code sandboxes (#7)

* code sandboxes

* lint

* fix: test

* fix: test

* lint

* lint

* fix: tests

Author: echarles

agent_codemode/__version__.py

@@ -4,4 +4,4 @@
 
 """Agent Codemode."""
 
-__version__ = "0.1.6"
+__version__ = "1.0.0"

agent_codemode/composition/executor.py

@@ -323,9 +323,9 @@ async def handle_execute_code(arguments: dict[str, Any]) -> dict[str, Any]:
         await executor.setup()
 
     # Inject context variables if provided
-    if context and executor._sandbox:
+    if context and executor.sandbox_client:
         for name, value in context.items():
-            executor._sandbox.set_variable(name, value)
+            executor.sandbox_client.set_variable(name, value)
 
     try:
         execution = await executor.execute(code, timeout=timeout)

agent_codemode/server.py

@@ -99,7 +99,7 @@ class CodemodeToolset(AbstractToolset):
 
         registry: ToolRegistry | None = None
         config: CodeModeConfig = field(default_factory=CodeModeConfig)
-        sandbox: Any | None = None  # Optional pre-configured sandbox (e.g., EvalSandbox)
+        sandbox_client: Any | None = None
         allow_direct_tool_calls: bool | None = None
         allow_discovery_tools: bool = True
         tool_reranker: Callable[[list, str, Optional[str]], Awaitable[list]] | None = None
@@ -190,7 +190,7 @@ async def _ensure_initialized(self) -> None:
                 self._executor = CodeModeExecutor(
                     registry=registry,
                     config=self.config,
-                    sandbox=self.sandbox,
+                    sandbox_client=self.sandbox_client,
                 )
                 await self._executor.setup()
                 logger.info(

agent_codemode/toolset.py

@@ -372,10 +372,10 @@ For Pydantic AI agents, use the `AgentSkillsToolset`:
 ```python
 from pydantic_ai import Agent
 from agent_skills import AgentSkillsToolset, SandboxExecutor
-from code_sandboxes.eval_sandbox import EvalSandbox
+from code_sandboxes import CodeSandboxClient
 
 # Create toolset with sandbox execution
-sandbox = EvalSandbox()
+sandbox = CodeSandboxClient.create(variant="eval")
 toolset = AgentSkillsToolset(
     directories=["./skills"],
     executor=SandboxExecutor(sandbox),

docs/docs/skills/index.mdx

@@ -34,7 +34,7 @@
 
 try:
     from agent_skills import AgentSkillsToolset, SandboxExecutor
-    from code_sandboxes.eval_sandbox import EvalSandbox
+    from code_sandboxes import CodeSandboxClient
 
     HAS_AGENT_SKILLS = True
 except ImportError:
@@ -351,16 +351,16 @@ def create_agent(model: str, codemode: bool) -> tuple[Agent, object | None, obje
         )
 
         # Create shared sandbox for both CodemodeToolset and AgentSkillsToolset
-        shared_sandbox = None
+        shared_client = None
         skills_toolset = None
         if HAS_AGENT_SKILLS:
-            shared_sandbox = EvalSandbox()
-            logger.info("Created shared EvalSandbox for codemode and skills toolsets")
+            shared_client = CodeSandboxClient.create(variant="eval")
+            logger.info("Created shared CodeSandboxClient for codemode and skills")
 
         toolset = CodemodeToolset(
             registry=registry,
             config=config,
-            sandbox=shared_sandbox,
+            sandbox_client=shared_client,
             allow_discovery_tools=True,  # Enable discovery tools (search_tools, get_tool_details, list_tool_names, list_servers)
         )
         toolsets = [toolset]
@@ -369,7 +369,7 @@ def create_agent(model: str, codemode: bool) -> tuple[Agent, object | None, obje
         if HAS_AGENT_SKILLS:
             skills_toolset = AgentSkillsToolset(
                 directories=[str((repo_root / "skills").resolve())],
-                executor=SandboxExecutor(shared_sandbox),
+                executor=SandboxExecutor(shared_client),
             )
             toolsets.append(skills_toolset)
             logger.info(

examples/simple/agent_cli.py

@@ -21,21 +21,21 @@ classifiers = [
   "Programming Language :: Python :: 3",
 ]
 dependencies = [
-    "agent_skills",
-    "code_sandboxes",
-    "mcp[cli]>=1.0",
+    "agent-skills",
+    "code-sandboxes",
+    "mcp[cli]>=1.10.1,<2",
     "pydantic>=2.0",
     "httpx>=0.24",
 ]
 
 [project.optional-dependencies]
 pydantic-ai = [
-  "pydantic-graph>=1.94.0",
-  "pydantic-ai-slim>=1.94.0",
+  "pydantic-ai-slim>=2.21.0,<3",
+  "pydantic-graph>=2.21.0,<3",
 ]
 test = [
   "ipykernel",
-  "jupyter_server>=1.6,<3",
+  "jupyter-server>=2.10,<3",
   "pytest>=7.0",
   "pytest-asyncio>=0.21",
   "pytest-cov>=4.1",

pyproject.toml

@@ -9,6 +9,28 @@
 import pytest
 
 
+def _rebuild_fastmcp_settings() -> None:
+    """Resolve the forward reference in FastMCP's ``Settings`` model.
+
+    ``mcp.server.fastmcp.server.Settings.lifespan`` is annotated with
+    ``FastMCP``, which is defined further down the same module, and upstream
+    never calls ``model_rebuild()``. Recent pydantic-settings releases warn
+    (``IncompleteFieldDefinitionWarning``) when such a model is instantiated,
+    and this suite turns warnings into errors, so collection fails as soon as
+    anything constructs a FastMCP server. Rebuilding the model once resolves
+    the reference for real instead of muting the warning.
+    """
+    try:
+        from mcp.server.fastmcp.server import Settings
+    except ImportError:  # pragma: no cover - mcp layout changed
+        return
+    if not getattr(Settings, "__pydantic_complete__", True):
+        Settings.model_rebuild()
+
+
+_rebuild_fastmcp_settings()
+
+
 @pytest.fixture
 def skills_dir(tmp_path: Path) -> Path:
     """Create a temporary skills directory."""

tests/conftest.py

@@ -3,13 +3,13 @@
 
 import asyncio
 
-from code_sandboxes import Sandbox
+from code_sandboxes import CodeSandboxClient
 
 
 async def main():
     # Create a sandbox
-    sandbox = Sandbox.create(variant="eval")
-    sandbox.start()
+    client = CodeSandboxClient.create(variant="eval")
+    client.start()
 
     # Set up an executor mock
     class MockExecutor:
@@ -18,11 +18,11 @@ async def call_tool(self, name, args):
             await asyncio.sleep(0.01)
             return {"status": "success", "data": f"Result for {name}"}
 
-    sandbox.set_variable("__executor__", MockExecutor())
+    client.set_variable("__executor__", MockExecutor())
 
     # First: Define __call_tool__
     print("\n=== Step 1: Define __call_tool__ ===")
-    result1 = sandbox.run_code("""
+    result1 = client.execute_code("""
 async def __call_tool__(tool_name, arguments):
     '''Call a tool through the executor.'''
     return await __executor__.call_tool(tool_name, arguments)
@@ -34,7 +34,7 @@ async def __call_tool__(tool_name, arguments):
 
     # Second: Use __call_tool__
     print("\n=== Step 2: Use __call_tool__ with await ===")
-    result2 = sandbox.run_code("""
+    result2 = client.execute_code("""
 result = await __call_tool__("test_tool", {"arg": "value"})
 print(f"Tool result: {result}")
 result
@@ -43,7 +43,7 @@ async def __call_tool__(tool_name, arguments):
     print(f"Result 2 - Stdout: {result2.stdout}")
     print(f"Result 2 - Results: {result2.results}")
 
-    sandbox.stop()
+    client.stop()
     print("\n=== Test completed successfully! ===")
 
 

tests/test_executor_async.py

@@ -10,65 +10,74 @@
 from code_sandboxes import ExecutionResult, Logs, OutputMessage
 from code_sandboxes.models import Result
 
+from agent_codemode.composition import executor as executor_module
 from agent_codemode.composition.executor import CodeModeExecutor
 from agent_codemode.discovery.registry import ToolRegistry
 
 
-class _StreamingSandbox:
+class _StreamingClient:
+    variant = "jupyter"
+
     def __init__(self) -> None:
         self.run_code_calls = 0
         self.streaming_called = False
 
-    def run_code(self, code: str, **kwargs) -> ExecutionResult:
+    def execute_code(self, code: str, **kwargs) -> ExecutionResult:
         _ = (code, kwargs.get("timeout"), kwargs.get("language"), kwargs.get("envs"))
         self.run_code_calls += 1
         return ExecutionResult(logs=Logs())
 
-    def run_code_streaming(self, code: str, **kwargs):
+    def execute_code_streaming(self, code: str, **kwargs):
         _ = (code, kwargs.get("timeout"), kwargs.get("language"), kwargs.get("envs"))
         self.streaming_called = True
         yield OutputMessage(line="status: RUNNING", timestamp=0.0, error=False)
         yield OutputMessage(line="hello", timestamp=0.0, error=False)
         yield Result(data={"text/plain": "42"}, is_main_result=True, extra={})
 
 
-class _NonStreamingSandbox:
+class _FailingStreamingClient:
+    variant = "jupyter"
+
     def __init__(self) -> None:
         self.run_code_calls = 0
 
-    def run_code(self, code: str, **kwargs) -> ExecutionResult:
+    def execute_code(self, code: str, **kwargs) -> ExecutionResult:
         _ = (code, kwargs.get("timeout"), kwargs.get("language"), kwargs.get("envs"))
         self.run_code_calls += 1
-        if self.run_code_calls >= 3:
-            return ExecutionResult(
-                logs=Logs(stdout=[OutputMessage(line="fallback", timestamp=0.0, error=False)]),
-            )
         return ExecutionResult(logs=Logs())
 
+    def execute_code_streaming(self, code: str, **kwargs):
+        _ = (code, kwargs)
+        raise RuntimeError("sandbox unavailable")
+        yield
+
 
 @pytest.mark.asyncio
-async def test_execute_uses_streaming_when_supported():
+async def test_execute_uses_streaming_when_supported(monkeypatch):
+    monkeypatch.setattr(executor_module, "_get_identity_env", lambda: {})
     executor = CodeModeExecutor(registry=ToolRegistry())
-    sandbox = _StreamingSandbox()
-    executor._sandbox = sandbox
+    client = _StreamingClient()
+    executor._sandbox_client = client
     executor._setup_done = True
 
     result = await executor.execute("print('hi')")
 
-    assert sandbox.streaming_called is True
+    assert client.streaming_called is True
     assert "status: RUNNING" in result.logs.stdout_text
     assert "hello" in result.logs.stdout_text
     assert result.results and result.results[0].data["text/plain"] == "42"
 
 
 @pytest.mark.asyncio
-async def test_execute_falls_back_to_run_code_without_streaming():
+async def test_execute_reports_streaming_infrastructure_failure(monkeypatch):
+    monkeypatch.setattr(executor_module, "_get_identity_env", lambda: {})
     executor = CodeModeExecutor(registry=ToolRegistry())
-    sandbox = _NonStreamingSandbox()
-    executor._sandbox = sandbox
+    client = _FailingStreamingClient()
+    executor._sandbox_client = client
     executor._setup_done = True
 
     result = await executor.execute("print('hi')")
 
-    assert sandbox.run_code_calls >= 3
-    assert result.logs.stdout_text == "fallback"
+    assert client.run_code_calls >= 2
+    assert result.execution_ok is False
+    assert result.execution_error == "sandbox unavailable"