Kernel

kernel is a local-first brain and workflow OS for coding agents.

It gives you one canonical place on your machine to define skills, agents, and commands, then syncs that brain into the dot-directories your agent hosts already use.

It also gives each repo a lightweight local workflow system under kernel/work/ so planning and execution stay visible, durable, and tool-agnostic.

what i

• Stores your canonical agent brain in ~/.kernel/brain/
• Syncs that brain into enabled hosts like .codex, .claude, .copilot, .pi, and OpenCode
• Keeps host-specific formatting at the edge through small materializers
• Manages local work items in kernel/work/<id>/
• Avoids remote SaaS dependencies and host-specific authoring

Layout

User brain

~/.kernel/
  config.yaml
  brain/
    skills/<id>/SKILL.md
    agents/<id>/AGENT.md
    commands/<id>.yaml
  state/
    sync-manifest.json

Repo workflow

kernel/
  work/<id>/
    work.yaml
    brief.md
    plan.md
    tasks.md
    journal.md

CLI

kernel init
kernel sync
kernel doctor
kernel host list
kernel work new "<goal>"
kernel work plan [id]
kernel work next [id]
kernel work status [id]
kernel work done <task>
kernel work archive [id]

Quick Start

npm install -g @hackefeller/kernel
kernel init
kernel host list
kernel work new "build analytics dashboard"
kernel work next

When To Use What

System

• kernel-init: first-time setup for the local Kernel home and host integration
• kernel-sync: push the current catalog into enabled hosts after template or configuration changes
• kernel-doctor: diagnose drift, missing generated files, or host setup problems

Workflow

• kernel-work-new: start a new repo-local work item from a natural-language goal
• kernel-work-plan: tighten the brief, plan, tasks, risks, and acceptance criteria
• kernel-work-next: pick the next unchecked task as the single execution target
• kernel-work-done: mark a verified task complete
• kernel-work-status: inspect current progress and the next task
• kernel-work-archive: close and preserve a completed work item

Specialist

• gh-pr-errors: inspect the latest GitHub Actions failure on the open pull request

Design Principles

• Define once locally, sync everywhere
• Local files are the source of truth
• Host-specific behavior stays in small adapters/materializers
• Tags categorize skills and agents for discovery
• Local work state should be visible in the repo, not hidden in chat

Notes

• The canonical workflow surface is kernel work; older spec and change command families are no longer shipped.
• kernel init seeds the built-in brain, imports existing ~/.codex/skills content when present, and syncs enabled hosts.
• kernel sync is idempotent and tracks generated output in ~/.kernel/state/sync-manifest.json.