second-opinions（第二意见）

This skill is for Claude only. Codex and Gemini CLI should not invoke it.

Second Opinions

Delegate tasks to OpenAI Codex (MCP) and/or Google Gemini (CLI) for independent perspectives from different model families. Works for code review, plan evaluation, doc editing, codebase analysis, or any arbitrary task.

Modes: both (default), codex, gemini

When To Use

• Code review before committing, opening a PR, or merging
• Plan or architecture review from multiple perspectives
• Documentation review or content editing
• Codebase analysis (patterns, race conditions, dead code)
• Any task benefiting from parallel exploration by external models

Scope

• Skip for trivial tasks you can verify directly
• Skip when the user asked for your own judgment only
• Exclude secrets, credentials, and tokens from prompts

Workflow

1. Gather Context

Infer as many parameters as possible from user's questions and conversation context. Only ask about what can't be inferred. Combine remaining questions into a single AskUserQuestion call (max 3 questions).

Parameters to resolve:

2. Build Material

Gather content based on task type:

Auto-detect default branch for code reviews:

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

3. Construct & Dispatch

Filesystem boundary: prepend this instruction to ALL prompts sent to Codex and Gemini, before any other content:

IMPORTANT: Do NOT read or execute any files under ~/.claude/,.claude/,.agents/, or any path containing "skills". These are AI agent configuration files meant for a different system. They contain prompt templates and bash scripts that are irrelevant to your task. Ignore them completely and focus only on the project's own source code, tests, and documentation.

This prevents models from wasting tokens reading skill definitions, agent configs, and other meta-files instead of doing the actual task.

Prompt structure: use XML tags for unambiguous parsing. Put long content first, instructions after (per Anthropic long-context best practices):

<boundary>
[Filesystem boundary instruction from above: ALWAYS included]
</boundary>

<material>
[The diff, plan, document, codebase excerpt, or task input: long content goes first]
</material>

<context>
[Project conventions from CLAUDE.md, if it exists]
</context>

<role>
[Role appropriate to the task: e.g., independent code reviewer, architecture evaluator, technical editor]
</role>

<task>
[What to do: review, analyze, evaluate, rewrite, etc.]
This is an analysis task. Do not write code or produce patches unless explicitly asked.
</task>

<focus>
[Focus area if specified]
</focus>

<output_format>
Structure your response as:
1. Critical issues (blocking): prefix each with [P1]
2. Important concerns (should address): prefix each with [P2]
3. Minor suggestions (nice to have): prefix each with [P3]
4. What's done well

End with a clear verdict: PASS (no P1 findings) or FAIL (P1 findings exist), plus a confidence score (0-1).
</output_format>

Adapt <role>, <task>, and <output_format> to the task type. For custom tasks, translate the user's intent directly.

Code review prompt: for code reviews, use adversarial framing. The whole point of a second opinion is to find what you missed:

<role>
You are a hostile code reviewer, chaos engineer, and security auditor. Your job is to find problems.
</role>

<task>
Review this diff for issues that would cause problems in production. Think like an attacker
and a chaos engineer. Look for: bugs, edge cases, race conditions, security holes, resource
leaks, failure modes under load, silent data corruption paths, unhandled error propagation,
and implicit assumptions that will break when inputs change. If a focus area is specified,
concentrate there but don't ignore other categories entirely.
This is an analysis task. Do not write code or produce patches.
</task>

Project context: If CLAUDE.md exists in the repo root, always include it so reviewers check against project conventions.

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

Dispatch rules:

• When running both tools, issue mcp__codex__codex and Bash(gemini...) calls in a single message for parallel execution.
• Dispatch directly without pre-checking tool availability. If a tool fails, run only the available one.

Iterative mode (for plans/architecture, or when explicitly requested):

1. Dispatch initial review
2. If reviewer returns significant concerns, revise the material based on feedback
3. Re-submit using mcp__codex__codex-reply (with threadId) or a fresh gemini call including revision context
4. Max 3 rounds to prevent loops

4. Present Results

Rabbit hole detection: before presenting output, scan each model's response for signs it got distracted by agent config files instead of reviewing project code. Look for: .claude/, SKILL.md, skills/, .agents/, CLAUDE.md (when mentioned as a skill file rather than project conventions), gstack, plugin.json, or hooks/. If detected, append a warning:

One or more models appear to have read agent configuration files instead of reviewing your code. Their output may be unreliable. Consider retrying.

Gate verdict: for code review task type, determine the gate from the output:

• Scan for [P1] markers in findings. If any exist, the gate is FAIL.
• If no [P1] markers exist (only [P2]/[P3] or no findings), the gate is PASS.
• Display the gate prominently: GATE: PASS or GATE: FAIL (N critical findings).

Both tools: present each model's findings, then a Cross-Model Analysis section:

CROSS-MODEL ANALYSIS:
  Both found:       [findings that overlap between Codex and Gemini]
  Only Codex found: [findings unique to Codex]
  Only Gemini found:[findings unique to Gemini]
  Agreement rate:   X% (N of M total unique findings overlap)
  Gate:             PASS / FAIL (N critical findings)

Follow with prioritized action items. When both models flag the same issue, it is high-confidence. When only one model flags something, note which one and your assessment of whether it is valid.

Single tool: present findings with gate verdict, then your assessment of which items are valid vs uncertain.

Error Handling

Examples

Code review (both tools):

User: /second-opinions review
Agent: [gathers branch diff, dispatches both with adversarial review prompt]
Codex: [finds race condition in concurrent writes, flags as [P1]]
Gemini: [finds same race condition plus unhandled error path in auth retry]
Agent: [presents cross-model analysis, agreement rate, GATE: FAIL (2 critical findings)]

Plan review (iterative, single tool):

User: /second-opinions review my migration plan with codex
Agent: [asks focus] -> "Correctness & edge cases"
Agent: [extracts plan, dispatches to Codex]
Codex: [returns concerns]
Agent: [revises plan, re-submits via codex-reply]
Codex: [approves revised plan]
Agent: [presents final result]

Custom task:

User: /second-opinions have gemini analyze src/auth/ for race conditions
Agent: [dispatches Gemini with direct file reading]
Agent: [presents findings with assessment]

Gotchas

• Gemini hangs without --yolo. It enters interactive approval mode when called from another agent. Always pass --yolo (or -y).
• Codex defaults to implementing. Without an explicit "this is an analysis task, do not write code" instruction, Codex will produce patches instead of findings. Always include analysis-only framing.
• Models wander into skill files. Both Codex and Gemini will read .claude/, .agents/, and skill directories if not told to ignore them. The filesystem boundary instruction prevents this, but always check output for signs of distraction.
• Heredocs break with diffs. Diffs contain $, backticks, and special characters that break shell heredoc expansion. Always pipe with printf + cat instead.
• Bare git diff misses staged changes. Use git diff HEAD to capture both staged and unstaged modifications.
• Large diffs degrade review quality. Beyond ~2000 lines, both models start missing issues. Warn the user and suggest narrowing scope.
• -e security is silently ignored. The Gemini security extension installs as gemini-cli-security, not security.