🌳 ai-devlog

Turn your AI coding chats into a living idea tree

Your sessions with Claude Code, Codex, Cursor & co. hold every decision you made — buried in thousands of messages. ai-devlog digs them out and grows them into one interactive, self‑contained HTML map of what you built and why.

Project at the center → categories → ideas → 💡 AI insights. Click any node for the full prompt, what the AI did, the diff, and the commit. Export to Markdown/JSON, switch EN ⇄ 中, light/dark.

🤔 Why

Traditional diffs tell you what changed; they lose why. Your chat history has the why — but it's an endless scroll. ai-devlog reads it, throws away the noise, and turns it into a tree you can actually navigate.

把 AI 编程聊天变成项目的可审计「想法树」：每个想法、每次分叉、每个实现都能追溯。

✨ Features

• 🌐 Auto‑discovers your Claude Code & Codex sessions — no file hunting.
• 🧹 Cuts the noise — drops tool‑results, IDE notifications and injected reminders, so only the prompts you actually typed become nodes.
• 🧠 LLM‑organized — groups the whole project into categories → ideas → sub‑ideas, nesting related ideas across conversations and time. Runs on your Claude subscription (the claude CLI) — no API key.
• 💡 AI insights — pulls out the genuine findings, decisions & recommendations the AI gave you (incl. from its web research), not the busywork.
• 🔗 Git‑correlated — folds the real commit + diff into the idea it belongs to.
• 🌳 Radial canvas — pan/zoom mind‑map that fans out in every direction.
• 🌏 Bilingual (EN ⇄ 中) and 🌙 light/dark, one click each.
• ⬇ Export the whole tree to Markdown or JSON.
• 📦 One file, zero deps, offline — share an index.html, open it anywhere. Nothing is uploaded.

🚀 Quick start

No install, no API key — just Node 18+.

# see it instantly with sample data
npx ai-devlog demo
# → opens ai-history-export/index.html

# …or on YOUR project: discover sessions + correlate git + LLM-organize, in one go
cd /path/to/your/project
npx ai-devlog auto --git --summarize

Or install it: npm install -g ai-devlog → then run ai-devlog anywhere.

🎨 What you get

A single self‑contained index.html — the whole UI is the idea tree:

| | | |---|---| | 🌳 Radial tree | drag to pan, scroll to zoom; the project sits at the center and ideas fan out in all directions, color‑coded by type | | 🖱️ Click a node | a drawer shows the full prompt and what the AI did — its answer, files changed, red/green diff, and the git commits that landed | | 🔎 Search | filter the tree to matching ideas/prompts/code and their ancestors | | 🌏 / 🌙 | language toggle (EN ⇄ 中) and light/dark mode |

🧩 How it works

chats ──▶ normalize ──▶ filter to real prompts ──▶ LLM: classify + nest + insights ──▶ radial HTML

1. Ingest sessions from each source and normalize them to a common message stream.
2. Filter — keep only genuine human prompts (Claude Code's type:"user" is mostly tool‑results & injected context; those are dropped).
3. Organize — the LLM classifies the project into 3–7 categories, nests each idea under the one it relates to, writes a bilingual title, and extracts the AI's insights. Without the LLM, a chronological keyword heuristic is used.
4. Render — everything inlines into one offline HTML tree.

| --source | Input | Auto? | |---|---|:--:| | claude-code | Claude Code *.jsonl (~/.claude/projects/…) | ✅ | | codex | Codex rollout-*.jsonl (~/.codex/sessions/…) | ✅ | | chatgpt | ChatGPT export conversations.json | — | | aider | .aider.chat.history.md | — | | markdown | any markdown with ## User / ## Assistant headings | — | | generic | JSON in ai-devlog's own schema | — |

node ai-devlog.mjs discover            # list this project's local sessions
node ai-devlog.mjs auto                # import them all + build the HTML
node ai-devlog.mjs auto --all          # every project on this machine
node ai-devlog.mjs import --source chatgpt conversations.json   # web exports stay manual

node ai-devlog.mjs auto --git --summarize     # one shot
# or, on an existing store:
node ai-devlog.mjs summarize                  # arrange the whole project
node ai-devlog.mjs summarize --refresh        # re-arrange from scratch

• Drives the headless Claude Code CLI (claude -p) — your subscription, no API key, no SDK.
• Classifies into 3–7 categories, nests ideas by meaning across conversations, and writes bilingual titles phrased as ideas (not commands).
• Insights use an include/exclude codebook (an insight answers "what does this mean / why it matters", not "what was done") and read the head and tail of each response so research conclusions aren't missed.
• Default model haiku (override with --model); cached in the store. This is the only step that sends data off your machine.

node ai-devlog.mjs auto --git                 # correlate commits while building
node ai-devlog.mjs scan-git --since "30 days ago"   # then: export

Each commit (with its diff) is attached to the nearest preceding idea by time (window configurable via --window <hours>). Commits with no nearby chat are grouped under an Unlinked commits node — nothing is silently dropped.

🔒 Privacy

Local‑first and read‑only: ai-devlog reads files you point it at and writes a static folder. Nothing is uploaded — except the optional summarize step, which sends idea text to your own Claude subscription. Redact secrets before sharing an export.

🛠️ Run it as a command

npm install -g ai-devlog     # then: ai-devlog auto
# or no install at all:
npx ai-devlog auto

🗺️ Roadmap

Cursor SQLite reader · richer insight ranking · lazy‑rendered tree for huge histories · MCP server exposing the history as queryable resources.

🧱 Project layout

ai-devlog.mjs        CLI (auto · discover · scan-git · summarize · import · export · demo)
lib/discover.mjs     find local Claude Code + Codex sessions
lib/git.mjs          read git commits/diffs for correlation
lib/summarize.mjs    LLM categories + hierarchy + insights, via the claude CLI
lib/parsers.mjs      source → normalized messages
lib/tree.mjs         messages → project-wide idea tree
lib/exporter.mjs     tree → single self-contained HTML
template/            index.html · style.css · app.js  (SVG radial canvas, inlined on export)

📄 License

MIT · made with Claude Code