fix: add MCP tool timeout + running state to prevent silent stall after hub_repo_details (#127)

agent/core/agent_loop.py

@@ -1108,6 +1108,14 @@ async def _exec_tool(
                     ) -> tuple[ToolCall, str, dict, str, bool]:
                         if not valid:
                             return (tc, name, args, err, False)
+                        # Emit "running" immediately so the CLI/frontend shows
+                        # progress instead of going silent (fixes issue #127).
+                        await session.send_event(
+                            Event(
+                                event_type="tool_state_change",
+                                data={"tool_call_id": tc.id, "tool": name, "state": "running"},
+                            )
+                        )
                         out, ok = await session.tool_router.call_tool(
                             name, args, session=session, tool_call_id=tc.id
                         )
@@ -1625,4 +1633,4 @@ async def submission_loop(
                 if local_path:
                     logger.info("Emergency save successful, upload in progress")
             except Exception as e:
-                logger.error(f"Emergency save failed: {e}")
+                logger.error(f"Emergency save failed: {e}")

agent/core/tools.py

@@ -3,11 +3,18 @@
 Provides ToolSpec and ToolRouter for managing both built-in and MCP tools
 """
 
+import asyncio
 import logging
 import warnings
 from dataclasses import dataclass
 from typing import Any, Awaitable, Callable, Optional
 
+# Timeout (seconds) for a single MCP tool call.
+# hub_repo_details fetches model-card + siblings list from huggingface.co/api
+# which can stall indefinitely on slow Hub responses.  30 s is generous but
+# prevents the agent from silently hanging for 5+ minutes.
+MCP_TOOL_TIMEOUT = 30
+
 logger = logging.getLogger(__name__)
 
 from fastmcp import Client
@@ -265,9 +272,21 @@ async def call_tool(
         # Otherwise, use MCP client
         if self._mcp_initialized:
             try:
-                result = await self.mcp_client.call_tool(tool_name, arguments)
+                result = await asyncio.wait_for(
+                    self.mcp_client.call_tool(tool_name, arguments),
+                    timeout=MCP_TOOL_TIMEOUT,
+                )
                 output = convert_mcp_content_to_string(result.content)
                 return output, not result.is_error
+            except asyncio.TimeoutError:
+                timeout_msg = (
+                    f"Tool '{tool_name}' timed out after {MCP_TOOL_TIMEOUT}s "
+                    "with no response from the HF MCP server. "
+                    "The Hub API may be slow or rate-limiting. "
+                    "You can retry, or proceed with the information already available."
+                )
+                logger.warning("MCP tool %s timed out after %ds", tool_name, MCP_TOOL_TIMEOUT)
+                return timeout_msg, False
             except ToolError as e:
                 # Catch MCP tool errors and return them to the agent
                 error_msg = f"Tool error: {str(e)}"
@@ -387,4 +406,4 @@ def create_builtin_tools(local_mode: bool = False) -> list[ToolSpec]:
     tool_names = ", ".join([t.name for t in tools])
     logger.info(f"Loaded {len(tools)} built-in tools: {tool_names}")
 
-    return tools
+    return tools