Verify each finding against the current code and only fix it if needed.

Nitpick comments:
In `@docs/public/llama-stack/llama-stack_quickstart.ipynb`:
- Around line 197-214: Replace the fixed sleep that waits for the MCP server
with an active readiness poll: after starting mcp_process (the Process created
with target _run_mcp_weather_server), repeatedly attempt an HTTP GET to the MCP
endpoint (use MCP_SERVER_URL if set or construct http://{host}:8002/mcp like the
current logic) with a short timeout and small backoff, stop polling when you
receive a successful response (2xx) or after a reasonable overall timeout, and
only then set os.environ["MCP_SERVER_URL"] and print the success message; keep
the same host/port (/mcp) and error/timeout handling so the notebook doesn’t
hang indefinitely.
- Around line 289-298: The cell re-imports LlamaStackClient and Agent and
references the previously-created client, causing redundancy or confusion;
either remove the duplicate imports and rely on the earlier client variable
(keep AGENT_TOOLS and base_url handling) or make the cell self-contained by
checking/creating a client when missing (e.g., test for existing client variable
and call LlamaStackClient(...) to initialize it) so references to
LlamaStackClient, Agent, client, AGENT_TOOLS and base_url are consistent; update
cell text to document the chosen approach (dependent vs standalone cells).
- Around line 205-210: The IP autodetection using
socket.gethostbyname(socket.gethostname()) to set MCP_SERVER_URL is unreliable
in containers/multi-interface hosts; update the logic around MCP_SERVER_URL so
it prefers an explicit MCP_SERVER_HOST env var (MCP_SERVER_HOST) or defaults to
127.0.0.1 when MCP_SERVER_HOST is missing, and add a brief inline comment or
documentation note next to the MCP_SERVER_URL assignment explaining the
container/network limitation and instructing users to set MCP_SERVER_HOST when
running inside containers; reference the MCP_SERVER_URL assignment and the
socket.gethostbyname/socket.gethostname calls when making the change.

Verify each finding against the current code and only fix it if needed.

Inline comments:
In `@docs/public/llama-stack/llama-stack_quickstart.ipynb`:
- Around line 233-240: The code currently overwrites
os.environ["MCP_SERVER_URL"] without preserving the caller's original value;
capture the original value (e.g., orig = os.environ.get("MCP_SERVER_URL"))
before setting MCP_SERVER_URL and set os.environ["MCP_SERVER_URL"] to the
computed value, then in the corresponding cleanup/teardown section restore the
environment to its prior state (if orig is None delete the key, otherwise set it
back to orig); apply the same capture-and-restore pattern for the other
occurrence referenced (the block around the later cleanup).
- Around line 104-106: The notebook can reuse a stale AGENT_TOOLS when
TOOL_OPTION is switched; update the TOOL_OPTION selector cell to reset/clear
AGENT_TOOLS (e.g., set AGENT_TOOLS = None or {}) whenever TOOL_OPTION is
assigned, and add a validation immediately before the agent construction step to
assert AGENT_TOOLS matches the chosen TOOL_OPTION (or raise an explicit error)
so reruns fail fast and the wrong backend is never used.
- Around line 114-116: The shared FastAPI test flow calls requests.post(...) but
the requests import is currently inside the Option A-only cell (which may be
skipped); move the requests import out of the Option A-only cell into a
common/non-optional setup cell (or add an explicit requests import in the API
test cell) so requests is always defined for the shared flow; update references
around get_weather and AGENT_TOOLS cells accordingly to ensure requests is
available regardless of TOOL_OPTION.
- Around line 183-241: The notebook spawns multiprocessing.Process with a nested
target (_run_mcp_weather_server) which is not picklable on spawn-based kernels
and uses a brittle time.sleep(5) readiness wait; move the server entrypoint out
of the cell into a top-level, importable function (e.g., module-level
run_mcp_weather_server) or launch the FastMCP server via subprocess instead of
Process, then replace the fixed sleep with an explicit readiness poll against
the MCP endpoint (MCP_SERVER_URL) with retry/backoff and timeout before setting
os.environ and printing; update usages of mcp_process, get_weather_mcp, and the
Process creation to reference the new top-level entrypoint.

Verify each finding against the current code and only fix it if needed.

Inline comments:
In `@docs/public/llama-stack/llama-stack_quickstart.ipynb`:
- Around line 255-261: Wrap the call to
socket.gethostbyname(socket.gethostname()) in a try/except that catches
socket.gaierror and socket.herror (and optionally a broad Exception) and on
exception fall back to using os.getenv("MCP_SERVER_HOST", "127.0.0.1") for
_host; then proceed to build MCP_SERVER_URL as before (look for the
MCP_SERVER_URL assignment and the _host logic around socket.gethostbyname to
apply this change).
- Around line 296-309: AGENT_TOOLS is using an invalid dict format for MCP;
after calling client.toolgroups.register(...) you must fetch the actual tool
objects or identifiers and pass those to Agent(tools=...), e.g. call
client.tools.list(toolgroup_id=toolgroup_id) and assign AGENT_TOOLS to that
returned list (or build a list of explicit MCP identifiers like
["mcp::filesystem::read_file"]) so Agent(tools=AGENT_TOOLS) receives valid tool
entries instead of the current {"type":..., "server_label":...,
"server_url":...} structure.