One agent layer, one persistence layer, one PDF pipeline

flowchart TB
  subgraph clients [Clients]
    Slack["Slack user"]
    Web["Web chat UI (existing)"]
    WA["WhatsApp (via bridge, optional)"]
  end

  subgraph surface [Surface layer]
    SlackSvc["slack.service (Socket Mode)"]
    WebApi["api.routes / chat_handler"]
    Bridge["bridge (Node, Baileys)"]
  end

  subgraph handler [process_chat_message]
    Phase0{"thread_ts maps to fraud_reports row?"}
    Followup["ReportFollowupAgent"]
    Memory["MemoryAgent (intent + complexity)"]
  end

  subgraph fraud [Fraud agent stack]
    FraudOrch["FraudOrchestratorAgent"]
    DataAg["FraudDataAnalystAgent"]
    PatAg["FraudPatternAgent"]
    RepAg["FraudReportAgent"]
    AlertAg["FraudAlertAgent"]
    VerAg["FraudVerificationAgent"]
  end

  subgraph followup [Follow-up stack]
    QAAg["ReportQAAgent"]
    EditAg["ReportEditorAgent"]
    TexEdit["tex_editor (pure Python)"]
  end

  subgraph storage [Local storage]
    Sqlite[("SQLite: pyguard.db
messages, fraud_transactions, fraud_reports")]
    Bundles[("backend/reports/fraud_report_<ts>/
report.tex, PDF, charts, history/")]
  end

  subgraph external [External services]
    OAI["OpenAI Agents SDK + Responses API"]
    Comp["Composio (Gmail, Calendar, Docs)"]
    Tec["tectonic (LaTeX to PDF)"]
  end

  Slack --> SlackSvc --> handler
  Web --> WebApi --> handler
  WA --> Bridge --> WebApi

  handler --> Phase0
  Phase0 -->|yes| Followup
  Phase0 -->|no| Memory

  Followup --> QAAg
  Followup --> EditAg
  EditAg --> TexEdit --> Tec
  TexEdit --> Bundles

  Memory -->|fraud intent| FraudOrch
  FraudOrch --> DataAg
  FraudOrch --> PatAg
  FraudOrch --> RepAg
  FraudOrch --> AlertAg
  FraudOrch --> VerAg

  RepAg --> Bundles
  RepAg --> Tec
  DataAg --> Sqlite
  QAAg --> Sqlite
  QAAg --> Bundles

  FraudOrch --> OAI
  Followup --> OAI
  Memory --> OAI

  SlackSvc --> Bundles

  Memory -.-> Comp

  classDef ext fill:transparent,stroke:#2a3356,stroke-dasharray:4 4;
  class OAI,Comp,Tec ext;