feat: migrate npm registry from Verdaccio to Gitea Packages

packages/concept-engine/src/steps/00a-site-audit.ts

@@ -0,0 +1,65 @@
+// ============================================================================
+// Step 00a: Site Audit (DataForSEO + AI)
+// ============================================================================
+
+import { PageAuditor } from "@mintel/page-audit";
+import type { ConceptState, StepResult, PipelineConfig } from "../types.js";
+
+export async function executeSiteAudit(
+    state: ConceptState,
+    config: PipelineConfig,
+): Promise<StepResult> {
+    const startTime = Date.now();
+
+    if (!state.url) {
+        return {
+            success: true,
+            data: null,
+            usage: { step: "00a-site-audit", model: "none", promptTokens: 0, completionTokens: 0, cost: 0, durationMs: Date.now() - startTime },
+        };
+    }
+
+    try {
+        const login = process.env.DATA_FOR_SEO_LOGIN || process.env.DATA_FOR_SEO_API_KEY?.split(":")?.[0];
+        const password = process.env.DATA_FOR_SEO_PASSWORD || process.env.DATA_FOR_SEO_API_KEY?.split(":")?.slice(1)?.join(":");
+
+        if (!login || !password) {
+            console.warn("  ⚠️ Site Audit skipped: DataForSEO credentials missing from environment.");
+            return {
+                success: true,
+                data: null,
+                usage: { step: "00a-site-audit", model: "none", promptTokens: 0, completionTokens: 0, cost: 0, durationMs: Date.now() - startTime },
+            };
+        }
+
+        const auditor = new PageAuditor({
+            dataForSeoLogin: login,
+            dataForSeoPassword: password,
+            openrouterKey: config.openrouterKey,
+            outputDir: config.outputDir ? `${config.outputDir}/audits` : undefined,
+        });
+
+        // Run audit (max 20 pages for the estimation phase to keep it fast)
+        const result = await auditor.audit(state.url, { maxPages: 20 });
+
+        return {
+            success: true,
+            data: result,
+            usage: {
+                step: "00a-site-audit",
+                model: "dataforseo",
+                cost: 0, // DataForSEO cost tracking could be added later
+                promptTokens: 0,
+                completionTokens: 0,
+                durationMs: Date.now() - startTime,
+            },
+        };
+    } catch (err: any) {
+        console.warn(`  ⚠️ Site Audit failed, skipping: ${err.message}`);
+        return {
+            success: true,
+            data: null,
+            usage: { step: "00a-site-audit", model: "none", promptTokens: 0, completionTokens: 0, cost: 0, durationMs: Date.now() - startTime },
+        };
+    }
+}

packages/concept-engine/src/steps/00b-research.ts

@@ -0,0 +1,121 @@
+// ============================================================================
+// Step 00b: Research — Industry Research via @mintel/journaling (No LLM hallus)
+// Uses Serper API for real web search results about the industry/company.
+// ============================================================================
+
+import type { ConceptState, StepResult } from "../types.js";
+
+interface ResearchResult {
+    companyContext: string[];
+    industryInsights: string[];
+    competitorInfo: string[];
+}
+
+/**
+ * Research the company and industry using real web search data.
+ * Uses @mintel/journaling's ResearchAgent — results are grounded in real sources.
+ *
+ * NOTE: The journaling package can cause unhandled rejections that crash the process.
+ * We wrap each call in an additional safety layer.
+ */
+export async function executeResearch(
+    state: ConceptState,
+): Promise<StepResult<ResearchResult>> {
+    const startTime = Date.now();
+
+    const companyName = state.siteProfile?.companyInfo?.name || "";
+    const websiteTopic = state.siteProfile?.services?.slice(0, 3).join(", ") || "";
+    const domain = state.siteProfile?.domain || "";
+
+    if (!companyName && !websiteTopic && !domain) {
+        return {
+            success: true,
+            data: { companyContext: [], industryInsights: [], competitorInfo: [] },
+            usage: { step: "00b-research", model: "none", promptTokens: 0, completionTokens: 0, cost: 0, durationMs: 0 },
+        };
+    }
+
+    // Safety wrapper: catch ANY unhandled rejections during this step
+    const safeCall = <T>(fn: () => Promise<T>, fallback: T): Promise<T> => {
+        return new Promise<T>((resolve) => {
+            const handler = (err: any) => {
+                console.warn(`  ⚠️ Unhandled rejection caught in research: ${err?.message || err}`);
+                process.removeListener("unhandledRejection", handler);
+                resolve(fallback);
+            };
+            process.on("unhandledRejection", handler);
+
+            fn()
+                .then((result) => {
+                    process.removeListener("unhandledRejection", handler);
+                    resolve(result);
+                })
+                .catch((err) => {
+                    process.removeListener("unhandledRejection", handler);
+                    console.warn(`  ⚠️ Research call failed: ${err?.message || err}`);
+                    resolve(fallback);
+                });
+        });
+    };
+
+    try {
+        const { ResearchAgent } = await import("@mintel/journaling");
+        const agent = new ResearchAgent(process.env.OPENROUTER_API_KEY || "");
+
+        const results: ResearchResult = {
+            companyContext: [],
+            industryInsights: [],
+            competitorInfo: [],
+        };
+
+        // 1. Research the company itself
+        if (companyName || domain) {
+            const searchQuery = companyName
+                ? `${companyName} ${websiteTopic} Unternehmen`
+                : `site:${domain}`;
+
+            console.log(`  🔍 Researching: "${searchQuery}"...`);
+            const facts = await safeCall(
+                () => agent.researchTopic(searchQuery),
+                [] as any[],
+            );
+            results.companyContext = (facts || [])
+                .filter((f: any) => f?.fact || f?.value || f?.text || f?.statement)
+                .map((f: any) => f.fact || f.value || f.text || f.statement)
+                .slice(0, 5);
+        }
+
+        // 2. Industry research
+        if (websiteTopic) {
+            console.log(`  🔍 Researching industry: "${websiteTopic}"...`);
+            const insights = await safeCall(
+                () => agent.researchCompetitors(websiteTopic),
+                [] as any[],
+            );
+            results.industryInsights = (insights || []).slice(0, 5);
+        }
+
+        const totalFacts = results.companyContext.length + results.industryInsights.length + results.competitorInfo.length;
+        console.log(`  📊 Research found ${totalFacts} data points.`);
+
+        return {
+            success: true,
+            data: results,
+            usage: {
+                step: "00b-research",
+                model: "serper/datacommons",
+                promptTokens: 0,
+                completionTokens: 0,
+                cost: 0,
+                durationMs: Date.now() - startTime,
+            },
+        };
+    } catch (err) {
+        console.warn(`  ⚠️ Research step skipped: ${(err as Error).message}`);
+        return {
+            success: true,
+            data: { companyContext: [], industryInsights: [], competitorInfo: [] },
+            usage: { step: "00b-research", model: "none", promptTokens: 0, completionTokens: 0, cost: 0, durationMs: Date.now() - startTime },
+        };
+    }
+}

packages/concept-engine/src/steps/01-extract.ts

@@ -0,0 +1,108 @@
+// ============================================================================
+// Step 01: Extract — Briefing Fact Extraction (Gemini Flash)
+// ============================================================================
+
+import { llmJsonRequest } from "../llm-client.js";
+import type { ConceptState, StepResult, PipelineConfig } from "../types.js";
+import { DEFAULT_MODELS } from "../types.js";
+
+export async function executeExtract(
+    state: ConceptState,
+    config: PipelineConfig,
+): Promise<StepResult> {
+    const models = { ...DEFAULT_MODELS, ...config.modelsOverride };
+    const startTime = Date.now();
+
+    // Build site context from the deterministic analyzer
+    const siteContext = state.siteProfile
+        ? `
+EXISTING WEBSITE ANALYSIS (FACTS — verifiably crawled, NOT guessed):
+- Domain: ${state.siteProfile.domain}
+- Total pages crawled: ${state.siteProfile.totalPages}
+- Navigation items: ${state.siteProfile.navigation.map((n) => n.label).join(", ") || "nicht erkannt"}
+- Existing features: ${state.siteProfile.existingFeatures.join(", ") || "keine"}
+- Services / Kompetenzen: ${state.siteProfile.services.join(" | ") || "keine"}
+- Employee count (from website text): ${(state.siteProfile as any).employeeCount || "nicht genannt"}
+- Company name: ${state.siteProfile.companyInfo.name || "unbekannt"}
+- Address: ${state.siteProfile.companyInfo.address || "unbekannt"}
+- Tax ID (USt-ID): ${state.siteProfile.companyInfo.taxId || "unbekannt"}
+- HRB: ${state.siteProfile.companyInfo.registerNumber || "unbekannt"}
+- Managing Director: ${state.siteProfile.companyInfo.managingDirector || "unbekannt"}
+- External related domains (HAVE OWN WEBSITES — DO NOT include as sub-pages!): ${state.siteProfile.externalDomains.join(", ") || "keine"}
+- Social links: ${Object.entries(state.siteProfile.socialLinks).map(([k, v]) => `${k}: ${v}`).join(", ") || "keine"}
+`
+        : "No existing website data available.";
+
+    const systemPrompt = `
+You are a precision fact extractor. Your only job: extract verifiable facts from the BRIEFING.
+Output language: GERMAN (strict).
+Output format: flat JSON at root level. No nesting except arrays.
+
+### CRITICAL RULES:
+1. "employeeCount": take from SITE ANALYSIS if available. Only override if briefing states something more specific.
+2. External domains (e.g. "etib-ing.com") have their OWN website. NEVER include them as sub-pages.
+3. Videos (Messefilm, Imagefilm) are CONTENT ASSETS, not pages.
+4. If existing site already has search, include "search" in functions.
+5. DO NOT invent pages not mentioned in briefing or existing navigation.
+
+### CONSERVATIVE RULE:
+- simple lists (Jobs, Referenzen, Messen) = pages, NOT features
+- Assume "page" as default. Only add "feature" for complex interactive systems.
+
+### OUTPUT FORMAT:
+{
+  "companyName": string,
+  "companyAddress": string,
+  "personName": string,
+  "email": string,
+  "existingWebsite": string,
+  "websiteTopic": string,         // MAX 3 words
+  "isRelaunch": boolean,
+  "employeeCount": string,        // from site analysis, e.g. "über 50"
+  "pages": string[],              // ALL pages: ["Startseite", "Über Uns", "Leistungen", ...]
+  "functions": string[],          // search, forms, maps, video, cookie_consent, etc.
+  "assets": string[],             // existing_website, logo, media, photos, videos
+  "deadline": string,
+  "targetAudience": string,
+  "cmsSetup": boolean,
+  "multilang": boolean
+}
+
+BANNED OUTPUT KEYS: "selectedPages", "otherPages", "features", "apiSystems" — use pages[] and functions[] ONLY.
+`;
+
+    const userPrompt = `BRIEFING (TRUTH SOURCE):
+${state.briefing}
+
+COMMENTS:
+${state.comments || "keine"}
+
+${siteContext}`;
+
+    try {
+        const { data, usage } = await llmJsonRequest({
+            model: models.flash,
+            systemPrompt,
+            userPrompt,
+            apiKey: config.openrouterKey,
+        });
+
+        return {
+            success: true,
+            data,
+            usage: {
+                step: "01-extract",
+                model: models.flash,
+                promptTokens: usage.promptTokens,
+                completionTokens: usage.completionTokens,
+                cost: usage.cost,
+                durationMs: Date.now() - startTime,
+            },
+        };
+    } catch (err) {
+        return {
+            success: false,
+            error: `Extract step failed: ${(err as Error).message}`,
+        };
+    }
+}

packages/concept-engine/src/steps/02-audit.ts

@@ -0,0 +1,110 @@
+// ============================================================================
+// Step 02: Audit — Feature Auditor + Skeptical Review (Gemini Flash)
+// ============================================================================
+
+import { llmJsonRequest } from "../llm-client.js";
+import type { ConceptState, StepResult, PipelineConfig } from "../types.js";
+import { DEFAULT_MODELS } from "../types.js";
+
+export async function executeAudit(
+    state: ConceptState,
+    config: PipelineConfig,
+): Promise<StepResult> {
+    const models = { ...DEFAULT_MODELS, ...config.modelsOverride };
+    const startTime = Date.now();
+
+    if (!state.facts) {
+        return { success: false, error: "No facts from Step 01 available." };
+    }
+
+    const systemPrompt = `
+You are a "Strict Cost Controller". Your mission is to prevent over-billing.
+Review the extracted FEATURES against the BRIEFING and the EXISTING SITE ANALYSIS.
+
+### RULE OF THUMB:
+- A "Feature" (1.500 €) is ONLY justified for complex, dynamic systems (logic, database, CMS-driven management, advanced filtering).
+- Simple lists, information sections, or static descriptions (e.g., "Messen", "Team", "Historie", "Jobs" as mere text) are ALWAYS "Pages" (600 €).
+- If the briefing doesn't explicitly mention "Management System", "Filterable Database", or "Client Login", it is a PAGE.
+
+### ADDITIONAL CHECKS:
+1. If any feature maps to an entity that has its own external website (listed in EXTERNAL_DOMAINS), remove it entirely — it's out of scope.
+2. Videos are ASSETS not pages. Remove any video-related entries from pages.
+3. If the existing site has features (search, forms, etc.), ensure they are in the functions list.
+
+### MISSION:
+Return the corrected 'features', 'otherPages', and 'functions' arrays.
+
+### OUTPUT FORMAT:
+{
+  "features": string[],
+  "otherPages": string[],
+  "functions": string[],
+  "removedItems": [{ "item": string, "reason": string }],
+  "addedItems": [{ "item": string, "reason": string }]
+}
+`;
+
+    const userPrompt = `
+EXTRACTED FACTS:
+${JSON.stringify(state.facts, null, 2)}
+
+BRIEFING:
+${state.briefing}
+
+EXTERNAL DOMAINS (have own websites, OUT OF SCOPE):
+${state.siteProfile?.externalDomains?.join(", ") || "none"}
+
+EXISTING FEATURES ON CURRENT SITE:
+${state.siteProfile?.existingFeatures?.join(", ") || "none"}
+`;
+
+    try {
+        const { data, usage } = await llmJsonRequest({
+            model: models.flash,
+            systemPrompt,
+            userPrompt,
+            apiKey: config.openrouterKey,
+        });
+
+        // Apply audit results to facts
+        const auditedFacts = { ...state.facts };
+        auditedFacts.features = data.features || [];
+        auditedFacts.otherPages = [
+            ...new Set([...(auditedFacts.otherPages || []), ...(data.otherPages || [])]),
+        ];
+        if (data.functions) {
+            auditedFacts.functions = [
+                ...new Set([...(auditedFacts.functions || []), ...data.functions]),
+            ];
+        }
+
+        // Log changes
+        if (data.removedItems?.length) {
+            console.log("  📉 Audit removed:");
+            for (const item of data.removedItems) {
+                console.log(`     - ${item.item}: ${item.reason}`);
+            }
+        }
+        if (data.addedItems?.length) {
+            console.log("  📈 Audit added:");
+            for (const item of data.addedItems) {
+                console.log(`     + ${item.item}: ${item.reason}`);
+            }
+        }
+
+        return {
+            success: true,
+            data: auditedFacts,
+            usage: {
+                step: "02-audit",
+                model: models.flash,
+                promptTokens: usage.promptTokens,
+                completionTokens: usage.completionTokens,
+                cost: usage.cost,
+                durationMs: Date.now() - startTime,
+            },
+        };
+    } catch (err) {
+        return { success: false, error: `Audit step failed: ${(err as Error).message}` };
+    }
+}

packages/concept-engine/src/steps/03-strategize.ts

@@ -0,0 +1,99 @@
+// ============================================================================
+// Step 03: Strategize — Briefing Summary + Design Vision (Gemini Pro)
+// ============================================================================
+
+import { llmJsonRequest } from "../llm-client.js";
+import type { ConceptState, StepResult, PipelineConfig } from "../types.js";
+import { DEFAULT_MODELS } from "../types.js";
+
+export async function executeStrategize(
+    state: ConceptState,
+    config: PipelineConfig,
+): Promise<StepResult> {
+    const models = { ...DEFAULT_MODELS, ...config.modelsOverride };
+    const startTime = Date.now();
+
+    if (!state.auditedFacts) {
+        return { success: false, error: "No audited facts from Step 02 available." };
+    }
+
+    const systemPrompt = `
+You are a high-end Digital Architect. Your goal is to make the CUSTOMER feel 100% understood.
+Analyze the BRIEFING and the EXISTING WEBSITE context.
+
+### OBJECTIVE:
+1. **briefingSummary**: Ein sachlicher, tiefgehender Überblick der Unternehmenslage.
+   - STIL: Keine Ich-Form. Keine Marketing-Floskeln. Nutze präzise Fachbegriffe. Sei prägnant.
+   - FORM: EXAKT ZWEI ABSÄTZE. Insgesamt ca. 6 Sätze.
+   - INHALT: Status Quo, was der Kunde will, welcher Sprung notwendig ist.
+   - ABSOLUTE REGEL: Keine Halluzinationen. Keine namentlichen Nennungen von Personen.
+   - RELAUNCH-REGEL: Wenn isRelaunch=true, NICHT sagen "keine digitale Präsenz". Es GIBT eine Seite.
+   - SORGLOS BETRIEB: MUSS erwähnt werden als Teil des Gesamtpakets.
+
+2. **designVision**: Ein abstraktes, strategisches Konzept.
+   - STIL: Rein konzeptionell. Keine Umsetzungsschritte. Keine Ich-Form. Sei prägnant.
+   - FORM: EXAKT ZWEI ABSÄTZE. Insgesamt ca. 4 Sätze.
+   - DATENSCHUTZ: KEINERLEI namentliche Nennungen.
+   - FOKUS: Welche strategische Wirkung soll erzielt werden?
+
+### RULES:
+- NO "wir/unser". NO "Ich/Mein". Objective, fact-oriented narrative.
+- NO marketing lingo. NO "innovativ", "revolutionär", "state-of-the-art".
+- NO hallucinations about features not in the briefing.
+- NO "SEO-Standards zur Fachkräftesicherung" or "B2B-Nutzerströme" — das ist Schwachsinn.
+  Use specific industry terms from the briefing (e.g. "Kabeltiefbau", "HDD-Bohrverfahren").
+- LANGUAGE: Professional German. Simple but expert-level.
+
+### OUTPUT FORMAT:
+{
+  "briefingSummary": string,
+  "designVision": string
+}
+`;
+
+    const userPrompt = `
+BRIEFING (TRUTH SOURCE):
+${state.briefing}
+
+EXISTING WEBSITE DATA:
+- Services: ${state.siteProfile?.services?.join(", ") || "unbekannt"}
+- Navigation: ${state.siteProfile?.navigation?.map((n) => n.label).join(", ") || "unbekannt"}
+- Company: ${state.auditedFacts.companyName || "unbekannt"}
+
+EXTRACTED & AUDITED FACTS:
+${JSON.stringify(state.auditedFacts, null, 2)}
+
+${state.siteAudit?.report ? `
+TECHNICAL SITE AUDIT (IST-Analyse):
+Health: ${state.siteAudit.report.overallHealth} (SEO: ${state.siteAudit.report.seoScore}, UX: ${state.siteAudit.report.uxScore}, Perf: ${state.siteAudit.report.performanceScore})
+- Executive Summary: ${state.siteAudit.report.executiveSummary}
+- Strengths: ${state.siteAudit.report.strengths.join(", ")}
+- Critical Issues: ${state.siteAudit.report.criticalIssues.join(", ")}
+- Quick Wins: ${state.siteAudit.report.quickWins.join(", ")}
+` : ""}
+`;
+
+    try {
+        const { data, usage } = await llmJsonRequest({
+            model: models.pro,
+            systemPrompt,
+            userPrompt,
+            apiKey: config.openrouterKey,
+        });
+
+        return {
+            success: true,
+            data,
+            usage: {
+                step: "03-strategize",
+                model: models.pro,
+                promptTokens: usage.promptTokens,
+                completionTokens: usage.completionTokens,
+                cost: usage.cost,
+                durationMs: Date.now() - startTime,
+            },
+        };
+    } catch (err) {
+        return { success: false, error: `Strategize step failed: ${(err as Error).message}` };
+    }
+}

packages/concept-engine/src/steps/04-architect.ts

@@ -0,0 +1,133 @@
+// ============================================================================
+// Step 04: Architect — Sitemap & Information Architecture (Gemini Pro)
+// ============================================================================
+
+import { llmJsonRequest } from "../llm-client.js";
+import type { ConceptState, StepResult, PipelineConfig } from "../types.js";
+import { DEFAULT_MODELS } from "../types.js";
+
+export async function executeArchitect(
+    state: ConceptState,
+    config: PipelineConfig,
+): Promise<StepResult> {
+    const models = { ...DEFAULT_MODELS, ...config.modelsOverride };
+    const startTime = Date.now();
+
+    if (!state.auditedFacts) {
+        return { success: false, error: "No audited facts available." };
+    }
+
+    // Build navigation constraint from the real site
+    const existingNav = state.siteProfile?.navigation?.map((n) => n.label).join(", ") || "unbekannt";
+    const existingServices = state.siteProfile?.services?.join(", ") || "unbekannt";
+    const externalDomains = state.siteProfile?.externalDomains?.join(", ") || "keine";
+
+    const systemPrompt = `
+Du bist ein Senior UX Architekt. Erstelle einen ECHTEN SEITENBAUM für die neue Website.
+Regelwerk für den Output:
+
+### SEITENBAUM-REGELN:
+1. KEIN MARKETINGSPRECH als Kategoriename. Gültige Kategorien sind nur die echten Navigationspunkte der Website.
+   ERLAUBT: "Startseite", "Leistungen", "Über uns", "Karriere", "Referenzen", "Kontakt", "Rechtliches"
+   VERBOTEN: "Kern-Präsenz", "Vertrauen", "Business Areas", "Digitaler Auftritt"
+
+2. LEISTUNGEN muss in ECHTE UNTERSEITEN aufgeteilt werden — nicht eine einzige "Leistungen"-Seite.
+   Jede Kompetenz aus dem existierenden Leistungsspektrum = eine eigene Seite.
+   Beispiel statt:
+     { category: "Leistungen", pages: [{ title: "Leistungen", desc: "..." }] }
+   So:
+     { category: "Leistungen", pages: [
+       { title: "Kabeltiefbau", desc: "Mittelspannung, Niederspannung, Kabelpflugarbeiten..." },
+       { title: "Horizontalspülbohrungen", desc: "HDD in allen Bodenklassen..." },
+       { title: "Elektromontagen", desc: "Bis 110 kV, Glasfaserkabelmontagen..." },
+       { title: "Planung & Dokumentation", desc: "Genehmigungs- und Ausführungsplanung, Vermessung..." }
+     ]}
+
+3. SEITENTITEL: Kurz, klar, faktisch. Kein Werbejargon.
+   ERLAUBT: "Kabeltiefbau", "Über uns", "Karriere"
+   VERBOTEN: "Unsere Expertise", "Kompetenzspektrum", "Community"
+
+4. Gruppe die Leistungen nach dem ECHTEN Kompetenzkatalog der bestehenden Site — nicht erfinden.
+
+5. Keine doppelten Seiten. Keine Phantomseiten.
+
+6. Videos = Content-Assets, keine eigene Seite.
+
+7. Entitäten mit eigener Domain (${externalDomains}) = NICHT als Seite. Nur als Teaser/Link wenn nötig.
+
+### KONTEXT:
+Bestehende Navigation: ${existingNav}
+Bestehende Services: ${existingServices}
+Externe Domains (haben eigene Website): ${externalDomains}
+Angeforderte zusätzliche Seiten aus Briefing: ${(state.auditedFacts as any)?.pages?.join(", ") || "keine spezifischen"}
+
+### OUTPUT FORMAT (JSON):
+{
+  "websiteTopic": string,  // MAX 3 Wörter, beschreibend
+  "sitemap": [
+    {
+      "category": string,   // Echter Nav-Eintrag. KEIN Marketingsprech.
+      "pages": [
+        { "title": string, "desc": string }  // Echte Unterseite, 1-2 Sätze Zweck
+      ]
+    }
+  ]
+}
+`;
+
+    const userPrompt = `
+BRIEFING:
+${state.briefing}
+
+FAKTEN (aus Extraktion):
+${JSON.stringify({ facts: state.auditedFacts, strategy: { briefingSummary: state.briefingSummary } }, null, 2)}
+
+Erstelle den Seitenbaum. Baue die Leistungen DETAILLIERT aus — echte Unterseiten pro Kompetenzbereich.
+`;
+
+    try {
+        const { data, usage } = await llmJsonRequest({
+            model: models.pro,
+            systemPrompt,
+            userPrompt,
+            apiKey: config.openrouterKey,
+        });
+
+        // Normalize sitemap structure
+        let sitemap = data.sitemap;
+        if (sitemap && !Array.isArray(sitemap)) {
+            if (sitemap.categories) sitemap = sitemap.categories;
+            else {
+                const entries = Object.entries(sitemap);
+                if (entries.every(([, v]) => Array.isArray(v))) {
+                    sitemap = entries.map(([category, pages]) => ({ category, pages }));
+                }
+            }
+        }
+
+        if (Array.isArray(sitemap)) {
+            sitemap = sitemap.map((cat: any) => ({
+                category: cat.category || cat.kategorie || cat.Kategorie || "Allgemein",
+                pages: (cat.pages || cat.seiten || []).map((page: any) => ({
+                    title: page.title || page.titel || "Seite",
+                    desc: page.desc || page.beschreibung || page.description || "",
+                })),
+            }));
+        }
+
+        return {
+            success: true,
+            data: { websiteTopic: data.websiteTopic, sitemap },
+            usage: {
+                step: "04-architect",
+                model: models.pro,
+                promptTokens: usage.promptTokens,
+                completionTokens: usage.completionTokens,
+                cost: usage.cost,
+                durationMs: Date.now() - startTime,
+            },
+        };
+    } catch (err) {
+        return { success: false, error: `Architect step failed: ${(err as Error).message}` };
+    }
+}