GH·RADAR

Score how far an AI agent's tool permissions drift from least privilege.

Compare three deterministic ways to shrink an agent trace and see which keeps it replayable.

Catch dangerous CORS policy combinations before they ship — from the config, not the wire.

Find the crontab lines that will bite you someday — statically.

Run AI coding agent CLIs (claude, codex, ...) sandboxed with bubblewrap, with logins separate from session history: any account on any history, usage limits at a glance. One Python file.

Grade how expensive a pull request is to review — before a human opens it.

Claude Code plugins from Akal Software — SupaForge: find and fix drift between Supabase environments

Catch primary-key drift before bad IDs reach your database.

Turn a CODEOWNERS file and a file list into a path-by-owner map — and see where ownership is broken.

Score how hard it would be to walk away from each proprietary dependency you run.

Reusable Python CLI template: streaming file checksums, tests, CI, and Dependabot.

Contract-driven agent execution with durable replay and signed evidence.

Migrate Vue 2 to Vue 3 with one command. Finds every breaking change file by file, auto fixes the safe ones and prints the @vue/compat flags you need. npx vueleap

Native Android companion for multiple Herdr connections on Windows. SSH, terminal control, and bigger fonts. GPLv3.

A browser cockpit for Claude Code: one main session, every subagent beside it with steer and stop, and everything they did.

Email for AI agents — its own address, not your Gmail. MCP client configs and examples for Atmark (invite-only).

Python service that reports a suspected leaked key, rotates a temporary replacement, and searches logs to summarize blast radius.

Agent skills for making cited, data-driven maps with the ShowMeOnMap MCP server

Checks saved answers of a generative system against an evaluation set, as plain string checking. Reads JSONL or CSV cases and lists missing facts, sources outside the allowed set, forbidden phrases, leaked secrets and personal data, with a pass rate per category a build can fail on. Never calls a model.

General household-finance skill for Hermes Agent — Obsidian vault, append-only JSONL ledger, pluggable bank-email adapters, verification gates, month close.

A reality-first loop for work that must be answered for: a stdlib-only CLI and a Claude Code skill over one roadmap, with the manifesto that states the practice.

Specification-driven curriculum: specs, black-box tests, and templates for eleven cross-platform CLI projects.

A validator that keeps your Markdown release notes honest.

Catch broken invoice JSON before it reaches your ledger — with arithmetic, not AI. The validator checks required fields, date and currency formats, and the amou