This directory publishes the current 60-task Standard60 snapshot as
BenchFlow-native task packages. The exact package list is tracked in
STANDARD60_MANIFEST.txt.
The public packages use the env0 runtime contract:
task.mduses BenchFlowschema_version: '1.3'.benchflow.env0.servicesdeclares the required publicmock-*services.environment/Dockerfilepinsghcr.io/benchflow-ai/env0:0.2.0.oracle/andverifier/use the native BenchFlow package layout.- Seed payloads under
/tasksare locked after build-time seeding.
The snapshot contains 11 auth, 8 email, 3 calendar, 10 document, 3 Drive,
12 multi-service, 8 Slack, and 5 Stripe tasks. Discord is not part of
Standard60; its runtime fixture remains under
example_tasks/discord-incident-followup.
tasks/ is a runnable public snapshot, not the task-authoring source of truth.
example_tasks/ remains the smaller fixture/template surface for env0 runtime
development.
Verify the exact snapshot and public-runtime invariants:
python3 -m unittest tests/test_standard60_tasks.pyRun BenchFlow structural checks:
while IFS= read -r task; do
bench tasks check "tasks/${task}" --level structural
done < tasks/STANDARD60_MANIFEST.txtRun the full oracle baseline:
BENCHFLOW_REWARD_LENIENT=1 bench eval run \
--tasks-dir tasks \
--agent oracle \
--sandbox docker \
--context-root . \
--concurrency 1 \
--build-concurrency 1 \
--jobs-dir .local/bf-oracle-standard60End-to-end evaluation requires Docker, the BenchFlow CLI, and either a locally
built or remotely published ghcr.io/benchflow-ai/env0:0.2.0 image.