Skip to content
Skills Manager/DOCS
English

Health and repair

Terminal window
skills status
skills agents status
skills agents status claude

skills status (also available as skills rescan) checks Library content, metadata, and Agent directories. It reports problems without repairing them. Opening metadata written by an older version can still upgrade it, with a backup; see metadata format upgrades.

The TUI loads a fast inventory first. Repository skills appear as unchecked until you open Health or run Verify local content from the command palette; Health also uses Ctrl-R for this check. Verification compares each local directory with its recorded baseline and reports repository or modified. A skill stays unchecked if its files cannot be read. Inventory still compares Agent shadow copies with Library skills so deployment safety stays current, which can require reading those copies. Ctrl-R outside Health and automatic filesystem refreshes rebuild only the inventory, returning repository skills to unchecked; the TUI does not run full content checks while idle. A verified status reflects the last completed check, so changes limited to auxiliary files require another explicit verification because lightweight polling does not walk every skill file.

State Action
local, repository No action; the Library entry is healthy.
unchecked Run Verify local content; if it remains unchecked, check file access.
modified Review the difference, then accept or update it.
missing-baseline, missing-source Check the recorded source and metadata.
missing, renamed? Confirm whether the content moved or was removed.
invalid, corrupt-meta Fix the reported SKILL.md or TOML error.
broken Repair or remove the Agent link.
shadow Compare the Agent copy with the Library skill before replacing it.
foreign, agent-only Leave it alone unless you intend to adopt or remove it.
Missing or read-only Agent directory Check the configured path. “Read-only” is a safety classification for Library overlap, a symlinked path, or a non-directory; it is not an operating-system permission test.

Deployment state and Library content health are reported separately.

Health separates Library problems, deployment issues, and available actions.

Start with a preview:

Terminal window
skills repair

Repair covers broken per-skill links and outdated link names. A unique declared-name match becomes the repair candidate; otherwise, specify the mapping:

Terminal window
skills repair --deployment 'claude/review=repos/toolbox/review'
skills repair --deployment 'claude/review=repos/toolbox/review' --apply

In this example, replace claude/review with the Agent link from the report and repos/toolbox/review with its Library key. When applied, the repair revalidates the source and destination. It skips stale plan entries and reports paths it cannot change. Repair does not restore missing skill content, rewrite source metadata, or rename Library entries.

Run the checks again after applying repairs:

Terminal window
skills status
skills agents status claude

The repaired entry should no longer appear as broken.

Choose the operation that matches the reported state. Preview it before applying it:

Broken links

Terminal window
skills agents clean claude --dry-run

Foreign links

Terminal window
skills agents remove-link claude external-skill --dry-run
skills agents adopt-link claude external-skill --dry-run

Agent-only or shadow copies

Terminal window
skills agents relink claude hello-skill --dry-run

Remove --dry-run and add --yes to apply the selected operation.

Command Result
clean Removes broken links; without skill names, checks every broken entry for that Agent.
remove-link Removes a foreign symlink without deleting its external target.
adopt-link Copies a valid foreign skill into the Library and repoints the link; the external target remains.
relink Replaces a matching Agent copy with a Library link after validating that no differing content will be lost.

Repair, cleanup, and relink change deployments. Adopt-link changes Library content and can trigger a root backup. Back up independent content before you adopt or replace it.