.env.example

# Copy this file to .env and fill in the values you need.
# Scripts in this repo source ${REPO_ROOT}/.env automatically.
# Everything below the "Required" section is commented out — uncomment only
# what you actually use.

# ─── Required ─────────────────────────────────────────────────────────────────
# Auditor agent API key. Default agent_cli is "claude" in all shipped configs.
ANTHROPIC_API_KEY=


# ─── Common paths ─────────────────────────────────────────────────────────────
# Where benchmark repos + datasets live, referenced from configs/**/*.yaml as
# ${BENCHMARK_REPOS_DIR}/<name>/{repo,data}. Wrapper (scripts/audit_one.sh)
# falls back to ./data/benchmark_repos when unset; running bench-audit
# subcommands directly requires this to be set to an absolute path.
# BENCHMARK_REPOS_DIR=/absolute/path/to/benchmark_repos

# Where audit outputs land. Required when running bench-audit subcommands
# directly; wrapper falls back to ./data/audit_runs when unset.
# AUDIT_RUN_DIR=/absolute/path/to/audit_runs


# ─── Optional ─────────────────────────────────────────────────────────────────
# Stream prompts + claude commands to stderr for debugging (0/1).
# BENCH_AUDIT_DEBUG_PROMPTS=0

# Batch static audit pipeline output dir (scripts/batch_static_audit_pipeline.sh).
# Defaults to ${REPO_ROOT}/data/multi_domain_all when unset.
# MULTI_DOMAIN_AUDIT_RUN_DIR=


# ─── Alternative agent backends ───────────────────────────────────────────────
# Only needed when overriding agent_cli to "codex" / "cursor" in a config,
# or when generating trajectories via Harbor.
# OPENAI_API_KEY=
# CLAUDE_CODE_OAUTH_TOKEN=


# ─── Per-suite paths ──────────────────────────────────────────────────────────
# Only needed when running the matching configs/ subtree.
# HARBOR_DATASETS_DIR=        # configs/harbor/*.yaml, configs/tb2/*.yaml
# TB2_JOBS_DIR=               # configs/tb2/eval_claude_code_sonnet.yaml
# SWEBENCH_DIR=               # configs/swe_bench/*.yaml


# ─── BenchGuard eval ──────────────────────────────────────────────────────────
# Only needed for benchmarks/benchguard/run_eval.sh (Gemini-based LLM judge).
# GEMINI_API_KEY=
# GOOGLE_API=                 # alias remapped to GEMINI_API_KEY by the script
# JUDGE_MODEL=gemini/gemini-3-flash-preview
# JUDGE_MAX_TOKENS=8192