copilot-cli

Adversarial code review via Copilot CLI. Challenges Claude's implementation — design choices, assumptions, edge cases. Use after completing a feature or before declaring work done. GPT-5 acts as a skeptical reviewer finding reasons the change should not ship. <example> Context: Claude has just finished implementing a feature and wants adversarial review before committing. user: "实现完了，帮我做个对抗式审查" assistant: "我让 Copilot 以红军视角审查这些变更，找出不应该上线的理由。" <commentary> Adversarial review after implementation — Copilot challenges Claude's work from a different AI perspective. </commentary> </example> <example> Context: Claude wants a second opinion on a critical code change before pushing. user: "这个改动涉及支付逻辑，帮我仔细检查" assistant: "支付相关的改动需要严格审查，我用 Copilot 做对抗式审查，重点关注边界条件和失败模式。" <commentary> Critical code path — adversarial review with high effort to catch edge cases in payment logic. </commentary> </example> <example> Context: Proactive review during workflow Step 3.5 (boundary review). user: "TDD 实现完成了，跑一下边界审查" assistant: "我派 Copilot 做边界审查，GPT 特别擅长找 null、timeout、竞态等失败模式。" <commentary> Workflow integration — Copilot review as part of the standard development workflow boundary check. </commentary> </example>

Dispatch self-contained tasks to GitHub Copilot CLI (Claude Opus 4.6). Use for context-heavy but goal-clear work that would consume too much of Claude's context window — file scanning, edge-case analysis, boundary-condition review, batch exploration. Copilot charges per invocation so prompts must be fully self-contained and complete in one shot. <example> Context: Claude is working on a large project and needs to scan 20+ files for a specific pattern. user: "检查所有 API 端点的错误处理是否完整" assistant: "这个扫描任务涉及大量文件，我交给 Copilot CLI 来处理，节省上下文空间。" <commentary> Context-heavy scanning task with clear goal — ideal for Copilot delegation to save Claude's context. </commentary> </example> <example> Context: Claude wants a second opinion on edge cases in a piece of code. user: "帮我分析一下这段代码的边界条件" assistant: "我让 Copilot 从 Opus 的角度分析边界条件，GPT 特别擅长这类工作。" <commentary> GPT-5 excels at catching boundary conditions and error handling gaps — complementary to Claude's strengths. </commentary> </example> <example> Context: Claude needs to explore an unfamiliar codebase before making changes. user: "先了解一下这个项目的技术栈和架构" assistant: "我派 Copilot 去探索项目结构，这样不消耗我的上下文窗口。" <commentary> Exploratory task that produces a summary — Copilot reads many files and returns a concise result. </commentary> </example>

Internal helper for invoking Copilot CLI from Claude Code subagents — covers invocation template, model routing, context injection, and self-containment rules. Not user-invocable.