judgment-day

name	judgment-day
description	Dual-model adversarial review: Claude and Copilot review independently, verdicts are synthesized.
when_to_use	After sdd-apply, as an enhanced alternative to sdd-verify. Also usable standalone for any code review.
allowed-tools	["Read","Bash","Glob","Grep","mcp__invencible-memory__memory_recall","mcp__invencible-memory__memory_save","mcp__invencible-hub__delegate_task","mcp__invencible-hub__get_task"]
model	sonnet
effort	high
user-invocable	true
hub-skill-ids	["review","reasoning"]

Goal

Two independent reviewers — Claude (you) and Copilot CLI (via hub) — review the same code without seeing each other's findings. Verdicts are synthesized into a confidence-weighted table.

Protocol

Phase 1 — Blind Delegation

1. Delegate a review task to Copilot CLI via the hub:

   delegate_task(
     skill_id="review",
     description="Independent code review (judgment-day blind pass).\n\n"
                 + "Review scope: <files changed in this SDD run>\n"
                 + "Spec to check against: <sdd-<change_id>-spec from memory>\n\n"
                 + "Output format — for each finding:\n"
                 + "  SEVERITY: CRITICAL|WARNING|SUGGESTION\n"
                 + "  FILE: <path>:<line>\n"
                 + "  ISSUE: <description>\n"
                 + "  FIX: <recommended fix>\n\n"
                 + "Save findings to memory: topic_key=sdd-<change_id>-review-copilot"
   )
   

2. While Copilot reviews, run your own independent review:
   • Load spec: memory_recall(topic_key="sdd-<change_id>-spec")
   • Read all changed files
   • List every finding with SEVERITY, FILE:LINE, ISSUE, FIX
   • Save your findings: memory_save(topic_key="sdd-<change_id>-review-claude", ...)

Phase 2 — Wait and Collect

3. Poll for Copilot's result: get_task(task_id=<delegated_task_id>)
   • Poll up to 5 times with ~10s between attempts
   • If no result after 5 polls: note "Copilot review unavailable" and proceed with Claude-only verdict

Phase 3 — Synthesis

4. Compare findings from both reviewers. Classify each finding:

   • CONFIRMED: found by BOTH reviewers → high confidence, must fix
   • SUSPECT-A: found by Claude only → triage, likely fix
   • SUSPECT-B: found by Copilot only → triage, likely fix
   • CONTRADICTION: reviewers disagree on the same location → escalate to human

5. Save synthesis to memory:

   • title: Judgment Day: <change_id>
   • topic_key: sdd-<change_id>-judgment
   • type: context
   • content:

     ## Verdict Table
     
     | Finding | Severity | Reviewer | Location | Issue |
     |---------|----------|----------|----------|-------|
     | <desc>  | CRITICAL | CONFIRMED | file:line | ... |
     ...
     
     ## Fix Loop Status
     Iteration: <N> / max 2
     
     ## Escalations
     <contradictions requiring human judgment>
     

Phase 4 — Fix Loop

6. Fix all CONFIRMED and SUSPECT findings (severity CRITICAL or WARNING)
7. Re-run the synthesis — maximum 2 iterations
8. If issues remain after 2 iterations: escalate to human with the verdict table

Standalone Usage

When used outside the SDD workflow, replace sdd-<change_id>-spec references with a description of what the code is supposed to do, and scope the review to the files you specify.