Multi-Tool AI Workflow Generator

Generate skills and native agents for multiple coding assistants from one shared template set.

Run kernel init once to detect supported tools in the current project, write global kernel configuration, and emit the files each tool expects.

---

Quick Start Guide

Installation

npm install -g @hackefeller/kernel
# or
bunx kernel init

Quick Start

1. Initialize Your Project

# Navigate to your project
cd your-project

# Initialize with auto-detection
kernel init --yes

# Or specify tools manually
kernel init --tools opencode,claude,cursor

2. Use The Installed Workflows

In your AI tool, invoke the generated workflows and agents for planning, exploration, execution, and review.

3. Update Generated Files

# Regenerate all files
kernel update

# Update specific tool
kernel update --tool cursor

Configuration

Initialization creates ~/.kernel/config.yaml. The full supported schema is documented in Configuration.

Supported Tools

Supported output targets currently include 6 AI coding tools:

| Tool | ID | Directory | | -------------- | ---------------- | ------------ | | OpenCode | opencode | .opencode/ | | Claude Code | claude | .claude/ | | OpenAI Codex | codex | .codex/ | | GitHub Copilot | github-copilot | .github/ | | Google Gemini | gemini | .gemini/ | | Cursor | cursor | .cursor/ |

Example Workflows

Starting New Work

1. Use the propose workflow to define the change
2. Use the explore workflow if you need to research
3. Use the apply workflow to implement
4. Use the archive workflow when done

Code Review

1. Invoke the review workflow or review agent
2. The generated review system will analyze and provide feedback

Performance Optimization

1. Use the planning or execution workflow to scope the optimization
2. Delegate to the relevant specialist or skill

Troubleshooting

No tools detected

# Check what tools are available
kernel detect

Config issues

# View current config
kernel config show

Regenerate files

# Force regeneration
kernel update --force

Next Steps

• Review the CLI Reference below

---

The Core Workflow

propose → explore → apply → archive — all powered by Linear via Linear MCP.

| Command | What it does | | ---------- | ------------------------------------------------------------------------------------- | | /propose | Create a Linear project with top-level issues and sub-issues for implementation tasks | | /explore | Read Linear project state, explore tradeoffs, write decisions back to Linear | | /apply | Pick up the next unblocked sub-issue from Linear, implement it, update Linear | | /archive | Close out the Linear project, finish remaining issues, report deferred work |

Linear is the source of truth. No local proposal.md files. No sticky notes. No "I'll just jot this down in a text file." The issue tracker IS the plan.

Releases

Releases are driven by Changesets.

1. Add a changeset file in .changeset/ for any change that should ship in a release.
2. Commit the changeset with the code it describes.
3. When changesets land on main, GitHub Actions opens a version PR that updates package.json and the changelog.
4. Merging that version PR triggers the release workflow, which tags the repo and creates the GitHub release.

---

CLI Reference

init

Initialize a project. Detect installed tools, write ~/.kernel/config.yaml, and generate the configured output for the current repository.

kernel init [options]

  -t, --tools <tools>      Tool IDs to configure: opencode, claude, codex, github-copilot, gemini, cursor, "all"
                           (default: all detected)
  -p, --profile <profile>  Profile: core | extended  (default: core)
  -d, --delivery <mode>    What to generate: skills | both  (default: both)
  -y, --yes                Use all detected tools without prompting
      --path <path>        Project path (default: current directory)

Delivery modes:

| Mode | Skills | Native Agents | | ---------------- | ------ | ------------- | | both (default) | Yes | Yes | | skills | Yes | No |

When you choose both, the generator emits skills for every configured tool and native agents only for tools that support them. Skills-only tools such as Cursor still receive the full skill set.

update

Regenerate all files from the current templates. Use it after editing configuration, changing templates, or refreshing generated output.

kernel update [options]

  -f, --force              Regenerate everything, no questions
  -t, --tool <tool>        Update a single tool only
      --path <path>        Project path (default: current directory)

config

View and modify ~/.kernel/config.yaml.

kernel config [action] [key] [value]

  show                    Print current config
  add-tool <tool-id>      Add a tool
  remove-tool <tool-id>   Remove a tool
  set <key> <value>       Set any config value

detect

Scan the project for supported tool directories and report which ones are present.

kernel detect [--path <path>]

vault compile

Compile your personal vault skills into each configured AI tool's native format.

The vault is a directory, often ~/.codex/skills/, containing skills to reuse across projects. Each skill is a SKILL.md with an optional references/ subdirectory. kernel vault compile reads that vault and writes tool-native output into the current project for every configured tool.

kernel vault compile [options]

  -v, --vault <path>      Vault root (overrides vaultPath in config; ~ supported)
  -t, --tools <tools>    Tools to compile for (default: all configured tools)
      --dry-run           Preview what would be written

Vault structure:

~/.codex/skills/
  my-secret-weapon/
    SKILL.md              # YAML frontmatter + markdown body
    references/
      context.md          # Supporting docs (loaded automatically)
      patterns.md

The source material stays in one place; the command handles the per-tool translation.

---

Configuration

Kernel reads global configuration from ~/.kernel/config.yaml.

The minimal valid config is:

version: "1.0.0"
tools:
  - opencode
profile: core
delivery: both

The full accepted schema is:

version: "1.0.0"
tools:
  - opencode
  - claude
profile: core
delivery: both
customWorkflows:
  - propose
featureFlags:
  example: false
vaultPath: ~/.codex
metadata:
  name: my-project
  description: Important work
  version: "0.1.0"

| Field | Required | Allowed values | Current behavior | | ----- | -------- | -------------- | ---------------- | | version | no | any string, defaults to "1.0.0" | Schema version marker. Parsed and preserved. | | tools | yes | opencode, claude, codex, github-copilot, gemini, cursor | Drives detection, adapter selection, generation, update, and vault compile. Must contain at least one valid tool ID. | | profile | no | core, extended, custom; defaults to core | Accepted and shown by the CLI, but generation currently uses the same built-in skill and agent catalog for every profile. | | customWorkflows | no | array of strings | Accepted by the schema and saved if present, but not currently consumed by generation. | | delivery | no | skills, both; defaults to both | Active. skills skips native agent generation. both generates skills plus native agents for tools that support them. | | featureFlags | no | map of string keys to boolean values | Accepted and saved, but not currently consumed elsewhere in the codebase. | | vaultPath | no | path string | Active only for kernel vault compile when --vault is not passed. ~ expansion is supported there. | | metadata | no | object with optional name, description, version | Accepted and saved, but not currently used by generation. |

Defaults

If you omit optional fields, kernel defaults to:

version: "1.0.0"
profile: core
delivery: both
featureFlags: {}

You still must provide at least one entry in tools.

Tool IDs

Use these IDs in tools, kernel init --tools, kernel update --tool, and kernel vault compile --tools:

| Tool | ID | Detected by | | ---- | -- | ----------- | | OpenCode | opencode | .opencode/ | | Claude Code | claude | .claude/ | | OpenAI Codex | codex | .codex/ | | GitHub Copilot | github-copilot | .github/ | | Google Gemini | gemini | .gemini/ | | Cursor | cursor | .cursor/ |

Runtime Notes

• kernel init writes ~/.kernel/config.yaml from the selected tools plus profile and delivery.
• kernel config show, kernel update, and kernel vault compile all read ~/.kernel/config.yaml.
• kernel config add-tool, remove-tool, and set write raw YAML updates without re-validating the full schema.
• The deprecated commands delivery mode is not part of the supported schema and should not be used.

---

Supported Tools

Generated output is supported for 6 AI coding tools:

| Tool | ID | Directory | Notes | | -------------- | ---------------- | ------------ | ----------------------------------------------------- | | OpenCode | opencode | .opencode/ | Native agent files plus skills discovery manifest | | Claude Code | claude | .claude/ | Native agent format with skills: preloading | | OpenAI Codex | codex | .codex/ | TOML agent format with `[[skills.config]