add-ai-webapi

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

Add AI Web API

Note

AI summarization APIs are a preview feature. Preview features aren't meant for production use and may have restricted functionality. These features are available before an official release so that customers can get early access and provide feedback.

Surface this note to the user verbatim during Phase 1 and again in the Phase 8 summary — copy the exact **Note** block above (including its wording about "available before an official release so that customers can get early access and provide feedback"). Do not paraphrase it into your own "Preview-feature note: ..." sentence; the wording matches the Microsoft Learn preview disclaimer and rephrasing it loses that fidelity.

Integrate Power Pages generative-AI summarization APIs into a SPA site. This skill focuses on the AI layer (Layer 3): the summarization service code and the Summarization/* site settings. The underlying Web API prerequisites — Webapi/<table>/enabled, Webapi/<table>/fields, table permissions, and web roles — are delegated to /integrate-webapi and /create-webroles so there is a single source of truth for every layer.

The two APIs covered

Example: Microsoft-shipped Copilot summary on a support-case page. Data Summarization can be called with any combination of entity set, columns, and prompt — but Microsoft documents and ships one specific configuration for the standard incident table: POST /_api/summarization/data/v1.0/incidents(<caseId>)?$select=description,title&$expand=incident_adx_portalcomments($select=description) with body { "InstructionIdentifier": "Summarization/prompt/case_summary" }. This is sometimes called the "Case-page Copilot preset" in Microsoft Learn. Treat it as one possible Data Summarization recipe — useful when the user explicitly wants to mirror the Microsoft sample — not as an automatic recommendation. A custom case-like table (cr363_servicerequest, adx_case), or the standard incident table summarised on different facets (priority, owner, SLA timer), is just a regular Data Summarization call with maker-defined values.

Reference: ${PLUGIN_ROOT}/skills/add-ai-webapi/references/ai-api-reference.md — canonical API shapes, required headers, site-setting names, error codes, and the documented support-case example. Read this at the start of the workflow; fetch the Microsoft Learn source pages with mcp__plugin_power-pages_microsoft-learn__microsoft_docs_fetch if the user asks for the latest.

Admin governance hierarchy: both APIs are gated by a three-level admin hierarchy — tenant PowerShell setting (enableGenerativeAIFeaturesForSiteUsers), Copilot Hub environment/site governance, and the site-level maker toggle (for Search Summary: Set up workspace → Copilot → Site search (preview) → Enable Site search with generative AI (preview)). Each level overrides the one below it, so "the maker toggle is on but the API still says disabled" is a real scenario — admin-level governance wins.

The two endpoints surface disablement differently:

• Search Summary → HTTP 200 with an embedded envelope { Code: 400, Message: "Gen AI Search is disabled." }. The generated fetchSearchSummary detects this and throws SearchSummaryApiError; the UI renders a remediation card.
• Data Summarization → HTTP 400 with error.code = 90041001 (admin-level disabled) or 90041003 (per-site Summarization/Data/Enable=false).

Full troubleshooting checklist (tenant → environment → site, plus runtime version, Bing dependency, and cross-region data movement) lives in references/ai-api-reference.md §1 "Troubleshooting: AI feature appears disabled (admin hierarchy)" — point users there when either disablement shape surfaces. Mention this governance hierarchy explicitly to the user before Phase 7, and again in the Phase 8 summary.

Built-in search control vs. custom code path: if the site uses the Microsoft-shipped Power Pages search control and only wants AI-summarised search results on that page, they don't need this skill — just the Copilot workspace toggle and the Search/Summary/Title content snippet. This skill is for sites that build their own search UI or need to call /_api/search/v1.0/summary from custom code. Confirm which path the user is on in Phase 1.

Core principles

• Layer 3 only, delegate the rest. Web API site settings, table permissions, and web roles all belong to /integrate-webapi and /create-webroles. This skill creates the summarization service code and the Summarization/* site settings — nothing else.
• Sequential agent spawning. Per plugins/power-pages/AGENTS.md, spawn the ai-webapi-integration agent sequentially per target (never in parallel). The first call establishes the shared summarization service file and CSRF helper; subsequent calls extend it. ai-webapi-settings-architect runs alone, after all code integrations land.
• Raw fetch + CSRF. Every summarization request attaches __RequestVerificationToken (from /_layout/tokenhtml) and X-Requested-With: XMLHttpRequest. Never route through an OData wrapper.
• Skip /integrate-webapi when it's not needed. If every confirmed target is Search Summary (which has no per-table Web API prerequisites), or every Layer 1/2 prerequisite already exists on disk, the skill goes straight from Phase 3 to Phase 5.
• Use TaskCreate/TaskUpdate — create the todo list upfront with all phases before starting.

Prerequisites:

• An existing Power Pages SPA site created via /create-site
• A Dataverse data model (tables + columns) set up via /setup-datamodel or manually — for any Data Summarization target
• The site must have been deployed at least once (.powerpages-site folder must exist) for the settings phase

Initial request: $ARGUMENTS

---

Workflow

(Phase headings below the workflow keep the technical "Layer 1+2 / Layer 3" names because they describe the runtime layering and are what maintainers grep for. The titles here mirror the user-facing task list.)

1. Check site is ready — locate project, detect framework, check data model, deployment status, and web-role presence.
2. Find where AI summaries fit — scan code for search / data summarization candidates.
3. Confirm what to add — review the manifest and pick which APIs / targets to integrate.
4. Set up data access for AI — invoke /create-webroles if needed, then invoke /integrate-webapi in AI-only read mode for data/case targets. Skip entirely for search-only or when prerequisites already exist.
5. Add AI summary code — invoke the ai-webapi-integration agent sequentially per target.
6. Register AI prompts — invoke the ai-webapi-settings-architect agent.
7. Verify everything — header-contract grep, $select grep, npm run build, validator script.
8. Review and deploy — record skill usage, summarise, offer /deploy-site.

---

Iteration mode (after first run)

This skill is a one-shot setup skill — Phases 1–8 run end-to-end the first time the user asks to integrate a summarization API. Once an AI surface is in place (service file, framework wrapper, UI call site, and Summarization/* settings all exist), follow-up requests to tweak the rendered UI (colours, spacing, copy, moving a button, a different empty-state message, wiring a second recommendation into the hook, etc.) are not a reason to re-enter this skill mentally and run every phase again. Doing so triggers a full pac pages upload-code-site and a chain of git commits for each tweak, which is exactly the noisy cadence the Phase 5.5 / 6.4 prompts above are there to avoid.

When the user asks for follow-up UI changes to an already-integrated AI surface:

• Edit the file(s) and run npm run build locally to verify the tweak compiles. That is the whole validation loop for a UI change.
• Do NOT automatically run pac pages upload-code-site. Uploading should happen once, at the end of the session, when the user has finished tweaking.
• Do NOT automatically git commit. Let the user batch related tweaks into a single commit.
• Batch the deployment and commit into a single end-of-session prompt once the user signals they're done (or when you've completed the last requested change).

🚦 Gate (consent · add-ai-webapi:iter.deploy-commit): End-of-iteration batched deploy + commit prompt — avoids a noisy per-tweak upload/commit cadence.

Trigger: User signals they're done with UI tweaks for the session. Why we ask: Auto-deploying or committing after every small edit produces one git commit + one pac pages upload-code-site per tweak; batching keeps history readable and avoids redundant deploys. Cancel leaves: Nothing — source files already edited; no deploy or commit fired.

Use AskUserQuestion:

Re-enter the full skill flow only when the user is adding a new AI surface (a new page, a new table, a second API). If you're unsure whether a request is a tweak or a new surface, ask.

---

Phase 1: Verify Site Exists

Goal: Locate the Power Pages project root and confirm prerequisites.

1.0 Detect iteration mode (before anything else)

Re-entry detection comes first because the rest of the skill assumes a first-time setup. Capture two signals about the project state:

1. Service signal: a summarization service exists — src/services/aiSummaryService.*, or any source file under src/ that contains /_api/search/v1.0/summary or /_api/summarization/data/v1.0/. When the signal is present, also note which API surface(s) the service code references — search-only, data-only, or both. This sub-classification matters below.
2. Settings signal: at least one Layer 3 site setting exists in .powerpages-site/site-settings/Summarization-*.sitesetting.yml.

Then route on the combination:

• Both signals present → re-entry. A previous /add-ai-webapi run completed end-to-end. Show the iteration-mode prompt below.
• Service signal present and the existing service code is search-only (it references /_api/search/v1.0/summary but not /_api/summarization/data/v1.0/) → re-entry. Search-only sites legitimately have no Summarization/* settings (Search Summary uses the workspace toggle, not per-call settings), so the absence of the settings signal is the steady state, not a failed run. Show the iteration-mode prompt below.
• Service signal present and the existing service code includes Data Summarization but the settings signal is absent → in-flight first run (a previous attempt failed before Phase 6 landed). Continue with the full flow without prompting.
• Neither signal present → first-time run. Continue with Phase 1.1.

When you reach the iteration-mode branch, ask:

• Tweak the existing surface: stop running this skill. Switch into the workflow described in the Iteration mode section above — Edit + npm run build, no auto upload-code-site, no auto commit, batched end-of-session prompt for deploy + commit.
• Add a new surface: continue with Phase 1.1 (full flow). The downstream phases will detect existing infrastructure (CSRF helper, summarization service, settings) and extend rather than duplicate.
• Not sure — show me what's already wired: list the existing service file(s), wired UI components, and Summarization/* settings, then re-ask the same question.

1.1 Create todo list

Create all 8 phase tasks upfront via TaskCreate — see Progress Tracking.

1.2 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.

1.3 Detect framework

Read package.json and detect React / Vue / Angular / Astro. See ${PLUGIN_ROOT}/references/framework-conventions.md.

1.4 Check for data model

Look for .datamodel-manifest.json. If found, read it — tables listed here are candidates for the Data Summarization API. The standard incident table is a candidate like any other; do not treat it specially.

1.5 Check deployment status — hard prerequisite

Look for .powerpages-site. Phase 4 (/integrate-webapi) and Phase 6 (ai-webapi-settings-architect) both require this folder to exist. Deferring the deploy until later is not a viable workaround: once Phase 5 has written the AI-calling service code, deploying a site whose Layer 1/2/3 settings aren't yet on disk publishes runtime-broken code (every summarization call 403/500s until a second deploy lands). The cleanest sequence is to deploy the clean scaffold now, before any AI code exists.

If .powerpages-site does NOT exist:

On Yes: invoke the Skill tool for power-pages:deploy-site and wait for completion. Then re-check .powerpages-site exists before proceeding. If it is still absent after the sub-skill returns (deploy failed mid-flow, the user cancelled it, or an upload completed but the local folder wasn't created), stop here with a clear message — surface the deploy-site outcome verbatim so the user can debug it, and tell them to re-run /deploy-site followed by /add-ai-webapi. Do NOT silently fall through to Phase 2; the downstream sub-skills require this folder.

On Stop: end the skill with a clear next-step message ("Run /deploy-site, then re-invoke /add-ai-webapi to continue"). Do NOT continue to Phase 2 — the downstream sub-skills can't run.

1.6 Check web roles

Look for .powerpages-site/web-roles/*.yml. Record whether any roles exist — the Phase 4 delegation needs at least one role before /integrate-webapi can create table permissions.

Output: confirmed project root, framework, data-model availability, deployment status, web-role inventory.

---

Phase 2: Explore AI integration points

Goal: Find every candidate for each of the two APIs — scoped to AI only.

The full Explore-agent prompt body, manifest shape, and delegation-decision rules live in ${PLUGIN_ROOT}/skills/add-ai-webapi/references/explore-prompt.md. Read that file first, then invoke the Explore agent (via Task with subagent_type: "Explore", thoroughness medium) and pass the prompt body verbatim.

What the Explore agent reports back (summary — see the reference for the exact prompt):

• Reserved-slot markers (POWERPAGES:AI-SLOT kind=<pick>) — authoritative placement hints planted by /create-site; orphan markers are flagged for Phase 3 to resolve.
• Search Summary candidates including related-record-discovery targets on detail pages ("suggested KB articles", "similar cases").
• Data Summarization candidates classified single-record / list / intent-mismatch, each with the existing fetch's OData query and a scope classification (matches-existing-fetch / scope-extends-beyond-existing-fetch / needs-definition / intent-mismatch).
• Existing infrastructure — CSRF helper, powerPagesApi.ts, prior aiSummaryService.*.
• Layer 1/2 status per Data Summarization target plus every $expand target — ready / missing / n/a (search).
• Fields-list breadth advisory — for Layer 1/2 ready rows whose existing Webapi/<table>/fields is broader than the AI surface needs (primary key included, relationship Navigation Properties unused by @odata.bind, unused columns), flagged fields-broader-than-ai-mode so Phase 3 can surface it.
• Layer 3 status per Data Summarization target — Summarization/Data/Enable and the specific Summarization/prompt/<id> the code will send.

Compile the integration manifest from the agent's output (one row per candidate, columns: #, API, Target file, Target kind, Entity Set, $select / $expand, Source, Layer 1/2 status, Layer 3 status) — the reference file contains a worked example.

Delegation decisions (compute directly from the status columns):

• Run /integrate-webapi? → True if any row's Layer 1/2 status is missing. Send only the missing-status rows in the tables= sentinel; don't re-audit settled tables.
• Run ai-webapi-settings-architect? → True if any row's Layer 3 status is missing.

Output: integration manifest + delegation decisions + existing-infra report + fields-broader-than-ai-mode advisory list (if any).

---

Phase 3: Review AI plan

Goal: Present the manifest and confirm which APIs / targets to integrate.

Show the user:

1. The list of APIs and targets found.
2. For each: which file references it and what the service will do.
3. The two delegation decisions from Phase 2 ("Will invoke /integrate-webapi for [tables]", "Will invoke ai-webapi-settings-architect for Layer 3").
4. Existing-infrastructure notes (CSRF helper reuse, powerPagesApi.ts, previous aiSummaryService.*).

Approval cadence (set expectations up-front)

Before asking the integration question, briefly tell the user how many more decision points are coming so the "one-shot" run isn't surprising. Count them from the Key decision points list, subtracting the ones that don't apply for this run:

• Search-only run: drop the /integrate-webapi delegation (Phase 4) and the settings architect (Phase 6) — 4–5 more pauses after this one.
• Layer 1/2 already ready: drop the Phase 4.3 architect approvals.
• No list / scope-extends / intent-mismatch rows: drop the per-target Phase 3 follow-ups.
• First-time run with all branches firing: 7–10 more pauses (web-role choice, per-target list-trigger and scope-confirmation, two integrate-webapi architect plans, the AI settings architect plan, two commit prompts, final deploy).

Phrase it as a heads-up, not a warning — e.g., "I'll pause for your input ~6 more times after this one (web-role choice, two architect plans, two commit prompts, final deploy). Could be a couple more if any list target needs a trigger or scope decision."

The integration question

Use AskUserQuestion and build the option list dynamically from the Phase 2 manifest — do not hardcode "Search summary and Data summarization" when only one category has candidates. Construct the question text from what was found:

• If both categories have candidates: "I found candidates for Search Summary and Data Summarization. Which should I integrate?"
• If only one category with one target: skip "All of them" — just confirm the single target ("Wire Search Summary into <page>?").

The default option list, with rows present only when the corresponding category has candidates:

If the user chooses "Let me select specific ones", follow up with a multi-select question listing each row of the integration manifest. When a detail-page candidate was flagged in Phase 2 as a related-record-discovery target, include it as a dedicated option (in addition to any Data Summarization option for the same page) so the user can consciously pick the AI-grounded path rather than a hand-rolled OData match — e.g.:

• Search Summary on CaseDetail.tsx (finds related KB articles via generative AI)
• Search Summary on ProductDetail.tsx (finds related products via generative AI)

Label the option with the page name and the outcome it produces, not just "Search Summary", so the user sees exactly where the AI surface will appear.

Per-target follow-up questions

After the user confirms targets, ask per-target follow-up questions only when needed:

• List-trigger — every list target gets a "load on open" vs "manual button" question.
• Scope confirmation — only when Phase 2 flagged the row scope-extends-beyond-existing-fetch, needs-definition, or intent-mismatch. Resolve intent-mismatch first because the chosen target kind decides which scope question (LIST or SINGLE-RECORD) to ask next.

Question text, option lists, and the scope-classification → question mapping live in ${PLUGIN_ROOT}/skills/add-ai-webapi/references/scope-classification.md. Read it when any per-target follow-up is required.

The chosen values flow into the Phase 5 agent-invocation prompt as the Scope for the summary call block ($filter, $orderby, Scope source, Target's existing list fetch). The existing UI fetch on the target component stays in place unchanged — the summary URL is an addition, never a replacement.

If the user picks a scope that adds new $expand targets (e.g., "include its line items"), re-evaluate the Phase 2 Layer 1/2 delegation decision — the new expansion is a new prerequisite (Web API enabled + parent-scope permission on the child table).

Handling "None — cancel"

When the user picks None — cancel:

1. Mark the remaining tasks as completed with a (skipped — cancelled by user) suffix in the activeForm so the task list reads cleanly rather than leaving them stuck pending.
2. Jump straight to Phase 8.1 (record skill usage with --skillName "AddAiWebapi" and an outcome of cancelled) and Phase 8.2 (present a one-line summary: "No changes made — you cancelled at the plan-review step").
3. Skip Phases 4, 5, 6, 7, and 8.3 entirely. Do not invoke /integrate-webapi, ai-webapi-integration, ai-webapi-settings-architect, or /deploy-site. Do not commit.

Output: user-confirmed integration manifest, or a clean cancellation.

---

Phase 4: Delegate Layer 1 + Layer 2

Goal: Ensure every Web API prerequisite for the AI target tables exists, by delegating to /create-webroles and /integrate-webapi (AI-only read mode) instead of writing Layer 1/2 files directly.

4.1 Skip-check

Skip this entire phase when any of the following is true:

• Every confirmed target is Search Summary (search has no per-table Web API prerequisites).
• The Phase 2 delegation decision said "Run /integrate-webapi? No" — all Layer 1/2 prerequisites are already on disk from prior runs.
• .powerpages-site does not exist (the sub-skills both require it).

In the skip case, proceed to Phase 5 and note this in the final summary.

4.2 Create missing web roles (if needed)

From Phase 1.6: if no web roles exist in .powerpages-site/web-roles/, or the roles that exist don't match the site's auth model, ask the user:

On Yes: invoke the Skill tool for power-pages:create-webroles with a prompt that includes the caller-suppress sentinel so the sub-skill does not issue its own deploy prompts (the orchestrator owns the single end-of-run deploy decision):

[CALLED-BY-PARENT-SKILL] caller=add-ai-webapi

Create web roles for this Power Pages SPA site. The parent skill /add-ai-webapi will later attach AI-only read-mode table permissions to the role(s) you create. Do not issue deploy-now prompts — the orchestrator batches the deploy at the end.

Wait for it to complete. Then re-check .powerpages-site/web-roles/ before proceeding to 4.3.

On Skip: this puts the run on a known-broken path — the AI endpoints will return 403 at runtime until the user manually creates a web role + table permissions. Don't fall through silently.

🚦 Gate (consent · add-ai-webapi:4.2.skip-webrole): Explicit acknowledgement before continuing without a web role — skipping leaves Layer 1/2 broken at runtime.

Trigger: User chose "Skip" on the web-role creation offer. Why we ask: Proceeding silently means the AI API endpoints return 403 at runtime; surfacing the trade-off lets the user make an informed stop-vs-continue choice. Cancel leaves: Nothing — no code written yet.

Confirm the trade-off with a second AskUserQuestion:

On Stop here: end the skill cleanly. Tell the user to run /create-webroles (and optionally /integrate-webapi) first, then re-invoke /add-ai-webapi.

On Continue: skip the rest of Phase 4 (the /integrate-webapi delegation needs a web role to attach permissions to, so running 4.3 would fail) and jump to Phase 5. In the Phase 8 summary, flag the Layer 1/2 gap loudly — list the exact files the user still needs to create (Webapi/<table>/enabled, Webapi/<table>/fields, table permissions per target) so the path to a working runtime is obvious from the final message.

4.3 Invoke /integrate-webapi in AI-only read mode

Build the sentinel arguments from the Phase 2 manifest:

• primary=<primary table logical name> (the table whose record / collection is being summarised — e.g., incident, cr4fc_product, cr363_workorder)
• tables=<primary plus every $expand target, comma-separated>
• expand-targets=<every $expand target, comma-separated; empty for pure data-summary targets with no $expand>
• caller=add-ai-webapi

Invoke the Skill tool for power-pages:integrate-webapi with a single prompt:

[AI-READ-ONLY] mode=ai-read-only primary=<primary> tables=<list> expand-targets=<list> caller=add-ai-webapi

Configure Layer 1/2 (Web API site settings + table permissions) for the following Power Pages AI summarization targets: . This is a read-only integration — the /_api/summarization/data/v1.0/ endpoint never mutates Dataverse. Return when all Web API site settings, table permissions, and the shared powerPagesApi.ts client are written to disk.

Per the AI-only read-mode contract (your SKILL.md Phase 1.6):

• Do not commit. Skip Phase 4.4 and Phase 6.5 — print the file list you would have committed so this orchestrator can stage it later. The parent skill batches commits.
• Do not deploy. Skip Phase 6.1 deploy ask, Phase 7.3 deploy ask, and Phase 7.4 post-deploy notes. Return the Phase 7.2 summary and stop. The parent skill owns the single end-of-orchestration deploy.

/integrate-webapi detects the [AI-READ-ONLY] sentinel (its Phase 1.6) and runs its full flow with hardened prompts: read-only table permissions, minimal fields list (no PK, only _<col>_value for lookups), and a read-only service layer. It still presents plan-mode approval prompts to the user for each architect — this skill does not suppress those.

Wait for /integrate-webapi to complete. Re-check the file system:

• src/shared/powerPagesApi.ts exists
• For every target table: Webapi/<table>/enabled exists, Webapi/<table>/fields exists
• For every target table: at least one table permission with read: true exists; Parent-scope permission present for every $expand target

If any prerequisite is still missing, surface this to the user before moving on — something in the delegated flow didn't land (for example, the user declined the architect's plan). Do NOT silently fall back to writing Layer 1/2 files here.

Output: Layer 1/2 prerequisites are on disk; shared powerPagesApi.ts + read-only service exists; web roles exist.

---

Phase 5: Implement Layer 3 code

Goal: Create the AI summarization service and wire it into each target's UI.

5.1 Invoke the ai-webapi-integration agent — first target (sequential)

For the first target in the confirmed manifest, invoke the agent at ${PLUGIN_ROOT}/agents/ai-webapi-integration.md via Task. The full prompt template — every field the agent expects, with notes on which orchestrator phase resolved each value — lives in ${PLUGIN_ROOT}/skills/add-ai-webapi/references/agent-invocation-prompt.md. Read that file, copy the template, replace every <…> placeholder with the concrete value for the current target, and pass it via Task. The agent does not interpret placeholders; sending the literal text <search | data> will confuse it.

The first call is sequential because it establishes the shared summarization service file (src/services/aiSummaryService.*) and the CSRF helper that subsequent targets reuse. The agent returns a structured file-modification list (see its "Return value" section); record it for Phase 5.5's per-file commit.

5.2 Verify service + CSRF helper exist

Before spawning more agents, verify:

• The summarization service file exists (default src/services/aiSummaryService.ts).
• getCsrfToken is defined once in the codebase (or imported from a pre-existing helper).

5.3 Invoke the agent for remaining targets — sequentially

Per plugins/power-pages/AGENTS.md, agent spawning is sequential. Invoke ai-webapi-integration once per remaining target, waiting for each completion before starting the next. Each target only adds an independent exported function, a framework wrapper (if not already present), and wires one UI file — there are no merge conflicts, but the sequential rule keeps failure modes simple.

If there is only one target total, skip 5.3.

5.4 Replace placeholder POSTs

For any placeholder InstructionIdentifier body the Explore agent flagged in Phase 2, the sub-agent will have replaced them. Confirm by grepping for InstructionIdentifier in the affected files and verifying each resolved call uses the real entity set and id.

5.5 Offer to commit

Don't commit automatically — on iterative runs an unprompted git commit here creates a noisy series of commits for what is effectively one set of changes.

🚦 Gate (consent · add-ai-webapi:5.5.commit): Explicit commit decision after Phase 5 summarization-service + UI wiring is complete.

Trigger: All Phase 5 targets have been wired (service, framework wrapper, UI call sites). Why we ask: Auto-committing on every integration run creates noisy one-commit-per-tweak history; letting the user batch is safer. Cancel leaves: Nothing — source files written; no git commit fired.

Use AskUserQuestion:

On Yes: stage only the files modified during Phase 5 — the summarization service, framework wrapper(s), each wired UI page, and any safe-markdown renderer component the agent emitted. The orchestrator already has this list from each ai-webapi-integration invocation's file-modification report. Use explicit git add <path> per file; do not use git add -A — unrelated work-in-progress files in src/ could otherwise be swept into the commit. Substitute <targets> in the commit message with a short human-readable list (e.g., CaseDetail and SearchResults, or 3 pages when there are many):

git add <each file modified by the Phase 5 agents>
git commit -m "Add AI summarization integration for <targets>"

On Skip: proceed without committing. The user will batch the commit themselves at the end.

Output: summarization service + framework wrappers + UI call sites created for every confirmed target.

---

Phase 6: Configure Layer 3 settings

Goal: Register the Summarization/* site settings via the ai-webapi-settings-architect agent.

6.1 Skip-check

Skip this phase when both of the following are true:

• Every confirmed target is Search Summary (search has no per-call Summarization/* site settings; see ai-api-reference.md).
• No Data Summarization target was added in Phase 3.

For search-only, remind the user to enable Site search with generative AI (preview) in the site's Copilot workspace after deploy, and proceed to Phase 7.

6.2 Confirm deployment prerequisite still holds

Phase 1.5 already gated on .powerpages-site existing. Re-check it here as a guard — if it has disappeared between Phase 1 and now (rare, but possible if the user manually cleaned the folder), stop and re-run the Phase 1.5 bootstrap-deploy prompt. Do NOT silently fall through into the architect with a missing folder.

6.3 Invoke ai-webapi-settings-architect

Invoke the agent at ${PLUGIN_ROOT}/agents/ai-webapi-settings-architect.md via Task:

"Analyse this Power Pages SPA site and propose generative-AI summarization site settings. The following Data Summarization targets were integrated in Phase 5: [list each target with its entity set, per-target InstructionIdentifier value, and target kind (single-record or list) — the architect needs this to decide on Summarization/Data/ContentSizeLimit]. If any target is list, the plan MUST include Summarization/Data/ContentSizeLimit=200000 — the 100k server default silently truncates list content; this is non-negotiable. Check for existing Summarization/* settings. Layer 1 (Webapi/<table>/*) and Layer 2 (table permissions) were configured in Phase 4 via /integrate-webapi in AI-only read mode — verify they are present on disk and cite them as met in your plan's prerequisite table. Propose the AI plan via plan mode, and on approval create the YAMLs with create-site-setting.js."

Wait for the agent to complete. If it reports missing Layer 1/2 prerequisites, something in Phase 4 didn't land — read the file system, identify the gap, and surface it to the user rather than attempting to create Layer 1/2 files here.

6.4 Offer to commit

🚦 Gate (consent · add-ai-webapi:6.4.commit): Explicit commit decision after Summarization/* site settings are created by the architect.

Trigger: ai-webapi-settings-architect has written all Summarization/Data/Enable + Summarization/prompt/<id> YAMLs. Why we ask: Auto-committing could bundle dirty pre-existing YAMLs into the commit; explicit consent scopes the commit to just the architect's output. Cancel leaves: Nothing — YAML files written to disk; no git commit fired.

Use AskUserQuestion:

On Yes: stage only the files the architect just created — use the filePath list returned by ai-webapi-settings-architect (or the per-file paths printed by each create-site-setting.js invocation), not a glob. A glob can sweep in pre-existing Summarization-* YAMLs that are dirty for unrelated reasons (e.g., a prior partial run, or a maker-edited prompt) and bundle them into the commit by accident.

Use explicit git add <path> per file:

git add <each Summarization-*.sitesetting.yml the architect just wrote>
git commit -m "Add AI summarization site settings"

On Skip: proceed without committing.

Output: Summarization/Data/Enable, Summarization/prompt/<id> settings created.

---

Phase 7: Verify

Goal: Confirm every expected file exists, all POSTs set both required headers, and the project builds.

Preview-feature reminder. A green build doesn't mean the API will return a summary at runtime — admin-level governance (tenant PowerShell, Copilot Hub) or the site-level maker toggle for Search Summary can still block it. See the Preview-feature note at the top of this skill and the admin-hierarchy checklist in references/ai-api-reference.md §1 "Troubleshooting: AI feature appears disabled". Tell the user now so the post-deploy test isn't a surprise.

7.1 File inventory

For each confirmed target, confirm:

• Service file: src/services/aiSummaryService.ts (or project-convention equivalent) contains the expected exported function (fetchSearchSummary or fetchDataSummary; the agent may also emit a thin wrapper such as fetchCaseSummary when the user picked the support-case scenario).
• Framework wrapper (non-Astro): React hook in src/hooks/, Vue composable in src/composables/, or Angular service in src/app/services/.
• UI wiring: at least one page/component imports the service or wrapper and calls it.
• Shared API client src/shared/powerPagesApi.ts exists when any Data Summarization target was in scope.
• Summarization/Data/ContentSizeLimit site setting when any list-summary target was integrated. Grep the source for fetchListSummary — if any match exists, confirm .powerpages-site/site-settings/Summarization-Data-ContentSizeLimit.sitesetting.yml is present with value: 200000 (or higher). Missing this setting silently caps list summaries at the 100k server default and ships the user truncated input. If it's missing, surface the gap to the user before completing Phase 7 — re-run the architect or create the YAML manually.

7.2 Header contract grep

Grep: "_api/search/v1\\.0/summary|_api/summarization/data/v1\\.0/" in src/**/*.{ts,tsx,js,jsx,vue,astro}

For every file that matches, verify the surrounding fetch includes:

• __RequestVerificationToken (CSRF token, fetched from /_layout/tokenhtml) — hard rule
• X-Requested-With: XMLHttpRequest — recommended (the validator warns when it's missing but does not block; it matches shell.ajaxSafePost and every other Power Pages POST)

For data summarization calls, additionally verify (both hard rules):

• The URL contains $select= (no wildcards)
• OData-MaxVersion: 4.0 and OData-Version: 4.0 headers are set

Fix any missing hard-rule header before proceeding. Missing CSRF produces 500s; missing $select or OData headers produces 403/400s. A missing X-Requested-With only produces a validator warning — add it for consistency, but it does not break the run.

7.3 Build check

cd "<PROJECT_ROOT>"
npm run build

Fix any type or import errors. Common issues: missing import of the summarization service in a wired page; type mismatch between DataSummaryResponse and the UI consumer; duplicate getCsrfToken declarations (if Phase 5 failed to reuse the existing helper).

7.4 Present verification results

(Same row order and example file paths as the Phase 2 manifest example, so a maintainer reading both tables can trace each row top-to-bottom. Row 2 mirrors the Microsoft-shipped support-case recipe — a Data Summarization call configured for incidents with the case_summary prompt identifier.)

Build status: Pass / Fail (with details).

Output: all integration files verified; project builds.

---

Phase 8: Review & Deploy

Goal: Record skill usage, present a summary, and offer deployment.

8.1 Record skill usage

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

Use --skillName "AddAiWebapi".

8.2 Present summary

8.3 Ask to deploy

Yes: invoke /deploy-site. No: acknowledge. Remind that the API calls will not work until the site is deployed with the new settings and permissions.

8.4 Post-deploy notes

Surface these to the user at the end of the run:

• Search Summary toggle — for any Search Summary integration, remind the user to flip Set up workspace → Copilot → Site search (preview) → "Enable Site search with generative AI (preview)". Without it, the endpoint returns the disabled envelope and the UI renders the remediation card instead of a summary.
• Test recipe — for Data Summarization, open a record-detail or list page, trigger the summary, confirm 200 + rendered text, then click a recommendation chip and confirm the follow-up sends RecommendationConfig. For Search Summary, perform a search and confirm the summary renders above keyword hits with citation anchors pointing at the SPA route (not /page-not-found/). The 90041005 "nothing to summarize" branch is normal for the Microsoft-shipped support-case recipe on a freshly-created case with no comments — test with a record that has substantive content in every selected / expanded column.
• When the disabled-state card or 90041001 shows up — walk the admin-hierarchy checklist in references/ai-api-reference.md §1 "Troubleshooting: AI feature appears disabled (admin hierarchy)". Retry doesn't help; an admin or maker has to change governance / flip the site toggle.
• 403 on any summarization call is always a Layer 1/2 issue (column casing in Webapi/<table>/fields, or missing read: true table permission) — re-run /integrate-webapi in AI-only read mode rather than hand-editing YAML.
• Full error-code reference (90041001 … 90041006) is in references/ai-api-reference.md §2 "Error codes". Open it when the user reports a 400.
• Column permission profiles can silently hide content. If a summary has obvious omissions, check Dataverse column permission profiles on the web role before suspecting the prompt or fields list.

Output: summary presented, deployment completed or deferred, post-deploy guidance given.

---

Important Notes

Throughout all phases

• Use TaskCreate/TaskUpdate to track progress at every phase.
• Ask for user confirmation at key decision points (list below).
• Sequential agent spawning — per the "Agent spawning" rule in plugins/power-pages/AGENTS.md. Never spawn ai-webapi-integration in parallel across targets (every target extends the same aiSummaryService.* file, so parallel runs would conflict).
• Commit at milestones — after implementation (Phase 5) and after settings creation (Phase 6).
• Never use an OData wrapper for summarization fetches — raw fetch only.
• Never write Layer 1/2 files directly — always delegate to /integrate-webapi / /create-webroles. This skill is Layer 3.

Key decision points (wait for user)

The list below is the conservative upper bound — many runs hit fewer prompts because skip checks (search-only run, Layer 1/2 ready, no list/intent-mismatch targets) eliminate whole branches. Realistic worst case for a multi-target first-time run is ~10 prompts; a clean re-entry tweak is just 1 prompt (the Phase 1.0 mode question).

1. At Phase 1.0 (re-entry detection): tweak existing surface, add a new surface, or review what's already wired. Only fires when a previous /add-ai-webapi run left a service file or Summarization/* settings on disk.
2. At Phase 1.5: bootstrap deploy or stop (if .powerpages-site is missing).
3. After Phase 3 (main): confirm which APIs / targets to integrate.
4. Phase 3 per-target follow-ups (variable count): list-trigger choice (one per list target), and scope-confirmation (one per scope-extends-beyond-existing-fetch, needs-definition, or intent-mismatch row). Question text and option lists live in references/scope-classification.md.
5. At Phase 4.2: create missing web role via /create-webroles (if needed). The sub-skill is invoked with the caller-suppress sentinel — its own deploy prompts don't fire.
6. At Phase 4.2 (Skip path only): confirm continuing despite known broken-runtime risk.
7. Inside the Phase 4.3 /integrate-webapi delegation: approve its table-permissions-architect plan and its webapi-settings-architect plan (each architect owns its own plan-mode prompt; the sub-skill is in AI-only read mode so its commit and deploy prompts are suppressed).
8. At Phase 5.5: commit the integration changes now or later.
9. Inside the Phase 6.3 ai-webapi-settings-architect call: approve its plan.
10. At Phase 6.4: commit the new settings now or later.
11. At Phase 8.3: deploy now or later.

List-summary use case playbook

When the target is a LIST of records (not a single record), the defaults for a single-record Copilot card are the wrong defaults — collection endpoint, tabular-insight prompt, ContentSizeLimit=200000, normalizeSummaryString + safe-markdown renderer, etc.

The full 10-rule playbook (collection endpoint, scope mirroring, prompt size, nav-property casing, mandatory ContentSizeLimit) lives in ${PLUGIN_ROOT}/skills/add-ai-webapi/references/ai-api-reference.md §2 "List-summary playbook". Read that section before any list-summary target reaches Phase 5 — the Phase 5 agent and the Phase 6 settings architect both reference the same playbook.

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