mcp_tool hooks get their own runner, can block actions, and are restricted from project settings

What

A hook is a command you set up to run automatically at certain moments while Claude Code works. An mcp_tool hook calls a tool on an MCP server, an outside service that gives Claude tools. Claude Code could already read this hook type when a remote list allowed it. What changed now:

• Own runner: mcp_tool hooks now run through their own path instead of the general one. They wait for plugin MCP servers to connect, up to the hook's time limit, and report timed out if the servers do not connect. They report MCP server '...' not connected if the server is missing, and "The organization's MCP server lists could not be applied to" a server when that check fails.
• Blocking: a response that is not empty is now recorded, and a new check can mark the result as blocked, using the response as the output. A blocking failure has the new code hook_mcp_tool_failed. Before, these hooks only failed when aborted, on an error, or when the call was not ok, and they never blocked.
• Project settings: hooks from projectSettings are skipped when userSettings exist and a new check is true. They may also go through a different runner when an input stream is available. The check for restricted hook sources now also matches mcp_tool hooks from projectSettings. It also matches plugin agent and http hooks whenever pluginId is set; before, agent hooks also needed a further check.
• Hook-type check: mcp_tool is now included, and the check for agent hooks now looks at the right value, fixing an earlier slip.
• New options: hook conditions gain an if_pattern option controlled by tengu_woven_ladder, which defaults to an empty list. Nothing has been read about that flag. A hook settings schema gains needsCtrlYFix, which can only be true.
• Abort check: the hook list is now passed to a new function before the abort check.

Why

If you write mcp_tool hooks, they can now stop an action, and their failures say what went wrong. Hooks that come from a project's settings are treated more strictly, so a project cannot always run them the way your own settings can.