[Feature Request] Agent trajectory / trace collection for Agentic RL training

[khazic]

Hi maintainers,

I've been exploring OpenSandbox as an execution environment for Agentic RL training,
and I think it's a great fit — isolated sandboxes with multi-language support are
exactly what's needed for safe agent execution.

One capability I'd find very valuable is native agent trajectory collection:
the ability to automatically record agent interaction traces (inputs, tool calls,
stdout/stderr, code execution results, file changes, errors, timestamps) within
a sandbox session, and export them in a structured format (e.g. JSON / JSONL)
suitable for RL training pipelines.

Currently this requires wrapping every SDK call manually, which is error-prone
and not portable across projects.

My questions:

1. Is there an existing plan or roadmap item for built-in trace/trajectory
   collection in OpenSandbox?
2. If not, would the maintainers be open to accepting a contribution that adds
   this feature? My intended use case is Agentic RL training — collecting
   (observation, action, reward signal) trajectories from agent runs inside
   sandboxes.

I'm willing to implement this and submit a PR. Before doing so, I'd love to
align on the design (where it lives, what format, opt-in vs always-on, etc.)
to make sure it fits the project's direction.

Thanks for building this — looking forward to your thoughts!

[hittyt]

Hi @khazic, thanks for the thoughtful proposal — this is a really interesting direction.

At the moment, we don’t have a concrete plan to build native trace/trajectory collection directly into OpenSandbox. That said, improving the observability of the sandbox is definitely on our roadmap, and process-level behavior collection is something we’re actively thinking about.

If this can be designed in a way that naturally supports Agentic RL use cases (e.g., trajectory collection for training), that would be a great addition to the ecosystem.

We absolutely welcome external contributions. It would be great if you could start with a draft (or the project's OSEP) outlining your design — especially around scope, data format, and whether it should be opt-in. That will help us align early and ensure it fits well with the project’s direction.

Looking forward to your ideas!