pi-zellij-tools

Pi package that adds tools for running shell commands and independent Pi sessions in new terminal targets.

What it adds

• run_shell: runs a shell command in a new terminal target.
• spawn_pi: creates a new terminal target and runs interactive pi inside it.
• spawn_agent: starts a preconfigured agent session from agents/*.md files.
• handoff_agent: starts a visible child agent with file-backed result handoff.
• /run-shell [command]: slash command to run a shell command.
• /spawn-pi [task prompt]: slash command to spawn a child Pi session.
• /spawn-agent <agent-name> [task]: slash command to start a configured agent.
• The parent Pi does not wait for the child process to finish.
• The child process is directly visible and interactive in the terminal.
• If prompt is empty, the child Pi starts in plain interactive mode.
• If prompt is provided, the child Pi handles it first and then waits for follow-up input.
• floating and direction are mutually exclusive and cannot be used together.

Current backend

The current implementation uses Zellij as the terminal backend. The default target type is pane.

Requirements

• zellij available on PATH
• pi available on PATH (for spawn_pi / spawn_agent)
• The parent pi process must already be running inside Zellij

Installation

NPM

pi install npm:pi-zellij-tools

Git

pi install git:github.com/x1any/pi-zellij-tools

Security / Trust Boundary

Installing this package grants Pi the ability to spawn arbitrary shell commands and independent Pi sessions in new Zellij panes. The run_shell tool can execute any shell command, and spawn_pi/spawn_agent can start new Pi instances with arbitrary prompts.

Project-level agents

Project-level agents (from .pi/agents/ in the repo) are treated as potentially untrusted:

• By default (agentScope: "user"), only user-level agents from ~/.pi/agent/agents/ are loaded.
• To use project-level agents, explicitly set agentScope: "project" or "both".
• When a project-level agent is selected and confirmProjectAgents is true (default), a confirmation dialog is shown before launching.
• If no UI is available (e.g., RPC mode), project-level agents are rejected unless confirmProjectAgents: false is explicitly set.

General

• Only install this package in projects and with AI models you trust.
• The -- separator is used to prevent Zellij argument injection, but it does not limit the shell commands or Pi prompts themselves.
• Review generated commands before allowing execution if your client supports confirmation.

Tools

run_shell

Run a shell command in a new terminal target.

Parameters:

• command (required): shell command to execute
• cwd: working directory (defaults to current cwd)
• name: name for the terminal target (defaults to shell-command)
• shell: shell to use (defaults to sh)
• target: terminal target configuration

spawn_pi

Start an independent pi session in a new terminal target.

Parameters:

• prompt: initial prompt passed to the child pi process
• cwd: working directory (defaults to current cwd)
• name: name for the terminal target (defaults to pi-child)
• target: terminal target configuration
• model: model pattern or ID passed to pi via --model
• thinkingLevel: thinking level passed to pi via --thinking

spawn_agent

Start a preconfigured agent session from agents/*.md files in a new terminal target.

Parameters:

• agent: agent name from agents/*.md. Omit to list available agents. Supports optional scope prefix: user:<name>, project:<name>, proj:<name>, both:<name>.
• task: initial task prompt for the agent.
• agentScope: which agent directories to use: "user" (default), "project", or "both". If agent uses a scope prefix, that prefix takes precedence.
• confirmProjectAgents: whether to confirm before using project-level agents (default: true).
• cwd: working directory (defaults to current cwd)
• name: name for the terminal target (defaults to agent-<agent-name>)
• target: terminal target configuration

handoff_agent

Start a visible child agent and collect a structured result from .pi/agent-runs/<runId>/.

Parameters:

• agent (required): agent name from agents/*.md
• task (required): delegated task
• context: filtered relevant context for child
• agentScope: "user" (default), "project", or "both"
• confirmProjectAgents: confirm project-level agent usage (default: true)
• cwd, name, target: same meaning as other spawn tools
• wait: wait for completion (default: true)
• timeoutMs: wait timeout (default: 600000)
• pollIntervalMs: polling interval (default: 2000)

Run protocol directory:

.pi/agent-runs/<runId>/
  manifest.json
  task.md
  context.md
  instructions.md
  status.json
  result.md
  inbox.md
  notes.md
  artifacts/

Timeout behavior:

• timeoutMs only controls how long the parent waits for status.json.
• A timeout does not stop or close the child pane.
• If timeout happens, inspect .pi/agent-runs/<runId>/status.json and result.md.
• Child panes are named handoff-<agent> by default, or by the provided name.
• Close stale panes manually from Zellij if no longer needed.

Tool returns run metadata and latest status.

Target configuration

All tools accept an optional target parameter:

{
  type?: "pane";       // Currently only "pane" is supported
  direction?: "right" | "down";
  floating?: boolean;
}

Slash commands

• /run-shell [command]: run a shell command in a new terminal target.
• /spawn-pi [task prompt]: create a new child Pi pane with default settings.
• /spawn-agent <agent-name> [task]: start a configured agent session.

Agent configuration

Agents are defined as Markdown files with YAML frontmatter in an agents/ directory.

Paths

• User level: ~/.pi/agent/agents/*.md — always available.
• Project level: <project>/.pi/agents/*.md — discovered by searching upward from cwd.

File format

---
name: reviewer
description: Code review specialist
model: claude-sonnet-4-20250514
thinkingLevel: high
tools: read,bash,grep,find,ls
---

You are a code review specialist. Focus on:

- Code quality and best practices
- Security vulnerabilities
- Performance issues

tools can also be a YAML array:

tools:
  - read
  - grep
  - find
  - ls

Frontmatter fields

| Field | Required | Description | | --- | --- | --- | | name | Yes | Agent name, used for matching. Letters, digits, -, _, . | | description | Yes | Description, shown in agent list | | model | No | Passed to pi --model | | thinkingLevel | No | Passed to pi --thinking. One of: off, minimal, low, medium, high, xhigh | | tools | No | Passed to pi --tools. Comma string or YAML array |

The Markdown body (after frontmatter) is used as the agent's system prompt.

Example

Create ~/.pi/agent/agents/reviewer.md:

---
name: reviewer
description: Review code in read-only mode
model: claude-sonnet-4-20250514
thinkingLevel: high
tools: read,grep,find,ls
---

You are a reviewer. Do not edit files. Review for bugs and risks.

Then ask: "Use spawn_agent to start reviewer for the current diff."

Minimal verification

Inside Zellij:

zellij action list-panes --json
zellij action new-pane --cwd "$PWD" -n pi-smoke -d right -- pi "Reply OK and wait for follow-up instructions."

After installing the package, ask Pi to use run_shell, spawn_pi, or spawn_agent. Success means a new Zellij pane appears with the command running.