15:I[981,["658","static/chunks/658-a698693f56d0116a.js","183","static/chunks/183-6beeeb801c274aaa.js","474","static/chunks/app/%5Btype%5D/layout-631c4730b83c8223.js"],"Posts"] 16:T521,Built this to manage the full Claude Code workflow autonomously, from initial plan creation to reviewed PR.

Tested it on a 25-task plan (5-7 items each) - ran overnight, woke up to a surprisingly solid PR-ready branch.

Just run `ralphex` (single binary) in your project directory.

Without a plan file, it starts interactive mode. Asks what to build, explores the codebase, writes the plan, then executes.

The loop: plan creation (asks questions, writes markdown) → task execution (fresh session each, no context bloat) → multi-phase reviews (Claude + optional Codex) → branch ready for PR.

Prompts for each phase are injected with minimal context needed for continuation: plan file, progress log, goal. Progress file accumulates Q&A during planning, execution logs during tasks, review findings between iterations. Output streams to terminal and progress log file simultaneously.

Everything customizable at user level (~/.config/ralphex/) or per-project (.ralphex/): prompts, agents, colors.

Web dashboard monitors progress files across directories. See active sessions, stream output live, spot failures without switching terminals.17:T4fc,Hi HN,

I’m a solo engineer working on a project that came out of frustration with traditional camera systems.

Most cameras answer one question well: “Did something move?” But many practical needs are more semantic:

Is the door open?

Is someone inside the room?

Is a package still there?

Did the light turn off?

So I built a system where you define a natural-language question, connect a camera (webcam / RTSP / edge device), and get notified only when the answer becomes “yes.”

Some high-level design choice

Show HN: I Gave Claude Code a Job as a Quant Researcher