.env.example

# Chief Engineer — config template. Copy to `.env` and fill in as needed.
# `.env` is LOADED AUTOMATICALLY at import (core/__init__.py); real shell env vars
# always win over the file. `.env` is gitignored — never committed, never uploaded
# to the Space (keep your HF_TOKEN here safely).
#
# TWO CONTEXTS:
#   • LOCAL (this .env)        → dev + recording via Ollama. Use the LOCAL block.
#   • SPACE (HF Space)         → do NOT push a .env. Set these same KEY=VALUE pairs
#                                as Space *Variables* (and HF_TOKEN as a *Secret*)
#                                in Settings, or via the huggingface_hub API
#                                (see docs/SPACE-DEPLOY.md). Use the ZEROGPU block.

# ── LOCAL (Ollama) ────────────────────────────────────────────────────────
# Primary model tag. gemma4:e4b = gemma4:latest (~9.6GB). NOTE: gemma4:4b does NOT exist.
CHIEF_ENGINEER_MODEL=gemma4:e4b          # Kyle's pick (6/11). gemma4:e2b is ~2× faster if you want snappier.
# Force the deterministic fallback (skip Ollama) — handy for fast offline B-roll:
# CHIEF_ENGINEER_OFFLINE=1

# ── ZEROGPU / SPACE (live transformers inference) ──────────────────────────
# Set these on the ZeroGPU Space (Variables), not locally, unless you have a local GPU.
# CHIEF_ENGINEER_BACKEND=zerogpu                       # default 'ollama'; 'zerogpu' = live HF model
# CHIEF_ENGINEER_HF_MODEL=google/gemma-4-E2B-it        # transformers-native (resolved 6/10)
# CHIEF_ENGINEER_GPU_SECONDS=90                         # 1st call loads the model
# CHIEF_ENGINEER_MAX_NEW_TOKENS=512
# HF_TOKEN=hf_xxx                                       # ONLY if the model is gated (Space SECRET; keep out of git)