Please provide an in-depth description of the question you have:
When I go though this getting started guidance: https://github.com/volcano-sh/agentcube/blob/main/docs/getting-started.md, I have question with the CodeInterpreter CRD. The CodeInterpreter CRD is designed to work hand-in-hand with the Python SDK(or future other SDKs), providing an out-of-the-box code execution experience. The SDK exposes some methods, all of which communicate with the PicoD daemon running inside the sandbox.
However, the current CRD design exposes template.image as a user-configurable field:
spec:
template:
image: ghcr.io/volcano-sh/picod:latest # User can change this
args:
- --workspace=/root
This quite confuse me:
- The Python SDK (CodeInterpreterClient) depends on PicoD's HTTP API (/api/execute, /api/files) inside the sandbox. If a user sets a custom image that doesn't include PicoD, the SDK simply won't work.
- If the user needs a fully custom runtime, AgentRuntime is the right choice — but it doesn't support warmPoolSize.
- SandboxClaim could combine warm-pool with custom images, but it's an internal detail that shouldn't be user-facing.
What do you think about this question?:
Here are some directions I think:
- Make image default to the official PicoD image and clearly document that custom images MUST implement the PicoD HTTP API contract (/api/execute, /api/files, /health), otherwise the SDK will not work.
- Ship a default CodeInterpreter CR with the Helm chart — so that SDK users get a zero-config experience out of the box, and only platform admins who understand the PicoD contract would create custom CodeInterpreter CRDs.
Would like to hear suggestions :) What do you guys think?
Environment:
- agentcube version:
- Kubernetes version:
- Others:
Please provide an in-depth description of the question you have:
When I go though this getting started guidance: https://github.com/volcano-sh/agentcube/blob/main/docs/getting-started.md, I have question with the CodeInterpreter CRD. The CodeInterpreter CRD is designed to work hand-in-hand with the Python SDK(or future other SDKs), providing an out-of-the-box code execution experience. The SDK exposes some methods, all of which communicate with the PicoD daemon running inside the sandbox.
However, the current CRD design exposes template.image as a user-configurable field:
This quite confuse me:
What do you think about this question?:
Here are some directions I think:
Would like to hear suggestions :) What do you guys think?
Environment: