.env.example

# Copy to .env and fill in. .env is git-ignored and never logged.

# --- OpenRouter (optional: without a key AgentReel runs its local templates) ---
OPENROUTER_API_KEY=
# Pin capable free models first. "openrouter/free" picks any free model with capacity (including
# tiny ones that often return broken JSON), so it is only the last resort. Check openrouter.ai/models.
OPENROUTER_MODEL=nvidia/nemotron-3-super-120b-a12b:free
OPENROUTER_FALLBACK_MODELS=google/gemma-4-31b-it:free,poolside/laguna-s-2.1:free,nvidia/nemotron-3-ultra-550b-a55b:free,openrouter/free
# Paid models are skipped unless this is true. Models whose price cannot be confirmed count as paid.
ALLOW_PAID_MODELS=false
# Global ceiling across all jobs (tracked in outputs/.spend_ledger.json).
MAX_API_SPEND_USD=5
OPENROUTER_TIMEOUT_S=90

# AI_PREFERRED | LOCAL_ONLY | AI_REQUIRED
AGENTREEL_EXECUTION_MODE=AI_PREFERRED

# --- Narration ---
# chatterbox (default): Chatterbox-Turbo, local on the GPU. Needs .venv-tts and the model in models/chatterbox-turbo
# (see README). Voices: "default" plus every 5-20 s reference clip in assets/voices.
TTS_ENGINE=chatterbox
CHATTERBOX_VOICE=default
TTS_RATE=+10%
# The old online Microsoft voices are kept but switched off. To use them again set TTS_ENGINE=edge;
# then TTS_PROVIDER (edge or sapi) and TTS_VOICE apply.
TTS_PROVIDER=edge
TTS_VOICE=en-US-AriaNeural
# GPU rendering: auto uses NVENC encoding + Chrome's GPU backend when an NVIDIA GPU works, on forces it,
# off keeps everything on the CPU. run.py --gpu overrides this.
AGENTREEL_GPU=auto

# --- Limits ---
MAX_REPAIR_ATTEMPTS=2
MAX_OUTPUT_MB=100
# Optional online research via the public Wikipedia API (article URLs are kept as sources).
# Off by default: research then uses only the documents and text you supply.
ONLINE_RESEARCH=false
# Required by the Wikimedia User-Agent policy when ONLINE_RESEARCH=true (your email or a URL).
AGENTREEL_CONTACT=
# AGENTREEL_OUTPUT_DIR=outputs

# --- Gemini (optional second provider; pick it with --provider gemini) ---
# Free key from Google AI Studio. Free-tier prompts may be used by Google to improve its models.
GEMINI_API_KEY=
GEMINI_MODEL=gemini-3.8-flash
GEMINI_FALLBACK_MODELS=gemini-3.5-flash-lite
# Rate limiting for the free tier: minimum seconds between requests (6 = at most 10 per minute),
# how long to wait out a per-minute limit before switching model, and the longest wait when every
# model is cooling down. Daily quota errors are never retried; that model is skipped until midnight Pacific.
GEMINI_MIN_INTERVAL_S=6
GEMINI_RETRY_WAIT_S=20
GEMINI_MAX_WAIT_S=90
# openrouter | gemini (overridden by --provider on run.py or the Streamlit launch)
LLM_PROVIDER=gemini