v0.58.4

Patch Changes

• Connecting an editor in Settings → AI tools no longer creates a config directory for a tool you don't have installed. For Cursor, Codex, Copilot, OpenCode and Claude Desktop, that directory is also how Open Knowledge decides the tool is installed, so switching one on used to make the app report the editor as present everywhere from then on, permanently and with no way to undo it. Those rows now say the tool wasn't found and point at their setup guide instead. Claude Code is unaffected, since its config file sits next to its detection folder rather than inside it.

• Starter packs now install their skills in the packaged desktop app. Seeding a pack wrote its folders, templates and root files but installed none of the pack's skills — picking the Knowledge base pack left .claude/skills/ holding only the platform open-knowledge skill, with no knowledge-base, research-with-sources or consolidate-notes. The pack's skill sources ship inside the app's app.asar archive, and the recursive copy the installer used cannot read out of that archive, so every copy failed and the seed carried on without them. The installer now copies through the same archive-aware walk the rest of the skill machinery uses. Pack skills carry the guidance that tells agents how to work in the folders a pack creates, so until now every packaged pack install shipped the layout without the behavior.

• Skills now live on a settings page named after them. Installing a skill used to be filed under AI tools & CLI, while the page called Skills held nothing but folder symlinks and never said what symlinking was for — so people looking for a skill looked in the wrong place, and the right place explained nothing. Both scopes now have Skills Studio: the skills OpenKnowledge ships (or the project's own skill) on top, and the folders your AI tools read them from below, under a heading that leads with the reason rather than the mechanism. The AI tools pages keep the connections and point at where skills went.

  Each skill row now says what the skill does in a line written for a person. It previously printed the skill's own description, which is trigger text written for an agent — the discovery skill's runs several hundred characters and ends by telling a model what not to do. The full text is still one click away in the skill's preview, and still what the install confirmation quotes.

  ok init no longer installs open-knowledge-write-skill unless you ask for it with --skills write-skill, matching what the desktop's first launch already does. A skill nobody was asked about is not installed on their behalf, and an existing copy is never removed. Because that leaves the authoring skill undiscoverable rather than merely unrequested, the first time you open Skills Studio it explains the page and offers the skill once — dismissible, remembered, and never shown for a skill you already declined.