claude-ui

When this skill is used:

1. Treat the user's exact request as the UI task to delegate.
2. Require the user to explicitly choose both a Claude model and a reasoning effort before running anything.
3. Use the invoking agent's structured user-input or selection tool to collect those choices when that tool exists in the current harness or mode.
4. Do not ask for model or effort in plain text if the structured tool is available. USE YOUR USER INPUT TOOL TO ASK FOR THE MODEL AND EFFORT!
5. If the structured tool is unavailable, ask the user directly in plain text and do not proceed until both values are provided.
6. Do not add extra repository context unless it is necessary. Claude Code can inspect the repository itself.
7. Run the bundled helper script outside the Codex sandbox so Claude Code can use the user's existing local auth:
   • bun run .agents/skills/claude-ui/src/run-claude-ui.ts --model "<model>" --effort "<effort>" -- "<user request>"
8. While the helper is running, prefer observation over intervention. Inspect the helper state and outputs rather than taking over the task.
9. Do not interrupt, replace, or second-guess the delegated Claude run just because it is slow.
10. Only intervene if there is a concrete failure such as a non-zero exit, an auth failure, a permission or tool error, or a clearly exceeded timeout.
11. If intervention is needed, preserve any partial work already produced and explain briefly why intervention was necessary.
12. Wait for Claude Code to finish.
13. Inspect the resulting changes with:

• git status --short
• git diff --stat

14. Summarize what changed.
15. If no files changed, say so clearly.

Rules:

• The helper script must run in the current Codex workspace or worktree.
• Run the helper command with escalated permissions so it can access the user's normal Claude CLI login context.
• The skill must not run until both --model and --effort have been explicitly selected.
• Prefer the harness selection/input tool over plain-text questioning whenever that tool is available.
• Check what models and effort values are available before asking, rather than relying on a hard-coded list in this skill.
• Once the helper script starts, the invoking agent's role is to monitor, inspect, verify, and summarize. Claude should remain the implementer unless there is a concrete failure.
• Do not try to convert Claude output into a fake Codex event format.
• Prefer direct in-place file edits so the Codex app can show the Git diff naturally.
• Keep the change minimal and focused on the user's request.