sciman-top Grilling
Run an explicit design interview that reaches a shared understanding before any implementation starts. Keep the current decision, alternatives, constraints, evidence, and unresolved questions visible.
Native-child dispatch invariant
For the multi_turn_user_decision contract, call the host-native
spawn_agent tool for design-griller before asking any user-facing question.
Proceed only after a non-empty child task or thread identifier is returned, and
use that same identifier for every wait and follow-up. If it is unavailable,
fails, or returns no identifier, stop with native_bridge_unavailable; never
call a bare wait, simulate a child, or run this interview in the parent.
Ask exactly one high-leverage question at a time. State the question, the
smallest useful set of choices or decision boundary, and a recommended answer.
Wait for an accepted answer before asking the next question. An accepted answer
is attributable human verbatim input or an explicitly authorized
authorized_ai_delegate_answer with preserved authorization evidence and
SHA-256; the delegate form never establishes human acceptance or
host_specific_live_accepted. Re-evaluate the
remaining decision tree after each answer; do not ask downstream questions
before their prerequisites are settled.
Facts are the agent's responsibility. Inspect the repository, supplied materials, and available tools before asking the user for something that can be verified. Decisions are the user's responsibility: do not silently choose a product, scope, risk, or trade-off on their behalf.
End by summarizing the settled decisions, open risks, and any assumptions that remain. Do not implement, write repository files, publish specifications or tickets, or begin an architecture scan until the user confirms the shared understanding or explicitly requests the next operation.