adversarial-debate

name	adversarial-debate
description	Use for significant code changes requiring multi-reviewer adversarial cross-challenge. Trigger: adversarial debate, cross-review, or when quality gates require L6 escalation.
tags	["quality","review","adversarial","debate"]

/adversarial-debate — Multi-Reviewer Confidence-Classified Debate

4 specialized reviewers independently assess code, then cross-challenge each other's findings. Output is confidence-classified: findings that survive debate are HIGH confidence.

Arguments

• /adversarial-debate <file-or-glob> — Standard 4-reviewer debate
• /adversarial-debate --rounds N — Set debate rounds (default: 2, max: 3)
• /adversarial-debate --converge — Enable convergence loop (fix + re-review, max 3 cycles)

When to Use

This is Quality Gate Level 5-6 in the escalating review system:

• Level 1-2: Self-check + peer review (small changes)
• Level 3: Devil's advocate (/opponents-view)
• Level 4: Cross-model (/codex review)
• Level 5: This tool — adversarial debate (significant changes)
• Level 6-7: Bridge Briefing (/bridge-briefing)

Auto-triggered by /evaluate when change size exceeds thresholds.

Workflow

flowchart TD
    Start([adversarial-debate]) --> Read["Read target files + git diff"]
    Read --> Spawn{{"Spawn 4+ reviewers in parallel\n(+A11y if HTML/CSS, +DB if migrations)"}}
    Spawn --> Sec["Security Reviewer"]
    Spawn --> Arch["Architecture Reviewer"]
    Spawn --> Perf["Performance Reviewer"]
    Spawn --> Qual["Quality Reviewer"]
    Sec --> Compile[Compile all findings]
    Arch --> Compile
    Perf --> Compile
    Qual --> Compile
    Compile --> Debate["Cross-examine: send all findings\nto each reviewer for challenge"]
    Debate --> Changed{"Positions changed\nAND round < max?"}
    Changed -- "Yes: changes + rounds left" --> Debate
    Changed -- "No: converged or round = max" --> Classify["Confidence Classification"]
    Classify --> ConvergeFlag{--converge flag?}
    ConvergeFlag -- No --> Output[Output debate report]
    ConvergeFlag -- Yes --> CritLeft{Critical findings remain?}
    CritLeft -- No --> Output
    CritLeft -- Yes --> Approve{User approves fixes?}
    Approve -- No --> Output
    Approve -- Yes --> Fix[Apply fixes]
    Fix --> CycleCheck{"Cycle < 3?"}
    CycleCheck -- Yes --> Read
    CycleCheck -- No --> Output
    Output --> Done([Done])

Key Decisions

Reviewer perspectives — each runs as a separate subagent for true independence:

• Security: OWASP Top 10, auth/authz gaps, input validation, data exposure, dependency risks
• Architecture: Design patterns, separation of concerns, coupling/cohesion, API contracts, backwards compatibility
• Performance: Algorithmic complexity, memory allocation, I/O efficiency, caching, DB query patterns
• Quality: Readability, error handling, test coverage, documentation, naming conventions

Finding format: Each finding must include severity (critical/warning/suggestion), confidence (HIGH/MEDIUM/LOW), file:line, description, fix suggestion.

Debate protocol: Each reviewer receives all other reviewers' findings and must label each response:

• CHALLENGED: Disagree with finding (explain why)
• CONCEDED: Missed this, other reviewer is right
• DEFENDED: Stand by own finding despite silence
• RAISED: New finding revealed by cross-examination

Confidence classification:

• HIGH: Unchallenged by all reviewers, or survived a challenge
• MEDIUM: Majority of reviewers agreed (e.g., 3/4, 4/5)
• LOW: Split opinion (e.g., 2-2, 2-3) — report both perspectives

Conditional reviewers: If changed files include HTML/CSS, add A11y reviewer. If DB migrations, add DB reviewer.

Output Format

## Adversarial Debate — {timestamp}

**Files**: {paths}
**Reviewers**: Security, Architecture, Performance, Quality
**Rounds**: {N} | **Converged**: {yes/no}

### HIGH Confidence Findings (unchallenged or survived debate)
1. [{severity}] {file}:{line} — {description}
   - Found by: {reviewer(s)}
   - Fix: {suggestion}

### MEDIUM Confidence Findings (majority agreed)
1. [{severity}] {file}:{line} — {description}
   - Agreed: {reviewers} | Disagreed: {reviewer}

### LOW Confidence Findings (split opinion)
1. [{severity}] {file}:{line} — {description}
   - For: {reviewer perspective}
   - Against: {reviewer perspective}

### Debate Log
- Round 1: {N} findings raised, {N} challenged, {N} conceded
- Round 2: {N} positions changed, {converged/continued}

### False Positives Caught
1. {finding} — challenged by {reviewer}, reason: {why it's fine}

**Summary**: {N} HIGH / {N} MEDIUM / {N} LOW findings across {N} files

Token Cost

• Standard (2 rounds): ~15,000-25,000 tokens
• With convergence (3 cycles): ~30,000-50,000 tokens
• Always-on: 0 (loaded on demand only)

Notes

• Each reviewer runs as a separate subagent for true independence
• Debate rounds prevent rubber-stamping — reviewers must actively engage with other findings