Claude Code
Wire the hooks, agents and deny rules into Claude Code.
install.sh copies agents/ into ~/.claude and writes the four skills hooks into Claude Code's settings.json (in CLAUDE_CONFIG_DIR, else ~/.claude). With AGENT_RULES set in skills.conf, it also manages permission rules in that file. The agents need nothing more.
Hooks
The installer adds missing skills hooks and leaves personal entries alone. Rerunning it with no changes leaves the file untouched. If it refuses the file, it prints the reason and a block containing only hooks still missing for you to merge into hooks yourself. Declined and unwired targets are excluded. When none are missing, it prints no block. The complete block below uses the default AGENTS_DIR, ~/.agents/skills. With AGENTS_DIR set, the installer writes absolute commands for that directory (Configuration).
{
"hooks": {
"SessionStart": [{ "matcher": "startup|resume|clear|compact", "hooks": [{ "type": "command", "command": "~/.agents/skills/playbook/bin/skills hook session-start" }] }],
"PreToolUse": [{ "matcher": "^Bash$", "hooks": [{ "type": "command", "command": "~/.agents/skills/playbook/bin/skills hook pre-tool-use" }] }],
"PostToolUse": [{ "matcher": "^(Edit|MultiEdit|Write)$", "hooks": [
{ "type": "command", "command": "~/.agents/skills/playbook/bin/skills hook post-tool-use" } ] }],
"Stop": [{ "hooks": [{ "type": "command", "command": "~/.agents/skills/playbook/bin/skills hook stop" }] }]
}
}What each hook does, and when it stays silent, is on the Hooks pages, along with the switches that turn them off.
All four hook commands run the skills CLI from the checkout through the AGENTS_DIR link. Both the default ~/ spelling and its absolute spelling count as present, so neither is duplicated.
The installer replaces retired shell hooks whose copies match a version this repository committed. A missing copy also gets replaced. It replaces session-brief.sh on SessionStart, commit-guard.sh on PreToolUse, and reply-guard.sh on Stop with their skills commands.
On PostToolUse, it replaces no-comments.sh and no-em-dash.sh with one skills hook post-tool-use command. A copy you edited stays wired and gets a skip line rather than a second hook beside it. A retired hook path outside the configured hooks home also gets a skip line, even if it names the default ~/.claude/hooks directory.
Hook changes appear under Add or Update in the terminal drift prompt. Unticking a hook saves CLAUDE_HOOK_<TARGET>=off in skills.conf, for example CLAUDE_HOOK_PRE_TOOL_USE=off. This prevents later installs from adding or replacing that target. Delete the line to re-enable it. It does not remove an already wired skills command or affect Codex.
If the playbook link does not run this checkout, the installer skips its hooks. Without Bun, it skips all but pre-tool-use, which denies shell commands until Bun is installed.
An unreplaced shipped copy still referenced by your settings gets a stale line. This can happen when the file was refused or the command uses an unrecognised spelling such as bash ~/.claude/hooks/session-brief.sh. Replaced entries and declined targets get no stale line.
Deny Rules
With AGENT_RULES set, each install run writes strengthening and neutral permission changes for your delivery mode. Only owned entries change, though formatting can be normalised to the file's indent unit. At a terminal, weakening changes are listed for you to confirm, and an untick is saved as that group's override. Other runs give them a left line telling you to edit the file by hand. With the key unset or the file refused, it names the reason instead, and --verbose prints the set for you to paste into permissions. Deny Rules renders the set and says which entry to drop once the commit guard is wired.
Add entries of your own: EnterWorktree, and any package manager your lockfile does not sanction, such as Bash(npm install:*) in a repository pnpm owns.
A deny rule is an exact prefix match on a tool call, so it does not misfire the way a hook grepping the command string does. A rule in prose is a suggestion. A rule in settings is a rule.
Allowed CLI verbs
With AGENT_RULES set, the installer allows these verbs through the Claude skills link and the agents store after you confirm the change. Adding an allow is a weakening change.
skills plans waveskills plans resumeskills plans scheduleskills plans frontierskills plans set-rowskills plans addskills plans logskills plans importskills plans showskills plans historyskills plans checkoutskills recallskills threads logskills threads resumeskills plans closeskills plans sweepskills plans startskills plans pinskills plans lintskills plans handoffskills plans chainskills plans stack-baseskills plans belowskills briefskills reviewersskills round gateskills round decideskills settingsskills review readskills pr greenskills pr watchskills deliveryskills fast-modeskills armsskills checkskills auditskills deviceskills dash fetchskills dashskills digestskills machine showskills machine labelskills machine refreshskills issues fileskills issues listskills issues showskills issues closeskills issues linkskills sessions scanskills sessions listskills sessions showskills sessions restoreskills hub joinskills hub devicesskills hub statusskills hub sync
Set AGENT_RULES_SKILLS_CLI=off in skills.conf to stop the installer adding these entries. Off leaves the group unmanaged, so entries an earlier run wrote stay until you remove them from permissions.allow in settings.json.
Agents
The Claude agents in agents/ load only in Claude Code, from the copies in ~/.claude/agents. The panel skills (interrogate, how, architect) fan out to them for independent perspectives next to the Codex arms, /no-comments sends its reviewer over a diff, and the playbook agent takes delegated work that follows the playbook.
| agent | what it does |
|---|---|
comment-sicko | Read-only comment reviewer. Reads a diff or a file set and reports every code comment that should die, plus scoped lint and type suppressions. Spawned by the no-comments skill. Reports findings; it never edits. |
fable-judgment | Fable at high reasoning, the level its own guidance recommends as the starting point. Use for judgment calls, taste, product and API design, UI and UX direction, naming, user-facing copy, plan and design review, and briefs whose intent is vague enough that the hard part is deciding what to build. |
opus-review | Opus at high reasoning. Second Claude arm for adversarial panels, plan and implementation review, hard debugging, and long-context codebase analysis. Use alongside fable-judgment when a decision needs two independent perspectives. |
playbook-agent | Default subagent for investigation, implementation and multi-step work. Reads the playbook skill in full before touching anything, then works to its principles. Use instead of general-purpose whenever the task follows those conventions. |