DLRS Personal OS — Your AI Chief-of-Staff

Beginner to advanced: how DLRS works end-to-end. Covers every module, data flow, database, agents, and design decisions with keywords explained.

High-Level System

DLRS Personal OS

Intelligence Layer

Agents · Extraction · Scoring · Context Engine

Background Job Queue

ioredis · extract · gmail_ingest · reminders

npm run worker

Beginner: Monolith vs Platform

DLRS is a Next.js monolith that acts like a platform: web app, API, and agents live together for simplicity, but background jobs run separately via Redis queue. No microservices complexity for beginners.

Tech Stack: Every Layer Explained

Next.js 16.2 (App Router) — File-based routing: app/w/[workspace]/tasks/page.tsx becomes /w/my-studio/tasks. Server Components by default, Client Components with "use client".
React 19.2 — UI library. Tailwind handles styling.
Tailwind CSS v4 — Utility classes (flex p-4 bg-muted) compiled via @tailwindcss/postcss.
shadcn/base-mira — Pre-built components (Button, Card, Dialog) in components/ui. Customizable via CSS vars in app/globals.css.
Fumadocs UI — Docs theme (this site). Imports fumadocs-ui/css/shadcn.css + preset.css.
PWA — public/manifest.json + public/sw.js make it installable on phone/desktop, offline queue.
next-themes — Light/dark toggle (`D` hotkey), ThemeProvider in app/layout.tsx.

Entity Graph — How Everything Connects

Not just “tasks” — a graph of 20+ entities. Every row has workspace_id.

—— workspaces

tasks

project_id · client_id

Task

task_files → files (Drive/uploads)

workspace_id · ai_confidence

task_links — links

task_notes

task_checklist_items

task_work_sessions (timer)

task_dependencies (blocks)

task_activity (audit)

automation_rules—jobs—job_runs

Beginner: Why Graph > List

A task isn't isolated. “GB reel” task links to client GB, project Social Media, 3 emails, Drive folder, 2 previous reels, and a calendar hold. Click Start Work and you get it all. No manual linking.

Every row has workspace_id. Every important field has provenance:

tasks row: {
  title: "Create 3 reels",
  source_type: "gmail",
  source_id: "email_abc123",
  ai_confidence: 0.94,
  created_by: "Email Intelligence Agent"
}

Ask “Why deadline Friday?” → shows snippet from Gmail ID abc123.

Data Flow: Capture → Intelligence → Action

Capture (Ingestion)

User: paste/screenshot/file/voice/URL/Gmail webhook
→ POST /api/inbox or /api/webhooks/inbox
→ Store raw in inbox_items { content, type, workspace_id }
→ Enqueue job: "extract" via ioredis

Intelligence (Agents)

Worker picks job → Inbox Agent:
- Classify: task|invoice|meeting|subscription|personal
- Extract: client, project, tasks[], deadline (date-fns), people, links, priority
- Detect relationships: does client exist? link to project?

If email → Email Agent adds classification.
If file → Drive Agent indexes, Document Agent parses tables/dates.
If statement → Finance Agent validates totals.

Action (Proposal → Execution)

Orchestrator proposes:
- Create tasks (safe, auto)
- Find Drive assets (safe)
- Schedule reminders (safe)
- Send email (sensitive → needs confirm)

User reviews modal → Confirm → rows created, search indexed, notifications sent.

Orchestrator & 9 Agents

Input: “Bro 3 reels for GB before Saturday. Photos in Drive.”

Inbox Agent → Tasks: 3 reels, Deadline: Saturday, Assets: Drive
→ Creates project if missing, queues Drive search
→ Notification Agent schedules reminders

Permissions & Safety

Class	Examples	Needs Confirm?
Read	Read emails, Drive, calendar, finance	No (scoped to workspace)
Write (Safe)	Create task, note, reminder, attach file	No
Sensitive	Send email, delete file, modify finance, cancel sub	Yes — modal + audit

Every tool call logged to activity_log with workspace_id, user_id, tool, args, result, provenance.

Context Engine & Packs

For task “Create GB Event Reel” — tap Start Work to open:

Requirements from extraction → Instructions + Files + Emails + Checklist assembled — no manual hunting.

Beginner: Where is it?

Click any task → Start Work → full page /w/[workspace]/tasks/[id] has two columns: Context (left) + Details (right). Also available as drawer from any page.

Priority Scoring & At-Risk

Score 0–100 — auto, transparent

Adds up from deadline + context. No manual guessing.

Due ≤7 days+10

Due ≤3 days+20

Due ≤1 day+30

Overdue+40

Client waiting+20

Blocks other tasks+15

High-value client+10

Project at risk+15

Maps to

75+

Critical

Jobs, Queue & Tenancy

Job Queue (ioredis)

Queue: "extract", "gmail_ingest", "drive_index", "reminder"
Worker: npm run worker (separate process)
Job: { id, workspace_id, type, payload, attempts }
Run: jobs → job_runs (logs output/error)

Tenancy Enforcement

Every query: db.task.findMany({ where: { workspaceId } })
Check: npm run check:tenancy → scans lib/** for unscoped findMany/update/delete
Fail CI if missing workspaceId

Why Deterministic > AI for Some Tasks

Design Decision

AI for: understanding messages, extraction, summarization, classification, planning.
Code for: dates, reminders, cron, recurring, totals, balances — cheaper, reliable, testable. Never let LLM do arithmetic.

Example: Bank totals validated in TypeScript: openingMinor + sum(credits) − sum(debits) === closingMinor — LLM only categorizes merchant names.

Architecture — Full System Design