AI Suite Dataflows

Stand: 2026-05-30

OpenCode Flow

sequenceDiagram
  participant User
  participant OpenCode
  participant LiteLLM
  participant MCP as Remote MCP
  participant CE as Controlled Execution
  User->>OpenCode: Prompt
  OpenCode->>LiteLLM: Chat / Coding
  OpenCode->>MCP: read-only / preflight tools
  MCP->>CE: scoped write only with approval
  CE-->>MCP: policy / audit / result
  MCP-->>OpenCode: operator response

Status: produktiv. OpenCode läuft aktuell bewusst direkt gegen LiteLLM/MCP, nicht zwingend über den Lanstyle Agent Gateway.

OpenWebUI Flow

sequenceDiagram
  participant User
  participant OpenWebUI
  participant LiteLLM
  participant Access as OpenCode Self-Service
  User->>OpenWebUI: Chat oder Zugriff anfordern
  OpenWebUI->>LiteLLM: Chat Completion
  OpenWebUI->>Access: Link/Button zu /opencode-access
  Access-->>User: One-time Download

Status: produktiv. Normale Secret-Ausgabe findet nicht im Chat statt.

Voice / Jarvis Flow

sequenceDiagram
  participant Device as Voice Client
  participant Gateway as Lanstyle Agent Gateway
  participant Qdrant
  participant Core as Nous Hermes Core
  participant LiteLLM
  Device->>Gateway: /voice/message
  Gateway->>Qdrant: Retrieval
  Gateway->>Core: Message + Kontext
  Core->>LiteLLM: Model call
  Core-->>Gateway: Antwort
  Gateway-->>Device: Textantwort

Status: vorbereitet/Backend aktiv. Audio bleibt Client-seitig.

Controlled Execution Flow

flowchart LR
  Plan["Plan"] --> Risk["Risk/Policy"]
  Risk --> Backup["Backup/Export"]
  Backup --> Approval["Approval ID"]
  Approval --> Lock["Dependency Lock"]
  Lock --> Execute["Scoped Execute"]
  Execute --> Verify["Verification"]
  Verify --> Audit["Audit"]
  Audit --> Rollback["Rollback Reference"]

Status: produktiv validiert für begrenzte Proxmox-/NetBox-Aktionen.