[POC] Add MCP tool schema regression testing with mcp-recorder

[caballeto]

[POC] Adds regression test coverage for MCP tool schemas using mcp-recorder -- like VCR.py but for MCP servers. Records the full protocol exchange into a JSON cassette and verifies it hasn't changed on every PR.

Why this matters

The MCP server exposes 20+ tools to AI agents. The tool names, descriptions, input schemas, and annotations are the contract -- when they drift, downstream agents break silently. Today there's no test that catches "someone changed a .describe() string" or "a tool got accidentally filtered out."

What this covers

A single cassette (list_tools.json, ~3200 lines) captures 3 interactions:

• Protocol handshake -- initialize response with protocol version, capabilities, server info
• Tool schemas -- tools/list with complete input schemas, descriptions, and annotations for all registered tools

If a tool is renamed, a parameter is removed, a description changes, or an annotation flips, the CI diff shows exactly what broke.

How it works

The server is spawned via stdio with MONDAY_TOKEN=test-token. The token is never validated for initialize + tools/list -- these only enumerate the in-memory tool registry. No network calls, no secrets, no real API access needed.

Why this is a POC

This PR adds a Python dependency (mcp-recorder) to a TypeScript repo, so it may not fit the project's tooling preferences. However, it demonstrates the benefits of a snapshot-based testing:

• Schema drift detection is cheap -- one YAML file, one cassette, zero secrets
• The cassette diff is the review -- any tool schema change shows up as a JSON diff in the PR, making it explicit and reviewable
• The full tool surface captured -- the committed cassette serves as living documentation of every tool's public interface

Changes

All additive -- only .github/workflows/pull_request.yml and package.json modified.

integration/
  scenarios.yml                      # list_tools scenario in ~14 lines of YAML
  cassettes/list_tools.json          # golden cassette (3 interactions, ~3200 lines)
  requirements.txt                   # mcp-recorder>=0.4.1
.github/workflows/pull_request.yml   # CI: verify step after build
package.json                         # mcp:record and mcp:verify convenience scripts

Run locally

pip install -r integration/requirements.txt
yarn build

mcp-recorder verify \
  --cassette integration/cassettes/list_tools.json \
  --target-stdio "node packages/monday-api-mcp/dist/index.js" \
  --target-env MONDAY_TOKEN=test-token

Update after intentional changes

yarn build
mcp-recorder verify \
  --cassette integration/cassettes/list_tools.json \
  --target-stdio "node packages/monday-api-mcp/dist/index.js" \
  --target-env MONDAY_TOKEN=test-token \
  --update

The cassette diff in the PR review shows exactly what changed in the tool surface.