Second Opinion — Independent AI Review

Spawn a sub-agent with fresh context — no bias from your current conversation. Three modes.

---

Mode 1: Review

Independent diff review against the base branch.

1. Get the diff:

   git diff origin/<base>

2. Assemble a context block: branch name, diff stat, commit messages, what the PR claims to do
3. Spawn a sub-agent with this prompt:

You are an independent code reviewer seeing this diff for the first time. No prior context.

Context: {branch}, {N} files changed, {stat summary} Commits: {commit messages} Stated intent: {what the PR says it does}

Diff: {full diff}

Review this diff. Be direct. Name specific files and line numbers.

1. What's GOOD about this code?
2. What's BROKEN or will break in production?
3. What's MISSING (tests, error handling, edge cases)?
4. Verdict: PASS / CHANGES REQUESTED / BLOCKED

4. Present the output verbatim under

   SECOND OPINION:

   header
5. Synthesize: where you agree, where you disagree, whether it changes your verdict

---

Mode 2: Challenge

Adversarial mode — the sub-agent tries to break your code.

1. Read the changed files fully
2. Spawn a sub-agent:

You are an adversarial code reviewer. Your job is to BREAK this code.

Changed files: {file contents}

For each file:

1. Find inputs that crash it
2. Find race conditions
3. Find security holes
4. Find edge cases the author didn't consider
5. Find performance traps

Be creative. Be mean. Think like a bug bounty hunter. For each finding: file:line, the attack vector, expected vs actual behavior.

3. Present findings. For each confirmed vulnerability, offer to fix it.

---

Mode 3: Consult

Ask the sub-agent anything about your codebase.

1. Assemble context: relevant files, current problem, what you've tried
2. Spawn a sub-agent:

Context: {relevant code and problem description}

The user asks: {question}

Be direct. Name specific files and functions. If you're guessing, say so. If you need more context, say exactly what you need.

3. Present the answer. If it needs follow-up, loop.

---

How to Invoke

Ask the user:

What do you want?

• Review — independent diff review (pass/fail)
• Challenge — adversarial mode, try to break my code
• Consult — ask anything about the codebase

Then proceed with the matching mode.

Important

• The sub-agent has no memory of your conversation. That's the feature — fresh eyes.
• Present its output verbatim. Don't summarize or filter.
• After presenting, synthesize: agree/disagree, what changes.
• If the sub-agent fails or times out: "Second opinion unavailable. Continuing without it."