Claude Code’s agent playbook, ported to pi

What it does

This pi extension spawns isolated sub-agents inside your terminal session, each with its own model, system prompt, tools, and thinking level. Agents can run in the foreground or background while you keep working, turning a single conversation into a fleet of specialists that report back rather than blocking the main thread.

The interesting bit

The UI layer is unusually thorough for a terminal plugin: a persistent widget tracks live token counts and context-window pressure, a FleetView lets you hop between running agents with arrow keys, and you can inject messages into a background agent mid-flight to redirect it. It even handles the boring infrastructure—git worktree isolation, graceful turn-limit warnings, and cron-style scheduling—so the agents behave more like reliable coworkers than runaway processes.

Key highlights

• Background agents run in parallel with a configurable concurrency limit and queue, returning styled notification boxes instead of raw XML dumps.
• Mid-run steering and session resume let you redirect or pick up agents without restarting their context.
• Custom agent types via Markdown files with YAML frontmatter, plus fuzzy model selection and tool denylists.
• Git worktree isolation automatically commits agent changes to separate branches.
• An event bus and cross-extension RPC let other pi extensions spawn and kill subagents programmatically.

Caveats

• Scheduling is session-scoped: jobs reset on /new and persist via JSON files with PID-based file locking for cross-instance safety.
• Headless pi -p explicitly does not wait for scheduled subagents, so unattended automation can drop tasks.
• Transcript files land in the OS temp directory with owner-only permissions and are cleared on reboot.

Verdict

Worth a look if you use pi daily and have been jealous of Claude Code’s ability to spin up parallel research or refactoring tasks. Skip it if you’re happy with a single linear chat or your workflow rarely leaves the main thread.