integrate-backend

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

Backend Integration

Analyze the user's business problem and recommend the right backend integration approach — Web API, AI Web API, Server Logic, Cloud Flows, or a combination — then route to the appropriate skill(s) to implement the solution.

Core Principles

• Understand the problem first: Never jump to a technology choice. Analyze the user's intent, data flow, security needs, and performance requirements before recommending.
• Recommend the simplest approach that works: Web API for straightforward Dataverse CRUD, AI Web API for generative summaries or grounded search over existing data, Server Logic when server-side processing is needed, Cloud Flows for async background work. Don't over-engineer.
• Secure actions belong on the server: When a write depends on a business rule that must be tamper-proof (state transitions, approval workflows, computed values), the server logic must validate AND execute the write — not just validate and leave the write to a client-side Web API call. See the Secure Action Principle in the decision framework.
• AI Web API sits on top of Web API: The Data Summarization and Case-preset endpoints read through the same _api layer as regular Web API, so they inherit the same table permissions, column permissions, and Webapi/<table>/* site settings. When a plan has both a Web API item and an AI Web API item for the same table, the AI item depends on (and goes in a later phase than) the Web API item. Search Summary has no per-table prereqs and can stand alone.
• Combinations are normal: Many real scenarios need more than one approach. Recommend combinations when justified, but explain why each piece is needed.
• Route, don't implement: This skill recommends and invokes the right skill(s). It does not create backend files itself.

Initial request: $ARGUMENTS

Workflow

1. Verify Site Exists — Locate the Power Pages project and check prerequisites
2. Understand the Business Problem — Analyze what the user needs and why
3. Recommend Integration Approach — Present the recommendation with reasoning
4. Route to Skill(s) — Invoke the appropriate backend skill(s) to implement

Phase 1: Verify Site Exists

Goal: Locate the Power Pages project root and confirm prerequisites

Actions:

1. Create todo list with all 4 phases (see Progress Tracking table)

1.1 Locate Project

Look for powerpages.config.json in the current directory or immediate subdirectories.

If not found: Tell the user to create a site first with /create-site.

1.2 Explore Current State

Use the Explore agent to quickly scan the site for existing backend integrations:

"Analyze this Power Pages code site for existing backend integrations:

1. Check .powerpages-site/server-logic/ — list any existing server logic endpoints
2. Check .powerpages-site/cloud-flow-consumer/ — list any registered cloud flows
3. Search frontend code (src/**/*.{ts,tsx,js,jsx,vue,astro}) for calls to /_api/ (Web API) and /_api/serverlogics/ (Server Logic) and /_api/cloudflow/ (Cloud Flows)
4. Check for existing service layers or API utilities in src/services/, src/shared/, or similar
5. List available web roles from .powerpages-site/web-roles/*.webrole.yml Report what backend integrations already exist so we can build on them."

1.3 Discover Dataverse Custom Actions

Check whether the user's Dataverse environment has existing custom actions that could be leveraged in the integration:

node "${PLUGIN_ROOT}/scripts/list-custom-actions.js" "<ENV_URL>"

The script returns Custom APIs (modern) and Custom Process Actions (legacy) with their names, descriptions, binding types, and parameters. If custom actions are found, note them — they will be factored into the recommendation in Phase 3.

Output: Project root confirmed, existing backend integrations identified, Dataverse custom actions discovered

Phase 2: Understand the Business Problem

Goal: Analyze the user's request to understand the underlying business problem, not just the technical ask

Actions:

2.1 Analyze the Request

From the user's request and the existing site state, determine:

• What is the user trying to accomplish? (business outcome, not technology)
• What data is involved? (Dataverse tables, external systems, user input)
• Who triggers the operation? (user action, form submit, page load, scheduled)
• Does the user need an immediate response? (real-time UI update vs. background processing)
• Are external services involved? (payment gateways, email, Graph, SharePoint, third-party APIs)
• Are credentials or secrets involved? (API keys, client secrets, tokens)
• Must logic be hidden from the browser? (pricing rules, validation algorithms, business rules)
• Is this a simple data operation or complex business logic? (CRUD vs. multi-step processing)
• Does any write depend on a business rule that must be tamper-proof? (state transitions, approval conditions, computed values) — if yes, the server logic must validate AND execute the write, not just validate
• Does the UI want an AI-generated summary, grounded AI search, or related-record discovery? (e.g., "summarize this case", "summarize open orders", "suggest KB articles on the case page", "AI-powered search") — if yes, AI Web API is the right fit. Watch for the phrasing signals: summarize, summary of, Copilot, related / similar / suggested , AI search, semantic search.
• Can existing Dataverse custom actions handle part of the requirement? If custom actions were discovered in Phase 1.3, check whether any align with the user's needs — server logic can wrap existing custom actions via InvokeCustomApi instead of building equivalent logic from scratch

2.2 Clarify if Ambiguous

If the request could map to multiple approaches and the right choice isn't clear, use AskUserQuestion to clarify:

Output: Clear understanding of the business problem and technical requirements

Phase 3: Recommend Integration Approach

Goal: Present a recommendation with clear reasoning

Actions:

3.1 Apply the Decision Framework

Reference: ${PLUGIN_ROOT}/skills/integrate-backend/references/decision-framework.md

Use the decision matrix, intent mapping, and Secure Action Principle from the reference to determine the right approach. Consider:

1. Can Web API alone handle this? If it's straightforward Dataverse CRUD with no external calls, no secrets, no server-side logic, and no business rules governing the write — recommend Web API. It's the simplest option.

2. Does it need AI Web API? If any of these apply, AI Web API is the right fit:

   • The UI wants an AI-generated summary of a record on its detail page (e.g., "summarize this case", "Copilot summary")
   • The UI wants an AI summary of a list or collection (e.g., "summarize open orders", "highlight trends in this week's cases")
   • A detail page wants related-record discovery (e.g., "suggest KB articles for this case", "similar products", "similar cases") — Search Summary's grounded retrieval is a better fit than a hand-rolled OData keyword match
   • The site wants AI-grounded search with citations, replacing or augmenting keyword-only search

   AI Web API is read-only — the Secure Action Principle does not apply. If an AI item covers a Dataverse table that is also covered by a Web API item, put the AI item in a later phase (it depends on the Web API Layer 1/2 prereqs being in place). Search Summary items have no per-table prereqs and can stand alone.

3. Does it need Server Logic? If any of these apply, Server Logic is needed:

   • External API calls (HttpClient)
   • Credentials/secrets must stay on the server
   • Business logic must be hidden from the browser
   • Multiple Dataverse queries should be batched into one endpoint
   • Server-side validation that can't be bypassed
   • Wrapping a Dataverse Custom API/Action for portal consumption — if custom actions were found in Phase 1.3, check whether any match the requirement before recommending building from scratch
   • The write depends on a business rule that must be tamper-proof (state transitions, approval conditions, computed values) — server logic must validate AND execute the write

4. Does it need Cloud Flows? If any of these apply, Cloud Flows are the right fit:

   • The operation is async — the user doesn't need an immediate result
   • Background processing: sending emails, notifications, processing orders
   • Multi-step workflows across systems with Power Automate connectors
   • Long-running processes that exceed the 120-second server logic timeout
   • Non-developers should be able to modify the workflow

5. Does it need a combination? Common combinations:

   • Web API + AI Web API: UI displays raw Dataverse records and an AI summary of the same data (most common AI pattern — dashboards, case/order detail pages)
   • Web API + Cloud Flow: UI reads/writes non-sensitive Dataverse fields, some actions trigger background flows
   • Server Logic + Cloud Flow: Real-time endpoint validates and executes the action, async flow does follow-up (e.g., server logic transitions status, Cloud Flow sends notification)
   • Web API + Server Logic: Web API for safe direct reads/writes, server logic for operations that need business rule enforcement (server logic validates AND writes for those operations)
   • Web API + AI Web API + Cloud Flow: support portal with browsable cases, Copilot summary + KB discovery, and async notifications

3.1.1 Security Review — Apply the Secure Action Principle

Before finalizing the plan, review every item assigned to Web API and ask: "If a user skipped any preceding server logic validation and called this Web API endpoint directly, could they violate a business rule?"

If the answer is yes, that write does not belong in a Web API item. Move the write into the server logic item that validates it. The server logic should validate AND execute the write using Server.Connector.Dataverse.

Common patterns that must use validate-and-execute server logic (not Web API):

Correct plan structure for state transitions:

Phase 1: Server Logic — "transition-order" endpoint
  - POST: accepts { entityId, targetStatus }
  - Reads current record, validates transition is allowed, writes new status
  - Returns { success, previousStatus, newStatus }

Phase 2: Web API — Order table CRUD
  - Read: list/filter orders (safe for Web API)
  - Create: new orders in Draft status (safe — initial state, no rule to enforce)
  - Update: description, notes, dates (safe — no business rules on these fields)
  - NOTE: Status changes are NOT here — they go through the server logic endpoint

Incorrect plan structure (anti-pattern):

3.2 Render the HTML Plan

Build the plan data and render an HTML plan before asking for approval. The plan visualizes:

• Key Concepts — Educational overview of Web API, Server Logic, and Cloud Flows (hardcoded in template)
• Overview — Stats per approach, approach chips, design rationale
• Data Flow — Visual flow diagrams showing how data moves for each user action, with steps color-coded by approach
• Implementation Order — Phase-grouped items with dependencies, complexity badges, and implementation commands
• Integration Items — Each item with its approach, reasoning, and implementation details

Prepare a JSON object with these keys:

ITEMS_DATA format:

Phase assignment rules — assign a phase number to each item based on dependencies:

1. Items with no dependencies go in the earliest phase appropriate for their approach
2. Items that depend on other items go in a later phase than their dependency
3. Items in the same phase have no dependencies on each other and can be built in parallel
4. Recommended default ordering: Server Logic foundations first (validate-and-execute endpoints for state transitions, batch queries), then Web API CRUD for non-sensitive fields (reads, creates with safe defaults, updates to fields with no business rules), then advanced Server Logic (multi-table transactions), then AI Web API for summaries/grounded search over the tables set up by Web API, then Cloud Flows (async follow-ups)
5. Security constraint: A Web API item must never write a field whose value is governed by a business rule enforced in a server logic item. If a field needs validation, the server logic item should write it directly — the Web API item should exclude that field from its scope
6. AI layering constraint: An AI Web API item for a Dataverse table that is also covered by a Web API item must be in a later phase than the Web API item — AI Web API's Data Summarization endpoint reuses the Web API's Layer 1/2 prereqs (table permissions, Webapi/<table>/* site settings). Search Summary items have no per-table prereqs and can go in any phase.

DATA_FLOWS_DATA format (each step's approach is one of webapi, aiwebapi, serverlogic, cloudflow):

Important: In data flow diagrams, when a server logic step validates a business rule, the next step should NOT be a Web API write for the same field. The server logic step should validate AND execute. For example:

Write the plan to <PROJECT_ROOT>/docs/backend-plan.html (create docs/ if needed). Use the render script:

The render script refuses to overwrite existing files. If the default path exists, choose a new descriptive filename (e.g., backend-plan-payments.html).

After rendering, open the HTML plan in the user's default browser:

3.3 Present Plan Summary

Do not restate the full plan in the CLI. The HTML file is the single detailed plan artifact.

In the CLI, give only a brief summary:

• Total items and how many per approach
• Whether the plan uses one approach or a combination
• The actual output path
• A note that the browser-opened HTML contains the full details including data flow diagrams

3.4 Confirm with User

🚦 Gate (plan · integrate-backend:3.4.plan-approval): Approve the integration plan before invoking the appropriate child skill (integrate-webapi / add-server-logic / add-cloud-flow). The plan HTML stays on disk regardless of choice — Cancel just stops the dispatch.

Trigger: Phase 3.3 has rendered the HTML plan and surfaced a brief CLI summary. Why we ask: Wrong child skill gets dispatched — add-server-logic for a Web API task wastes minutes; add-cloud-flow for a Web API task creates orphaned flow YAML. Cancel leaves: Nothing — no child skill invoked, HTML plan stays at its saved path.

Use AskUserQuestion:

If "Change approach": Ask what they'd prefer and why, update the plan, and present again.

If "Cancel": Stop the workflow.

Output: User-approved integration approach

Phase 4: Route to Skill(s)

Goal: Invoke the appropriate skill(s) to implement the approved approach, respecting phase ordering and building in parallel within each phase

Actions:

4.1 Build the Phase Execution Plan

Group the approved items by their phase number from the plan. Each phase is a batch of independent items — items within a phase have no dependencies on each other and can be built in parallel.

4.2 Execute Phase by Phase

Process phases in order (Phase 1, then Phase 2, etc.). Complete all items in a phase before moving to the next — later phases depend on earlier phases.

Within each phase, maximize parallelism:

• Single approach in the phase: Invoke the skill once with all items for that phase. Tell the skill: "These N items are independent — implement them in parallel where possible."
• Multiple approaches in the same phase: Invoke each skill for its items. Since items in the same phase have no cross-dependencies, the order of skill invocation within a phase does not matter. When invoking the second skill, pass context from the first so it can follow the same frontend patterns (e.g., naming conventions, file organization).

Example — a plan with 4 phases:

When invoking each skill, include:

1. Which items to implement — list the specific item names from the plan for this phase
2. Parallelism guidance — "These items are in the same phase and have no dependencies on each other. Implement them in parallel where possible."
3. Context from previous phases — what was created so far (files, patterns, services) so the skill can build on it

4.3 Summary

After all phases complete, present a brief summary of everything that was created:

Remind the user to deploy with /deploy-site if they haven't already.

4.4 Record Skill Usage

Reference: ${PLUGIN_ROOT}/references/skill-tracking-reference.md

Follow the skill tracking instructions in the reference to record this skill's usage. Use --skillName "IntegrateBackend".

Output: All recommended backend integrations implemented

Important Notes

When NOT to Use This Skill

If the user's request clearly and unambiguously maps to a single approach, skip this skill and go directly to the implementation skill:

• "Create a server logic endpoint for..." → /add-server-logic
• "Integrate Web API for the contacts table" → /integrate-webapi
• "Add a cloud flow for sending emails" → /add-cloud-flow
• "Add an AI summary to the case detail page" / "summarize open cases" / "suggest KB articles on this page" → /add-ai-webapi

This skill is for ambiguous requests where the user describes a business problem and needs help choosing the right approach.

Examples: What Routing Looks Like

Example 1: Simple Dataverse CRUD → Web API

Example 2: External API with credentials → Server Logic

Example 3: Background email → Cloud Flow

Example 4: Server-side validation → Server Logic (validate-and-execute)

Example 5: Dashboard performance → Server Logic

Example 6: CRUD + background processing → Web API + Cloud Flow

Example 7: Validate + process + notify → Server Logic + Cloud Flow

Example 8: State transitions + CRUD → Server Logic (validate-and-execute) + Web API

Example 9: AI summary on detail page → AI Web API (Data Summarization, support-case recipe)

Example 10: Related KB articles on detail page → AI Web API (search summary)

Example 11: List summary + browsable records → Web API + AI Web API

Progress Tracking

Before starting Phase 1, create a task list with all phases using TaskCreate:

Mark each task in_progress when starting and completed when done via TaskUpdate.

Begin with Phase 1: Verify Site Exists