菜单
Run an adversarial review of code or a diff through independent reviewers, then lead-judge consensus, disagreements, and false positives. Use when the user asks to interrogate, challenge, stress-test, find blind spots, perform multi-model review, or when `henry-mode` needs adversarial judgment of a risky change.
默认使用会先检查来源的 Prompt；你也可以切换为直接命令，或下载本地副本。
决定是否安装前，请先阅读 SKILL.md，以及 SkillsMP 当前展示的配套文件。
用 Codex 或 Claude 帮你安装 复制这段 Prompt，粘贴到 Codex、Claude 或其他助手里，让它检查 Skill 页面并帮你完成安装。
直接命令不会经过审查 Prompt；运行前请先检查来源。
npx skills add https://github.com/caezium/skills --skill interrogate命令会保持在同一行。复制前请横向滚动并检查完整内容。
想先保存到本地？可下载 SkillsMP 当前能够提供的文件。
正在显示 SKILL.md
| name | interrogate |
| description | Run an adversarial review of code or a diff through independent reviewers, then lead-judge consensus, disagreements, and false positives. Use when the user asks to interrogate, challenge, stress-test, find blind spots, perform multi-model review, or when `henry-mode` needs adversarial judgment of a risky change. |
Give independent reviewers the same intent, code, and rubric. Diversity comes from independent reasoning and, where genuinely available, different model families—not invented personas or model aliases. The deliverable is a lead-reviewed verdict; never auto-apply findings.
| Harness | Parallel mechanism | Model selection | Fallback |
|---|---|---|---|
| Cursor | Available Task/subagent calls | Valid models exposed by the harness | Independent inherited-model reviewers or sequential passes |
| Claude Code | Available Agent/Task calls | Valid models exposed by the harness | Independent inherited-model reviewers or sequential passes |
| Codex / T3 | spawn_agent, drained with wait_agent | Valid overrides only when offered | Run waves within slot limits; inherit the parent model |
| No delegation | None | None | Apply the rubric through separate sequential passes and label the result self-reviewed |
Only delegate when current user and system policy permit it. If true model diversity is unavailable, disclose that; never pretend reviewer labels are different models.
Use the exact files or diff the user names. Otherwise inspect the full feature-branch diff against its appropriate merge base. Include only surrounding files needed to trace the changed behavior; give every reviewer the same package.
Write one paragraph describing what the change is meant to accomplish, derived from the user's request, commit/PR context, and code. Ask only if ambiguity would materially change the review. Reviewers challenge execution, not the chosen product goal.
Read and fill these references:
Launch as many reviewers as the user requested or the harness can support usefully; default to three independent passes. Each reviewer is read-only and receives the identical filled prompt. Use model overrides only from valid harness-advertised choices. Require structured findings with severity, location, evidence, and an optional concrete suggestion.
Drain every reviewer. Merge duplicate findings, identify independent agreement, retain lone findings at lower confidence, and surface explicit disagreement. Inspect relevant code yourself before accepting a claim.
Read references/lead-judgment.md, then categorize every finding:
Name the reviewers that raised each finding and give a one-line rationale. Consensus is strong evidence, not a substitute for tracing the code.
IntentReviewers—configuration, model-diversity status, and finding countAct OnConsiderNotedDismissedAgreement MapCoverage GapsDo not edit code or create review comments unless the user separately authorizes those actions.