Crate coding_agent_hooks 

Agent-agnostic hook protocol types and adapters for AI coding agents.

This crate provides the building blocks for writing hook-based middleware that works with any AI coding agent (Claude Code, Gemini CLI, Codex CLI, Amazon Q, OpenCode, Copilot CLI, etc.).

Core Concepts

• AgentKind identifies which agent is calling.
• HookProtocol abstracts agent-specific JSON formats.
• ToolUseHookInput and friends are the normalized input types parsed from any agent’s hook JSON.
• HookOutput is a structured response type (used directly by Claude Code; other agents convert via protocol methods).
• Tool and permission-mode name resolution normalizes across agents.

Adding a New Agent

Implement HookProtocol — only two methods are required (agent and parse_tool_use). Add entries to the tool and mode alias tables in agents so policy evaluation uses consistent names.

Re-exports

pub use agents::AgentKind;

pub use input::HookInput;

pub use input::SessionStartHookInput;

pub use input::StopHookInput;

pub use input::ToolUseHookInput;

pub use output::Effect;

pub use output::HookOutput;

pub use output::HookSpecificOutput;

pub use output::PermissionBehavior;

pub use output::PermissionDecision;

pub use output::PermissionRequestOutput;

pub use output::PostToolUseOutput;

pub use output::PreToolUseOutput;

pub use output::SessionStartOutput;

pub use protocol::HookProtocol;

Modules

agents

Multi-agent support — agent identification, tool name normalization, and protocol implementations.

input

Hook input types — normalized representations of what agents send via stdin.

output

Hook output types — structured responses sent back to agents via stdout.

protocol

Hook protocol abstraction for multi-agent support.