challenge

Challenge

Adversarial review with two modes. Standard: three native challenger subagents (convergent · systems · critical) stress-test reasoning, Gemini retrieves documented constraints — all run independently, Claude merges and judges. Council (--council): Four parallel perspective agents (devil's advocate, optimist, pragmatist, operator) plus Gemini — Claude synthesizes across all five.

Deep (--deep, or auto on high-stakes): any mode, plus an adversarial verification stage — every merged finding is independently re-attacked by N skeptics (verify-findings workflow), so plausible-but-wrong findings are dropped to FALSE_POSITIVE and survivors get a calibrated severity before you see them. All agents are native subagents on the subscription — no ruflo, no API key.

1. Gather context about what to challenge:

   • If $ARGUMENTS provided, use it as the description of what to challenge.
   • If $ARGUMENTS contains --council: strip --council from the target description, activate council mode — step 4a spawns 4 parallel perspective agents instead of one Opus agent. All other steps unchanged.
   • If $ARGUMENTS contains --hats: strip --hats from the target description, activate hats mode — step 4a spawns 4 parallel hat agents (White/Black/Yellow/Green) instead of hive-mind workers. Mutually exclusive with --council; if both present, --hats wins.
   • If $ARGUMENTS contains --deep: strip --deep from the target description, activate deep mode — runs the adversarial verification stage (step 4d) over the merged findings before synthesis. Combinable with any mode (standard / --council / --hats).
   • Auto-deep: even without --deep, activate deep mode automatically when the target is high-stakes — an architectural or irreversible decision, cost-of-being-wrong rated expensive/catastrophic in step 2, or an M+ effort plan. When auto-triggering, tell the user one line: "High-stakes target — running deep verification."
   • If no arguments: conversation-context inference — scan recent conversation turns (both user and assistant) to identify the most significant unchalllenged decision, plan, or proposal. Priority order:
     1. A plan or architecture decision being actively discussed
     2. A proposal the user described or asked about
     3. Your own most recent substantive recommendation
     4. A trade-off or choice where alternatives weren't explored
   • Extract: the decision/plan itself, the key constraints mentioned, and any assumptions stated or implied.
   • Frame it naturally: "I see we're discussing [X]. Let me stress-test that." — not "Let me challenge my last answer."
   • If the conversation has no substantive decision context (e.g., session just started, only greetings), then ask what to target.

2. Scope discovery — before launching challengers, ask 2-4 targeted questions that could change the challenge outcome.

   First, think (silently): given what's being challenged, what dimensions are non-obvious but consequential? What does Claude not know yet that would change which findings matter?

   Then surface 2-4 questions via a single AskUserQuestion call. Choose from the dimensions below — pick the ones most relevant to the target. Skip generic process questions.

   Dimension	When to ask	Example option set
   Deadline pressure	Plan or implementation work	"this week" / "this month" / "no fixed date"
   Cost of being wrong	Architecture or migration	"cheap (refactor)" / "expensive (data migration)" / "catastrophic (prod incident)"
   What's been tried	Recurring problem or refactor	"first attempt" / "one prior approach" / "multiple stalled attempts"
   Stakeholder breadth	Anything visible externally	"just me" / "one other person" / "wider team" / "external users"
   Constraints already locked	Mid-project decisions	"nothing locked yet" / specific commitments listed
   Hard ceiling on scope	Sprawling proposals	"S" / "M" / "L" / "explicitly unbounded"
   Gemini grounding	Brana-domain decisions — user wants to skip	"run agy (Recommended)" / "skip — no Brana docs apply"