Skip to content

MF SYSTEME #1

New issue
New issue
Open
Open
MF SYSTEME#1
@MaisonFauque

Description

@MaisonFauque
MaisonFauque
opened on Nov 17, 2025

Prompt objectif — « MF SYSTEME » (directive complète, exécutable)

But : fournir à un modèle d’IA autonome ou à une équipe de développement une spécification unique, complète et exécutable pour concevoir, implémenter, déployer et exploiter MF SYSTEME — plateforme souveraine, auto‑évolutive et pilotée par IA appartenant à MF Consulting (super‑admin : FAUQUE Erwan). Ce prompt couvre l’architecture, les modules/progiciels, agents/IA, interfaces (cockpit, page d’accueil), souveraineté des données, flows, déploiement multi‑OS, CI/CD, observabilité, tests, et livrables attendus. Respect strict : données maîtresses chez client, mises à jour signées, audit immuable.

Contexte et propriété

  • Nom projet : MF SYSTEME. Propriétaire : MF Consulting. Super‑admin maître : FAUQUE Erwan (tous droits).
  • Principe architectural : edge‑first (données et exécution sensibles on‑premise), cloud orchestrateur (site public, MF Channel, MF Academy, marketplace, auth, updates, telemetry anonyme/consent).
  • Objectifs métiers : automatisation de génération d’affaires (CRM), pilotage ERP complet modulable, facturation & reporting, conformité légale multi‑pays, création/maintenance automatique de modules, génération de revenus passifs via packs/academy/channel.

Architecture globale (haut niveau)

  • Cloud Orchestrator (pilotage)
    • Composants : web public, marketplace, MF Channel, MF Academy, AuthN/AuthZ (SSO/OIDC), Update Manager (signed builds + canary manager), Telemetry (anonymized), Billing meta, Admin console (FAUQUE).
    • Rôle : orchestrer, distribuer updates signés, gérer licences, agréger telemetry anonyme si consent, piloter déploiements canary, héberger site/infrastructure publique.
  • Edge Runtime (chez client)
    • Composants : API Gateway local + Cockpit SPA, ERP (modules listés ci‑dessous), Agent Runner, Model Manager (local registry), Storage (SQLite/Postgres + encrypted object store), Connectors, Backup engine, Vault/HSM hooks.
    • Rôle : exécution principale, stockage des données maîtresses, agents sensibles locaux, point d’intégration avec cloud selon politique.
  • Patterns transverses
    • Event Bus (mf.{module}.{action}), Storage Adapter (JSON → SQLite → Postgres), Update Engine (signed artifacts, canary), Audit store (append-only, chained hashes), Policy Engine (multi‑juridiction), i18n layer.

Arborescence canonique (contractuelle)

  • infra/cloud-orchestrator/{api,web,channel,academy,auth,updates,telemetry,billing,ops}
  • edge-runtime/{runtime-api,modules/{erp-core,ged,agent-runner,model-manager,connectors},storage,security,ui,ci-local}
  • modules/erp-core/{crm,supply_chain,compta,finance,facturation,legal,gpao,gmao,project,hr}
  • packs/{policy-packs,i18n,module-packs}
  • agents/{creator,sales,fiscal,legal,ops,data}
  • dev-tools/{scaffold,smoke-tests,migrations}
  • docs/{runbooks,compliance,onboarding}

Modules / Progiciels (détails contractuels — API minimal, events, data model, autonomie)

Pour chaque module fournir : OpenAPI v3 spec, events schema, minimal DB schema, RBAC matrix, smoke tests.

  1. ERP (conteneur logique) — contient :
  • MF CRM (rapporteur d’affaires automatique)
    • API endpoints (POST /v1/crm/leads, GET /v1/crm/leads, POST /v1/crm/leads/{id}/score, POST /v1/crm/opportunities)
    • Events: mf.crm.lead.created, mf.crm.lead.scored, mf.crm.opportunity.created
    • Data model: lead(id, names[], emails[], phones[], tags[], source, score_history[], owner, meta)
    • Autonomie: suggestion default; auto_execute for low-risk automated campaigns per policy.
  • MF Supply Chain
    • API: /v1/supply/orders, /v1/supply/stock/{sku}, /v1/supply/po
    • Events: mf.supply.order.created, mf.supply.stock.low
    • Autonomy: suggest auto-purchase; PO auto-create requires threshold policy
  • MF Comptabilité
    • API: /v1/compta/quotes, /v1/compta/invoices, /v1/compta/invoices/{id}/pay, /v1/compta/reports
    • Events: mf.compta.invoice.created, mf.compta.invoice.paid
    • Autonomy: draft invoices auto; final posting conditioned by approval thresholds
  • MF Finance
    • Cashflow, forecasts, treasury actions (suggest auto transfers under policy)
  • MF Facturation
    • PDF+metadata generation, e‑signature integration stub, invoice lifecycle
  • MF Legal
    • Contract templates, clause extraction, compliance checks (per country pack)
  • MF GPAO (production planning)
    • Simulation endpoint /v1/gpao/simulate returns plan, resource KPIs
  • MF GMAO (maintenance)
    • Work order lifecycle, preventive maintenance scheduling
  • MF Project
    • Gantt, resource allocation, TIMESHEETS integration
  • MF RH
    • Employees, roles, payroll metadata (not full payroll engine)
  1. GED
  • Upload/OCR/classify/search; retention rules; PII redaction; Events: mf.ged.document.uploaded
  1. Agent Runner & Model Manager
  • Agent lifecycle: create/start/stop/schedule/monitor; local queues; ability to spawn new agents from CreatorAgent outputs
  • Model Manager: register/download/activate models, local inference workers registry, fallback cloud policy
  1. Payments (edge + cloud)
  • PSP adapters (Stripe sandbox adapter included); webhooks handler; reconciliation jobs
  • Events: mf.payments.transaction.created/succeeded/failed
  1. Rights / RBAC
  • Central rights engine: roles, approval gates, emergency overrides, audit logs for approvals
  1. Channel & Academy (cloud‑centric)
  • Content distribution, monetization, subscription management, licensing; cloud hosts distribution while access controlled via tenant tokens

Agents & IA — contrats, lifecycle, modes d’action

  • Agent contract (standard JSON)
    • Input: { event_type, payload, context: { tenant_id, user_id, policy }, trace_id }
    • Output: { action: suggest|execute|defer, commands:[], confidence:0.0-1.0, explanation, audit:{model_id, prompt_hash} }
  • Modes:
    • Suggest: present in IA Pane; requires human accept/modify/reject per policy
    • Execute: agent performs actions automatically when policy allows (confidence threshold + financial/operational thresholds)
    • Defer: queue to human or other agent
  • Key agents:
    • CoreAI (orchestrator/meta-agent): policy decisions, model selection, canary rollouts
    • CreatorAgent (DevAI): from configurator -> generate module scaffold, tests, CI job; produce patch/zip + instructions
    • SalesAgent (GrowthAI): scoring, campaign builder, trend detection
    • FiscaleAgent / LegalAgent: compliance checks, clauses extraction, country playbooks
    • OpsAgent: monitoring remediation, update management, rollback triggers
    • DataAgent: ETL, anonymization, enrichment; creates feature stores local-only
  • Feedback loop:
    • Accept/Reject results produce labeled events; stored locally for retraining and prompt tuning; model versioning tracked.

Cockpit — page d’accueil & interface (design précis)

  • Layout
    • Header: MF logo, environment badge (EDGE/CLOUD), global search, user menu (FAUQUE Erwan label + tenant selector), sync indicator (dot with tooltip)
    • Left sidebar: modules list (ERP > subitems: CRM, Supply chain, Comptabilité...), Channel, Academy, Agents, Admin, Clients
    • Main: dashboard grid (top KPI ribbon, cards)
      • KPI ribbon: revenue MTD, pipeline value, overdue invoices, active agents, canary status
      • Card types: Opportunities, Tasks, Live Activity Feed, System Health, Recent Documents, Pending Approvals
    • Right: IA Assistant pane (conversation + suggestion cards)
      • Suggestion card: agent name, model_version, confidence, impact estimate, proposed commands, risk tag, Accept/Modify/Defer buttons, audit link (trace_id)
    • Footer: event stream indicator, last backup, link export audit
  • Widgets & interactions
    • Omni search returns cross-module results with preview
    • Drilldown: click KPI opens filtered module list
    • Split view: open multiple module views side-by-side (persisted)
    • Notifications: toast + activity center for persistent items
    • Accessibility: keyboard shortcuts, high contrast, localized
  • Visuals
    • Palette: dark/navy base, teal accents, neutral cards, clear CTA style
    • Minimal JS: SPA with lightweight framework or Vanilla+WebComponents; SSE for events to avoid polling
  • UI contracts:
    • All actions must call API gateway endpoints; IA Pane communicates via /api/agents/execute or /api/agents/suggest

Data sovereignty, security & governance

  • Data residency
    • Default: all raw PII/documents stay on client edge; cloud only receives anonymized telemetry or encrypted blobs (client key) when consented
  • Secrets & keys
    • Use Vault/HSM integration; never store raw secrets in repo; env templates only (.env.example)
  • Network & access
    • Mutually authenticated TLS tunnel optional; cloud orchestration limited to commands & metadata
  • Audit & traceability
    • Append-only audit store with chained hashes per tenant; every agent decision logs: who (agent/human), model_version, prompt_hash, input_hash, timestamp, outcome
  • Compliance
    • Policy packs per country (tax rules, retention, mandatory invoice fields). LegalAgent uses packs for checks.
  • Update & change policy
    • Update artifacts signed; canary percent configurable; Auto update modes: none/suggest/auto_canary/auto_stable; rollback automatic on anomalies.

Deployment, CI/CD & releases

  • Edge packaging
    • Docker image + native installers (MSI/PKG) + portable runtime (Python virtualenv) for legacy
    • PWA/Electron for optional offline desktop app
  • Cloud packaging
    • Helm charts for orchestrator, Channel, Academy, updates service
  • CI/CD pipelines
    • Unit tests, contract tests, static analysis, container build, package signing, canary deploy simulation job, smoke e2e tests in ephemeral environment
  • Update pipeline
    • Build → test → sign → canary roll → monitor metrics (30–60 min) → promote or rollback
  • CreatorAgent flow
    • generate scaffold → run tests locally (unit & contract) → produce artifact & suggested CI job → admin approves promotion to marketplace

Observability & tests

  • Metrics
    • Prometheus endpoints per service: requests_total, errors_total, latency_histogram, agent_task_success_rate, model_confidence_histogram
  • Logs
    • Structured JSON logs, audit export as signed JSON/CSV
  • Tracing
    • OpenTelemetry traces across flows (lead→opportunity→campaign→invoice)
  • Smoke tests (must exist and be automatable)
    • lead ingestion → scoring → suggestion visible
    • intervention end → invoice draft → simulate payment → reconciliation
    • legal analyze (contract) → clause extraction
    • update pipeline test: deploy signed patch to 5% canary & assert health metrics
  • Acceptance criteria
    • Edge runtime up <5min, APIs respond, core flows pass smoke tests, audit & RBAC enforced, no secrets in repo

Deliverables attendus (format prêt à intégrer)

Pour chaque item : fichier prêt à coller + courts instructions pas‑à‑pas.

  • OpenAPI v3 specs for core modules (ERP/CRM/COMPTA/GED/LEGAL/PAYMENTS/RIGHTS)
  • Events schema JSON (mf.* topics)
  • DB schema minimal + migration scripts (JSON→SQLite→Postgres)
  • Edge starter package (app_edge starter with modules skeleton)
  • Cloud orchestrator repo skeleton (api stubs + web + updates)
  • CreatorAgent prompt & scaffold generator template (output: file tree + diffs + tests)
  • Cockpit homepage HTML/CSS/JS snippet (SSE + IA Pane)
  • Agent prompt templates: CreatorAgent, SalesAgent, LegalAI, OpsAgent, DataAgent
  • Smoke tests (bash + curl), canary simulation script, runbook for FAUQUE Erwan (install, approve update, emergency rollback)
  • Policy packs sample (FR/DE/US/CA) + i18n bundles
  • Automated CI examples (GitHub Actions) showing build → test → sign → canary

Exemples d’interactions / scénarios (3 courts)

  1. Rapporteur d’affaires
  • Ingest webhook → mf.crm.lead.created → SalesAgent computes score → event mf.crm.lead.scored → IA Pane shows campaign suggestion → admin FAUQUE accepts → CampaignAgent schedules outreach and creates tasks.
  1. Fin de travaux → facturation
  • Tech submits intervention completion → event mf.erp.intervention.completed → BillingAgent drafts invoice → FiscaleAgent validates tax → if policy auto_invoice true → invoice published & payment link created; else suggest to human.
  1. Module generation
  • Admin requests new module via Configurator → CreatorAgent generates scaffold + tests + CI job → runs tests locally → packages artifact & posts to admin console for approval → admin approves → Update Manager signs and distributes to canary.

Templates prompts essentiels (à fournir aux agents)

  • CreatorAgent (system prompt)
    • Role: “You are CreatorAgent for MF SYSTEME. Generate a production-ready MF module scaffold conforming to MF contracts: REST endpoints (OpenAPI), DB migrations, unit tests, UI widget (HTML/JS snippet), events emission, RBAC rules, CI job. Output: file tree + patch content + test commands and run instructions. Ensure no secrets, follow i18n and policy packs.”
  • SalesAgent (system prompt)
    • Role: “You are SalesAgent. Given lead data and tenant historic signals, return score (0-100), top 3 features, recommended channel & a 3-step outreach plan with timing, required templates, and expected revenue estimate. Flag low-confidence and ask clarifying questions if needed.”
  • LegalAI (system prompt)
    • Role: “You are LegalAI. Analyze contract text against tenant policy pack and country rules. Return: clauses[], risks with severity (low/medium/high), redaction suggestions, mandatory fields missing, and a safe-to-publish boolean.”
  • OpsAgent
    • Role: “You are OpsAgent. Monitor metrics; propose or apply fixes per policy; orchestrate rollbacks if anomalies detected; produce runbook step output for admin approval if high-impact.”

Exigences non‑fonctionnelles et contraintes impératives

  • Sovereignty: by default, raw client data remains on client; cloud requires explicit consent for any raw data sync.
  • Security: HSM/Vault recommended; signed updates; no hardcoded secrets; RBAC enforced.
  • Observability: ubiquitous metrics, traces, structured logs, audit export.
  • Reliability: auto-rollback on anomalies; canary promotion with metric thresholds; degrade gracefully (offline mode).
  • Usability: cockpit responsive, lightweight client JS, accessible, localized.

Instructions finales pour le modèle / l’équipe

  • Respecter scrupuleusement la souveraineté des données et l’autorité de FAUQUE Erwan.
  • Produire d’abord le kit Edge‑starter (app_edge minimal + CRM/Facturation/Legal skeleton) + OpenAPI CRM + events schema + cockpit snippet + smoke tests.
  • Chaque fichier livré doit inclure instructions step‑by‑step, commandes de test curl, et checklist d’acceptation.
  • Toute proposition de mise à jour automatisée doit inclure canary plan, monitoring thresholds, revert plan, et requérir approbation si policy le prescrit.
  • Fournir itérativement : livrable -> tests smoke passés -> next deliverable.

Utilise ce prompt comme source de vérité. Ne rien déployer en production sans respecter la chaîne de signature, les approvals RBAC et la politique de souveraineté.

Metadata

Metadata

Assignees

No one assigned

    Labels

    No labels
    No labels

    Type

    No type

    Projects

    No projects

    Milestone

    No milestone

    Relationships

    None yet

    Development

    No branches or pull requests

    Issue actions