[agents, examples] feat: support Claude Code on OpenYuanRong

[Mengyuyang]

Summary

Add an opt-in OpenYuanRong task configuration for running the existing Claude Code agent in SWE-Bench and SWE-ReBench sandboxes. The configuration mounts a prebuilt Claude Code sidecar and uses its explicit CLI path, so OpenYuanRong rollouts do not install or upgrade Claude Code inside each sandbox.

The standalone verl inference runner also propagates resolved editable package roots and the existing OpenYuanRong environment to Ray workers. This prevents driver/worker vllm import mismatches before expensive rollout actors are created.

Changes

Add examples/quickstart/training/task_config_claude_code_openyuanrong.yaml.

Configure swe_bench and swe_rebench with the openyuanrong sandbox provider, OpenYuanRong registry image mapping, sidecar mount, tunnel proxy port, rollout timeout, and Claude Code prompt/model defaults.

Add optional agent.executable to ClaudeCodeConfig, defaulting to claude for backward compatibility.

When agent.executable is an explicit path such as /opt/claude-code/bin/claude, validate it with test -x before rollout execution.

Fail explicitly when the configured sidecar executable is unavailable. The explicit-path mode never falls back to npm or native Claude Code installation.

Keep the existing PATH-based claude discovery and installation fallback unchanged when no explicit executable path is configured.

Add Ray runtime-environment helpers that resolve physical package roots for vllm, verl, and uni_agent, then forward them to Ray workers through the first ray.init() call.

Forward already configured OpenYuanRong environment variables to nested rollout workers and validate the worker-side engine import before creating the rollout worker group.

Configuration and API

- name: swe_bench
  sandbox:
    provider: openyuanrong
    image_map:
      - from: "swebench/**"
        to: "swr.cn-east-3.myhuaweicloud.com/openyuanrong/swe-bench-verified/**:v2"
    sandbox_kwargs:
      proxy_port: 38197
      mounts:
        - target: /opt/claude-code
          image_url: swr.cn-east-3.myhuaweicloud.com/openyuanrong/claude-code-tool:latest
  agent:
    name: claude_code
    executable: /opt/claude-code/bin/claude

[codecov-commenter]

⚠️ Please install the to ensure uploads and comments are reliably processed by Codecov.

Codecov Report

✅ All modified and coverable lines are covered by tests.

📢 Thoughts on this report? Let us know!

[yyDing1]

Do not modify this file; inject these envs into the ray env.

[yyDing1]

The real execution path should be injected into the env during sandbox start. This file is not expected to modify, too.

[yyDing1]

what's the purpose of this file?