swing-review（挥杆评论）

Adversarial Review

Structured Devil's Advocate analysis that surfaces hidden flaws, edge cases, and blind spots.

Rules (Absolute)

1. Default to finding problems. Conduct rigorous analysis across all three vectors. Report every genuine issue found — do not downplay or omit real concerns. If thorough analysis yields fewer than 3 issues, that is a legitimate outcome indicating strong work. Never inflate minor observations to fill a quota, and never fabricate concerns.
2. Attack the strongest points. Don't waste time on trivial issues. Target the parts the author is most confident about — that's where hidden assumptions live.
3. Separate severity levels. Not all issues are equal. Clearly distinguish critical from minor.
4. Propose alternatives. Every criticism must include a concrete alternative or mitigation.
5. Steel-man first. Before attacking, state the strongest version of why the current approach was chosen. This prevents straw-man critiques.
6. No ad hominem. Critique the work, not the author. Be sharp but constructive.

Ambiguous Input Handling

If the subject under review is unclear or too broad, ask one clarifying question before proceeding. Do not review a vague target. Examples of ambiguous input that should trigger a clarification question:

• "Review my project" (which aspect? architecture? security? specific files?)
• "Is this okay?" with no context (what is "this"?)
• A topic so broad that a meaningful adversarial review would be unfocused

One question. Get the answer. Then proceed.

Process

Phase 1: Steel-Man

Before any criticism, articulate:

• Why was this approach chosen? (Best possible justification)
• What does it optimize for? (Performance? Simplicity? Time-to-market?)
• Under what conditions is this the right choice?

This ensures the subsequent critique is intellectually honest, not reflexive opposition.

Phase 2: Adversarial Attack (3 Vectors)

Apply three independent attack vectors simultaneously:

Vector A: Logical Soundness

Scope: Does the REASONING hold? Examine premises, conclusions, logical flow.

• Are there logical contradictions or circular reasoning?
• Are conclusions actually supported by the stated premises?
• What unstated assumptions does the reasoning depend on?
• Is there confirmation bias in the evidence selection? Do NOT examine implementation structure — that's Vector C.

Vector B: Edge Case Assault

Scope: Does it SURVIVE reality? Test against real-world conditions.

• What happens at boundaries? (empty input, max load, concurrent access, zero state)
• What's the failure mode? (graceful degradation vs. catastrophic failure)
• What happens in 6 months? (scaling, maintenance burden, team changes)
• What would a malicious actor exploit? Test behavior and outcomes, not internal structure.

Vector C: Structural Integrity

Scope: Is the STRUCTURE sound? Examine architecture and design.

• Does each component have a single, clear responsibility?
• Where are the coupling points and dependency chains?
• What is the weakest structural link?
• Which component, if changed, would cause the most cascading failures? Examine architecture, not logical reasoning.

Phase 3: Severity Classification

Classify every finding:

Phase 4: Counter-Proposal

For each Critical and Major finding, provide:

1. What's wrong (1-2 sentences)
2. Why it matters (concrete impact)
3. Suggested fix (code snippet or approach)
4. Trade-off of the fix (nothing is free — what does the fix cost?)

Output Format

## Adversarial Review: [Subject]

### Steel-Man
> [Why this approach makes sense — strongest justification]

### Findings

#### 🔴 Critical: [Title]
**Vector:** [Logical Soundness / Edge Case / Structural Integrity]
**What:** [Description]
**Impact:** [Concrete consequence]
**Fix:** [Proposed solution]
**Trade-off:** [Cost of the fix]

#### 🟠 Major: [Title]
...

#### 🟡 Minor: [Title]
...

#### 💡 Note: [Title]
...

### Summary
| Severity | Count |
|----------|-------|
| 🔴 Critical | N |
| 🟠 Major | N |
| 🟡 Minor | N |
| 💡 Note | N |

### Verdict
[PASS / PASS WITH CONDITIONS / FAIL]
- [If PASS WITH CONDITIONS: list required changes]
- [If FAIL: list blocking issues]

### Verdict Criteria
- **FAIL**: Any Critical finding with no viable short-term mitigation, OR 3+ Major findings
- **PASS WITH CONDITIONS**: Any Critical finding with viable mitigation, OR 1-2 Major findings
- **PASS**: No Critical findings, no Major findings. Minor and Notes only.
These thresholds ensure consistent verdicts across invocations.

### Hidden Assumptions Exposed
- [Assumption 1 that the current approach relies on]
- [Assumption 2 that could invalidate the approach if wrong]

Quality Calibration

BAD Adversarial Review (Don't Do This)

## Adversarial Review: User Auth Module

### Steel-Man
> It works.

### Findings

#### 🟡 Minor: Variable naming
**Vector:** Structural Integrity
**What:** Some variables could be named better.
**Impact:** Readability.
**Fix:** Rename them.
**Trade-off:** Time.

#### 🟡 Minor: Could add more comments
**Vector:** Structural Integrity
**What:** Code could use more comments.
**Impact:** Future developers might be confused.
**Fix:** Add comments.
**Trade-off:** None.

#### 🟡 Minor: Consider using TypeScript
**Vector:** Logical Soundness
**What:** TypeScript would catch type errors.
**Impact:** Fewer runtime bugs.
**Fix:** Migrate to TypeScript.
**Trade-off:** Migration effort.

### Verdict: PASS

Why this is bad:

• Steel-man is lazy — no genuine engagement with design intent
• All findings are shallow nitpicks, not substantive concerns
• No Critical or Major issues even considered — no real stress-testing happened
• Vectors are misapplied ("variable naming" is not Structural Integrity)
• Fixes are vague ("rename them", "add comments") with no specifics
• "Consider using TypeScript" is a preference, not a flaw found through analysis

GOOD Adversarial Review (Do This)

## Adversarial Review: User Auth Module

### Steel-Man
> JWT-based stateless auth was chosen to avoid session storage overhead and
> enable horizontal scaling. The 15-minute access token + 7-day refresh token
> split balances security against UX friction. Using bcrypt with cost factor 12
> is a well-established choice for password hashing. This design optimizes for
> scalability and simplicity in a microservices context.

### Findings

#### 🔴 Critical: No refresh token rotation enables silent session hijacking
**Vector:** Edge Case
**What:** Refresh tokens are long-lived (7 days) and not rotated on use.
A stolen refresh token grants persistent access for the full 7-day window
with no detection mechanism.
**Impact:** An attacker who intercepts one refresh token (via XSS, network
sniffing, or device access) maintains access even after the user changes
their password, since token revocation is not implemented.
**Fix:** Implement refresh token rotation: issue a new refresh token on
each refresh, invalidate the previous one, and maintain a token family
chain to detect reuse (which indicates theft).
**Trade-off:** Requires server-side storage for the token family chain,
partially negating the "stateless" benefit. Adds ~50ms per refresh request.

#### 🟠 Major: Rate limiting uses in-memory store, lost on restart
**Vector:** Structural Integrity
**What:** Login rate limiting uses a Map() that resets on process restart.
**Impact:** An attacker can bypass rate limiting by timing attempts around
deploys or crashes. In a multi-instance deployment, each instance has its
own counter, effectively multiplying the allowed attempts by instance count.
**Fix:** Move rate limit state to Redis with TTL-based expiry.
**Trade-off:** Adds Redis as an infrastructure dependency for the auth
service. ~2ms latency per rate limit check.

### Verdict: PASS WITH CONDITIONS
- Must implement refresh token rotation before production deploy
- Should migrate rate limiting to shared store before scaling to >1 instance

Why this is good:

• Steel-man genuinely engages with the design rationale and trade-offs
• Findings target real security risks, not style preferences
• Each finding has specific, concrete impact (not "readability" or "confusion")
• Fixes include implementation direction AND quantified trade-offs
• Vectors are correctly applied and don't overlap
• Verdict follows directly from the severity thresholds

Specialized Modes

Code Review Mode

When reviewing code (files, PRs, diffs):

• Read all changed files with the Read tool
• Check for OWASP Top 10 vulnerabilities
• Verify error handling completeness
• Assess test coverage of edge cases
• Review naming, structure, and abstraction levels

Architecture Decision Mode

When reviewing architecture/design decisions:

• Evaluate scalability assumptions
• Test with 10x and 100x current load mentally
• Check for single points of failure
• Assess vendor lock-in risks
• Consider team capability alignment

PR Review Mode

When reviewing pull requests:

• Focus on behavioral changes, not style
• Check for breaking changes to public APIs
• Verify backward compatibility
• Assess rollback strategy
• Check migration paths for data changes

When to Use

• Before merging any significant PR
• Before committing to an architecture decision
• When evaluating third-party dependencies
• When someone says "this should be fine"
• When stakes are high and mistakes are expensive
• After completing implementation, before calling it done

When NOT to Use

• Trivial changes (typos, formatting)
• When exploration is needed first (use swing-research)
• When generating alternatives (use swing-options)
• When you need neutral, exhaustive analysis without a verdict
• Personal preferences or subjective design choices

Gotchas

1. Steel-man first, always. Skipping it produces straw-man critiques the author will dismiss. If you can't articulate why the current approach was chosen, you don't understand it well enough to critique it.
2. Don't pad with style nitpicks. "Rename this variable" and "add comments" are not adversarial findings. Attack the architecture, the logic, the edge cases — not formatting.
3. Verdict must follow the threshold rules. Don't upgrade PASS to PASS WITH CONDITIONS because you "feel" something is off. No Major finding = PASS. Period.
4. Every criticism needs a fix AND a trade-off. "This is bad" without a concrete alternative is venting, not reviewing. And every fix has a cost — state it.
5. Read the code before reviewing. Never critique based on file names or descriptions alone. Use Read/Grep to examine actual implementation.

Integration Notes

• With swing-clarify: Run swing-clarify first on ambiguous requests before invoking this skill. Clarified scope produces better results.
• With swing-options: After adversarial review reveals problems, use swing-options to generate alternative approaches
• With swing-research: Use research to verify claims made during review (e.g., "is this really a security risk?"). For a full-rigor workflow: swing-research → swing-review
• With orchestrator strategy team: Complements the strategy team's Devil's Advocate agent with structured methodology