fix(mcp): default inputSchema to {type:"object"} for registered tools in tools/list

[vishalhabib99]

MCPMock lists a registered tool exactly as it was registered. A tool added without inputSchema is therefore listed without one, but the MCP schema requires inputSchema on every Tool. The official clients reject the whole tools/list result, not just that tool.

Two documented paths produce such a tool:

• The docs Quick Start: mcp.addTool({ name: "search", description: "Search the web" })
• mcp.onToolCall("name", handler) for a name that was never added. That stores { name }.

Repro (@copilotkit/aimock 1.45.0; current main has the same handler):

const mcp = new MCPMock({ port: 38733 });
mcp.addTool({ name: "search", description: "Search", inputSchema: { type: "object", properties: { query: { type: "string" } } } });
mcp.addTool({ name: "rich", description: "rich" });
await mcp.start();

• Python SDK 2.3.0 list_tools(): ValidationError: ... ListToolsResult tools.1.inputSchema Field required
• TypeScript SDK 1.30.0 client.listTools(): Invalid input: expected object, received undefined at tools[1].inputSchema

So a test that points a real MCP client at the mock fails at listTools, before it reaches the tool it's testing.

Fix: in tools/list, a registered tool with no inputSchema is listed with { type: "object" }. This is the default McpFakes.listTools() already uses ("A tool with no inputSchema lists { type: "object" }"), so registered tools and fake tools now behave the same. A tool that has a schema is listed unchanged. The recorded-list path also gets the default, because it reuses the same tools array.

Tests: mcp-mock.test.ts.

• The existing "tools/list returns registered tools" case asserted that calc is listed without inputSchema. I updated it to expect { type: "object" }.
• I added a case covering both the addTool and onToolCall paths. Both fail on main and pass here.
• prettier --check, eslint and pnpm typecheck pass.
• The full pnpm test run on macOS shows the same 36 failures with and without this change, all in validate-cli*, mcp-fakes-report-route, mcp-fakes-doc-samples and mcp-fakes-redgreen. They look like they need a built CLI or other local setup.

I left out one related gap: resources/read for a resource registered without content returns no text or blob. That would be a separate PR if you want it.

[chatgpt-codex-connector]

Codex Review Summary

This comment shows the latest Codex review activity on this pull request.

Review	Status	Commit	Review trigger
📝 Code Review	⚠️ Failed 2026-10-11T23:08:54.880074Z	5e8074c	PR opened

ℹ️ About Codex in GitHub

Your team has set up Codex to review pull requests in this repo. Reviews are triggered when you

• Open a pull request for review
• Mark a draft as ready
• Comment "@codex review" or "@codex security review".

Codex reacts with 👀 while any review is running, comments if it has suggestions, and reacts with 👍 once all reviews finish with no findings.

[pkg-pr-new]

Open in StackBlitz

npm i https://pkg.pr.new/@copilotkit/aimock@529

commit: 3b13230

[jpr5]

Thank you @vishalhabib99! 🙏