Skip to main content
command CX Codex

verify-changes — Claude Code command: what it does, how to install

verify-changes executes adversarial multi-subagent verification across recent code diffs to weed out bugs and false alarms.

MIT ★ 1,424 updated 2026-10-06

Install into Claude Code · Manually: put the files in place All methods and platforms ↓
.claude/commands/verify-changes.md
Open on GitHub

What verify-changes does

The command inspects the latest repository changes via git diff HEAD~1 and launches five background subagents in a single message. These subagents simultaneously validate syntax with tools like tsc or mypy, execute targeted tests, inspect style via linters, compile build targets, and grep for leaked credentials.

Raw findings do not go straight into the report. A secondary trio of adversarial subagents reviews every flagged item to filter false positives, uncover missing edge cases, and ensure reported issues are not intentional patterns. The final summary presents pass/fail verdicts alongside confirmed flaws.

Parallel execution relies strictly on the Task tool supporting run_in_background: true in your environment.

What it helps with

  • verify recent commits across typing, test suites, builds, and linters
  • filter out false positives from automated verification passes
  • scan modified files for committed tokens and credentials
  • catch missed edge cases and faulty error handling after edits

Example prompts for the agent

  • /project-starter:verify-changes
  • Run /project-starter:verify-changes to inspect the latest diff
  • Verify my recent edits using /project-starter:verify-changes

How to install verify-changes

Manually: put the files in place

.claude/commands/verify-changes.md

Call it with /verify-changes. For all projects use ~/.claude/commands/

Questions about verify-changes

How to install verify-changes in Claude Code?

Place the command file at .claude/commands/verify-changes.md.

What does verify-changes do?

It reads changed files from Git, runs five automated verification subagents simultaneously, and refines the findings using three adversarial reviewers.

Why does it use an adversarial review phase?

Three specialized agents review the initial findings to discard non-issues, check project context, and catch edge cases missed during the first pass.

Source text

verify-changes.md · MIT
---
description: Comprehensive verification after code changes. Uses Boris Cherny's multi-subagent adversarial approach.
---

# Verify Changes

Run comprehensive verification of recent code changes using multi-subagent adversarial approach.

## Phase 1: Gather Context

Run these commands to understand what changed:

```bash
git diff HEAD~1 --stat
git log -3 --oneline
git diff --name-only HEAD~1
```

Identify affected files and their corresponding test files.

## Phase 2: Spawn Verification Subagents (Parallel)

Use the Task tool with `run_in_background: true` to spawn subagents in parallel.

**CRITICAL**: Launch ALL subagents in a SINGLE message for true parallelism:

```
[All Task calls in ONE message]

Task 1: {
  description: "Syntax & Type Check",
  prompt: "Run language-specific type checker (tsc, mypy, go vet)...",
  run_in_background: true
}

Task 2: {
  description: "Test Runner",
  prompt: "Run tests for affected files first, then integration tests...",
  run_in_background: true
}

Task 3: {
  description: "Lint & Style Check",
  prompt: "Run project linter (eslint, ruff, golangci-lint)...",
  run_in_background: true
}

Task 4: {
  description: "Security Scan",
  prompt: "Grep for hardcoded secrets, check for vulnerabilities...",
  run_in_background: true
}

Task 5: {
  description: "Build Validator",
  prompt: "Run build command, verify artifacts exist...",
  run_in_background: true
}
```

Each subagent returns its result automatically when it completes — there is no
separate retrieval call. Read each verifier's report as it returns.

### Subagent Details

### Subagent 1: Syntax & Type Check

- Run language-specific type checker (tsc, mypy, go vet)
- Report any type errors or warnings
- Exit with list of issues found

### Subagent 2: Test Runner

- Run tests for affected files first
- Run related integration tests
- Report failures with file:line context

### Subagent 3: Lint & Style Check

- Run project linter (eslint, ruff, golangci-lint)
- Check for formatting issues
- Report violations with severity

### Subagent 4: Security Scan

- Grep for hardcoded secrets (passwords, API keys, tokens)
- Check for common vulnerabilities in new code
- Review any new dependencies

### Subagent 5: Build Validator

- Run build command
- Verify build artifacts exist
- Check for build warnings

## Phase 3: Adversarial Review

Spawn 3 adversarial subagents to review Phase 2 findings:

### Adversarial A: False Positive Filter

- Review each finding from Phase 2
- Determine if it's a real issue or false alarm
- Provide reasoning for each determination

### Adversarial B: Missing Issues Finder

- Look for issues the first pass might have missed
- Check edge cases in changed code
- Verify error handling is adequate

### Adversarial C: Context Validator

- Verify findings make sense in project context
- Check if "issues" are actually intentional patterns
- Consider project conventions

## Phase 4: Synthesize Results

Combine all subagent outputs into final report:

```
## Verification Results: [PASS/FAIL]

### Confirmed Issues
1. [Issue] - [Location] - [Why it's confirmed real]

### Warnings (Review Recommended)
1. [Warning] - [Location] - [Context]

### All Checks
- [ ] Type checking: [PASS/FAIL]
- [ ] Tests: [PASS/FAIL]
- [ ] Linting: [PASS/FAIL]
- [ ] Security: [PASS/FAIL]
- [ ] Build: [PASS/FAIL]

### Subagent Summary
- Initial findings: X issues
- After adversarial review: Y confirmed issues
- False positives filtered: Z
```

## Usage

This command ships with the project-starter plugin. Invoke with: `/project-starter:verify-changes`

© CloudAI-X · MIT · CloudAI-X/claude-workflow-v2

Similar slash commands

All slash commands for Code review and refactoring →

From the same set46

The whole set →
run-tests run-tests runs the project's tests in tiers: the ones your changes touch first, then the whole unit suite, then integration. CX Codex +1 ★ 1,424repo stars parallel-review parallel-review runs code review across multiple directories or files simultaneously, one subagent per target. CX Codex +1 ★ 1,424repo stars code-simplifier The code-simplifier command reviews recent changes and proposes cleaner code without changing behavior. Use it after implementing a feature or refactoring. CX Codex +1 ★ 1,424repo stars security-scan security-scan is a command for Claude Code that checks a repository for hardcoded secrets, vulnerable dependencies, injection, XSS and weak config before a commit or PR. CX Codex +1 ★ 1,424repo stars save-session-learnings The save-session-learnings command for Claude Code saves useful findings from a coding session to CLAUDE.md and AGENTS.md after debugging or a substantial task. CX Codex +1 ★ 1,424repo stars analyzing-projects analyzing-projects walks through an unfamiliar codebase step by step: stack, structure, patterns, dev commands — useful when onboarding onto a project. CX Codex +1 ★ 1,424repo stars plan Creates and maintains a structured PLAN.md file in your project root to track phases, task dependencies, and progress logs. CX Codex +1 ★ 1,424repo stars refactor-guided refactor-guided guides structural refactoring in small steps and runs tests after each change. It keeps feature work separate from code restructuring. CX Codex +1 ★ 1,424repo stars

Project sponsors

Companies that keep this analytics open