pi-kumite

Pi package for the Kumite project setup and review workflow.

Packages

Package details

extensionprompt

Install pi-kumite from npm and Pi will load the resources declared by the package manifest.

npm repo home report

$ pi install npm:pi-kumite

Package: pi-kumite
Version: 0.1.0
Published: Jun 9, 2026
Downloads: not available
Author: onishell
License: MIT
Types: extension, prompt
Size: 133.4 KB
Dependencies: 1 dependency · 1 peer

Pi manifest JSON

{
  "extensions": [
    "./extensions/kumite/index.js"
  ],
  "prompts": [
    "./prompts"
  ]
}

Security note

Pi packages can execute code and influence agent behavior. Review the source before installing third-party packages.

README

pi-kumite

Pi package for the Kumite workflow.

Kumite is the host CLI and installer for setting up the Pi agent coding harness. This package adds the Pi-native entrypoints that let Pi initialize a project, check Kumite installation, and run the Kumite setup flow from inside a trusted Pi session.

Install

From this local checkout, run the install command from the project that should load the extension:

cd /path/to/your-project
pi install /home/dev/projects/pi-kumite -l

From npm after publishing:

pi install npm:pi-kumite

Commands

/kumite-doctor
/kumite-init
/kumite-continue
/kumite-next
/kumite-models
/kumi-models

/kumite-init writes .pi/settings.json for the target project so future Pi sessions load the Kumite extension, prompts, and skills from the same package.

/kumite-init also upserts agents.md, the curated agent entry point and index for project memory. It links to deeper memory files instead of duplicating them. A tiny root AGENTS.md bridge is written only because Pi auto-loads uppercase AGENTS.md; it points agents back to canonical agents.md.

/kumite-continue resumes a stalled Kumite loop by asking the extension for the next workflow step and then launching the returned subagent, for example reviewer after an implementer handoff. If the planner has produced a draft plan for grill, it pauses for parent-session questions before final Gherkin/plan writing.

/kumite-next only displays the next workflow step decision for diagnostics.

/kumite-models and /kumi-models open a Pi-native selector for project-local Kumite subagent routing. They write .pi/kumite/models.json and patch model: / thinking: frontmatter into the generated .pi/agents/{orchestrator,scout,planner,planner-fallback,implementer,reviewer,curator}.md files. Use inherit to make an agent use the active/default Pi session model.

Agents can also apply routing non-interactively with the kumite_apply_agent_models tool.

Requirements

pi must already be installed.
The kumite binary should be on PATH for commands that execute setup/init.
Install Kumite first from the main repository when needed:

curl -fsSL https://raw.githubusercontent.com/OnishellT/kumite/main/install.sh | sh