ODS · decision courseENPT
Architecture · minimum M5 Max pilot

ODS architecture — from click to model

In the minimum pilot, the web UI receives the request, Docker services route it, and the native macOS llama-server performs inference. Extensions add capabilities and also processes, data, and ports.

By the end: explain, locate evidence, and apply it to the pilot.
01

Complete Simple + Technical

In the minimum pilot, the web UI receives the request, Docker services route it, and the native macOS llama-server performs inference. Extensions add capabilities and also processes, data, and ports.

Bounded analogy: The UI is the counter, LiteLLM is routing control, and llama-server is the runway. Unlike an airport, each service can fail and retain state independently.

Under the hood

On macOS, documentation describes host Metal and containers reaching host.docker.internal:8080. Compose combines base, Apple overlay, and extensions. Local, cloud, and hybrid modes change routing and egress.

Boundary: installation, runtime, and Docker/Compose were not executed.

02

14 semantic microdiagrams

Each figure teaches one relationship; the sequence is a decision map, not a health dashboard.

local browser
01 · local browser — decision relationship
Open WebUI :3000
02 · Open WebUI :3000 — decision relationship
dashboard :3001
03 · dashboard :3001 — decision relationship
API :3002
04 · API :3002 — decision relationship
LiteLLM
05 · LiteLLM — decision relationship
Docker network
06 · Docker network — decision relationship
host bridge
07 · host bridge — decision relationship
llama-server :8080
08 · llama-server :8080 — decision relationship
host Metal
09 · host Metal — decision relationship
Compose base
10 · Compose base — decision relationship
Apple overlay
11 · Apple overlay — decision relationship
extensions
12 · extensions — decision relationship
local mode
13 · local mode — decision relationship
hybrid egress
14 · hybrid egress — decision relationship
03

Command and proof boundary

rg -n 'host.docker.internal|8080|local|hybrid' work/ods-latest/ARCHITECTURE.md
Do not execute by inference: install commands are only a future user-gated plan.

Allows claiming

checkreading
localexercised contract
upstreamdeclared capability

Remains open

runtimegate
health/chat/UIlive pilot
lifecyclebackup/restore/uninstall

Local sources

  • outputs/ods-adoption-dossier.md:39-60
  • work/ods-latest/ARCHITECTURE.md:114-151
  • work/ods-latest/ods/docs/MACOS-QUICKSTART.md:49-71
  • outputs/ods-check-receipt.md:131-182

Includes dossier, receipt, and official source by path:line.

04

Practice with feedback

1 · Prediction

2 · Stepper

1. classify the need

2. locate the proof

3. decide without overclaiming

3 · Scenarios

4 · Checklist

5 · Slide deck

In the minimum pilot, the web UI receives the request, Docker services route it, and the native macOS llama-server performs inference. Extensions add capabilities and also processes, data, and ports.

A check is not runtime.

decide without overclaiming

6 · Quiz

Which route preserves the minimum pilot?

7 · Active recall

Demo debate: code-understanding + prototype-animation; flowchart-diagram + research-feature-explainer; runner-up visual variants