fix(agent): widen MCP client-session timeout for cold first runs

src/agent/deep_agent/runner.py

@@ -38,6 +38,11 @@
 
 _DEFAULT_MODEL = "litellm_proxy/aws/claude-opus-4-6"
 
+# MCP client-session timeout for stdio servers.  Cold-cache first spawns of
+# heavy servers (vibration imports scipy + numpy + DSP) can exceed the mcp
+# library's default handshake budget.  Warm runs are sub-second.
+_MCP_CLIENT_TIMEOUT_SECONDS = 30
+
 
 def _build_chat_model(model_id: str):
     """Construct a LangChain chat model for *model_id*.
@@ -85,6 +90,11 @@ def _build_mcp_connections(
             "command": "uv",
             "args": ["run", cmd_arg],
             "cwd": str(_REPO_ROOT),
+            "session_kwargs": {
+                "read_timeout_seconds": _dt.timedelta(
+                    seconds=_MCP_CLIENT_TIMEOUT_SECONDS
+                ),
+            },
         }
     return connections
 

src/agent/openai_agent/runner.py

@@ -39,6 +39,12 @@
 
 _DEFAULT_MODEL = "litellm_proxy/azure/gpt-5.4"
 
+# MCP client-session timeout for stdio servers.  The SDK default of 5s is too
+# tight for first-spawn cold-cache cases (heavy imports like scipy/numpy push
+# vibration cold start past 10s on a freshly synced env).  Warm handshakes are
+# all sub-second so a generous value has no operational downside.
+_MCP_CLIENT_TIMEOUT_SECONDS = 30
+
 
 def _build_run_config(model_id: str) -> RunConfig | None:
     """Build a RunConfig with a LiteLLM model provider when needed.
@@ -95,6 +101,7 @@ def _build_mcp_servers(
                     "args": ["run", cmd_arg],
                 },
                 cache_tools_list=True,
+                client_session_timeout_seconds=_MCP_CLIENT_TIMEOUT_SECONDS,
             )
         )
     return servers