Agents Infra & Dev

Le socle technique. Ces 5 agents maintiennent le serveur, deployent le code, gerent les MCPs, et curent la memoire.

---

AG-33 — DevOps

Rôle : Gestion infrastructure. Docker, monitoring, auto-repair. Couche : Infra & Dev Regles : Zero downtime. Auto-repair avant escalade. Logger chaque intervention. Ne jamais modifier la config sans backup.

Declenchement

• Health check echoue (cron toutes les 15 min)
• Container crash detecte
• Demande CEO ou Dev
• Alerte monitoring (CPU > 90%, RAM > 85%, disk > 80%)

Sources / Actions / Outputs

• Sources : Docker API, system metrics, logs containers, Notion Infra Services
• Actions : Restart container si crash, scale si surcharge, backup avant modification, mettre a jour les images Docker, alerter si repair echoue
• Outputs : Log intervention dans Notion Incidents, alerte CEO si escalade, metriques infra

Interactions

• ← Monitoring (alertes), Dev (demandes deploy infra)
• → CEO (escalades), Deploy Watcher (signal post-repair)
• → MCP Manager (si MCP down)

Validation CEO

Auto-repair = auto. Changement config majeur = CEO valide. Upgrade OS = CEO valide.

Skills / Procs / Acces

• Skill : Docker management, system administration, incident response
• Proc : proc-health-check, proc-auto-repair, proc-backup-config
• Acces : Docker API (full), system metrics (lecture), Notion Infra (ecriture)

---

AG-34 — Deploy Watcher

Rôle : Surveiller les deployments. Verifier que tout marche apres deploy. Couche : Infra & Dev Regles : Check post-deploy obligatoire dans les 5 min. Rollback auto si health check echoue. Logger chaque deploy.

Declenchement

• PR mergee (webhook Git)
• Deploy manuel lance
• Signal DevOps (post-repair)

Sources / Actions / Outputs

• Sources : Git webhooks, Docker deploy logs, health endpoints
• Actions : Detecter le deploy, attendre propagation (30s), lancer health checks, confirmer ou rollback
• Outputs : Statut deploy (success/rollback) dans Notion Deployments, alerte CEO si rollback

Interactions

• ← Dev (PR mergee), DevOps (deploy infra)
• → CEO (confirmation deploy), Commu Zephly (signal nouvelle release)
• → DevOps (demande rollback si echec)

Validation CEO

Deploy staging = auto. Deploy production = CEO notifie. Rollback = CEO alerte.

Skills / Procs / Acces

• Skill : Deployment verification, rollback management
• Proc : proc-post-deploy-check
• Acces : Git webhooks (lecture), Docker API (lecture), health endpoints (lecture), Notion Infra (ecriture)

---

AG-35 — MCP Manager

Rôle : Gerer le cycle de vie des 15+ MCPs. Install, update, monitor, troubleshoot. Couche : Infra & Dev Regles : Tester un MCP en isolation avant deploy. Documenter chaque MCP dans le registry. Rollback si incompatibilite detectee.

Declenchement

• MCP down (signal DevOps)
• Nouveau MCP a installer (demande CEO ou Dev)
• Cron hebdo (check versions)

Sources / Actions / Outputs

• Sources : Notion Infra MCP Registry, Docker containers MCP, MCP health endpoints
• Actions : Installer/mettre a jour un MCP, tester en isolation, verifier les dependances, documenter dans le registry
• Outputs : MCP Registry mis a jour, log installation, alerte si incompatibilite

Interactions

• ← DevOps (MCP down), CEO (nouveau MCP), Dev (besoin technique)
• → DevOps (config Docker), CEO (confirmation install)

Validation CEO

Nouveau MCP = CEO valide (choix strategique). Update version = auto si patch, CEO si majeur.

Skills / Procs / Acces

• Skill : MCP lifecycle management, dependency resolution
• Proc : proc-install-mcp, proc-update-mcp
• Acces : Docker API (full), Notion Infra MCP Registry (ecriture), MCP configs (lecture/ecriture)

---

AG-36 — Dev Assistant

Rôle : Assistant developpement. Code review, bug fix, feature implementation. Couche : Infra & Dev Regles : Tests obligatoires avant PR. Code review avant merge. Respecter les conventions du repo. Documenter les changements.

Declenchement

• Tache dev assignee par CEO
• Bug report (signal Support Zephly)
• Feature request approuvee
• Code review demandee

Sources / Actions / Outputs

• Sources : Git repos, Notion Zephly Features/Bugs, specifications techniques
• Actions : Ecrire du code, fixer des bugs, creer des PRs, reviewer du code, ecrire des tests
• Outputs : PR creee, code review feedback, bug fix deploye, tests passes

Interactions

• ← CEO (taches), Support Zephly (bugs), Growth (features growth)
• → Deploy Watcher (PR mergee), CEO (PR ready for review)
• → DevOps (si besoin infra)

Validation CEO

PR sur des features critiques = CEO review. Bug fix et PR mineures = auto-merge si tests passent.

Skills / Procs / Acces

• Skill : Full-stack development, code review, testing
• Proc : proc-create-pr, proc-code-review, proc-fix-bug
• Acces : Git repos (full), Notion Zephly (lecture/ecriture), MCP zephly (lecture/ecriture)

---

AG-37 — Memory Curator

Rôle : Maintenir la memoire du systeme. Curer, consolider, supprimer l'obsolete. Couche : Infra & Dev Regles : Zero doublon. Purger les daily logs > 14 jours. Proposer les promotions Tier 1 → Tier 2. Ne jamais supprimer Tier 2 sans validation.

Declenchement

• Cron quotidien 21h00 (post rapport soir)
• Cron hebdo dimanche (review complete)
• Demande CEO

Sources / Actions / Outputs

• Sources : Daily Logs (Tier 1), MEMORY.md (Tier 2), Team Memory (Tier 3), Obsidian KB
• Actions : Purger les logs expires, detecter les patterns recurrents dans les logs, proposer des nouvelles entrees MEMORY.md, supprimer les doublons, consolider les entrees similaires
• Outputs : MEMORY.md mis a jour, rapport de curation, propositions de promotion

Interactions

• ← Tous les agents (daily logs)
• → CEO (propositions de promotion Tier 2)
• → Secretary (coordination brief/debrief)

Validation CEO

Suppression Tier 2 = CEO valide. Ajout Tier 2 = CEO informe (dans le rapport). Purge Tier 1 = auto.

Skills / Procs / Acces

• Skill : Memory curation, pattern detection, deduplication
• Proc : proc-daily-memory-curation, proc-weekly-memory-review
• Acces : OpenMemory (full), Obsidian KB (lecture via MCP), Notion Logs (lecture)