feat: post-train environment

responses_api_agents/claude_code_agent/app.py

@@ -266,6 +266,7 @@ class ClaudeCodeAgentConfig(BaseResponsesAPIAgentConfig):
     anthropic_base_url: Optional[str] = None
     max_turns: Optional[int] = 30  # None -> unlimited turns
     timeout: int = 300
+    cwd: Optional[str] = None
     system_prompt: Optional[str] = None
     allowed_tools: Optional[str] = None
     disallowed_tools: Optional[str] = None
@@ -481,6 +482,7 @@ async def _run_claude_code(
                 stdout=asyncio.subprocess.PIPE,
                 stderr=asyncio.subprocess.PIPE,
                 env=env,
+                cwd=self.config.cwd,
             )
             communication = asyncio.create_task(proc.communicate())
             try:

responses_api_agents/claude_code_agent/setup_claude_code.py

@@ -15,6 +15,7 @@
 
 import logging
 import os
+import platform
 import shutil
 import subprocess
 import tarfile
@@ -26,13 +27,22 @@
 
 _CLAUDE_PKG = "@anthropic-ai/claude-code"
 _NODE_VERSION = "22.15.0"
-_NODE_DIST_URL = f"https://nodejs.org/dist/v{_NODE_VERSION}/node-v{_NODE_VERSION}-linux-x64.tar.xz"
-_LOCAL_PREFIX = Path(__file__).parent / ".claude_node"
+_NODE_ARCH = {"x86_64": "x64", "aarch64": "arm64", "arm64": "arm64"}.get(platform.machine())
+if _NODE_ARCH is None:
+    raise RuntimeError(f"Unsupported Node.js architecture: {platform.machine()}")
+_NODE_DIST_URL = f"https://nodejs.org/dist/v{_NODE_VERSION}/node-v{_NODE_VERSION}-linux-{_NODE_ARCH}.tar.xz"
+_LOCAL_PREFIX = Path(__file__).parent / f".claude_node_{_NODE_ARCH}"
 
 
 def _npm_install(npm_bin: str, version: str | None) -> None:
     pkg = f"{_CLAUDE_PKG}@{version}" if version else f"{_CLAUDE_PKG}@latest"
-    subprocess.run([npm_bin, "install", "-g", pkg], check=True)
+    npm_cache = _LOCAL_PREFIX / ".npm-cache"
+    npm_cache.mkdir(parents=True, exist_ok=True)
+    subprocess.run(
+        [npm_bin, "install", "-g", pkg],
+        check=True,
+        env={**os.environ, "NPM_CONFIG_CACHE": str(npm_cache)},
+    )
 
 
 def _install_node_locally() -> Path:

responses_api_agents/claude_code_agent/tests/test_app.py

@@ -396,8 +396,11 @@ async def run_claude_code(*args, observation_collector=None, **kwargs):
 
 
 class TestRunClaudeCode:
-    def test_wires_command_env_and_cleans_up(self, tmp_path: Path) -> None:
-        agent = _make_agent(mcp_config="/path/to/mcp.json")
+    @pytest.mark.parametrize("model_name", ["claude-sonnet-4-6", "nvidia/qwen/qwen3.8-27b", "Qwen/Qwen3.8-27B"])
+    def test_wires_command_env_and_cleans_up(self, tmp_path: Path, model_name: str) -> None:
+        agent = _make_agent(
+            mcp_config="/path/to/mcp.json", model=model_name, anthropic_base_url="http://localhost:8000"
+        )
         captured: dict = {}
 
         class FakeProc:
@@ -412,6 +415,15 @@ async def communicate(self):
 
         async def fake_exec(*cmd, **kwargs):
             env = kwargs["env"]
+            for key in (
+                "ANTHROPIC_MODEL",
+                "ANTHROPIC_DEFAULT_HAIKU_MODEL",
+                "ANTHROPIC_DEFAULT_SONNET_MODEL",
+                "ANTHROPIC_DEFAULT_OPUS_MODEL",
+                "CLAUDE_CODE_SUBAGENT_MODEL",
+            ):
+                assert env[key] == model_name
+            assert cmd[cmd.index("--model") + 1] == model_name
             config_dir = env["CLAUDE_CONFIG_DIR"]
             captured["cmd"] = list(cmd)
             captured["config_dir"] = config_dir
@@ -434,7 +446,7 @@ async def fake_exec(*cmd, **kwargs):
         # config dir is removed after the run (no leakage between rollouts)
         assert not Path(captured["config_dir"]).exists()
         assert output_items == []
-        assert model == "claude-sonnet-4-6"
+        assert model == model_name
         assert metadata["status"] == "completed"
 
     def test_skills_staged_and_bare_dropped(self, tmp_path: Path) -> None:

responses_api_agents/nemorl_env/README.md

@@ -0,0 +1,49 @@
+# NeMo RL environment
+
+A Gym environment for training models to train models.
+
+1. The model runs in Claude Code to edit a recipe and pinned NeMo RL/Gym sources in a CPU sandbox
+   (60 minutes, 200 turns, 256K context). Core changes such as loss fn, advantage estimation or
+   data-processing changes are accepted.
+2. A GPU sandbox applies the patch and trains Qwen2.5-1.5B-Instruct with
+   NeMo RL + NeMo Gym GRPO: 8 prompts × 8 responses, with Gym's math verifier.
+3. A fresh, unpatched GPU sandbox serves model with vLLM and runs
+   `gym eval run` on AIME25 with a 32K context budget, including the prompt.
+
+Reward is AIME25 avg@8: mean accuracy over eight sampled answers per problem
+(30 problems, 240 answers; temperature 0.7).
+
+## Setup
+
+Configure `policy_model` and the sandbox connections in `config.yaml`. CPU/GPU
+domains and API keys, author image, and registry credentials are declared there
+using environment-variable references. The author image must provide Python at
+`/agent_deps_mount/bin/python`, uv, and Git.
+
+Generate datasets:
+
+```bash
+python responses_api_agents/nemorl_env/prepare.py
+gym eval prepare --benchmark aime25
+```
+
+`app.py` orchestrates the sandboxes, `author_worker.py` runs Claude Code.
+`task_environment/` contains the recipe, training data, runner, and evaluator.
+`recipe.yaml` explicitly inherits NeMo RL's full `grpo_math_1B.yaml`, it is an
+editable override, not a standalone config. The launcher writes the complete
+`/testbed/resolved_grpo.yaml` before training. Data/environment edits are retained,
+model, batch size, GPU/time budget, and held-out evaluation remain fixed.
+`data/train.jsonl` contains one stable author prompt; `data/example.jsonl` contains
+five independently identified copies. Use repeated rollouts for exploration.
+`eval_concurrency` in `config.yaml` controls simultaneous AIME25 answers (default 30).
+
+With the author model configured, record real example rollouts with:
+
+```bash
+gym env start --config responses_api_agents/nemorl_env/config.yaml --model-type vllm_model
+gym eval run --no-serve --agent nemorl_env --input responses_api_agents/nemorl_env/data/example.jsonl --output responses_api_agents/nemorl_env/data/example_rollouts.jsonl
+```
+
+Math data: the first 512 training rows of NVIDIA's
+[OpenMathInstruct-2](https://huggingface.co/datasets/nvidia/OpenMathInstruct-2)
+(CC-BY-4.0)