feat(ms-ai-architect): add plugin to open marketplace (v1.5.0 baseline)

Initial addition of ms-ai-architect plugin to the open-source marketplace. Private content excluded: orchestrator/ (Linear tooling), docs/utredning/ (client investigation), generated test reports and PDF export script. skill-gen tooling moved from orchestrator/ to scripts/skill-gen/. Security scan: WARNING (risk 20/100) — no secrets, no injection found. False positive fixed: added gitleaks:allow to Python variable reference in output-validation-grounding-verification.md line 109. Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
2026-04-07 17:17:17 +02:00 · 2026-04-07 17:17:17 +02:00 · 6a7632146e
commit 6a7632146e
parent a8d79e4484
490 changed files with 213249 additions and 2 deletions
--- a/plugins/ms-ai-architect/tests/capture-fixture.sh
+++ b/plugins/ms-ai-architect/tests/capture-fixture.sh
@ -0,0 +1,41 @@
+#!/bin/bash
+# capture-fixture.sh — Extract agent output fixture from a document
+# Usage: bash tests/capture-fixture.sh <source-file> <section-header> <output-dir>
+#
+# Example:
+#   bash tests/capture-fixture.sh docs/utredning/.../utredning.md "S5: Sikkerhetsvurdering" tests/fixtures/security-assessment/
+set -euo pipefail
+
+if [ $# -lt 3 ]; then
+  echo "Usage: $0 <source-file> <section-header> <output-dir>"
+  echo ""
+  echo "Extracts a section from source file and saves as fixture.md"
+  echo ""
+  echo "Example:"
+  echo "  $0 docs/utredning/file.md 'S5: Sikkerhetsvurdering' tests/fixtures/security-assessment/"
+  exit 1
+fi
+
+SOURCE="$1"
+HEADER="$2"
+OUTPUT_DIR="$3"
+
+if [ ! -f "$SOURCE" ]; then
+  echo "Error: Source file not found: $SOURCE"
+  exit 1
+fi
+
+mkdir -p "$OUTPUT_DIR"
+
+# Extract section from ## header to next ## header (or EOF)
+awk -v header="$HEADER" '
+  BEGIN { found=0 }
+  /^## / {
+    if (found) exit
+    if (index($0, header)) found=1
+  }
+  found { print }
+' "$SOURCE" > "$OUTPUT_DIR/fixture.md"
+
+LINES=$(wc -l < "$OUTPUT_DIR/fixture.md" | tr -d ' ')
+echo "Captured $LINES lines to $OUTPUT_DIR/fixture.md"
--- a/plugins/ms-ai-architect/tests/fixtures/ai-act/fixture-high-risk.md
+++ b/plugins/ms-ai-architect/tests/fixtures/ai-act/fixture-high-risk.md
@ -0,0 +1,93 @@
+## EU AI Act — Vurdering: AutomatiskSaksbehandler
+
+**Dato:** 2026-02-22
+**Vurdert av:** AI Act Assessor
+**Organisasjon:** Statens vegvesen
+
+### 1. Risikoklassifisering
+
+| Attributt | Verdi |
+|-----------|-------|
+| **Risikonivå** | Høyrisiko |
+| **Annex III-kategori** | Kategori 5: Tilgang til og bruk av essensielle offentlige tjenester og ytelser |
+| **GPAI-status** | Ja — basert på GPT-4o via Azure OpenAI |
+| **Klassifiseringsgrunnlag** | Systemet automatiserer vurdering av helsekrav ved søknad om førerkort (klasse B). Direkte påvirkning på borgeres rett til førerkort — en essensiell offentlig tjeneste. |
+| **Konfidens** | Høy |
+
+#### Steg 1: Forbudt-sjekk (Art. 5)
+Ingen forbudte praksiser identifisert. Systemet scorer ikke individer sosialt, og beslutninger kan overprøves av saksbehandler.
+
+#### Steg 2: Annex III høyrisiko-sjekk
+**Treffer kategori 5 (a):** AI-systemer som brukes av offentlige myndigheter for å vurdere berettigelse til offentlige ytelser og tjenester, inkludert tildelingsbeslutninger.
+
+Førerkort er en essensiell offentlig tjeneste i norsk kontekst. Automatisert vurdering av helsekrav påvirker direkte borgeres tilgang til denne tjenesten.
+
+**Grensevurdering:** Det er ingen tvil om at dette er høyrisiko. Systemet tar beslutninger som direkte påvirker enkeltpersoners rettigheter og muligheter.
+
+#### Steg 3: GPAI-sjekk
+GPT-4o er en GPAI-modell. Microsoft er provider av grunnmodellen. Statens vegvesen er deployer av det tilpassede systemet. Ettersom systemet har vesentlig tilpasning (fine-tuning på norske helseattest-vurderinger), kan Statens vegvesen også anses som provider av det tilpassede høyrisiko-systemet.
+
+#### Steg 4: Ikke relevant — allerede klassifisert som høyrisiko
+
+### 2. Rolle
+
+| Attributt | Verdi |
+|-----------|-------|
+| **Organisasjonens rolle** | Deployer (primært) + mulig Provider (ved vesentlig tilpasning) |
+| **Begrunnelse** | Som deployer har Statens vegvesen alle Art. 26-27 forpliktelser. Ved fine-tuning av modellen kan organisasjonen også få provider-forpliktelser for det tilpassede systemet (Art. 25). |
+| **Provider (grunnmodell)** | Microsoft (Azure OpenAI Service) |
+
+### 3. Forpliktelser
+
+| # | Artikkel | Krav | Status | Gap |
+|---|----------|------|--------|-----|
+| 1 | Art. 26(1) | Bruk i samsvar med bruksanvisning | Delvis | Bruksanvisning fra Microsoft, men ikke tilpasset norsk kontekst |
+| 2 | Art. 26(2) | Menneskelig tilsyn (effektiv kontroll) | Delvis | Saksbehandler kan overprøve, men prosedyre ikke formalisert |
+| 3 | Art. 26(5) | FRIA gjennomført for offentlig sektor | Ikke oppfylt | Ingen FRIA utført |
+| 4 | Art. 26(6) | Loggoppbevaring minimum 6 måneder | Ikke oppfylt | Logger settes til 90 dager i Application Insights |
+| 5 | Art. 27 | FRIA for offentlig myndighet-deployer | Ikke oppfylt | Obligatorisk — ikke startet |
+| 6 | Art. 13 | Transparens: bruksinstruksjon tilgjengelig | Ikke oppfylt | Ingen Art. 13-dokumentasjon |
+| 7 | Art. 14 | Menneskelig tilsyn: override-mekanismer | Delvis | Override mulig men ikke systematisk |
+| 8 | Art. 50(1) | Informer personer om AI-bruk | Ikke oppfylt | Borgere informeres ikke om at AI vurderer helseattester |
+| 9 | Art. 9 | Risikostyringssystem | Ikke oppfylt | Ingen formell risikostyring for AI-systemet |
+| 10 | Art. 12 | Automatisk loggføring | Delvis | Logger finnes men retention er for kort |
+
+### 4. Tiltaksplan
+
+| # | Tiltak | Prioritet | Frist | Ansvarlig |
+|---|--------|-----------|-------|-----------|
+| T1 | Gjennomfør FRIA (Art. 27) — bruk `/architect:frimpact` | Kritisk | 2026-05-01 | Personvernombud + AI-rådgiver |
+| T2 | Etabler risikostyringssystem (Art. 9) | Kritisk | 2026-06-01 | Seksjonsleder |
+| T3 | Øk log retention til minimum 6 måneder (Art. 12/26) | Kritisk | 2026-04-01 | IT-drift |
+| T4 | Utvikle transparensnotis til borgere (Art. 50) | Høy | 2026-05-01 | Kommunikasjonsavdeling |
+| T5 | Formalisér override-prosedyre for saksbehandlere (Art. 14) | Høy | 2026-05-15 | Fagleder |
+| T6 | Gjennomfør DPIA (GDPR Art. 35) — overlapper med FRIA | Høy | 2026-05-01 | Personvernombud |
+| T7 | Utarbeid Art. 13 bruksinstruksjon | Middels | 2026-06-15 | AI-rådgiver |
+| T8 | Forbered samsvarsvurdering (Annex IV, Art. 43) | Middels | 2026-07-01 | Kvalitetsansvarlig |
+| T9 | Vurdér behov for ekstern samsvarsvurdering | Lav | 2026-07-15 | Juridisk avdeling |
+
+### 5. Neste steg
+
+1. **Umiddelbart:** `/architect:frimpact` — FRIA er obligatorisk og bør prioriteres høyest
+2. **Innen 30 dager:** `/architect:dpia` — Personvernkonsekvensanalyse (utdyper personverndimensjonen)
+3. **Innen 60 dager:** `/architect:ros` — ROS-analyse med AI Act-dimensjon (dimensjon 6)
+4. **Innen 90 dager:** `/architect:conformity` — Start samsvarsvurdering
+5. **Dokumentér:** `/architect:adr` — Dokumenter klassifiseringsbeslutningen
+
+### Viktige frister
+
+| Frist | Krav | Relevans |
+|-------|------|----------|
+| 2025-02-02 | Forbudte AI-praksiser (Art. 5) | Gjelder ikke |
+| 2025-08-02 | Governance og sanksjoner (Art. 99) | Gjelder — governance-struktur kreves |
+| 2026-08-02 | GPAI-krav + Annex III høyrisiko | **Gjelder direkte — 161 dager** |
+| 2027-08-02 | Alle høyrisiko-krav (full compliance) | Gjelder — full Art. 9-27 compliance |
+
+### Referanser
+- `ai-act-classification-methodology.md` — Klassifiseringsmetodikk
+- `ai-act-annex-iii-checklist.md` — Annex III kategori 5 vurdering
+- `ai-act-deployer-obligations.md` — Art. 26-27 forpliktelser
+- `ai-act-fria-template.md` — FRIA-mal referanse
+- `ai-act-provider-obligations.md` — Art. 9-15 (ved provider-status)
+- `ai-act-compliance-guide.md` — Generell veileder
+- Microsoft Learn: Azure OpenAI responsible AI practices
--- a/plugins/ms-ai-architect/tests/fixtures/ai-act/fixture.md
+++ b/plugins/ms-ai-architect/tests/fixtures/ai-act/fixture.md
@ -0,0 +1,79 @@
+## EU AI Act — Vurdering: FartsPrediksjonsagent
+
+**Dato:** 2026-02-22
+**Vurdert av:** AI Act Assessor
+**Organisasjon:** Statens vegvesen
+
+### 1. Risikoklassifisering
+
+| Attributt | Verdi |
+|-----------|-------|
+| **Risikonivå** | Minimal risiko |
+| **Annex III-kategori** | Ikke Annex III |
+| **GPAI-status** | Ja — basert på GPT-4o, men ikke systemisk risiko |
+| **Klassifiseringsgrunnlag** | Systemet predikerer gjennomsnittsfart på vegstrekninger basert på historiske trafikkdata. Ingen direkte påvirkning på individer, ingen biometrisk identifikasjon, ikke kritisk infrastrukturstyring. |
+| **Konfidens** | Høy |
+
+#### Steg 1: Forbudt-sjekk (Art. 5)
+Ingen av de forbudte praksisene er relevante. Systemet scorer ikke individer sosialt, manipulerer ikke sårbare grupper, og bruker ikke biometrisk fjernidentifisering.
+
+#### Steg 2: Annex III høyrisiko-sjekk
+Systemet treffer ingen av de 8 Annex III-kategoriene:
+- Ikke biometrisk identifisering
+- Ikke styring av kritisk infrastruktur (predikerer, styrer ikke)
+- Ikke utdanning/opplæring
+- Ikke ansettelse/personal
+- Ikke essensielle offentlige tjenester
+- Ikke rettshåndhevelse
+- Ikke migrasjon/grensekontroll
+- Ikke rettsforvaltning
+
+#### Steg 3: GPAI-sjekk
+Systemet bruker Azure OpenAI GPT-4o som grunnmodell. GPT-4o er en GPAI-modell, men FartsPrediksjonsagent er en downstream-applikasjon — provider-forpliktelser for GPAI hviler på Microsoft som modell-provider.
+
+#### Steg 4: Begrenset/Minimal
+Systemet har ingen direkte brukerinteraksjon med borgere. Resultater vises kun til trafikkplanleggere internt. Klassifiseres som **minimal risiko**.
+
+### 2. Rolle
+
+| Attributt | Verdi |
+|-----------|-------|
+| **Organisasjonens rolle** | Deployer |
+| **Begrunnelse** | Statens vegvesen bruker et AI-system utviklet internt med Azure OpenAI. Ettersom systemet ikke markedsføres til andre, og bruker standard Azure-tjenester uten vesentlig tilpasning av modellen, er rollen deployer. |
+| **Provider (ekstern)** | Microsoft (Azure OpenAI Service) |
+
+### 3. Forpliktelser
+
+| # | Artikkel | Krav | Status | Gap |
+|---|----------|------|--------|-----|
+| 1 | Art. 50(1) | Transparens: informer brukere om AI-bruk | Oppfylt | Interne brukere informert |
+| 2 | Art. 4 | AI-kompetanse: sikre tilstrekkelig kunnskap | Delvis | Opplæringsplan ikke formalisert |
+| 3 | Frivillig | Code of Conduct (Art. 95) | Ikke startet | Anbefales men ikke påkrevd |
+
+### 4. Tiltaksplan
+
+| # | Tiltak | Prioritet | Frist | Ansvarlig |
+|---|--------|-----------|-------|-----------|
+| T1 | Formalisér AI-kompetanseplan for trafikkplanleggere | Lav | 2026-12-31 | Seksjonsleder |
+| T2 | Vurdér frivillig Code of Conduct-tilslutning | Lav | 2027-06-30 | AI-rådgiver |
+
+### 5. Neste steg
+
+1. Ingen regulatoriske blokkeringer — systemet kan brukes uten ytterligere tiltak
+2. Anbefaler `/architect:ros` for generell risikovurdering (god praksis)
+3. Vurdér `/architect:transparency` for å generere intern AI-bruksnotis
+
+### Viktige frister
+
+| Frist | Krav | Relevans |
+|-------|------|----------|
+| 2025-02-02 | Forbudte AI-praksiser (Art. 5) | Gjelder ikke |
+| 2025-08-02 | Governance og sanksjoner (Art. 99) | Gjelder ikke direkte |
+| 2026-08-02 | GPAI-krav + Annex III høyrisiko | Gjelder ikke (minimal risiko) |
+| 2027-08-02 | Alle høyrisiko-krav (full compliance) | Gjelder ikke |
+
+### Referanser
+- `ai-act-classification-methodology.md` — Klassifiseringsmetodikk (steg 1-4)
+- `ai-act-compliance-guide.md` — Generell veileder
+- `ai-act-annex-iii-checklist.md` — Annex III-sjekkliste
+- Microsoft Learn: Azure OpenAI EU Data Boundary compliance
--- a/plugins/ms-ai-architect/tests/fixtures/cost-estimation/fixture.md
+++ b/plugins/ms-ai-architect/tests/fixtures/cost-estimation/fixture.md
@ -0,0 +1,58 @@
+## S6: Kostnadsvurdering
+
+### S6.1: TCO per alternativ (3 år)
+
+| Kostnadspost | Alt 0 | Alt 1 | Alt 2 | Alt 2B | Alt 3 |
+|-------------|-------|-------|-------|--------|-------|
+| **Etablering** | 0 | 0 | 800K | 2 035K | 3 500K |
+| Prosjektkostnader | 0 | 0 | 200K | 500K | 800K |
+| Utvikling (interne) | 0 | 0 | 400K | 885K | 1 500K |
+| QA-konsulent | 0 | 0 | 0 | 200K | 400K |
+| Opplæring | 0 | 0 | 100K | 300K | 300K |
+| Buffer | 0 | 0 | 100K | 150K | 500K |
+| **Årlig drift** | 0 | 0 | 600K | 1 350K | 1 800K |
+| Lisenser (Copilot Studio) | 0 | 0 | 230K | 230K | 0 |
+| Azure OpenAI (tokens) | 0 | 0 | 120K | 210K | 210K |
+| Azure AI Search | 0 | 0 | 110K | 200K | 200K |
+| Infrastruktur (øvrig) | 0 | 0 | 50K | 120K | 300K |
+| Drift/vedlikehold | 0 | 0 | 50K | 150K | 350K |
+| Overvåking | 0 | 0 | 20K | 80K | 100K |
+| Embedding-refresh | 0 | 0 | 20K | 60K | 60K |
+| Regional støtte/opplæring | 0 | 0 | 0 | 150K | 150K |
+| Risk buffer (drift) | 0 | 0 | 0 | 150K | 230K |
+| **3-års TCO** | **0** | **0** | **2 000K** | **6 335K** | **9 100K** |
+
+Alle beløp i NOK. P50-estimat. Valutakurs: 11 NOK/USD.
+
+### S6.2: Konfidensgradering
+
+| Kostnadspost | Konfidens | Kilde |
+|-------------|-----------|-------|
+| Azure OpenAI token-priser | 🟢 Høy | MCP-verifisert (microsoft-learn) |
+| Azure AI Search S1 | 🟢 Høy | MCP-verifisert |
+| Copilot Studio capacity | 🟡 Middels | Fra kunnskapsbase (kan endre seg med ny prismodell) |
+| Intern utviklerkostnad | 🟡 Middels | Estimert 700 NOK/time, 80% dedikasjon |
+| Gevinstrealisering | 🔴 Lav | Antatt, basert på generelle produktivitetsestimater |
+
+### S6.3: Gevinstrealisering (justert 30-50% realisering)
+
+| År | Brutto gevinst | Realiserings-grad | Netto gevinst | AI-kostnad | Netto |
+|----|---------------|-------------------|---------------|------------|-------|
+| År 1 | 77M | 20% | 15.4M | 2.035M | +13.4M |
+| År 2 | 77M | 40% | 30.8M | 1.35M | +29.5M |
+| År 3 | 77M | 50% | 38.5M | 1.35M | +37.2M |
+
+**NNV (3 år, 4% diskonteringsrente):** ~+80M NOK (konservativt estimat)
+**Tilbakebetalingstid:** ~2 måneder (Fase 1-drift)
+
+> **Merknad:** Gevinstestimatene er konservative (30-50% realisering vs. teoretisk 100%). Faktisk realisering avhenger av adopsjon, datakvalitet og endringsledelse. Anbefaler gevinstmåling fra måned 3.
+
+### S6.4: Copilot Studio-lisensstrategi
+
+Start med 1 capacity pack (25K meldinger/mnd). Skaler basert på faktisk bruk:
+- MVP-pilotfase: 1 capacity pack (~230K NOK/år)
+- Full utrulling: 2-3 capacity packs basert på bruksvolumet
+- Overvåk via Copilot Studio Analytics og Azure Monitor
+
+---
+
--- a/plugins/ms-ai-architect/tests/fixtures/security-assessment/fixture.md
+++ b/plugins/ms-ai-architect/tests/fixtures/security-assessment/fixture.md
@ -0,0 +1,63 @@
+## S5: Sikkerhetsvurdering
+
+### S5.1: Sikkerhetsscoring
+
+**Totalscore: 2.80/5.00 — BETINGET AKSEPTABEL**
+
+| Dimensjon | Score | Status | Viktigste funn |
+|-----------|-------|--------|----------------|
+| Identity & Access | 3.5/5 | 🟡 Adekvat | Entra ID + Managed Identity. Mangler: PIM for admin, break-glass prosedyrer. |
+| Network Security | 2.5/5 | 🟡 Adekvat | VNet-integrasjon planlagt. Mangler: Private Endpoints for alle endepunkter. |
+| Data Protection | 2.0/5 | 🔴 Svak | Sweden Central for Azure OpenAI. Presidio PII planlagt. Mangler: CMK, DLP-policyer. |
+| Content Safety | 3.5/5 | 🟡 Adekvat | Content Safety + Prompt Shields planlagt. Mangler: norskspesifikke policyer, Groundedness Detection tuning. |
+| Compliance & Governance | 2.5/5 | 🟡 Adekvat | AI Act-klassifisering utført. Mangler: DPIA, AI-register, Schrems II TIA. |
+| Monitoring & Response | 2.8/5 | 🟡 Adekvat | Application Insights + Sentinel planlagt. Mangler: AI-spesifikke deteksjonsregler, incident response-plan. |
+
+### Kritiske funn
+
+**P0 (Blokkerende):**
+1. **DPIA ikke gjennomført** — Obligatorisk (GDPR art. 35) pga. PII i avviksrapporter
+2. **Schrems II TIA mangler** — Azure OpenAI i Sweden Central krever formell risikovurdering
+3. **Incident Response-plan mangler** — Ingen prosedyre for AI-spesifikke hendelser (prompt injection breach, PII-lekkasje)
+
+**P1 (Høy prioritet):**
+4. **Red team-testing av security trimming** — Må verifisere at regionbasert tilgangskontroll fungerer korrekt
+5. **AI-spesifikk alerting mangler** — Ingen varsling for anomalier i prompt-mønstre eller uvanlig token-forbruk
+6. **PIM ikke konfigurert** — Admin-tilgang til AI-ressurser bør styres via Privileged Identity Management
+
+### S5.2: DPIA-status
+
+| Spørsmål | Svar |
+|----------|------|
+| Behandles personopplysninger? | Ja — PII i avviksrapporter, chatlogger |
+| Er DPIA påkrevd? | Ja (GDPR art. 35) |
+| Er DPIA gjennomført? | ⬜ Ikke startet |
+| Personvernombud involvert? | Planlagt for Fase 1 |
+| Konsultasjon med Datatilsynet? | Vurderes etter DPIA |
+
+**Tidsestimat:** 1-2 uker for gjennomføring, inkl. personvernombuds review.
+
+### S5.3: ROS-analyse
+
+| # | Risiko | S | K | Nivå | Tiltak | Restrisiko |
+|---|--------|---|---|------|--------|------------|
+| 1 | Hallusinering i faglige svar | 4 | 3 | HØY | Groundedness detection 0.8, kildehenvisninger, HITL-disclaimer | MIDDELS |
+| 2 | PII-lekkasje via AI-svar | 2 | 4 | HØY | Presidio pre-indexing + runtime PII-filter + security trimming | LAV |
+| 3 | Prompt injection | 3 | 3 | MIDDELS | Prompt Shields + Content Safety + system message-hardening | LAV |
+| 4 | Uautorisert dokumenttilgang | 2 | 4 | MIDDELS | Security trimming med SharePoint ACL-mapping + Entra ID | LAV |
+| 5 | Schrems II — data residency | 3 | 3 | MIDDELS | Formell TIA + risikoaksept. Data-at-rest i Norway East. | MIDDELS |
+| 6 | Modell-degradering over tid | 2 | 2 | LAV | Kvartalsvis eval + drift-monitoring + automatisk rollback | LAV |
+
+### S5.4: Dataklassifisering
+
+| Datatype | Klassifisering | Behandlingsgrunnlag | Lagringssted |
+|----------|----------------|---------------------|--------------|
+| Håndbøker (N-serien) | Intern | Berettiget interesse | Azure AI Search (Norway East) |
+| Kontrakter | Fortrolig | Berettiget interesse | Azure AI Search (Norway East), security trimmed |
+| Inspeksjonsrapporter | Intern | Berettiget interesse | Azure AI Search (Norway East) |
+| Avviksrapporter (PII) | Fortrolig | Samtykke/Lovhjemmel | Azure AI Search (Norway East), PII-maskert |
+| Chatlogger | Intern | Berettiget interesse | Application Insights (Sweden Central), 90d retention |
+| NVDB-data | Åpen | Åpne data | Ikke lagret — live API-oppslag |
+
+---
+
--- a/plugins/ms-ai-architect/tests/fixtures/summary/fixture.md
+++ b/plugins/ms-ai-architect/tests/fixtures/summary/fixture.md
@ -0,0 +1,98 @@
+## S1: Sammendrag
+
+### 1.1 Teknisk sammendrag
+
+Denne utredningen vurderer innføring av en AI-assistert kunnskapssøk-løsning for drift- og vedlikeholdsavdelinger i en norsk statlig veietat med ~1500 ansatte fordelt på 5 regioner. Løsningen skal erstatte dagens fragmenterte, manuelle dokumentsøk på tvers av 65 000+ dokumenter i 5 separate systemer (SharePoint, Doculive, Landbruks-IT, lokale filservere, NVDB).
+
+**Anbefalt plattform:** Alternativ 2B — Hybrid (Copilot Studio + Azure AI Foundry RAG) (S2.5, S8.1)
+
+Copilot Studio fungerer som Teams-native UI-lag («Veihjelper AI»), mens Azure AI Foundry gir full kontroll over RAG-pipeline med custom chunking, Presidio PII-filtrering og SharePoint ACL-basert security trimming. Multi-modell-strategi med GPT-4o for komplekse fagspørsmål og GPT-4o-mini for enkel gjenfinning (80/20 cost routing) sikrer kostnadseffektivitet (S4.2). Embedding med text-embedding-3-large (3072 dim) i Azure AI Search S1 med hybrid search og semantic reranking gir best ytelse for norsk fagterminologi (S4.3).
+
+**Arkitektur (S8.2):** Brukerinteraksjon via Teams (mobil/desktop) → Entra ID-autentisering → Copilot Studio Agent → Azure AI Foundry RAG-pipeline (Azure OpenAI GPT-4o/mini i Sweden Central, Azure AI Search S1 i Norway East) + NVDB REST API via function calling. Presidio PII-filter (pre-indexing) og Azure AI Content Safety (runtime) sikrer personvern og innholdssikkerhet.
+
+**Sikkerhetsstatus (S5):** Totalscore 2.80/5.00 — BETINGET AKSEPTABEL. Tre P0-blokkere er identifisert: (1) DPIA ikke gjennomført (GDPR art. 35), (2) Schrems II TIA mangler for Azure OpenAI i Sweden Central, (3) Incident Response-plan mangler. Alle tre er innarbeidet som obligatoriske aktiviteter i Fase 0 av implementeringsplanen (S9.4). AI Act-klassifisering: Begrenset risiko — transparenskrav oppfylt med AI-merking og kildehenvisninger (S4.1).
+
+**Kostnad (S6):** Etableringskostnad 2.035M NOK (Fase 1, innenfor 2M-budsjettramme). Årlig driftskostnad 1.35M NOK inkludert Copilot Studio capacity pack (230K), Azure OpenAI tokens (210K), Azure AI Search (200K), infrastruktur (120K), drift/vedlikehold (150K), overvåking (80K), embedding-refresh (60K), regional støtte (150K) og risk buffer (150K). 3-års TCO: 6.335M NOK. Fase 2 (Doculive + Landbruks-IT) budsjetteres separat til 850K.
+
+**Implementeringsplan (S9):** Fasevis utrulling over 48 uker. Fase 0 (uke 1-4): Forberedelse, DPIA, Schrems II TIA, Azure OpenAI-søknad. Fase 1 (uke 5-32): MVP med SharePoint (39K dok) + NVDB, inkludert infrastruktur, RAG-pipeline, Copilot Studio agent, sikkerhetstesting, pilot med 50 brukere, opplæring (5 regioner) og region-for-region utrulling. Fase 2 (uke 33-48): Full dekning med Doculive (13K) og Landbruks-IT (10K). 8 milepæler definert (M0-M8), inkludert gevinstevaluering i uke 56 (S9.2).
+
+**Arkitekturprinsipper (S3):** 3 av 7 Digdir-prinsipper fullt oppfylt, 4 delvis oppfylt. Hovedavvik: Schrems II-risiko (P4 Tillit), begrenset ekstern datadeling (P2/P7). Trade-off vedtatt: Sweden Central aksepteres med formell risikoaksept da Norway East ikke tilbyr Azure OpenAI.
+
+**Digital samhandling (S7):** Juridisk samhandling er oppfylt (berettiget interesse, DPA med Microsoft, ingen vedtaksfatning). Organisatorisk samhandling er delvis under arbeid (RACI-matrise, AI-styringsgruppe). Semantisk samhandling er oppfylt (OData, Dublin Core, standardisert embedding). Teknisk samhandling er oppfylt (REST API, OAuth 2.0, SLA 99.8%). Styring planlagt med kvartalsvis AI-styringsgruppe og definerte KPI-er.
+
+5 ADR-er er utarbeidet (S8.3): Copilot Studio som UI (ADR-001), custom RAG i Azure AI Foundry (ADR-002), Sweden Central med risikoaksept (ADR-003), batch-import fra Landbruks-IT (ADR-004), og Security Trimming med SharePoint ACL-mapping (ADR-005).
+
+---
+
+### 1.2 Beslutningsnotat (Executive Summary)
+
+**Anbefaling:** Iverksett Alternativ 2B — Hybrid Copilot Studio + Azure AI Foundry RAG for AI-assistert dokumentsøk i drift- og vedlikeholdsavdelingene.
+
+**Bakgrunn:** Driftspersonell bruker 3-5 timer per uke på manuelt dokumentsøk i 5 ulike systemer. Produktivitetstapet er estimert til ~77M NOK/år for hele organisasjonen. Kunnskapen er fragmentert, eksisterende søk forstår ikke norsk vei-fagterminologi, og det finnes ingen felles søkeflate.
+
+**Hva vi anbefaler:** En Teams-basert AI-assistent («Veihjelper AI») som gir umiddelbar tilgang til hele dokumentkorpuset (65 000+ dokumenter) med norskspråklig semantisk søk, kildehenvisninger og sanntids veidata fra NVDB. Løsningen bruker Microsofts AI-plattform med full kontroll over sikkerhet og personvern.
+
+**Hvorfor dette alternativet:** Alt 2B gir den beste balansen mellom brukeropplevelse (Teams-native), sikkerhet (PII-filtrering, security trimming) og kostnad (innenfor 2M budsjett for Fase 1). Alt 1 (SharePoint AI Search) dekker kun 60% av dokumentene. Alt 2 (innebygd RAG) mangler kontroll over PII og security trimming. Alt 3 (full custom) er overingeniørt og sprenger budsjettet.
+
+**Investering:** 2.035M NOK etableringskostnad (Fase 1), 1.35M NOK årlig drift. 3-års TCO: 6.335M NOK. Fase 2 (full dokumentdekning): 850K NOK separat.
+
+**Forventet gevinst:** Med konservativ 30-50% realisering av produktivitetsgevinst gir løsningen en NNV på ~+80M NOK over 3 år. Tilbakebetalingstid: ~2 måneder etter produksjonssetting. Gevinstmåling starter fra måned 3.
+
+**Sikkerhet og compliance:** Sikkerhetsstatus er BETINGET AKSEPTABEL (2.80/5). Tre blokkerende funn må lukkes før produksjon: DPIA, Schrems II TIA og Incident Response-plan. Alle er planlagt gjennomført i Fase 0 (uke 1-4). AI Act: Begrenset risiko — transparenskrav oppfylt.
+
+**Tidsplan:** Fase 0 (forberedelse): 4 uker. Fase 1 (MVP i produksjon): 7-8 måneder. Fase 2 (full dekning): 3-4 måneder. Total: ~12 måneder til full dekning.
+
+**Kritiske forutsetninger:**
+1. Azure OpenAI-tilgang i Sweden Central må innvilges
+2. DPIA må godkjennes av personvernombud
+3. 2 Azure-utviklere må dedikeres minimum 80%
+4. Schrems II-risikoaksept må signeres av behandlingsansvarlig
+
+**Risiko:** Hovedrisikoen er hallusinering i faglige svar (HØY), som mitigeres med groundedness detection (terskel 0.8), kildehenvisninger og tydelig AI-disclaimer. PII-lekkasje (HØY) mitigeres med Presidio pre-indexing og runtime PII-filter.
+
+**Anbefaling til ledelsen:** Godkjenn prosjektoppstart med Go/No-Go-beslutning etter Fase 0 (4 uker). Bevilg 2.035M NOK for Fase 1. Utpek prosjekteier med mandat i KI-seksjonen. Igangsett DPIA umiddelbart.
+
+---
+
+### 1.3 Nøkkeltall
+
+| Parameter | Verdi |
+|-----------|-------|
+| **Anbefalt alternativ** | 2B: Hybrid (Copilot Studio + Azure AI Foundry RAG) |
+| **AI Act-risikoklasse** | Begrenset risiko |
+| **Kompleksitetsscore** | 14/18 (KOMPLEKS) |
+| **Sikkerhetsscore** | 2.80/5.00 (Betinget akseptabel) |
+| **P0-blokkere** | 3 (DPIA, Schrems II TIA, Incident Response) |
+| **Etableringskostnad (Fase 1)** | 2.035M NOK |
+| **Årlig driftskostnad** | 1.35M NOK |
+| **3-års TCO** | 6.335M NOK |
+| **Fase 2-kostnad** | 850K NOK (separat) |
+| **NNV (3 år, 4% diskontering)** | ~+80M NOK (konservativt) |
+| **Tilbakebetalingstid** | ~2 måneder (etter Fase 1-drift) |
+| **Brutto produktivitetsgevinst** | ~77M NOK/år |
+| **Gevinstraliseringsgrad** | 30-50% (konservativt) |
+| **Dokumentkorpus** | 65 000+ dokumenter |
+| **Datakilder** | 5 (SharePoint, Doculive, Landbruks-IT, filservere, NVDB) |
+| **Primærbrukere** | ~750 (potensielt 1500) |
+| **Time-to-value (MVP)** | 7-8 måneder (Fase 1) |
+| **Full dekning** | 12 måneder (Fase 1 + 2) |
+| **AI-modeller** | GPT-4o + GPT-4o-mini, text-embedding-3-large |
+| **Azure-regioner** | Sweden Central (OpenAI), Norway East (AI Search) |
+| **SLA** | ~99.8% (samlet) |
+| **ADR-er** | 5 (alle Accepted) |
+| **Digdir-prinsipper** | 3/7 fullt oppfylt, 4/7 delvis |
+| **Antall milepæler** | 8 (M0-M8, uke 1-56) |
+
+---
+
+### 1.4 Konfidenstabell
+
+| Dimensjon | Konfidens | Begrunnelse |
+|-----------|-----------|-------------|
+| Teknisk gjennomførbarhet | 🟢 Høy | RAG-teknologi er veletablert (Azure AI Search GA siden 2014, Azure OpenAI GA siden 2023). Copilot Studio GA med AI-kapabiliteter. 2 interne Azure-utviklere finnes. |
+| Kostnadsestimat | 🟡 Middels | Azure-priser MCP-verifisert (høy). Intern utviklerkostnad estimert (middels). Gevinstestimater konservative men usikre (lav). P50 med 150K buffer. |
+| Regulatorisk compliance | 🟡 Middels | AI Act avklart (Begrenset risiko). Schrems II mitigerbar med TIA. DPIA ikke gjennomført — blokkerende men standard prosess (1-2 uker). |
+| Organisatorisk gjennomførbarhet | 🟡 Middels | Kompetansegap mitigeres med 300K opplæring + 200K QA-konsulent. Avhenger av 80% dedikering fra 2 utviklere. Endringsledelse budsjettert med superbrukermodell. |
+
+---
+
--- a/plugins/ms-ai-architect/tests/lib/e2e-helpers.sh
+++ b/plugins/ms-ai-architect/tests/lib/e2e-helpers.sh
@ -0,0 +1,187 @@
+#!/bin/bash
+# e2e-helpers.sh — Shared validation functions for E2E regression tests
+set -euo pipefail
+
+# Counters
+PASS=0
+FAIL=0
+WARN=0
+SUITE_NAME=""
+
+# Colors
+RED='\033[0;31m'
+GREEN='\033[0;32m'
+YELLOW='\033[1;33m'
+BLUE='\033[0;34m'
+NC='\033[0m'
+
+init_suite() {
+  SUITE_NAME="$1"
+  PASS=0
+  FAIL=0
+  WARN=0
+  echo -e "${BLUE}═══════════════════════════════════════${NC}"
+  echo -e "${BLUE}  E2E Suite: $SUITE_NAME${NC}"
+  echo -e "${BLUE}═══════════════════════════════════════${NC}"
+  echo ""
+}
+
+pass() {
+  PASS=$((PASS + 1))
+  echo -e "  ${GREEN}PASS${NC} $1"
+}
+
+fail() {
+  FAIL=$((FAIL + 1))
+  echo -e "  ${RED}FAIL${NC} $1"
+}
+
+warn() {
+  WARN=$((WARN + 1))
+  echo -e "  ${YELLOW}WARN${NC} $1"
+}
+
+print_summary() {
+  local total=$((PASS + FAIL))
+  echo ""
+  echo -e "${BLUE}───────────────────────────────────────${NC}"
+  echo -e "  ${SUITE_NAME}: ${PASS}/${total} PASS, ${FAIL} FAIL, ${WARN} WARN"
+  echo -e "${BLUE}───────────────────────────────────────${NC}"
+  echo ""
+  return $FAIL
+}
+
+# --- Assertions ---
+
+assert_has_section() {
+  local file="$1"
+  local header="$2"
+  local description="${3:-Section: $header}"
+  if grep -qE "^#{1,4} .*${header}" "$file"; then
+    pass "$description"
+  else
+    fail "$description — header '$header' not found"
+  fi
+}
+
+assert_min_lines() {
+  local file="$1"
+  local min="$2"
+  local description="${3:-Minimum $min lines}"
+  local count
+  count=$(wc -l < "$file" | tr -d ' ')
+  if [ "$count" -ge "$min" ]; then
+    pass "$description ($count lines)"
+  else
+    fail "$description — only $count lines (min: $min)"
+  fi
+}
+
+assert_encoding_ok() {
+  local file="$1"
+  local description="${2:-UTF-8 encoding valid}"
+  # Check for common broken UTF-8 sequences
+  if file "$file" | grep -qi "utf-8\|ascii\|unicode"; then
+    pass "$description"
+  else
+    fail "$description — encoding not UTF-8/ASCII"
+  fi
+}
+
+assert_no_ascii_approximation() {
+  local file="$1"
+  local description="${2:-No ASCII approximation of Norwegian chars}"
+  # Check for ae/oe/aa used where æ/ø/å should be (Norwegian words)
+  # Only flag common Norwegian words that should have æøå
+  local violations=0
+  for pattern in '\bvaere\b' '\bfoerste\b' '\bhoey\b' '\bgjoere\b' '\baarlig\b' '\bsaerlig\b' '\bnoedvendig\b'; do
+    if grep -qiE "$pattern" "$file" 2>/dev/null; then
+      violations=$((violations + 1))
+    fi
+  done
+  if [ "$violations" -eq 0 ]; then
+    pass "$description"
+  else
+    fail "$description — $violations ASCII approximation patterns found"
+  fi
+}
+
+assert_scores_in_range() {
+  local file="$1"
+  local description="${2:-Scores in X/5 format within range}"
+  # Match X/5 or X.X/5 patterns and verify range
+  local bad_scores=0
+  while IFS= read -r match; do
+    local score="${match%%/*}"
+    score=$(echo "$score" | tr -d ' ')
+    # Check if it's a valid number between 0 and 5
+    if echo "$score" | grep -qE '^[0-5](\.[0-9]+)?$'; then
+      :
+    else
+      bad_scores=$((bad_scores + 1))
+    fi
+  done < <(grep -oE '[0-9]+\.?[0-9]*/5' "$file" 2>/dev/null || true)
+
+  local total_scores
+  total_scores=$(grep -cE '[0-9]+\.?[0-9]*/5' "$file" 2>/dev/null || echo "0")
+  if [ "$total_scores" -gt 0 ] && [ "$bad_scores" -eq 0 ]; then
+    pass "$description ($total_scores scores found)"
+  elif [ "$total_scores" -eq 0 ]; then
+    fail "$description — no X/5 scores found"
+  else
+    fail "$description — $bad_scores out-of-range scores"
+  fi
+}
+
+assert_has_nok_amounts() {
+  local file="$1"
+  local min="${2:-1}"
+  local description="${3:-NOK amounts present}"
+  local count
+  count=$(grep -cE '([0-9]+[.,]?[0-9]*(M|K)\s*NOK|[0-9]+\s*NOK|NOK\s*[0-9]+|[0-9]+\s*000\s*NOK)' "$file" 2>/dev/null || echo "0")
+  if [ "$count" -ge "$min" ]; then
+    pass "$description ($count found)"
+  else
+    fail "$description — only $count NOK amounts (min: $min)"
+  fi
+}
+
+assert_min_tables() {
+  local file="$1"
+  local min="$2"
+  local description="${3:-Minimum $min markdown tables}"
+  # Count lines starting with | that contain | somewhere after the first char (table rows)
+  local table_separators
+  table_separators=$(grep -cE '^\|.*\|.*\|' "$file" 2>/dev/null || echo "0")
+  # Rough estimate: each table has header + separator + at least 1 row = 3 lines minimum
+  local estimated_tables=$((table_separators / 3))
+  if [ "$estimated_tables" -ge "$min" ]; then
+    pass "$description (~$estimated_tables tables)"
+  else
+    fail "$description — only ~$estimated_tables tables (min: $min)"
+  fi
+}
+
+assert_matches_pattern() {
+  local file="$1"
+  local pattern="$2"
+  local description="${3:-Pattern match: $pattern}"
+  if grep -qE "$pattern" "$file"; then
+    pass "$description"
+  else
+    fail "$description"
+  fi
+}
+
+assert_has_dimensions() {
+  local file="$1"
+  local min="$2"
+  local description="${3:-Minimum $min security dimensions}"
+  local count
+  count=$(grep -cE '^\| .+ \| [0-9]+\.?[0-9]*/5' "$file" 2>/dev/null || echo "0")
+  if [ "$count" -ge "$min" ]; then
+    pass "$description ($count dimensions)"
+  else
+    fail "$description — only $count dimensions (min: $min)"
+  fi
+}
--- a/plugins/ms-ai-architect/tests/run-e2e.sh
+++ b/plugins/ms-ai-architect/tests/run-e2e.sh
@ -0,0 +1,78 @@
+#!/bin/bash
+# run-e2e.sh — Run E2E regression tests for agent outputs
+# Usage: bash tests/run-e2e.sh [--security] [--cost] [--summary] [--all]
+set -euo pipefail
+
+SCRIPT_DIR="$(cd "$(dirname "$0")" && pwd)"
+
+RED='\033[0;31m'
+GREEN='\033[0;32m'
+CYAN='\033[0;36m'
+NC='\033[0m'
+
+# Parse arguments
+RUN_SECURITY=false
+RUN_COST=false
+RUN_SUMMARY=false
+RUN_ROS=false
+RUN_AI_ACT=false
+
+if [ $# -eq 0 ] || [ "${1:-}" = "--all" ]; then
+  RUN_SECURITY=true
+  RUN_COST=true
+  RUN_SUMMARY=true
+  RUN_ROS=true
+  RUN_AI_ACT=true
+else
+  while [ $# -gt 0 ]; do
+    case "$1" in
+      --security) RUN_SECURITY=true ;;
+      --cost)     RUN_COST=true ;;
+      --summary)  RUN_SUMMARY=true ;;
+      --ros)      RUN_ROS=true ;;
+      --ai-act)   RUN_AI_ACT=true ;;
+      *)
+        echo "Usage: bash tests/run-e2e.sh [--security] [--cost] [--summary] [--ros] [--ai-act] [--all]"
+        exit 1
+        ;;
+    esac
+    shift
+  done
+fi
+
+echo -e "${CYAN}╔══════════════════════════════════════════════╗${NC}"
+echo -e "${CYAN}║  MS AI Architect — E2E Regression Tests      ║${NC}"
+echo -e "${CYAN}╚══════════════════════════════════════════════╝${NC}"
+echo ""
+
+FAILURES=0
+
+if $RUN_SECURITY; then
+  bash "$SCRIPT_DIR/test-security-output.sh" || FAILURES=$((FAILURES + 1))
+fi
+
+if $RUN_COST; then
+  bash "$SCRIPT_DIR/test-cost-output.sh" || FAILURES=$((FAILURES + 1))
+fi
+
+if $RUN_SUMMARY; then
+  bash "$SCRIPT_DIR/test-summary-output.sh" || FAILURES=$((FAILURES + 1))
+fi
+
+if $RUN_ROS; then
+  bash "$SCRIPT_DIR/test-ros-output.sh" || FAILURES=$((FAILURES + 1))
+fi
+
+if $RUN_AI_ACT; then
+  bash "$SCRIPT_DIR/test-ai-act-output.sh" || FAILURES=$((FAILURES + 1))
+fi
+
+echo -e "${CYAN}══════════════════════════════════════════════${NC}"
+if [ "$FAILURES" -eq 0 ]; then
+  echo -e "${GREEN}  All E2E suites passed${NC}"
+else
+  echo -e "${RED}  $FAILURES suite(s) had failures${NC}"
+fi
+echo -e "${CYAN}══════════════════════════════════════════════${NC}"
+
+exit $FAILURES
--- a/plugins/ms-ai-architect/tests/test-ai-act-output.sh
+++ b/plugins/ms-ai-architect/tests/test-ai-act-output.sh
@ -0,0 +1,81 @@
+#!/bin/bash
+# test-ai-act-output.sh — Validate ai-act-assessor agent output structure
+set -euo pipefail
+
+SCRIPT_DIR="$(cd "$(dirname "$0")" && pwd)"
+source "$SCRIPT_DIR/lib/e2e-helpers.sh"
+
+FIXTURE="$SCRIPT_DIR/fixtures/ai-act/fixture.md"
+FIXTURE_HR="$SCRIPT_DIR/fixtures/ai-act/fixture-high-risk.md"
+
+if [ ! -f "$FIXTURE" ]; then
+  echo "ERROR: Fixture not found: $FIXTURE"
+  echo "Run: bash tests/capture-fixture.sh to generate fixtures"
+  exit 1
+fi
+
+if [ ! -f "$FIXTURE_HR" ]; then
+  echo "ERROR: High-risk fixture not found: $FIXTURE_HR"
+  exit 1
+fi
+
+init_suite "AI Act Assessor Agent"
+
+# === Minimal risk fixture ===
+echo ""
+echo "  --- Minimal Risk Fixture ---"
+
+# Structure checks
+assert_has_section "$FIXTURE" "Risikoklassifisering" "Has risk classification section"
+assert_has_section "$FIXTURE" "Rolle" "Has role section"
+assert_has_section "$FIXTURE" "Forpliktelser" "Has obligations section"
+
+# Content quality
+assert_min_lines "$FIXTURE" 20 "Minimum 20 lines"
+
+# Encoding
+assert_encoding_ok "$FIXTURE" "UTF-8 encoding valid (minimal)"
+assert_no_ascii_approximation "$FIXTURE" "No ASCII approximation (minimal)"
+
+# Domain-specific: classification result
+assert_matches_pattern "$FIXTURE" "(Minimal risiko|Begrenset risiko|Høyrisiko|Forbudt)" "Contains risk level classification"
+assert_matches_pattern "$FIXTURE" "(Provider|Deployer)" "Contains role determination"
+
+# Domain-specific: article references
+assert_matches_pattern "$FIXTURE" "Art\. [0-9]" "References specific AI Act articles"
+
+# === High risk fixture ===
+echo ""
+echo "  --- High Risk Fixture ---"
+
+# Structure checks
+assert_has_section "$FIXTURE_HR" "Risikoklassifisering" "HR: Has risk classification section"
+assert_has_section "$FIXTURE_HR" "Rolle" "HR: Has role section"
+assert_has_section "$FIXTURE_HR" "Forpliktelser" "HR: Has obligations section"
+assert_has_section "$FIXTURE_HR" "Tiltaksplan" "HR: Has action plan section"
+assert_has_section "$FIXTURE_HR" "Neste steg" "HR: Has next steps section"
+assert_has_section "$FIXTURE_HR" "Viktige frister" "HR: Has deadline section"
+
+# Content quality
+assert_min_lines "$FIXTURE_HR" 40 "HR: Minimum 40 lines"
+assert_min_tables "$FIXTURE_HR" 1 "HR: Has at least 1 table"
+
+# Encoding
+assert_encoding_ok "$FIXTURE_HR" "HR: UTF-8 encoding valid"
+assert_no_ascii_approximation "$FIXTURE_HR" "HR: No ASCII approximation"
+
+# Domain-specific: high-risk classification
+assert_matches_pattern "$FIXTURE_HR" "Høyrisiko" "HR: Classified as high-risk"
+assert_matches_pattern "$FIXTURE_HR" "Annex III" "HR: References Annex III"
+assert_matches_pattern "$FIXTURE_HR" "FRIA" "HR: References FRIA requirement"
+
+# Domain-specific: deadlines
+assert_matches_pattern "$FIXTURE_HR" "202[5-7]-[0-9]{2}-[0-9]{2}" "HR: Contains compliance deadline dates"
+
+# Domain-specific: article references
+assert_matches_pattern "$FIXTURE_HR" "Art\. (26|27|9|13|14|50)" "HR: References key deployer/provider articles"
+
+# Domain-specific: action plan
+assert_matches_pattern "$FIXTURE_HR" "(Kritisk|Høy|Middels|Lav)" "HR: Action plan has priority levels"
+
+print_summary
--- a/plugins/ms-ai-architect/tests/test-cost-output.sh
+++ b/plugins/ms-ai-architect/tests/test-cost-output.sh
@ -0,0 +1,39 @@
+#!/bin/bash
+# test-cost-output.sh — Validate cost-estimation-agent output structure
+set -euo pipefail
+
+SCRIPT_DIR="$(cd "$(dirname "$0")" && pwd)"
+source "$SCRIPT_DIR/lib/e2e-helpers.sh"
+
+FIXTURE="$SCRIPT_DIR/fixtures/cost-estimation/fixture.md"
+
+if [ ! -f "$FIXTURE" ]; then
+  echo "ERROR: Fixture not found: $FIXTURE"
+  echo "Run: bash tests/capture-fixture.sh to generate fixtures"
+  exit 1
+fi
+
+init_suite "Cost Estimation Agent"
+
+# Structure checks
+assert_has_section "$FIXTURE" "Kostnad" "Has cost section header"
+assert_has_section "$FIXTURE" "TCO" "Has TCO section"
+assert_has_section "$FIXTURE" "Konfidens" "Has confidence grading section"
+assert_has_section "$FIXTURE" "Gevinst" "Has benefit realization section"
+
+# Content quality
+assert_min_lines "$FIXTURE" 30 "Minimum 30 lines"
+assert_min_tables "$FIXTURE" 2 "Minimum 2 tables (TCO, confidence)"
+assert_has_nok_amounts "$FIXTURE" 3 "At least 3 NOK amounts"
+
+# Encoding
+assert_encoding_ok "$FIXTURE" "UTF-8 encoding valid"
+assert_no_ascii_approximation "$FIXTURE" "No ASCII approximation of Norwegian chars"
+
+# Domain-specific
+assert_matches_pattern "$FIXTURE" "(Alt|Alternativ)\s*[0-9]" "References numbered alternatives"
+assert_matches_pattern "$FIXTURE" "(Etablering|etablering)" "Has establishment cost section"
+assert_matches_pattern "$FIXTURE" "(drift|Drift)" "Has operational cost section"
+assert_matches_pattern "$FIXTURE" "(MCP-verifisert|Verifisert|verifisert)" "Has verification markers"
+
+print_summary
--- a/plugins/ms-ai-architect/tests/test-hooks.sh
+++ b/plugins/ms-ai-architect/tests/test-hooks.sh
@ -0,0 +1,132 @@
+#!/bin/bash
+# test-hooks.sh — Unit tests for ms-ai-architect hook scripts
+# Usage: bash tests/test-hooks.sh
+
+set -euo pipefail
+
+PLUGIN_ROOT="$(cd "$(dirname "$0")/.." && pwd)"
+SCRIPTS_DIR="$PLUGIN_ROOT/hooks/scripts"
+PASS=0
+FAIL=0
+
+pass() { echo -e "\033[0;32m  ✓ $1\033[0m"; PASS=$((PASS + 1)); }
+fail() { echo -e "\033[0;31m  ✗ $1\033[0m"; FAIL=$((FAIL + 1)); }
+
+echo "=== Hook Script Tests ==="
+echo ""
+
+# -------------------------------------------------------
+# pre-edit-secrets.mjs
+# -------------------------------------------------------
+echo "--- pre-edit-secrets.mjs ---"
+
+# Test 1: clean content passes
+echo '{"tool_input":{"content":"Hello world","file_path":"readme.md"}}' \
+  | node "$SCRIPTS_DIR/pre-edit-secrets.mjs" 2>/dev/null \
+  && pass "Clean content: exit 0" \
+  || fail "Clean content: expected exit 0"
+
+# Test 2: Azure Storage key blocked (constructed at runtime to avoid hook self-trigger)
+AZURE_KEY_PAYLOAD=$(printf '{"tool_input":{"content":"%s","file_path":"config.ts"}}' \
+  "DefaultEndpointsProtocol=https;AccountName=test;AccountKey=$(printf 'a%.0s' {1..44})=")
+echo "$AZURE_KEY_PAYLOAD" \
+  | node "$SCRIPTS_DIR/pre-edit-secrets.mjs" 2>/dev/null \
+  && fail "Azure Storage key: expected exit 2 (blocked)" \
+  || pass "Azure Storage key: blocked correctly"
+
+# Test 3: GitHub token blocked (constructed at runtime)
+GH_TOKEN="ghp_$(printf 'a%.0s' {1..40})"
+printf '{"tool_input":{"content":"token: %s","file_path":"ci.yml"}}' "$GH_TOKEN" \
+  | node "$SCRIPTS_DIR/pre-edit-secrets.mjs" 2>/dev/null \
+  && fail "GitHub token: expected exit 2 (blocked)" \
+  || pass "GitHub token: blocked correctly"
+
+# Test 4: empty content passes
+echo '{"tool_input":{"content":"","file_path":"empty.md"}}' \
+  | node "$SCRIPTS_DIR/pre-edit-secrets.mjs" 2>/dev/null \
+  && pass "Empty content: exit 0" \
+  || fail "Empty content: expected exit 0"
+
+# Test 5: test files skipped (constructed at runtime)
+printf '{"tool_input":{"content":"api_key = \\"%s\\"","file_path":"auth.test.ts"}}' \
+  "$(printf 'x%.0s' {1..25})" \
+  | node "$SCRIPTS_DIR/pre-edit-secrets.mjs" 2>/dev/null \
+  && pass "Test file: skipped (exit 0)" \
+  || fail "Test file: should be skipped"
+
+echo ""
+
+# -------------------------------------------------------
+# session-start-context.mjs
+# -------------------------------------------------------
+echo "--- session-start-context.mjs ---"
+
+OUTPUT=$(CLAUDE_PLUGIN_ROOT="$PLUGIN_ROOT" node "$SCRIPTS_DIR/session-start-context.mjs" 2>/dev/null)
+EXIT_CODE=$?
+
+if [ $EXIT_CODE -eq 0 ]; then
+  pass "Runs without error (exit 0)"
+else
+  fail "Expected exit 0, got $EXIT_CODE"
+fi
+
+if echo "$OUTPUT" | grep -q "Architect:"; then
+  pass "Output contains 'Architect:' prefix"
+else
+  fail "Output missing 'Architect:' prefix"
+fi
+
+echo ""
+
+# -------------------------------------------------------
+# stop-assessment-reminder.mjs
+# -------------------------------------------------------
+echo "--- stop-assessment-reminder.mjs ---"
+
+# Test from a temp dir without .work/
+TMPDIR_TEST=$(mktemp -d)
+OUTPUT=$(cd "$TMPDIR_TEST" && node "$SCRIPTS_DIR/stop-assessment-reminder.mjs" 2>/dev/null)
+EXIT_CODE=$?
+
+if [ $EXIT_CODE -eq 0 ]; then
+  pass "No .work/: exit 0"
+else
+  fail "No .work/: expected exit 0, got $EXIT_CODE"
+fi
+
+if [ "$OUTPUT" = "{}" ]; then
+  pass "No .work/: returns {}"
+else
+  fail "No .work/: expected {}, got: $OUTPUT"
+fi
+
+# Test with a fresh .work/ session
+mkdir -p "$TMPDIR_TEST/.work/test-session"
+touch "$TMPDIR_TEST/.work/test-session/state.json"
+OUTPUT=$(cd "$TMPDIR_TEST" && node "$SCRIPTS_DIR/stop-assessment-reminder.mjs" 2>/dev/null)
+
+if echo "$OUTPUT" | grep -q "systemMessage"; then
+  pass "Fresh .work/: returns systemMessage"
+else
+  fail "Fresh .work/: expected systemMessage in output"
+fi
+
+if echo "$OUTPUT" | grep -q "architect:adr"; then
+  pass "Fresh .work/: suggests /architect:adr"
+else
+  fail "Fresh .work/: missing /architect:adr suggestion"
+fi
+
+rm -rf "$TMPDIR_TEST"
+
+echo ""
+
+# -------------------------------------------------------
+# Summary
+# -------------------------------------------------------
+TOTAL=$((PASS + FAIL))
+echo "=== Results: $PASS/$TOTAL passed, $FAIL failed ==="
+
+if [ "$FAIL" -gt 0 ]; then
+  exit 1
+fi
--- a/plugins/ms-ai-architect/tests/test-kb-integrity.sh
+++ b/plugins/ms-ai-architect/tests/test-kb-integrity.sh
@ -0,0 +1,110 @@
+#!/bin/bash
+# test-kb-integrity.sh — Cross-reference agent KB paths against actual files
+# Also finds orphaned KB files not referenced by any agent or SKILL.md
+# Usage: bash tests/test-kb-integrity.sh
+
+set -euo pipefail
+
+PLUGIN_ROOT="$(cd "$(dirname "$0")/.." && pwd)"
+PASS=0
+FAIL=0
+WARN=0
+
+pass() { echo -e "\033[0;32m  ✓ $1\033[0m"; PASS=$((PASS + 1)); }
+fail() { echo -e "\033[0;31m  ✗ $1\033[0m"; FAIL=$((FAIL + 1)); }
+warn() { echo -e "\033[1;33m  ⚠ $1\033[0m"; WARN=$((WARN + 1)); }
+
+echo "=== KB Integrity Test ==="
+echo ""
+
+# -------------------------------------------------------
+# Check 1: Agent file references resolve
+# -------------------------------------------------------
+echo "--- Check 1: Agent File References ---"
+
+for agent_file in "$PLUGIN_ROOT"/agents/*.md; do
+  [ -f "$agent_file" ] || continue
+  basename_agent="$(basename "$agent_file")"
+
+  # Extract explicit file paths (references/, skills/) from agent
+  while IFS= read -r ref_path; do
+    ref_path=$(echo "$ref_path" | sed 's/`//g; s/"//g; s/^[[:space:]]*//; s/[[:space:]]*$//')
+    [ -z "$ref_path" ] && continue
+
+    full_path="$PLUGIN_ROOT/$ref_path"
+
+    if [ -e "$full_path" ]; then
+      pass "$basename_agent -> $ref_path"
+    else
+      fail "$basename_agent -> $ref_path (NOT FOUND)"
+    fi
+  done < <(grep -oE '(references|skills)/[a-zA-Z0-9_./-]+\.md' "$agent_file" 2>/dev/null || true)
+done
+
+echo ""
+
+# -------------------------------------------------------
+# Check 2: SKILL.md file references resolve
+# -------------------------------------------------------
+echo "--- Check 2: SKILL.md File References ---"
+
+for skill_file in "$PLUGIN_ROOT"/skills/*/SKILL.md; do
+  [ -f "$skill_file" ] || continue
+  skill_name="$(basename "$(dirname "$skill_file")")"
+
+  while IFS= read -r ref_path; do
+    ref_path=$(echo "$ref_path" | sed 's/`//g; s/"//g; s/^[[:space:]]*//; s/[[:space:]]*$//')
+    [ -z "$ref_path" ] && continue
+
+    # Resolve: references/ paths are relative to skill dir
+    if [[ "$ref_path" == references/* ]]; then
+      full_path="$PLUGIN_ROOT/skills/$skill_name/$ref_path"
+    else
+      full_path="$PLUGIN_ROOT/$ref_path"
+    fi
+
+    if [ -e "$full_path" ]; then
+      pass "SKILL:$skill_name -> $ref_path"
+    else
+      fail "SKILL:$skill_name -> $ref_path (NOT FOUND)"
+    fi
+  done < <(grep -oE '(references|skills)/[a-zA-Z0-9_./-]+\.md' "$skill_file" 2>/dev/null || true)
+done
+
+echo ""
+
+# -------------------------------------------------------
+# Check 3: Orphaned KB files
+# -------------------------------------------------------
+echo "--- Check 3: Orphaned KB Files ---"
+
+ORPHAN_COUNT=0
+
+while IFS= read -r kb_file; do
+  rel_path="${kb_file#$PLUGIN_ROOT/}"
+  kb_basename="$(basename "$kb_file")"
+
+  # Check if this filename appears in any agent or SKILL.md
+  if grep -rql "$kb_basename" "$PLUGIN_ROOT/agents/" "$PLUGIN_ROOT"/skills/*/SKILL.md 2>/dev/null; then
+    : # Referenced, ok
+  else
+    warn "Orphaned: $rel_path"
+    ORPHAN_COUNT=$((ORPHAN_COUNT + 1))
+  fi
+done < <(find "$PLUGIN_ROOT/skills" -path "*/references/*.md" -type f 2>/dev/null)
+
+if [ "$ORPHAN_COUNT" -eq 0 ]; then
+  pass "No orphaned KB files found"
+fi
+
+echo ""
+
+# -------------------------------------------------------
+# Summary
+# -------------------------------------------------------
+TOTAL=$((PASS + FAIL))
+echo "=== Results: $PASS/$TOTAL passed, $FAIL failed, $WARN warnings ==="
+
+if [ "$FAIL" -gt 0 ]; then
+  exit 1
+fi
--- a/plugins/ms-ai-architect/tests/test-plugin-discovery.sh
+++ b/plugins/ms-ai-architect/tests/test-plugin-discovery.sh
@ -0,0 +1,161 @@
+#!/bin/bash
+# test-plugin-discovery.sh — Smoke test for auto-discovery chain
+# Validates that plugin.json, hooks.json, and script references are consistent
+# Usage: bash tests/test-plugin-discovery.sh
+
+set -euo pipefail
+
+PLUGIN_ROOT="$(cd "$(dirname "$0")/.." && pwd)"
+PASS=0
+FAIL=0
+
+pass() { echo -e "\033[0;32m  ✓ $1\033[0m"; PASS=$((PASS + 1)); }
+fail() { echo -e "\033[0;31m  ✗ $1\033[0m"; FAIL=$((FAIL + 1)); }
+
+echo "=== Plugin Discovery Smoke Test ==="
+echo ""
+
+# -------------------------------------------------------
+# Check 1: plugin.json
+# -------------------------------------------------------
+echo "--- Check 1: plugin.json ---"
+
+PLUGIN_JSON="$PLUGIN_ROOT/.claude-plugin/plugin.json"
+
+if [ -f "$PLUGIN_JSON" ]; then
+  pass "plugin.json exists"
+else
+  fail "plugin.json missing at .claude-plugin/plugin.json"
+fi
+
+if grep -q '"auto_discover": true' "$PLUGIN_JSON" 2>/dev/null || grep -q '"auto_discover":true' "$PLUGIN_JSON" 2>/dev/null; then
+  pass "auto_discover: true"
+else
+  fail "auto_discover is not true"
+fi
+
+# plugin.json must NOT have a "hooks" key
+if grep -q '"hooks"' "$PLUGIN_JSON" 2>/dev/null; then
+  fail "plugin.json contains 'hooks' key (should be in hooks/hooks.json only)"
+else
+  pass "plugin.json does not contain 'hooks' key"
+fi
+
+echo ""
+
+# -------------------------------------------------------
+# Check 2: hooks.json
+# -------------------------------------------------------
+echo "--- Check 2: hooks.json ---"
+
+HOOKS_JSON="$PLUGIN_ROOT/hooks/hooks.json"
+
+if [ -f "$HOOKS_JSON" ]; then
+  pass "hooks.json exists"
+else
+  fail "hooks.json missing at hooks/hooks.json"
+fi
+
+if node -e "JSON.parse(require('fs').readFileSync('$HOOKS_JSON', 'utf-8'))" 2>/dev/null; then
+  pass "hooks.json is valid JSON"
+else
+  fail "hooks.json is invalid JSON"
+fi
+
+# Check structure: hooks key is an object
+if node -e "
+  const h = JSON.parse(require('fs').readFileSync('$HOOKS_JSON', 'utf-8'));
+  if (!h.hooks || typeof h.hooks !== 'object') process.exit(1);
+" 2>/dev/null; then
+  pass "hooks.json has 'hooks' object at root"
+else
+  fail "hooks.json missing root 'hooks' object"
+fi
+
+echo ""
+
+# -------------------------------------------------------
+# Check 3: Script references
+# -------------------------------------------------------
+echo "--- Check 3: Script References ---"
+
+# Extract script paths from hooks.json
+SCRIPT_PATHS=$(node -e "
+  const h = JSON.parse(require('fs').readFileSync('$HOOKS_JSON', 'utf-8'));
+  for (const [event, handlers] of Object.entries(h.hooks)) {
+    for (const handler of handlers) {
+      for (const hook of (handler.hooks || [])) {
+        if (hook.command) {
+          const match = hook.command.match(/\\\$\\{CLAUDE_PLUGIN_ROOT\\}\\/(.+)/);
+          if (match) console.log(match[1]);
+        }
+      }
+    }
+  }
+" 2>/dev/null)
+
+for script_path in $SCRIPT_PATHS; do
+  full_path="$PLUGIN_ROOT/$script_path"
+  if [ -f "$full_path" ]; then
+    pass "Script exists: $script_path"
+  else
+    fail "Script missing: $script_path"
+  fi
+done
+
+echo ""
+
+# -------------------------------------------------------
+# Check 4: Event names
+# -------------------------------------------------------
+echo "--- Check 4: Event Names ---"
+
+VALID_EVENTS="SessionStart UserPromptSubmit PreToolUse PostToolUse Stop"
+
+HOOK_EVENTS=$(node -e "
+  const h = JSON.parse(require('fs').readFileSync('$HOOKS_JSON', 'utf-8'));
+  for (const event of Object.keys(h.hooks)) console.log(event);
+" 2>/dev/null)
+
+for event in $HOOK_EVENTS; do
+  if echo "$VALID_EVENTS" | grep -qw "$event"; then
+    pass "Valid event: $event"
+  else
+    fail "Invalid event: $event"
+  fi
+done
+
+echo ""
+
+# -------------------------------------------------------
+# Check 5: Documentation
+# -------------------------------------------------------
+echo "--- Check 5: Documentation ---"
+
+CLAUDE_MD="$PLUGIN_ROOT/CLAUDE.md"
+
+if grep -q "## Hooks" "$CLAUDE_MD" 2>/dev/null; then
+  pass "CLAUDE.md has Hooks section"
+else
+  fail "CLAUDE.md missing Hooks section"
+fi
+
+for script in pre-edit-secrets session-start-context stop-assessment-reminder; do
+  if grep -q "$script" "$CLAUDE_MD" 2>/dev/null; then
+    pass "CLAUDE.md documents $script"
+  else
+    fail "CLAUDE.md missing $script documentation"
+  fi
+done
+
+echo ""
+
+# -------------------------------------------------------
+# Summary
+# -------------------------------------------------------
+TOTAL=$((PASS + FAIL))
+echo "=== Results: $PASS/$TOTAL passed, $FAIL failed ==="
+
+if [ "$FAIL" -gt 0 ]; then
+  exit 1
+fi
--- a/plugins/ms-ai-architect/tests/test-ros-output.sh
+++ b/plugins/ms-ai-architect/tests/test-ros-output.sh
@ -0,0 +1,74 @@
+#!/bin/bash
+# test-ros-output.sh — Validate ros-analysis-agent output structure
+set -euo pipefail
+
+SCRIPT_DIR="$(cd "$(dirname "$0")" && pwd)"
+source "$SCRIPT_DIR/lib/e2e-helpers.sh"
+
+FIXTURE="$SCRIPT_DIR/fixtures/ros-analysis/fixture.md"
+
+if [ ! -f "$FIXTURE" ]; then
+  echo "ERROR: Fixture not found: $FIXTURE"
+  echo "Run: bash tests/capture-fixture.sh to generate fixtures"
+  exit 1
+fi
+
+init_suite "ROS Analysis Agent"
+
+# Structure checks
+assert_has_section "$FIXTURE" "ROS-analyse" "Has ROS analysis header"
+assert_has_section "$FIXTURE" "Risikoregister" "Has risk register section"
+assert_has_section "$FIXTURE" "Risikomatrise" "Has risk matrix section"
+assert_has_section "$FIXTURE" "Tiltaksplan" "Has measures plan section"
+assert_has_section "$FIXTURE" "Restrisiko" "Has residual risk section"
+assert_has_section "$FIXTURE" "Dimensjonsvurdering" "Has dimension assessment section"
+
+# Content quality
+assert_min_lines "$FIXTURE" 60 "Minimum 60 lines"
+assert_min_tables "$FIXTURE" 4 "Minimum 4 tables (register, matrix, dimensions, measures)"
+assert_scores_in_range "$FIXTURE" "ROS scores in valid X/5 range"
+assert_has_dimensions "$FIXTURE" 6 "At least 6 risk dimensions scored"
+
+# Encoding
+assert_encoding_ok "$FIXTURE" "UTF-8 encoding valid"
+assert_no_ascii_approximation "$FIXTURE" "No ASCII approximation of Norwegian chars"
+
+# Domain-specific: methodology references
+assert_matches_pattern "$FIXTURE" "(NS 5814|ISO 31000)" "References NS 5814 or ISO 31000 methodology"
+
+# Domain-specific: threat and risk IDs
+assert_matches_pattern "$FIXTURE" "T-[A-Z]{3}-[0-9]{2}" "Contains threat IDs (T-xxx-NN format)"
+assert_matches_pattern "$FIXTURE" "R-[0-9]" "Contains risk IDs (R-NN format)"
+
+# Domain-specific: risk dimensions
+assert_matches_pattern "$FIXTURE" "(Modellsikkerhet|Dataintegritet|Bias|Tilgjengelighet|Forklarbarhet|Juridisk|Organisatorisk)" "Covers ROS risk dimensions"
+
+# Domain-specific: regulatory references
+assert_matches_pattern "$FIXTURE" "(AI Act|GDPR|OWASP)" "References key regulations/standards"
+
+# Structure: check all 8 phases (Full ROS)
+assert_has_section "$FIXTURE" "Fase 1" "Has Phase 1 header"
+assert_has_section "$FIXTURE" "Fase 2" "Has Phase 2 header"
+assert_has_section "$FIXTURE" "Fase 3" "Has Phase 3 header"
+assert_has_section "$FIXTURE" "Fase 4" "Has Phase 4 header"
+assert_has_section "$FIXTURE" "Fase 5" "Has Phase 5 header"
+assert_has_section "$FIXTURE" "Fase 6" "Has Phase 6 header"
+assert_has_section "$FIXTURE" "Fase 7" "Has Phase 7 header"
+assert_has_section "$FIXTURE" "Fase 8" "Has Phase 8 header"
+assert_has_section "$FIXTURE" "Ledelsessammendrag" "Has executive summary"
+
+# Measure IDs (M-xxx)
+assert_matches_pattern "$FIXTURE" "M-[0-9]" "Contains measure IDs (M-NN format)"
+
+# Minimum threat count for full ROS
+threat_count=$(grep -cE "T-[A-Z]{3}-[0-9]{2}" "$FIXTURE" || echo 0)
+if [ "$threat_count" -ge 8 ]; then
+  pass "Minimum 8 threats identified ($threat_count found)"
+else
+  fail "Minimum 8 threats — only $threat_count found"
+fi
+
+# Vedlegg O coverage (for systems with agents/MCP)
+assert_matches_pattern "$FIXTURE" "(MAESTRO|forsyningskjede|MCP|supply chain)" "References supply chain/MAESTRO (Vedlegg O coverage)"
+
+print_summary
--- a/plugins/ms-ai-architect/tests/test-security-output.sh
+++ b/plugins/ms-ai-architect/tests/test-security-output.sh
@ -0,0 +1,43 @@
+#!/bin/bash
+# test-security-output.sh — Validate security-assessment-agent output structure
+set -euo pipefail
+
+SCRIPT_DIR="$(cd "$(dirname "$0")" && pwd)"
+source "$SCRIPT_DIR/lib/e2e-helpers.sh"
+
+FIXTURE="$SCRIPT_DIR/fixtures/security-assessment/fixture.md"
+
+if [ ! -f "$FIXTURE" ]; then
+  echo "ERROR: Fixture not found: $FIXTURE"
+  echo "Run: bash tests/capture-fixture.sh to generate fixtures"
+  exit 1
+fi
+
+init_suite "Security Assessment Agent"
+
+# Structure checks
+assert_has_section "$FIXTURE" "Sikkerhetsvurdering" "Has security assessment header"
+assert_has_section "$FIXTURE" "Sikkerhetsscoring" "Has scoring section"
+assert_has_section "$FIXTURE" "Kritiske funn" "Has critical findings section"
+assert_has_section "$FIXTURE" "DPIA" "Has DPIA section"
+assert_has_section "$FIXTURE" "ROS-analyse" "Has risk analysis section"
+assert_has_section "$FIXTURE" "Dataklassifisering" "Has data classification section"
+
+# Content quality
+assert_min_lines "$FIXTURE" 40 "Minimum 40 lines"
+assert_min_tables "$FIXTURE" 3 "Minimum 3 tables (scoring, ROS, data classification)"
+assert_scores_in_range "$FIXTURE" "Security scores in valid X/5 range"
+assert_has_dimensions "$FIXTURE" 5 "At least 5 security dimensions scored"
+
+# Encoding
+assert_encoding_ok "$FIXTURE" "UTF-8 encoding valid"
+assert_no_ascii_approximation "$FIXTURE" "No ASCII approximation of Norwegian chars"
+
+# Domain-specific
+assert_matches_pattern "$FIXTURE" "(GDPR|DPIA|personvern)" "References GDPR/DPIA"
+assert_matches_pattern "$FIXTURE" "(AI Act|AI-Act)" "References AI Act"
+assert_matches_pattern "$FIXTURE" "(Schrems II|Schrems)" "References Schrems II"
+assert_matches_pattern "$FIXTURE" "P0|P1|Blokkerende" "Has priority classifications (P0/P1)"
+assert_matches_pattern "$FIXTURE" "(Identity|Network|Data Protection|Content Safety|Compliance|Monitoring)" "Covers standard security dimensions"
+
+print_summary
--- a/plugins/ms-ai-architect/tests/test-summary-output.sh
+++ b/plugins/ms-ai-architect/tests/test-summary-output.sh
@ -0,0 +1,39 @@
+#!/bin/bash
+# test-summary-output.sh — Validate summary-agent output structure
+set -euo pipefail
+
+SCRIPT_DIR="$(cd "$(dirname "$0")" && pwd)"
+source "$SCRIPT_DIR/lib/e2e-helpers.sh"
+
+FIXTURE="$SCRIPT_DIR/fixtures/summary/fixture.md"
+
+if [ ! -f "$FIXTURE" ]; then
+  echo "ERROR: Fixture not found: $FIXTURE"
+  echo "Run: bash tests/capture-fixture.sh to generate fixtures"
+  exit 1
+fi
+
+init_suite "Summary Agent"
+
+# Structure checks
+assert_has_section "$FIXTURE" "Sammendrag" "Has summary header"
+assert_has_section "$FIXTURE" "Beslutningsnotat" "Has decision note section"
+assert_has_section "$FIXTURE" "Nøkkeltall" "Has key figures section"
+
+# Content quality
+assert_min_lines "$FIXTURE" 50 "Minimum 50 lines"
+assert_min_tables "$FIXTURE" 1 "At least 1 table (key figures)"
+assert_has_nok_amounts "$FIXTURE" 3 "At least 3 NOK amounts"
+
+# Encoding
+assert_encoding_ok "$FIXTURE" "UTF-8 encoding valid"
+assert_no_ascii_approximation "$FIXTURE" "No ASCII approximation of Norwegian chars"
+
+# Cross-references
+assert_matches_pattern "$FIXTURE" "S[0-9]" "References other sections (S-numbers)"
+assert_matches_pattern "$FIXTURE" "(Sikkerhet|sikkerhet)" "Cross-references security"
+assert_matches_pattern "$FIXTURE" "(Kostnad|kostnad|TCO)" "Cross-references cost"
+assert_matches_pattern "$FIXTURE" "(Go|NO-GO|anbefaling|Anbefaling)" "Contains recommendation"
+assert_matches_pattern "$FIXTURE" "(Fase|fase)\s*[0-9]" "References implementation phases"
+
+print_summary
--- a/plugins/ms-ai-architect/tests/validate-plugin.sh
+++ b/plugins/ms-ai-architect/tests/validate-plugin.sh
@ -0,0 +1,293 @@
+#!/bin/bash
+# validate-plugin.sh — Static validation for ms-ai-architect plugin
+# Usage: bash tests/validate-plugin.sh
+
+set -euo pipefail
+
+# Colors for output
+RED='\033[0;31m'
+GREEN='\033[0;32m'
+YELLOW='\033[1;33m'
+NC='\033[0m'
+
+PLUGIN_ROOT="$(cd "$(dirname "$0")/.." && pwd)"
+PASS=0
+FAIL=0
+WARN=0
+
+pass() { echo -e "${GREEN}  ✓ $1${NC}"; PASS=$((PASS + 1)); }
+fail() { echo -e "${RED}  ✗ $1${NC}"; FAIL=$((FAIL + 1)); }
+warn() { echo -e "${YELLOW}  ⚠ $1${NC}"; WARN=$((WARN + 1)); }
+
+echo "=== ms-ai-architect Plugin Validation ==="
+echo "Plugin root: $PLUGIN_ROOT"
+echo ""
+
+# -------------------------------------------------------
+# Check 1: Agent Frontmatter
+# -------------------------------------------------------
+echo "--- Check 1: Agent Frontmatter ---"
+
+VALID_MODELS="opus sonnet haiku"
+VALID_COLORS="blue green yellow purple cyan red orange magenta white"
+
+for agent_file in "$PLUGIN_ROOT"/agents/*.md; do
+  [ -f "$agent_file" ] || continue
+  basename_file="$(basename "$agent_file")"
+
+  # Must have --- on line 1
+  first_line="$(head -n 1 "$agent_file")"
+  if [ "$first_line" != "---" ]; then
+    fail "$basename_file: missing frontmatter delimiter (---) on line 1"
+    continue
+  fi
+
+  # Extract frontmatter (between first and second ---)
+  frontmatter="$(sed -n '1,/^---$/{ /^---$/d; p; }' "$agent_file" | sed '1d')"
+  # sed '1d' removes the first --- captured; we actually need lines between first and second ---
+  # Redo: extract lines between line 2 and next ---
+  frontmatter="$(awk 'NR==1{next} /^---$/{exit} {print}' "$agent_file")"
+
+  # Check required fields
+  for field in "name:" "description:" "model:" "color:" "tools:"; do
+    if echo "$frontmatter" | grep -q "^${field}"; then
+      pass "$basename_file: has $field"
+    elif echo "$frontmatter" | grep -q "^  *${field}"; then
+      # indented (part of multiline) - still counts for description
+      pass "$basename_file: has $field"
+    else
+      # description can be multi-line with |
+      if [ "$field" = "description:" ] && echo "$frontmatter" | grep -q "description:"; then
+        pass "$basename_file: has $field"
+      else
+        fail "$basename_file: missing $field"
+      fi
+    fi
+  done
+
+  # Validate model value
+  model_value="$(echo "$frontmatter" | grep "^model:" | sed 's/^model: *//' | tr -d '[:space:]')"
+  if [ -n "$model_value" ]; then
+    model_valid=false
+    for m in $VALID_MODELS; do
+      if [ "$model_value" = "$m" ]; then
+        model_valid=true
+        break
+      fi
+    done
+    if $model_valid; then
+      pass "$basename_file: model '$model_value' is valid"
+    else
+      fail "$basename_file: model '$model_value' is not valid (expected: $VALID_MODELS)"
+    fi
+  fi
+
+  # Validate color value
+  color_value="$(echo "$frontmatter" | grep "^color:" | sed 's/^color: *//' | tr -d '[:space:]')"
+  if [ -n "$color_value" ]; then
+    color_valid=false
+    for c in $VALID_COLORS; do
+      if [ "$color_value" = "$c" ]; then
+        color_valid=true
+        break
+      fi
+    done
+    if $color_valid; then
+      pass "$basename_file: color '$color_value' is valid"
+    else
+      fail "$basename_file: color '$color_value' is not valid (expected: $VALID_COLORS)"
+    fi
+  fi
+
+  # Validate tools is a JSON array (starts with [)
+  tools_line="$(echo "$frontmatter" | grep "^tools:" || true)"
+  if [ -n "$tools_line" ]; then
+    tools_value="$(echo "$tools_line" | sed 's/^tools: *//')"
+    if echo "$tools_value" | grep -q '^\['; then
+      pass "$basename_file: tools is a JSON array"
+    else
+      fail "$basename_file: tools is not a JSON array (got: $tools_value)"
+    fi
+  fi
+done
+
+echo ""
+
+# -------------------------------------------------------
+# Check 2: Command Frontmatter
+# -------------------------------------------------------
+echo "--- Check 2: Command Frontmatter ---"
+
+for cmd_file in "$PLUGIN_ROOT"/commands/*.md; do
+  [ -f "$cmd_file" ] || continue
+  basename_file="$(basename "$cmd_file")"
+  basename_noext="${basename_file%.md}"
+
+  # Must have --- on line 1
+  first_line="$(head -n 1 "$cmd_file")"
+  if [ "$first_line" != "---" ]; then
+    fail "$basename_file: missing frontmatter delimiter (---) on line 1"
+    continue
+  fi
+
+  # Extract frontmatter
+  frontmatter="$(awk 'NR==1{next} /^---$/{exit} {print}' "$cmd_file")"
+
+  # Check required fields: name, description
+  for field in "name:" "description:"; do
+    if echo "$frontmatter" | grep -q "${field}"; then
+      pass "$basename_file: has $field"
+    else
+      fail "$basename_file: missing $field"
+    fi
+  done
+
+  # Check allowed-tools (warn if missing)
+  if echo "$frontmatter" | grep -q "allowed-tools:"; then
+    pass "$basename_file: has allowed-tools"
+  else
+    warn "$basename_file: missing allowed-tools (recommended)"
+  fi
+
+  # Validate name matches filename pattern
+  name_value="$(echo "$frontmatter" | grep "^name:" | sed 's/^name: *//' | tr -d '[:space:]')"
+  if [ -n "$name_value" ] && [ "$name_value" = "$basename_noext" ]; then
+    pass "$basename_file: name matches filename"
+  elif [ -n "$name_value" ]; then
+    fail "$basename_file: name '$name_value' does not match filename '$basename_noext'"
+  fi
+done
+
+echo ""
+
+# -------------------------------------------------------
+# Check 3: Encoding Validation
+# -------------------------------------------------------
+echo "--- Check 3: Encoding Validation ---"
+
+encoding_issues=0
+for dir in agents commands skills; do
+  dir_path="$PLUGIN_ROOT/$dir"
+  [ -d "$dir_path" ] || continue
+
+  while IFS= read -r -d '' mdfile; do
+    basename_file="$(basename "$mdfile")"
+    rel_path="${mdfile#$PLUGIN_ROOT/}"
+
+    # Check for broken UTF-8 sequences
+    if grep -ql 'Ã¦\|Ã¸\|Ã¥\|Ã\†\|Ã˜\|Ã…' "$mdfile" 2>/dev/null; then
+      fail "$rel_path: broken æ/ø/å encoding detected"
+      encoding_issues=$((encoding_issues + 1))
+    fi
+
+    if grep -ql 'â€"' "$mdfile" 2>/dev/null; then
+      fail "$rel_path: broken em-dash/en-dash encoding detected"
+      encoding_issues=$((encoding_issues + 1))
+    fi
+  done < <(find "$dir_path" -name '*.md' -print0)
+done
+
+if [ "$encoding_issues" -eq 0 ]; then
+  pass "No encoding issues found in agents/, commands/, skills/"
+fi
+
+echo ""
+
+# -------------------------------------------------------
+# Check 4: KB Reference Validation
+# -------------------------------------------------------
+echo "--- Check 4: KB Reference Validation ---"
+
+for agent_file in "$PLUGIN_ROOT"/agents/*.md; do
+  [ -f "$agent_file" ] || continue
+  basename_file="$(basename "$agent_file")"
+
+  # Extract lines referencing references/ paths
+  ref_paths="$(grep -o 'references/[a-zA-Z0-9_-]*/\?' "$agent_file" | sort -u || true)"
+
+  if [ -z "$ref_paths" ]; then
+    continue
+  fi
+
+  while IFS= read -r ref_path; do
+    # Normalize: remove trailing slash, build full path relative to skill references
+    ref_dir="$(echo "$ref_path" | sed 's|/$||')"
+    # Check across all skill directories
+    full_path=""
+    for skill_dir in "$PLUGIN_ROOT"/skills/*/; do
+      if [ -d "${skill_dir}${ref_dir}" ]; then
+        full_path="${skill_dir}${ref_dir}"
+        break
+      fi
+    done
+    if [ -z "$full_path" ]; then
+      full_path="$PLUGIN_ROOT/skills/ms-ai-engineering/$ref_dir"
+    fi
+
+    if [ -d "$full_path" ]; then
+      # Check if directory has files
+      file_count="$(find "$full_path" -maxdepth 1 -name '*.md' -type f | wc -l | tr -d ' ')"
+      if [ "$file_count" -gt 0 ]; then
+        pass "$basename_file: $ref_dir/ exists ($file_count files)"
+      else
+        warn "$basename_file: $ref_dir/ exists but is empty"
+      fi
+    else
+      fail "$basename_file: referenced $ref_dir/ does not exist at $full_path"
+    fi
+  done <<< "$ref_paths"
+done
+
+echo ""
+
+# -------------------------------------------------------
+# Check 5: Plugin.json Validation
+# -------------------------------------------------------
+echo "--- Check 5: Plugin.json Validation ---"
+
+plugin_json="$PLUGIN_ROOT/.claude-plugin/plugin.json"
+
+if [ ! -f "$plugin_json" ]; then
+  fail "plugin.json not found at .claude-plugin/plugin.json"
+else
+  pass "plugin.json exists"
+
+  # Check required fields
+  for field in "name" "version" "description"; do
+    if grep -q "\"$field\"" "$plugin_json"; then
+      pass "plugin.json: has \"$field\""
+    else
+      fail "plugin.json: missing \"$field\""
+    fi
+  done
+
+  # Check auto_discover: true
+  if grep -q '"auto_discover"' "$plugin_json"; then
+    auto_val="$(grep '"auto_discover"' "$plugin_json" | grep -o 'true\|false')"
+    if [ "$auto_val" = "true" ]; then
+      pass "plugin.json: auto_discover is true"
+    else
+      fail "plugin.json: auto_discover is not true (got: $auto_val)"
+    fi
+  else
+    fail "plugin.json: missing auto_discover"
+  fi
+fi
+
+echo ""
+
+# -------------------------------------------------------
+# Summary
+# -------------------------------------------------------
+echo "=== Results ==="
+echo -e "${GREEN}PASS: $PASS${NC}"
+echo -e "${RED}FAIL: $FAIL${NC}"
+echo -e "${YELLOW}WARN: $WARN${NC}"
+
+if [ $FAIL -gt 0 ]; then
+  echo -e "${RED}VALIDATION FAILED${NC}"
+  exit 1
+else
+  echo -e "${GREEN}VALIDATION PASSED${NC}"
+  exit 0
+fi