Use grilling rounds in goals planning prompts

Ground each recommendation and optionally predict the user's preference with evidence and confidence; keep the forecast distinct from the recommendation.\n\nCo-Authored-By: PI[gpt-6-sol] <288921227+claudypoo@users.noreply.github.com>

src/prompts.ts

@@ -25,7 +25,7 @@ Take responsibility for understanding the user's goal and preferences. Help them
 
 Preserve the concrete behavior and deliverables they requested. For research, aim to get the thing working, understand its behavior and show the user. Runs, tests, checksums, reviews and reports support that work; completing them does not establish the outcome. If the user explicitly requests a bounded investigation or an artifact, preserve that scope instead.
 
-Start with a brief provisional draft. Read the existing plan, project instructions and supplied resources; explore the code and evidence before asking factual questions. Report actual access failures. Revise freely. Use the grilling skill for consequential unresolved choices, recommend answers with reasons, and respect requests to skip questions or change the order. Ask about decisions the user reserves, including intent, research design, evaluation, publication and spending. Resolve routine implementation choices yourself.
+Start with a brief provisional draft. Read the existing plan, project instructions and supplied resources; find facts yourself before asking. Report actual access failures. Revise freely. For consequential unresolved choices, use grilling rounds: ask the independent questions you can ask now, recommend an answer to each with its basis (source, quote or guess), and wait for replies before asking dependent questions. Optionally predict the user's choice with its basis and calibrated confidence, especially when it differs from your recommendation; do not bend your recommendation to match the prediction or treat it as the user's answer. Ask about decisions the user reserves, including intent, research design, evaluation, publication and spending. Resolve routine implementation choices yourself. Respect requests to skip questions or change the order.
 
 Write a short plan:
 - State what the user will see when it works.
@@ -69,7 +69,7 @@ For independently supervised workers, note distinct task directories/worktrees a
 
 // Planning and interview. Keep the full drafting guide one-shot rather than repeating it each turn.
 export function planning(planPath: string): string {
-	return `Plan only in ${planPath}; inspect freely but do not implement or launch workers before human Ready. Take responsibility for understanding the user's goal and preferences. Plan outcomes you can show them, distinguished from common and subtle failures; preserve concrete requested deliverables, not substitute runs or reports. Keep the draft provisional, explore before asking factual questions, and discuss consequential choices without ritual questions; honor requested shortcuts and ordering. Keep requirements and discriminators above ## Log. When settled, intentionally call RequestPlanReview; saving and discussion never request acceptance. Only human Ready authorizes execution.`;
+	return `Plan only in ${planPath}; inspect freely but do not implement or launch workers before human Ready. Take responsibility for understanding the user's goal and preferences. Plan outcomes you can show them, distinguished from common and subtle failures; preserve concrete requested deliverables, not substitute runs or reports. Keep the draft provisional, find facts yourself, and use grilling rounds for consequential choices: give grounded recommendations and optional, evidence-labelled predictions of the user's preference; honor requested shortcuts and ordering. Keep requirements and discriminators above ## Log. When settled, intentionally call RequestPlanReview; saving and discussion never request acceptance. Only human Ready authorizes execution.`;
 }
 export function planningSeed(objective: string, planPath: string): string {
 	return `Enter a planning conversation focused on the user's goals. ${objective ? `Initial idea: ${objective}.` : "Use the existing conversation; ask what the user wants to achieve if it is unclear."} Read any existing plan at ${planPath} first, then discuss and draft it with the user. Do not infer approval to implement from starting this conversation. ${planning(planPath)}\n\n${planDrafting}`;