Mini Jarvis — Data skill for Claude Code

mini-jarvis

An **advisory-only**, human-in-the-loop AI agent system for small-business operations. It is not tied to one industry: describe your business and currency in config, and the agents adapt their prompts. The bundled demo data models a small Sydney plumbing business. The pattern:

agent reads business data → an AI model (Claude, Gemini or OpenAI) produces a recommendation → **a human must approve it** → everything is logged.

Advisory-only design

• Agents never perform actions. They read data and produce recommendations with status PENDING.
• A human approves or rejects each recommendation through the API. Approving only changes the status and writes an audit row. It does not execute anything. A test checks that approving calls no AI provider or notifier.
• Every step goes to the audit_log table with a timestamp and actor: agent run started/finished/failed, every AI call, fallback used, recommendation created/approved/rejected, notification sent/failed.
• Agents depend only on the AiProvider abstraction, never on a vendor SDK.

Architecture

                 ┌──────────────┐   POST /api/agents/{name}/run      ┌───────────────────┐
  Director ─────▶│  REST API    │◀── (JWT, single admin account) ────│  AgentScheduler   │
  (curl/UI)      │  controllers │                                    │ (off by default)  │
                 └──────┬───────┘                                    └─────────┬─────────┘
                        │                                                      │
                        ▼                                                      ▼
                 ┌──────────────────────────────────────────────────────────────────┐
                 │ AgentRunService: run record, audit trail, saves PENDING recs,    │
                 │                  optional notification (never fails the run)     │
                 └──────┬───────────────────────────────────────────────┬───────────┘
                        │ Agent.r