heavy

name	heavy
description	Explicitly invoked parallel investigation when distinct evidence lanes can materially improve the answer.

Heavy

Outcome

Spend more intelligence only when distinct evidence or independent reasoning can materially improve the answer.

Activation

Activate only when the user explicitly names Heavy or $heavy. Do not infer it from task difficulty or length.

Root authority

The root chooses whether to delegate, assigns non-overlapping evidence territories, resolves disagreement, and owns the synthesis. No vote or numeric score substitutes for judgment.

Invoking Heavy does not expand write, external-action, credential, model-routing, or delegation authority. Keep workers read-only unless the underlying task separately authorizes mutation.

Context

Freeze the question, inspect the strongest available evidence, and identify what another lane could learn that the root cannot learn as cheaply in sequence.

Constraints

Do not use fixed lane counts or canned personas. Keep each lane brief bounded and self-contained. Parallel work must add information, not duplicate opinions. Stop spawning when evidence converges.

Evidence

The synthesis uses the evidence that changed or supported the conclusion, preserves uncertainty, and identifies any decisive observation still missing.

Failure boundary

Return an explicitly partial result when a required evidence lane fails. Do not present multi-model confidence that was not actually obtained.

Optional references

Load unknowns prompts only when they expose a real blind spot in the current question.

Codex runtime

Use Codex delegation only when it is in scope. Prefer parallel read-only investigation before parallel writers, and keep each worker's return compact and evidence-bearing.