preetamnath/panel-review

Panel Review

Three parallel reads on N focused questions, with disagreement preserved.

Steps

Step 1 — Dispatch

• Reviewers: R0 (you, concurrent) + R1, R2 (two general-purpose subagents in parallel).
• Prompt (identical for all three): artifact path(s), the questions, confidence score 0.00–1.00 per recommendation.
• Preflight: if any question is vague, sharpen via AskUserQuestion first.
• Output schema per question: recommendation, confidence, 1-2 sentence reason.
• Unprompted observations (any reviewer): 1-sentence claim, raising reviewer(s), confidence 0.00–1.00; merge duplicates.

Step 2 — Synthesize and confirm

• Sweep: for questions with max < 0.80 AND any reviewer < 0.70, run second-opinion on the highest-confidence reviewer's pick and raise Max to the synthesized score if higher.
• Order: sort by max confidence descending (post-sweep).
• Dependency override: if a decision can invalidate later questions, walk it first.
• Table:

| # | Question | R0 | R1 | R2 | Max |

• Low-consensus filter: drop questions where post-sweep max < 0.60 AND no reviewer ≥ 0.75; list in Step 5.
• Checkpoint: use AskUserQuestion to confirm the walk order before walking.

Step 3 — Walk questions one at a time

Queue rule: drop later questions invalidated by an approved decision, with a one-line reason.

For each question:

• Present: restate the question, quote the artifact span if location-specific, and show R0/R1/R2 side-by-side (plus second-opinion's synthesized recommendation, if Sweep ran).
• Decide: AskUserQuestion with options: apply / alternative / defer. Apply if actionable; otherwise record.
  • On pushback: run second-opinion on the challenged pick.

Step 4 — Walk unprompted observations

List all observations. Walk those with confidence ≥ 0.70 using Step 3's sub-procedure (skip R0/R1/R2 split).

Step 5 — Summary

One line each:

• Decisions applied — question # + pick.
• Decisions deferred — reason.
• Questions skipped (upstream invalidation) — reason.
• Questions dropped (low panel consensus) — question + max score.
• Unprompted observations — applied / deferred / dropped.