Ultrathink

Status as of 2026-09-13: trimmed to inputs, outputs, and constraints. The instructions on how to reason first-principles were removed — this generation does that unprompted, and scripting it crowded out the context that actually helps.

Use for an architectural fork, a design that feels wrong but works, a recurring bug whose cause keeps moving. Not for anything where the answer is a lookup or the shape of the solution is already agreed.

Before answering: recall this realm's wisdom, failures, and patterns. Arriving at a conclusion the project already reached and discarded is the main failure mode here, and chitta is the only place that history exists.

Constraints:

• Run in the orchestrator. Ultrathink in a subagent starts without the context that makes the exercise worth doing; if it must be delegated, use subagent_type: "fork" so it inherits this session.
• Name the assumption you are discarding, not just the conclusion. A recommendation that doesn't say what it stopped believing can't be audited.
• Contradicting an existing chitta memory is a finding, not an inconvenience — surface the conflict rather than quietly picking a side.

Output: the recommendation, the assumption it overturns, and the cost of being wrong. Promote anything that generalizes beyond this problem with grow(type=wisdom); leave the rest in the response.