Agent Playbook: local-first Behavior CI for agents

Who it is for

Use Agent Playbook when you want to add local-first Behavior CI — that is, continuous integration of behavior — to a coding Agent. The official one-line position is Local-first Behavior CI for coding agents. It configures skills for Claude Code, Codex, Gemini, and DeepSeek Harness. It fits people already using these four hosts who accept that init will write configuration into the repository. The blurb also says it turns repeated corrections into reversible behavior changes, but that introductory sentence does not unpack the mechanism, so do not treat an unwritten inbox or self-improvement flow as a proven capability. If you cannot accept any hooks capture, stay with the default: a new install turns Claude hooks off.

First try in 3 minutes

Prefer only the official first-choice init:

pnpm dlx @codeharbor/agent-playbook init

The same page also gives the npm form:

npm exec -- @codeharbor/agent-playbook init

The first time, use the pnpm line. This writes files; it is not read-only. First have the Agent restate: it will run init in the current repository, configure skills for Claude Code, Codex, Gemini, and DeepSeek Harness, and leave Claude hooks off by default. After you confirm the workspace path, allow it to run. Do not rewrite it as a skills add that never appears, and do not invent a skill directory.

Success signal: init finishes, and hooks stay off unless you explicitly asked for --hooks at that moment. Officially apb is a short alias for agent-playbook; the first time, do not invent apb subcommands — the page only guarantees the alias itself. After init, ask one read-only question and have the Agent explain what Behavior CI means in this repository, without changing behavior rules yet.

Usage examples

Example 1: Initialize skills for the four hosts. Run init at the project root so Claude Code, Codex, Gemini, or DeepSeek Harness can read Playbook skills. It fits people who want behavior constraints to stay in the local repository. The notes do not name the files that will be generated, so before you run you can only confirm “configuration will be written,” not pre-fill paths.

Example 2: Turn Claude hooks on explicitly. Officially a new install disables them by default; only adding --hooks enables bounded, redacted private session summaries and failure capture. It fits people who need to look back at failed sessions and accept summary capture. Before turning them on, have the Agent restate: it will enable bounded, redacted private summaries and failure capture, not the default install. Do not add --hooks onto an undocumented switch for another client. The blurb mentions a behavior inbox and self-improvement; the page gives no command, so do not invent one.

Limits

The default command is pnpm dlx @codeharbor/agent-playbook init. npm exec is a parallel spelling, not a third installer. The host names are Claude Code, Codex, Gemini, and DeepSeek Harness — the scope that will get skills configured, not a worldwide claim that “only these four can read SKILL.md.” New installs leave hooks off; --hooks is what captures redacted summaries. apb is written only as an alias. Do not fill in unwritten subcommands, directories, or store installs. Repository page:

https://github.com/zhaono1/agent-playbook