.env.sample

AI_GATEWAY_API_KEY=
# Any Vercel AI Gateway model id works here, e.g. openai/gpt-4.1-mini,
# anthropic/claude-sonnet-4-5, or another gateway-supported provider/model.
AGENT_BENCH_JUDGE_MODEL=openai/gpt-4.1-mini
AGENT_BENCH_COST_PER_1K_TOKENS_USD=0.003
LLM_JUDGE_RESPONSE_CACHE=true
LLM_JUDGE_SYSTEM_PROMPT=

# Judge robustness (all opt-in; require AI_GATEWAY_API_KEY to take effect).
# Number of independent judge passes; >1 runs a panel with distinct reviewer
# lenses and takes the median score/qualityScore (reduces single-judge bias).
AGENT_BENCH_JUDGE_SAMPLES=1
# Attach a rendered screenshot of the produced index.html as multimodal evidence
# for visual/frontend tasks (needs Playwright + a multimodal-capable judge model).
AGENT_BENCH_JUDGE_SCREENSHOT=false
# Optional explicit Chromium binary for the screenshot judge (else Playwright's).
AGENT_BENCH_CHROMIUM_PATH=

# Score profile override. Leave empty to auto-pick per task (sandbox runs use
# "hybrid", trace tasks use "trace"). Set to force one of:
#   hybrid | artifact | trace | judge | state | craft
# "craft" is an opt-in composite that also weights the judge's code-quality
# score and the workflow/process score the default profiles leave unweighted
# (outcome .45 / process .15 / review .15 / quality .15 / efficiency .1).
AGENT_BENCH_SCORE_PROFILE=

# Sandbox execution: auto (default) prefers Docker when the daemon + image are
# available, then macOS seatbelt, then falls back to a plain host process.
# Set to "docker" to require Docker explicitly.
AGENT_BENCH_SANDBOX_PROVIDER=auto
AGENT_BENCH_SANDBOX_DOCKER_IMAGE=node:22-bookworm-slim