.env.example

# Copy this to .env. docs/configuration.md explains every knob in
# here.
DOMAIN=localhost

EMAIL=admin@localhost

TLS_MODE=off

# Address we publish on: nginx under docker, caddy on bare metal
# windows. 127.0.0.1 = this machine only. 0.0.0.0 = anyone who can
# reach this box, which is what a public TLS_MODE needs, what
# reading Jun on your phone over the house wifi needs, and what
# nothing else should have. Either way it's HTTPS, on a
# self-signed cert unless TLS_MODE=on, and plain http only
# redirects to it.
BIND_ADDR=127.0.0.1

# Every other name or address this box answers to. Both nginx and
# php refuse a request whose Host is not DOMAIN, localhost or
# 127.0.0.1 (nginx with a 444, php with a 421), so opening the
# phone at https://192.168.1.42 needs that exact address in the
# list. You do NOT normally have to fill this in: once BIND_ADDR
# is off loopback, start.sh finds this machine's own private
# addresses and adds them, and prints them as "reachable as:".
# This line is for what it can't guess: an mDNS name, a
# tailscale address, whatever a proxy calls you. Space or comma
# separated. BIND_ADDR still has to be 0.0.0.0, this alone does
# nothing.
# OMEGA_EXTRA_HOSTS=jun.local

# Signing up asks for this key when it is set. Both installers
# write one when the line is missing and print it at the end, and
# they never fill in a line you emptied yourself, so an upgrade
# cannot turn the gate back on behind your back. The first
# account skips the key on a fresh install. Later accounts must
# present it. Empty or no line means anyone who reaches the page
# can sign up.
# OMEGA_REGISTRATION_KEY=

AI_PROVIDER=ollama

# Careful, OpenRouter sends what you type to their cloud.
# OPENROUTER_API_KEY=
# OPENROUTER_MODEL=openrouter/auto

# LLAMACPP_URL=http://llamacpp:8080
# LLAMACPP_MODEL_HF=efficiencyx/Jun-LoRA-E2B-GGUF:Q4_K_M
# LLAMACPP_PORT=8081
# Serve a GGUF (the model file llama.cpp loads) from disk
# instead of pulling MODEL_HF. The file must sit inside
# MODELS_DIR. start.sh mounts that directory and checks the file
# exists.
# LLAMACPP_MODELS_DIR=./models
# LLAMACPP_MODEL_FILE=Jun-12B-Q4_K_M.gguf
# Multi-token prediction, Gemma 4 only. Point this at the MTP
# assistant repo for the SAME Gemma 4 size Jun was fine-tuned
# from (E2B here) and llama.cpp drafts a few tokens ahead with
# the little model, then checks them against the real one in a
# single pass. Nothing gets accepted that Jun wouldn't have said
# anyway, the tokens the drafter got right just came cheap. It
# costs its own VRAM though. Empty turns it off, this is not a
# flag, it is the repo name.
# ./mtp-autotune.sh fills it in from LLAMACPP_MODEL_HF when it
# is empty, same repo with -MTP in the name. Set it here only
# when the drafter is elsewhere.
# LLAMACPP_MTP=amaranus/Gemma-4-E2B-it-qat-assistant-MTP-Q8_0-GGUF
# How many tokens to draft ahead. Shallow usually wins (on a
# 3060 depth 1 beat depth 4 by 28%), and ./mtp-autotune.sh
# measures it for your card instead of guessing. llama.cpp takes
# this at startup, so the tuner restarts llama-server once per
# depth.
# LLAMACPP_MTP_N_MAX=1
# Turn tools off for fine-tunes whose tool-call syntax llama.cpp
# cannot parse. llama.cpp 500s the request instead of degrading.
# Off costs cross-chat recall, which runs on tool calls.
# LLAMACPP_TOOLS=on

COMPOSE_PROFILES=ollama

OLLAMA_URL=http://ollama:11434

# JUN_PORT=8080
# JUN_PHP_PORT=8079

OLLAMA_MODELS_TO_PULL=hf.co/efficiencyx/Jun-LoRA-E2B-GGUF:Q4_K_M

# Dedicated model used to title new conversations. Pinned to CPU
# and kept loaded so it never takes VRAM from the chat model.
# Empty falls back to truncating the first user message instead.
TITLE_MODEL=hf.co/efficiencyx/Titlewen-GGUF:F16

# Multi-token prediction for the Ollama side, Gemma 4 only. Name
# the MTP assistant repo for the same Gemma 4 size Jun was
# fine-tuned from and the entrypoint pulls it, then builds a
# `jun-mtp` model that carries it as a DRAFT layer. The little
# model guesses a few tokens ahead, Jun checks them in one pass,
# the ones she agrees with came cheap. Nothing gets said that she
# wouldn't have said anyway. Empty turns it off, this is the repo
# name and not a flag.
#
# Off by default because whether it pays depends on the card, not
# on Jun. The installer offers it, and ./mtp-autotune.sh measures
# every depth here and keeps the fastest. Pair the drafter with
# the SAME Gemma 4 size AND the same QAT (quantization-aware
# trained) branch she was fine-tuned from. A mismatched one still
# runs, it just guesses wrong much more often (2.10 accepted
# tokens per pass against 2.74) and nothing says why. Left empty
# the tuner derives it from the chat model, same repo with -MTP in
# the name, so Jun-LoRA-12B-GGUF drafts off Jun-LoRA-12B-MTP-GGUF.
# Set it here only when the drafter lives somewhere else, a value
# in this file is never overwritten.
# OLLAMA_MTP=hf.co/Janvitos/gemma-4-12B-it-qat-assistant-MTP-Q8_0-GGUF:Q8_0
# How many tokens the drafter proposes per pass, baked in as
# draft_num_predict. Has to be a number, "auto" is a question for
# ./mtp-autotune.sh and not a value. The entrypoint falls back
# to 1 for anything non-numeric. Measured on a 3060 with the 12B
# on prose: no drafter 36.1 tok/s, depth 1 45.3, depth 2 42.0,
# depth 3 38.2, depth 4 35.4. Deeper is not better, every extra
# token in the batch she has to check costs about 10ms more.
# OLLAMA_MTP_N_MAX=1

# Written by the tuner, not by hand. It names the card the depth
# above was measured on, and start.sh / start.ps1 hold it against
# whatever is in the box now. Different card, they run the tuner
# again before handing you the app.
# MTP_TUNED_GPU=
# on by default. off stops the start scripts re-measuring after a
# GPU change. Worth knowing before you leave it on: the llamacpp
# sweep restarts llama-server once per depth, so that boot is not
# a quick one.
# MTP_AUTOTUNE=on

VOICE=on

# off lets Jun still walk out mid-scene without locking Anon out
# afterwards.
FLEE_BANS=on

# Going out (the shop, karaoke, a meal) is Jun's decision: the
# pages bounce back home unless she agreed in chat. on drops that
# gate and shows the "Force her" buttons to every account, not
# just admins.
FREE_ROAM=off

# The voice image is CPU-only: cuda costs ~2GB VRAM for little
# synthesis speedup.
TTS_DEVICE=cpu

# Karaoke is a separate sidecar so it can take the GPU while
# voice stays on CPU. off leaves that container (a few GB) out of
# the build entirely.
KARAOKE=on

# cpu | cuda | auto: device for splitting a song into stems.
# Worth the VRAM here: minutes on CPU vs seconds on a card, and
# it's handed back afterwards.
SEP_DEVICE=auto

# base + empty STT_LANG = multilingual with per-utterance
# auto-detect. Use base.en + STT_LANG=en for English-only (a bit
# faster/sharper), or size up the multilingual model (small,
# medium, large-v3) for better non-English STT.
STT_MODEL=base
STT_LANG=

# Decoded-audio limits apply after decompression, so a tiny
# compressed upload cannot expand until it exhausts memory.
# Expensive jobs are serialized.
# STT_MAX_DURATION_S=120
# STT_MAX_CONCURRENT=1
# TTS_MAX_CONCURRENT=2
# TTS_MAX_QUEUE=8
# SEP_MAX_DURATION_S=900
# SEP_MAX_CONCURRENT=1
# SEP_MAX_JOBS=4

OLLAMA_FLASH_ATTENTION=1
OLLAMA_KV_CACHE_TYPE=q8_0

# auto orders cards biggest-VRAM-first, so the largest is
# device 0. A comma list (UUIDs or indices) picks specific
# cards. all leaves the driver's order alone.
# GPU_DEVICES=auto

# on splits one model across every GPU (ollama:
# OLLAMA_SCHED_SPREAD, llama.cpp on CUDA: --split-mode row).
# Slower per token, but fits more.
# TENSOR_PARALLEL=off

# OLLAMA_NUM_PARALLEL=1
# OLLAMA_MAX_LOADED_MODELS=3
# OLLAMA_KEEP_ALIVE=5m

# Context window. Sized from the VRAM left over after the model
# weights, falling back to system RAM when the card is unknown or
# too small to matter. Set this by hand to override both.
# OMEGA_NUM_CTX=8192

# Ceilings on one chat turn against the model server. TURN is the
# wall clock for the whole turn, tool rounds included. IDLE hangs
# up when nothing has arrived for that long, and has to cover a
# cold load plus prompt eval on a big context.
# OMEGA_TURN_TIMEOUT_S=900
# OMEGA_STREAM_IDLE_S=300

# Card size in MiB. start.sh probes this from nvidia-smi/rocm-smi.
# Set it only to report less than the card really has, e.g. when
# another app permanently owns part of the VRAM. Also gates the
# partial-offload refit in api/lib/providers/ollama.php.
# OMEGA_GPU_VRAM_MB=12288

# OMEGA_STATE_DIR=/var/lib/omega
# MEMORY_DIR=/var/lib/omega/memory

# Localhost port the consolidation worker listens on for per-user
# data keys pushed by php (api/lib/crypto.php key_push). Only matters on
# a bare metal install where something else may already hold it.
# Inside the php container it is never reachable from outside.
# OMEGA_KEY_PORT=9099

TTS_URL=http://tts:8001
KARAOKE_URL=http://karaoke:8001

# Shared secret php sends the tts/karaoke sidecars in
# X-Sidecar-Secret. start.sh writes a random one here when the
# line is missing. Leave it empty to run the sidecars on their
# Host allowlist alone.
# SIDECAR_SECRET=

# Container resource ceilings. Raise a model limit if you
# deliberately serve a larger model. Do not remove the ceiling on
# a network-reachable deployment.
# PHP_MEMORY_LIMIT=1g
# OLLAMA_MEMORY_LIMIT=16g
# LLAMACPP_MEMORY_LIMIT=16g
# TTS_MEMORY_LIMIT=8g
# KARAOKE_MEMORY_LIMIT=12g


# A signed-in account that sends this key gets promoted to admin.
# Empty turns promotion off.
# OMEGA_DEV_KEY=

#                                                          ░░▒░░                          
#                                                        ░░▒▒▒░░░                         
#                                                      ░▒▒▒▒▒▒░░░                         
#                         ░░░░░░░░░░      ░░░░░░░    ░░▒▒▒▒▒▒▒▒░░░                        
#               ░           ░░░▒▒▒▒▒░░░░▒▒▒▒░░░    ░░▒▒▒▒▒▒▒▒▒▒░░░░░                      
#      ░░▒▒▒▒▒▒▒▒░░░░░░░░  ░░░▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒░░▒▒▒▒▒▒▒▒▒▒▒▒▒░░▓█▒                      
#      ░░░▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒░▓██▒                      
#      ░░░▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒██▒                       
#      ░░░░▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒░░░▒▒▒▒▒░▓▒░                       
#      ░░░░░▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒░▒░▒▒▒▒▒▓▓▓░                      
#       ░▒▓▒░░▒▒▒▒▒▒▒░░▒▒░░▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒░░▒▒▒▒▒▒▒▒▒▒░▒▒▒▓▓▒░░░░░░                 
#       ░▓██▓░░▒▒▒▒▒░░▒▒░░▒▒▒▒▒▒▓▒▒▒▒▓▓▓▒▒▒▒▒░░▒░░▒░▒▒▒▒▒▒▒▒▒░▒▒░░░▒░░░░░░░               
#         ░░▓▓░▒▒▒▒░░▒▒░░▒▓░▓▒▒▓▓▒▒▓███▒▒▒▒▒▒░░░░▒▓▒░▒▒▒▒▒░▒▒▒░▒░░▒▒▒▒▒░░░░░              
#          ░░█▓▒▒▒░░▒▒▒░░░░▒▒▒▒▒▒▒▒▓▓▓▓▒▒▒▒▒░▒▒░░▒█▓▒▒▒▒▒▒▒░▒▒░░░░░▒▒▒▒▒▒▒▒░              
#    ░░░░░▒░░▒▒▒▒░░░▒▒░░░░▒▒▒▒▒▒▒▒▒▒▒▒▒▓▓▒░▒▓█▒░░▒██▒░▒▒▒▒▒░▒▒▒░▒▒░░▒▒▒▒▒▒░               
#   ░░░░░▒▒▒▒░░▒▒░░▒▒▒░░░░▒▒▒▒▒▒▒▒░░▒▒▒▒▒░▒███▒░▒███▓░▒▒▒▒▒░░▒▒░▒░░░░▒▒░░░                
#   ░░░▒▒▒▒▒░░▒▒▒░░▒▒▒░░░░▒░░░░░░▒░░▒▒░▒▓██▓▒▒░░░▒▒▒▒░░▒▒▒░░░▒▒░▒░░░░░▒░░                 
#   ░▒▒▒▒▒░░░▒▒▒▒░░▒▒░░░░░░░░░░░░░░░░░▒█████▒░░░▒▒▒▒▒░░░▒▒░░░▒▒░▒▒░░░░▒▒░                 
#   ░▒▒▒░░░░░▒▒▒▒░░▒▒░░░▒▒▒▒░░░░░░░▒▓██████▓▒▒▒▒▒▒▒▒▒▒░░▒▒░░░▒▒░▒▒░▒▒░░▒░                 
#     ░░▒▒░░░░▒▒▒░░▒░░░░▒▒▒░░▒░░░▓█████████▓▒▒▒▒▒▒▒▒▒▒░░▒▒░░░▒▒░▒▒░▒▒▒▒▒░                 
#      ░▒▒░░░░▒▒▒▒░░░░░░▒▒░▒▒░░▒████████████▒▒▒▓▓▓▓▓▒░ ░▒░░░░▒▒░▒▒░▒▒▒▒▒░                 
#     ░▒▒░░░░░▒▒▒▒░░░░░░░░▒▓▒░▒██████████████▒▒▒▓▓▓▒░░░░░░░░▒▒▒░▒░░▒▒▒▒░░                 
#     ░▒▒░░▒░░▒▒▒▒▒░░░░░░░▒▒░▓████████████████▓▒▒░▒▓▓▓░░░░░░▒▒░▒▒░▒▒▒▒░░                  
#      ░▒▒▒▒░░▒▒▒▒▒▒░░░░▒▓▓▓█████████████████████████▓░░░░░░▒▒░░░▒▒▒▒░                    
#       ░▒▒▒▒░░▒▒▒▒▒░░░░▒████████████████████████████▒░░░░░▒▒░░░░▒▒░░                     
#        ░░▒▒░░░▒▒▒▒░░░░▒▓███████████████████████████▒░░░░░░░░▒░░░░                       
#          ░░▒░░░▒▒▒░░░░░░░░▒▓▓███████████████████▓▓▒░░░▒░░░░░░░░░░                       
#           ░░░░░▒░░░░░░░▒▒░░░░░░░▒▒▓▓▓▓▓▓▓▓▒▒▒░░░░░░░░▒▒░░░░░░░                          
#            ░░ ░░░░░░░▒░░░▒▒▒▒▒░░░░░▒▒▒░░░░░▒▒▒▒░░▒░▒▒▒▒░░░░                             
#               ░░░░░░░▒▒▒▒░░░▒▒▒▒▒▒░▒█▓▒▒▒▒▒▒░░▒▒▒▒▒▒▒▒▒▒░░                              
#                ░░  ░░░░▒▒░░▒▒▒░▒▒▒▒▒▒▒▒▒▒▒░░▒▒▒▒▒▒▒▒░░░░                                
#                         ░░▒▒▒▒░▒▒▒▒▒▒▒▒▒▒▒░▒▒▒▒▒▒░░░                                    
#                         ░▒▒▒▒▒░▒▒▒▒▒▒▒▒▒▒▒░▒▒▒▒▒▒▒▒░░                                   
#                       ░░▒▒▒▒▒▒░▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒░░                                     
#                      ░░▒▒▒▒▒▒▒░▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒░░                                     
#                      ░▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒░░                                     
#                    ░░▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒░░                                     
#                    ░▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒░                                     
#                   ░▒▒▒▒▒▒▒░▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒░░                                    
#                  ░░▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒░                                    
#                 ░░▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒░                                   
#                 ░░▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒░                                   
#                  ░▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒░                                   
#                 ░░░▒▒▒▒▒▒▒▒▒▒▒▒▓▓▓▓▓▓▓▓▓▓▓▓▓▒▒▒▒▒▒▒▒░                                   
#                ░░▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▒▒░░                                    
#                ░▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▒░                                   
#               ░▒▒▒▒▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▒░                                  
#              ░▒▓▓▒░▒▓▓▓▓▓▓▓▓▓▓▓▓▒▒▒▒▒▒▒▒▒▒▒▓▓▓▓▓▓▓▓▓▓▒░░                                
#                ░▒▒▒▓▓▓▓▒▒▒▓▓▓▓▒░         ▒▓▓▒▒▓▓▓▓▒▒▒▓▓▓░                               
#                   ░▒▒▒▓▓▒▒▒▒░            ▒▓█▒▒▒▒▒▒▒▒▓▒░░                                
#                   ░▒▒▒▒▒▒▒▒▒░            ░▒▒▒▒▒▒▒▒▒░░                                   
#                  ░░░▒▒░░▒▒▒░░            ░░▒▒▒░░░▒▒▒░                                   
#                   ░░░░░░░░░               ░░░░░░░▒░░                                    
#                                                 ░░░