@aglyn/plugins-ai 1.0.0-beta.143

package/src/lib/jobs/ai-job-seo-step.js

@@ -0,0 +1,809 @@
+import { _ as _extends } from "@swc/helpers/_/_extends";
+/**
+ * @license
+ * Copyright 2026 Aglyn LLC
+ *
+ * Licensed under the Apache License, Version 2.0 (the "License");
+ * you may not use this file except in compliance with the License.
+ * You may obtain a copy of the License at
+ *
+ *   http://www.apache.org/licenses/LICENSE-2.0
+ *
+ * Unless required by applicable law or agreed to in writing, software
+ * distributed under the License is distributed on an "AS IS" BASIS,
+ * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+ * See the License for the specific language governing permissions and
+ * limitations under the License.
+ */ import { hostCollectionKind } from "@aglyn/aglyn/app-utils/collection-kind";
+import { hostPublicOrigin } from "@aglyn/aglyn/app-utils/host-naming";
+import { buildLlmsTxt } from "@aglyn/aglyn/app-utils/llms-txt";
+import { nodesReferenceScreen } from "@aglyn/aglyn/app-utils/screen-link-value";
+import { screenRoutePathToUrl } from "@aglyn/aglyn/app-utils/screen-route";
+import { isScreenIndexable, isSearchDiscouraged, statusPageScreenIds } from "@aglyn/aglyn/app-utils/search-indexing";
+import { SCREEN_SEO_LISTING_FIELDS, isSeoListingFieldKey } from "@aglyn/aglyn/app-utils/seo-listing-fields";
+import { decodeStoredNodes } from "@aglyn/aglyn/app-utils/stored-nodes";
+import { getTemplateScreenIds } from "@aglyn/tenant-runtime/template-screens";
+import { AI_SEO_FINDING_LABELS, AI_SEO_OUTPUT_IDS, aiSeoAuditView, aiSeoJobTarget } from "../model/ai-seo.js";
+import { aiModelForStep } from "../providers/routing.js";
+import { AI_SEO_AUDIT_MAX_PAGES, aiSeoAudit, aiSeoAuditBatches, aiSeoHeadingDemotions, aiSeoNormalizePath, parseAiSeoKeywordLines } from "../runtime/seo-audit.js";
+import { runValidatedGeneration } from "../runtime/ai-doctrine.js";
+import { AI_SEO_FIELDS_MAX_TOKENS, generateSeoFields } from "../runtime/seo-fields.js";
+import { AI_SEO_PAGE_TEXT_MAX_CHARS, aiSeoPageFacts } from "../runtime/seo-page-facts.js";
+import { AI_SEO_FIXES_TOOL_NAME, AI_SEO_SITE_TOOL_NAME, aiSeoFixImages, aiSeoFixesTool, aiSeoKeywordCoverage, aiSeoKeywordList, aiSeoSiteTool, checkAiSeoFixes, checkAiSeoSite } from "../tools/ai-seo-tool.js";
+import { AI_SEO_FIXES_MAX_TOKENS, AI_SEO_SITE_MAX_TOKENS, aiSeoGenerationMaxTokens } from "./ai-job-seo-budget.js";
+/**
+ * The `seo` step (AGL-2910): search listings and a site audit, as proposals.
+ *
+ * `inputs.target` says what the job is about:
+ *
+ * - `screen` — one page's listing, written from the version the person is
+ *   looking at (`screenId`, `versionId`), for the page's SEO card;
+ * - `product` — one product's listing, from the name and description the
+ *   product editor handed over (`name`, `text`), for the product editor;
+ * - `site` — the audit of every published page the sitemap lists, with a
+ *   proposed fix per finding, structured data for the site's entity, and the
+ *   agent guidance `/llms.txt` leads with.
+ *
+ * ## It writes nothing
+ *
+ * Like every runner, this reads, asks and returns. A listing is served
+ * straight from the screen document, the structured data and `/llms.txt`
+ * from the host document, and a page's content from its published version,
+ * so the values ride on the job's outputs and nothing else: a person puts
+ * them in the page's SEO card or the site SEO form and saves them there, and
+ * content fixes become a new version only through the apply door.
+ *
+ * ## An audit, a unit at a time
+ *
+ * The first pass reads the site — every published page's version and the
+ * shared layouts, for the links between them — scores it without a model,
+ * records the report, and runs the first unit of generated work. Each later
+ * pass runs one more unit: the site-wide proposal, then the page fixes in
+ * batches. Every pass is one reservation and one provider exchange, and it
+ * asks the machine to continue until nothing is left, so a large site is a
+ * few beats of work rather than one request that cannot finish.
+ */ export const AI_SEO_NO_SITE_COPY = 'This SEO job names a site this workspace does not have.';
+export const AI_SEO_NO_PAGE_COPY = 'This page is no longer on the site.';
+export const AI_SEO_NO_TARGET_COPY = 'This SEO job does not say what to write about.';
+export const AI_SEO_NO_PRODUCT_COPY = 'Give the product a name before writing its listing.';
+/** Shared layouts an audit reads for the links in a site's navigation. */ export const AI_SEO_LAYOUT_SCAN_LIMIT = 20;
+/** How much of each page a batch of fixes carries. */ export const AI_SEO_BATCH_PAGE_TEXT_CHARS = 700;
+/**
+ * Output budgets, declared with the step's time (`ai-job-seo-budget.ts`,
+ * AGL-3035). A batch's largest answer — every page with a title, description,
+ * heading and image descriptions at their limits — and the site proposal's are
+ * measured in `ai-job-seo-step.spec.ts`.
+ */ export { AI_SEO_FIXES_MAX_TOKENS, AI_SEO_SITE_MAX_TOKENS };
+/** Documents read at once. */ const READ_CHUNK = 10;
+/** Pages the site prompt and the llms.txt preview list. */ const SITE_PAGES_LISTED = 25;
+/** The fields a product's listing holds. */ const PRODUCT_LISTING_FIELDS = [
+    'title',
+    'description'
+];
+const ZERO_USAGE = {
+    inputTokens: 0,
+    outputTokens: 0,
+    cacheReadTokens: 0,
+    cacheWriteTokens: 0
+};
+const str = (value)=>typeof value === 'string' ? value.trim() : '';
+const round6 = (value)=>Math.round(value * 1000000) / 1000000;
+function modelOf(context) {
+    var _ref;
+    return (_ref = context.modelFor == null ? void 0 : context.modelFor.call(context, 'job.seo')) != null ? _ref : aiModelForStep('job.seo');
+}
+function spentNothing(model) {
+    return {
+        usage: ZERO_USAGE,
+        estCostUsd: 0,
+        model,
+        stopReason: null
+    };
+}
+function spentBy(generation) {
+    return {
+        usage: generation.usage,
+        estCostUsd: generation.estCostUsd,
+        model: generation.model,
+        stopReason: generation.stopReason
+    };
+}
+function addSpent(a, b) {
+    var _b_stopReason;
+    return {
+        usage: {
+            inputTokens: a.usage.inputTokens + b.usage.inputTokens,
+            outputTokens: a.usage.outputTokens + b.usage.outputTokens,
+            cacheReadTokens: a.usage.cacheReadTokens + b.usage.cacheReadTokens,
+            cacheWriteTokens: a.usage.cacheWriteTokens + b.usage.cacheWriteTokens
+        },
+        estCostUsd: round6(a.estCostUsd + b.estCostUsd),
+        model: b.model || a.model,
+        stopReason: (_b_stopReason = b.stopReason) != null ? _b_stopReason : a.stopReason
+    };
+}
+/** Through JSON, so no `undefined` reaches the job document's write. */ const plain = (value)=>JSON.parse(JSON.stringify(value));
+/**
+ * The fields a listing job writes: the ones the editor named (`inputs.fields`,
+ * comma-separated) that the target's editor holds, or all of those.
+ */ export function aiSeoRequestedFields(raw, allowed) {
+    const named = String(raw != null ? raw : '').split(',').map((part)=>part.trim()).filter(isSeoListingFieldKey);
+    const wanted = new Set(named.length ? named : allowed);
+    return allowed.filter((key)=>wanted.has(key));
+}
+/** The site's name, as its visitors read it. */ function brandOf(host) {
+    var _host_seo;
+    return str((_host_seo = host.seo) == null ? void 0 : _host_seo.title) || str(host.displayName) || str(host.subdomain);
+}
+/** The site a job names, read and checked against the job's org. */ async function loadSite(firestore, job) {
+    if (!job.hostId) return {
+        failure: AI_SEO_NO_SITE_COPY
+    };
+    const snapshot = await firestore.collection('hosts').doc(job.hostId).get();
+    const host = snapshot.exists ? snapshot.data() : null;
+    // The route checks the caller's membership of the org the job is metered
+    // against, not that the site belongs to it. A site of another org is
+    // refused here, before anything about it reaches a prompt.
+    if (!host || host.orgId !== job.orgId) return {
+        failure: AI_SEO_NO_SITE_COPY
+    };
+    return {
+        host
+    };
+}
+/** A version's node map and root, or `null` when it cannot be read. */ async function readVersionNodes(hostRef, screenId, versionId) {
+    if (!versionId) return null;
+    const snapshot = await hostRef.collection('screens').doc(screenId).collection('versions').doc(versionId).get();
+    if (!snapshot.exists) return null;
+    const nodes = decodeStoredNodes(snapshot.get('nodes'));
+    if (!nodes) return null;
+    const rootId = snapshot.get('rootId');
+    return _extends({
+        nodes
+    }, typeof rootId === 'string' ? {
+        rootId
+    } : {});
+}
+/** Every other page's search title, for a proposal not to repeat. */ async function siteTitles(hostRef, except) {
+    const snapshot = await hostRef.collection('screens').select('seo', 'deletedAt').limit(1000).get();
+    return snapshot.docs.filter((doc)=>!except.has(doc.id) && !doc.get('deletedAt')).map((doc)=>{
+        var _doc_get;
+        return str((_doc_get = doc.get('seo')) == null ? void 0 : _doc_get.title);
+    }).filter(Boolean);
+}
+/* ------------------------------------------------------------------------ *
+ * One listing: a page or a product
+ * ------------------------------------------------------------------------ */ function listingOutcome(generation, job, subject, context) {
+    var _job_hostId;
+    var _context_facts_h1s_, _context_facts, _context_facts1;
+    const spent = spentBy(generation);
+    if (generation.status === 'refused') return _extends({
+        outputs: []
+    }, spent, {
+        refused: true
+    });
+    if (generation.status === 'needs_input') return _extends({
+        outputs: []
+    }, spent, {
+        failure: generation.message
+    });
+    const values = generation.value;
+    const notes = [];
+    if (context.fields.includes('imageAlt') && context.hasImage && !values.imageAlt) {
+        notes.push('No image description was proposed: nothing on the page says what the picture shows.');
+    }
+    const proposal = {
+        kind: 'fields',
+        subject,
+        values,
+        keywords: aiSeoKeywordCoverage(context.keywords, {
+            title: values.title,
+            description: values.description,
+            h1: (_context_facts = context.facts) == null ? void 0 : (_context_facts_h1s_ = _context_facts.h1s[0]) == null ? void 0 : _context_facts_h1s_.text,
+            body: (_context_facts1 = context.facts) == null ? void 0 : _context_facts1.text
+        }),
+        notes
+    };
+    const output = {
+        resource: 'seo',
+        id: AI_SEO_OUTPUT_IDS.fields(subject.kind, subject.id),
+        hostId: (_job_hostId = job.hostId) != null ? _job_hostId : null,
+        label: `SEO proposal · ${subject.name}`,
+        proposal: plain(proposal)
+    };
+    return _extends({
+        outputs: [
+            output
+        ]
+    }, spent);
+}
+async function runScreenListing(context, host) {
+    var _screen_seo;
+    var _job_inputs, _job_inputs1, _job_inputs2, _host_screens, _job_inputs3;
+    const { job, firestore, signal } = context;
+    const model = modelOf(context);
+    const hostRef = firestore.collection('hosts').doc(job.hostId);
+    const screenId = str((_job_inputs = job.inputs) == null ? void 0 : _job_inputs['screenId']);
+    const snapshot = screenId ? await hostRef.collection('screens').doc(screenId).get() : null;
+    const screen = (snapshot == null ? void 0 : snapshot.exists) ? snapshot.data() : null;
+    if (!screen || screen.deletedAt) {
+        return _extends({
+            outputs: []
+        }, spentNothing(model), {
+            failure: AI_SEO_NO_PAGE_COPY
+        });
+    }
+    const versionId = str((_job_inputs1 = job.inputs) == null ? void 0 : _job_inputs1['versionId']) || str(screen.versionId);
+    const version = await readVersionNodes(hostRef, screenId, versionId);
+    const facts = aiSeoPageFacts(version == null ? void 0 : version.nodes, {
+        rootId: version == null ? void 0 : version.rootId
+    });
+    const fields = aiSeoRequestedFields((_job_inputs2 = job.inputs) == null ? void 0 : _job_inputs2['fields'], SCREEN_SEO_LISTING_FIELDS);
+    const seo = (_screen_seo = screen.seo) != null ? _screen_seo : {};
+    const hasImage = Boolean(str(seo.image));
+    // The share image is often a picture the page shows too: its description
+    // and the text beside it there are what is known about it.
+    const onPage = hasImage ? facts.images.find((image)=>image.src === str(seo.image)) : undefined;
+    const route = (_host_screens = host.screens) == null ? void 0 : _host_screens[screenId];
+    const path = typeof route === 'string' ? screenRoutePathToUrl(route) : null;
+    const name = str(screen.displayName) || path || 'Untitled page';
+    const keywords = aiSeoKeywordList((_job_inputs3 = job.inputs) == null ? void 0 : _job_inputs3['keywords']);
+    const generation = await generateSeoFields(_extends({
+        subject: {
+            kind: 'screen',
+            name,
+            path
+        },
+        brand: brandOf(host),
+        text: facts.text || str(screen.description),
+        fields,
+        current: {
+            title: str(seo.title),
+            description: str(seo.description),
+            breadcrumb: str(seo.breadcrumb),
+            imageAlt: str(seo.imageAlt)
+        },
+        image: hasImage ? {
+            alt: str(seo.imageAlt) || (onPage == null ? void 0 : onPage.alt) || null,
+            context: (onPage == null ? void 0 : onPage.context) || null
+        } : null,
+        keywords,
+        otherTitles: await siteTitles(hostRef, new Set([
+            screenId
+        ])),
+        model,
+        maxTokens: aiSeoGenerationMaxTokens(model, AI_SEO_FIELDS_MAX_TOKENS)
+    }, signal ? {
+        signal
+    } : {}));
+    return listingOutcome(generation, job, {
+        kind: 'screen',
+        id: screenId,
+        name,
+        path
+    }, {
+        keywords,
+        facts,
+        fields,
+        hasImage
+    });
+}
+async function runProductListing(context, host) {
+    var _job_inputs, _job_inputs1, _job_inputs2, _job_inputs3, _job_inputs4, _job_inputs5, _job_inputs6;
+    const { job, signal } = context;
+    const model = modelOf(context);
+    const name = str((_job_inputs = job.inputs) == null ? void 0 : _job_inputs['name']).slice(0, 200);
+    if (!name) return _extends({
+        outputs: []
+    }, spentNothing(model), {
+        failure: AI_SEO_NO_PRODUCT_COPY
+    });
+    const fields = aiSeoRequestedFields((_job_inputs1 = job.inputs) == null ? void 0 : _job_inputs1['fields'], PRODUCT_LISTING_FIELDS);
+    const keywords = aiSeoKeywordList((_job_inputs2 = job.inputs) == null ? void 0 : _job_inputs2['keywords']);
+    const productId = str((_job_inputs3 = job.inputs) == null ? void 0 : _job_inputs3['productId']) || null;
+    const generation = await generateSeoFields(_extends({
+        subject: {
+            kind: 'product',
+            name
+        },
+        brand: brandOf(host),
+        // The product's own words, as its editor handed them over: the product
+        // document is the commerce plugin's, and this step never reads it.
+        text: str((_job_inputs4 = job.inputs) == null ? void 0 : _job_inputs4['text']).slice(0, AI_SEO_PAGE_TEXT_MAX_CHARS),
+        fields,
+        current: {
+            title: str((_job_inputs5 = job.inputs) == null ? void 0 : _job_inputs5['currentTitle']),
+            description: str((_job_inputs6 = job.inputs) == null ? void 0 : _job_inputs6['currentDescription'])
+        },
+        image: null,
+        keywords,
+        model,
+        maxTokens: aiSeoGenerationMaxTokens(model, AI_SEO_FIELDS_MAX_TOKENS)
+    }, signal ? {
+        signal
+    } : {}));
+    return listingOutcome(generation, job, {
+        kind: 'product',
+        id: productId,
+        name
+    }, {
+        keywords,
+        facts: null,
+        fields,
+        hasImage: false
+    });
+}
+/**
+ * Read the site the way its sitemap lists it: every screen the routing map
+ * publishes, less the template screens, the status pages and every screen
+ * that is not public — through the same shared predicates the tenant's
+ * `sitemap.xml` and `/llms.txt` apply, so the audit covers exactly the pages
+ * a crawler is handed. Then each page's published version, and the shared
+ * layouts, for the links that decide which pages nothing points to.
+ */ export async function scanAiSeoSite(firestore, hostId, host, rawKeywords) {
+    var _host_screens;
+    const hostRef = firestore.collection('hosts').doc(hostId);
+    const routing = (_host_screens = host.screens) != null ? _host_screens : {};
+    const [screenDocs, templateScreenIds, layoutDocs] = await Promise.all([
+        hostRef.collection('screens').select('visibility', 'seo', 'description', 'displayName', 'versionId', 'deletedAt').limit(1000).get(),
+        getTemplateScreenIds({
+            hostId
+        }),
+        hostRef.collection('layouts').select('versionId').limit(AI_SEO_LAYOUT_SCAN_LIMIT).get()
+    ]);
+    const screens = new Map(screenDocs.docs.map((doc)=>[
+            doc.id,
+            doc.data()
+        ]));
+    const excluded = new Set([
+        ...statusPageScreenIds(host),
+        ...templateScreenIds
+    ]);
+    const published = Object.entries(routing).filter(([screenId])=>!excluded.has(screenId)).filter(([screenId])=>{
+        const screen = screens.get(screenId);
+        return Boolean(screen) && !(screen == null ? void 0 : screen.deletedAt) && isScreenIndexable(screen);
+    }).map(([screenId, route])=>({
+            screenId,
+            path: screenRoutePathToUrl(String(route))
+        })).sort((a, b)=>a.path < b.path ? -1 : a.path > b.path ? 1 : 0);
+    const audited = published.slice(0, AI_SEO_AUDIT_MAX_PAGES);
+    const versions = new Map();
+    for(let start = 0; start < audited.length; start += READ_CHUNK){
+        await Promise.all(audited.slice(start, start + READ_CHUNK).map(async ({ screenId })=>{
+            var _screens_get;
+            versions.set(screenId, await readVersionNodes(hostRef, screenId, str((_screens_get = screens.get(screenId)) == null ? void 0 : _screens_get.versionId)));
+        }));
+    }
+    const layoutNodes = [];
+    const layouts = layoutDocs.docs.filter((doc)=>str(doc.get('versionId')));
+    for(let start = 0; start < layouts.length; start += READ_CHUNK){
+        await Promise.all(layouts.slice(start, start + READ_CHUNK).map(async (doc)=>{
+            const snapshot = await doc.ref.collection('versions').doc(str(doc.get('versionId'))).get();
+            const nodes = snapshot.exists ? decodeStoredNodes(snapshot.get('nodes')) : null;
+            if (nodes) layoutNodes.push(nodes);
+        }));
+    }
+    const keywordsByPath = parseAiSeoKeywordLines(rawKeywords);
+    const auditedPaths = new Set(audited.map((page)=>aiSeoNormalizePath(page.path)));
+    const notes = Object.keys(keywordsByPath).filter((path)=>!auditedPaths.has(path)).map((path)=>`Keywords for ${path} were not used: no audited page is published at that address.`);
+    const pages = audited.map(({ screenId, path })=>{
+        var _screens_get, _versions_get, _screen_seo, _keywordsByPath_aiSeoNormalizePath, _ref;
+        const screen = (_screens_get = screens.get(screenId)) != null ? _screens_get : {};
+        const version = (_versions_get = versions.get(screenId)) != null ? _versions_get : null;
+        const linkedFrom = layoutNodes.some((nodes)=>nodesReferenceScreen(nodes, screenId)) || audited.some((other)=>{
+            var _ref;
+            var _versions_get;
+            return other.screenId !== screenId && nodesReferenceScreen((_ref = (_versions_get = versions.get(other.screenId)) == null ? void 0 : _versions_get.nodes) != null ? _ref : null, screenId);
+        });
+        return {
+            screenId,
+            path,
+            name: str(screen.displayName) || path,
+            versionId: str(screen.versionId) || null,
+            seo: (_screen_seo = screen.seo) != null ? _screen_seo : {},
+            description: str(screen.description),
+            facts: aiSeoPageFacts(version == null ? void 0 : version.nodes, {
+                rootId: version == null ? void 0 : version.rootId
+            }),
+            linkedFrom,
+            keywords: (_keywordsByPath_aiSeoNormalizePath = keywordsByPath[aiSeoNormalizePath(path)]) != null ? _keywordsByPath_aiSeoNormalizePath : [],
+            nodes: (_ref = version == null ? void 0 : version.nodes) != null ? _ref : null
+        };
+    });
+    return {
+        pages,
+        skipped: Math.max(0, published.length - audited.length),
+        notes
+    };
+}
+/** The units an audit owes after the outputs recorded so far, in the order they run. */ export function aiSeoPendingUnits(report, outputs) {
+    const done = new Set(outputs.map((output)=>output.id));
+    const units = [];
+    if (report.siteProposal && !done.has(AI_SEO_OUTPUT_IDS.site)) units.push({
+        kind: 'site'
+    });
+    const batches = aiSeoAuditBatches(report);
+    batches.forEach((screenIds, index)=>{
+        if (!done.has(AI_SEO_OUTPUT_IDS.fixes(index + 1))) {
+            units.push({
+                kind: 'batch',
+                batch: index + 1,
+                of: batches.length,
+                screenIds
+            });
+        }
+    });
+    return units;
+}
+/* ---- The site-wide proposal ---------------------------------------------- */ export const AI_SEO_SITE_INSTRUCTIONS = [
+    {
+        text: 'You propose two things for a website: the structured data that says who publishes it, and ' + 'the guidance an AI agent reads first in its /llms.txt.\n\n' + `Answer by calling ${AI_SEO_SITE_TOOL_NAME} exactly once. A reply in prose cannot be used.\n\n` + 'Rules:\n' + '- Fill only the fields you are told are empty; answer null for every other one.\n' + '- Use only what the pages you are given state. A contact email or telephone number goes in ' + 'only when a page shows it exactly; never compose one.\n' + '- The entity description says what the publisher is, in one plain sentence, not what one page ' + 'offers.\n' + '- The agent guidance is checkable and specific: the questions this site answers and which ' + 'pages answer them. No marketing language: an agent discounts a claim it cannot check.\n' + '- Keep every length the tool states, and write in the language of the pages.',
+        cacheBreakpoint: true
+    }
+];
+/** The site SEO form fields blank on the site now. */ export function aiSeoBlankSiteFields(host) {
+    var _ref, _ref1;
+    var _host_seo, _host_seo1;
+    const entity = (_ref = (_host_seo = host.seo) == null ? void 0 : _host_seo.entity) != null ? _ref : {};
+    const agent = (_ref1 = (_host_seo1 = host.seo) == null ? void 0 : _host_seo1.agent) != null ? _ref1 : {};
+    const blank = new Set();
+    if (entity.type === undefined || entity.type === null || entity.type === '') blank.add('seo.entity.type');
+    if (!str(entity.name)) blank.add('seo.entity.name');
+    if (!str(entity.description)) blank.add('seo.entity.description');
+    if (!str(entity.email)) blank.add('seo.entity.email');
+    if (!str(entity.telephone)) blank.add('seo.entity.telephone');
+    if (!str(entity.contactType)) blank.add('seo.entity.contactType');
+    if (!str(agent.whenToUse)) blank.add('seo.agent.whenToUse');
+    if (!str(agent.howToUse)) blank.add('seo.agent.howToUse');
+    return blank;
+}
+/** The site's content collections, for the llms.txt preview. */ async function readContentCollections(hostRef) {
+    try {
+        const snapshot = await hostRef.collection('collections').select('slug', 'displayName', 'kind').limit(50).get();
+        return snapshot.docs.filter((doc)=>hostCollectionKind(doc.data()) === 'content' && str(doc.get('slug'))).map((doc)=>({
+                slug: str(doc.get('slug')),
+                name: str(doc.get('displayName')) || undefined
+            }));
+    } catch (unused) {
+        // A preview without the feeds is still a preview of the guidance.
+        return [];
+    }
+}
+/** `/llms.txt` as it reads with `values` saved over what the site stores now. */ export function aiSeoLlmsPreview(host, values, pages, collections) {
+    var _hostPublicOrigin, _values_seoagentwhenToUse, _values_seoagenthowToUse, _values_seoentityemail;
+    var _host_seo, _host_seo1, _host_seo_agent, _host_seo2, _host_seo_agent1, _host_seo3, _host_seo_entity, _host_seo4;
+    return buildLlmsTxt({
+        siteName: str((_host_seo = host.seo) == null ? void 0 : _host_seo.title) || str(host.displayName) || 'Site',
+        origin: (_hostPublicOrigin = hostPublicOrigin(host)) != null ? _hostPublicOrigin : '',
+        description: str((_host_seo1 = host.seo) == null ? void 0 : _host_seo1.description) || undefined,
+        agent: {
+            whenToUse: (_values_seoagentwhenToUse = values['seo.agent.whenToUse']) != null ? _values_seoagentwhenToUse : (_host_seo2 = host.seo) == null ? void 0 : (_host_seo_agent = _host_seo2.agent) == null ? void 0 : _host_seo_agent.whenToUse,
+            howToUse: (_values_seoagenthowToUse = values['seo.agent.howToUse']) != null ? _values_seoagenthowToUse : (_host_seo3 = host.seo) == null ? void 0 : (_host_seo_agent1 = _host_seo3.agent) == null ? void 0 : _host_seo_agent1.howToUse
+        },
+        pages,
+        collections,
+        hasSearch: true,
+        contactEmail: (_values_seoentityemail = values['seo.entity.email']) != null ? _values_seoentityemail : str((_host_seo4 = host.seo) == null ? void 0 : (_host_seo_entity = _host_seo4.entity) == null ? void 0 : _host_seo_entity.email) || undefined
+    });
+}
+async function proposeSite(context, host, scan) {
+    var _job_hostId;
+    var _host_seo;
+    const { job, firestore, signal } = context;
+    const model = modelOf(context);
+    const hostRef = firestore.collection('hosts').doc(job.hostId);
+    const blank = aiSeoBlankSiteFields(host);
+    const collections = await readContentCollections(hostRef);
+    const topLevel = scan.pages.filter((page)=>!page.path.replace(/^\//, '').includes('/')).slice(0, SITE_PAGES_LISTED).map((page)=>_extends({
+            path: page.path
+        }, page.name ? {
+            title: page.name
+        } : {}));
+    // The pages that say who publishes a site: the home page, and whatever is
+    // about the company or how to reach it.
+    const telling = scan.pages.filter((page)=>page.path === '/' || /about|contact|company|team|who/i.test(page.path)).slice(0, 4);
+    const prompt = [
+        `Site: ${brandOf(host) || 'untitled site'}`,
+        ((_host_seo = host.seo) == null ? void 0 : _host_seo.description) ? `Site description: ${str(host.seo.description)}` : '',
+        `Empty fields you may fill: ${[
+            ...blank
+        ].join(', ') || 'none'}`,
+        'Pages:',
+        ...scan.pages.slice(0, SITE_PAGES_LISTED).map((page)=>`- ${page.path} — ${page.name}${str(page.seo.title) ? ` — ${str(page.seo.title)}` : ''}`),
+        collections.length ? `Content collections: ${collections.map((entry)=>entry.name || entry.slug).join(', ')}` : '',
+        ...telling.flatMap((page)=>[
+                '',
+                `Text of ${page.path}:`,
+                page.facts.text.slice(0, 1200) || '(no text)'
+            ])
+    ].filter((line)=>line !== '').join('\n');
+    const siteText = scan.pages.map((page)=>page.facts.text).join('\n');
+    const generation = await runValidatedGeneration('seo-site', _extends({
+        step: 'job.seo',
+        model,
+        instructions: AI_SEO_SITE_INSTRUCTIONS,
+        messages: [
+            {
+                role: 'user',
+                content: prompt
+            }
+        ],
+        tool: aiSeoSiteTool(),
+        maxTokens: aiSeoGenerationMaxTokens(model, AI_SEO_SITE_MAX_TOKENS)
+    }, signal ? {
+        signal
+    } : {}, {
+        check: (answer)=>checkAiSeoSite(answer, {
+                blank,
+                siteText
+            })
+    }));
+    const proposed = generation.status === 'ok' ? generation.value : {
+        values: {},
+        notes: []
+    };
+    const notes = generation.status === 'ok' ? proposed.notes : [
+        generation.status === 'refused' ? 'The AI declined to propose structured data for this site.' : generation.message
+    ];
+    const site = {
+        values: proposed.values,
+        llmsPreview: aiSeoLlmsPreview(host, proposed.values, topLevel, collections),
+        notes
+    };
+    return {
+        output: {
+            resource: 'seo',
+            id: AI_SEO_OUTPUT_IDS.site,
+            hostId: (_job_hostId = job.hostId) != null ? _job_hostId : null,
+            label: 'SEO audit · structured data and agent guidance',
+            proposal: plain({
+                kind: 'site',
+                site
+            })
+        },
+        spent: spentBy(generation)
+    };
+}
+/* ---- A batch of page fixes ------------------------------------------------ */ export const AI_SEO_FIXES_INSTRUCTIONS = [
+    {
+        text: 'You fix the search listings and headings of a website’s pages. For each page you are given ' + 'what an audit found, what the page shows now, and its text.\n\n' + `Answer by calling ${AI_SEO_FIXES_TOOL_NAME} exactly once, with one entry per page. A reply in ` + 'prose cannot be used.\n\n' + 'Rules:\n' + '- Fix only what the audit found. Answer null for a value it did not find a problem with.\n' + '- Write only from the page’s own text. Never invent a fact, a price, a name, a place or a claim.\n' + '- A title says what the page is, specifically. No two pages share a title, and none repeats a ' + 'title you are told another page uses.\n' + '- A description is one or two sentences telling a searcher what they will find.\n' + '- The main heading names what the page is about in a few plain words.\n' + '- An image description says what the picture most likely shows, from the text beside it. ' + 'Leave an image out when the text gives no idea.\n' + '- A target keyword goes in only where the page is about it and it reads naturally: at most ' + 'once in a title and once in a description. Never list keywords or repeat a word to rank.\n' + '- Keep every length the tool states, and write in the language of each page.',
+        cacheBreakpoint: true
+    }
+];
+/** The fixes one page gets: the model's values, its content fixes, and the guidance no fix can give. */ export function aiSeoPageFixFor(page, answer) {
+    var _ref;
+    const values = {};
+    if (answer == null ? void 0 : answer.title) values.title = answer.title;
+    if (answer == null ? void 0 : answer.description) values.description = answer.description;
+    const content = [];
+    for (const alt of (_ref = answer == null ? void 0 : answer.imageAlts) != null ? _ref : [])content.push({
+        kind: 'image-alt',
+        nodeId: alt.nodeId,
+        alt: alt.alt
+    });
+    if (answer == null ? void 0 : answer.h1) {
+        const first = page.facts.h1s[0];
+        const promotable = page.facts.headings.find((heading)=>heading.editable && heading.nodeId);
+        if (page.codes.has('h1-thin') && (first == null ? void 0 : first.editable) && first.nodeId) {
+            content.push({
+                kind: 'h1-set',
+                nodeId: first.nodeId,
+                text: answer.h1
+            });
+        } else if (page.codes.has('h1-missing') && (promotable == null ? void 0 : promotable.nodeId)) {
+            content.push({
+                kind: 'h1-set',
+                nodeId: promotable.nodeId,
+                text: answer.h1
+            });
+        } else if (page.codes.has('h1-missing')) {
+            content.push({
+                kind: 'h1-insert',
+                text: answer.h1
+            });
+        }
+    }
+    if (page.codes.has('h1-multiple')) content.push(...aiSeoHeadingDemotions(page.facts));
+    const guidance = [];
+    if (page.codes.has('orphan')) {
+        guidance.push('Link to this page from your navigation or a related page. Where it belongs is yours to choose.');
+    }
+    return {
+        screenId: page.screenId,
+        values,
+        content,
+        guidance
+    };
+}
+/** The user turn for one batch: each page's findings, what it shows now, and its text. */ export function aiSeoFixesPrompt(pages, titlesElsewhere) {
+    const titles = titlesElsewhere.length ? `Titles other pages use:\n${titlesElsewhere.slice(0, 40).map((title)=>`- ${title}`).join('\n')}\n\n` : '';
+    const blocks = pages.map((page)=>{
+        var _page_seo, _page_seo1, _page_facts_h1s_;
+        const lines = [
+            `### Page ${page.screenId}: ${page.name} (${page.path})`,
+            `Found: ${[
+                ...page.codes
+            ].map((code)=>AI_SEO_FINDING_LABELS[code]).join('; ')}`,
+            `Title now: ${str((_page_seo = page.seo) == null ? void 0 : _page_seo.title) || '(none)'}`,
+            `Description now: ${str((_page_seo1 = page.seo) == null ? void 0 : _page_seo1.description) || '(none)'}`,
+            `Main heading now: ${((_page_facts_h1s_ = page.facts.h1s[0]) == null ? void 0 : _page_facts_h1s_.text) || '(none)'}`
+        ];
+        if (page.keywords.length) lines.push(`Target keywords: ${page.keywords.join(', ')}`);
+        const images = aiSeoFixImages(page);
+        if (images.length) {
+            lines.push('Images without a description:', ...images.map((image)=>`- ${image.nodeId}: beside "${image.context || 'no text'}"`));
+        }
+        lines.push('Text:', page.facts.text.slice(0, AI_SEO_BATCH_PAGE_TEXT_CHARS) || '(no text)');
+        return lines.join('\n');
+    });
+    return `${titles}${blocks.join('\n\n')}`;
+}
+async function proposeFixes(context, report, unit) {
+    var _job_hostId;
+    const { job, firestore, signal } = context;
+    const model = modelOf(context);
+    const hostRef = firestore.collection('hosts').doc(job.hostId);
+    const reports = new Map(report.pages.map((page)=>[
+            page.screenId,
+            page
+        ]));
+    // Read now, not as audited: a page edited since keeps its edits, and one
+    // deleted since is skipped.
+    const loaded = await Promise.all(unit.screenIds.map(async (screenId)=>{
+        var _screen_seo;
+        const pageReport = reports.get(screenId);
+        const snapshot = await hostRef.collection('screens').doc(screenId).get();
+        const screen = snapshot.exists ? snapshot.data() : null;
+        if (!pageReport || !screen || screen.deletedAt) return null;
+        const version = await readVersionNodes(hostRef, screenId, str(screen.versionId));
+        return {
+            screenId,
+            path: pageReport.path,
+            name: pageReport.name,
+            codes: new Set(pageReport.findings.map((entry)=>entry.code)),
+            keywords: pageReport.keywords.map((entry)=>entry.keyword),
+            seo: (_screen_seo = screen.seo) != null ? _screen_seo : {},
+            facts: aiSeoPageFacts(version == null ? void 0 : version.nodes, {
+                rootId: version == null ? void 0 : version.rootId
+            })
+        };
+    }));
+    const pages = loaded.filter((page)=>page !== null);
+    const titlesElsewhere = await siteTitles(hostRef, new Set(pages.map((page)=>page.screenId)));
+    let spent = spentNothing(model);
+    let answers = {};
+    const notes = [];
+    if (pages.length) {
+        const generation = await runValidatedGeneration('seo-fixes', _extends({
+            step: 'job.seo',
+            model,
+            instructions: AI_SEO_FIXES_INSTRUCTIONS,
+            messages: [
+                {
+                    role: 'user',
+                    content: aiSeoFixesPrompt(pages, titlesElsewhere)
+                }
+            ],
+            tool: aiSeoFixesTool(pages.map((page)=>page.screenId)),
+            // Lowered on a tier too slow to answer a whole batch and ask again
+            // inside the least time the step registers.
+            maxTokens: aiSeoGenerationMaxTokens(model, AI_SEO_FIXES_MAX_TOKENS)
+        }, signal ? {
+            signal
+        } : {}, {
+            check: (answer)=>checkAiSeoFixes(answer, pages, titlesElsewhere)
+        }));
+        spent = spentBy(generation);
+        if (generation.status === 'ok') {
+            answers = generation.value;
+        } else {
+            notes.push(generation.status === 'refused' ? 'The AI declined to propose fixes for these pages.' : generation.message);
+        }
+    }
+    const fixes = pages.map((page)=>{
+        var _answers_page_screenId;
+        return aiSeoPageFixFor(page, (_answers_page_screenId = answers[page.screenId]) != null ? _answers_page_screenId : null);
+    });
+    const skipped = unit.screenIds.length - pages.length;
+    if (skipped > 0) {
+        notes.push(`${skipped} ${skipped === 1 ? 'page was' : 'pages were'} deleted since the audit and left out.`);
+    }
+    return {
+        output: {
+            resource: 'seo',
+            id: AI_SEO_OUTPUT_IDS.fixes(unit.batch),
+            hostId: (_job_hostId = job.hostId) != null ? _job_hostId : null,
+            label: `SEO audit · fixes ${unit.batch} of ${unit.of}`,
+            proposal: plain({
+                kind: 'fixes',
+                batch: unit.batch,
+                fixes,
+                notes
+            })
+        },
+        spent
+    };
+}
+async function runSiteAuditPass(context, host) {
+    var _job_outputs, _job_outputs1;
+    var _job_inputs;
+    const { job, firestore } = context;
+    const model = modelOf(context);
+    const recorded = aiSeoAuditView((_job_outputs = job.outputs) != null ? _job_outputs : []);
+    const outputs = [];
+    let spent = spentNothing(model);
+    let report;
+    let scan = null;
+    if (recorded) {
+        report = recorded.report;
+    } else {
+        var _ref, _ref1, _job_hostId;
+        var _job_inputs1, _host_seo, _host_seo1;
+        scan = await scanAiSeoSite(firestore, job.hostId, host, (_job_inputs1 = job.inputs) == null ? void 0 : _job_inputs1['keywords']);
+        report = aiSeoAudit(scan.pages, {
+            discouraged: isSearchDiscouraged(host),
+            entity: (_ref = (_host_seo = host.seo) == null ? void 0 : _host_seo.entity) != null ? _ref : {},
+            agent: (_ref1 = (_host_seo1 = host.seo) == null ? void 0 : _host_seo1.agent) != null ? _ref1 : {}
+        }, {
+            skipped: scan.skipped
+        });
+        report.notes.push(...scan.notes);
+        const findings = report.pages.reduce((sum, page)=>sum + page.findings.length, report.site.length);
+        outputs.push({
+            resource: 'seo',
+            id: AI_SEO_OUTPUT_IDS.report,
+            hostId: (_job_hostId = job.hostId) != null ? _job_hostId : null,
+            label: `SEO audit · ${report.pages.length} ${report.pages.length === 1 ? 'page' : 'pages'} · ${findings} ${findings === 1 ? 'finding' : 'findings'}`,
+            proposal: plain(report)
+        });
+    }
+    const units = aiSeoPendingUnits(report, [
+        ...(_job_outputs1 = job.outputs) != null ? _job_outputs1 : [],
+        ...outputs
+    ]);
+    if (!units.length) return _extends({
+        outputs
+    }, spent);
+    const [unit] = units;
+    const result = unit.kind === 'site' ? await proposeSite(context, host, scan != null ? scan : await scanAiSeoSite(firestore, job.hostId, host, (_job_inputs = job.inputs) == null ? void 0 : _job_inputs['keywords'])) : await proposeFixes(context, report, unit);
+    outputs.push(result.output);
+    spent = addSpent(spent, result.spent);
+    return _extends({
+        outputs
+    }, spent, units.length > 1 ? {
+        continue: true
+    } : {});
+}
+export const runAiJobSeoStep = async (context)=>{
+    const { job, firestore } = context;
+    if (!firestore) {
+        // The machine always hands one in; a runner called without it is a
+        // wiring fault, not a customer path.
+        throw new Error('the SEO step reads the site it writes about and was given no Firestore');
+    }
+    const target = aiSeoJobTarget(job.inputs);
+    if (!target) return _extends({
+        outputs: []
+    }, spentNothing(modelOf(context)), {
+        failure: AI_SEO_NO_TARGET_COPY
+    });
+    const site = await loadSite(firestore, job);
+    if ('failure' in site) return _extends({
+        outputs: []
+    }, spentNothing(modelOf(context)), {
+        failure: site.failure
+    });
+    const withFirestore = _extends({}, context, {
+        firestore
+    });
+    switch(target){
+        case 'screen':
+            return runScreenListing(withFirestore, site.host);
+        case 'product':
+            return runProductListing(withFirestore, site.host);
+        case 'site':
+            return runSiteAuditPass(withFirestore, site.host);
+    }
+};
+
+//# sourceMappingURL=ai-job-seo-step.js.map