fuck ai-shell in the bin

README.md

ai.fish

Because @builder-io/ai-shell is terrible. Not their fault, really, OpenAI's models just aren't as good.

Dependencies

Tool	Notes	URL
Claude Code	You need it set up and functional.	docs.anthropic.com
A JavaScript package manager	We try all of them in order, see NOTES	oven-sh/bun
gum (optional, recommended)	gum is a terminal toolbox for lovely UI glue.	charmbracelet/gum

Invocation

ai: Request a prompt with an inline editor, and generate the command. This requires gum.

ai [PROMPT]: Provide the prompt on the command line. You don't need to quote it, but you can if you like.

Parameters

-x / --execute: DANGEROUS: Run the generated command immediately. Tries to catch dangerous commands.

-f / --force: VERY DANGEROUS: Disable enforcement of the guardrails, executing literally anything the model outputs.

Notes

The script will attempt to call Claude Code using these tools in this order:

bun → deno → pnpm → yarn → npx → claude directly

Updates

2025-07-28: I've improved the system prompt, and added execution. Now, when you run it, you'll get asked if you want to run the command. If you specify -x, it'll auto-run, for those of you who prefer yolo mode.

22025-07-28: Added a -f / --force flag for bypassing safety checks, and implemented configurable guardrails with a pattern array to catch dangerous commands like rm -rf / and sudo rm. Now ai -x will refuse to run dangerous commands unless you also specify -f / --force for true yolo mode.

ai.fish

function ai --description "AI assistant for generating shell commands"
    argparse 'x/execute' 'f/force' -- $argv
    or return 1

    # Configurable guardrails: patterns to match against potentially dangerous commands
    set -l dangerous_patterns \
        "*rm *" \
        "*chmod -R 777 /*" \
        "*> /dev/sd*" \
        "*dd if=*of=/dev/*" \
        "*mkfs*" \
        "*fdisk*" \
        "*:(){ :|:& };:*" \
        "*curl*|*sh*" \
        "*wget*|*sh*"

    set -l system_prompt "You must output ONLY a single shell command that accomplishes the requested task. Check the --help for the command, and the man page with 'man foo | cat'. NOTABLE PITFALL: BSD vs. GNU versions of tools, which have the same name but different parameters. Adapt your command if necessary. Do not perform the task yourself. Do not output any explanation, markdown formatting, or multiple lines. Output exactly one executable shell command and nothing else."
    set -l ai_output
    set -l prompt

    if test (count $argv) -eq 0
        # No arguments provided, check if gum is available
        if not command -q gum
            echo "❌  Error: gum is not available for interactive prompts."
            echo
            echo "    Install gum from https://github.com/charmbracelet/gum or"
            echo "    invoke this function by specifying the prompt directly:"
            echo
            echo "    ai [PROMPT]"
            echo
            return 1
        end

        set prompt (gum write --header "Enter your prompt" --placeholder "Type your prompt here..." --width 80 --height 10)

        # Check if user cancelled (empty prompt)
        if test -z "$prompt"
            echo "❌ Cancelled by user"
            return 1
        end
    else
        # Arguments provided, use them as the prompt
        set prompt "$argv"
    end

    # Escape quotes and special characters in the prompt to prevent command injection
    set prompt (string escape -- $prompt)

    echo "🤖 Generating command..."

    # Try different package managers to run claude code
    if command -q bun
        set ai_output (bun x @anthropic-ai/claude-code --append-system-prompt "$system_prompt" -p "$prompt" 2>/dev/null)
    else if command -q deno
        set ai_output (deno run -A npm:@anthropic-ai/claude-code --append-system-prompt "$system_prompt" -p "$prompt" 2>/dev/null)
    else if command -q pnpm
        set ai_output (pnpm dlx @anthropic-ai/claude-code --append-system-prompt "$system_prompt" -p "$prompt" 2>/dev/null)
    else if command -q yarn
        set ai_output (yarn dlx @anthropic-ai/claude-code --append-system-prompt "$system_prompt" -p "$prompt" 2>/dev/null)
    else if command -q npx
        set ai_output (npx -y @anthropic-ai/claude-code --append-system-prompt "$system_prompt" -p "$prompt" 2>/dev/null)
    else if command -q claude
        set ai_output (claude --append-system-prompt "$system_prompt" -p "$prompt" 2>/dev/null)
    else
        echo "❌  Error: No suitable package manager or claude command found."
        echo "    Please install bun, deno, pnpm, yarn, npm, or claude code directly"
        return 1
    end

    # Check if AI command failed or returned empty output
    if test $status -ne 0
        echo "❌  Error: Failed to communicate with AI service"
        return 1
    end

    if test -z "$ai_output"
        echo "❌  Error: AI service returned no output"
        return 1
    end

    # Remove code blocks, trim whitespace, and get the actual command
    # Handle various markdown formats and clean up output
    set -l command (echo "$ai_output" | sed -E 's/```[a-zA-Z0-9]*//g; s/```//g' | string trim | head -n 1)

    # Validate that we got a non-empty command
    if test -z "$command"
        echo "❌  Error: AI returned empty command after processing"
        echo "Raw output: $ai_output"
        return 1
    end

    # Check for potentially dangerous commands using configurable patterns
    set -l is_dangerous false
    for pattern in $dangerous_patterns
        if string match -q "$pattern" "$command"
            set is_dangerous true
            break
        end
    end

    if set -q _flag_execute
        # Execute mode (-x flag specified)
        if test "$is_dangerous" = true
            if set -q _flag_force
                # Both -x and -f specified: skip guardrails and execute
                echo "⚠️  FORCE MODE: Executing potentially dangerous command without guardrails"
                echo "⚡ Executing: $command"
                eval $command
            else
                # Only -x specified: refuse to execute dangerous command
                echo "❌  SAFETY: Potentially destructive command detected, refusing to auto-execute"
                echo "🤖 Generated command:"
                echo "$command"
                echo ""
                echo "💡 Use --force (-f) with --execute (-x) to bypass safety checks, or run without --execute for confirmation prompt"
                return 1
            end
        else
            # Not dangerous: execute immediately
            echo "⚡ Executing: $command"
            eval $command
        end
    else
        # Interactive mode (no -x flag): show command and ask for confirmation
        echo "🤖 Generated command:"
        echo "$command"

        # Show warning if dangerous (ignoring -f flag in interactive mode)
        if test "$is_dangerous" = true
            echo ""
            echo "⚠️  WARNING: This command appears potentially destructive!"
        end

        echo ""
        read -l -P "Execute this command? [y/N] " confirm
        if test "$confirm" = y -o "$confirm" = Y
            echo "⚡ Executing..."
            eval $command
        else
            echo "❌ Execution cancelled"
        end
    end
end