pi-loadout

Interactive tool and skill loadouts for Pi sessions.

pi-loadout adds a /loadout command for choosing which tools and skills are active in the current Pi session, grouped by source extension/package. Toggles apply live as you change them, the session branch is updated on close, and you can save the current selection as a global default for future sessions.

Install

pi install npm:pi-loadout

Usage

/loadout          # open interactive picker
/loadout status   # print current active tools and skills
/loadout reset    # re-enable every available tool and skill in this session
/loadout help     # show subcommand list

Picker controls

• Tab — switch between Tools and Skills panes
• Space — toggle the selected group, tool, or skill (applies immediately)
• Enter — collapse/expand the selected group
• Ctrl+S — save current selection as the global default (~/.pi/agent/loadout.json)
• ↑ / ↓ or J / K — navigate
• Esc — close the picker (live changes are kept and persisted to the session branch)

Why use /loadout?

Pi enables every installed tool and skill by default. As your package list grows, that becomes harder to manage. /loadout gives you grounded control over what each session can do:

• Hide skills you don't want auto-suggested, keep them callable manually. Disabled skills are removed from the model's <available_skills> system prompt block, but /skill:name invocation still works. Useful when a skill is helpful occasionally but distracting in normal turns.
• Disable tools you don't want the model to reach for in this session without uninstalling the package that provides them. Saved selection persists in the current session branch and is restored when you resume.
• Run Pi on smaller or cheaper models by trimming the tool definitions and skill block down to what the task actually needs.
• Debug extensions by toggling individual tools or whole packages on and off without an install/reinstall cycle.
• Set a sensible default once. Ctrl+S in the picker writes the current selection to ~/.pi/agent/loadout.json, which seeds new sessions until you override it per-branch.

How it works

• Live apply. Each toggle in the picker calls Pi's active-tool API immediately. There is no separate "save" step inside the picker for the session; changes are in effect as you make them.
• Branch-persisted on close. When the picker closes (Esc, or selecting a confirmation), the current selection is appended to the session branch as a custom entry (pi-loadout:selection). Resuming or branching the session restores it.
• Global default (Ctrl+S). Writes the current selection to ~/.pi/agent/loadout.json. New sessions without their own branch entry read this file.
• Skill filtering. The skill list shown to the model is filtered in before_agent_start by replacing the <available_skills> block. Explicit /skill:name invocations are unaffected.
• Session log entries. Loadout changes are logged as visible session messages (pi-loadout:loadout changed) for resume/export readability. These log entries are filtered out of context, compaction, and tree summarization, so they don't pollute model input.
• Status bar. While installed, the footer shows N/total tools · N/total skills for the current session.

Group rows

• ● enabled — every item in the group is enabled
• ◐ partial — some items in the group are enabled
• ○ disabled — no items in the group are enabled

Toggling a group toggles every item in that group. Toggling an item affects only that item.

Caveats

• Prompt-cache miss on change. Changing tool definitions or available skills invalidates provider prompt caches. The next response after a loadout change is typically slower and more expensive. If you toggle frequently within a session, expect repeated cache misses.
• Mid-session changes are silent to the model. Loadout log entries are filtered out of the model's context. If a tool disappears mid-session the model may still attempt to call it on the next turn before adapting.
• No named profiles yet. Only one global default and one branch override. Multiple named loadouts (e.g. coding, review) are not implemented; see the issue tracker.

License

MIT