Pratiyush/llm-wiki

Pratiyush/llm-wiki (218 stars) is a Python CLI tool that auto-compiles coding agent session transcripts into a Karpathy-style LLM Wiki + static HTML site. It bridges the gap between agent session history and persistent wiki knowledge.

Features (confidence: high)

  • Multi-agent support: Detects and ingests from Claude Code, Codex CLI, Cursor, Copilot Chat, Gemini CLI, and Obsidian
  • Session auto-detect: Looks for agent session stores at standard paths
  • Privacy redaction: Strips usernames, API keys, tokens, emails from transcripts
  • SessionStart hook: Optional auto-sync on every claude code launch
  • Three-layer output:
    • raw/sessions/ — immutable markdown transcripts
    • wiki/ — LLM-maintained interlinked pages (index, log, sources, entities, concepts)
    • site/ — static HTML with dark mode, global search, keyboard shortcuts
  • 14 lint rules: Wiki quality checks
  • Graphify AI knowledge graph: Communities, god nodes, hyperedges
  • Approval workflow: Candidate pages with list/promote/merge/discard
  • 11 CLI commands: init, sync, build, serve, adapters, graph, query, export, lint, candidates, synthesize

Architecture (confidence: high)

Three layers per Karpathy’s spec, plus AI-consumable exports (llms.txt, JSON-LD, per-page .txt + .json siblings). Deterministic build with Python’s markdown library as the only runtime dependency. 2,329 tests.

Implications

This is the most production-ready session-to-wiki pipeline in the ecosystem. Unlike most implementations that require manual ingest, Pratiyush auto-discovers agent sessions and converts them continuously. The approval workflow (candidates → promote → merge) solves the quality gate problem that manual wikis face. Its 2,329 tests make it the most rigorously tested LLM Wiki implementation.

Sources

  • raw/articles/pratiyush-llm-wiki.md