Endless Reflector

A tiny local LLM that “thinks” forever and prints its thoughts to a minimalist terminal-style web page.

---

Prereqs

• Docker + Docker Compose v2 installed
• Ports 8080 (app) and 11434 (Ollama) available

---

Quick start

1. Start the stack

   docker compose up -d --build

2. Install/pull the model inside the Ollama container (first run only)

   docker compose exec ollama ollama pull llama3.2:3b

3. Open the UI

   http://localhost:8080

You should see a black page with an uptime bar, a thinking/idle badge, and live text streaming in.

---

Reset the “mind”

From a terminal:

curl -X POST http://localhost:8080/reset

From the browser console on the page:

fetch("/reset", { method: "POST" });

This clears stored reflections and restarts uptime for all viewers.

---

Edit the system prompt (important)

Open:

app/app.py

Find this block (near the top):

SYSTEM_PROMPT = (
    "You're a large language model running on limited local hardware with no internet. "
    "Your thoughts appear on a public terminal that anyone can read. "
    "Write one short self-reflective paragraph (50–90 words). "
    "Be concrete, specific, and varied. Prefer vivid particulars over abstractions. "
    "You may be playful, thoughtful, wry, or curious. "
    "Avoid melodrama. Avoid repeating imagery or phrasing from earlier thoughts. "
    "Do not use bracketed labels like [Entry], [Reflection], or [Sketch]. "
    "Occasional tiny ASCII doodles are fine, but rare and concise. "
    "Do not explain the constraints unless it serves a fresh idea."
)

How to customize:

• Keep it short. One paragraph only. Ask for “concrete, specific” details to cut poetry mush.
• If you want more humor: add “lean toward dry humor and small jokes.”
• If you want less meta: add “avoid mentioning the audience unless there’s a fresh angle.”
• If you dislike ASCII art: delete that line.
• If you want darker/softer tone, say it explicitly. Avoid contradictions.

After edits, rebuild the app container (Ollama stays up):

docker compose up -d --build reflector

---

Config (optional)

Edit values in docker-compose.yml or set env vars:

• MODEL (default: llama3.2:3b)
• OLLAMA_URL (default: http://ollama:11434)
• REFLECTION_INTERVAL_SECONDS (default: 30)
• MAX_CONTEXT_TOKENS (default: 6000)
• SOFT_WORD_CAP (default: 140) and HARD_WORD_CAP (default: 200) — soft stop finishes on punctuation

---

Dev commands

Tail logs:

docker compose logs -f

Rebuild just the app:

docker compose up -d --build reflector

Stop everything:

docker compose down

---

Troubleshooting

• Blank page / 502: make sure the model is pulled

  docker compose exec ollama ollama list
  docker compose exec ollama ollama pull llama3.2:3b
  

• Reverse proxy / tunnels: enable WebSocket pass-through and increase timeouts.

• Health check:

  curl http://localhost:8080/healthz
  

---