skillshare v0.25.13 Release Notes

TL;DR

1. mcp serve can edit the skills you list. An Agent previews the full diff and audit result, then saves exactly that preview. Editing stays off until you list skills.
2. list_skills and read_skill are marked read-only, so clients that ask before tools that change data stop asking for them.

Edit skills over mcp serve

An Agent that reaches your skills only through mcp serve could read them but had to hand every change back to you. Skills you list in mcp.serve.edit.skills can now be changed one existing file at a time:

mcp:
  serve:
    edit:
      skills: [pdf, _team__*]   # flat names, matched like a target's include
      preview_ttl: 30m          # how long a preview can be applied; default 30m

The server then offers two more tools:

1. preview_skill_edit takes a file, the revision the Agent read, and the complete new content. It writes nothing and returns the full diff, the audit findings and a preview id.
2. apply_skill_edit takes only that preview id and saves exactly what was previewed, once. It is marked destructiveHint, so a client that asks before such tools asks you.

An edit is refused, and nothing is written, when:

• the skill is not listed, sits in a tracked repo, or was installed from upstream, since update would overwrite the change;
• the file changed after it was read or previewed;
• the skill would fail validation, or the file is not UTF-8 text;
• the audit finds something at or above audit.block_threshold. As with install, a finding you accepted with install --force does not block, and each finding is marked new or existing so you can see what the edit adds;
• the preview is older than preview_ttl, was already applied, or the server restarted.

An edit cannot create, delete or rename files. Each save goes to the operations log and each audit to skillshare log --audit. Targets get the change on the next sync. skillshare mcp serve --check reports how many served skills are editable.

Thanks to @salmonumbrella for proposing safe skill editing over MCP (#540).

Read-only skill tools

Clients and gateways that gate tools by annotation treat a tool without readOnlyHint as one that changes data, so every list_skills or read_skill call could stop for approval. Both tools are now marked read-only and are called without asking.

Contributors

@star-nebula