Running Harbor Job in ACK

1. Base Environment

Start your container instance for development using the following base image:

alibaba-cloud-linux-3-registry.cn-hangzhou.cr.aliyuncs.com/alinux3/python:3.11.1

This image comes with Python 3.11 pre-installed. Harbor requires Python >= 3.12, which will be installed via uv.

2. Environment Initialization

Version requirement: You must use the feat/add-ack-environment branch:
https://github.com/alibaba/harbor/tree/feat/add-ack-environment

Run the following commands on your instance to set up the environment:

yum install git vim -y
curl -LsSf https://astral.sh/uv/install.sh | sh
curl -sL "https://github.com/moby/buildkit/releases/download/v0.18.2/buildkit-v0.18.2.linux-amd64.tar.gz" | \
    tar -xz -C /usr/local/bin --strip-components=1 bin/buildctl

git clone https://github.com/alibaba/harbor.git
cd harbor
git checkout feat/add-ack-environment
source $HOME/.local/bin/env
uv venv --python=3.12
source .venv/bin/activate
uv pip install -e .
uv pip install kubernetes
export BAILIAN_API_KEY=<YOUR API KEY>
mkdir ~/.kube
cat > ~/.kube/config << EOF
apiVersion: v1
clusters:
- cluster:
   ...
EOF

3. Start BuildKit Service (Optional)

If using BuildKit for image building (recommended), start the BuildKit daemon on your instance:

# Install buildkitd
curl -sL "https://github.com/moby/buildkit/releases/download/v0.18.2/buildkit-v0.18.2.linux-amd64.tar.gz" | \
    tar -xz -C /usr/local/bin --strip-components=1 bin/buildkitd

# Start BuildKit daemon (listening on port 1234)
buildkitd --addr tcp://0.0.0.0:1234 &

4. Submit a Harbor Job

Once the environment is ready, submit a SWE-Bench Verified evaluation job:

OPENAI_API_KEY=$BAILIAN_API_KEY \
OPENAI_BASE_URL=https://dashscope.aliyuncs.com/compatible-mode/v1 \
harbor run \
  -d swe-bench/swe-bench-verified \
  -a swe-agent \
  --ak total_cost_limit=0 \
  --ak per_instance_cost_limit=0 \
  -m openai/qwen-max \
  -e ack \
  --ek kubeconfig=~/.kube/config \
  --ek namespace=default \
  --ek registry=registry.cn-hongkong.aliyuncs.com/swe-bench \
  --ek use_sandbox_claim=true \
  --ek use_buildkit=true \
  --ek buildkit_address=tcp://<buildkit service> \
  --ek image_pull_secret=acr-registry \
  -n 1 \
  -l 1 \
  --force-build \
  --debug

Parameter Reference

Parameter	Description
-d swe-bench/swe-bench-verified	Dataset: SWE-Bench Verified
-a swe-agent	Agent: SWE-Agent
--ak total_cost_limit=0	Remove total cost limit (unlimited)
--ak per_instance_cost_limit=0	Remove per-instance cost limit (unlimited)
-m openai/qwen-max	Model: Qwen-Max via Bailian OpenAI-compatible endpoint
-e ack	Environment type: ACK (Kubernetes)
--ek kubeconfig=~/.kube/config	K8s kubeconfig path
--ek namespace=default	K8s namespace
--ek registry=...	Image registry URL (Hong Kong ACR)
--ek use_sandbox_claim=true	Enable OpenKruise SandboxClaim for warm sandboxes
--ek use_buildkit=true	Enable BuildKit for image building
--ek buildkit_address=...	BuildKit daemon address
--ek image_pull_secret=acr-registry	Private registry pull secret name
-n 1	Number of concurrent trials
-l 1	Run only 1 task (for debugging)
--force-build	Force rebuild the image
--debug	Enable debug logging

Version

• --ek use_sandbox_claim=true is the key to enable openkruise agents
• -e ack running openkruise agents in ACK cluster