openkruise/agents

Note: this is a candidate lfx mentorship program, and will be assign to the selected mentee What would you like to be added: load testing framework that can evaluate the performance of OpenKruise Agents in large scale cases using Kwok GitHub workflow and scripts to regularly run the load testing and report the evaluation results. Possible sandbox controller modifications to enable lightweight load testing. Why is this needed: rapid and large scale sandbox provision is crucial for agentic-workloads. OpenKruise agents had optimized rapid sandbox provisioning using techniques such as pooling and efficient sandbox discovering. To ensure no regression occurs, it is important to continuously test and evaluate the performance of OpenKruise Agents in a lightweight and resource efficient way. This project aims to build a load testing framework using tools such as Kwok to evaluate the performance of OpenKruise Agents in large scale cases. (e.g. 100,000 sandboxes), and construct a workflow to r

Summary Request to add comprehensive Prometheus metrics to the SandboxManager component to improve observability, monitoring, and debugging capabilities. These metrics will enable operators to track the performance and health of sandbox operations including pause/resume, claim, clone, snapshot, and routing. If you are new to OpenKruise Agents, this is a good chance to quickly familiarize yourself with the working mechanism of SandboxManager. Motivation Currently, the SandboxManager only exposes basic sandbox creation metrics (sandbox_creation_latency_ms and sandbox_creation_responses). As the system scales, we need detailed visibility into: Operation success rates and latencies Performance bottlenecks in different stages Routing synchronization health Resource lifecycle management Proposed Metrics Below are some example metrics expected to be exposed. If you believe there are other metrics worth exposing, please feel free to add them. 1. Pause Operation Metrics Metric Name Type La

What would you like to be added: Currently, when the Sandbox enters hibernation, it goes directly into the “Paused” state, and the success of hibernation is determined by whether the Condition SandboxPaused Status is true. This causes significant confusion for users. We would like to introduce a “Pausing Phase” so that the Sandbox only enters the “Paused” state after hibernation has successfully completed. Why is this needed:

What would you like to be added: add csi sidecar implementation for common opensource shared storage, e.g. ceph, minio Why is this needed: enable data sharing across sandboxes (such as common skills) and sandbox specific data backup

What would you like to be added: enable E2B sandbox api to accept an extra metadata, e.g. e2b.agents.kruise.io/cpu-scale-factor: 2 sandbox manager try to scale the cpu of the allocated sandbox based on the cpu-scale-factor [optional] return the sandbox immediately once we are sure that the resizing is feasible(PodResizingInProgress condition) Why is this needed: To reduce overall cost of maintaining a warm-up pool, the users desire to keep a pool of sandboxes with minimum resources, and once the sandboxes is claimed, the allocated sandbox can be resized inplace to a larger size. And the sandbox can even be return to the agent user even if the sandbox is still in the process of inplace resizing.