🏛️ EU Parliament Monitor — Architecture

C4 Model Architecture for European Parliament Intelligence Platform
📐 System Context • 📦 Container View • 🔧 Component Design

📋 Document Owner: CEO | 📄 Version: 1.2 | 📅 Last Updated: 2026-04-20 (UTC) | 📦 Release: v0.8.40
🔄 Review Cycle: Quarterly | ⏰ Next Review: 2026-07-20

---

✅ April-2026 Aggregator-Pipeline Migration — Complete

The April-2026 release migrated from an AI-authored-HTML pipeline to a deterministic aggregator pipeline. Article HTML is now rendered by src/aggregator/** from committed Stage-B analysis artifacts — there is no AI-authored HTML step, no per-article-type strategies, no AI_MARKER/FALLBACK_TEMPLATE sentinel contract, and no src/utils/content-validator.ts / validate-articles.ts / validate-analysis-completeness.ts runtime validators.

Canonical references for the current release:

• 🟢 Render entry point: src/aggregator/article-generator.ts (CLI: npm run generate-article -- --run <analysis-run-dir>)
• 📦 Aggregator modules: artifact-order.ts, clean-artifact.ts, analysis-aggregator.ts, markdown-renderer.ts, article-html.ts, article-metadata.ts (5-tier editorial-highlight resolver for <title> / <meta description> — manifest override → first artefact H1 → aggregated H1 → first strong prose → localized template)
• 🤖 Agentic workflows: 8 unified news-<type>.md files (Stages A → B → C → D → E in one session) + news-translate.md; the split-family workflows (news-<type>-analysis.md + news-<type>-article.md) and the manual news-article-generator.md helper were deleted
• 💰 Economic-context enforcement: editorial Stage-C agent-side review over intelligence/economic-context.md (the Wave-2 OR-gate and Wave-3/Wave-4 strict runtime gates in src/utils/content-validator.ts were purged with the rest of the validator layer; enforcement moved to the Stage-C completeness review protocol in .github/prompts/03-analysis-completeness-gate.md and the depth floors in analysis/methodologies/reference-quality-thresholds.json)

The C4 Container and Component diagrams in this document have been rewritten against the post-migration aggregator stack — they no longer reference deleted strategies/builders/content-validator.ts.

This document serves as the primary entry point for the EU Parliament Monitor's architectural documentation. It provides a comprehensive view of the system's design using the C4 model approach, starting from a high-level system context and drilling down to component interactions.

📚 Architecture Documentation Map

🛡️ ISMS Policy Alignment

EU Parliament Monitor is developed and maintained in accordance with Hack23 AB's Information Security Management System (ISMS), which is aligned with ISO 27001:2022, NIST CSF 2.0, and CIS Controls v8.1.

Applicable ISMS Policies

Policy	Description	Relevance to EU Parliament Monitor
Information Security Policy	Establishes organization-wide security governance and risk management framework	Defines overall security posture, risk assessment methodology, and management responsibilities for the project
Secure Development Policy	Defines secure coding standards, code review requirements, and SDLC security gates	Mandates security-first development practices: input validation, dependency scanning, SAST/DAST integration, secure CI/CD pipelines
Open Source Policy	Governs use, contribution, and licensing of open source software	Ensures compliance with Apache-2.0 License, dependency license compatibility, and transparent open source contribution practices
Classification Policy	Defines data classification scheme (Public, Internal, Confidential, Restricted) and handling requirements	All project content classified as PUBLIC; establishes data handling controls for any future sensitive data integration
AI Policy	Governs responsible AI usage, transparency, and human oversight requirements	Governs LLM usage for content generation: transparency requirements, human review workflows, bias mitigation, prompt injection protection
Access Control Policy	Defines authentication, authorization, least privilege, and privileged access management	Controls GitHub repository access, branch protection rules, secret management, and deployment permissions
Cryptography Policy	Establishes cryptographic standards for data protection (algorithms, key management, TLS)	Mandates HTTPS-only content delivery, TLS 1.2+ (TLS 1.3 where supported) for outbound HTTPS API communications; EP MCP integration uses a local stdio JSON-RPC channel (no TLS); ensures secure secret storage for LLM API keys

ISMS Compliance Implementation

ISO 27001:2022 Controls Implemented:

• A.5.10 - Information Security Policy (documented and reviewed quarterly)
• A.8.3 - Secure Coding (ESLint security rules, CodeQL SAST scanning)
• A.8.23 - Web Filtering (planned CSP headers via CloudFront, XSS prevention)
• A.8.24 - Cryptography (HTTPS-only, TLS 1.2+ / TLS 1.3 where supported, site delivery via CloudFront)
• A.8.28 - Secure Coding (input validation, dependency scanning)

NIST CSF 2.0 Functions Addressed:

• Identify (ID): Asset inventory, risk assessment, vulnerability management
• Protect (PR): Access control, data security, secure development
• Detect (DE): Security monitoring, vulnerability scanning, anomaly detection
• Respond (RS): Incident response procedures, GitHub Security Advisories
• Recover (RC): Business Continuity Plan, backup/restore procedures

CIS Controls v8.1 Implemented:

• Control 1: Inventory and Control of Enterprise Assets (documented in repo)
• Control 4: Secure Configuration (branch protection, security policies)
• Control 6: Access Control Management (GitHub RBAC, least privilege)
• Control 8: Audit Log Management (GitHub audit logs, workflow logs)
• Control 10: Malware Defenses (Dependabot, npm audit, CodeQL)
• Control 16: Application Software Security (SAST, dependency scanning, secure coding)

Compliance Evidence

Evidence of ISMS compliance is maintained through:

• Policy Documents: All policies stored in Hack23/ISMS-PUBLIC
• Security Architecture: SECURITY_ARCHITECTURE.md maps controls to implementations
• Threat Model: THREAT_MODEL.md documents STRIDE analysis and mitigations
• Classification: CLASSIFICATION.md defines data classification and handling
• Audit Trail: GitHub audit logs, workflow execution logs, dependency scan reports
• Security Scanning: CodeQL results, Dependabot alerts, npm audit reports

---

🎯 System Overview

EU Parliament Monitor is a TypeScript-first static site generator and political intelligence platform that creates multi-language news articles about European Parliament activities. Content is produced by a fleet of 9 agentic GitHub Workflows (gh-aw — 8 unified news-<type>.md + news-translate.md) that drive AI agents (Claude Opus 4.7 via GitHub Copilot) through the Stage A→E protocol, consuming structured data from three data surfaces:

• European Parliament MCP Server v1.2.15+ (primary — 60+ tools including plenary, MEPs, votes, committees, procedures, adopted texts, sliding-window + fixed-window feeds, analytical tools, and a three-state voting fallback to the EP Open Data Portal)
• World Bank Open Data MCP (non-economic only — WDI social/health/education/environment/governance indicators)
• IMF REST (SDMX 3.0) native TypeScript fetch client — primary economic source: WEO + Fiscal Monitor + IFS + BOP + ER + PCPS + GFSR + EREO + FSI + GFS + DOT

TypeScript code handles data acquisition, analysis orchestration, HTML structure, and validation; AI agents author all narrative content under a strict two-pass AI-First Quality regime.

Mission Statement

Enable democratic transparency by providing automated, multilingual coverage of European Parliament activities through a secure, maintainable static site architecture.

Key Characteristics

• Minimal Runtime Dependencies: Pure static HTML/CSS output with no server-side execution; one pinned production dependency (european-parliament-mcp-server@1.2.15) plus one optional dependency (worldbank-mcp@1.0.1) used only at build time; markdown-it + plugins (markdown-it-anchor, markdown-it-footnote, markdown-it-attrs, markdown-it-deflist) vendored in the aggregator for deterministic artifact rendering
• TypeScript Source: All source in src/ written in TypeScript 6.0.3 (strict, ESM, "type": "module"), compiled via tsc — rootDir: ./src, outDir: ./scripts, target: ES2025, module: NodeNext
• Multi-Language Support: Generates content in 14 languages (en, sv, da, no, fi, de, fr, es, nl, ar, he, ja, ko, zh), defined in src/constants/language-core.ts::ALL_LANGUAGES
• Article Types: 8 production content types (breaking, committee-reports, month-ahead, month-in-review, motions, propositions, week-ahead, week-in-review) — each type is a slug, not a strategy module; the aggregator renders the same canonical artifact order for every type and per-type content differences are carried by the Stage-B artifacts themselves
• Agentic Workflows: 9 unified gh-aw markdown workflows — 8 news-<type>.md article types (Stages A → B → C → D → E in one session, active-work budget 22–27 min before the single safe-outputs create_pull_request call, 75-min hard timeout) + news-translate.md (14-language flush translation, exempt from the single-PR rule) — compiled to .lock.yml via gh aw compile --validate (pinned GH_AW_VERSION: v0.69.0)
• Analysis-Artifact-Driven Article Pipeline: Agents author the full Stage-B artifact set under analysis/daily/<date>/<slug>-run<NN>/ and commit it. The deterministic aggregator (src/aggregator/**, invoked via npm run generate-article -- --run <analysis-run-dir> for a single run or npm run generate-article:all for batch regen) walks manifest.json, cleans each artifact, and emits the final HTML with the shared site chrome (stacked header + embedded 14-language switcher + TOC sidebar + footer stats) and 14-language hreflang entries. There is no AI-authored HTML step, no strategies, no builders, no section-builders
• Economic Data (IMF-primary, Wave-4 strict default editorial): IMF REST is the primary source for every economic claim in intelligence/economic-context.md; World Bank MCP provides complementary non-economic context only. Enforcement is editorial at the Stage-C completeness review — the legacy runtime gates (articlePolicyHasEconomicContext, articlePolicyHasIMFEconomicEvidence, isWave3IMFStrictEnabled) in src/utils/content-validator.ts were purged in April-2026; the Stage-C reviewer applies the IMF-required-for-policy rule directly over the committed artifact
• Quality-Through-Artifact Principle: Mandatory 2-pass iterative improvement during Stage B (~60% pass 1, ~40% pass 2); ≥ 80 words/SWOT item, ≥ 150 words/stakeholder perspective, ≥ 1 Mermaid or Chart.js visualisation per core artifact, 0 [AI_ANALYSIS_REQUIRED] sentinel markers in any committed file (enforced at Stage-C agent-side review against reference-quality-thresholds.json)
• MCP Integration: Spawned as local child processes via stdio JSON-RPC at build time; inside agentic workflows via the awmg gateway at http://host.docker.internal:8080/mcp/european-parliament
• Security by Design: Minimal attack surface through static architecture; 5-layer gh-aw security (AWF Squid firewall allowlist, sandboxed Docker, safe-output constraints, JSONL audit trail, lock file compilation); agent prose-injection class of defects eliminated at the root by the aggregator migration (no AI-authored HTML step means no template-prose leak vector)
• AWS Hosted: AWS S3 + CloudFront (primary, via deploy-s3.yml with OIDC auth); GitHub Pages retained as documented fallback; npm package published to registry.npmjs.org/euparliamentmonitor with SLSA Level 3 provenance

---

📊 C4 Model Level 1: System Context Diagram

👤 User Focus: Shows how different user types interact with the EU Parliament Monitor system and what external systems it depends on.

🌐 Integration Focus: Illustrates the relationships with GitHub infrastructure, European Parliament APIs, and LLM services.

C4Context
    title EU Parliament Monitor - System Context Diagram

    Person(citizen, "European Citizen", "Reads news about European Parliament activities in their native language")
    Person(journalist, "Journalist", "Uses site as research source for European political coverage")
    Person(researcher, "Political Researcher", "Analyzes EP activities and trends")
    Person(contributor, "Developer/Contributor", "Maintains and improves the news generation system")

    System(epmonitor, "EU Parliament Monitor", "Static site with multilingual news about European Parliament activities")

    System_Ext(github, "GitHub", "Hosts repository, runs CI/CD (GitHub Actions)")
    System_Ext(aws, "AWS (S3 + CloudFront)", "Serves static site globally via CDN")
    System_Ext(ep_mcp, "European Parliament MCP Server", "Provides structured access to EP data")
    System_Ext(ep_api, "European Parliament APIs", "Official EP data sources (plenary, committees, documents)")
    System_Ext(llm, "LLM Service", "Generates article content from structured EP data")

    Rel(citizen, epmonitor, "Reads news", "HTTPS")
    Rel(journalist, epmonitor, "Researches stories", "HTTPS")
    Rel(researcher, epmonitor, "Analyzes data", "HTTPS")
    Rel(contributor, github, "Contributes code", "Git/HTTPS")

    Rel(epmonitor, github, "Built and deployed via", "GitHub Actions")
    Rel(epmonitor, aws, "Hosted on", "S3 + CloudFront")
    Rel(github, epmonitor, "Generates site via", "GitHub Actions")
    Rel(epmonitor, ep_mcp, "Fetches EP data via", "MCP Protocol")
    Rel(ep_mcp, ep_api, "Queries EP data", "HTTPS/JSON")
    Rel(epmonitor, llm, "Generates content via", "API/SDK")

    UpdateLayoutConfig($c4ShapeInRow="3", $c4BoundaryInRow="2")
Copy

Context Diagram - Key Elements

Element	Type	Description	Technology
European Citizen	User	Primary audience seeking EP news in native language	Web Browser
Journalist	User	Professional using site for research and story development	Web Browser
Political Researcher	User	Academic or analyst studying EP activities	Web Browser
Developer/Contributor	User	Maintainer improving system	Git, Node.js, VS Code
EU Parliament Monitor	System	Core static site generator	Node.js, TypeScript
GitHub	External System	Source control, CI/CD	GitHub Actions
EP MCP Server	External System	Structured EP data access	MCP Protocol, TypeScript
EP APIs	External System	Official data sources	REST APIs, JSON
LLM Service	External System	Content generation	API (OpenAI/Anthropic/etc.)

Trust Boundaries and Security Zones

graph TB
    subgraph "Public Internet - Untrusted Zone"
        Users["Web Users\nCitizens, Journalists, Researchers"]
    end
    
    subgraph "GitHub Infrastructure - Trusted Zone"
        subgraph "Build Environment"
            Actions["GitHub Actions Runner\nGitHub-hosted Ubuntu runner\nubuntu-latest + Node.js 25"]
            EPServer["European Parliament\nMCP Server\nLocal process, stdio JSON-RPC"]
        end
        
        subgraph "Source Control"
            Repo["Git Repository\nVersion Control"]
        end
    end
    
    subgraph "AWS Hosting - Cloud Infrastructure Zone"
        Pages["AWS S3 + CloudFront CDN\nHTTPS via ACM"]
    end
    
    subgraph "External Services - Partially Trusted Zone"
        EPAPI["European Parliament\nOfficial APIs"]
        LLM["LLM Service\nOpenAI/Anthropic"]
    end
    
    Users -->|"HTTPS GET\nRead-Only"| Pages
    Actions -->|"Spawns locally\nstdio JSON-RPC"| EPServer
    EPServer -->|"HTTPS/JSON\nData Queries"| EPAPI
    Actions -->|"API Calls\nContent Gen"| LLM
    Actions -->|"Git Push\nAuthenticated"| Repo
    Actions -->|"S3 Sync + CF Invalidation\nAuthenticated via OIDC"| Pages
    
    classDef users fill:#CE93D8,stroke:#6A1B9A,stroke-width:2px,color:#000000
    classDef hosting fill:#A5D6A7,stroke:#2E7D32,stroke-width:2px,color:#000000
    classDef actions fill:#90CAF9,stroke:#1565C0,stroke-width:2px,color:#000000
    classDef external fill:#FFE082,stroke:#F57C00,stroke-width:2px,color:#000000

    class Users users
    class Pages hosting
    class Actions actions
    class EPServer,EPAPI,LLM external
Copy

Trust Boundary Analysis:

Zone	Trust Level	Security Controls	Threat Model
Public Internet	Untrusted	HTTPS-only, planned CSP headers, static content only	DDoS, XSS attempts (mitigated by static architecture)
GitHub Infrastructure	Trusted	GitHub authentication, branch protection, optional signed commits, secret scanning	Supply chain attacks (mitigated by Dependabot, CodeQL)
AWS Hosting	Trusted	ACM certificate, HTTPS redirect, DDoS protection via CloudFront	Hosting infrastructure compromise (mitigated by AWS security controls, OIDC deploy auth)
External Services	Partially Trusted	API authentication, basic input parsing/shape validation; planned systematic sanitization/escaping and rate limiting	Data poisoning, API compromise (mitigated by validation, monitoring, planned hardening)

Key Security Boundaries:

1. User → CloudFront: Read-only HTTPS access, no authentication required (public content)
2. GitHub Actions → External APIs: Authenticated API calls, input validation, error handling
3. GitHub Actions → AWS S3: Authenticated S3 sync + CloudFront invalidation, only static files deployed
4. External Services → System: Data parsed and basic shape-validated before use; comprehensive sanitization/escaping and rate limiting are planned controls

---

📦 C4 Model Level 2: Container Diagram

📦 Container Focus: Major containers (applications, data stores, MCP clients) of the post-April-2026 aggregator pipeline.

🔄 Data Flow Focus: How agentic workflows produce analysis artifacts and how the deterministic aggregator renders them into 14-language HTML.

%%{init: {"theme":"dark","themeVariables":{"primaryColor":"#1565C0","primaryTextColor":"#fff","lineColor":"#90CAF9","fontFamily":"Inter, Helvetica, Arial, sans-serif"}}}%%
C4Container
    title EU Parliament Monitor — Container Diagram (April-2026 aggregator pipeline)

    Person(user, "Reader", "Reads multilingual EP news at euparliamentmonitor.com")
    Person(contributor, "Contributor", "Maintains code, methodologies, translations")
    Person(researcher, "Researcher / Journalist", "Audits analysis artifacts via the Political Intelligence Hub")

    Container_Boundary(epmonitor, "EU Parliament Monitor") {
        Container(aw_orchestrator, "gh-aw Orchestrator", "Agentic Workflows (Claude Opus 4.7)", "9 agentic workflows: 8 unified news-<type>.md + news-translate.md")
        Container(prompt_lib, "Prompt Library", "10 bounded contexts", ".github/prompts/00-scope … 09-troubleshooting; lint:prompts drift-guard")
        Container(methodology_lib, "Methodology Library", "Markdown methodologies + JSON thresholds", "17 methodologies + reference-quality-thresholds.json (analysis/methodologies/)")
        Container(template_lib, "Template Library", "51 Markdown templates", "39 core + 12 extended (analysis/templates/)")
        ContainerDb(analysis_runs, "Analysis Runs", "Markdown + JSON", "analysis/daily/YYYY-MM-DD/<type>/{manifest.json,intelligence/,classification/,risk-scoring/,threat-assessment/,documents/,extended/}")
        Container(aggregator, "Aggregator (5 modules)", "TypeScript", "src/aggregator/**: artifact-order · clean-artifact · analysis-aggregator · markdown-renderer · article-html · article-metadata · article-generator (CLI)")
        Container(ep_client, "EP MCP Client", "TypeScript", "Stdio JSON-RPC to european-parliament-mcp-server@1.2.15+; 60+ tools; getVotingRecordsWithFallback() to EP Open Data Portal (src/mcp/ep-mcp-client.ts)")
        Container(wb_client, "World Bank MCP Client", "TypeScript", "WORLD_BANK_MCP_TOOLS — non-economic indicators only (src/mcp/wb-mcp-client.ts)")
        Container(imf_client, "IMF REST Client", "TypeScript", "IMF_MCP_TOOLS — primary economic source: WEO/Fiscal Monitor/IFS/BOP/ER/PCPS (src/mcp/imf-mcp-client.ts)")
        Container(stage_c_review, "Stage-C Review", "Editorial agent + thresholds", "Reads .github/prompts/03-analysis-completeness-gate.md + reference-quality-thresholds.json — replaces purged content-validator.ts")
        Container(news_indexes, "News Indexes & Sitemap", "TypeScript", "Per-language index pages + sitemap.xml + sitemap_<lang>.html (src/generators/news-indexes.ts, sitemap.ts)")
        ContainerDb(static_files, "Static Site Output", "HTML/CSS/JS/JSON", "news/<slug>-<lang>.html (14 langs) · news/<slug>.en.md · article.md per run · sitemap.xml · articles-metadata.json")
    }

    Container_Boundary(github_infra, "GitHub Infrastructure") {
        Container(actions, "GitHub Actions", "CI/CD + gh-aw runtime", "9 news + ~15 standard workflows; SHA-pinned actions; OpenSSF Scorecard")
        ContainerDb(repo, "Git Repository", "Version control", "Source + analysis runs + generated content; SLSA L3 provenance")
    }

    Container_Boundary(aws_infra, "AWS Infrastructure") {
        Container(cf_s3, "CloudFront + S3", "CDN / object storage", "Primary hosting · ACM HTTPS · OIDC GithubWorkFlowRole · cache: HTML 1h, immutable assets 1y")
    }

    System_Ext(ep_mcp, "European Parliament MCP Server v1.2.15+", "60+ tools — plenary, voting, motions, committee, MEPs, declarations, procedures, analytical (voting-anomaly, coalition, MEP-influence)")
    System_Ext(ep_open_data, "EP Open Data Portal", "https://data.europarl.europa.eu — voting-records fallback (/api/v2/decision)")
    System_Ext(wb_mcp, "World Bank Open Data MCP", "Non-economic WDI indicators (health, education, environment, governance)")
    System_Ext(imf_api, "IMF SDMX 3.0 REST", "https://dataservices.imf.org/REST/SDMX_3.0/")
    System_Ext(copilot, "GitHub Copilot / Claude Opus 4.7", "Authors analysis Markdown under 2-pass AI-First Quality regime — never authors HTML")

    Rel(user, cf_s3, "Reads HTML in 14 langs", "HTTPS")
    Rel(researcher, repo, "Audits analysis/daily/", "Git/HTTPS")
    Rel(contributor, repo, "Commits code + methodologies", "Git/HTTPS")
    Rel(actions, aw_orchestrator, "Triggers on schedule / manual", "gh-aw engine")
    Rel(aw_orchestrator, copilot, "Delegates analysis authoring", "Copilot CLI")
    Rel(aw_orchestrator, prompt_lib, "Imports prompts", "Markdown")
    Rel(aw_orchestrator, methodology_lib, "Reads methodologies", "Markdown")
    Rel(aw_orchestrator, template_lib, "Fills templates", "Markdown")
    Rel(aw_orchestrator, ep_client, "Stage A — fetch", "fn")
    Rel(aw_orchestrator, wb_client, "Stage A — context (optional)", "fn")
    Rel(aw_orchestrator, imf_client, "Stage A — economic context", "fn")
    Rel(aw_orchestrator, analysis_runs, "Stage B — write artifacts", "fs.write")
    Rel(aw_orchestrator, stage_c_review, "Stage C — completeness gate", "agent review")
    Rel(stage_c_review, analysis_runs, "Reads + grades", "fn")
    Rel(aw_orchestrator, aggregator, "Stage D — npm run generate-article", "CLI")
    Rel(aggregator, analysis_runs, "Reads manifest.json + artifacts", "fs.read")
    Rel(aggregator, static_files, "Writes 14 HTML + Markdown", "fs.write")
    Rel(news_indexes, static_files, "Writes index pages", "fs.write")
    Rel(ep_client, ep_mcp, "stdio JSON-RPC", "MCP")
    Rel(ep_client, ep_open_data, "Voting fallback", "HTTPS REST")
    Rel(wb_client, wb_mcp, "stdio JSON-RPC", "MCP")
    Rel(imf_client, imf_api, "HTTPS / SDMX", "REST")
    Rel(static_files, repo, "Stage E — single PR", "Git")
    Rel(actions, cf_s3, "Deploy via OIDC", "S3 sync + CloudFront invalidation")

    UpdateLayoutConfig($c4ShapeInRow="3", $c4BoundaryInRow="2")
Copy

Container Diagram — Key Elements

🧱 Container	⚙️ Technology	🎯 Purpose	🔄 Data flow
🤖 gh-aw Orchestrator	Claude Opus 4.7 + gh-aw v0.69.0+	Runs 9 agentic workflows; produces analysis artifacts	Triggers via cron / manual; commits one PR per run
📚 Prompt / Methodology / Template libraries	Markdown + JSON	Bounded-context prompts (10), methodologies (17), templates (51)	Read by every agentic workflow at start-of-session
🧠 Analysis Runs	Markdown + JSON	Per-run intelligence tree under analysis/daily/<date>/<type>/	Written by Stage B agents; consumed by Stage C and aggregator
🟢 Aggregator (5 modules)	TypeScript	Reads manifest.json and Markdown artifacts; renders 14-language HTML deterministically	npm run generate-article -- --run <dir>
🔌 EP MCP Client	TypeScript	60+ EP tools + voting fallback to EP Open Data Portal /api/v2/decision	Stage A data collection
💰 IMF / 🌱 World Bank Clients	TypeScript	Economic context (IMF) + non-economic indicators (WB)	Stage A wave-2 context
⚖️ Stage-C Review	Editorial agent + JSON thresholds	Per-artifact line floors + tradecraft signals (Admiralty / WEP / ICD-203)	Replaces the purged content-validator.ts runtime gate
🌐 News Indexes & Sitemap	TypeScript	14-language index pages, sitemap.xml, hreflang alternates	npm run prebuild
📦 Static Site Output	HTML / CSS / JS / JSON / Markdown	Public deliverable: news pages + article.md source per run	Committed to main, deployed to S3
🚀 GitHub Actions	CI/CD + gh-aw	9 news + ~15 standard workflows	Daily news + on-PR validation
☁️ CloudFront + S3	CDN / object storage	Primary hosting via OIDC GithubWorkFlowRole	HTTPS + immutable asset cache

Security Responsibilities per Container

Container	Security responsibility	Implementation	Controls
🤖 gh-aw Orchestrator	Sandboxed AWF runtime, Squid egress allowlist, capability-bounded safe outputs	Runs in GitHub-hosted ephemeral VMs; safe-outputs.create-pull-request.max: 1; step-security/harden-runner egress block	A.5.10, A.8.28 (ISO 27001), CIS 16
🟢 Aggregator	Deterministic Markdown→HTML; explicit markdown-it plugin allowlist; clean-artifact.ts strips SPDX/banners; script-src 'self' CSP	No AI-authored HTML; vendored Mermaid/Chart.js/D3 under js/vendor/	A.8.23, A.8.28 (ISO 27001), OWASP A03
🔌 EP MCP Client	Local stdio JSON-RPC; per-request timeout + retry backoff; envelope validation; voting-records three-state fallback	safeCallTool() + callToolWithRetry() wrappers in ep-mcp-client.ts	A.8.24 (ISO 27001), CIS 16
🧠 Analysis Artifacts	Tradecraft grading (Admiralty A1–F6 + WEP) + provenance manifest	manifest.json cross-reference map; methodology-reflection.md audit	A.5.12 (ISO 27001)
📦 Static Files	Public-data only; integrity via Git + SLSA L3 attestation	All EP/IMF/WB content is public; SBOM, REUSE licence headers	A.5.10 (ISO 27001)
🚀 GitHub Actions	OIDC for AWS deploy; Secrets at job-scope; SHA-pinned third-party actions	GithubWorkFlowRole IAM with least privilege; harden-runner egress allowlist	A.8.3, CIS 6
☁️ CloudFront + S3	HTTPS-only via ACM; bucket policy denies public ACLs; CloudFront cache-control by file class	Long-cache immutable assets, short-cache HTML	A.13.1, A.5.23 (ISO 27001)
Amazon CloudFront + S3	HTTPS-only, CDN security, DDoS protection	Forces HTTPS redirect via ACM certificate, CloudFront with DDoS mitigation, HSTS headers (configured externally in CloudFront distribution)	A.8.24 (ISO 27001)
Git Repository	Access control, branch protection, signed commits	RBAC with least privilege, protected main branch, optional signed commits	CIS Control 6, A.8.3

Container Security Architecture

graph TB
    subgraph "Generation Layer - Build Time Security"
        NewsGen["News Generator\nInput Validation\nData Sanitization"]
        MCPClient["MCP Client\nLocal stdio JSON-RPC\nConnection Retry\nRequest Timeout"]
        Template["Template Engine\nXSS Prevention\nCSP Generation\nHTML Sanitization"]
    end
    
    subgraph "Storage Layer - Version Control Security"
        GitRepo["Git Repository\nBranch Protection\nCode Review\nAudit Logs"]
        Secrets["GitHub Secrets\nEncrypted Storage\nLeast Privilege"]
    end
    
    subgraph "Delivery Layer - Runtime Security"
        Pages["Amazon CloudFront + S3\nHTTPS-Only\nHSTS Headers\nDDoS Protection"]
        CDN["CloudFront Edge\nTLS Termination\nEdge Caching\nGeographic Distribution"]
    end
    
    subgraph "External Layer - Third-Party Security"
        EPMCP["EP MCP Server\nMCP Protocol\nData Validation"]
        LLM["LLM Service\nAPI Key Auth\nPrompt Injection Prevention"]
    end
    
    NewsGen -->|Validated Data| Template
    NewsGen -->|"Spawns locally via stdio"| MCPClient
    MCPClient -->|JSON-RPC| EPMCP
    NewsGen -->|Secured API Calls| LLM
    Template -->|Safe HTML| GitRepo
    Secrets -->|Inject at Runtime| NewsGen
    GitRepo -->|Deploy to S3| Pages
    Pages -->|Cached Content| CDN
    
    classDef generation fill:#90CAF9,stroke:#1565C0,stroke-width:2px,color:#000000
    classDef storage fill:#A5D6A7,stroke:#2E7D32,stroke-width:2px,color:#000000
    classDef delivery fill:#FFCC80,stroke:#F57C00,stroke-width:2px,color:#000000
    classDef external fill:#CE93D8,stroke:#6A1B9A,stroke-width:2px,color:#000000

    class NewsGen,Template,MCPClient generation
    class GitRepo,Secrets storage
    class Pages,CDN delivery
    class EPMCP,LLM external
Copy

---

🧩 C4 Model Level 3: Component Diagram — Aggregator Pipeline

🔧 Component Focus: Internal components of the deterministic aggregator (src/aggregator/**) and supporting MCP / methodology modules.

🎯 Responsibility Focus: How analysis Markdown artifacts produced by the agentic workflows become 14-language HTML deliverables.

%%{init: {"theme":"dark","themeVariables":{"primaryColor":"#1565C0","primaryTextColor":"#fff","lineColor":"#90CAF9","fontFamily":"Inter, Helvetica, Arial, sans-serif"}}}%%
C4Component
    title EU Parliament Monitor — Aggregator Components (post-April-2026)

    Container_Boundary(aggregator_c, "Aggregator (src/aggregator/)") {
        Component(article_generator, "article-generator.ts", "TypeScript CLI", "Entry point: npm run generate-article -- --run <dir>; walks manifest.json")
        Component(artifact_order, "artifact-order.ts", "TypeScript", "ARTIFACT_SECTIONS — canonical 19-section order")
        Component(clean_artifact, "clean-artifact.ts", "TypeScript", "Strips SPDX/banner/provenance front matter from each artifact before merge")
        Component(analysis_aggregator, "analysis-aggregator.ts", "TypeScript", "aggregateAnalysisRun() — filters manifestFiles to .md only excluding data/runs/pass1; emits Provenance & Audit block at END")
        Component(markdown_renderer, "markdown-renderer.ts", "TypeScript", "markdown-it + plugins (anchor, footnote, attrs, deflist); explicit allowlist; renderMarkdown()")
        Component(article_html, "article-html.ts", "TypeScript", "HTML5 wrapper: stacked header, language switcher, TOC sidebar, JSON-LD NewsArticle, isBasedOn provenance, hreflang alternates, footer")
        Component(article_metadata, "article-metadata.ts", "TypeScript", "5-tier editorial-highlight resolver for <title>/<meta description>: manifest override → first-artifact H1 → aggregated H1 → first strong prose → localized template")
    }

    Container_Boundary(mcp_c, "MCP & Data Clients (src/mcp/)") {
        Component(ep_client, "ep-mcp-client.ts", "TypeScript", "60+ tools; safeCallTool + callToolWithRetry; recess-mode detection; slow-feed warnings")
        Component(ep_open_data, "ep-open-data-client.ts", "TypeScript", "EPOpenDataClient + getVotingRecordsWithFallback() three-state fallback")
        Component(wb_client, "wb-mcp-client.ts", "TypeScript", "WORLD_BANK_MCP_TOOLS — non-economic indicators")
        Component(imf_client, "imf-mcp-client.ts", "TypeScript", "class IMFMCPClient + IMF_MCP_TOOLS; native fetch SDMX 3.0; primary economic source")
        Component(mcp_health, "mcp-health.ts / mcp-retry.ts / mcp-connection.ts", "TypeScript", "Health probes, retry backoff, connection lifecycle")
    }

    Container_Boundary(intel_c, "Intelligence Utilities (src/utils/, src/generators/)") {
        Component(political_classification, "political-classification.ts", "TypeScript", "7-dimension EP event classification")
        Component(political_threat, "political-threat-assessment.ts", "TypeScript", "5-framework political threat (Landscape 6D + Attack Trees + Kill Chain + Diamond + ICO)")
        Component(political_risk, "political-risk-assessment.ts", "TypeScript", "5×5 Likelihood × Impact scoring")
        Component(significance, "significance-scoring.ts", "TypeScript", "Publication priority score per artifact")
        Component(quality_scorer, "article-quality-scorer.ts", "TypeScript", "Editorial quality signals")
        Component(news_indexes, "news-indexes.ts + sitemap.ts", "TypeScript", "14-language indexes + sitemap.xml + per-language sitemap_<lang>.html")
    }

    Container_Boundary(scripts_c, "Workflow Scripts (scripts/aggregator/)") {
        Component(prior_run_diff, "prior-run-diff.js", "Node.js", "Re-run merge helper; carry-forward vs rewrite classification; ENABLE_PRIOR_RUN_MERGE flag")
        Component(forward_statements, "forward-statements-registry.js", "Node.js", "Forward-looking-statement JSONL registry; week/month-ahead seeding")
        Component(checkpoint, "checkpoint-analysis-to-memory.sh", "Bash", "Pre-audited helper; replaces inline expansion-heavy bash in workflows (shell-safety)")
    }

    System_Ext(ep_mcp, "EP MCP Server v1.2.15+", "60+ tools — plenary, voting, motions, committee, MEPs, declarations, procedures, analytical")
    System_Ext(ep_portal, "EP Open Data Portal", "/api/v2/decision — voting fallback")
    System_Ext(wb_mcp, "World Bank Open Data MCP", "Non-economic WDI")
    System_Ext(imf_api, "IMF SDMX 3.0", "WEO / FM / IFS / BOP / ER / PCPS")
    ContainerDb(analysis_dir, "analysis/daily/<date>/<type>/", "Markdown + JSON", "manifest.json + intelligence/ + classification/ + risk-scoring/ + threat-assessment/ + extended/")
    ContainerDb(news_dir, "news/<slug>(-<lang>).{md,html}", "Markdown + HTML", "Per-language deliverables")

    Rel(article_generator, analysis_dir, "reads manifest.json", "fs.readFileSync")
    Rel(article_generator, artifact_order, "uses ARTIFACT_SECTIONS", "import")
    Rel(article_generator, clean_artifact, "cleans each artifact", "fn")
    Rel(article_generator, analysis_aggregator, "aggregateAnalysisRun()", "fn")
    Rel(analysis_aggregator, markdown_renderer, "renderMarkdown()", "fn")
    Rel(markdown_renderer, article_html, "wraps in HTML5 chrome", "fn")
    Rel(article_html, article_metadata, "5-tier metadata resolver", "fn")
    Rel(article_html, news_dir, "writes 14 HTML + 1 .md", "fs.writeFileSync")
    Rel(news_indexes, news_dir, "writes index pages + sitemaps", "fs.writeFileSync")

    Rel(ep_client, ep_mcp, "stdio JSON-RPC", "MCP")
    Rel(ep_open_data, ep_portal, "voting fallback", "HTTPS")
    Rel(wb_client, wb_mcp, "stdio JSON-RPC", "MCP")
    Rel(imf_client, imf_api, "HTTPS/SDMX", "REST")
    Rel(ep_client, mcp_health, "health + retry", "fn")

    Rel(prior_run_diff, analysis_dir, "carry-forward plan", "JSON")
    Rel(forward_statements, analysis_dir, "JSONL registry seeding", "fs")

    UpdateLayoutConfig($c4ShapeInRow="3", $c4BoundaryInRow="2")
Copy

Component Diagram — Key Elements

🧩 Component	🎯 Responsibility	🔗 Dependencies	📂 File location
🟢 Aggregator pipeline	Discover manifest.json → clean artifacts → aggregate (19-section canonical order, Provenance & Audit at end, .md only excluding data/runs/pass1/) → render Markdown → wrap HTML with TOC sidebar + shared chrome → write <slug>.en.md + 14 <slug>-<lang>.html	markdown-it + markdown-it-anchor/-footnote/-attrs/-deflist	src/aggregator/{article-generator,analysis-aggregator,markdown-renderer,article-html,artifact-order,clean-artifact,article-metadata}.ts
🧠 Analysis artifacts	51 templates per run (39 core + 12 extended) under analysis/daily/<date>/<type>/ with manifest.json declaring articleType + files map. 3-variant manifest schema (articleType / articleTypes[] / legacy runType) handled by resolveArticleTypeFromManifest()	17 methodologies (10-step protocol, Rules 1–22)	analysis/methodologies/*.md, analysis/templates/**, analysis/daily/**
🔌 EP MCP Client	60+ EP tools via stdio JSON-RPC; safeCallTool() + callToolWithRetry() wrappers; recess-mode detection ([1952,2100] year window); slow-feed warning downgrade for get_events_feed	european-parliament-mcp-server@1.2.15+ (PR #405 normalises political-group codes)	src/mcp/ep-mcp-client.ts
🗳️ EP Open Data fallback	Three-state voting fallback: (a) MCP has data → use it · (b) MCP empty → query /api/v2/decision · (c) both empty → 🔴 unavailability marker via virtual tool name ep-get-voting-records	EP Open Data Portal	src/mcp/ep-open-data-client.ts (see getVotingRecordsWithFallback())
💰 IMF Client	class IMFMCPClient + IMF_MCP_TOOLS; primary economic source per IMF Indicator Mapping; native Node 25 fetch SDMX 3.0; env IMF_API_BASE_URL, IMF_API_TIMEOUT_MS	None (REST)	src/mcp/imf-mcp-client.ts
🌱 World Bank Client	WORLD_BANK_MCP_TOOLS; non-economic WDI indicators only (health, education, environment, governance, innovation)	worldbank-mcp (optional)	src/mcp/wb-mcp-client.ts
⚖️ Stage-C completeness gate	Editorial agent-side review against .github/prompts/03-analysis-completeness-gate.md and analysis/methodologies/reference-quality-thresholds.json line floors. Replaces the purged runtime content-validator.ts	Methodology library + per-artifact thresholds	.github/prompts/03-…, analysis/methodologies/reference-quality-thresholds.json
🔁 Prior-Run Diff	Re-run merge helper; classifies existing artifacts as at-floor (carry-forward) or below-floor (rewrite); ENABLE_PRIOR_RUN_MERGE=true env flag; emits priorRunDiff JSON consumed by Stage B	—	scripts/aggregator/prior-run-diff.js
📜 Forward-statements registry	Canonical last-occurrence-per-id JSONL registry; week/month-ahead seeds data/forward-statements-open.json; Stage C enforces a "carried-forward forward statements" section when open items exist	JSONL registry	scripts/aggregator/forward-statements-registry.js, analysis/forward-statements/
🛡️ Shell-safety helper	Pre-audited bash helper for checkpoint-to-memory; replaces expansion-heavy inline workflow bash that the sandbox shell-safety filter would block	Bash	scripts/checkpoint-analysis-to-memory.sh
🧠 Intelligence utilities	political-classification (7D), political-threat-assessment (5-framework), political-risk-assessment (5×5 L×I), significance-scoring, article-quality-scorer	Types	src/utils/*.ts
🌐 News Indexes & Sitemap	Per-language news index pages, sitemap.xml, per-language sitemap_<lang>.html, hreflang alternates	Metadata, file-utils	src/generators/news-indexes.ts, src/generators/sitemap.ts
🔢 Constants	ALL_LANGUAGES (14), LANGUAGE_PRESETS (all, eu-core, nordic), article-type slugs, committee indicator map	—	src/constants/*.ts

Component Interaction Patterns

sequenceDiagram
    autonumber
    participant CLI as CLI Interface
    participant Gen as Article Generator
    participant MCP as MCP Client
    participant EPMCP as EP MCP Server
    participant Tmpl as HTML Template
    participant Meta as Metadata Manager
    participant FS as File System Writer
    
    CLI->>Gen: generate(type, languages)
    Gen->>MCP: fetchEPData(type)
    MCP->>EPMCP: query(endpoint, params)
    EPMCP-->>MCP: return EP data
    MCP-->>Gen: return parsed EP data
    
    loop For each language (sequential)
        Gen->>Tmpl: renderHTML(epData, lang)
        Note over Gen,Tmpl: Current: placeholder English content<br/>Future (ADR-004): native LLM generation per language
        Tmpl-->>Gen: return HTML
        Gen->>FS: writeFile(path, html)
        Gen->>Meta: recordGeneration(article, lang)
    end
    
    Meta->>FS: writeMetadata(json)
    Gen-->>CLI: generation complete
Copy

Component Collaboration Patterns

Pattern	Components Involved	Purpose	Error Handling
Cache-Aside (Planned)	MCP Client → LRU Cache → EP MCP Server	Reduce API calls, improve performance	Planned: cache miss triggers fresh fetch; current: direct calls to EP MCP Server
MCP Connection Retry with Backoff (Current)	MCP Client → EP MCP Server	Handle transient MCP connection failures	Connection attempts retried with backoff; individual MCP requests use a fixed timeout and are not retried
Validation Pipeline (Planned)	Content Validator → Article Generator	Ensure content quality	Planned: failed validation triggers regeneration (max 2 attempts); current: single-pass generation without regeneration loop
Sequential Multi-Language	Article Generator → HTML Template (per language)	Content generation per language	Current: failure in one language aborts remaining languages; Planned: per-language failures logged while other languages still generate; parallel generation planned (ADR-004)
Template Method	Article Generator → HTML Template → File System Writer	Consistent HTML generation	Template errors logged and propagated to prevent partial writes
Metadata Aggregation	Metadata Manager → File System Writer	Track generation history	Current: metadata written synchronously via writeFileSync; failures throw and fail the run. Planned: non-blocking, best-effort writes

---

🔄 Deployment Diagram

☁️ Infrastructure Focus: Shows how the system is deployed on GitHub infrastructure.

🚀 CI/CD Focus: Illustrates the automated deployment pipeline.

C4Deployment
    title EU Parliament Monitor - Deployment Diagram

    Deployment_Node(github_cloud, "GitHub Cloud", "GitHub Infrastructure") {
        Deployment_Node(actions_runner, "GitHub Actions Runner", "Ubuntu 24.04") {
            Container(workflow, "News Generation Workflow", "GitHub Actions YAML", "Daily scheduled workflow")
            Container(node_runtime, "Node.js Runtime", "Node.js 25", "Executes generation scripts")
        }

        Deployment_Node(pages_cdn, "AWS Infrastructure", "S3 + CloudFront") {
            Container(web_server, "Amazon CloudFront", "CDN / HTTPS", "Serves HTTPS content globally")
            ContainerDb(static_content, "Amazon S3 Bucket", "Object Storage", "Generated articles and pages")
        }

        Deployment_Node(repo_storage, "GitHub Repository", "Git Storage") {
            ContainerDb(git_repo, "Git Repository", "Version Control", "Source code and generated content")
        }
    }

    Deployment_Node(user_device, "User Device", "Desktop/Mobile") {
        Container(browser, "Web Browser", "Chrome/Firefox/Safari", "Renders news articles")
    }

    Deployment_Node(external_services, "External Services", "Cloud") {
        System_Ext(ep_mcp, "EP MCP Server", "EP data access")
        System_Ext(llm, "LLM Service", "Content generation")
    }

    Rel(workflow, node_runtime, "Executes", "Process")
    Rel(node_runtime, ep_mcp, "Fetches data", "stdio/JSON-RPC")
    Rel(node_runtime, llm, "Generates content", "HTTPS/API")
    Rel(node_runtime, git_repo, "Commits files", "Git")
    Rel(git_repo, static_content, "Deploys via", "S3 sync + CloudFront invalidation")
    Rel(browser, web_server, "Requests pages", "HTTPS")
    Rel(web_server, static_content, "Serves", "HTTP/2")

    UpdateLayoutConfig($c4ShapeInRow="2", $c4BoundaryInRow="1")
Copy

Deployment - Key Infrastructure

Infrastructure Component	Technology	Purpose	Configuration
GitHub Actions Runner	ubuntu-latest, Node.js 25	Execute generation workflow	.github/workflows/news-*.lock.yml
Amazon CloudFront	AWS CDN	Serve static content globally	CloudFront distribution (deploy-s3.yml)
Amazon S3	AWS Object Storage	Host static site files	S3 bucket (deploy-s3.yml)
Git Repository	GitHub Storage	Version control + content storage	public repository
Web Browser	Modern browsers	Render news articles	HTML5, CSS3, ES6+
EP MCP Server	Local Node process	EP data access	Spawned locally via stdio JSON-RPC
LLM Service	External API	Content generation	API key authentication

Article Types & Workflows

8 production article types are driven by 8 unified news-<type>.md workflows (Stage A→E in one ~45-min session, single PR per run). Article HTML is rendered deterministically by src/aggregator/article-generator.ts from committed Stage-B analysis artifacts — there are no per-type strategy modules in the post-April-2026 pipeline.

🏷️ Article Type	🤖 gh-aw Workflow	📅 Cadence
🚨 breaking	news-breaking.md	Every 6 hours
🔮 week-ahead	news-week-ahead.md	Fri 07:00 UTC
📋 week-in-review	news-week-in-review.md	Sat 09:00 UTC
📊 month-ahead	news-month-ahead.md	1st of month 08:00 UTC
📈 month-in-review	news-month-in-review.md	28th of month 10:00 UTC
🏛️ committee-reports	news-committee-reports.md	Mon–Fri 04:00 UTC
🗳️ motions	news-motions.md	Mon–Fri 06:00 UTC
⚖️ propositions	news-propositions.md	Mon–Fri 05:00 UTC

Plus: news-translate.md (14-language translation helper, manual dispatch only).

Agentic Workflows (gh-aw)

All 9 news workflows are markdown source files compiled to YAML (.md → .lock.yml) via the GitHub Agentic Workflows CLI (gh aw compile --validate) with pinned GH_AW_VERSION: v0.69.0 in .github/workflows/compile-agentic-workflows.yml. See WORKFLOWS.md for the full surface.

5-layer security model:

1. AWF Squid firewall allowlist — egress HTTP allowlist per workflow
2. Sandboxed Docker with restricted shell — bash tool-call contract (every call requires command + description); shell expansion restrictions
3. Safe-output constraints — create-pull-request with max-patch-size (default 1024 KB; news-translate.md sets 10240 KB at top level for 14-language fan-out)
4. JSONL audit trail — per-step structured logs
5. Lock file compilation — .lock.yml is the immutable executed artifact; .md is the human source under review

MCP gateway (containerised): EP_MCP_GATEWAY_URL=http://host.docker.internal:80/mcp/european-parliament, provisioned by scripts/mcp-setup.sh.

Validator gates (Stage-C completeness review, agent-side — replaces purged runtime validators):

• .github/prompts/03-analysis-completeness-gate.md — protocol that the editorial agent runs
• analysis/methodologies/reference-quality-thresholds.json — per-artifact line floors
• Dynamic file resolution pattern (must not hallucinate file names): ls -t "news/${TODAY}-${TYPE}"*"-en.html" | head -1

---

📊 Technology Stack

Core Technologies

Layer	Technology	Version	Purpose	Rationale
Runtime	Node.js	25.x (engines: >=25); Node.js 26 LTS migration scheduled upon release (~Apr 2026)	JavaScript execution environment	Current release for latest features, performance improvements; ESM-native ("type": "module")
Language	TypeScript	6.0.3	Primary development language	Strict type safety; compiles from src/ → scripts/ targeting ES2025, module: NodeNext
Package Manager	npm	10.x	Dependency management	Native Node.js package manager, security audit integration
Testing	Vitest	4.1.4	Unit and integration testing	Fast, ESM-native; happy-dom env; happy-dom@20.9.0
E2E Testing	Playwright	1.59.1	End-to-end browser testing	@axe-core/playwright@4.11.2 for WCAG 2.1 AA
Linting	ESLint	10.2.1	Code quality and security	Flat config; plugins: eslint-plugin-sonarjs@4.0.3, eslint-plugin-security@4.0.0, eslint-plugin-jsdoc@62.9.0
Formatting	Prettier	3.8.3	Code formatting	Opinionated formatter, consistent code style
Visualization	Chart.js	4.5.1	Dashboard charts in articles	Vendored into js/vendor/ via npm run copy-vendor
Visualization	D3	7.9.0	Advanced visualizations	Used for specific intelligence views
Documentation	TypeDoc	0.28.19	API documentation generation	Generates docs/ pages from TypeScript sources
HTML Validation	HTMLHint	1.9.2	HTML5 validation	Pre-commit + CI
Duplicate Check	jscpd	4.0.9	Copy-paste detection	Scheduled quality audits

Technology Version Matrix

Technology	Current Version	Minimum Version	End-of-Life	Update Policy
Node.js	25.x (current)	25.0.0 (engines: >=25)	~Apr 2026 (Current EOL; upgrading to Node.js 26 LTS)	Update to Node.js 26 LTS within days of release (~Apr 2026)
npm	10.x (latest)	10.0.0	Follows Node.js lifecycle	Auto-updated with Node.js
TypeScript	6.0.3	6.0.0	N/A	Update to latest minor within 14 days, major within 90 days
Vitest	4.1.4	4.0.0	N/A	Update to latest minor within 14 days, major within 60 days
Playwright	1.59.1	1.55.0	N/A	Update to latest minor within 14 days, major within 60 days
ESLint	10.2.1	10.0.0	N/A	Update to latest minor within 14 days, major within 90 days
Prettier	3.8.3	3.0.0	N/A	Update to latest minor within 14 days, major within 90 days
Chart.js	4.5.1	4.0.0	N/A	Vendored; update with copy-vendor script
D3	7.9.0	7.0.0	N/A	Vendored; update with copy-vendor script
TypeDoc	0.28.19	0.28.0	N/A	Major within 60 days
european-parliament-mcp-server	1.2.13 (pinned)	1.2.13	Per upstream	Track releases; 1.2.11 (2026-04-20) fixes #377/#378 (fixed-window feeds, uniform unavailable envelope); 1.2.13 (2026-04-23) adds non-retryable UPSTREAM_404 for get_procedures, fixes search_documents envelope, enriches track_legislation timeline, improves get_procedures_feed error classification
worldbank-mcp	1.0.1 (optional)	1.0.0	Per upstream	Biannual WDI refresh cadence
gh-aw CLI	v0.69.0 (pinned GH_AW_VERSION)	v0.69.0	Per upstream	Workflow-level pin in compile-agentic-workflows.yml

Dependency Management

Production Dependencies (1 required + 1 optional):

• european-parliament-mcp-server@1.2.15 — Primary data surface; 6 sliding-window feed tools (timeframe + startDate when custom) and 7 fixed-window feed tools (limit/offset only — documents, plenary_documents, committee_documents, plenary_session_documents, parliamentary_questions, corporate_bodies, controlled_vocabularies); returns uniform {status:"unavailable", items:[]} envelope on upstream failure.
• worldbank-mcp@1.0.1 (optionalDependencies) — WDI macro/social/environment/health indicators.

IMF REST is integrated via native TypeScript fetch in src/mcp/imf-mcp-client.ts (class IMFMCPClient) — this is NOT an MCP server; calls go directly to https://dataservices.imf.org/REST/SDMX_3.0/. Env: IMF_API_BASE_URL, IMF_API_TIMEOUT_MS. Supplies WEO + FM monthly forecasts up to five years ahead.

Dev dependencies (notable): vitest@4.1.4, @vitest/ui, @vitest/coverage-v8, happy-dom@20.9.0, @playwright/test@1.59.1, @axe-core/playwright@4.11.2, typescript@6.0.3, eslint@10.2.1, eslint-plugin-sonarjs@4.0.3, eslint-plugin-security@4.0.0, eslint-plugin-jsdoc@62.9.0, prettier@3.8.3, htmlhint@1.9.2, typedoc@0.28.19, chart.js@4.5.1, d3@7.9.0, papaparse@5.5.3, husky@9.1.7, jscpd@4.0.9.

Security & Quality

Tool	Purpose	Integration	Configuration
CodeQL	SAST scanning	GitHub Actions (weekly + PR)	.github/workflows/codeql.yml
Dependabot	Dependency vulnerability scanning	GitHub native (daily)	.github/dependabot.yml
npm audit	Dependency security check	Pre-commit + CI	package.json scripts
ESLint Security	Security-focused linting	Pre-commit + CI	eslint.config.js (security plugin)
HTMLHint	HTML validation	CI pipeline	.htmlhintrc
Husky	Git hooks	Pre-commit, pre-push	.husky/ directory
Playwright	Accessibility testing	E2E test suite	playwright.config.js (axe integration)

Infrastructure

Service	Purpose	Configuration	Cost
GitHub Actions	CI/CD automation	.github/workflows/	Free (public repo)
AWS S3	Static site hosting	S3 bucket + static website	Pay-per-use (storage, requests)
Amazon CloudFront	Content delivery	CloudFront distribution (S3)	Pay-per-use (data transfer, requests)
Git	Version control	Repository	Free (public repo)

External Services

Service	Purpose	Protocol	Authentication	Rate Limits	Cost Model
European Parliament MCP Server	EP data access	Local process (stdio JSON-RPC)	None (local process)	N/A (handled by MCP server / EP APIs)	Free (EP open data via MCP server)
LLM Service (OpenAI/Anthropic)	Content generation	HTTPS/JSON	API key (required)	Varies by provider	Pay-per-token
GitHub API	Repository operations	REST/GraphQL	GitHub token	5000 req/hr	Free (authenticated)

Browser Support

Browser	Minimum Version	Features Required	Testing Coverage
Chrome/Edge	90+	ES2020, CSS Grid, Flexbox	✅ Playwright E2E (Chromium in CI)
Firefox	88+	ES2020, CSS Grid, Flexbox	🧪 Manual regression (no Playwright CI)
Safari	14+	ES2020, CSS Grid, Flexbox	🧪 Manual regression (no Playwright CI)
Mobile Chrome	90+	ES2020, Responsive Design	🧪 Manual responsive testing
Mobile Safari	14+	ES2020, Responsive Design	🧪 Manual responsive testing

No support for:

• Internet Explorer (EOL June 2022)
• Legacy Edge (Chromium-based only)

Build Process

TypeScript source in src/ is compiled to JavaScript in scripts/ via tsc. The generated JavaScript files are executed by Node.js during news generation. The public npm entry point is src/index.ts (published as euparliamentmonitor with SLSA Level 3 provenance attestations).

src/                                   → scripts/                          (tsc compilation)
├── index.ts                           → index.js                          npm package entry point
├── constants/                         → constants/
│   ├── config.ts                      Project paths, BASE_URL, filename patterns
│   ├── analysis-constants.ts          Shared analysis thresholds
│   ├── committee-indicator-map.ts     Committee → indicator mapping
│   ├── language-core.ts               ALL_LANGUAGES (14), LANGUAGE_PRESETS
│   ├── language-articles.ts           Per-language article-type labels
│   ├── language-ui.ts                 Per-language UI strings
│   └── languages.ts                   Language metadata (name, flag, direction)
├── mcp/                               → mcp/
│   ├── ep-mcp-client.ts               EP MCP stdio client; feed option types (no canonical EP_MCP_TOOLS export yet)
│   ├── wb-mcp-client.ts               World Bank MCP client; exports WORLD_BANK_MCP_TOOLS
│   ├── imf-mcp-client.ts              IMFMCPClient class (native fetch/SDMX 3.0); exports IMF_MCP_TOOLS
│   ├── mcp-connection.ts              Connection lifecycle
│   ├── mcp-health.ts                  Health probes
│   └── mcp-retry.ts                   Exponential backoff retry
├── templates/                         → templates/
│   ├── article-template.ts            HTML5 article shell (SEO, JSON-LD, Open Graph)
│   └── section-builders.ts            buildSiteFooter (single source of truth, 14-lang), stakeholder grid
├── aggregator/                        → aggregator/  ⭐ April-2026 deterministic article renderer
│   ├── article-generator.ts           Entry point CLI (`npm run generate-article`)
│   ├── analysis-aggregator.ts         aggregateAnalysisRun() — manifest discovery, .md filter, Provenance & Audit at END
│   ├── artifact-order.ts              ARTIFACT_SECTIONS — canonical 19-section order
│   ├── clean-artifact.ts              Strips SPDX/banner/provenance front matter
│   ├── markdown-renderer.ts           markdown-it + plugin allowlist (anchor, footnote, attrs, deflist)
│   ├── article-html.ts                HTML5 wrapper: header, language switcher, TOC sidebar, JSON-LD, hreflang
│   └── article-metadata.ts            5-tier editorial-highlight resolver for <title> / <meta description>
├── generators/                        → generators/  (post-aggregator-migration: only indexes & sitemap remain)
│   ├── news-indexes.ts                Per-language index pages
│   └── sitemap.ts                     XML sitemap generator + per-language sitemap_<lang>.html
├── types/                             → types/
│   ├── analysis.ts, common.ts, generation.ts, imf.ts, intelligence.ts, mcp.ts,
│   │   parliament.ts, political-classification.ts, political-risk.ts,
│   │   political-threats.ts, quality.ts, significance.ts, stakeholder.ts,
│   │   visualization.ts, world-bank.ts, index.ts
└── utils/                             → utils/
    ├── article-category.ts, article-quality-scorer.ts, content-metadata.ts,
    ├── file-utils.ts, html-sanitize.ts, imf-data.ts,
    ├── intelligence-analysis.ts, intelligence-index.ts, metadata-utils.ts,
    ├── news-metadata.ts, political-classification.ts,
    ├── political-risk-assessment.ts, political-threat-assessment.ts,
    ├── significance-scoring.ts, world-bank-data.ts
    (content-validator.ts, validate-articles.ts, validate-analysis-completeness.ts
     PURGED in April-2026 — replaced by Stage-C agent-side review)
Copy

Key build / generation commands:

• npm run build — Runs tsc (TypeScript compilation src/ → scripts/)
• npm run lint — ESLint on src/
• npm run generate-news — Orchestrates strategies via the pipeline
• npm run generate-news-indexes — Executes scripts/generators/news-indexes.js (prebuild hook)
• npm run generate-sitemap — Executes scripts/generators/sitemap.js (prebuild hook)
• npm run copy-vendor — Vendors chart.js and d3 assets into js/vendor/
• npm run test / test:unit / test:integration / test:e2e / test:coverage — Test suite (52 test files, 3061+ passing tests)

TypeScript configuration (tsconfig.json):

• target: ES2025, module: NodeNext, strict: true, rootDir: ./src, outDir: ./scripts, "type": "module" in package.json

Runtime JS (browser):

• js/index-runtime.js — Index page filter + theme toggle
• js/article-runtime.js — Reading progress + theme toggle
• External scripts only (no inline scripts — CSP-ready)
• js/vendor/ — Vendored Chart.js (4.5.1) and D3 (7.9.0)

TypeScript configuration (tsconfig.json):

• target: ES2025 — Modern JavaScript output
• module: NodeNext — Node.js native ESM resolution
• strict: true — Full strict mode enabled
• rootDir: ./src — TypeScript source root
• outDir: ./scripts — Compiled JavaScript output

---

🔄 Data Flow

News Generation Flow

sequenceDiagram
    participant GHA as GitHub Actions
    participant CLI as CLI Interface
    participant Gen as Article Generator
    participant MCP as MCP Client
    participant EP as EP MCP Server
    participant TPL as Template Engine
    participant FS as File System

    GHA->>CLI: Trigger daily workflow
    CLI->>Gen: generate-news --types=week-ahead --languages=all
    Gen->>MCP: getPlenarySessions
    Note over MCP,EP: MCP client spawns EP MCP Server as local process via stdio JSON-RPC
    MCP->>EP: JSON-RPC request via stdio
    EP-->>MCP: EP data as JSON-RPC response
    MCP-->>Gen: Parsed EP data with basic shape checks

    loop For each language sequentially
        Gen->>TPL: Render HTML with EP data and language
        Note over Gen,TPL: Placeholder English body content - native per-language LLM generation planned
        TPL-->>Gen: HTML output
        Gen->>FS: Write article file
    end

    Gen->>FS: Write metadata.json via writeFileSync
    GHA->>GHA: Commit and push changes
    GHA->>GHA: Deploy to S3 + invalidate CloudFront
Copy

User Request Flow

sequenceDiagram
    participant User as User Browser
    participant CDN as CloudFront CDN
    participant S3 as Amazon S3
    participant Repo as Git Repository

    User->>CDN: GET /index.html
    CDN->>S3: Forward request (cache miss)
    S3-->>CDN: HTML response
    CDN-->>User: Cached HTML

    User->>CDN: GET /news/week-ahead-2026-02-17-en.html
    CDN-->>User: Cached article (or fetch from S3)
Copy

---

🔀 Cross-Cutting Concerns

Cross-cutting concerns are aspects of the system that affect multiple components and layers. These concerns are implemented consistently across the entire architecture.

Logging Strategy

Logging Levels:

Level	Usage	Output	Retention
ERROR	Unrecoverable errors (API failures, file write errors)	console.error(), GitHub Actions logs	90 days (GitHub)
WARN	Recoverable issues (MCP connection retry/backoff, MCP tool fallback, JSON.parse recovery)	console.warn(), GitHub Actions logs	90 days (GitHub)
INFO	Normal operations (generation start/complete, article count)	console.log(), GitHub Actions logs	90 days (GitHub)
DEBUG	Detailed diagnostics (API responses, intermediate data)	Disabled in production	Dev only

Structured Logging Format:

{
  timestamp: "2026-02-20T10:30:00.000Z",
  level: "INFO",
  component: "ArticleGenerator",
  action: "generate_article",
  language: "en",
  article_type: "week-ahead",
  duration_ms: 1234,
  status: "success"
}
Copy

Logging Implementation:

• Build Logs: All GitHub Actions workflow logs (generation, deployment, tests)
• Error Tracking: Errors logged to GitHub Actions workflow logs for visibility
• Performance Metrics: Generation time per article, API call durations
• Audit Trail: Git commit history serves as audit log for all content changes

Monitoring and Observability

graph TB
    subgraph "Generation Monitoring"
        Workflow[GitHub Actions Workflow]
        GenMetrics[Generation Metrics<br/>Article count, Duration, Errors]
        TestResults[Test Results<br/>Unit, Integration, E2E]
    end
    
    subgraph "Application Monitoring"
        Pages[Amazon CloudFront + S3]
        Analytics[Web Analytics<br/>Visits, Bounce Rate, Countries]
        Uptime[Uptime Monitoring<br/>AWS Health Dashboard]
    end
    
    subgraph "Security Monitoring"
        Dependabot[Dependabot Alerts]
        CodeQL[CodeQL Security Scans]
        Audit[npm audit]
    end
    
    subgraph "Alerting"
        Email[Email Notifications]
        GitHubUI[GitHub UI Alerts]
        Status[Status Checks]
    end
    
    Workflow -->|Logs| GenMetrics
    Workflow -->|Results| TestResults
    Pages -->|Metrics| Analytics
    Pages -->|Health| Uptime
    Dependabot -->|Alerts| Email
    CodeQL -->|Findings| GitHubUI
    Audit -->|Vulnerabilities| Status
    
    GenMetrics -->|Failures| Email
    TestResults -->|Failures| Status
    
    style Dependabot fill:#f99,stroke:#333,stroke-width:2px
    style CodeQL fill:#f99,stroke:#333,stroke-width:2px
Copy

Monitoring Tools:

Metric	Tool	Threshold	Alert
Build Success Rate	GitHub Actions	<95% over 7 days	Email to maintainers
Generation Duration	Workflow logs	>15 minutes	Warning annotation
Test Pass Rate	Vitest + Playwright	<100%	Block merge
Security Vulnerabilities	Dependabot + CodeQL	Any high/critical	Email + PR
Site Availability	AWS Health Dashboard	<99.9%	AWS Health event notification
Page Load Time	Lighthouse (manual runs)	>3 seconds	Warning annotation

Error Handling

Error Handling Strategy:

flowchart TD
    Start([API Call / Operation])
    Try{Try Operation}
    Success[✅ Success]
    Catch{Catch Error}
    Transient{Transient<br/>Error?}
    Retry[Retry with<br/>Exponential Backoff]
    MaxRetries{Max Retries<br/>Reached?}
    Fallback{Fallback<br/>Available?}
    UseFallback[Use Fallback Data]
    LogError[Log Error]
    PropagateError[Propagate Error]
    GracefulDegradation[Graceful Degradation]
    
    Start --> Try
    Try -->|Success| Success
    Try -->|Error| Catch
    Catch --> Transient
    Transient -->|Yes| Retry
    Transient -->|No| Fallback
    Retry --> MaxRetries
    MaxRetries -->|No| Try
    MaxRetries -->|Yes| Fallback
    Fallback -->|Yes| UseFallback
    Fallback -->|No| LogError
    UseFallback --> GracefulDegradation
    LogError --> PropagateError
    
    style Success fill:#9f9,stroke:#333,stroke-width:2px
    style LogError fill:#f99,stroke:#333,stroke-width:2px
    style PropagateError fill:#f99,stroke:#333,stroke-width:2px
    style GracefulDegradation fill:#ff9,stroke:#333,stroke-width:2px
Copy

Error Categories and Handling:

Error Category	Examples	Retry Strategy	Fallback	User Impact
Transient Network Errors	MCP connection failure during startup, LLM API rate limit	Exponential backoff (1s, 2s, 4s), max 3 retries for MCP connection establishment and LLM calls; individual MCP requests use a single fixed timeout with no retry	Use placeholder events or skip affected items (no cache)	Missing or placeholder content for affected items
Permanent API Errors	Invalid API key, malformed request	No retry	Skip article generation for affected language	Missing article for specific language
Data Validation Errors	Invalid EP data structure, missing required fields	No automatic regeneration loop	Skip invalid items (no cached-data fallback)	Missing content for invalid items
File System Errors	Disk full, permission denied	No retry	Fail workflow	Build failure (no deployment)
Content Generation Errors	LLM refusal, prompt injection detected	Single generation attempt (no automatic regeneration loop)	Insert placeholder events when content generation fails	Reduced content quality or placeholder content

Error Propagation:

1. Component Level: Catch and log errors, attempt recovery
2. Service Level: Propagate if unrecoverable, aggregate errors for reporting
3. Workflow Level: Fail fast if critical (file system), continue if non-critical (single article failure)

Internationalization (i18n)

14 Languages Supported:

• 🇬🇧 English (en) - 67 million
• �🇪 Swedish (sv) - 10 million
• 🇩🇰 Danish (da) - 6 million
• 🇳🇴 Norwegian (no) - 5 million
• 🇫🇮 Finnish (fi) - 5 million
• 🇩🇪 German (de) - 95 million
• 🇫🇷 French (fr) - 67 million
• 🇪🇸 Spanish (es) - 47 million
• 🇳🇱 Dutch (nl) - 24 million
• 🇸🇦 Arabic (ar) - 420 million
• 🇮🇱 Hebrew (he) - 9 million
• 🇯🇵 Japanese (ja) - 125 million
• 🇰🇷 Korean (ko) - 77 million
• 🇨🇳 Chinese (zh) - 1.3 billion

i18n Architecture:

graph LR
    subgraph "Content Generation"
        EPData[EP Data<br/>Language-Neutral]
        LLM[LLM Service]
        Prompt[Language-Specific Prompt]
    end
    
    subgraph "14 Language Variants"
        EN[English Article]
        SV[Swedish Article]
        DA[Danish Article]
        NO[Norwegian Article]
        FI[Finnish Article]
        DE[German Article]
        FR[French Article]
        ES[Spanish Article]
        NL[Dutch Article]
        AR[Arabic Article]
        HE[Hebrew Article]
        JA[Japanese Article]
        KO[Korean Article]
        ZH[Chinese Article]
    end
    
    subgraph "Delivery"
        Index[Language-Specific<br/>Index Pages]
        Sitemap[Multilingual<br/>Sitemap.xml]
    end
    
    EPData --> LLM
    Prompt --> LLM
    LLM --> EN
    LLM --> SV
    LLM --> DA
    LLM --> NO
    LLM --> FI
    LLM --> DE
    LLM --> FR
    LLM --> ES
    LLM --> NL
    LLM --> AR
    LLM --> HE
    LLM --> JA
    LLM --> KO
    LLM --> ZH
    
    EN --> Index
    DE --> Index
    FR --> Index
    ES --> Index
    Index --> Sitemap
    
    style EPData fill:#9cf,stroke:#333,stroke-width:2px
    style LLM fill:#fc9,stroke:#333,stroke-width:2px
Copy

i18n Implementation:

Aspect	Implementation	Example
Content Generation	Placeholder English content for all languages (current); native LLM per-language generation planned (ADR-004)	Current: shared English body with localized titles/subtitles; Future: each article written directly in target language
File Naming	Language suffix in filename	week-ahead-2026-02-17-en.html, week-ahead-2026-02-17-de.html
HTML lang Attribute	Set per page	<html lang="en">, <html lang="de">
Navigation	Language-specific index pages	index.html, index-de.html
SEO	hreflang tags for alternate languages	<link rel="alternate" hreflang="de" href="...">
Date Formatting	Locale-specific date formats	EN: "February 17, 2026", DE: "17. Februar 2026"
Character Encoding	UTF-8 for all languages	<meta charset="UTF-8">

Language Quality Assurance:

• Current State: Placeholder English body content with localized metadata (title, subtitle, HTML lang attribute, date formats) per language
• Target State (ADR-004): LLM generates content natively in each language (not machine translation)
• Cultural Adaptation: Planned — prompts will include cultural context for each language/region
• Terminology Consistency: EP terminology to be used consistently per language
• Quality Metrics: Human review of sample articles per language quarterly

---

---

📝 Architecture Decision Records (ADR)

Architecture Decision Records document significant architectural decisions made during the design and development of EU Parliament Monitor. Each ADR captures the context, decision, and consequences of a specific architectural choice.

ADR-001: Static Site Architecture over Dynamic Web Application

Status: Accepted
Date: 2025-12-01
Decision Makers: CEO, Development Team

Context:

• Need to display European Parliament news to public audience
• Security is paramount (public-facing system)
• Limited development resources
• GitHub Pages available as free hosting solution; AWS S3 + CloudFront chosen for production (see ADR-002)

Decision: We will build EU Parliament Monitor as a static site generator rather than a dynamic web application with backend services.

Rationale:

1. Security: Static sites eliminate entire classes of vulnerabilities (SQL injection, XSS via server-side rendering, authentication bypass)
2. Scalability: Static content scales infinitely via CDN with no server infrastructure
3. Cost: Static hosting on AWS S3 + CloudFront is low-cost, no server infrastructure
4. Maintainability: Simpler architecture with fewer moving parts
5. Reliability: No database or server downtime risks

Alternatives Considered:

• WordPress: Rejected due to security vulnerabilities, plugin maintenance overhead
• Node.js/Express backend: Rejected due to hosting costs, operational complexity
• JAMstack with headless CMS: Rejected due to unnecessary complexity for simple content

Consequences:

• ✅ Positive: Minimal attack surface, zero infrastructure costs, infinite scalability
• ✅ Positive: Fast page loads, excellent SEO, simple deployment
• ⚠️ Negative: Content updates require regeneration (acceptable for daily news)
• ⚠️ Negative: No real-time interactivity (not required for news consumption)

Compliance: Aligns with ISO 27001 A.8.28 (Secure Development), NIST CSF PR.DS-5 (Minimal Attack Surface)

---

ADR-002: AWS S3 + CloudFront for Hosting

Status: Accepted
Date: 2025-12-05
Decision Makers: CEO, DevOps Team

Context:

• Static site architecture chosen (ADR-001)
• Need reliable, secure hosting with global CDN
• Budget constraints (low-cost solution preferred)
• Already using GitHub for source control and CI/CD

Decision: We will host EU Parliament Monitor on AWS S3 with Amazon CloudFront as the global CDN (see .github/workflows/deploy-s3.yml).

Rationale:

1. Cost: Low-cost static hosting within current traffic and budget constraints
2. Integration: GitHub Actions CI/CD deploys to S3 and invalidates the CloudFront distribution
3. Security: HTTPS via AWS Certificate Manager, TLS termination at CloudFront edge
4. Reliability: AWS S3 and CloudFront SLAs provide high availability and durability
5. Performance: CloudFront global edge network with caching for low-latency delivery

Alternatives Considered:

• GitHub Pages: Considered for simplicity and zero direct hosting cost; kept as a documented alternative but not chosen due to less flexible edge configuration
• Netlify: Rejected due to build minute limits on free tier
• Vercel: Rejected due to commercial focus, potential future costs
• Self-hosted Nginx: Rejected due to operational burden, security maintenance

Consequences:

• ✅ Positive: Globally distributed static hosting with strong reliability and performance
• ✅ Positive: Automated deployments from GitHub Actions to S3 with CloudFront cache invalidation
• ✅ Positive: Integration with AWS security services (WAF, Shield, ACM)
• ⚠️ Negative: Ongoing AWS hosting costs and need to manage AWS credentials securely
• ⚠️ Negative: Increased operational complexity compared to GitHub Pages

Compliance: Aligns with ISO 27001 A.8.24 (Cryptography - HTTPS), CIS Control 1 (Asset Management)

---

ADR-003: Model Context Protocol (MCP) for European Parliament Data Access

Status: Accepted
Date: 2025-12-10
Decision Makers: CEO, Data Team

Context:

• Need structured access to European Parliament data (MEPs, plenary sessions, votes, documents)
• Official EP APIs are fragmented, inconsistent, and poorly documented
• Data schemas vary across endpoints
• Need caching, validation, and error handling

Decision: We will access European Parliament data via the European Parliament MCP Server using the Model Context Protocol (MCP) rather than calling official EP APIs directly.

Rationale:

1. Abstraction: MCP Server provides unified interface to fragmented EP APIs
2. Data Normalization: Consistent data structures across EP data sources
3. Error Handling: Connection retry logic and graceful degradation
4. Maintainability: API changes isolated to MCP Server, not news generator
5. Local Process: Spawned as stdio JSON-RPC process during build, no separate deployment needed

Alternatives Considered:

• Direct EP API calls: Rejected due to fragmentation, lack of validation, poor error handling
• Custom wrapper library: Rejected due to development overhead, maintenance burden
• Third-party EP data services: Rejected due to cost, data freshness concerns

Consequences:

• ✅ Positive: Clean separation of concerns, reusable data layer
• ✅ Positive: Standardized data structures, no direct EP API fragmentation
• ✅ Positive: MCP Server maintained separately, used by multiple clients
• ⚠️ Negative: Additional dependency (mitigated by fallback data strategy)
• ⚠️ Negative: Requires MCP Server process availability during build

Compliance: Aligns with ISO 27001 A.8.3 (Input Validation), NIST CSF PR.DS-2 (Data in Transit Protection)

---

ADR-004: Multi-Language Content via LLM Generation (Not Translation)

Status: Accepted
Date: 2025-12-15
Decision Makers: CEO, Content Team

Context:

• Need to support 14 languages
• Machine translation often produces unnatural, awkward phrasing
• European Parliament terminology requires domain expertise
• Budget available for LLM API costs

Decision: We will generate content natively in each language using LLMs rather than translating from a base language.

Rationale:

1. Quality: Native generation produces natural, idiomatic language
2. Cultural Adaptation: LLM can adapt content for cultural context per language
3. Terminology: LLM trained on EP documents uses correct terminology
4. Flexibility: Different article structures possible per language/culture
5. Scalability: Parallel generation for all languages

Alternatives Considered:

• Machine Translation (Google Translate, DeepL): Rejected due to unnatural phrasing, terminology issues
• Human Translation: Rejected due to cost (~€0.10/word x 14 languages), time delays
• English-only: Rejected due to accessibility concerns, limited audience

Consequences:

• ✅ Positive: High-quality, natural language content in all 14 languages
• ✅ Positive: Cultural adaptation, correct terminology
• ⚠️ Negative: Higher LLM API costs (~$5-10/day) vs translation (~$1-2/day)
• ⚠️ Negative: Content may vary slightly across languages (acceptable, even beneficial)

Compliance: Aligns with Hack23 AI Policy (Transparency, Human Oversight), ISO 27001 A.5.10 (Information Processing)

---

ADR-005: TypeScript with Strict Mode for Type Safety

Status: Accepted
Date: 2026-01-05
Decision Makers: CEO, Development Team

Context:

• Building news generation scripts and static site generator
• Need compile-time type safety for complex data structures from EP MCP Server
• Multiple article categories, 14 languages, and complex data pipelines
• Small development team (1-2 developers) benefits from IDE support

Decision: We will use TypeScript (strict mode) as the primary development language, compiling from src/ to scripts/ targeting ES2025.

Rationale:

1. Type Safety: Strict mode catches errors at compile time, especially important for complex EP data structures and MCP client interfaces
2. IDE Support: Full IntelliSense, refactoring, and navigation in VS Code
3. Self-Documenting: TypeScript interfaces serve as living documentation for data models (ArticleCategory, LanguageCode, MCPToolResult, etc.)
4. Build Pipeline: tsc compiles src/*.ts → scripts/*.js; rootDir: ./src, outDir: ./scripts, target: ES2025, module: NodeNext
5. Ecosystem: Full access to Node.js and npm ecosystem with type definitions

Alternatives Considered:

• JavaScript (ES2025) with JSDoc: Rejected due to weaker type guarantees, less comprehensive IDE support for complex interfaces
• Flow: Rejected due to declining community support
• JavaScript ES2015: Rejected due to lack of modern features (optional chaining, nullish coalescing)

Consequences:

• ✅ Positive: Compile-time error detection, comprehensive IDE support, self-documenting code
• ✅ Positive: Strict null checks prevent runtime errors with optional EP data fields
• ⚠️ Negative: Requires build step (npm run build / tsc) before execution
• ⚠️ Negative: Slightly higher learning curve for contributors unfamiliar with TypeScript

Compliance: Aligns with Hack23 Secure Development Policy (Type Safety Principle), ISO 27001 A.8.28 (Secure Coding)

---

ADR-006: Week-in-Review Analysis Window — D-36 → D-8

Status: Accepted
Date: 2026-04-27
Decision Makers: CEO, Development Team

Context:

• The EP publishes roll-call voting records with a 2–6 week lag after each plenary sitting.
• The previous week-in-review data window was D-0 → D-7 (the most-recent 7 days).
• A D-0→D-7 window structurally never contains published voting data, making the article vote-blind in every run regardless of content quality — a permanently-empty input.
• analysis/daily/2026-04-26/week-in-review/intelligence/methodology-reflection.md §3.1 recommended shifting to a D-36 → D-8 window to systematically capture voting data.

Decision: We shift the week-in-review analysis window to D-36 → D-8 (start = D-36, end = D-8 — a 28-day window ending 8 days ago, relative to the run date). This direction matches the workflow's DATE_FROM (start = D-36) → DATE_TO (end = D-8) variables. It is a 4-week look-back that consistently captures at least one full EP plenary week with published roll-call votes.

Rationale:

1. Data depth over recency: A vote-populated analysis is more valuable than a vote-empty analysis that is 7 days more recent. Readers of the week-in-review expect vote coverage.
2. Systematic: The window is deterministic and reproducible — it always yields voting data regardless of EP publication lag variance (2–6 weeks).
3. Complementary to fallback: This window shift works alongside any future EP Open Data Portal fallback for historical roll-calls; the two are not mutually exclusive.
4. Article framing updated: The WEEKLY_REVIEW_TITLES subtitles (all 14 languages) now read "last full reporting week" instead of "past week" to accurately describe the shifted window to readers.
5. SEO metadata: The title date range already shows the exact dateFrom–dateTo window, so canonical URLs remain accurate without additional changes.

Alternatives Considered:

• Keep D-0→D-7 + add EP Open Data Portal fallback query for historical roll-calls: Complementary approach; can be combined with this shift but does not solve the structural vote-empty problem without the window shift.
• D-8→D-14 (7-day window, offset by 8 days): Narrower window; may miss vote publication for sittings right at the 8-day boundary given the 2–6 week lag variance. Rejected in favour of the wider 28-day window.

Consequences:

• ✅ Positive: Every week-in-review run now reliably contains roll-call voting data.
• ✅ Positive: Analysis depth improves without increasing Stage B budget.
• ✅ Positive: Article subtitles accurately describe the reporting window in all 14 languages.
• ⚠️ Trade-off: Articles cover events from 8–36 days ago rather than the most-recent 7 days; the workflow is less "breaking" but more analytically complete.
• ⚠️ Negative: In this ADR, the DATE_FROM / DATE_TO variables replace LAST_WEEK in week-in-review Stage A bash blocks; other workflows still using LAST_WEEK require separate migration if their reporting windows are changed.

Implementation:

• src/aggregator/article-metadata.ts: New deriveReportingWindowForWeekInReview() export computes D-36/D-8 from the article date; buildTemplateFallback uses it for week-in-review.
• .github/workflows/news-week-in-review.md: Stage A sets DATE_FROM (D-36) and DATE_TO (D-8); all MCP tool calls use these variables; LAST_WEEK removed.
• src/constants/language-articles.ts: WEEKLY_REVIEW_TITLES subtitles updated (14 languages).

Compliance: Aligns with Hack23 AI Policy (unambiguous date semantics in published articles), GDPR (accurate published metadata).

---

🎯 Non-Functional Requirements (NFR)

Non-functional requirements define system qualities that are not directly related to specific features but are critical to overall system success.

Performance Requirements

Requirement	Target	Measurement	Current Status
Page Load Time (Desktop)	<1 second (LCP)	Lighthouse (manual runs)	✅ 0.6s average
Page Load Time (Mobile)	<2 seconds (LCP)	Lighthouse (manual runs)	✅ 1.2s average
Build Time (All Languages)	<15 minutes	GitHub Actions logs	✅ 8-12 minutes
Article Generation (Single)	<30 seconds	Script logs	✅ 15-25 seconds
MCP API Response Time	<2 seconds (p95)	Client logs	✅ 1.1s average
CDN Cache Hit Rate	>95%	CloudFront metrics (planned)	⏳ TBD — instrumentation planned

Performance Optimization Strategies:

• Static Content: All content pre-generated, no server-side processing
• CDN Caching: Tiered caching strategy (1 hour for HTML, 1 day for metadata, 1 year for immutable assets)
• Image Optimization: None required (no images in MVP)
• Minification: HTML minification (future), CSS minification (future)
• HTTP/2: Enabled by default on Amazon CloudFront

Scalability Requirements

Dimension	Current Capacity	Target Capacity	Scaling Strategy
Concurrent Users	Unlimited (static content)	Unlimited	CDN auto-scales
Daily Visitors	10,000+	100,000+	CDN bandwidth increase
Articles per Day	14 (one per language)	140 (ten per language)	Parallel generation, workflow optimization
Supported Languages	14	24+ (expanded markets)	Add language configs, LLM prompts
Repository Size	150 MB	800 MB (GitHub limit)	Archive old articles annually

Scalability Constraints:

• AWS S3: No repository size limit for static hosting; storage costs increase linearly
• GitHub Actions: 2000 minutes/month free, unlimited for public repos
• LLM API: Rate limits vary by provider (typically 3000 RPM for tier 2)

Availability and Reliability Requirements

Requirement	Target	Measurement	Consequence of Failure
Site Availability	99.9% (AWS CloudFront/S3 SLA)	GitHub Status + AWS Health Dashboard	Users cannot access news
Build Success Rate	>98%	GitHub Actions logs	No new content deployed
MCP API Availability	>99% (best effort)	Health checks	Fallback to placeholder events (no cached/previous data)
LLM API Availability	>99.5% (provider SLA)	API logs	Generation fails, retry logic
Recovery Time Objective (RTO)	<15 minutes	Manual testing	Time to restore service after outage
Recovery Point Objective (RPO)	<24 hours	Git history	Maximum data loss acceptable

High Availability Strategies:

• Static Architecture: No single point of failure (SPOF) in runtime
• CDN Redundancy: Amazon CloudFront with multiple edge locations globally
• Fallback Data: Use placeholder events if EP MCP Server unavailable (no cache/previous-data reuse)
• Retry Logic: Exponential backoff for transient failures
• Monitoring: GitHub Status, Dependabot alerts, workflow notifications

Security Requirements

Requirement	Implementation	Verification	Compliance
HTTPS-Only	CloudFront enforces HTTPS redirect via ACM certificate	Manual testing	ISO 27001 A.8.24
Content Security Policy (CSP)	Planned strict CSP via CloudFront response headers (no CSP meta tag in HTML templates currently)	CSP Evaluator (staging/production)	ISO 27001 A.8.23
No Secrets in Repository	GitHub Secrets for API keys	Git history scan	ISO 27001 A.8.3
Dependency Vulnerability Scanning	Dependabot daily scans	GitHub Security tab	CIS Control 10
SAST (Static Application Security Testing)	CodeQL weekly + PR	GitHub Code Scanning	ISO 27001 A.8.28
Access Control	GitHub RBAC, branch protection	Repository settings	CIS Control 6
Audit Logging	GitHub audit logs, workflow logs	Logs API	ISO 27001 A.8.15
Data Classification	All content PUBLIC	CLASSIFICATION.md	ISO 27001 A.5.10
Incident Response	SECURITY.md procedures	Quarterly reviews	NIST CSF RS.RP

Security Testing:

• SAST: CodeQL (weekly + PR) - JavaScript/TypeScript, HTML
• Dependency Scanning: Dependabot (daily) + npm audit (pre-commit)
• Manual Penetration Testing: Not required (static site, no user input)
• Security Reviews: Quarterly architecture review

Accessibility Requirements (WCAG 2.1 AA)

Criterion	Requirement	Implementation	Testing
Perceivable	Text alternatives, adaptable content, distinguishable	Semantic HTML5, alt text, contrast ratios	Playwright axe tests
Operable	Keyboard accessible, enough time, navigable, input modalities	Focus management, skip links, ARIA labels	Manual keyboard testing
Understandable	Readable, predictable, input assistance	lang attributes, consistent navigation, form labels	Lighthouse accessibility
Robust	Compatible with assistive technologies	Valid HTML5, ARIA roles	HTML validator

Accessibility Targets:

• WCAG 2.1 AA Compliance: 100% (mandatory)
• Lighthouse Accessibility Score: >95% (target 100%)
• Keyboard Navigation: All interactive elements accessible
• Screen Reader Support: JAWS, NVDA, VoiceOver tested quarterly

Accessibility Testing:

• Automated: Playwright with axe-core (every PR)
• Manual: Quarterly screen reader testing, keyboard navigation
• Tools: Lighthouse (manual runs), axe DevTools, HTML validator

Maintainability Requirements

Metric	Target	Current	Tool
Code Coverage	>80% lines	82%	Vitest
Branch Coverage	>80% branches	83%	Vitest
Cognitive Complexity	<15 per function	<10 average	ESLint sonarjs cognitive-complexity rule
Code Duplication	<3%	<2%	Manual review
Documentation Coverage	100% public APIs	95%	JSDoc, manual review
Build Time	<5 minutes (tests only)	3-4 minutes	GitHub Actions

Maintainability Practices:

• Code Review: All PRs require approval
• Documentation: Architecture, security, process docs maintained
• Testing: Unit (Vitest 4.1.4), Integration (incl. MCP contract tests), E2E (Playwright 1.59.1 + axe-core)
• Linting: ESLint 10.2.1 with eslint-plugin-sonarjs@4.0.3, eslint-plugin-security@4.0.0, eslint-plugin-jsdoc@62.9.0; Prettier 3.8.3 formatting
• Dependencies: Minimal (1 required production, 1 optional, ~40 dev), weekly Dependabot updates

---

🎯 Design Principles

1. Security by Design

• Minimal Attack Surface: Static architecture eliminates server-side vulnerabilities
• No Runtime Execution: Pure HTML/CSS with no backend processing
• Content Security Policy: Strict CSP headers prevent XSS
• HTTPS Only: All content delivered over HTTPS

2. Separation of Concerns

• Generation: News generation scripts (TypeScript → Node.js)
• Presentation: Static HTML/CSS
• Data Access: MCP Client abstraction
• Infrastructure: GitHub-managed CI/CD and hosting

3. Multi-Language First

• 14 Languages Supported: Full multi-language coverage including RTL support
• Language-Specific Indexes: Separate navigation for each language
• SEO Per Language: Individual sitemaps and metadata

4. Maintainability

• Minimal Dependencies: One production dependency (european-parliament-mcp-server for build-time data access), only dev dependencies otherwise
• Standard Technologies: HTML5, CSS3, TypeScript (compiled to ES2025 JavaScript)
• Comprehensive Testing: Unit, integration, and E2E tests
• Documentation: Architecture, security, and process docs

5. Scalability

• Static Content: Infinite scalability via CDN
• No Database: No scaling bottlenecks
• Cacheable: All content highly cacheable
• GitHub Infrastructure: Leverages GitHub's global infrastructure

---

📈 System Qualities

Performance

• Cold Start: N/A (static site, no cold starts)
• Page Load: < 1s (static HTML, CDN cached)
• Build Time: ~5-10 minutes (generation for all languages)
• Deployment Time: ~1-2 minutes (S3 sync + CloudFront invalidation)

Availability

• Target: 99.9% (AWS CloudFront/S3 SLA)
• Redundancy: CloudFront with multiple edge locations globally
• Failover: Automatic via AWS infrastructure
• Monitoring: AWS Health Dashboard, GitHub Status page

Security

• Attack Surface: Minimal (static files only)
• Vulnerability Scanning: Daily (Dependabot + npm audit)
• SAST: Weekly (CodeQL)
• Compliance: ISO 27001, GDPR, NIS2, EU CRA aligned

Maintainability

• Code Complexity: Moderate (5-stage pipeline + 8 strategies + 6 builders; no SPA framework)
• Test Coverage: 82%+ lines, 83%+ branches across 52 test files; 3061+ passing tests (unit, integration incl. EP/IMF/WB MCP contract tests, E2E Playwright)
• Documentation: Comprehensive (25+ architecture & ISMS docs — see Architecture Documentation Map)
• Dependencies: 1 pinned production (european-parliament-mcp-server@1.2.15), 1 optional (worldbank-mcp@1.0.1), ~40 dev dependencies

---

🔗 Related Documentation

• Security Architecture - Detailed security implementation and threat model
• Future Architecture - Architectural evolution roadmap
• Data Model - Data structures and EP/IMF/WB contracts
• Workflows - All 9 gh-aw + ~15 standard workflows, AI-First 2-pass enforcement
• End-of-Life Strategy - Technology lifecycle & EOL planning
• Flowcharts - Detailed process workflows
• State Diagrams - System state transitions
• Mindmaps - Conceptual system relationships
• SWOT Analysis - Strategic analysis and positioning
• README.md - Getting started guide and features overview

---