LLM Provider Configuration & Agent Tool Loop

This document describes the design and implementation of the multi-model LLM provider configuration and the agentic tool execution loop in DOST.

---

1. System Overview

DOST provides a Bring Your Own Keys (BYOK) model layer, allowing users to configure API keys for multiple LLM providers (e.g. Groq, Google Gemini) and dynamically toggle models per-chat. The core controller for model selection and tool loop execution resides in the Electron main process:

• Model Manager: models.js (Class: AIModel)
• IPC Controller: ipcHandlers.js
• Agent Engine: Vercel AI SDK Core (ToolLoopAgent)

---

2. Model Discovery & API Key Management

A. Provider and Model Registry

On start, the model manager queries the central FastAPI database endpoint /api/v1/llm_models to fetch available providers, model lists, and configurations. It maps this data into the local reactive state (state.providers).

B. Isolated Environment Credential Injection

To prevent key leaks, provider API keys are stored in the user-isolated configuration file (users/{userId}/config.json) inside the aiConfig.envStore object:

1. Clear on Boot: During initialization, clearProviderEnvVars() actively deletes any existing provider API keys from the application's current process.env.
2. Context Injection: When the active user loads the chat, the manager reads their encrypted credentials from envStore and injects them back into process.env dynamically:
   javascriptCopy

   process.env[provider.env_var] = envStore[providerName][provider.env_var];

3. Multi-User Isolation: When switching profiles, the environment variable map is wiped, and keys for the next user are loaded, preventing cross-account access.

---

3. The Agentic Tool Execution Loop

The core execution engine is powered by Vercel AI SDK's ToolLoopAgent. It analyzes user requests, selects appropriate tools, executes them in sequence, and generates final responses.

Execution Cycle Sequence

Safety & Guardrails

• Step Limit Guardrail: To prevent infinite cycles (e.g., two tools repeatedly triggering each other), the loop is constrained by a step limit of 8:
  javascriptCopy

  stopWhen: stepCountIs(8)

• User Confirmation Interlock: The agent is instructed via system prompts to draft changes and prompt the user for confirmation before performing any destructive or irreversible actions, such as shutting down/restarting the PC, deleting scheduled automation tasks, or sending emails.

---

4. Response Styling & Formatting Rules

The system prompt enforces strict formatting guidelines on the LLM to ensure visual consistency: