Agent Task Verifier

Verify what your AI coding agent changed before you trust it.

What It Is

Agent Task Verifier is a local CLI for developers who use Claude Code, Codex, Cursor, or another AI coding agent. It runs after an agent has changed your repo and checks whether the result looks aligned, consistent, runnable, and safe enough to trust.

It is not an AI coding agent and not a code generator. It verifies AI coding agent output before you commit, merge, publish, or deploy.

Why It Exists

AI coding agents can modify many files quickly, but review is still the bottleneck. This tool gives you a fast, deterministic first pass over what changed, whether it matches the task, and which risks deserve manual review.

Quick Start

npm install -g agent-task-verifier
agent-task-verifier init
agent-task-verifier verify

v0.1.0 is fully local and deterministic. It has no LLM dependency, no paid API dependency, and no external AI API calls.

What It Checks

• Git change scope and changed files
• Rough task alignment against .agent-task-verifier/task.md
• Risky files such as env files, credentials, private keys, build outputs, and generated reports
• Package and lockfile consistency
• Suspicious package scripts and lifecycle scripts
• Documentation and implementation consistency
• Security-sensitive diff patterns
• Available local scripts such as typecheck, build, test, and lint

Local Development

npm install
npm run build
node dist/cli.js --help

Example Usage

agent-task-verifier verify --task-file ./TASK.md --base main
agent-task-verifier verify --run-checks
agent-task-verifier report

Commands

• agent-task-verifier init creates .agent-task-verifier/task.md and .agent-task-verifier/config.json.
• agent-task-verifier verify checks the current git diff against the task brief and writes Markdown and JSON reports.
• agent-task-verifier verify --task-file <path> uses a custom task brief.
• agent-task-verifier verify --base <git-ref> compares against a specific git ref. The default is HEAD.
• agent-task-verifier verify --run-checks runs safe local checks when matching npm scripts are available.
• agent-task-verifier report prints the latest report summary.
• agent-task-verifier --help prints command help.
• agent-task-verifier --version prints the package version.

Report Output

verify writes:

• AGENT_TASK_VERIFICATION_REPORT.md
• agent-task-verification-report.json

The report includes a status, trust score, task summary, changed files, check results, warnings, recommended manual review items, and a final recommendation.

Statuses

• TRUSTED: The changes look aligned and no major risks were detected.
• REVIEW_REQUIRED: The changes may be acceptable, but warnings or uncertainty require human review.
• HIGH_RISK: Significant risks or inconsistencies were detected.
• BLOCKED: The result should not be trusted until blocking issues are resolved.

What v0.1.0 Does Not Do

• It does not call an LLM or external AI API.
• It does not prove correctness or security.
• It does not replace human review.
• It does not install dependencies automatically.
• It does not publish, deploy, or modify remotes.