# Security The security and trust boundary HooKit draws. HooKit is a **policy layer, not a sandbox**. Understanding that boundary is the most important thing you can take from these docs. ## What HooKit does - Evaluates configured Hooks against bounded Pi Event data and returns the event-specific Event Outcome that may block, patch, cancel, or report. - Selects owned Actions against individual Hook Results and requests their Pi Effects separately, without changing the Event Outcome. - Reads project storage only after Pi **trusts** the project. Untrusted project storage is neither read nor written. - Records started Hook commands and selected Actions in bounded Execution Reports. ## What HooKit does not do **Hook shells execute locally as trusted code with your Pi process permissions.** A Hook's `shell` and `when` strings run through the real local shell with the same privileges as your Pi session—they can read, write, and execute anything your account can. The 5-second default timeout bounds commands and preconditions; it is **not a sandbox**. Reached Hook shells receive bounded Event data through `PI_*` environment variables. Tool inputs and tool-result text can contain source code, credentials, or other sensitive content, and the shell inherits the Pi process's ambient capabilities. Treat Hook logging and every repository Hook as code that can observe or disclose that data. This has two direct consequences: 1. **Treat repository Hooks as executable code.** Any Hook you install from a repository—or any policy another author shares—runs with your permissions. Review its Source (the plain-text Hook you are about to install) before enabling it, and prefer sources you trust. The install picker shows each candidate's `event:`/`shell:`/`when:` inline for exactly this reason. 2. **A policy check is not a sandbox for the tool it inspects.** When HooKit blocks a `write` to `.env`, it has inspected an input and refused—it has not quarantined the read or isolated the model. Do not store real secrets in files a policy merely attempts to guard, and do not rely on a guard Hook as a substitute for keeping secrets out of sensitive directories and model context. ## The trust model - Project configuration is read only after Pi trusts the project. A trusted project's entries replace global entries per Catalog merge rules; an untrusted project leaves your global Catalog and enablement unchanged. - Catalog eligibility is canonical: a Section is valid only as `local` or `owner/repo`; a project remote Section requires a declaration in that file's `repos`. Global canonical Sections do not require that declaration. Unusable Sections fail loudly instead of being silently filtered. - Nothing in this boundary is weakened by Action delivery: a decision (block, patch, cancellation, report) stands whether or not its Effect was delivered, so a failed delivery cannot let a blocked action through.