Fix/xinshi/agent runs

rock/sdk/sandbox/agent/iflow_cli.py

@@ -161,6 +161,17 @@ async def _install_npm(self):
             error_msg="npm installation failed",
         )
 
+        await self._sandbox.arun(
+            cmd=(
+                "mv runtime-env /opt/nodejs && "
+                "ln -sf /opt/nodejs/bin/npm /usr/local/bin/npm && "
+                "ln -sf /opt/nodejs/bin/npx /usr/local/bin/npx && "
+                "ln -sf /opt/nodejs/bin/corepack /usr/local/bin/corepack && "
+                "ln -sf /opt/nodejs/bin/node /usr/local/bin/node"
+            ),
+            session=self.agent_session,
+        )
+
         elapsed_step = time.time() - step_start
         self._log_step("NPM installation finished", step_name="NPM Install", is_complete=True, elapsed=elapsed_step)
 
@@ -202,6 +213,11 @@ async def _install_iflow_cli_package(self):
             error_msg="iflow-cli installation failed",
         )
 
+        await self._sandbox.arun(
+            cmd="ln -s /opt/nodejs/bin/iflow /usr/local/bin/iflow",
+            session=self.agent_session,
+        )
+
         elapsed_step = time.time() - step_start
         self._log_step(
             "IFlow CLI installation finished", step_name="IFlow Install", is_complete=True, elapsed=elapsed_step

rock/sdk/sandbox/agent/openhands.py

@@ -163,13 +163,13 @@ class OpenhandsConfig(DefaultAgentConfig):
 
     # Command to clone Openhands/benchmarks repository and install dependencies
     openhands_sdk_install_cmd_list: list[str] = [
-        f"/openhands/python/bin/pip config set global.index-url {env_vars.ROCK_PIP_INDEX_URL}",
-        "/openhands/python/bin/pip install openhands-agent-server==1.6.0 openhands-sdk==1.6.0",
-        "/openhands/python/bin/pip install openhands-tools==1.6.0 openhands-workspace==1.6.0",
+        f"/openhands/runtime-env/bin/pip config set global.index-url {env_vars.ROCK_PIP_INDEX_URL}",
+        "/openhands/runtime-env/bin/pip install openhands-agent-server==1.6.0 openhands-sdk==1.6.0",
+        "/openhands/runtime-env/bin/pip install openhands-tools==1.6.0 openhands-workspace==1.6.0",
         "rm -rf /openhands/benchmarks",
         "git clone -b features/local_workspace_fix_early_stop https://github.com/shayue-wt/benchmarks.git /openhands/benchmarks",
-        "/openhands/python/bin/pip install datasets huggingface-hub jinja2 pandas Pillow toml swebench",
-        "/openhands/python/bin/pip install tqdm 'unidiff>=0.7.5,<0.8.0' 'modal>=1.1.4' commit0 pytest-json-report",
+        "/openhands/runtime-env/bin/pip install datasets huggingface-hub jinja2 pandas Pillow toml swebench",
+        "/openhands/runtime-env/bin/pip install tqdm 'unidiff>=0.7.5,<0.8.0' 'modal>=1.1.4' commit0 pytest-json-report",
     ]
 
     python_install_timeout: int = 300
@@ -417,7 +417,7 @@ async def run(
                 agent_run_cmd = (
                     f"cd {self.config.agent_workdir}/benchmarks && "
                     "export PYTHONPATH='.' && "
-                    f"{self.config.agent_workdir}/python/bin/python "
+                    f"{self.config.agent_workdir}/runtime-env/bin/python "
                     "./benchmarks/swebench/run_infer.py "
                     ".llm_config.json --dataset eval --split test --note rock_rollout "
                     f"--select ./{instance_config} --max-iterations {self.config.max_iteration}"

rock/sdk/sandbox/agent/swe_agent.py

@@ -152,7 +152,7 @@ class SweAgentConfig(DefaultAgentConfig):
 
     swe_agent_workdir: str = "/tmp_sweagent"
 
-    python_install_cmd: str = env_vars.ROCK_RTENV_PYTHON_V31114_INSTALL_CMD
+    python_install_cmd: str = env_vars.ROCK_RTENV_PYTHON_V31212_INSTALL_CMD
 
     swe_agent_install_cmd: str = (
         "[ -d SWE-agent ] && rm -rf SWE-agent; "
@@ -273,7 +273,7 @@ async def _install_swe_agent_package(self):
         self._log_step("Installing SWE-agent repository", step_name="SWE-agent Install")
 
         swe_agent_install_cmd = (
-            f"export PATH={self.config.swe_agent_workdir}/python/bin:$PATH && "
+            f"export PATH={self.config.swe_agent_workdir}/runtime-env/bin:$PATH && "
             f"cd {self.config.swe_agent_workdir} && "
             f"{self.config.swe_agent_install_cmd}"
         )
@@ -428,7 +428,7 @@ async def run(
 
                 swe_agent_run_cmd = (
                     f"cd {self.config.swe_agent_workdir} && "
-                    f"{self.config.swe_agent_workdir}/python/bin/sweagent run --config {config_filename}"
+                    f"{self.config.swe_agent_workdir}/runtime-env/bin/sweagent run --config {config_filename}"
                 )
                 full_cmd = f"bash -c {shlex.quote(swe_agent_run_cmd)}"
                 logger.debug(

rock/sdk/sandbox/model_service/base.py

@@ -162,7 +162,7 @@ async def install(self) -> None:
             # Step 4: Install model service
             step_start_time = time.time()
             model_service_install_cmd = (
-                f"export PATH={self.config.workdir}/python/bin:$PATH && "
+                f"export PATH={self.config.workdir}/runtime-env/bin:$PATH && "
                 f"cd {self.config.workdir} && {self.config.model_service_install_cmd}"
             )
             bash_service_cmd = f"bash -c {shlex.quote(model_service_install_cmd)}"
@@ -219,8 +219,8 @@ async def start(self) -> None:
             start_cmd = (
                 f"export ROCK_LOGGING_PATH={self.config.logging_path} && "
                 f"export ROCK_LOGGING_FILE_NAME={self.config.logging_file_name} && "
-                f"{self.config.workdir}/python/bin/{self.config.stop_cmd} && "
-                f"{self.config.workdir}/python/bin/{self.config.start_cmd.format(model_service_type=self.config.model_service_type)}"
+                f"{self.config.workdir}/runtime-env/bin/{self.config.stop_cmd} && "
+                f"{self.config.workdir}/runtime-env/bin/{self.config.start_cmd.format(model_service_type=self.config.model_service_type)}"
             )
             bash_start_cmd = f"bash -c {shlex.quote(start_cmd)}"
             logger.debug(f"[{sandbox_id}] Model service Start command: {bash_start_cmd}")
@@ -262,7 +262,7 @@ async def stop(self) -> None:
         try:
             logger.info(f"[{sandbox_id}] Stopping model service")
 
-            stop_cmd = f"{self.config.workdir}/python/bin/{self.config.stop_cmd}"
+            stop_cmd = f"{self.config.workdir}/runtime-env/bin/{self.config.stop_cmd}"
             bash_stop_cmd = f"bash -c {shlex.quote(stop_cmd)}"
 
             await self._sandbox.arun(
@@ -301,7 +301,7 @@ async def watch_agent(self, pid: str) -> None:
             raise RuntimeError(error_msg)
 
         try:
-            watch_agent_cmd = f"{self.config.workdir}/python/bin/{self.config.watch_agent_cmd.format(pid=pid)}"
+            watch_agent_cmd = f"{self.config.workdir}/runtime-env/bin/{self.config.watch_agent_cmd.format(pid=pid)}"
             bash_watch_cmd = f"bash -c {shlex.quote(watch_agent_cmd)}"
             logger.debug(f"[{sandbox_id}] Model service watch agent with pid={pid}, cmd: {bash_watch_cmd}")
 
@@ -369,7 +369,7 @@ async def anti_call_llm(
             else:
                 cmd = self.config.anti_call_llm_cmd_no_response.format(index=index)
 
-            full_cmd = f"{self.config.workdir}/python/bin/{cmd}"
+            full_cmd = f"{self.config.workdir}/runtime-env/bin/{cmd}"
             bash_cmd = f"bash -c {shlex.quote(full_cmd)}"
             logger.debug(f"[{sandbox_id}] Executing command: {bash_cmd}")
 

tests/integration/sdk/sandbox/agent/swe_agent/test_init.py

@@ -29,7 +29,6 @@ async def _verify_exists(sandbox: Sandbox, directory_path: str, items: set[str])
     logger.info(f"Directory {directory_path} contents: {result.stdout}")
 
 
-@pytest.mark.skip
 @pytest.mark.need_admin
 @SKIP_IF_NO_DOCKER
 @pytest.mark.asyncio
@@ -52,14 +51,13 @@ async def test_swe_agent_initialization(sandbox_instance: Sandbox):
     await _verify_exists(sandbox_instance, "/", {agent_dir_name})
 
     # 4. Verify agent installation directories
-    await _verify_exists(sandbox_instance, swe_agent_config.swe_agent_workdir, {"python", "SWE-agent"})
+    await _verify_exists(sandbox_instance, swe_agent_config.swe_agent_workdir, {"runtime-env", "SWE-agent"})
 
     # 5. Verify Python executables
-    python_bin_path = f"{swe_agent_config.swe_agent_workdir}/python/bin"
+    python_bin_path = f"{swe_agent_config.swe_agent_workdir}/runtime-env/bin"
     await _verify_exists(sandbox_instance, python_bin_path, {"sweagent"})
 
 
-@pytest.mark.skip
 @pytest.mark.need_admin
 @SKIP_IF_NO_DOCKER
 @pytest.mark.asyncio
@@ -87,8 +85,8 @@ async def test_swe_agent_with_model_service(sandbox_instance: Sandbox):
     await _verify_exists(sandbox_instance, "/", {agent_dir_name, model_service_dir_name})
 
     # 4. Verify agent installation directories
-    await _verify_exists(sandbox_instance, swe_agent_config.swe_agent_workdir, {"python", "SWE-agent"})
+    await _verify_exists(sandbox_instance, swe_agent_config.swe_agent_workdir, {"runtime-env", "SWE-agent"})
 
     # 5. Verify Python executables
-    python_bin_path = f"{swe_agent_config.swe_agent_workdir}/python/bin"
+    python_bin_path = f"{swe_agent_config.swe_agent_workdir}/runtime-env/bin"
     await _verify_exists(sandbox_instance, python_bin_path, {"sweagent"})

tests/integration/sdk/sandbox/agent/swe_agent/test_run.py

@@ -106,7 +106,6 @@ async def _init_git_repository(sandbox: Sandbox, repo_path: str) -> None:
             raise RuntimeError(f"Failed to execute: {cmd}")
 
 
-@pytest.mark.skip
 @pytest.mark.need_admin
 @SKIP_IF_NO_DOCKER
 @pytest.mark.asyncio

tests/integration/sdk/sandbox/model_service/test_model_service.py

@@ -10,7 +10,6 @@
 logger = init_logger(__name__)
 
 
-@pytest.mark.skip
 @pytest.mark.need_admin
 @SKIP_IF_NO_DOCKER
 @pytest.mark.asyncio
@@ -36,10 +35,10 @@ async def test_model_service_install_and_start(sandbox_instance: Sandbox):
     logger.info(f"Work directory contents: {result.stdout}")
     assert result.exit_code == 0
     assert "cpython31114.tar.gz" in result.stdout, "Tar archive file missing"
-    assert "python" in result.stdout, "Python directory missing"
+    assert "runtime-env" in result.stdout, "Python directory missing"
 
     # 5. Verify Python executables
-    python_bin_path = f"{model_service_config.workdir}/python/bin"
+    python_bin_path = f"{model_service_config.workdir}/runtime-env/bin"
     result = await sandbox_instance.execute(Command(command="ls", cwd=python_bin_path))
     logger.info(f"Python bin directory contents: {result.stdout}")
     assert result.exit_code == 0