challenge

/challenge - GO / NO-GO verdict on a proposal

Judge in two phases whether a discovered problem is real and a proposed idea usable, so the next decision proceeds from a verified GO / NO-GO.

Input

$ARGUMENTS may contain the challenge target (a proposal file path or description). If empty, stop and ask the user to specify the target, since silent inference from the conversation has high misfire risk. If non-empty, treat it as the challenge target.

Phase 1 Grill

Grill the proposal from evidence on its own, then return only the unresolved residual to the user, sorted by reversibility. Each open question in the proposal is either a fact or a preference; facts get verified and may overturn the core. A preference that survives verification becomes the residual, routed to the user by reversibility.

1. Read OUTCOME.md if present. Its done state / non-goals / constraints stand in for part of the user's intent. If absent, infer the outcome from $ARGUMENTS and the conversation and confirm it via AskUserQuestion. Pass the confirmed outcome to the Phase 2 outcome critic as its evaluation axis
2. List the open questions in the proposal and sort each into fact or preference. A fact is a question evidence settles to one answer; a preference is one needing a choice (priority / scope / trade-off). Sort by the nature of the question, not by advisor confidence. Do not treat a preference as fact
3. Run a loop to verify the facts. subagents check facts in parallel → advisor synthesizes, re-checks the sorting, and names the next evidence to get → the main session organizes the results and decides whether to continue. Break when no further evidence would change the sorting. Cap at 3 rounds; questions left unresolved fall to the residual
4. If a verified fact overturns the proposal's core (the core targets a state that already holds, or a verified fact contradicts it), stop and skip Phase 2. If the core survives and only a sub-claim is broken, proceed on the surviving part. Stop only when fact evidence backs it, not on advisor opinion alone. Put the grounds for overturning in the Why
5. The questions left at break are the residual, genuine preferences proven through verification. advisor scores each residual with a best-guess plus reversibility / blast-radius, then asks via AskUserQuestion only the irreversible or high-impact ones (cap 7). For the rest, proceed on the best-guess as a stated assumption logged in the output Why. Keep the questions skipped via subagent and the residuals advanced on assumption in the Why, leaving the user a way to veto later

Output to Phase 2

Aggregate grill findings into critic-design input schema before spawning.

Field	Source
source	user-grill
artifact_type	inferred from $ARGUMENTS (spec / plan / design / ADR / doc)
approach	one-line summary of the proposal core
decisions	architectural-level decisions crystallised during grill (terminology checks and scope minutiae excluded)
trade-offs	trade-offs surfaced during grill
referenced_files	files cited or read during grill
outcome_ref	OUTCOME.md path plus a digest of its done state / non-goals / constraints. If OUTCOME.md is absent, use the outcome confirmed in Step 1

Phase 2 Devil

Land the Phase 1 material on two critic-design (internal attack / OUTCOME.md attack), adversarially probing for holes.

Pass	Role
advisor	Evidence synthesis / self-consistency check / sorting audit
critic-design (internal)	Attack the proposal on its own terms (hidden weaknesses, failure modes)
critic-design (outcome)	Attack whether it reaches the outcome (outcome fit / non-goal / constraint breach)

1. Compose the Phase 2 input from the Phase 1 aggregation and the original $ARGUMENTS context
2. Spawn two critic-design via Task in parallel (subagent_type: critic-design, background: false). One handles the internal attack, the other takes outcome_ref for the outcome attack (skip when no outcome is available). Mention ARCHITECTURE.md if present
3. Wait for both, reconcile verdicts and weaknesses, and dedupe overlap

Output

Lead with the Verdict, concentrate the backing in Why, name the next move in Actionable items.

Section	Content
Verdict	GO / NO-GO. Note the condition if conditional. One line, first
Why	The fact-verification evidence (reproduction / refutation), the two critic-design verdicts (internal / outcome), and the residuals advanced on a best-guess with their reversibility (the user's veto targets). Close with a one-line note that the verdict is a judgment within evidence range
Actionable items	Top 3 concrete actions (keep / remove / revise)