Claude Code

Install

nah hook claude install

Restart Claude Code and inspect the active PreToolUse hook with /hooks. Remove only nah's entry with:

nah hook claude uninstall

The installer preserves unrelated ~/.claude/settings.json content and is idempotent. It uses Claude Code's native executable-and-arguments form, with no copied script, shell quoting, or Python dependency.

Behavior

Known Bash, read, write, edit, search, and related tool calls enter the shared nah pipeline. A definite violation is returned as a block with nah-branded feedback. Everything else returns no decision and continues through Claude Code's normal permission flow. nah never returns Claude's explicit allow, so it never skips a permission prompt.

Boundaries

PreToolUse runs before Claude's permission layer. Current Claude Code documentation does not say that bypassPermissions (--dangerously-skip-permissions, often called YOLO mode) disables hooks. It skips normal permission prompts, so delegated calls can execute immediately. Explicit ask rules, selected connector and MCP interactions, and root/home removal circuit breakers can still prompt. Treat the hook as a guardrail, not as a sandbox guarantee.

--safe-mode disables all customizations, including hooks. --bare skips automatic hook discovery. Neither mode loads nah's user hook.

Claude continues a tool call when a command hook exits with anything other than the blocking exit code 2, including ordinary errors and timeouts. Evaluation failure delegates and returns a fixed systemMessage; malformed outer input returns no decision. Other matching hooks run in parallel, including hooks bundled by enabled plugins, and one hook cannot prevent another from starting. Users can set disableAllHooks; administrators can set allowManagedHooksOnly, which excludes this user hook unless nah is installed as managed configuration.

Hook coverage and loading remain runtime-owned. Inspect /hooks after installation and configuration changes. Remote Claude Code environments need their own installation. Trusted hooks or tools that perform work directly without the intercepted call remain outside nah.

While active, this adapter blocks visible lifecycle commands, mutations to its shared user settings.json, and visible child launches using --safe-mode or --bare. Permission modes such as --dangerously-skip-permissions do not disable hooks and are not self-protection findings. The agent is told not to retry protected changes; an operator can use nah nap from another terminal.

This integration is best effort: runtime APIs and hook behavior can change. After upgrades, verify the latest official upstream documentation linked below, inspect the loaded hook, and test it before relying on nah. See Claude Code's hooks reference, permission modes, and settings reference.