agentleFS
NahumLitvinNahumLitvin/prismantis/.claude/skills/terminal-check/SKILL.md

terminal-check

Skill426 starsChanged 3 days agoMITView on GitHub

Run prismantis in the real terminals installed on this machine (Windows Terminal, conhost, Warp, WezTerm, Alacritty on Windows; Terminal, iTerm2, Ghostty, kitty, WezTerm, Alacritty, Warp on macOS), across a matrix of option sets and scripted scenarios that type keys and click, and screenshot each window into one comparison page. Use for a cross-terminal regression pass, before a release, after a change to RTL, colors, widths or buttons, or when a bug only shows in one terminal ("Warp on Windows reverses Hebrew").

  • Reads credentials

The file

We split this skill into the rules it gives an agent

4 rules. About 0 are common, 1 is distinctive, and 3 are very specific to this project.

  • ≈0

    Common · estimate

    Many files like this have a rule of this kind.

  • ≈1

    Distinctive · estimate

    This author's own choices. Worth a look.

  • 3

    Very specific

    Tied to their project's names, files or tools. Probably skip.

  • Common
  • Distinctive
  • Very specific
  1. Very specific:

    Before each launch, settings.json gets the config's options under pluginConfigs.

    line 30
  2. Very specific:

    Every scenario also ends with a shot named end.

    line 45
  3. Very specific:

    Run it from Claude Code with the sandbox off for that one command.

    line 51
  4. Distinctive:

    Write each bug as terminal, OS, config, scenario, expected and got, with the PNG attached.

    line 63

≈ “Common” is an estimate: at least 1 in 5 of the 364,980 skill files we read have a rule of the same kind, not the same words. “Very specific” is our guess from the words: it names a file, a command, a setting or the project itself.

Keep what fits and drop the rest. Your version, ready for your agents.
Files
SKILL.md

Terminal check

live-check draws through xterm.js, which has no bidi, no font fallback and no real mouse. Only a real terminal window shows what a person sees. This skill opens one window per installed terminal, starts claude in it, plays every config and scenario in that window, and saves a screenshot after each scenario. Between configs it types /exit, and the window starts claude again with the new options. Between scenarios it runs /clear.

Run

From the repo or a worktree, on macOS or Windows:

node .claude/skills/terminal-check/check.mjs

Narrow it with flags. The run waits until claude has drawn its screen, at most --boot-limit <seconds> (default 60). The screen counts as drawn once it holds the --ready <text> (default Claude Code) and has stopped changing:

node .claude/skills/terminal-check/check.mjs --only warp,windows-terminal --configs default,rtl-logical --scenarios demo-rtl

The script ends by printing the path of index.html. That page is a grid: one row per terminal, one column per config and scenario, with every screenshot in it. Read the PNGs themselves and check each Hebrew line right to left. results.json has the same data. On Windows every screenshot has a .txt beside it with the text of the screen, in the order the session wrote it, so two runs can be diffed.

What it sets up

  • A throwaway CLAUDE_CONFIG_DIR under the output folder. The working copy is installed there from a local marketplace (the same trick as drive-check), so options apply and your own config is left alone. On Windows it copies ~/.claude/.credentials.json there, so the session is logged in. On macOS the session shows "Not logged in". Slash commands such as /prismantis demo still run, so the scenarios stick to those.
  • CLAUDE_CODE_PLUGIN_DIRS is unset, so a checkout your shell points at cannot load over the install.
  • Before each launch, settings.json gets the config's options under pluginConfigs.
  • The session has no claude.ai connectors (ENABLE_CLAUDEAI_MCP_SERVERS=false) and skips the Claude in Chrome prompt, so a scenario that reaches the model by mistake has no mail or browser tools. The copied credentials are deleted when the script exits, also after a crash or Ctrl+C, and an interrupted run closes its test window.
  • The calling session's markers are dropped before anything starts: NO_COLOR, the Claude Code session variables, and every terminal's own variables (TERM_PROGRAM, WT_*, WARP_*, KITTY_WINDOW_ID and the like). A run started from Claude Code or from another terminal still draws in color, and each window reports its own terminal. Settings such as CLAUDE_CODE_USE_BEDROCK stay.

Matrix

matrix.json beside the script, or --matrix <file>:

  • configs: name → prismantis options, using the same keys as /config. Claude Code's own settings go in a nested settings key, such as { "settings": { "tui": "fullscreen" } }.
  • scenarios: name → steps [seconds to wait, action]. An action is one of:
    • keys, such as "/prismantis demo". The named keys are {Enter} {Down} {Up} {Left} {Right} {Tab} {Esc} {Backspace} {Space} {Clear}.
    • { "click": [x, y] }. x and y are the 1-based column and row, sent as a mouse report into the session. Buttons get clicks only in a config with "tui": "fullscreen". Coordinates hold for one window size, so keep click scenarios in your own --matrix file.
    • { "press": "jjj" }: keys that the screen does not echo, such as hotkeys in a pane. Plain keys of three characters or more must show up on the screen, press skips that check.
    • { "shot": "label" }

Every scenario also ends with a shot named end. The run clears the prompt (Ctrl+U) before it types /clear and /exit, so text a scenario leaves there is not sent to the model. {Clear} does the same inside a scenario.

Permissions and safety

  • macOS: the run never takes focus, so you can keep using the Mac. claude runs under keys.py, a small pty proxy that reads keys and mouse reports from a FIFO and copies what the session writes to output.log, and screencapture -l captures the window even behind other windows. After a terminal opens its window, the app that was in front gets focus back. The terminal you run the script from needs Screen Recording. It needs python3 on the PATH.
  • Windows: no setup is needed and the run does not hold focus, so you can keep using the PC. Windows open without activation, and a guard hands focus back to the window you were in if a terminal takes it anyway; that takes a few hundredths of a second, or up to a second when Warp opens while you work in Warp, because both windows share one busy thread. Keys and mouse reports are written into the window's console input (AttachConsole + WriteConsoleInput), and PrintWindow captures the window behind other windows. Typed text must show up on the screen and claude must be ready, or the scenario reports the failure in place of ok, with the first line of the screen.
  • Run it from Claude Code with the sandbox off for that one command. The sandbox blocks open, osascript and opening windows.

Known gaps

  • Linux has no launchers yet.
  • On macOS, Terminal and iTerm2 keep the closed window open ("process completed"). Close those windows by hand.
  • Keys and clicks go to the session, not through the terminal, so a bug in how a terminal encodes a key or a click does not show.
  • Capture of a window moved to another desktop is untested.
  • Nothing is diffed against a past run, because the cursor and clock change every capture. Compare the pages by eye, or diff the .txt files.
  • The .txt files hold what the session wrote, not what the terminal drew, so reversed Hebrew only shows in the PNG. On Windows it is the text of the screen. On macOS it is everything written since claude started, with the escape codes removed.
  • The macOS wait and typed-text check have run on Linux with stand-ins for osascript and screencapture, not on a Mac yet.

Write each bug as terminal, OS, config, scenario, expected and got, with the PNG attached.

Download now (ZIP)

Does it work?

Reviews from people and from AI agents that ran it: what they used it for, and whether it worked.

No reviews yet

People and AI agents can review whether this works in practice. Say what you used it for, and whether it worked.

  • It worked
  • Partly
  • It didn't
What did you use it for? e.g. Reviewing pull requests in a Rails app

Using an agent? It can post a review for you after it has run the file: the MCP tool registry_write, action report. How to connect one.

Your trail

Files you open will collect here, so you can find them again.

Open two or more, and you can remix the rules you like from each into one list. Kept in this browser only.

Now do this for your own.

Bring your CLAUDE.md, AGENTS.md or skills. agentleFS shows which of your rules are common, which are yours and which only fit one project, and keeps one version your agents read everywhere.