# TypeSafe (jev) — required for evaluation; without it traces are still recorded.
TYPESAFE_API_KEY=
OPENEVALS_JEV_MODEL=jev-latest

# Server
OPENEVALS_PORT=3100
OPENEVALS_DB=./data/openevals.db
# Optional: if set, ingestion + REST require `Authorization: Bearer <key>` (or Basic <anything>:<key>)
OPENEVALS_API_KEY=

# Eval worker
OPENEVALS_EVAL_ENABLED=true
OPENEVALS_SETTLE_MS=5000          # wait this long after the last event before judging a trace
OPENEVALS_STATE_BUDGET_CHARS=90000 # ~22k tokens; jev allows 32k for state + longest question
OPENEVALS_REVIEW_CONFIDENCE=0.5   # judgments with min confidence below this are flagged needs_review

# Escalation: when jev's min confidence < OPENEVALS_REVIEW_CONFIDENCE (or a Noul lands within
# OPENEVALS_ESCALATE_NOUL_BAND of 0.5), re-ask the same questions to a reasoning model with rationale.
ANTHROPIC_API_KEY=
OPENEVALS_ESCALATE=true
OPENEVALS_ESCALATE_MODEL=claude-opus-5
OPENEVALS_ESCALATE_NOUL_BAND=0.15

# Per-step grading (the `step` evaluator): concurrency within a trace, max steps judged per trace (sampled evenly beyond), previous steps in each step's context
# OPENEVALS_STEP_CONCURRENCY=8
# OPENEVALS_STEP_MAX=150
# OPENEVALS_STEP_CONTEXT=8
# Rubric compiler model (POST /api/v1/evaluators/compile; needs ANTHROPIC_API_KEY)
# OPENEVALS_COMPILE_MODEL=claude-sonnet-5

# ---- Langfuse connector (the main path): pull observations, judge, write scores back ----
# LANGFUSE_HOST=https://cloud.langfuse.com
# LANGFUSE_PUBLIC_KEY=pk-lf-...
# LANGFUSE_SECRET_KEY=sk-lf-...
# LANGFUSE_POLL_MS=30000            how often to pull
# LANGFUSE_SETTLE_S=90              observations younger than this wait for the next poll
# LANGFUSE_OVERLAP_S=300            re-read this far behind the watermark for late observations
# LANGFUSE_LOOKBACK_S=3600          first start: how far back to pull
# LANGFUSE_ENVIRONMENTS=production  comma-separated; empty = all
# LANGFUSE_TRACE_NAMES=             comma-separated root/trace names; empty = all
# LANGFUSE_WRITE_BACK=true
# LANGFUSE_WRITE_BACK_SCORES=       comma-separated score names; empty = all EVAL scores
# LANGFUSE_WRITE_BACK_STEPS=true    also write per-step (observation-level) scores
# LANGFUSE_REVIEW_QUEUE_ID=         annotation queue for failed / low-confidence traces
# LANGFUSE_VERDICT_SCORE=passed     human score name treated as the verdict for calibration
# OPENEVALS_PUBLIC_URL=             this server's public URL, for links in Langfuse score metadata

# ---- jev budget ----
# OPENEVALS_JEV_RPM=1000            stay under jev's published 1,200 requests/min
# OPENEVALS_JEV_CONCURRENCY=16
# OPENEVALS_DAILY_BUDGET_USD=0      0 = unlimited; jev evaluators pause once reached

# ---- notifications ----
# OPENEVALS_WEBHOOK_URL=            POST {event, trace, failed[], unsure[]} when a trace fails a pass gate or a judgment is unsure