Shell Yeah

A fork of Shell Yeah by Steve Castaneda (MIT), with two extra switches in the Run window: Auto-run and Auto forward results. See Running commands automatically and FORK.md, which also covers how it stays in sync with the original.

A plugin for Paseo 0.13 and later. Type /sh and a command in the message box to run it yourself, in the agent's folder, with live output in the chat. The agent sees the output only if you send it.

It's the Paseo version of Claude Code's ! mode. It's also how an agent hands you a command it can't run itself: a login that needs your browser, a deploy behind your VPN, a publish that needs your 2FA code. The command shows up under the agent's reply with a Run button.

Running a command

Type /sh npm test and press Enter. A card appears in the chat:

• While the command runs: a spinner, a timer, the output as it arrives in a box you can type into, and Stop. Progress bars and spinners redraw in place, the way they do in a terminal.
• When it ends: Done or Exit 1 with how long it took, the colored output, and Send, Copy and Rerun. Buttons and notes name the chat's agent: Send to Claude in a Claude chat, Send to Codex in a Codex chat.

The output box holds the last 300 lines and stays on the newest one while the command runs. Once a command has run for a second, the box opens to 20 lines, like a terminal window, so a setup wizard or a long build has room. When the command ends, the box shrinks to fit the output. On a phone the box always fits the output, so a prompt stays in view when the keyboard opens. Scroll up for earlier lines; Jump to latest brings you back. The arrow in the header makes the box taller, up to 60% of the window. Copy copies everything the plugin kept, up to 10,000 lines.

Rerun runs the command again in the same card. The arrows next to Run 2 of 2 go back through earlier runs, up to 10. Each shows its result and the end of its output. Send and Copy use the run you're looking at.

The command runs on the Paseo host as the user Paseo runs as, in the agent's working folder, with your shell ($SHELL -c). It skips interactive startup files like .zshrc, so aliases aren't loaded. Your PATH carries over. On Windows, the command runs in PowerShell 7 if it's installed, otherwise in the Windows PowerShell that comes with Windows. It skips your PowerShell profile for the same reason.

Each command gets a terminal of its own, so it behaves as it does in yours. When it asks for a 2FA code, a password or a yes or no, click anywhere in the output box and type, as you would in a terminal. The box's border lights up while your keys go to the command. On a computer, after you press Run or Rerun in a card, they go there without a click. Each key goes to the command as you press it, so a one-key menu takes one key and arrow-key menus work. What you type shows only if the command shows it, so a password stays hidden. Esc stops typing, and so does switching to another window. Click the output to type again. On a phone, tap the output. When a command prints a sign-in link, like npm's browser sign-in, click the link in the card. Stop works like Ctrl+C. Pagers are off, so git log prints everything instead of waiting for you to press q. When the command ends, the plugin closes its terminal. Like closing a terminal window, that also ends any background job the command started, unless you started it with nohup.

The plugin drops a leading ! or $ , so you can paste a command straight from a README.

Sending output to the agent

Nothing reaches the agent until you press Send. The chat then shows one line, like Output of `npm test` (exit 1, 2.9s) · all 29 lines attached. The agent gets the command and its output as an attachment, which Paseo doesn't display, so that line says what went along. If the agent is in the middle of a turn, the output joins that turn instead of interrupting it. If the agent is waiting for you to approve something, the plugin asks you to answer that first.

Output of up to 200 lines goes whole. Longer output goes as a summary: the first 10 lines, the last 60, and the lines that look like errors, with their line numbers. The plugin saves the full log to a file and gives the agent its path, so the agent can read or search the rest itself. The chat line then says summary attached, full 12,408-line log saved for the agent.

Saved logs go in ~/.local/state/shell-yeah/logs (%LOCALAPPDATA%\shell-yeah\logs on Windows), one folder per chat, never in your project. They clean up on their own. Archiving a chat deletes its logs, and archiving a workspace archives its chats. Any log older than a week is deleted, and past 100 MB the oldest logs go first.

The footer says where things stand: Claude can't see this before you send, and Sent to Claude 10:42 after.

You can also attach a command's output to your next message: open the attach menu in the message box and pick it under Command output.

Commands from the agent

Agents created after you install the plugin learn one rule: when they want you to run something, they put it alone in a fenced block marked sh run, under a one-line comment that says in plain words what it does. The rule also says never to leave a command for you only in a sentence.

```sh run
# Deploy the coupon fix to staging
npm run deploy:staging
```

When the reply ends, the plugin adds a card under it: the agent's comment as the title, the command, Suggested by Claude, and Run, Copy and Dismiss. Nothing runs until you press Run. The shell skips the comment line, and Copy leaves it out. Without the plugin, the block is an ordinary shell code block.

After Run, the card shows the output like any other run, and Send to Claude is the main button.

Running commands automatically

By default nothing an agent suggests runs until you press Run, and nothing goes back to the agent until you press Send. Three switches at the bottom of the Run window (the Run pill in the message box) change that. Both are off by default and apply to every chat, and a change takes effect for the next reply. A line under them says what the current setting does.

• Auto-run. The cards start themselves as soon as the reply ends, as if you had pressed Run on each.
• Auto forward results. When a command the agent handed you finishes, its output goes back to the agent, as Send sends it, so it can carry on without you. It works the same for a command Auto-run started and for one you started by pressing Run on its card.
• Forward my own commands. The same for the commands you type yourself, with /sh or in the Run window. It is its own switch because what one of your own commands prints, a token for instance, goes to the agent. It is independent of the other two.

With both on, an agent can run a command, read the result and write the next command with nobody pressing anything.

• Several commands in one reply run one after another, each after the one before finished, and the chain stops at the first that fails or is stopped. The ones left over stay as cards with Run.
• Turning the setting off stops the commands still waiting. A command already running keeps going; Stop it from its card.
• Only the sh run blocks an agent wrote count, the same ones that get a card. /sh, the Run window and plain shell blocks are not started for you.
• The commands run on this computer as you, with no confirmation, so a reply that quotes untrusted text could include one. Leave it off for agents you wouldn't give a shell.
• A command that asks for a login or a code waits for you to type it in its card, as it does after pressing Run.
• Each result goes back as soon as its command finishes. A failure is sent too, since it is what the agent needs. A command you stopped is not sent. If the agent is waiting for you to approve something, the result stays unsent and the card's Send button is still there. A send that fails for another reason is tried again a few times, and every outcome is written to the plugin's log (paseo plugin logs shell-yeah).
• At most 20 commands start this way per chat in an hour. With results going back, an agent can answer each one with the next, so this is the brake on a loop nobody is watching. Past it, the cards wait for Run.

The card also warns you, in amber, when the command is hard to undo: when it runs as administrator (sudo), deletes files, throws away uncommitted changes, force-pushes, or runs a script downloaded from the internet. The plugin spots these by itself, in bash and in PowerShell, so the warning shows even when the agent didn't mention it. It catches the common forms, not every disguise, so still read the command.

Agents that started before the plugin was installed don't know the rule. Type /sh on its own and the plugin looks through the last reply for any shell code block and offers each one, up to four.

The Run pill

Agents still write a command into a sentence now and then: "sign in with npm login, then publish with npm publish". The Run pill above the message box handles those. It opens a small window that lists the commands in the agent's last reply, up to eight, from Run cards, shell code blocks and sentences. Press one to run it as a card in the chat. The pencil puts it in the box first so you can change it. Or paste any command into the box and press Enter. Shift+Enter adds a line.

Code in a sentence, or in a block with no language, counts when it starts with a program on your PATH, a shell builtin like cd or a script in the agent's folder. On Windows, a PowerShell command like Get-ChildItem counts too. Code in a sentence also needs a second word, so a passing mention of git doesn't show up, and neither does a file name like package.json. The window shows the same amber warnings as the cards.

Agents

/sh, Copy and Rerun work in any chat, because the plugin runs the command, not the agent. So far I've tested Send only with Claude.

Run cards need the agent to know the sh run rule. Paseo passes it to new Claude, Codex, OpenCode, Pi and Oh My Pi chats. On Windows, the rule also tells the agent to write PowerShell. Agents Paseo runs over ACP, such as Copilot and Cursor, don't get it; type /sh on its own to turn the shell blocks in their last reply into Run cards, or open the Run pill. The pill works with every agent.

Limits

• Full-screen programs like vim or top don't work in a card. Use a terminal tab for those.
• Up to 8 commands run at once across all chats. The plugin keeps the last 30 finished ones for Copy and Send; older cards fall back to the lines they show.
• Paseo only treats /sh as a command when the message box has no attachments. With an attachment, /sh npm test goes to the agent as text.
• Cards show live output only while the plugin is running. After Paseo restarts, a card that was still running says Interrupted, and finished cards keep the lines they ended with.
• The terminal comes from node-pty, the library Paseo uses for its own terminals. Paseo doesn't promise plugins that library. If a Paseo update takes it away, commands still run, but they can't take typing.

Install

paseo plugin add git:Eldcerust/shell-yeah

This is a fork, and it has the same plugin ID as the original, so install one or the other. To use the original instead: paseo plugin add npm:@stevecastaneda/paseo-shell-yeah.

Enable plugins under Settings → Plugins on the Paseo host. You need Node.js 22.18 or later on the daemon host.

Changelog

1.3.1

• [Fixed] Screen readers on web and desktop. They now hear whether a run's output is expanded and when a button is busy. Before, they heard neither. Phones already did.

1.3.0

• [Improved] Needs Paseo 0.13 or later. Shell Yeah now uses the plugin interface from Paseo 0.13, because Paseo stops supporting the older one in April 2027. On an older Paseo, the version you already have keeps working. Nothing else changes.

1.2.3

• [Fixed] Shell Yeah installs and updates again. Paseo couldn't build 1.2.2. Updating to it failed and left 1.2.1 running, and a fresh install failed. 1.2.3 has the same fixes and builds.

1.2.2

• [Fixed] Run cards show up on Paseo 0.13, and each command gets one card. Paseo 0.13 stopped handing Claude's reply to plugins when a turn ends, so no Run card appeared under a reply. Shell Yeah now reads the reply back from Paseo, and only from the turn that just ended. Before, a turn started by another agent's message or by a schedule offered a second card for a command from the reply before it.

1.2.1

• Leaving the Paseo window now stops typing into a command. When you come back, click the output to type again, so a key meant for something else can't reach the command.

1.2.0

• You now type into the output box itself, as in a terminal, instead of into a bar under it. Click anywhere in the box and type. Its border lights up while your keys go to the command, and Esc stops typing. On a computer, after Run or Rerun, your keys go there without a click. A command that waits for input before it prints anything now gets an empty box to type into.
• Once a command has run for a second, its output box opens to 20 lines, like a terminal window, so a setup wizard or a long build has room. When the command ends, the box shrinks to fit the output. A command that finishes within a second never grows. On a phone the box still fits the output.
• Dismiss now removes a suggestion's card instead of leaving a "Dismissed" line behind.

1.1.0

• Rerun now runs the command again in the same card, so it no longer adds a card for each run. Arrows in the card's header show earlier runs, up to 10. Send and Copy use the run you're looking at.

1.0.1

• A command whose output moves the cursor very far no longer crashes Shell Yeah. A dozen bytes, like ESC[500000000C, could run it out of memory. The cursor now stops at 4,000 characters, the longest line Shell Yeah keeps, and a move down adds at most 40 rows, one screen's height.

1.0.0

• First version: /sh with live output, sending output to the agent, attaching it to a message, and Run cards for commands the agent hands you. Each command runs in a terminal of its own, and keys you type go straight to it, so you can answer its prompts, like a 2FA code or a one-key menu. Runs on macOS, Linux and Windows, where commands run in PowerShell. Long output goes to the agent as a summary plus a saved log it can search, and saved logs clean up on their own. Cards for suggested commands carry the agent's plain-words title and warn about commands that are hard to undo. The Run pill lists the commands in the agent's last reply, including ones written in a sentence, and runs any command you paste.