Brainstorming

Treat the request supplied with $brainstorming as the topic. Fan it out to fresh subagents whose prompts use genuinely different thinking operations, then turn their contributions into one coherent landscape.

1. Design the Perspectives

1. Extract the topic, goal, known facts, constraints, and desired outcome from the invoking prompt and relevant conversation.
2. Choose at least three pairwise-distinct angles suited to the topic. Consider stakeholder viewpoints, upside, failure, feasibility, assumption inversion, analogy, second-order effects, time horizons, or a wildcard—not as a checklist, but as raw material for task-specific lenses.
3. Keep adding an angle while it covers a consequential aspect the existing angles do not. Stop when another angle would mostly duplicate one already chosen.

The perspective set is complete when it contains at least three non-overlapping lenses and every additional credible lens would add little new territory.

2. Dispatch Fresh Thinkers

Write each subagent prompt independently so its wording, questions, and reasoning frame embody its lens. Include enough shared facts for the subagent to understand the task without conversation history, while varying how the problem is posed—for example, as a counterfactual, critique, analogy search, stakeholder narrative, opportunity expansion, or constraint challenge. Scope each subagent to analysis and response only, and ask for concrete insights with the reasoning behind them.

Spawn one subagent per prompt with fork_turns: "none". Omit the model and reasoning_effort arguments so each subagent inherits both from the caller. Launch independent prompts in parallel when slots permit; if capacity is constrained, continue as slots become available until every designed perspective has run.

Dispatch is complete when every designed perspective has an independently worded prompt and every prompt has been assigned to a fresh subagent.