Upgrade mcp dependency from 1.x to 2.0

[Copilot]

mcp 2.0 reworks the low-level Server API: @server.list_tools()/@server.call_tool() decorators are gone (replaced by on_list_tools/on_call_tool constructor kwargs), handlers now take (ctx, params) instead of (tool_name, arguments), and server.request_handlers[...] / response .root unwrapping are removed. The blast radius in dsagt is narrow: only the dispatch shell and test harness touch the SDK Server object directly.

Dependency

• Bumped mcp>=1.0.0,<2.0.0 → mcp>=2.0.0,<3.0.0 in pyproject.toml, re-synced lockfile.

Dispatch shell (src/dsagt/mcp/server.py)

• build_dispatch_server() now passes on_list_tools/on_call_tool callables to Server(...) instead of using the removed decorators.
• on_list_tools returns types.ListToolsResult(tools=...) instead of a bare list.
• on_call_tool(ctx, params) reads params.name/params.arguments (previously (tool_name, arguments)) and returns types.CallToolResult(content=[...]) instead of a bare content list.
• Per-tool dispatch logic (span opening, ValueError → error-dict, str-vs-dict formatting) is unchanged.

async def on_call_tool(ctx, params: types.CallToolRequestParams) -> types.CallToolResult:
    tool_name = params.name
    arguments = params.arguments
    handler = handlers[tool_name]
    ...
    return types.CallToolResult(content=[types.TextContent(type="text", text=text)])

return Server(name, on_list_tools=on_list_tools, on_call_tool=on_call_tool)

Test harness

• tests/mcp_helpers.py (call_tool_sync/call_tool_async) and the inline duplicates in test_dsagt_server.py, test_kb_search_filters.py, test_knowledge_server.py, test_memory_tools.py now use server.get_request_handler("tools/call"|"tools/list").handler instead of the removed server.request_handlers[...] dict, dropping the .root unwrap.
• Consolidated the three inline-copy helpers onto the shared mcp_helpers.py versions where practical, rather than patching each copy separately.
• Fixed tool.inputSchema → tool.input_schema reads in tests — mcp 2.0 renamed the types.Tool Python attribute (the inputSchema alias still works for construction, so production tool-building code is unaffected).

mcp.server.stdio.stdio_server, NotificationOptions, InitializationOptions, and server.get_capabilities(...) are unaffected by the bump, so _run_stdio, create_dsagt_server, and main() needed no changes.

• Fixes Upgrade mcp dependency from 1.x to 2.0 #39

[github-actions]

PR Preview Action v1.8.1
🚀 View preview at https://AI-ModCon.github.io/dsagt/pr-preview/pr-40/
Built to branch gh-pages at 2026-08-14 20:00 UTC. Preview will be ready when the GitHub Pages deployment is complete.

[jeanbez]

While testing, I saw these two:

1. Previously @server.call_tool() validated against inputSchema by default, which does not seem to be the case now, so every malformed agent call now reaches the handler raw. This should resolve that:

    jsonschema.validate(instance=arguments, schema=schemas[tool_name])
except jsonschema.ValidationError as e:
    result = {"status": "error", "error": f"Input validation error: {e.message}"}

Plus this in server.py line 92 (after tool_category):

schemas = {tool.name: tool.input_schema for tool in tools}

2. arguments can be None. The PR does arguments = params.arguments adding or {} should resolve the issue

arguments = params.arguments or {}

Together they turn a clean Input validation error: 'query' is a required property into Unexpected error: 'NoneType' object is not subscriptable, plus a full traceback in the server log for what is a client-side error.

[shreddd]

@jeanbez - we don't strictly need to accept this PR either. Let's discuss more today.