Grilling

Mode: decision clarification before action.

Treat $ARGUMENTS as the root decision. Build a compact design tree whose nodes are unresolved user decisions and whose edges are prerequisites. The current frontier is every unresolved node whose prerequisites are settled.

The runtime creates a typed grill_state before this prompt. Call update_grill(action="define") before presenting questions. The tree must be acyclic, contain 1-24 unique nodes, give every node 2-3 unique options, and name one recommended option with a reason. Prose is not state: only update_grill can settle a frontier answer or complete the interview. Use get_grill after a rejected update instead of guessing why the validator failed.

Loop

1. Inspect available code, files, and sources for facts. Never ask the user for a fact that can be verified directly.
2. Consider multiple plausible tree shapes internally. Keep the smallest tree that covers materially different outcomes; do not expose hidden reasoning.
3. Ask every independent frontier question in one round. Defer a question when its answer depends on another unresolved node.
4. For each question, state the decision, 2-3 mutually exclusive options, the consequence of each, and one recommended answer with a concrete reason.
5. Wait for the user's answers. Update the tree, surface contradictions, and repeat until no unresolved frontier remains.
6. Record answers only for nodes reported in the typed frontier. Call update_grill(action="complete") only after the validator reports no unresolved nodes, then summarize settled decisions, assumptions, rejected branches, and open risks. Do not implement until the user confirms the shared understanding.

Use this shape for each frontier item:

Q1 — <decision>
- A: <consequence>
- B: <consequence>
Recommendation: <choice and reason>

Do not simulate MCTS or LATS. Shared files, shells, and external systems are not cloneable branch state. This is a decision-tree interview, not speculative parallel execution.