dual-agent: 双智能体安全协同工作流

核心理念

WorkBuddy (WB) 负责执行、校验和最终决策；Claude Code (CC) 或其他外部推理助手只作为只读顾问，提供架构分析、代码审查、风险识别和第二意见。

这个发布版是 advisory-only：不内置自动安装脚本、不自动调用外部 CLI、不运行 bundled executable、不读取或传递私密认证材料、不自动执行顾问返回的步骤。

Roles

• WB: Executor and controller. Inspect context, make judgments, apply edits only when appropriate, and explain risk to the user.
• CC / external advisor: Read-only reasoning advisor. Provide analysis, critique, recommendations, assumptions, and risk notes only.

When to Use

Use this skill when any condition applies:

• Architecture decisions affect three or more files or modules.
• User asks for a second opinion, independent review, or adversarial critique.
• Bug diagnosis remains unresolved after initial WB investigation.
• Security, permissions, data integrity, finance controls, or compliance risks are involved.
• Performance optimization needs trade-off analysis.
• A design decision needs structured challenge before implementation.

Do not use this skill for simple edits, formatting, renaming, routine build/test execution, ordinary git operations, or tasks the user explicitly asks WB to complete directly.

Safety Rules

Follow these rules strictly:

1. Do not auto-install Claude Code, CC Switch, or any third-party software.
2. Do not execute advisor-provided commands automatically.
3. Do not give the advisor file-write access.
4. Do not send secrets, credentials, tokens, private keys, or unnecessary personal data to the advisor.
5. Share only minimal, relevant context.
6. Validate every advisor recommendation before implementation.
7. Ask the user before applying material or risky changes.
8. If WB and the advisor disagree, present both views and make a reasoned recommendation.

Workflow Modes

Semi-Auto Advisory Mode (Default)

WB suggests advisor involvement and asks for confirmation before sharing context.

WB: 建议进行外部顾问审查。
原因: [why a second opinion is valuable]
将共享的信息: [brief summary of context]
将询问的问题: [focused review question]
是否确认继续?

After receiving advisor input, WB reviews it and presents a decision plan. WB does not blindly execute advisor output.

WB-Only Mode

If the user declines advisor involvement, or if external advisor access is unavailable, WB continues solo and notes that the second-opinion layer was skipped.

Advisor Output Protocol

Ask the advisor to return:

1. Free-form analysis for human review.
2. Key risks and assumptions.
3. Recommended next steps.
4. Items that require verification before implementation.

Treat advisor output as advisory material only. WB must validate all recommendations.

Effort Guidance

When the user chooses to consult Claude Code or another reasoning advisor, prefer high-reasoning settings for complex work.

Recommended default for dual-agent review: xhigh, because this skill is reserved for complex tasks.

Sub-skill Integration

Grill

Use a design-review or grill-style workflow when the user discusses planning, architecture, workflow design, major refactoring, or unclear trade-offs.

Debug Session

Use a structured debugging workflow when the user reports bugs, errors, unexpected behavior, failing tests, or unresolved diagnosis.

Status and Reporting

After advisor review, report:

• WB assessment
• Advisor assessment
• Agreed findings
• Disagreements or uncertainty
• Recommended next action
• Risk controls and rollback considerations

Reference Files

Load references only when needed:

• references/cc-caller.md: advisory prompt templates and review protocol
• references/context-monitor.md: context usage display guidance
• references/cost-tracker.md: high-level cost-reporting guidance
• references/env-detector.md: safe environment detection and manual setup guidance

Constraints

• Advisor input is never authoritative by itself.
• WB remains accountable for final recommendations and actions.
• No auto-installation, silent installation, privileged installation, external download, or bundled executable runner is included in this skill.
• No bundled script is required for the published advisory workflow.