space_env.example

BROWSER_ENV_LLM_PROVIDER=openai_compatible
OPENAI_BASE_URL=https://test--browser-env-vllm-a3.modal.run/v1
OPENAI_MODEL=Qwen3-30B-A3B-Instruct-2507-FP8
OPENAI_API_KEY=dummy
# LLM prompt/window controls (recommended for MiniWoB form tasks):
BROWSER_ENV_LLM_MAX_NEW_TOKENS=128
BROWSER_ENV_LLM_MAX_ELEMENTS=24
BROWSER_ENV_LLM_MAX_HISTORY=5
BROWSER_ENV_LLM_MAX_JUDGE_STEPS=16
# Optional LLM/oracle budget override aliases used by the env:
# BROWSER_ENV_ORACLE_BUDGET=3
# BROWSER_ENV_LLM_BUDGET=3
BROWSER_ENV_ORACLE_SUBMIT_AFTER_SELECT_REPEAT=1
BROWSER_ENV_CURRICULUM_ADVANCE_ON_FAILURE_AFTER=3
# Curriculum mode:
# - curated (default): fixed 4-task ladder
# - miniwob_full: include all registered MiniWoB tasks, ordered easy->medium->hard
# BROWSER_ENV_CURRICULUM_MODE=miniwob_full
# Optional cap for quick tests:
# Use 0 for no cap, or set 32/64 for a bounded broad run.
# BROWSER_ENV_CURRICULUM_MAX_TASKS=0
# Optional per-difficulty step budgets:
# BROWSER_ENV_MAX_STEPS_EASY=12
# BROWSER_ENV_MAX_STEPS_MEDIUM=24
# BROWSER_ENV_MAX_STEPS_HARD=48
# Optional extra headers as a JSON object:
# OPENAI_EXTRA_HEADERS_JSON={"Authorization":"Bearer ..."}

# MiniWoB must still be available to the environment runtime.
# If your Dockerfile sets it internally, you may not need this secret.
# MINIWOB_URL=file:///app/external/miniwob-plusplus/miniwob/html/miniwob/

# Disable LLM services entirely for pure env bring-up:
# BROWSER_ENV_DISABLE_LLM=1

# -----------------------------
# Additional optional variables
# -----------------------------

# Runtime/logging
# BROWSER_ENV_LOG_LEVEL=INFO
# BROWSER_ENV_HEADLESS=1
# BROWSER_ENV_ARTIFACT_DIR=artifacts

# LLM provider tuning
# BROWSER_ENV_LLM_TEMPERATURE=0.0
# BROWSER_ENV_LLM_TIMEOUT_SECONDS=60
# BROWSER_ENV_LLM_USE_JSON_MODE=1
# For hf_local provider:
# BROWSER_ENV_LLM_MODEL_ID=Qwen/Qwen2.5-3B-Instruct
# BROWSER_ENV_LLM_LOAD_IN_4BIT=0

# Curriculum controls
# BROWSER_ENV_CURRICULUM_PRESET=
# BROWSER_ENV_CURRICULUM_SUCCESS_MODE=cycle
# BROWSER_ENV_CURRICULUM_VARY_SEED=1
# BROWSER_ENV_CURRICULUM_SEED_STRIDE=997
# BROWSER_ENV_CURRICULUM_SHARED_START=0
# BROWSER_ENV_CURRICULUM_ALLOW_HTML_TASK_DISCOVERY=0

# Step budget adaptation and safety gates
# BROWSER_ENV_ENABLE_STEP_BUDGET_ADAPTATION=1
# BROWSER_ENV_STEP_BUDGET_INCREMENT=2
# BROWSER_ENV_STEP_BUDGET_MAX_EXTRA=8
# BROWSER_ENV_STEP_BUDGET_TRIGGER_FAILURES=max_steps_exceeded,low_progress_abort
# BROWSER_ENV_ORACLE_SUBMIT_AFTER_TYPE_REPEAT=1
# BROWSER_ENV_MAX_REPEATED_ACTIONS=3
# BROWSER_ENV_MAX_INVALID_ACTIONS=3
# BROWSER_ENV_MAX_NO_PROGRESS_STEPS=4
# BROWSER_ENV_MAX_NO_PROGRESS_STEPS_CAP=8
# BROWSER_ENV_NO_PROGRESS_STEP_RATIO=0.4

# Trackio wiring (if using hosted dashboard links)
# TRACKIO_SPACE_ID=
# TRACKIO_DASHBOARD_URL=
# BROWSER_ENV_TRACKIO_SPACE_ID=

# Hugging Face / Space runtime tokens and IDs
# HF_TOKEN=
# HUGGING_FACE_HUB_TOKEN=
# HF_SPACE_ID=
# SPACE_ID=

# Optional network override
# NO_PROXY=localhost,127.0.0.1

# AgentLab study output root
# AGENTLAB_EXP_ROOT=artifacts/agentlab

# -----------------------------
# Modal/vLLM server variables
# (used by modal_vllm_server.py)
# -----------------------------
# MODEL_NAME=Qwen/Qwen3-30B-A3B-Instruct-2507-FP8
# MODEL_REVISION=
# SERVED_MODEL_NAME=Qwen3-30B-A3B-Instruct-2507-FP8
# MODAL_GPU=L40S
# TENSOR_PARALLEL_SIZE=1
# MAX_MODEL_LEN=32768
# DTYPE=auto
# FAST_BOOT=1
# MODAL_MIN_CONTAINERS=1
# MODAL_APP_NAME=browser-env-vllm-a3
# MODAL_WEB_LABEL=browser-env-vllm-a3

# Internal hub behavior toggle used in Modal image setup
# HF_HUB_ENABLE_HF_TRANSFER=0