nebi-mod-verifylisted
Install: claude install-skill romanized/nebi-skill
# Mod verify: check what the screen actually shows
`claude plugin test` checks the tree a hook returns, never the painted screen. This skill runs a real session, captures the screen and the debug log, and checks both. It is step 5 of the done bar in `<repo-root>/mods/CONTRACT.md`.
## Run
Scripts live in this skill's `scripts/` folder (deployed under `~/.claude/skills/nebi-mod-verify/scripts/`).
1. **Paint check, mods as installed:** `node scripts/capture.mjs --label after`. It takes 1 to 3 minutes. The first run installs two pinned npm packages into `~/.cache/nebi-mod-verify` and needs the network once. It recreates a demo git repo there, starts `claude` with `--debug-file` in a 150x46 pseudo-terminal, captures two fixed scenes (turn end on a ❓ ask, and mid-turn during a slow Bash call), exits with `/exit` so SessionEnd hooks run, renders PNGs in Nebi's light Ghostty theme, then runs [check.mjs](scripts/check.mjs). Exit 0 means every check passed.
2. **Stock baseline:** `node scripts/capture.mjs --safe-mode` (label `baseline`). No CLAUDE.md, hooks, status line, skills or installed mods. It is recorded, never asserted.
3. **Compare:** `node scripts/check.mjs <run folder> --baseline <baseline folder>` prints each string as baseline -> run.
4. **Add what the change promises:** `--present "TEXT"` (checked on the final turn-end screen) and `--absent "TEXT"` (screens and scrollback), repeatable, on either script.
Runs land in `~/.cache/nebi-mod-verify/runs/<label>/` unless `--