Code Advocate

English | 中文

AI-generated code must prove itself before entering your codebase.

The Problem

AI coding assistants generate code fast, but nobody validates it. Code review catches style issues, tests cover happy paths, but the dangerous stuff — hidden assumptions, race conditions, missing error handling, side-effect ordering — slips through because a single reviewer (human or AI) tends to agree with what's in front of them.

The Idea

Code Advocate subjects code to adversarial validation: two opposing roles argue against each other, forcing the truth out.

Role	Job
Defender	Constructs the best possible case for why the code is correct
Prosecutor	Actively searches for flaws, edge cases, hidden assumptions

The tension between these roles surfaces issues that neither would find alone.

Single-model review says: "Looks good to me." Adversarial review says: "The Defender claims this is safe, but the Prosecutor found that when userID is empty from the cron job caller, line 42 will panic."

How It Works

AI generates code
│
▼
┌──────────────────┐
│ Phase 1: Defense │ Defender builds a structured argument:
│ │ intent, correctness, assumptions, edge cases
└────────┬─────────┘
│
▼
┌──────────────────┐
│ Phase 2: Attack │ Prosecutor challenges every defense point:
│ │ counterexamples, violated assumptions, risks
└────────┬─────────┘
│
▼
┌──────────────────┐
│ Phase 3: Cross- │ Defender concedes or refutes each challenge
│ Examination │ with evidence. No second rebuttals.
└────────┬─────────┘
│
▼
┌──────────────────┐
│ Phase 4: Verdict │ PASS / CONDITIONAL PASS / REJECT
│ │ with specific actions for each issue
└──────────────────┘

Verdict Levels

Level	Meaning	Action
PASS	All prosecution points refuted	Commit with confidence
CONDITIONAL PASS	Some conceded/disputed points, none critical	Fix conceded items, review disputed, then commit
REJECT	Critical issues that could cause incidents	Do not commit. Fix and re-run

Example Output

For a simple GetUserInfo function, the verdict might look like:

ADMITTED:
✅ Empty userID validation — correct
✅ Nil user handling — correct
CONCEDED (must fix):
❌ Error message leaks DB details — replace err.Error() with generic message
❌ Missing error path logging — add log.InfoCtx calls
DISPUTED (developer decision):
⚠️ Email sanitization responsibility — verify API layer escaping
VERDICT: CONDITIONAL PASS

See examples/ for full verdict documents.

Installation

# Add as a Claude Code plugin
claude --plugin-dir /path/to/code-advocate

Or add to your project's .claude/plugins.json:

{
"plugins": ["github:happli-sys/code-advocate"]
}

Usage

Once installed, trigger Code Advocate by saying:

• "advocate this code"
• "put this code on trial"
• "review AI generated code"
• "cross-examine this function"

Structure

code-advocate/
├── SKILL.md # Core workflow definition
├── references/
│ ├── verdict-format.md # Verdict template and size guidelines
│ ├── challenge-patterns.md # 8 prosecution attack patterns
│ └── defense-patterns.md # 7 defense strategies with evidence hierarchy
├── examples/
│ ├── simple-function-verdict.md # Verdict for a simple Go function
│ └── complex-service-verdict.md # Verdict for a multi-service interaction
└── scripts/
└── collect-context.sh # Collect git/AST context for validation

Prosecution Attack Patterns

The Prosecutor uses 8 systematic patterns to find weaknesses:

1. Nil/Zero Value Attack — Trace what happens when every value is nil/empty/zero
2. Boundary Condition Attack — Off-by-one, overflow, empty results, max size
3. Concurrency Attack — Shared mutable state, lock granularity, race conditions
4. Error Handling Attack — Partial failures, silent failures, cleanup on error
5. Dependency Attack — What happens when every external call fails
6. Assumption Violation Attack — Find implicit assumptions and violate them
7. Side Effect Attack — Ordering, idempotency, compensation
8. Convention Violation Attack — Compare against project-specific rules

See references/challenge-patterns.md for details.

Why This Is Different

	Code Review	LLM Code Review	Code Advocate
Perspective	One reviewer	One model	Two opposing roles
Bias	Agreement bias	Self-agreement bias	Deliberate opposition
Output	"Looks good" or comments	"No issues found"	Structured verdict with evidence
Coverage	What reviewer thinks of	What model thinks of	What adversarial tension reveals

License

MIT