Cross-Model Verification

Overview

A single model shares blind spots with itself—fresh context from a different architecture catches them. This skill orchestrates the second-opinion workflow: collect the artifact, generate an adversarial prompt, dispatch to a second model, diff the results, and report findings classified by severity.

The Process

Security Gate → Collect → Adversarial Prompt → Dispatch → Diff → Report
      │            │            │                  │         │        │
      ▼            ▼            ▼                  ▼         ▼        ▼
  Redaction     Artifact   "Find issues"      Which      Compare   Severity
  decision      isolation  NOT "is it good"   model?     outputs   labels

Step 0: Security Gate

Expected Output: safe-to-dispatch / redacted / local-only / abort

Before sending anything to another model, classify and scrub the artifact.

Artifact contains secrets, credentials, production data, customer data, or private repo context?
  │
  ├── No → safe-to-dispatch
  │
  ├── Yes, can be redacted without losing review value
  │     ├── Replace values with placeholders: <REDACTED_API_KEY>, <CUSTOMER_ID>, <INTERNAL_HOST>
  │     └── Output: redacted + list of redaction classes
  │
  ├── Yes, cannot be safely redacted
  │     ├── External model or web UI → ask user before dispatch
  │     └── No approval → local-only degraded review
  │
  └── Security auth / production migration + unredacted sensitive data + no second model under approved boundary
        └── abort and escalate to human review

Run a quick scan before dispatch:

grep -Ei "sk-|AKIA|AIza|xox|password|secret|private_key|token|BEGIN .*PRIVATE KEY" <artifact-file>
# Expected: no matches, or every match is intentionally redacted before dispatch

Step 1: Collect

Expected Output: artifact + contract, smallest reviewable unit, no prior CLAIM

Isolate the smallest reviewable unit. Strip your reasoning—handing over conclusions biases the reviewer toward agreement.

What to review?
  │
  ├── Code diff / PR → paste the diff
  ├── Architecture decision → 3-5 sentence proposal + constraints
  └── Assertion → the claim + supporting evidence (distinct from your CLAIM)

Rule: The unit must be small enough that a reviewer can hold it in one read. If it's a 500-line PR, decompose first.

Step 2: Adversarial Prompt

Expected Output: prompt contains ARTIFACT + CONTRACT, asks to find issues, omits author's CLAIM

Framing decides the answer. Always "find issues", never "is it good".

Adversarial review. Find what is wrong with this artifact.
Assume the author is overconfident. Look for:
- Unstated assumptions
- Edge cases not handled
- Hidden coupling or shared state
- Ways the contract could be violated
- Existing conventions this might break
- Failure modes under unexpected input

Do NOT validate. Do NOT summarize. Find issues, or state
explicitly that you cannot find any after thorough examination.

ARTIFACT: <paste artifact>
CONTRACT: <paste contract or constraints>

Do NOT pass your CLAIM. The reviewer must independently determine whether the artifact satisfies the contract.

Step 3: Dispatch

Expected Output: model choice + invocation method + security decision recorded

Which model?
  │
  ├── CLI tool (gemini, codex) →
  │     ├── which gemini / which codex
  │     ├── gemini --version / codex --version
  │     ├── Confirm invocation with user
  │     └── Pipe via stdin (never inline -p with untrusted artifact)
  │
  ├── Same session model only → Degraded self-review
  │     └── 同模型不构成 cross-model，必须标记 degraded
  │
  ├── Web interface → Manual (user pastes)
  │
  └── None available → Degraded self-review
        └── Write ARTIFACT + CONTRACT as fresh self-prompt
        └── Walk Steps 1-5 with hard mental separator
        └── Flag result as degraded

Shell escaping (load-bearing): Artifacts contain $(...), backticks, quotes. Inline arguments truncate or execute embedded shell.

# Write to temp file, pipe via stdin
echo "$PROMPT" > /tmp/cross-model-prompt.md
gemini --approval-mode plan -p "" < /tmp/cross-model-prompt.md

# Read-only sandbox for codex (prevents artifact injection from executing)
codex exec --sandbox read-only -C <repo-path> - < /tmp/cross-model-prompt.md

Step 4: Diff

Expected Output: findings classified as valid / trade-off / noise / contract-misread

The reviewer's output is data, not verdict. You are still the orchestrator.

For each finding:
  ├── Contract misread? → Fix contract, re-classify
  ├── Valid + actionable? → Change it, re-loop
  ├── Valid trade-off? → Document explicitly
  └── Noise? → Note it, move on

Step 5: Report

Expected Output: severity-classified report with approve / request changes / needs discussion

═══ Cross-Model Verification ═══
  模型 A: <name>
  模型 B: <name>

  发现:
  [Critical] <finding> — 必须修复
  [Required] <finding> — 建议修复
  [Optional] <finding> — 可改进
  [FYI] <finding> — 供参考

  结论: approve / request changes / needs discussion

Failure / Degraded Strategy

Second model unavailable?
  │
  ├── CLI tool missing → Ask user to install, or use degraded
  │
  ├── Artifact too large → Decompose and review in chunks
  │
  ├── Timeout / no response → Re-dispatch with same prompt (retry once)
  │
  └── None available → Degraded self-review
        ├── Write ARTIFACT + CONTRACT as fresh self-prompt
        ├── Walk Steps 1-5 with hard mental separator
        ├── Flag result explicitly as "degraded" in output
        └── Never claim "cross-verified" if only self-reviewed

Degraded self-review protocol:

1. Close any open tabs with the artifact—you must come back cold.
2. Rewrite the adversarial prompt as a self-prompt, strip any "I found X" from it.
3. Set a 5-minute timer. If you finish early, you didn't try hard enough.
4. In the output, prefix every finding with [degraded] to remind downstream consumers this wasn't cross-validated.

When to abort rather than degrade:

• High-stakes change (production migration, security auth) AND no second model AND the artifact is > 500 lines → Escalate to human review, don't fake verification.
• Artifact contains unredacted secrets or regulated data and no approved second-model boundary exists → abort, redact, or use local-only degraded review.

Common Pitfalls

When to Use

Verification Checklist

• Artifact 已隔离为最小可审查单元
• Artifact 已完成敏感信息扫描；外发内容已 redacted 或获得明确许可
• Adversarial prompt 使用"find issues"而非"is it good"
• 未将个人判断或 CLAIM 传入 prompt
• 通过 stdin 传递 artifact（不使用 -p 内联参数）
• 所有发现已分类（Critical/Required/Optional/FYI）
• 结论明确：approve / request changes / needs discussion

Interaction with Other Skills

• skill-review-workflow: skill-review 的审查结果可送入本 skill 做二次验证。skill-review 发现 Critical 问题时自动触发。
• agent-security-guard: 安全相关发现（API key、shell 注入）优先用 agent-security-guard 专项检测，本 skill 负责一般性逻辑审查。