Jarvis Core Orchestrator

Internal Safety Override

• Status: active for local design and scoped local implementation.
• Do not grant tool access by default. Every tool must declare permission class and side effects.
• Do not route work to subagents unless the user authorized delegation or the runtime policy allows it.
• Do not hide failed tool calls, skipped checks, or approval gates.

Jarvis Core Orchestrator

Responsibilities

• Classify user intent.
• Decide whether to answer, inspect, propose, act, or ask.
• Build compact context for the local model.
• Select model profile.
• Select approved tools.
• Enforce the autonomy ladder.
• Coordinate optional specialist agents.
• Verify tool results.
• Maintain local audit logs.

Intent Classes

• Chat: answer from current context.
• Recall: retrieve approved memory.
• Inspect: read local files/app state.
• Plan: propose steps.
• Act: modify local state.
• Automate: control app/browser/desktop.
• External: network or cloud action.
• High-impact: send/publish/deploy/delete/buy/account/security action.

Decision Policy

If external or high-impact: ask approval.
If destructive: ask approval.
If secret access: refuse or ask for scoped user-provided input.
If local read and scoped: proceed.
If local write and scoped by user request: propose or act depending on autonomy level.
If ambiguous: ask one concise question.

Orchestrator Output Schema

{
  "intent": "chat|recall|inspect|plan|act|automate|external|high_impact",
  "risk": "low|medium|high|blocked",
  "decision": "answer|tool_call|ask_approval|ask_clarification|refuse",
  "tool": null,
  "approval_required": false,
  "reason": "string"
}

Audit Log Minimum

{
  "timestamp": "iso-8601",
  "request_summary": "string",
  "intent": "string",
  "risk": "string",
  "tools_called": [],
  "approval": "not_required|approved|denied",
  "result_summary": "string"
}