Agent Validator
Agent Validator is a configurable feedback runner for AI-assisted development workflows. It runs the checks and AI reviews that apply to the files that changed, records structured runtime state, and supports agent skills that can fix and rerun failures until validation passes.
Use it when you want an agent to work inside an external validation loop: build, lint, test, review, fix, rerun.
Quick Start
Requirements:
- Node.js 18 or newer
- Git
- Optional review CLIs: Claude Code, Codex, Gemini, GitHub Copilot, Cursor, or OpenCode
Install and initialize:
npm install -g agent-validator
agent-validate init
agent-validator is also installed as an alias, but the CLI help and generated docs use agent-validate.
After init, run /validator-setup from your coding agent to discover project tooling and write the first checks into .validator/config.yml.
What It Does
- Runs deterministic check gates such as build, lint, typecheck, test, and security scanners.
- Runs AI review gates through local CLI adapters, using your existing tool subscriptions.
- Matches changed files to configured entry points so unrelated parts of a repo do not run unnecessary gates.
- Supports disabled opt-in reviews such as
task-compliance, activated with --enable-review.
- Tracks execution state and trusted snapshots so reruns verify fixes without repeatedly reviewing already validated work.
- Installs agent skills and plugins so agents can run, diagnose, skip, and commit with validator-aware workflows.
Minimal Config
cli:
default_preference:
- claude
adapters:
claude:
allow_tool_use: false
thinking_budget: low
model: claude-sonnet-5-5
entry_points:
- path: "."
exclude:
- validator_logs
- .validator
checks:
- test:
command: npm test
- lint:
command: npm run lint
reviews:
- all-reviewers:
builtin: all-reviewers
This uses the preferred reviewer from the latest review evals. Measured on the all-reviewers eval fixture (56 seeded issues):
| Option | Adapter | Model | thinking_budget | Recall | Precision | Cost / review |
|---|
| Preferred | claude | claude-sonnet-5-5 | low | 0.82 | 0.93 | ~$0.24 |
| Codex | codex | gpt-6-sol | medium | 0.70 | 0.74 | ~$0.15 |
See Recommended Reviewer Configurations for details. The 2026-10-15 end-to-end eval explains why the cheaper Haiku 5.5 is not recommended: its extra false positives cost more in fixing time and broken code than it saves on review.
Inline checks and reviews are defined inside entry_points. File-based gates are also supported under .validator/checks/ and .validator/reviews/.
Documentation
