second-opinions

Compare original and translation side by side

🇺🇸

Original

English

🇨🇳

Translation

Chinese

If you are Codex or Gemini CLI, stop and do not use this skill.

如果你是Codex或Gemini CLI，请停止操作，不要使用此skill。

Second Opinions

第二意见评审

Run independent reviews using OpenAI Codex (via MCP) and/or Google Gemini (via CLI) for second opinions from different model families.

Modes:

both

(default),

codex

gemini

-- selected via

AskUserQuestion

使用OpenAI Codex（通过MCP）和/或Google Gemini（通过CLI）执行独立评审，获取来自不同模型家族的第二意见。

模式：

both

（默认）、

codex

、

gemini

——通过

AskUserQuestion

选择。

When To Use

使用场景

Before opening a PR or merging a branch.
Before committing significant changes.
Plan or architecture review from multiple perspectives.
Security, performance, or correctness audit.

发起PR或合并分支之前。
提交重大代码变更之前。
从多视角进行方案或架构评审。
安全性、性能或正确性审计。

When Not To Use

不适用场景

Trivial changes you can verify directly.
When the user asked for your own judgment only.
Do not send secrets, credentials, or tokens in prompts.

可直接验证的微小变更。
用户仅要求你提供自身判断时。
不要在提示词中发送密钥、凭证或令牌。

Workflow

工作流程

1. Gather Context

1. 收集上下文

Use

AskUserQuestion

to collect review parameters in one call. Combine applicable questions (max 4).

Tool (always ask):

header: "Review tool"
question: "Which tool should run the review?"
options:
  - "Both Codex and Gemini (Recommended)" -> run both in parallel
  - "Codex only"                          -> mcp__codex__codex
  - "Gemini only"                         -> gemini CLI

Scope (always ask):

header: "Review scope"
question: "What should be reviewed?"
options:
  - "Uncommitted changes"  -> git diff HEAD + untracked files
  - "Branch diff vs main"  -> git diff <branch>...HEAD
  - "Specific commit"      -> follow up for SHA

Focus (always ask):

header: "Review focus"
question: "Any specific focus areas?"
options:
  - "General review"
  - "Security & auth"
  - "Performance"
  - "Error handling"

使用

AskUserQuestion

在一次调用中收集评审参数。可组合相关问题（最多4个）。

工具（必问）:

header: "Review tool"
question: "Which tool should run the review?"
options:
  - "Both Codex and Gemini (Recommended)" -> run both in parallel
  - "Codex only"                          -> mcp__codex__codex
  - "Gemini only"                         -> gemini CLI

评审范围（必问）:

header: "Review scope"
question: "What should be reviewed?"
options:
  - "Uncommitted changes"  -> git diff HEAD + untracked files
  - "Branch diff vs main"  -> git diff <branch>...HEAD
  - "Specific commit"      -> follow up for SHA

评审重点（必问）:

header: "Review focus"
question: "Any specific focus areas?"
options:
  - "General review"
  - "Security & auth"
  - "Performance"
  - "Error handling"

2. Preview and Validate

2. 预览与验证

Auto-detect default branch for branch diffs:

bash

git symbolic-ref refs/remotes/origin/HEAD 2>/dev/null | sed 's@^refs/remotes/origin/@@' || echo main

Show diff stats for the selected scope. If empty, stop and tell the user. If >2000 lines changed, warn and ask whether to proceed or narrow scope.

自动检测分支对比的默认分支：

bash

git symbolic-ref refs/remotes/origin/HEAD 2>/dev/null | sed 's@^refs/remotes/origin/@@' || echo main

展示所选范围的差异统计信息。如果无差异，停止操作并告知用户。如果变更行数超过2000行，发出警告并询问用户是否继续或缩小范围。

3. Dispatch Reviews

3. 发起评审

Project context: If a

CLAUDE.md

exists in the repo root, always include it in the review prompt so reviewers check against project conventions.

Run the selected tool(s). When running both, issue

mcp__codex__codex

and

Bash(gemini ...)

calls in a single message for parallel execution. Both are read-only operations with no shared state.

See references/codex.md for Codex MCP patterns. See references/gemini.md for Gemini CLI patterns.

Do not pre-check tool availability. If a tool fails, report the install instructions from the error handling table and run only the available tool.

项目上下文： 如果仓库根目录存在

CLAUDE.md

，请务必将其包含在评审提示词中，以便评审者对照项目规范进行检查。

运行选定的工具。如果同时运行两个工具，请在同一条消息中调用

mcp__codex__codex

和

Bash(gemini ...)

以实现并行执行。两者均为只读操作，无共享状态。

关于Codex MCP的使用模式，请参考references/codex.md。关于Gemini CLI的使用模式，请参考references/gemini.md。

无需预先检查工具可用性。如果某工具运行失败，请根据错误处理表格中的说明告知用户安装方法，仅运行可用的工具。

4. Present Results

4. 展示结果

When running both, present with clear headers:

undefined

如果同时运行两个工具，请使用清晰的标题展示结果：

undefined

Codex Review

Gemini Review

Summary

Where the two reviews agree and differ. Prioritized action items.


When running a single tool, present its findings directly with your assessment of which findings are valid vs uncertain.

Where the two reviews agree and differ. Prioritized action items.


如果仅运行单个工具，请直接展示其评审结果，并对结果的有效性或不确定性给出你的评估。

Error Handling

错误处理

Error	Action
`mcp__codex__codex` unavailable	Run Gemini only,
`gemini: command not found`	Run Codex only.
Both unavailable	Stop and inform the user
Gemini extension missing	Install: `gemini extensions install <github-url>` (see gemini reference)
Empty diff	Stop: no changes to review
Timeout	Suggest narrowing the scope
One tool fails	Present the other's results, note the failure

错误	处理操作
`mcp__codex__codex` 不可用	仅运行Gemini
`gemini: command not found`	仅运行Codex
两者均不可用	停止操作并告知用户
缺少Gemini扩展	安装： `gemini extensions install <github-url>` （参考Gemini相关文档）
差异内容为空	停止操作：无变更可评审
超时	建议缩小评审范围
其中一个工具运行失败	展示另一个工具的结果，并说明该工具运行失败

Examples

示例

Both tools (default):

User: /second-opinions
Agent: [asks 3 questions: tool, scope, focus]
User: picks "Both", "Branch diff", "Security"
Agent: [detects default branch = main, shows diff --stat]
Agent: [auto-includes CLAUDE.md, dispatches mcp__codex__codex + gemini in parallel]
Agent: [presents both reviews, highlights agreements/differences]

Single tool:

User: /second-opinions
Agent: [asks 3 questions: tool, scope, focus]
User: picks "Codex only", "Uncommitted", "General"
Agent: [shows diff --stat]
Agent: [calls mcp__codex__codex with review prompt]
Agent: [presents findings with assessment]

Large diff warning:

User: /second-opinions
Agent: [asks questions] -> user picks "Both", "Uncommitted", "General"
Agent: [shows diff --stat: 45 files, +3200 -890]
Agent: "Large diff (3200+ lines). Proceed, or narrow the scope?"
User: "proceed"
Agent: [runs both reviews]

双工具模式（默认）:

User: /second-opinions
Agent: [询问3个问题：工具、范围、重点]
User: 选择“双工具”、“分支差异”、“安全性”
Agent: [检测到默认分支为main，展示diff --stat结果]
Agent: [自动包含CLAUDE.md，并行调用mcp__codex__codex和gemini]
Agent: [展示两份评审结果，突出共识与差异]

单工具模式:

User: /second-opinions
Agent: [询问3个问题：工具、范围、重点]
User: 选择“仅Codex”、“未提交变更”、“通用评审”
Agent: [展示diff --stat结果]
Agent: [调用mcp__codex__codex并传入评审提示词]
Agent: [展示评审结果并给出评估]

大差异警告:

User: /second-opinions
Agent: [询问问题] -> 用户选择“双工具”、“未提交变更”、“通用评审”
Agent: [展示diff --stat结果：45个文件，+3200 -890]
Agent: “差异内容过大（超过3200行）。是否继续，还是缩小评审范围？”
User: “继续”
Agent: [运行双工具评审]