agents-hooks_skill

Overview

This skill provides event-driven hooks for AI coding agents, focused on automating Claude Code or Codex CLI workflows using stdin JSON and decision-control responses. It defines hook events, input/output schemas, execution rules, and practical templates to guard, modify, or augment tool runs. Use it to enforce safety, run fast checks, and integrate agent runs into CI/CD pipelines.

How this skill works

Hooks are configured in a settings.json mapping events to command or prompt handlers. Each hook receives a JSON payload via stdin and runs as a normal user process with a 60s default timeout; matching hooks run in parallel and identical commands are deduplicated. PreToolUse hooks can return structured decision-control JSON (allow/deny/ask and updatedInput) and prompt hooks allow LLM-evaluated decisions for nuanced cases.

When to use it

• Automating event-driven workflows for Claude Code or Codex CLI
• Blocking or modifying dangerous shell/tool invocations (PreToolUse)
• Formatting, auditing, or notifying after successful tool runs (PostToolUse)
• Initializing runtime or repo state at session start or setup
• Running tests or summaries when the agent stops

Best practices

• Treat stdin JSON as untrusted input and validate it before use
• Use absolute paths, quote variables, and avoid eval to prevent injection
• Keep runtime preflight checks deterministic and fast (<1s target) and fail with clear remediation commands
• Prefer command hooks for deterministic checks and prompt hooks for nuanced policy decisions
• If strict ordering is required, wrap multiple steps in a single script rather than relying on parallel execution

Example use cases

• PreToolUse guard that blocks destructive git or rm commands and suggests safer alternatives
• PostToolUse formatter plus linter run on edited files before committing changes
• SessionStart preflight that verifies required binaries and versions and prints remediation steps
• Stop hook that runs test suite and writes a summary report to stdout for injection into the final context
• Permission auto-allow rules for repetitive, low-risk command prefixes paired with deny rules for dangerous patterns