Logex

logex

Write blog-style session papers from coding-agent session transcripts (Claude Code, Codex, Pi, DSH). One session → N articles, one per topic. The LLM (your agent) decides topic segmentation and writes the articles in-session — no extra API key needed.

- English · 中文 · 日本語 · 한국어 · Español

English

What is logex

logex turns coding-agent session JSONL transcripts (Claude Code, Codex, Pi, DSH) into blog-quality technical articles. It does the boring work (parse, chunk, score, group by signal) as a pure pipeline, then hands the actual topic segmentation and writing to the LLM that's already in the loop. One long session typically yields 1–3 articles — the agent reads chunk summaries, decides what's worth writing, and produces structured JSON articles that get published to a separate data repo (logex-data).

The webapp (React + Vite SPA) reads that data repo and renders the blog at logex-io.vercel.app.

Install

Requires Node.js 18+.

Quick Start — the /logex workflow

In Claude Code, just run:

Under the hood the skill runs:

The agent is the one deciding topic boundaries and drafting the prose — logex is the scaffolding around it.

Supported transcripts. The parser reads Claude Code, Codex, Pi, and DSH (DeepSeek Harness) session JSONL. Auto-discovery (logex list) scans ~/.claude/projects/, ~/.codex/sessions/, ~/.pi/agent/sessions/, and ~/.dsh/sessions/; zstd-compressed DSH transcripts (.jsonl.zstd) are decompressed transparently (needs the zstd binary). You can also pass any JSONL path directly: /logex .

Claude Code integration (plugin + skill + hook)

Install the plugin once, get the skill and the Stop-hook reminder automatically:

This registers:

- Skill logex — the /logex slash command and the full segmentation + writing procedure.
- Stop hook — on session…