0.65.8 — Faster MCP tool init

Building Langroid tools from an MCP server now costs one tools/list
round-trip instead of 1 + N. If you connect to MCP servers over the network
(SSE / HTTP), especially ones exposing many tools, tool initialization is now
materially faster. No API or behavior change for well-formed tools, and no
configuration needed.

pip install -U langroid

Performance — one list_tools() instead of 1 + N (#1054)

FastMCPClient.get_tools_async() previously issued 1 + N tools/list
round-trips to build N tools: one initial list_tools(), then — inside
get_tool_async → get_mcp_tool_async — one more full re-list per tool, just
to pick a single entry by name. Because fastmcp's Client.list_tools() is
uncached, every one of those is a real network call, so tool init scaled as
O(N) round-trips and dominated startup latency on slow or remote servers.

The network fetch is now split from the pure, synchronous, network-free
schema-to-ToolMessage conversion:

• get_tools_async() calls list_tools() once and maps the conversion
  over the result — 1 round-trip total.
• get_tool_async(name) still fetches its single Tool, then converts it.

Generated tool models are identical to before — verified by a new
behavior-equivalence test asserting get_tools_async() yields the same models
(names, fields, defaults, renamed fields, and order) as building each tool via
get_tool_async().

New — public tool_model_from_mcp_tool()

FastMCPClient.tool_model_from_mcp_tool(target) is now public: a pure,
synchronous, network-free conversion from an already-fetched mcp.types.Tool
into a Langroid ToolMessage subclass. Allow-list / subset consumers can fetch
one list_tools() snapshot, filter it, and build models for just the tools
they want — without re-listing the server once per tool:

async with FastMCPClient(server) as client:
    tools = await client.client.list_tools()
    wanted = [t for t in tools if t.name in {"search", "fetch"}]
    models = [client.tool_model_from_mcp_tool(t) for t in wanted]

Robustness

The now-public builder tolerates malformed or missing MCP tool metadata that
the previous inline code dereferenced unguarded — a null / non-dict
inputSchema, missing name / inputSchema / description, null nested
properties / items, and non-string required entries. An invalid tool name
raises a clear ValueError; a malformed schema now degrades to a permissive
model instead of raising AttributeError / TypeError and failing the whole
get_tools_async() batch on one bad tool.

Behavior change: a genuinely malformed tool that previously crashed tool
loading is now tolerated (built as a permissive model). Well-formed tools are
unaffected.

Credit

The optimization was originally proposed by
@alexagr in
#1053; this release ships it
with added hardening, tests, and docs.

Upgrade

pip install -U langroid

Full changelog:
0.65.7...0.65.8