npx skills add: --agent targets silently no-op in non-interactive mode (reports success, writes nothing) #4

Summary

npx skills add <pkg> --skill <name> --agent <a> <b> -g -y reports a successful install with status: "installed" and lists the target agents under "agents": [...], but when run non-interactively (auto-detected claude-code_2-1-275_agent) it never actually creates the per-agent skill file/symlink for any agent except the one it auto-detected as invoking it. Exit code is 0 and there is no warning/error in stdout or stderr.

Repro

npx --yes skills add typesafe-ai/skills --skill typesafe-ai --agent codex cursor -g -y --json

Output (trimmed):

●   claude-code_2-1-275_agent  Agent detected — installing non-interactively
◇  Installation Summary
│  ~/.agents/skills/typesafe-ai
│    copy → Codex, Cursor
◇  Installation complete
◇  Installed 1 skill
│  ✓ typesafe-ai (copied)
│    → ~/.agents/skills/typesafe-ai
└  Done!

[
  {
    "name": "typesafe-ai",
    "status": "installed",
    "path": "/Users/mishaalmurawala/.agents/skills/typesafe-ai",
    "scope": "global",
    "agents": ["Codex", "Cursor"],
    "mode": "copy"
  }
]

Exit code: 0. No stderr output. Ran it twice — same result both times.

Expected

Per the documented per-agent shapes (confirmed by inspecting other already-working skills on the same machine, e.g. origami):

• Codex gets a real copied directory at ~/.codex/skills/typesafe-ai/
• Cursor gets a symlink at ~/.cursor/skills/typesafe-ai → ../../.agents/skills/typesafe-ai

Actual

Only the shared source directory ~/.agents/skills/typesafe-ai (containing SKILL.md + LICENSE) is created. Neither ~/.codex/skills/typesafe-ai nor ~/.cursor/skills/typesafe-ai exists on disk:

$ ls ~/.codex/skills/typesafe-ai
ls: /Users/mishaalmurawala/.codex/skills/typesafe-ai: No such file or directory

$ ls ~/.cursor/skills/typesafe-ai
ls: /Users/mishaalmurawala/.cursor/skills/typesafe-ai: No such file or directory

npx skills ls -g -a codex and npx skills ls -g -a cursor both list typesafe-ai as installed for that agent (Agents: Codex, Cursor) even though the on-disk artifact the agent would actually read is missing — so list is reporting from the same source-of-truth the add command already trusted, not verifying the per-agent copy actually landed.

I confirmed this isn't a one-off race: reproduced identically on two separate runs, both with -y and --json.

Workaround

Manually replicating the install shape used for other working skills closes the gap:

cp -R ~/.agents/skills/typesafe-ai ~/.codex/skills/typesafe-ai
ln -s ../../.agents/skills/typesafe-ai ~/.cursor/skills/typesafe-ai

Environment

• skills CLI: 1.7.0 (via npx --yes skills)
• Invoking agent: Claude Code 2.1.275 (auto-detected by the CLI as claude-code_2-1-275_agent, which is likely why only its own non-interactive path ran and the explicit --agent codex cursor targets were skipped)
• OS: macOS (Darwin 27.0.0)
• Node: run via npx, no local pin

Suspected root cause

The "Agent detected — installing non-interactively" branch appears to special-case the auto-detected invoking agent and skip the explicit --agent <list> targets entirely when running non-interactively — i.e. --agent codex cursor is silently ignored once the tool decides it's running inside an agent session, rather than being honored in addition to (or instead of) the auto-detected one. The success report/exit code should not be 0/"installed" for agents whose files were never written.

Happy to provide the full raw terminal capture if useful.