Hook Manager Skill | Agent Skills

Hook Manager

Manage the hooks that ship with the core plugin across Claude Code, Codex, and Grok Build. The catalog source of truth is hooks/manifest.json in the installed plugin — read it live rather than trusting a memorized list:

cat "$(ls -d ~/.claude/plugins/cache/b-open-io/core/*/ | sort -V | tail -1)hooks/manifest.json"

Each entry carries name, event, matcher, tier, runtimes, summary, and description. Hooks register automatically when the plugin is installed; this skill controls which ones actually run.

The config file

Per-hook enable/disable lives in a JSON config. Resolution order (first file with an explicit true/false verdict for a hook wins):

$BOPEN_HOOKS_CONFIG — explicit override (tests, scripts)
<project>/.claude/bopen-hooks.json — per-project (Claude)
<project>/.grok/bopen-hooks.json — per-project (Grok)
~/.claude/core/hooks-config.json — per-user (Claude)
~/.grok/core/hooks-config.json — per-user (Grok)

{
  "version": 1,
  "hooks": {
    "bouncer": true,
    "damage-control": true,
    "publish-gate": true,
    "agent-browser-solo": true,
    "session-context": true,
    "hammertime": true,
    "browser-intent": true
  }
}

A hook is disabled ONLY by an explicit false. Missing files, missing keys, or unparseable JSON all mean enabled — a broken config must never silently switch the guards off. Changes take effect on the next hook invocation; no restart is required.

First-time setup (the [BOPEN-HOOKS-SETUP] directive)

When session context carries [BOPEN-HOOKS-SETUP], the user has no config yet. Do not interrupt their task; offer setup at a natural pause. The flow:

Read the manifest (command above) and present the hooks in two tiers via AskUserQuestion (multiSelect):
- Guards (recommended on): entries with tier: "guard" — they prevent work loss, secret exposure, and unticketed publishes.
- Workflow (preference): entries with tier: "workflow".
Run the prerequisite checks below and fold findings into the recommendation (e.g. agent-browser-solo without agent-browser installed silently falls back to native WebFetch — still safe to leave on).
Write ~/.claude/core/hooks-config.json (Claude/Codex) or ~/.grok/core/hooks-config.json (Grok) with the full hooks map — include every hook with an explicit boolean, even the all-defaults case. Writing the file is what dismisses the setup notice permanently.

Enabling / disabling a hook

Edit the user config (or project config for repo-scoped changes) with the Read and Write tools — read, flip the boolean, write back. Create the file from the template above when it does not exist. Never edit files inside the plugin cache; updates overwrite them.

Expect a confirmation prompt on every config write: damage-control treats these files as ask-tier so guards can never be disarmed without the user's explicit yes (on Codex the write is denied outright — have the user edit the file themselves). This is by design; do not route around it.

Warn before disabling guards: bouncer stops git reset --hard-class work loss, damage-control enforces zero-access paths like .env, publish-gate blocks unticketed npm/on-chain publishes. Disabling them is the user's call, but say plainly what protection goes away.

Diagnosis

When a hook misbehaves or the user asks "why did X get blocked/skipped":

# Which config verdict applies to a hook?
for f in "$BOPEN_HOOKS_CONFIG" ./.claude/bopen-hooks.json ./.grok/bopen-hooks.json ~/.claude/core/hooks-config.json ~/.grok/core/hooks-config.json; do
  [ -f "$f" ] && echo "$f: $(jq -r '.hooks["<name>"] // "no verdict"' "$f")"
done

# Prerequisites
command -v jq || echo "jq missing — hooks fail open without it"
command -v agent-browser || echo "agent-browser missing — agent-browser-solo falls back to native WebFetch"
[ -n "$LINEAR_API_KEY" ] || echo "LINEAR_API_KEY unset — publish-gate fails closed on gated publishes"
command -v python3 || echo "python3 missing — hammertime and JSON escaping degrade"

Hard denies surface to the model as structured permission denials; on the Codex runtime they arrive as stderr JSON with exit 2. Both are by design.
hammertime also has its own controls — Skill(core:hammertime) and the hammertime:manage skill — for rule-level tuning beyond on/off.
The definitive behavior reference for every hook is its script header in the plugin's hooks/ directory; read it before guessing.

What this skill never does

Never copies hook files into ~/.claude — hooks ship with the plugin and update through claude plugin update core@b-open-io.
Never edits hooks/*.sh, claude-hooks.json, or anything in the plugin cache.
Never disables a guard hook without telling the user what it protected.

Agent Skills: Hook Manager