feat(mcp): Add searchable tool catalog

[dcramer]

Add a searchable MCP tool catalog so ordinary Sentry operations can live behind experimental search_tools and execute_tool instead of all needing top-level MCP registration. The direct surface remains centrally declared, and the shared availability path keeps skill, constraint, capability, release-mode, and schema filtering consistent across direct and catalog execution.

Tool Catalog

Move ordinary tools into a flat tools/catalog/ tree, keep gateway/wrapper tools under tools/special/, and move shared implementation helpers into tools/support/. Each catalog tool remains its own module while the central surface policy decides which tools are exposed directly through tools/list.

Experimental Gateway

search_tools and execute_tool are marked experimental for now. Standard tools/list omits them, experimental mode includes them, and the eval mock stdio server opts into experimental mode so behavior evals can exercise discovery plus execution.

Search/Execute Contract

search_tools returns structured JSON instead of markdown, with only the data needed to call execute_tool: name, description, executable input schema, and MCP safety annotations. The returned input schema is filtered for the active session, so organization, project, and region arguments that will be injected from constraints are not shown to the model.

{
  "query": "issue",
  "results": [
    {
      "name": "get_issue_details",
      "description": "...",
      "inputSchema": { "type": "object", "properties": {} },
      "annotations": { "readOnlyHint": true }
    }
  ]
}

MCP Disclosure And Behavior

Tests cover standard and experimental tool disclosure, agent-mode use_sentry disclosure, catalog-only discovery and execution, lean search result metadata, constrained schema filtering, per-tool snapshot baselines, and eval cases that require model behavior through search_tools plus execute_tool.

Validated with pnpm run tsc, pnpm run lint, pnpm run test, pnpm --filter @sentry/mcp-server-evals run build, pnpm run measure-tokens, and a mock stdio MCP smoke check. Token measurement reports 22 stable direct tools and 14,012 total tokens.

[cursor]

Cursor Bugbot has reviewed your changes and found 1 potential issue.

^{❌ Bugbot Autofix is OFF. To automatically fix reported issues with cloud agents, enable autofix in the Cursor dashboard.}

^{Reviewed by Cursor Bugbot for commit 6cf7fce. Configure here.}