Judgment Day

Activation Contract

Load this skill only when the user explicitly asks for Judgment Day, dual/adversarial review, or equivalent Spanish trigger (juzgar, que lo juzguen). Review a specific target: files, feature, PR, or architecture slice.

Within this QA boilerplate, this skill is cited as an OPTIONAL adversarial augmentation by:

• /test-automation — Review phase for high-risk test changes
• /git-flow-master — pre-PR gate when the diff is large or touches shared fixtures / base classes
• /framework-development — pre-archive review of framework evolution diffs

It is never invoked automatically. Users opt in via explicit trigger.

Inputs — read these first, in this order

Canonical reading order for any AI starting cold on a Judgment Day review. Read in order; stop earlier when the target is small enough that later inputs add no signal.

1. The diff / files / PR / architecture slice under review — the literal target the user named.
2. AGENTS.md — repo conventions, Critical Rules, behavioral layer (the judges must score against these, not generic best-practice).
3. .agents/skills/REGISTRY.md — skill registry; resolve which project skills apply to the target's file paths + task type, and inject the same Skills to load before work block into both judge prompts.
4. The change's spec / PR description / Jira ticket — the stated intent. Judges score against intent, not their imagined intent.
5. references/prompts-and-formats.md — judge prompts, fix prompts, warning rubric, verdict table format.
6. Prior judge outputs from earlier rounds (Round 2+ only) — to detect regressions or stale findings vs. new ones.

Hard Rules

• Resolve project skills before launching agents: read skill registry, match skill paths by target files/task, and inject the same Skills to load before work block into both judge prompts and fix prompts.
• Launch two blind judges in parallel with identical target and criteria; never review the code yourself.
• Wait for both judges before synthesis; never accept a partial verdict.
• Classify warnings as WARNING (real) only if normal intended use can trigger them; otherwise downgrade to INFO as WARNING (theoretical).
• Ask before fixing Round 1 confirmed issues.
• After any fix agent runs, immediately re-launch both judges in parallel before commit/push/done/session summary.
• Terminal states are only JUDGMENT: APPROVED or JUDGMENT: ESCALATED.
• After 2 fix iterations with remaining issues, ask the user whether to continue.

Decision Gates

Execution Steps

1. Confirm target and optional custom criteria.
2. Resolve exact skill paths from registry or warn if missing.
3. Start Judge A and Judge B concurrently via delegation.
4. Synthesize findings into confirmed, suspect, contradiction, and INFO buckets.
5. Ask before Round 1 fixes; delegate a separate fix agent for confirmed approved fixes only.
6. Re-judge in parallel after fixes; repeat until approved, escalated, or user asks to stop.
7. Before any terminal action, verify every active Judgment Day has a terminal state.

Output Contract

Return ## Judgment Day — {target} with round number, verdict table, confirmed/suspect/contradiction counts, fixes applied, re-judgment result, Skill Resolution, and final JUDGMENT: APPROVED ✅ or JUDGMENT: ESCALATED ⚠️.

Anti-patterns — NEVER do these

• J1. NEVER let the two judges see each other's verdicts before synthesis — blind review is the entire value proposition; cross-contamination collapses two judges into one.
• J2. NEVER skip the re-judge step after fixes — fixes can regress other behavior; only a fresh dual verdict confirms the change actually landed clean.
• J3. NEVER run judgment-day on a diff with > 30 files in a single pass — signal degrades past that size. Split into focused per-area reviews and synthesize manually.
• J4. NEVER override a judge's reject without documenting WHY in the synthesis — silent overrides destroy the audit trail and the next reviewer cannot tell if the issue was triaged or missed.
• J5. NEVER use judgment-day for trivial typo / format-only / autogenerated diffs — orchestration overhead (two judge agents + synthesis + re-judge loop) exceeds the value of the verdict.
• J6. NEVER feed judges different prompts, different criteria, or different context — the blindness only works if both judges see the identical brief. Asymmetric briefs produce asymmetric verdicts that look like real disagreement but are artifacts.

References

• references/prompts-and-formats.md — judge/fix prompts, warning rubric, verdict tables, and language snippets.