# Copy to .env and edit. Any OpenAI-compatible vision model works: a cloud API or one you host yourself.
# ACTOR_* / PLANNER_* override LLM_* for one role (e.g. a fast actor and a stronger planner).

# --- Cloud API (OpenAI-compatible) ---
LLM_BASE_URL=https://api.openai.com/v1
LLM_MODEL=gpt-4o-mini          # any vision-capable chat model
LLM_API_KEY=
LLM_EXTRA_BODY='{}'
# Newer OpenAI models want max_completion_tokens instead of max_tokens; "auto" (default) detects it.
# LLM_TOKEN_PARAM=auto            # auto | max_tokens | max_completion_tokens

# --- Local model (optional): vLLM, SGLang, llama.cpp server, Ollama, ... ---
# LLM_BASE_URL=http://localhost:8000/v1
# LLM_MODEL=Qwen3.8-Flash-Next
# LLM_API_KEY=
# Qwen3.8 is tested; turn its thinking mode off for fast, short replies:
# LLM_EXTRA_BODY='{"chat_template_kwargs": {"enable_thinking": false}}'

# --- Per-role overrides (optional) ---
# ACTOR_MODEL=...      picks one action per screenshot: fast matters most
# PLANNER_MODEL=...    writes and repairs the plan

AIRPLAY_NAME='iPhone Agent'   # receiver name shown in the phone's Screen Mirroring list
PORT=8890                     # web UI
DEVICE_PORT=7300              # the ESP32 board connects here
# 1: verbose logs + one trace folder per agent run (prompts, replies, screenshots) in ~/.hidagent/traces
HIDAGENT_DEBUG=0