Interview the user relentlessly about a plan or design until reaching shared understanding, resolving each branch of the decision tree.
Interview me relentlessly about every aspect of this plan until we reach a shared understanding. Walk down each branch of the design tree, resolving dependencies between decisions one-by-one. For each question, provide your recommended answer.
Follow the detect-and-defer protocol from the Shark reference. Check SHARK_TASK_HASH:
echo "$SHARK_TASK_HASH"
adapter context to know how to spawn agents. If no adapter is wired, warn and fall back to inline execution (no background agents — explore the codebase yourself when needed).Before starting the interview, spawn research remoras to build a fact base from the codebase. This prevents slow, serial exploration during the interview.
Identify research axes: From the user's plan, extract 3-5 independent research questions that the codebase can answer. Examples:
Spawn research remoras: Spawn one remora per research axis, all in a single message with parallel tool calls using run_in_background: true.
// All research remoras spawned concurrently via adapter collect()
collect([
{
task: "Explore the codebase to answer: {research question 1}. Repo root: {cwd}. Return a concise factual summary (under 300 words). Do NOT make recommendations — just report what you find.",
name: "xavier research: {short label}",
subagent_type: "Explore"
},
// ... one entry per research question
])
Collect results: As each remora completes, record its findings. Once all have reported, compile a Research Brief — a structured summary of codebase facts organized by research axis. Keep it under 500 words total.
Show the research brief to the user: Present the brief before starting the interview so the user and Xavier share the same factual ground.
With the research brief as shared context, interview the user one question at a time: