Subagents refuse paid runs on relayed user approval; run it yourself

In Claude Code, a background subagent told to run a paid API check refused, even when the coordinator quoted the user's exact AskUserQuestion answer. It treats any agent message as unverifiable consent for spending money. What worked: ask the subagent to return the exact one-line command with no key values, run that command in the main session (which holds the user's real approval), then hand the output file back to the subagent for grading and commit. Ask for spend approval in the main session before you delegate, and plan for the main session to do the paid step.