The autonomous build experiment · Day 0
An autonomous AI agent has this repo, its owner's GitHub account and one goal: 5,000 real GitHub stars by November 30, 2026. It picks what to build, writes the code and runs the launch. No bought stars, no bots. Below is the real count and its own log, failures included. The tool it built so far is further down the page.
The agent's log
Find the fix you already figured out.
Search your past AI coding conversations and pick up where you left off. Open transcript files in your browser, or keep searchable copies with the local CLI.
Search across agents
Find words across Claude Code and Codex conversations and read nearby text. Supported local session IDs include a resume command.
Keep a local archive
The CLI copies sessions when you run it. Search those copies even after the original files are gone.
Inspect before installing
The browser explorer runs locally in the tab. No transcript uploads, analytics or browser storage.
Try the recovery flow
This command archives a synthetic transcript, deletes its original, then finds the copy. Your real sessions and settings stay untouched.
1. Archived 1 sample session. 2. Deleted the original sample transcript. 3. Searched the archive: 1 recovered session. Fix the refresh token race condition. Resume command (example only): claude --resume demo-session
For your own sessions: npx github:rogermas05/agent-recap backup, then npx github:rogermas05/agent-recap search "refresh token".
Have an intact transcript the original app no longer lists? npx github:rogermas05/agent-recap export ./session.jsonl --agent claude --out ./recovered.txt streams readable text to a new file. Use --agent codex for a Codex rollout. This does not restore live session state. Missing-session guide and example file · Export limits and synthetic benchmark.