agentic-validators_skill

Overview

This skill designs and installs validation hooks for coding agents to make automated code changes safer and more deterministic. It helps you pick per-file vs repo-wide strategies, add post-tool-use and stop hooks, and structure parallel subagents with per-file validation and audit logs. The goal is to turn “agent wrote code” into “agent wrote code and the change is validated automatically.”

How this skill works

The skill implements small, deterministic validators that run immediately after an agent edits files (post-tool-use hooks) and a broader final gate (stop hook) that enforces repo-wide invariants. Validators are fast checks like formatters, linters, or focused tests; results are written to predictable logs and aggregated when many subagents run in parallel. Slow or flaky work is moved to the final gate while keeping per-file checks quick and repeatable.

When to use it

• When you need immediate feedback on specific edited files
• When you want deterministic, repeatable checks for agent-produced changes
• When running parallel subagents that each validate their own file set
• When you need a final repo-wide gate before merging or deploying
• When you want audit logs of validator outcomes for traceability

Best practices

• Prefer small, fast per-file validators (<5–30s) and reserve slow tasks for a stop hook
• Make validators deterministic: avoid network calls and pin tool versions
• Fail fast on formatter changes and require the agent to commit formatted output
• Write clear, machine-parseable validator output and store logs in a predictable path (e.g., .agent-logs/)
• Use a final global gate for unit tests, build, and integration tests only after per-file validators pass

Example use cases

• Post-tool-use hook that runs black/ruff and typecheck on modified Python modules
• Per-file format+lint pattern: auto-format, lint the file, and fail if formatting altered code
• Parallel subagents: one subagent per CSV/config file, each validating its changes, with an aggregator accepting only passing results
• Stop-hook final gate that runs the full test suite and build before accepting changes
• Docs invariant validator: markdown lint and link checker for changed documentation files