writing

A live terminal dashboard for Claude Code sessions — per-session spend, tool activity, and risk — that can pause or kill a runaway one. 100% local, reading the transcripts already on your disk.

I built agentop — a terminal dashboard for coding-agent sessions — by having one agent write each task and a second, adversarial agent review it before the next task starts. Here are the real bugs that pipeline caught.

As agents run longer and more autonomously, the missing piece isn't a smarter model — it's a thin layer that watches a run and decides when to continue, pause, or kill it. Two tools I built, and the thesis underneath them.

Why I built a download CLI that auto-dispatches across HTTP, cloud, git, media, and torrents — and how the native parallel engine with per-chunk resume actually works.

How I built a dependency graph over MCP server tool catalogs — heuristic + LLM extraction, cross-server edges, and a cartographic SVG/dagre visualization.

The Patroni / Keepalived / GlusterFS patterns I used to build a multi-layer HA setup for APIs, databases, and Airflow.