Subagents get their own effort level and token budget

What

A subagent is a helper Claude starts with the Agent tool to work on part of a task. When the Agent tool now starts or resumes a subagent, a new helper works out two things for it: an effort level, meaning how hard the model reasons, and a total-token budget, meaning how many tokens the subagent may spend. Claude Code already showed Claude a running count of tokens left. Subagents now get a budget and countdown of their own.

• Forked subagents with a token budget get a total_tokens_reminder message in their prompt, and the budget is passed to the run as totalTokensReminderBudget. Resumed subagents carry the budget too.
• Effort can be lowered by a remote value, tengu_harmonic_riddle, or by the environment variable CLAUDE_CODE_HARMONIC_RIDDLE. An environment variable is a setting passed to a program from outside it. Text names an effort level. A negative whole number steps down from the session's own effort. It only ever lowers effort, never raises it. It does not apply when the subagent has an explicit effort or its effort is already "low". Nothing has been read about tengu_harmonic_riddle.
• Remote subagents are left out of this.
• Usage records for a subagent launch now include the session effort, the subagent effort, subagent_total_tokens_budget and subagent_budget_on_top_of_start.
• When a teammate is spawned, the result now includes its agentId and resolvedModel.

The Agent tool's description, as captured under this release, is unchanged from the release before.

Why

Subagents can now see how many tokens they have left, which can change how they pace their work. If the remote value or the environment variable is set, subagents may reason less hard than the main session.