Initial commit: arr-stack homelab + local AI

Docker Compose media stack (gluetun, *arr, Jellyfin), Ternary-Bonsai AI
(Open WebUI, Forgejo, optional OpenHands), and Ansible bootstrap for the
SER5 Ubuntu host.

.env.example

@@ -0,0 +1,74 @@
+# Copy to .env and adjust. The compose file reads these values.
+
+# --- user / group that owns your media (run `id` on the host) ---
+PUID=1000
+PGID=1000
+
+# --- your timezone (https://en.wikipedia.org/wiki/List_of_tz_database_time_zones) ---
+TZ=Europe/Berlin
+
+# --- host paths (relative to this folder is fine) ---
+CONFIG_DIR=./config
+DATA_DIR=./data
+
+# --- your LAN subnet, so the qBittorrent WebUI stays reachable through
+#     gluetun's firewall. If your machine is 192.168.1.50, use 192.168.1.0/24
+LAN_SUBNET=192.168.1.0/24
+
+# =============================================================================
+# Local AI (Ternary-Bonsai-27B on SER5 PRO / Ryzen 7735HS)
+# =============================================================================
+
+# GGUF weights directory (populated by scripts/download-bonsai-model.sh)
+BONSAI_MODELS_DIR=./models/bonsai
+BONSAI_MODEL_PATH=/models/Ternary-Bonsai-27B-Q2_0.gguf
+
+# cpu (default) or vulkan — rebuild image after changing:
+#   BONSAI_BACKEND=vulkan docker compose build --no-cache bonsai
+BONSAI_BACKEND=cpu
+
+# GPU layers: 0 = CPU only. With vulkan image + /dev/dri, try 99.
+BONSAI_NGL=0
+
+# Context size. 16384 is a solid default on 32–64 GB RAM with the Arr stack.
+# Raise for long docs / repo work; set BONSAI_KV4=1 if you run OOM.
+BONSAI_CTX=16384
+BONSAI_KV4=0
+
+# Leave 0 for auto threads, or e.g. 12 on 16-thread 7735HS to leave headroom
+BONSAI_THREADS=0
+
+# Sampling (Prism 27B thinking-mode defaults)
+BONSAI_TEMP=0.7
+BONSAI_TOP_P=0.95
+BONSAI_TOP_K=20
+
+# Host port for llama-server API (OpenAI-compatible at /v1)
+BONSAI_PORT=8081
+
+# Open WebUI
+OPEN_WEBUI_PORT=3000
+OPENAI_API_KEY=sk-local-bonsai
+# true = login required (recommended); false = open LAN access
+WEBUI_AUTH=true
+
+# Folder mounted read-only into Open WebUI as /workspace (RAG + file Q&A)
+# Point this at your git projects, notes, docs, etc.
+AI_WORKSPACE_DIR=./workspace
+
+# SearXNG (web search for RAG) — published on localhost only
+SEARXNG_PORT=8888
+SEARXNG_SECRET=change-me-to-a-long-random-string
+
+# OpenHands coding agent (docker compose --profile coding up -d)
+OPENHANDS_PORT=3001
+# LiteLLM-style id; after first start check http://host:8081/v1/models
+OPENHANDS_MODEL=openai/local
+
+# Forgejo — self-hosted Git (web + SSH clone)
+FORGEJO_HTTP_PORT=3002
+FORGEJO_SSH_PORT=2222
+FORGEJO_DOMAIN=192.168.8.123
+FORGEJO_ROOT_URL=http://192.168.8.123:3002/
+# true after you create the admin account (blocks random signups)
+FORGEJO_DISABLE_REGISTRATION=false