Execute the tool FOR REAL with the given params, through the same executor a production call takes: one HTTP call to the target, logged as a provider_calls row (source_service api, tool_type custom_tool), 0 credits, a fresh idempotency key per press. result carries the outcome verbatim: status, output {status, body, extracted}, the classified error_code and a message that opens with the token of a tool-level failure (custom_tool_disabled, input_binding...), and provider_call_sid for the snapshots. Use it to see what a column would get, or why a tool returns nothing (compare output.body with output_mapping). A dry-run that does not hit the target would validate nothing. Throttled 10 a minute.

Contract:

• MCP tool test_custom_tool, registry package mcp.integrations/custom_tools, mount integrations.providers.
• Operation action, response envelope action.
• Flags: none.