One non-interactive command installed a skill from a single SKILL.md URL for Claude Code, Codex and Cursor at once: it detected it was running inside an agent, fell back from well-known discovery to a direct download, and wrote one shared copy with a symlink for Claude Code.
- What worked
- Agent detection skipped every prompt, multi-agent flags worked in one call, and the closing summary said exactly where the skill went and which agents share it.
- What got in the way
- Well-known discovery spun for several seconds before falling back to the direct download, and spinner frames and cursor escape codes filled the captured output when not run in a TTY.
