v1.0.0-alpha
Pre-releaseKERNEL v1.0.0-alpha
Self-Evolving Project Intelligence for Claude Code
The first release of KERNEL - a Claude Code plugin that progressively builds your project configuration based on observed patterns instead of upfront speculation.
What's Included
Core Commands
| Command | Description |
|---|---|
/kernel-init |
Initialize KERNEL in any project - analyzes stack, detects tier, creates customized CLAUDE.md |
/discover |
Map codebase, find tooling, extract conventions, populate state.md |
/plan |
Get-it-right-first-time planning methodology |
/debug |
Systematic diagnosis and root cause fixing |
/review |
Correctness, consistency, completeness validation |
/kernel-status |
Show config health and staleness report |
/kernel-prune |
Review and remove stale config entries |
/handoff |
Generate context handoff brief for session continuation |
/parallelize |
Set up git worktrees for parallel development |
Knowledge Banks
Token-optimized methodology templates (~500-800 tokens each):
- PLANNING-BANK.md - Mental simulation before execution, interface definition first
- DEBUGGING-BANK.md - Reproduce → Isolate → Instrument → Root cause → Fix → Verify
- DISCOVERY-BANK.md - Reconnaissance before action, map terrain, extract conventions
- REVIEW-BANK.md - Correctness checks, convention adherence, invariant validation
Banks are NOT loaded by default (zero token cost until referenced via modes).
Skills
Lightweight capability pointers (~50-100 tokens each):
- worktree-parallelization - Detect tasks suitable for git worktree parallel development
Rules Templates
Starter templates for project-specific patterns:
preferences.md- Formatting, presentation, tool choicesassumptions.md- Extract assumptions before executinginvariants.md- Non-negotiable contractspatterns.md- Discovered coding patternsdecisions.md- Architecture decision log
Configuration Reference
- CONFIG-TYPES.md - Complete guide on when to use Agents vs Skills vs Commands vs Rules vs Hooks vs MCP (decision tree, comparison table, anti-patterns)
Sample Project
Included sample-project/ with a TaskMgr CLI demonstrating KERNEL in action with:
- CSV export functionality
- Database optimization
- Encryption support
- API sync capabilities
- Full test suite
Installation
Via Plugin Menu (Recommended)
- In Claude Code, run
/plugin - Navigate to Marketplaces tab
- Select + Add Marketplace
- Enter:
https://github.com/ariaxhan/kernel-plugin - Go to Discover tab and enable the KERNEL plugin
Programmatic
Add to ~/.claude/settings.json:
{
"extraKnownMarketplaces": {
"kernel-marketplace": {
"source": { "source": "github", "repo": "ariaxhan/kernel-plugin" }
}
},
"enabledPlugins": { "kernel@kernel-marketplace": {} }
}How It Works
Philosophy
- CORRECTNESS > SPEED - One working implementation beats three debug cycles
- INVESTIGATE BEFORE IMPLEMENT - Search for existing patterns first
- DETECT, THEN ACT - Don't assume tooling exists, find it
- PROTECT STATE - Backup before mutation, confirm before deletion
The Bank Architecture
Instead of loading 6000+ tokens of coding rules upfront, KERNEL uses:
- Banks (~500-800 tokens) - Methodology templates with slots to fill
- Skills (~50-100 tokens) - Lightweight pointers, auto-discovered by context
- Commands (0 tokens until used) - On-demand workflow invocation
- State (~200 tokens) - Single source of truth, updated by modes
Total baseline cost: ~330 tokens (vs ~6000+ traditional approach)
Pattern → Artifact Mapping
When patterns emerge during work, KERNEL creates the appropriate Claude Code artifact:
| Pattern | Artifact Created |
|---|---|
| Repeated workflow | .claude/commands/workflow-name.md |
| Specialized task | .claude/agents/specialist-name.md |
| External service | Entry in .mcp.json |
| Pre/post processing | Hook in .claude/settings.json |
| Domain capability | .claude/skills/capability.md |
| Explicit preference | .claude/rules/topic.md |
Requirements
- Claude Code CLI v1.0.33+
- Python 3.8+ (optional, for Claude Docs MCP server)
Platform Support
Works on macOS, Linux, and Windows.
Known Limitations
extraKnownMarketplacesonly works in interactive mode (not CI/headless)- Banks are project-agnostic templates - fill slots as you learn your codebase
- Config registry tracking (
memory/config_registry.jsonl) requires manual bootstrap for existing configs
What's Next
Planned for future releases:
- More MCP server specific support
- Mathematical compacting of context powered by ARBITER (https://github.com/ariaxhan/arbiter), a propositional logic validation and compression library.
- Surprises
Contributing
Contributions welcome. Read CONFIG-TYPES.md before creating artifacts and follow the philosophy in CLAUDE.md.
License
MIT
Author: Aria Han
Repository: https://github.com/ariaxhan/kernel-plugin