Overview

ToolboxMD/karpathy-wiki is a background auto-capture + detached ingest implementation. It is a Tier 4 starter template in the comprehensive report — focused on removing the manual “drop the source” step via background capture.

Key Facts

  • Feature: background auto-capture + detached ingest
  • Role: removes the human drop-step from the ingest pipeline
  • Architecture: background service with async processing
  • Part of the third-wave specialized implementations

Features

Core Architecture

  • Four-skill split — loader, capture, read, ingest; auto-loaded via SessionStart hook
  • Detached background ingest — sources land as markdown in .wiki-pending/, ingester spawns in background
  • Discovery-driven categories — any top-level mkdir <name>/ creates a category (no code changes)
  • Per-directory _index.md tree — recursive; root index.md is a small MOC

Ingest & Processing

  • Auto-capture — sources flow in without human triggering each ingest
  • Deep orientation (steps 1-9) in ingester — comprehensive source analysis
  • Raw-direct ingest — drop file into <wiki>/inbox/, next SessionStart ingests directly
  • Per-wiki .ingest-runs.jsonl — run history, fork-asymmetry detection
  • Ingest issues — surfaced to .ingest-issues.jsonl, displayed by wiki issues/wiki status

Search & Retrieval

  • Read-from-wiki protocol — 6-step orientation ladder (orient → candidate count → inline/subagent/web search → cite)

Quality & Governance

  • Validator — enforces frontmatter fields, type/path consistency, depth limits
  • Tier-1 lint — frontmatter, link resolution, source existence, quality blocks

MCP / Tools

  • CLI commandswiki status, wiki capture, wiki ingest-now, wiki issues, wiki use, wiki init-main
  • Project-wiki auto-resolutionwiki-resolve.sh picks right wiki per cwd (project vs main)

Integration

  • Multi-platform hooks — Claude Code, Cursor, Copilot CLI, OpenCode
  • Agent-compatible — works alongside CLI agents
  • Markdown output — standard LLM Wiki format

Data & Storage

  • Markdown files — standard wiki output
  • File-based — no database required

Relationships

  • llm-wiki — (type: belongs-to) implementation with automated capture
  • ingest — (type: enables) its detached ingest is an automation of this operation
  • ai-memex-cli — (type: sibling-of) sibling auto-capture approach (session distillation)
  • vault-roadmap — (type: belongs-to) maps to the Automation phase

Implications

ToolboxMD directly addresses the user’s “too manual” complaint: background auto-capture means sources flow in without the human triggering each ingest.

Sources