pi-agent-clilisted

# executing-pi-agent ## Purpose Use this skill to perform coding, review, research, or automation tasks through the `pi` CLI. Prefer explicit tool allowlists so each invocation receives only the capabilities required for the task. Pi implements the Agent Skills standard and can also consume compatible `SKILL.md` files directly. This skill covers the other direction: invoking Pi as an execution agent from another coding-agent workflow. ## Security Model Pi runs with the operating-system permissions of the user that launches it. Project trust controls whether project-local settings, skills, extensions, and related resources are loaded; it is **not** a sandbox and does not restrict what enabled tools can do. For untrusted repositories, unattended execution, or high-impact tasks, run Pi inside an external containment boundary such as a container, VM, micro-VM, or policy-controlled sandbox. Mount only the required workspace paths and pass only the minimum credentials and network access needed. ## Capability Tiers | Tier | Capability | Pi tools | Approval required | Typical tasks | | :--- | :--------- | :------- | :---------------- | :------------ | | **0** | Observe | `read,grep,find,ls` | No | Code review, audits, architecture analysis, local research | | **1** | Edit | `read,grep,find,ls,edit,write` | Yes | Refactoring, documentation, mechanical fixes | | **2** | Execute | Add `bash` only when required | Yes (High Risk) | Tests, builds, dependency changes, scripted autom