wksp

Open a new terminal tab, cd to the given folder or worktree, and launch a coding tool (Claude Code or OpenCode) with a chosen model. Optionally auto-enter handoff mode.

macOS only. Uses AppleScript to open a new tab in iTerm2 (preferred) or Terminal.app (fallback).

Usage

/wksp <name-or-path> [with|using|as <model>]

The <name-or-path> argument is required. If not provided, list worktrees and ask the user to pick one.

Natural language model specification:

• /wksp spawn2 with haiku
• /wksp spawn2 using sonnet
• /wksp spawn2 as opus

Sandbox: CRITICAL

The pretrust step writes to ~/.claude.json which requires dangerouslyDisableSandbox: true.

The osascript command MUST also run with dangerouslyDisableSandbox: true.

Workflow

Step 1: Parse natural language and resolve folder path

python3 -c "import sys; sys.path.insert(0, '~/.claude/skills/wksp'); from wksp_ops import parse_spawn_command; path, model = parse_spawn_command('<ARG>'); print(f'{path} {model or \"\"}'.strip())"

Then resolve path:

python3 ~/.claude/skills/wksp/wksp_ops.py resolve-spawn-path --arg '<PATH>'

If multiple matches are returned, ask user to choose. If none, stop.

Step 1.5: Detect available tools and build model list

Before presenting options, check which tools are installed and gather models:

which claude && echo "claude:yes" || echo "claude:no"
which opencode && echo "opencode:yes" || echo "opencode:no"

If neither is installed, stop with an error.

Claude Code models (hardcoded): opus, sonnet, haiku

OpenCode models (dynamic — only fetch if OpenCode is installed):

opencode models

Parse the output — each line is a model ID (e.g. openai/gpt-5.2).

Build the combined model list:

• Start with Claude models as ["Opus (Claude)","claude:opus"], ["Sonnet (Claude)","claude:sonnet"], ["Haiku (Claude)","claude:haiku"]
• Append OpenCode models as ["<model_id> (OpenCode)","opencode:<model_id>"] for each model
• Default selection: claude:opus

The model value uses tool:model format so Claude can parse which tool to use from the selection.

Step 2: Choose model and handoff

Present options using AskUserQuestion:

1. Model: Show numbered list of all available models (from Step 1.5) in tool:model format. If a model was already parsed from natural language in Step 1, pre-select it.
2. Handoff: Ask whether to auto-enter handoff mode (default: yes).

Parse tool and model from the selected value by splitting on : (first part = tool, rest = model).

Step 3: Ensure handoff skill is allowed

If the user chose to enable handoff, ensure Skill(handoff) is in the worktree's .claude/settings.local.json permissions so Claude won't prompt for it:

python3 ~/.claude/skills/wksp/wksp_ops.py ensure-skill-permission --path '<RESOLVED_PATH>' --skill handoff

This reads <RESOLVED_PATH>/.claude/settings.local.json (creating it if needed), checks if Skill(handoff) is in permissions.allow, and adds it if missing. For the handoff skill, it also adds a Bash(python3 "<scripts_dir>/"*) wildcard pattern so all handoff scripts run without manual approval. Must run with dangerouslyDisableSandbox: true since the path may be outside the project.

Step 4: Pre-trust workspace

python3 ~/.claude/skills/wksp/wksp_ops.py pretrust --path '<RESOLVED_PATH>'

This writes ~/.claude.json and must run with dangerouslyDisableSandbox: true.

Step 5: Open terminal tab and launch

Use launch.py to build the shell command and open a new terminal tab. The script auto-detects iTerm2 vs Terminal.app via the bundled AppleScript.

python3 ~/.claude/skills/wksp/launch.py --path '<RESOLVED_PATH>' --tool '<TOOL>' [--model '<MODEL>'] [--handoff]

Both tools accept --model for model selection. For the handoff prompt:

• claude: positional arg is the prompt — claude --model <M> "enter handoff no ask"
• opencode: positional arg is project path, use --prompt — opencode --model <M> --prompt "enter handoff no ask"

The script prints the generated shell command on stdout. Use this for the confirmation message.

Step 6: Confirm

Print confirmation: which path was opened, which tool, which model, and whether handoff was enabled.