architect and screen-planner only need Read/Write/Bash to draft markdown — they don't need EnterPlanMode/AskUserQuestion themselves. Spawn them; the orchestrator owns the gates.

3. Run the gates yourself — use EnterPlanMode four times (data model → native caps + connectors merged → screen graph 4a → screen specs 4b). Same gate prompts as the planner agent would use. Gate 4 is a markdown screen-graph review only — design picking happens unconditionally at Step 6.75 via /design-system (no separate style-picker handoff at Gate 4 even in inline mode).

4. Write the final approved native-app-plan.md with an ## Approvals block at the bottom listing each gate, who approved (user), and a timestamp.

   HARD RULES for the plan structure (mirror the planner agent's template at agents/native-app-planner.md Step 4):

   • Top-level headings are EXACTLY: ## Overview, ## App Requirements, ## Data Model, ## Native Capabilities, ## Design Direction, ## Connectors, ## Screens, ## Approvals. Do NOT invent a ## Brief super-section that nests the data model under it.
   • ## App Requirements is the user's confirmed brief verbatim (the <requirements_brief> from Step 2b), capped at ~80 lines. No expansion, no rewriting, no embedded preview of the data model.
   • Discovery failure notes (e.g. az login on the wrong tenant, 401 from dataverse-request.js, all entities classified Create) go to <working_dir>/memory-bank.md under ## Discovery Notes, NOT into the plan. Keep at most a single one-line breadcrumb in ## Data Model like > Discovery skipped — see memory-bank.md. if relevant.
   • Sample data notes, immutability plug-in notes, file-column setup notes, dispatch-block server rules go under a single ### Notes subsection in ## Data Model. Cap each at 2 sentences; link to post-deployment-tasks.md for longer write-ups instead of inlining.

If the orchestrator's OWN Task tool is unavailable (rare — would mean even leaf agents can't be spawned), fall further to fully-inline mode: orchestrator drafts the data model from a Dataverse OData probe (resolve-environment.js + table list filter on the brief's domain keywords), drafts native caps + connectors heuristically, drafts the screen graph + specs against shared/references/screen-templates.md, and runs the four gates against the user. This is the last-resort path — functional but slower because the orchestrator does work the architects normally parallelize.

Hard rule: never silently skip a gate just because the planner couldn't run. The user MUST approve each section through EnterPlanMode before any mutation step (Step 8 onwards) executes.

3.0 — Sub-agent return-status switch (canonical)

Use the plugin-wide protocol in AGENTS.md rule #10 for every Task return in this skill: planner, parallel screen-builders, and future agent spawns. Parse the literal first line and branch: DONE continues; DONE_WITH_CONCERNS: surfaces + records in memory-bank.md; NEEDS_CONTEXT: re-dispatches with missing context, capped at 2 retries; BLOCKED: stops and records under ## Blocks. Unknown first lines are malformed and must be treated as BLOCKED.

Planner-only early-return signals are handled before the status switch: INDUSTRY_CONFIRM_REQUESTED: routes to Step 3.0a; DESIGN_VIBE_REQUESTED: routes to Step 3a. After the handoff, re-spawn the planner and process its new first line through this switch.

Step 3.0a — Industry confirmation handoff (orchestrator-owned)

When the planner is uncertain about which industry the app belongs to (no keyword match, ambiguous match, or wizard-aesthetic conflict), it returns early with this single line as its message:

INDUSTRY_CONFIRM_REQUESTED: <inferred-industry>|<reason-code>|<top-3-alternatives-comma-sep>

Example: INDUSTRY_CONFIRM_REQUESTED: productivity|no-keywords|field-ops,healthcare,e-commerce

This fires before Gate 1 — it's not a gate, just a confidence check so the wrong industry doesn't silently lock in the design language for the entire app.

Skip this section if <design_vibe_opt_in> is yes or skip — in those cases the user is either driving design explicitly (yes) or has opted out of design entirely (skip), so industry inference doesn't matter.

When you see INDUSTRY_CONFIRM_REQUESTED: and <design_vibe_opt_in> is no:

1. Parse the three pipe-delimited fields. Map reason codes to a short user-facing explanation:

   • no-keywords → "no clear industry signal in your description"
   • ambiguous-match → "your description matches multiple industries"
   • wizard-conflict → "your aesthetic answer doesn't match the inferred industry"

2. Map each industry slug to a one-line description for the picker (use these exactly):

   Slug	Description
   field-ops	Field/Ops — high contrast, large targets, camera-forward (Uber Driver, ServiceTitan)
   finance	Finance — blue palette, conservative type, generous whitespace (banking apps)
   healthcare	Healthcare — warm palette, friendly type, compassionate copy (patient apps)
   education	Education — bright playful, gamification, streak/progress (Duolingo)
   productivity	Prod

…(truncated)