multi-ai-code-review

name	multi-ai-code-review
description	Run a comprehensive code review using available AI tools in parallel. Analyzes architecture, code quality, security, and test coverage. Dynamically detects and leverages available MCP servers for multi-perspective review. Use when reviewing code changes, before merging PRs, or for quality audits.
effort	xhigh
argument-hint	["branch-name"]
disable-model-invocation	true

Multi-AI Code Review

Run a comprehensive code review on $ARGUMENTS (or the current branch if no arguments are provided), leveraging any available AI MCP servers for multi-perspective analysis.

The bar for every finding (read first)

Every finding you report MUST clear the bar in refs/finding-bar.md: located (file:line), concretely fixable (minimal suggested diff), confident (high/medium only — drop the rest), and material (correctness / security / data / spec / explicit convention — never pure style). Prefer fewer, sharper findings: ten line-anchored issues with fixes beat forty generic observations. This bar is what makes the review as pointed as an inline PR reviewer instead of a wall of advice.

What the review is built around

The spine of this review is refs/high-signal-checklist.md — the concrete, recurring issues that automated PR reviewers (e.g. GitHub Copilot) flag and that generic "review for quality" prompts miss: spec↔code mismatch, enum/constant hardcoding, nil-on-nullable, time/range boundaries, get-or-insert races, missing indexes, XSS escaping, weak test assertions, message↔logic drift. Walking the diff against every category of that checklist is the primary job (Phase 3.b), and pre-empts the post-PR churn of fixing these one comment at a time.

The perspectives below are secondary lenses — a coverage net so nothing whole-cloth is missed. They are necessary but not sufficient; do not let them turn the review into generic advice. Any finding from a lens still has to clear the finding bar.

Perspective	Details
Architecture	Pattern appropriateness, SOLID principles, consistency with existing architecture
Quality	Readability, maintainability, duplication, complexity
Security	OWASP Top 10, input validation, authentication, authorization
Testing	Coverage gaps, coverage threshold compliance, test case sufficiency (normal/error/edge/boundary), edge cases
Performance	Inefficient data fetching, memory leaks, unnecessary computation, algorithmic complexity
Conventions	CLAUDE.md / AGENTS.md project convention compliance
Consistency	PR description / commit messages / linked issue vs. the actual diff (claimed scope, definitions, DB impact, behavior all match the code)

Phase 1: Collect Changes

Base Branch Detection

Determine the base branch automatically:

1. Run git remote show origin and look for the HEAD branch.
2. If that fails, check for main, master, or develop branches (in that order).
3. If multiple candidates exist, pick the first that exists locally or as a remote tracking branch.
4. If no base branch can be determined, ask the user.

Gather Change Information

git branch --show-current
git diff <base>...HEAD --stat
git diff <base>...HEAD
git status --porcelain
git log --oneline <base>...HEAD

PR description / intent (required for the Consistency perspective): capture the stated intent so the review can cross-check it against the diff (checklist category A).

# If a PR already exists for this branch:
gh pr view --json title,body 2>/dev/null

If no PR exists yet, use the commit messages (git log) and the linked issue (Linear/Jira ID in the branch name or commits) as the statement of intent. Fetch the issue body if the tooling is available.

Large diff handling: If git diff --stat shows more than 1000 lines changed or 30+ files, ask the user whether to review all changes or focus on specific directories/files.

Prepare Review Context

Read project conventions before starting the review:

• CLAUDE.md — coding conventions, style rules, project guidelines
• AGENTS.md — agent-specific instructions
• docs/ directory — architecture and design documents

Prepare the following context to pass to AI engines:

• diff: full output of git diff <base>...HEAD
• file_list: output of git diff --name-only <base>...HEAD
• commit_log: output of git log --oneline <base>...HEAD
• intent: PR title/body (or commit messages + linked issue) — the claims to verify
• conventions: contents of CLAUDE.md and AGENTS.md (if present)

Phase 2: Parallel AI Review (Dynamic MCP Detection)

Check your available tools for AI-powered MCP servers and dispatch reviews in parallel. Below are known integrations — use any that are available, and skip those that are not.

Pass the finding bar and the high-signal checklist into every dispatched prompt, and require each engine to anchor findings to file:line with a minimal fix. An engine that returns generic prose without locations is producing noise — re-prompt it for located findings or discard its output.

Codex — Architecture Review

Check if mcp__codex__spawn_agent is available. If so, spawn an agent in background with the diff, file list, commit log, and conventions, instructing it to review from an architecture perspective (design patterns, SOLID principles, architectural consistency, separation of concerns, dependencies, scalability).

Gemini — Code Quality & Security Review

Check if mcp__gemini__gemini-analyze-code is available. If so, call it with the full diff, instructing it to review code quality and security (vulnerabilities, error handling, duplication, test coverage, performance).

Other AI MCP Servers

If other AI-capable MCP tools are available, dispatch additional review perspectives to them.

Output Format for All Dispatched Reviews

Request each engine to return findings in this format:

• File: path and line range
• Severity: Critical / High / Medium
• Category: review perspective / checklist category
• Finding: what is wrong (not what the code does)
• Fix: the minimal concrete change (a diff if possible)
• Confidence: high / medium

If no AI MCP servers are available: Skip Phase 2. Claude performs a comprehensive review covering ALL perspectives in Phase 3.

Phase 3: Claude Integrated Review

a. Integrate Phase 2 Results

• Incorporate findings from any AI engines that returned results — but only those that clear the finding bar. Discard located-but-immaterial nits and unlocated prose.
• Consensus: findings flagged by multiple engines → elevate priority.
• Conflicts: if engines disagree, provide Claude's own assessment with reasoning.
• Gaps: identify any perspectives not covered by Phase 2 and review those areas directly.

b. Claude's Own Review — checklist-driven (the core pass)

This is the primary pass. Do it whether or not Phase 2 ran.

Step 1 — Spec ↔ implementation table (checklist category A). Take each factual claim in the captured intent (PR/commit/issue) and point at the line that proves or contradicts it. Emit the claim-by-claim table from the report template (Claim | file:line | VERIFIED / MISMATCH) before anything else. A MISMATCH that ships wrong behavior is Critical.

Step 2 — Walk every checklist category against the diff. Go through refs/high-signal-checklist.md categories B–I in order — B. enum/constant hardcoding, C. nil/nullable, D. time & range boundaries, E. concurrency & idempotency, F. indexes & query efficiency, G. web security & a11y, H. test rigor, I. message/i18n↔logic. Do not skim. For each category, either emit located findings or write "clear". Each finding gets file:line, severity, confidence, and a minimal-diff fix.

Step 3 — Secondary lenses. Sweep the perspectives table for anything the checklist didn't cover (architecture fit, integration risks, convention compliance, edge cases the checklist doesn't enumerate). Same bar applies — located, fixable, material, or it doesn't ship.

c. Auto-Detect and Run Quality Gates

Auto-detect the project's test, lint, and type-check commands from configuration files. Run each gate and report results. If a command is undetectable, ask the user. If the project does not use static typing, skip the type-check gate.

• Non-interactive mode: Always run commands in non-interactive mode to prevent watch-mode hangs.

d. Coverage Threshold Check

1. Detect the project's coverage command and configured thresholds from configuration files
2. If thresholds exist, run the coverage command and compare results against thresholds
3. Flag any coverage shortfalls as High severity findings with specific uncovered lines/branches

Output

Report findings the way an inline PR reviewer does — located, minimal, actionable — following refs/report-template.md. Print the report directly to the conversation. Do not create a file unless the user requests it.

Each finding is one block (outer fence shown with four backticks so the inner diff renders):

[Severity] path/to/file.rb:42 — one-line statement of the problem · category · confidence: high

Why it's wrong: one or two concrete sentences — what breaks, under what input/timing.

```diff
- offending line
+ minimal fix
```

• Lead with file:line. One finding = one location + one fix. The diff must be minimal — change only what the finding requires.
• Group findings Critical → High → Medium, then by file. Below Medium is not reported.
• For a design-level issue with no safe one-line fix, replace the diff with the smallest concrete next step and say why a patch isn't given.
• Always include the spec↔implementation table before the findings.
• No "general suggestions" / "keep in mind" section. If it has no line, it is not a finding.

Severity Levels

• Critical: Must fix before merge (security, data loss, breaking change, a spec mismatch that ships wrong behavior)
• High: Should fix before merge (bug, convention violation, missing index on a filtered/sorted column, weak test that passes while broken)
• Medium: Worth fixing, with a concrete fix (maintainability, a non-blocking checklist hit). Anything below Medium is dropped, not downgraded.

Notes

• Always run quality gates regardless of change size