Skill Rollcall

Skill roll call

Every agent that supports skills - Claude Code, Codex, Cursor, Gemini CLI, GitHub Copilot, OpenCode, Amp, Cline and the rest - reads each SKILL.md's frontmatter at startup, and most watch the folders afterwards, so adds, edits and removals normally show up within the session. None of them tell anyone why a folder was skipped, and each reads a different set of folders. This skill rebuilds the same index from disk, for the host you are running in, with the reasons attached, then repairs, lints or audits on request.

The script is scripts/rollcall.py next to this file - run it from wherever this skill was installed (~/.claude/skills/skill-rollcall/, ~/.agents/skills/skill-rollcall/, ~/.codex/skills/..., ~/.cursor/skills/..., or a project's .agents/skills/). It is stdlib-only Python 3.10+ and read-only unless --fix --apply is passed.

Which host

The script detects the host from the environment (CLAUDECODE, CODEX_SANDBOX, CURSOR_AGENT, GEMINI_CLI) and scans that host's folders. If nothing identifies the host it scans every known host's folders and labels each row with who reads it. Pass --agent when you know better:

Each host's folders, per its docs: Claude Code ~/.claude/skills and .claude/skills; Codex ~/.agents/skills, ~/.codex/skills and .agents/skills; Cursor ~/.cursor/skills and .agents/skills; Gemini CLI ~/.gemini/skills; Copilot ~/.copilot/skills; OpenCode ~/.config/opencode/skills; every host walks .agents/skills (or its own project folder) from the working directory up to the git root. Rows from the shared .agents/skills convention say agents, because several hosts read it. Under --agent all, a skill that appears in two hosts' folders is a copy, not a duplicate; the duplicate warning only fires when one host would see both.

Pick the mode from what the user asked

Flags combine. --problems-only drops the table when only findings matter; --json when you need to process the result.

Steps

1. Run it. For --known, pass the skill names from your own system context - the list your skill tool accepts (Claude Code's Skill tool, Codex's $skill menu, Cursor's /skill list). That turns "compare by eye" into a diff: rows become registered or NEW, and names the host lists but that no longer exist on disk are reported as stale.

2. Answer the question first, then summarize. If they asked about one skill, lead with that skill's line. Then, briefly:

   • errors - folders that will never register anywhere. Give the one-line fix each: add frontmatter, close it, add a name, move a nested folder up. If the script planned a fix, offer --fix --apply; show the dry run first so they can see the moves.
   • warnings - registers, but: no description (never auto-triggers), name not matching the folder or outside the Agent Skills spec (strict hosts skip it), a description over the spec's 1024 chars (hosts truncate it), duplicate names one host sees together, a skill Codex has switched off in config.toml, and under --lint the description-quality findings and Claude-only frontmatter other hosts ignore. These are why "it's installed but never fires".
   • NEW - on disk, not yet in your index. Usually it appears on the next turn. If they want it now, do step 3.
   • audit - high findings deserve a direct read of the file before the user runs that skill. review hits are phrases that are ordinary inside a security or prompt-testing skill (a jailbreak signature table will contain "ignore previous instructions"); read them in context and say what you found rather than just relaying the count. info is external hosts the skill references.

   Do not paste the whole table back when a few rows matter. Say the count, name the exceptions.

3. To use a NEW skill before the host registers it, read its SKILL.md and follow it. That is all any host's skill tool does: it puts the file body into context. Read any scripts/ or references/ it points to when the instructions call for them. Tell the user you loaded it manually so they know the slash or $ form may lag.

4. If a skill is on disk with no errors and still not registered after another turn, the host is not watching that folder. Claude Code's --bare mode has no watcher; Codex says to restart when a change does not appear. /clear or a restart reruns discovery but wipes the conversation, so let them finish anything in flight. For a Claude Code plugin, changes to hooks/, agents/ or .mcp.json need /reload-plugins; SKILL.md text does not.

5. If the user reports "/x isn't a command here" from the Claude desktop app but you have x in your own listing, the app's slash-command picker is stale, not the harness. Tell them to send the request as a plain message; do not suggest /clear for this.

Related, not overlapping

Claude Code's /skill-doctor (2.1.252+) reports what each skill costs and how often it is used. It does not say why a skill failed to register. When the user's real question is "which skills should I turn off", point them there; the context-cost line here is a rough stand-in for when /skill-doctor is unavailable or you are not in Claude Code. The skills CLI (npx skills ls) lists what it installed, not what the host loaded.

What this cannot do

It cannot make the host rescan; only the host does that. Host detection is best effort: it reads environment markers, and a host that sets none is reported as all, so pass --agent when you know. It does not evaluate whether a description will trigger for a given phrase - --lint catches the structural reasons a description under-triggers (too short, no "use when", overlaps with a sibling), not the semantic ones. It does not read Codex or Cursor plugin bundles, only skill folders. And --audit is a pattern scan, not a verdict: a clean audit means nothing obvious was found, not that the skill is safe.