Design Philosophy

This set of documents is the product design spec for ETOS LLM Studio, not marketing material. The goal:

• Help newcomers understand how the product works without reading Swift code
• Help current users know which behavior is deliberate design vs just a default
• Help external contributors understand the module boundaries between Daily Pulse, Memory, Worldbook, MCP, and Shortcuts

Before Reading This Set

We recommend you use ETOS for a few days first — configure a provider, send a few rounds of chat, install at least one Skill or MCP. Design docs are abstract on their own; hands-on first makes everything easier.

Not installed yet? Install & First Launch.

ETOS Is Not a Chat Wrapper

Most LLM clients position themselves as "wrap a model API in a usable app." ETOS targets something different: fit the following into a single native client.

Subsystem	Responsibility
Chat	Multi-provider, multi-model, multimodal, fine-grained request tuning
Context	Global prompt, topic prompt, enhancement prompt, long-term memory, cross-session summary, user profile, worldbook
Proactive	Daily Pulse generates "what to look at today" before you ask
Tools	Built-in tools, MCP, Shortcuts, Skills, sandbox file tools — unified under one governance
Two-device experience	iPhone for full configuration + management; Apple Watch for quick access + alerts + lightweight follow-up

Complexity is much higher than "API + text display." That's why explicit design principles matter — otherwise the system tears itself apart past a certain point.

Four Design Principles

1. Explainable

ETOS deliberately avoids "the model suddenly got smart, but you have no idea what it just saw."

Context is split into a clear hierarchy:

Layer	Responsibility
Global prompt	Long-term identity and overall rules ("you are AI X")
Topic prompt	Current session's subject constraints
Enhancement prompt	Per-turn auto-added instruction
Memory / Session summary / User profile	Long-term and cross-session background
Worldbook	Rule-based, conditionally-triggered targeted injection

Each layer can be toggled independently — you can run with everything on, or strip down to a bare model, instead of accepting a black box.

For the full assembly order, see Prompt & Context Assembly.

2. Proactive but Low-Interruption

Daily Pulse isn't "replace chat." It solves a real problem: a lot of people open the app each day not knowing what to ask.

Approach:

• One card stack per day, not an infinite feed
• Feedback-driven (like / downvote / hide / save), not forced recommendations
• Local persistence + local reminders, not cloud account push
• Best-effort morning delivery, not a promised cloud cron job

Details in Daily Pulse Internals.

3. Tools Are Governed, Not Maxed Out

ETOS doesn't market "number of tools." It treats "what tools are actually exposed in this session" as the more important question.

Tool Center has at least two state layers:

• Configured to be enabled
• Actually available in this session

The two often differ. Factors:

Factor	Effect
Approval policy	"Always Deny" blocks even configured-on tools
Worldbook isolation	The whole session disables tools
MCP server selected for chat	Unselected servers don't contribute tools
Per-tool disable	Same

Details in Worldbook & Tool Governance.

4. Two-Device Division of Labor, Not Mirror

iPhone and Apple Watch are not the same UI shrunk down.

Device	Job
iPhone	Provider config, tool governance, worldbook editing, memory management, import/export, debug/feedback
Apple Watch	Receive alerts, quick session start, follow up on Daily Pulse, voice or short text input

That's why ETOS pushes complex policy down to the ETOSCore layer (ETOSCore/ETOSCore/) rather than hard-coding it into a specific screen — ETOSCore lets each device take what it needs.

Data-Flow Overview

User input  /  Daily Pulse cards  /  External tool results
              │
              ▼
        ChatService request orchestration
              │
    ┌─────────┼─────────┐
    │         │         │
    ▼         ▼         ▼
 Prompt    Memory    Worldbook
              │
              ▼
        Tool exposure & approval
              │
              ▼
        Send to selected model
              │
              ▼
      Response, summary, profile, feedback writeback

Every model-bound request goes through this pipeline.

Where to Start

Your question	Read
What goes into the prompt before a message is sent?	Prompt & Context Assembly
Why is Daily Pulse "proactive"? What signals feed it?	Daily Pulse Internals
What do memory, session summary, and user profile each do?	Memory, Summary & Profile
Why does a worldbook affect tool availability?	Worldbook & Tool Governance

One Overall Take

If you treat ETOS purely as a chat wrapper, many entry points look scattered.

If you treat it as a native AI workstation, things click:

Module	Role
Chat tab	Execution
Settings tab	Governance
Daily Pulse	Proactive discovery
Memory + Worldbook	Long-term context
Tool Center	Capability exposure + risk control

Settings is long because it's the governance console for an entire AI system — not "too many features piled up," but "governance itself needs this many panels."