feat: add AutoRL-Bench for RL post-training evaluation

[couragec]

• Add AutoRL-Bench: an evaluation framework for RL post-training agents
• Merge latest main into rl-posttraining, resolve 46 conflicts (all accepting main's version)
• Includes benchmarks (GSM8K, HumanEval, ALFWorld), agents (OpenHands, OpenCode), and Streamlit UI

---

📚 Documentation preview 📚: https://RDAgent--1336.org.readthedocs.build/en/1336/

In general, the fix is to avoid returning the raw exception message from the server to the client. Instead, log the detailed error (possibly including the stack trace) on the server side and respond with a generic error message that does not reveal internal details.

The best targeted fix here is to change the except block in submit() so that:

• The server logs the detailed exception (and, if desired, stack trace) using the existing logger.
• The HTTP response uses a generic, localized error message, without including str(e).

Because we must not change imports except to add well-known libraries and we already have a logger, we can simply modify the returned JSON. A minimal change is to replace {"error": str(e)} with a fixed message such as {"error": "Internal server error during submission"} (or similar). If more diagnostics are needed for clients, we can also include a non-sensitive error code like {"error": "internal_error", "message": "Submission failed due to internal error"}. This keeps functionality the same from a control-flow perspective while removing sensitive information leakage. All of this happens in rdagent/scenarios/rl/autorl_bench/core/server.py within the submit function’s except block.

General fix: further validate the user-provided path string before converting it into a Path, and ensure that paths containing traversal sequences (..), absolute path indicators, or other clearly unsafe content are rejected early. This keeps the logic of constraining everything under safe_root while making the taint flow obviously safe to CodeQL and to human reviewers.

Best concrete fix here:

• Extend _safe_resolve to:
  • Reject path strings containing NUL bytes (which can cause surprising behavior in some filesystem APIs).
  • Split the normalized path into components and reject any component equal to "..".
  • Avoid constructing Path from unfiltered user input; instead, build the path by iterating over allowed components using Path() / .joinpath(...).
• Keep existing checks that reject absolute paths and ensure the final resolved path is under safe_root.

Specific changes in rdagent/app/rl/ui/app.py:

• Modify _safe_resolve (lines 23–36) to:
  • Normalize the input string.
  • Inspect the components of the normalized path (using Path(normalized).parts).
  • Reject ".." components before forming the path relative to safe_root.
  • Build path_obj from the filtered parts (rather than directly from the user string).
  • Keep the absolute path check and relative_to(safe_root) guard.

No new imports are strictly necessary; we can rely on the existing os and Path imports.

---