Changelog

All notable changes to skillshare are documented here. For the full commit history, see GitHub Releases.

[0.25.1] - 2026-10-07​

New Features​

One shared instruction file, and your own, in the same file​

• prepend and append modes for single-file extras — tools without @import, such as Codex, Gemini CLI and Cursor, could only take a shared instruction file as a symlink or copy, which replaced the tool's own file. prepend and append write the shared file's content into the target file as a managed block at the top or end and leave the rest of the file alone. When the source changes, the next sync replaces the block in place. A block edited by hand is not overwritten: extras list shows the target as modified until you copy the edit back to the source or reapply the block. Refs: #436.

  skillshare extras init team-agents --file AGENTS.md --source ~/dotfiles/team \
    --target ~/.codex --mode append
  skillshare sync extras
  

• The AGENTS.md tab offers prepend and append — the mode picker lists them with import as the modes that keep the tool's own file. Each target row's path opens the file it writes, with skillshare's blocks tinted, a hand-edited block in the warning colour with Collect and Reapply, and a link to the target page for editing.

Doctor​

• doctor honors OpenCode's switches for external skills — OpenCode also reads ~/.claude/skills and ~/.agents/skills, so doctor warned about every overlap and the warning could not be silenced. OpenCode loads one skill per name, so a skill synced to both folders is fine: doctor now warns only about skills OpenCode loads from another target's folder that are missing from its own, including ones you keep there by hand, and names them. When OPENCODE_DISABLE_CLAUDE_CODE_SKILLS, OPENCODE_DISABLE_CLAUDE_CODE or OPENCODE_DISABLE_EXTERNAL_SKILLS is set where skillshare runs, doctor, sync and the dashboard treat that folder as skipped. Refs: #458.

  ! opencode loads 1 skill missing from its own folder, from: claude
  

Dashboard​

• Uninstall a tracked repo with uncommitted changes from the dashboard — the dashboard refused such a repo, or one whose git status could not be read, with no way past it. It now says why and asks again, then moves the repo to the trash as it is, like uninstall --force.

Bug Fixes​

• Installing a git submodule path is refused instead of leaving an empty folder — skillshare does not fetch submodules, so installing one reported success with nothing in it, a path inside one failed with a misleading "does not exist", and whole-repo installs dropped its skills silently. The error now names the submodule, its pinned commit and its upstream URL, and install, --track and update warn about each submodule they skip. Credentials in a submodule URL are never shown. Refs: #433.
• Scripts stay executable when a skill is downloaded through the content API — when sparse checkout was not available, the GitHub, Gitea and CNB fallbacks wrote every file without the exec bit. A file that starts with #! is now executable. Refs: #439.
• Plugin changes keep a symlinked config.yaml — plugin add and plugin update replaced a config.yaml linked from a dotfiles folder with a regular file, so the shared config never got the change, and they ignored preserve_tilde_on_save. A plugin source spelled with ~ also made plugin update fail with "snapshot ownership conflict". Refs: #460, #461.
• A project's config can no longer be pointed outside the project — a cloned project could link .skillshare/config.yaml, or its sources.mcp file, to any file you can write, and the next save, plugin, hook or MCP change, or the dashboard's config editor, would rewrite it. These writes now refuse a destination outside the project and name the link. A global config still writes through its link.
• The dashboard recognizes a project made with init -p --visible — it looked only for .skillshare/config.yaml, so a project in skillshare/ was reported as having no config and the MCP page missed its override conflict. Refs: #437.
• The config schemas accept every valid key — editors flagged plugins, cnb_hosts, gitea_hosts, extras_source, log and a skill's kind as unknown. Refs: #438.
• Codex and Gemini agent conversion works inside ESM projects — in a project whose package.json sets "type": "module", the bundled Codex and Gemini extensions failed with require is not defined. The dashboard's Sync page now shows a short cause for a failure, with the full output expandable and copyable.
• The floating action bar stays on one line — on the Pi and Oh My Pi pending-changes bars, wide content wrapped and overflowed the bar.

[0.25.0] - 2026-10-06​

New Features​

Serve skills over MCP​

• skillshare mcp serve — serves your skills read-only over MCP, for Agents that cannot reach the synced folders, such as one in a throwaway VM or behind an MCP gateway. It implements the Skills extension (SEP-2640): each skill is one entry with its full frontmatter and a manifest of its files with their sha256 digest and size. Most Agents do not support the extension yet, so the server also offers two tools, list_skills and read_skill, which any Agent that uses MCP tools can call; a client that declares the extension gets the skills natively and does not see the tools. Changes in the source show up within 5 seconds without a restart. Refs: #428.

  skillshare mcp serve                       # stdio, every enabled skill
  skillshare mcp serve --target claude       # only what the claude target selects
  skillshare mcp serve --check               # list skipped skills and exit
  SKILLSHARE_MCP_TOKEN=change-me skillshare mcp serve --http 0.0.0.0:8765 \
    --tls-cert cert.pem --tls-key key.pem    # HTTPS for other machines
  

  • Global by default; -p serves the project in the current directory.
  • A skill is skipped, with a warning on stderr, when it breaks the Agent Skills format (for example its name differs from its directory), is over 512 files or 16 MiB, or contains a nested skill that is not served.
  • --http on a non-loopback address requires SKILLSHARE_MCP_TOKEN and HTTPS through --tls-cert and --tls-key, or a loopback address behind a TLS proxy. Cross-origin browser requests are refused.
  • Do not connect local Agents that already sync skills; they would see each skill twice.
• A Skillshare tab in the dashboard's Add server — adds skillshare mcp serve as an MCP server, with the target and scope to serve, so the Agents you sync it to can read your skills.

Followed source links​

• skillshare link and unlink — use a skills checkout where it already lives, or keep skills on an external drive, by linking the folder directly under the skills source. --enable also turns on follow_source_links; unlink removes only the link and leaves the folder alone. On Windows link creates a junction, which needs no Developer Mode. Refs: #274, #419.

  skillshare link ~/code/dev-skills --enable    # links it as _dev-skills
  skillshare sync
  skillshare unlink _dev-skills
  

• follow_source_links — with this setting on, globally or per project, list, sync, status, audit and the dashboard find the skills inside first-level links in the source. A link that would loop back into the source, overlaps a sync target, or is missing is skipped with a warning. While a linked drive is unmounted, that run deletes nothing: target links, copies and install metadata stay until it is back. update --all skips linked checkouts, because they are your own working copies; update one by name.

  follow_source_links: true
  

• Link and unlink folders from the dashboard — Link folder on the Skills page takes a path, an optional name and the follow_source_links switch, and shows the same reasons as the command when it refuses a folder. A linked folder is listed as its own group with the folder it points at, and its row unlinks it after a confirmation; the link can be restored from Trash.

Oh My Pi​

• MCP, code hooks, plugins and extensions for Oh My Pi (omp) — the omp target already received skills and instructions. It now also gets MCP servers in its own mcpServers format, code hooks as native extensions, plugins from reviewed local or Git sources or imported from its marketplace, and a guarded Extensions tab for choosing which extensions load. An account declared with agent: omp gets skills, instructions, MCP and code hooks. Refs: #409.

  skillshare mcp add docs --url https://example.com/mcp --target omp -g --sync
  skillshare plugin add ./my-omp-plugin --target omp -g
  

Targets​

• DeepSeek Harness and GitLab Duo — two new built-in targets, deepseek-harness and gitlab-duo.
• Unicode skill names under target_naming: standard — names in lowercase letters of any script, such as café or 日本語-tool, now sync like ASCII names instead of being skipped. Underscores are still rejected.

Bug Fixes​

• update no longer pulls a Git repository that a link points at — a link inside the skills source to a checkout elsewhere was treated as a tracked repo, so skillshare update and the dashboard ran git pull there, and with --force reset it. A linked checkout is now only updated when you name it. Git worktrees and submodules, which have a .git file, get the same guards. Refs: #410.
• Option values that look like a scope flag are kept — in commands such as link --name -g, the value was read as the -g flag, which changed the scope or reported a mode conflict. A value after an option that takes one now stays its value, and -- ends the options. Refs: #422.
• Moving a Windows junction to another drive keeps it a junction — when the skills source and the trash were on different drives, unlink and trash restore recreated a junction as a symlink, which needs Developer Mode, so they could fail after link had worked. Refs: #420.
• A plugin whose source changes without a new version shows the commits — plugin check compares the source, not the version, so a repository that pushes commits without changing its version showed the same version with an Update button. The version tag now adds the old → new commit.
• Pi MCP settings with invalid OAuth client registration are rejected before sync — oauth.clientRegistration accepts dcr or cimd; with cimd, a clientId or clientName, or a callbackUrl other than HTTP on localhost or 127.0.0.1 with path /callback, is reported instead of being written to Pi.
• Searching skills with no match shows the no-match state — linked folders no longer appear as empty groups while a search or filter excludes every skill, so Clear filters is shown.

[0.24.6] - 2026-10-05​

New Features​

• Memory has a Refresh button, and agents keep short facts in INDEX.md — an agent writes shared notes from outside the dashboard, so the note list and the open note stayed stale until you reloaded the page. Refresh sits next to the search box and reloads both. The active guidance also told agents to create a separate file for every durable fact, which left a folder of one-sentence files. A fact that fits in a sentence or two now goes as one bullet under ## Notes in INDEX.md; only a longer note gets its own file. Existing active guidance blocks show as Outdated on the Memory page; reconnect to update them.

  skillshare extras memory instructions --update-mode active -g
  

• The Targets list links each part of a target to its tab — the Syncing column described what a target gets as one line of text, and the whole row led to the Skills tab. MCP, Hooks and the instruction file are now separate links into their own tabs, and the connected shared AGENTS.md shows by its file name, so the list shows at a glance which targets are not connected yet.

Bug Fixes​

• diff and the Sync tab no longer list synced agents as pending when a target converts them — after a successful sync, a target with an agent extension (for example Codex writing reviewer.toml) still showed every agent as New in skillshare diff and in the dashboard's Sync tab, because they looked for the source name. They now check what the sync recorded, so converted agents drop out once synced and show again when the source changes. An output whose agent was removed from the source shows as a removal, and diff leaves out the agents a target would not receive because of its include/exclude or an agent's targets frontmatter, as sync does. Refs: #391.

  skillshare sync agents
  skillshare diff
  

• status no longer counts a plain copy as a converted agent — when a target that already held copied agents started using an agent extension, status, doctor, diff and the dashboard counted the old reviewer.md copy as the converted output, although the next sync writes reviewer.toml and removes the copy. Only outputs the extension wrote count now. A converted output written before 0.24.5 shows as pending until the next sync records its source.
• The dashboard rejects skill names that point outside the skills folder — the audit, update and batch uninstall routes joined the name from the request to the source without checking it, so a name such as ../../x made audit scan a folder outside the skills source and report it clean. These routes now answer with an error for such a name. Nested names such as group/skill work as before. Refs: #402.
• skillshare audit agents keeps the Skills tab on skills — the audit screen opened for agents only discovered skills inside the agents folder, and audit agents --json reported names such as ../agents/helper.md when the two folders are siblings. Agents are now named relative to the agents folder and the Skills tab reads the skills source.
• The hub filters sit on the title row — the title and the filters used two rows with a wide empty strip above and below the filters; they now share one row.

[0.24.5] - 2026-10-04​

New Features​

• Pi packages show their version, and Check updates finds new ones on npm — an npm package in Pi had no recorded version and no source, so the Plugins page showed none and Check updates could only say to check in Pi. The version now comes from the package Pi installed, on the Plugins page and in a Pi target's Extensions tab, for npm, git and local packages alike. For an npm package added without a version, check compares it with npm's latest. A newer version shows as old → new with an Update button on the row and in the check's result; it opens the update preview for that one plugin and Agent and runs pi update <source> for that package only. A package added with a version range or tag, or one that npm resolves to a registry other than npmjs, is left to Pi.

  skillshare plugin check -g
  skillshare plugin update pi-mcp-adapter --target pi -g
  

• The Plugins page's Sync box puts what needs a look first — after a run, failures come first with their message, then the other changes, and Agents that ended the same way share one row with their logos (each names itself on hover). What stayed the same folds into one Unchanged line that opens to one row per plugin, with Pi packages apart. The preview folds what it leaves alone the same way, below the changes it will make.

Bug Fixes​

• Agents synced through an agent extension no longer show as drift — a target that converts agents (for example to reviewer.toml) was reported as 0/N with drift by status, doctor and the dashboard after every successful sync, because the converted names never matched the sources. Converted outputs now count as synced, and editing an agent in the source shows as drift until the next sync. Outputs written by earlier releases count as current until that sync records their source. Refs: #391.

  skillshare sync agents
  skillshare status
  

• status expects only the agents a target would sync — status counted every agent against each target, while sync leaves out the ones .agentignore, the target's include/exclude and an agent's targets frontmatter exclude, so such a target showed N/M with drift after a complete sync, in the table and in --json. status now applies the same filters, and a leftover link for an agent the target no longer syncs no longer stands in for a missing one. doctor and the dashboard already filtered this way. Refs: #395.
• Updating a Codex plugin that ships a portable plugin.json no longer fails on every sync — Codex installs from a root plugin.json ahead of .codex-plugin/plugin.json, and reports 1.0.0 when that manifest has no version. Skillshare expected the version from .codex-plugin/plugin.json, so a package that ships both failed with native update returned version 1.0.0; expected 4.10.3. Skillshare now expects the version Codex reads, keeps the logo from .codex-plugin/plugin.json, and still checks the version of a package whose portable manifest has none.

  skillshare sync plugins
  

• Disabling a skill that .skillignore.local keeps on now fails with the reason — a ! rule in .skillignore.local is applied after .skillignore, so the skill stayed enabled while the dashboard reported success. Single and batch toggles now report such a skill as failed and name the file that overrides it, the attempt is logged, and Settings › Files shows the .local rules above the editor.
• Plugins Skillshare installs into Pi are listed and named as plugins — a plugin installed from a Git source and bound only to Pi landed in Pi packages, and Pi's Extensions tabs titled it by its state path, in global and project views. It now stays under the managed plugins, and the tabs show its name with the Managed by Skillshare tag. The Import dialog leaves out what Skillshare already manages and says when an Agent has nothing left to import.
• A folder switch that is partly on reads as partly on — in the resource tree, the mixed state looked like a stuck off switch; it now has a half-tone track and a dash, like an indeterminate checkbox.

[0.24.4] - 2026-10-04​

Bug Fixes​

• A skill that ships its own .claude/skills/ no longer floods your Agents with the author's dev skills — installing a repository whose root is a skill copies the whole repository into the source, including any .claude/skills/ the author keeps for working on it. Those showed up as nested skills such as ffmpeg-skill/.claude/skills/code-review and were synced to every target. Discovery now skips target directories (.claude, .cursor, .factory, ...) nested below the source root, as install already did. Host-style paths directly under the root, such as .cursor/skills/*, still work and still pick their target. The next sync removes the stray links.

  skillshare sync
  

• The shared-memory guidance says which notes go to the project and which to the shared folder — an Agent that reads both a global and a project guidance block saw the same text twice and had to guess where a user preference or a project decision belongs. The global block now points project notes at the project memory when one exists and keeps user preferences, decisions that apply across projects and pitfalls in a tool; the project block keeps decisions, conventions and pitfalls of that project and says that notes about you, your tools or other projects do not belong in it. Existing guidance blocks show as Outdated on the dashboard's Memory page; reconnect to update them.

  skillshare extras memory instructions --update-mode active -g
  

[0.24.3] - 2026-10-04​

New Features​

• doctor points out links at the top of the skills source — a symlink or Windows junction directly inside the skills source is not followed by discovery, so the skills behind it never reach your Agents. doctor now lists each one as a Source link info row, in global and project mode, without following it; doctor --json reports them as undeclared_source_links checks with status info. A write into the skills source whose path crosses such a link is refused instead of written through it. Refs: #274.

  skillshare doctor
  

Bug Fixes​

• Plugin commands find the Codex CLI that the desktop app ships — Codex installed only through the Codex desktop app, Homebrew or the Windows installer was not on PATH for Skillshare, so plugin commands for the codex target stopped with a missing-CLI error. Skillshare now also looks in those places (macOS ChatGPT.app, Windows %LOCALAPPDATA%\OpenAI\Codex\bin\<version>, /opt/homebrew/bin, /usr/local/bin), and the error lists every place it searched. On a machine where Codex lives elsewhere, set SKILLSHARE_CODEX_CLI; it is read from the environment, so a config.yaml shared between machines needs no machine-specific path. Refs: #338.

  export SKILLSHARE_CODEX_CLI=/path/to/codex
  skillshare sync plugins
  

• The shared-memory active guidance says where a lesson goes — a durable fact or decision gets its own note linked from INDEX.md; a lesson from a failure goes under a new heading in LEARNED.md, as the starter file and the sharing-memory guide describe. The --update-mode help now also says that passive saves notes when you ask, not only points them out.

Performance​

• Agents are told to save to the shared notes instead of their own memory — the guidance block connected to each Agent now says that the notes folder is shared by every coding tool, to read INDEX.md at the start of each task, and, in active mode, to save a fact worth keeping here rather than in the tool's own memory, with a note format and the INDEX.md line to add. passive still saves only on request, but points out facts worth keeping so you can ask. Existing guidance blocks show as Outdated in the dashboard's Memory page; reconnect to update them.

  skillshare extras memory instructions --update-mode active -g
  

[0.24.2] - 2026-10-04​

Performance​

• Upgrades and the first dashboard start download much faster on throttled networks — GitHub's release file server limits each connection on some networks, so skillshare upgrade could spend minutes on the 9 MB archive and the first skillshare ui of a new version waited again for the dashboard files. Downloads now split the file into 8 parts fetched at the same time, each on its own connection; in one measurement the darwin_arm64 archive went from 171 s to 25 s. Checksums are verified as before, and a server without partial download support gets a single download as before.

  skillshare upgrade
  

[0.24.1] - 2026-10-04​

New Features​

• Agents can keep the shared memory notes up to date — each connected agent now has an update mode. passive, the default, reads the notes and updates them only when you ask. active also lets the agent save facts that will matter in later sessions, such as a stated preference or a decision with its reason; it skips one-off details, asks before saving when unsure, updates an existing note instead of adding a duplicate, and tells you what it saved. Choose the mode per agent in the dashboard's Connect to agents, which now shows each file's change as one diff, or print the guidance from the CLI:

  skillshare extras memory instructions --update-mode active -g
  

  Agents that read the same file share one mode. The dashboard flags an agent that reads guidance in both modes from different files.
• Config files saved by the CLI are laid out like the dashboard's Beautify — when a command saves config.yaml, globally or in a project, its sections follow the dashboard's order with a blank line between them, instead of running together with mcp, plugins and hooks above the sources. A file that uses YAML aliases keeps its order where sorting would break them.

Bug Fixes​

• Pi turns a global MCP server off in a project with its own override — a disabled entry now writes "NAME": {"enabled": false} to .pi/mcp.json, the override Pi 1.0.1 and later read, instead of a copy of the global server's command or url. Turning the server back on with Pi's /mcp used to start it without its args, env or headers; now the global server keeps all of them. It also works in project mode, where pi in a disabled entry used to be an error. The next sync rewrites switches written by earlier releases, and an identical override made with Pi's /mcp is no longer a conflict. Older Pi reports the override as invalid. Refs: #378.
• Audit no longer blocks config keys named like role labels — prompt-injection-1 reported lines such as root: ./src, admin: false or ignore: { tags: ["design"] } as CRITICAL prompt overrides and blocked the install. Lowercase config keys with a plain value now pass, while directives such as SYSTEM: jailbreak, system: Override ... or an empty system: opening an indented block are still flagged. Refs: #374.
• The dashboard counts hooks — the global and project dashboards listed every resource type except hooks, so a source with hooks looked empty. Hooks now appear between MCP and Plugins, as in the sidebar.

[0.24.0] - 2026-10-03​

Breaking Changes​

• init without a terminal now sets up what pressing Enter would — in CI, scripts and AI agents, init used to do less than the interactive defaults. It now enables git, installs the built-in skill and selects every detected tool, and prints each decision with the flag that changes it. To keep the previous minimal setup, pass the flags explicitly:

  skillshare init --no-git --no-skill --no-targets --no-copy
  

• Output that is not a terminal is plain text — piped or redirected output no longer contains ANSI colors. Parse --json where a command offers it rather than the human-readable text, which changed throughout this release.

New Features​

Init​

• Init asks first, then writes — init collects every answer, shows a summary with Yes / Change settings / Cancel, and writes nothing until you confirm. Esc cancels cleanly, and --dry-run no longer creates tool folders. The built-in skill is installed by default, and the first sync runs at the end so the tools work right away.
• Connect an existing skillshare repo — on a second machine, choose "Connect my existing skillshare repo". Init checks the repo before writing, detects its layout, prefers the repo's version of a skill with the same name, and keeps local-only skills for the next push. Init no longer writes a git identity into a repo that already exists.

Command output​

• One output style for every command — sync, status, install, update, check, push, pull, doctor, audit, mcp, plugin, extras, target and the other commands report results as aligned rows and end with one closing line or the next step to take, without boxes, logos or trees.
• Help that fits on one screen — skillshare help lists one line per command, and --help after a subcommand shows that subcommand's help instead of running it.

  skillshare extras memory --help
  

• doctor warns about an outdated built-in skill — when a newer built-in skill is published, doctor says so and counts it in the summary.
• trash empty --force — empties the trash without asking.

Terminal UI​

• One frame and keymap for every full-screen TUI — list, trash, log, audit, analyze, diff, restore, search, target, extras and the MCP manager share one title line, one key line with the common keys and the position, and lowercase letters for actions. Less common keys are under ?, and confirmations ask on the key line so the list stays visible. In list, tab switches between skills and agents; the All tab and the s status cycle are gone, and status: in the filter or --status covers enabled and disabled.
• Small choices are asked inline — the install skill picker, a large repo's folder picker, the restore source, the new wizard, extras init and MCP and plugin values no longer take over the screen. The question appears in place and collapses into a ✓ line that stays in the scrollback. Long lists can be narrowed with /, and long descriptions are cut to one row.
• A file viewer in the TUIs — browse a skill's files from list, extras, audit (opened at the finding), analyze and trash (before restoring). Control characters and hidden Unicode are shown as symbols, and very large files are skipped.

Git sync​

• push --pull syncs both ways in one command — commits local changes, merges the remote the same way pull does, pushes, and syncs the scope's targets. A conflict stops before anything is pushed and keeps the local commit. It never rebases or force-pushes.

  skillshare push --pull
  

  The dashboard's Git Sync page has the same action as Sync both ways, which lists what it will do before changing anything.

Memory​

• Shared Markdown memory across agents — keep notes in extras/memory, globally or per project, and connect the agents that should read them. The dashboard has a two-pane Memory browser with search, moves and renames, version history and recovery from backups. From the CLI:

  skillshare extras memory init
  skillshare extras memory list --search deploy
  skillshare extras memory instructions   # guidance to add to your agent instructions
  

  This is a shared file overlay; each agent's own automatic memory stays separate.

Pi​

• Turn individual Pi extensions on and off — a Pi target's Extensions tab in the dashboard lists every package in Pi's settings, including ones installed with Pi itself, and switches each extension for the global, account or project target. Each change shows the exact edit to Pi's settings before it is written, changes only that package's extension list, and is refused if Pi would apply it differently. Turning a file back to what the package loads by default removes its rule instead of adding the opposite one. Editing needs Pi 0.99.2 or later; older versions are read-only and say which version is needed.
• Add npm packages from pi.dev — plugin add npm:<package> installs a Pi package through pi install and records it like an import. The preview says that Pi runs the package's install scripts, and only targets that run Pi take npm sources. The dashboard also accepts a pasted pi install command or a pi.dev address.

  skillshare plugin add npm:@scope/package --target pi -g
  

Install and check​

• check detects changes at local install sources — skills installed from a local path, such as one shipped inside an app, used to show only "local source". check now compares the files at that path with the ones installed and reports an available update, or an error naming the path when the source is gone. Project installs with a relative path are compared against the project root.
• Paste an npx skills add command in the install dialog — the dashboard reduces npx skills@latest add owner/repo --skill=name or owner/repo@skill to its source and preselects the named skills.

Bug Fixes​

Git sync​

• Pull no longer discards local commits that hold only files — a first pull treated an agents or extras repo whose content was root-level files as empty and reset it onto the remote without --force.
• This machine's config.yaml is protected at the root git scope — a first pull is refused when the remote tracks config.yaml, a later pull that brings in a tracked copy keeps the local file and warns, and a push is refused when unpushed commits add or change config.yaml. The refusal says how to repair the history.
• Pull syncs extras at the root git scope — in the CLI and the dashboard, merged extras changes now reach their targets.
• The pull conflict dialog shows what differs — both versions get taller panes, lines only one side has are tinted, each side scrolls to its first difference, and the header counts the differing lines.

Install and upgrade​

• upgrade verifies the download before replacing the binary — the release archive is checked against the release's checksums.txt; a mismatch or an unreachable checksums file stops the upgrade and leaves the running binary untouched.
• Local installs update from any directory — global installs record the absolute path of a local source, so update and check no longer fail with "source path does not exist" when run elsewhere. A skill installed as its SKILL.md alone is updated in that shape instead of copying the whole folder.

Windows​

• No more flashing console windows — started without a console, such as from a scheduled task or pythonw, a single pull opened about ten terminal windows. Skillshare now runs its child processes in a hidden console.
• Stopping skillshare stops its child processes — killing it mid-pull no longer leaves git, ssh or hooks running. The ui start server, browser and editor still keep running.
• install file:///C:/... works — the URL is no longer rewritten with backslashes before it is handed to git.
• Interrupted or oversized downloads no longer leave temp files in %TEMP%, audit no longer reports content-missing for absolute paths in skill metadata, and Pi's locks left by a crashed process are recovered.

Plugins and MCP​

• Symlinked config folders are accepted — a linked /var on macOS, a home on a linked volume, a dotfiles-linked ~/.pi, or a PI_CODING_AGENT_DIR, CLAUDE_CONFIG_DIR or CODEX_HOME behind a link no longer fails with "refusing to modify symlinked native path". A linked file or folder inside the config folder is still refused.
• Pi project MCP overrides are explained — a .pi/mcp.json entry that only enables or disables a global server, as /mcp in Pi 1.0.1 writes, is reported as a project override with nothing to import, and a sync conflict with it says to replace it or change it with /mcp in Pi.

[0.23.5] - 2026-10-02​

Bug Fixes​

Sync​

• An include pattern that matches no skill is reported — target filters match the source path name, such as frontend__dev, even when target_naming: standard shows the folder as dev. A pattern written with the name the target shows selected nothing, and sync removed what a working pattern had linked without saying why. Sync now warns about each include pattern that matches no skill, names the target, and suggests the source path name when one fits. The warning appears in skillshare sync and in the dashboard's sync results. --json output is unchanged.

  ! claude: include pattern "dev" matches no skill in the source (filters use the source path name; did you mean "frontend__dev"?)
  

[0.23.4] - 2026-10-01​

Bug Fixes​

Plugins​

• Claude marketplace cleanup no longer fails after settings lost the declaration — when a Skillshare marketplace was only recorded in Claude's known_marketplaces.json, for example after a dotfile manager overwrote ~/.claude/settings.json, every sync failed with "could not finish removing its marketplace" and the plugin stayed pending. The marketplace is now removed from every settings scope. A name also declared at another path is reported as a conflict and left in place.

  skillshare sync plugins -g
  

• Skill and plugin name clashes are explained before install — Claude reads a skill folder that has a plugin manifest as <name>@skills-dir and loads only one plugin per name, so adding a Claude plugin from the same repository as a synced skill made /plugin report the skill as not loaded. The install preview now says Claude will load the plugin and skip that folder until one of them is renamed or removed.

[0.23.3] - 2026-10-01​

Bug Fixes​

Plugins​

• Plugin marketplaces are named after their plugin — the local marketplace Skillshare registers for each Claude Code or Codex plugin was named skillshare-<hash>, which said nothing in /plugin or codex plugin marketplace list. New installs are named skillshare-<plugin>-<hash>, such as skillshare-humanizer-322a2e01808560f4. Existing installs keep their names.
• Excluding or removing a plugin removes its marketplace — the marketplace stayed registered after its plugin was excluded or removed, and for good when the plugin had already been uninstalled in the Agent or its first install had failed. Skillshare now removes the marketplace it registered in those cases too, and a failed cleanup runs again on the next sync. A marketplace with the same name at another path, and imported plugins' marketplaces, are left alone. An update registers a missing Skillshare marketplace again.

  skillshare plugin disable humanizer --target claude -g
  skillshare sync plugins -g
  

• Imported plugins whose marketplace is gone are skipped with the reason — when an imported plugin's native marketplace was no longer registered, for example after synced settings dropped it, sync failed with "claude command failed". Sync and update now skip that Agent and say how to recover, and the plugin's other Agents still sync. Skillshare does not switch an imported plugin to its recorded source by itself.
• Plugin errors are translated and shown once — the Plugins page showed failed operations in English, and repeated the same message once per failed Agent.

[0.23.2] - 2026-10-02​

New Features​

Hooks​

• Git config hooks — a hook entry can bind git to register named Git config hooks (Git 2.54+) globally or in a project. Skillshare writes the commands to its own include file, adds the include.path line, and writes the helper scripts listed under files; {files} expands to their directory on each machine. Preview, sync, backups and restore work as they do for Agent hooks, and managing hooks never runs them. Linked worktrees share their project's hooks.

  bindings:
    git:
      commands:
        project.check:
          events: [pre-commit]
          command: "{files}/check.sh"
      files:
        check.sh: |
          #!/bin/sh
          exec make check
  

  skillshare hooks add git-check --file git-check.yaml -g --dry-run
  skillshare hooks sync -g
  

  With Git older than 2.54 the files are still written, and the plan says why the hooks cannot run. parallel needs Git 2.55+. When the include target is a symlink or not writable, the plan prints the lines to add by hand. The git key always means Git, so an account target named git receives no hooks; the plan warns once an entry uses the git binding.

Doctor​

• Check MCP servers, hooks, plugins and extras — skillshare doctor and the dashboard's Doctor page now check these too, in global and project mode:
  • MCP: environment variables, commands, client rules and sync state, without DNS lookups or starting servers
  • Hooks: what hooks sync would still change or refuse. Git hooks that cannot run, for example because Git is missing, are reported with the reason instead of as unsynced
  • Plugins: what plugin sync would change, without fetching sources, when a plugin is configured
  • Extras: config errors, broken links in targets, and drift as diff reports it

  skillshare doctor --json
  

MCP​

• Pi 1.0 OAuth metadata URL — piOptions.oauth.authServerMetadataUrl, which Pi 1.0 uses instead of discovering a server's authorization server, is checked when you save: it must use https, or http on localhost. Pi 1.0 keeps OAuth sign-ins per server name and URL, so after renaming a server or changing its url, sign in again in Pi.

  mcp:
    servers:
      example:
        url: https://mcp.example.com/mcp
        piOptions:
          oauth:
            authServerMetadataUrl: https://example.okta.com/.well-known/openid-configuration
  

Backups​

• Backup limits and Delete all — backup.max_count and backup.max_size_mb in the global config set how many target folder snapshots to keep and their total size (defaults 10 and 500 MB, 0 = no limit). sync, the dashboard's sync and backup --cleanup apply them; backups older than 30 days are still removed. The dashboard's Target folders tab shows the limits, edits them, and adds Delete all behind a confirmation. Project snapshots keep the defaults, and file, MCP and hooks backups are not affected.

  backup:
    max_count: 20
    max_size_mb: 1000
  

Dashboard​

• Share several plugins as one command — Share on the Plugins page, in the header or a plugin's menu, lists the plugins added from an HTTPS source and copies one line that adds the ticked ones in order. Each add uses --no-tui -g, so the plugins land in the recipient's global config without a picker per plugin; they choose Agents on the Plugins page afterwards. An option keeps the per-plugin Agent picker. Plugins added from a local directory are left out, since the path only exists on your machine.

[0.23.1] - 2026-10-01​

New Features​

Plugins​

• Update Codex plugins — plugin update now updates a Codex binding. Codex has no update command, so Skillshare refreshes the reviewed snapshot, adds the plugin again, which replaces the installed copy, and checks the installed version. A plugin disabled in Codex is skipped with the reason instead of being turned back on.

  skillshare plugin update review --target codex --no-tui
  

• Run a compatible CLI for an account — an account target can set cli so its plugin commands run a compatible executable instead of the Agent's own, such as omo for a Pi account. It takes a name on PATH or an absolute path. If the CLI is missing, the command fails; Skillshare does not fall back to the Agent's CLI.

  skillshare target add omo --agent pi --config-dir ~/.omo/agent --cli omo
  

Hooks​

• Hooks for account targets — a global hook binding can name an account target declared with agent and config_dir, such as a second Codex home. Claude, Codex and Pi accounts use their Agent's native binding format, and sync, import and backups keep the account name.

  targets:
    codex-2:
      agent: codex
      config_dir: ~/.codex-2
  hooks:
    entries:
      check:
        bindings:
          codex-2:
            events:
              Stop:
                - hooks:
                    - type: command
                      command: "echo checked"
  

MCP​

• Turn off a global server for Pi in one project — a disabled entry for a project under mcp.projects now reaches Pi. Skillshare writes the global server's command, or its url without the query, with enabled: false to that project's .pi/mcp.json; args, env and headers stay out of the file. This replaces the 0.23.0 change that removed pi from the targets of disabled entries. If a 0.23.0 sync already removed pi from such an entry, add it back to targets. A project's own config still cannot turn off a global server for Pi.

  mcp:
    projects:
      ~/work/project01:
        servers:
          context7:            # off in this project only
            disabled: true
  

• Turn a server off in Pi from the dashboard — Turned on in Pi in a server's Pi settings writes enabled: false when you clear it, so Pi keeps the server without connecting to it, and removes it when you check it again. It sits above Tool exposure and stays in step with the Other Pi settings JSON.

Dashboard​

• Discard Git Sync changes — Discard changes restores tracked files in the source repository to the last commit and deletes untracked files and folders, after you confirm. Ignored files, nested repositories and the root config.yaml are kept, and a dry run shows what would change.
• Resolve Git pull conflicts across computers — when this computer and the remote both changed the same files, Git Sync no longer only stops with the merge undone. Review conflicts shows the local and remote version of each file side by side; keep one whole-file version per file, then Apply choices and pull. Both commit histories are kept, other files merge normally, and .metadata.json conflicts still resolve automatically. If either side gets new commits before you apply, the dialog asks you to choose again. Binary files and files over 16 KiB can be chosen but have no preview.
• Pull and merge, Commit and pull — when both computers have new commits, Pull becomes Pull and merge and the page explains that both histories are kept before you push. With uncommitted changes while the remote is ahead, Commit and push becomes Commit and pull.

Bug Fixes​

Sync​

• Sync no longer deletes skill links you made yourself — in merge mode, sync removed any link in a target that pointed into the skills source but was filtered out, including links you created by hand or with another tool, and the removed links could not be restored. A live link is now removed only when Skillshare created it, as recorded in .skillshare-manifest.json, or with --force; other links are kept and counted as local. Links created before 0.15.0 that are now filtered out are kept as well; remove them by hand or with --force.
• Agent and extra links to other places are kept — merge-mode sync and extras --remove-target --prune removed every agent .md link or extra link that was not expected, including links to files outside the source. Now only broken links and links into the source are removed. A link whose file cannot be read, for example because of permissions, is no longer treated as broken.
• Copy-mode agents keep your own files — in copy and extension modes, sync deleted every agent file in the target that did not match a source agent, including files you put there. Copies are now tracked in a .skillshare-manifest.json in that agents folder, and sync removes only copies it wrote that you have not edited. Copies that were already orphaned before upgrading are kept; delete them by hand.
• Failed agent prunes are reported — when sync could not delete an orphaned agent link or copy, it still counted it as pruned. It now shows a warning, and the other orphans are still removed.

Plugins​

• One Agent no longer blocks a plugin update — a plugin installed on several Agents could not update anywhere when one of them could not take the update, such as a plugin disabled in Codex or Copilot, or an imported package. Those Agents are now skipped with the reason, the others update, and the skipped update stays pending. Imported Codex plugins now update through codex plugin marketplace upgrade, and imported Pi packages in global mode through pi update. A plugin whose Agents hold different versions shows each version.
• Claude account targets list their plugins — a Claude account target could fail with "native plugin list contains an unsupported plugin identifier", for example when the dashboard ran from the home folder, because Claude also lists plugins of other scopes. Those entries are now ignored.
• Plugin results are translated — the dashboard's Plugins page showed last-action statuses and installation messages in English in every language.

Hooks​

• Editing a hook from a target tab keeps account bindings — saving a hook from a target page dropped its bindings for account targets such as codex-2, so the next sync removed that account's hook.
• Hooks report config errors — when the global config failed to load, skillshare hooks ran without account targets and reported unsupported Agent "codex-2" instead of the real error.
• Missing account folders are explained — the dashboard's hook config view showed nothing for an account whose config_dir is missing. It now shows a warning for that target.

MCP​

• Account shells keep the default home — when CLAUDE_CONFIG_DIR, CODEX_HOME or PI_CODING_AGENT_DIR pointed at a declared account's config_dir, sync, import and the dashboard sent the plain Agent target to that account's home and could prune entries in its default home. The plain target now keeps its default home, and sync warns about the shadowed variable.
• Project-mode MCP status is complete again — in project mode, mcp check, the dashboard and the target list dropped pending syncs and conflicts, and mcp remove --keep-files did not stop managing a project's Claude off switch, so the next sync removed it anyway.
• Pi 0.99.2 settings are kept — Pi 0.99.2's auth: {provider: NAME}, which sends a provider's /login token to an HTTP server, was refused in piOptions as a pi-mcp-adapter setting, and loading or importing a config dropped it, and the dashboard's Other Pi settings refused it too. It is now kept and checked: it needs an https url, or http on localhost, and global mode, because Pi reads it only from its global file. Only the adapter's string auth is still removed. oauth.clientName is checked as text, and description passes through.

  skillshare mcp add docs --url https://example.com/mcp --target pi --pi-options '{"auth":{"provider":"github"}}' --no-tui -g
  

• Server names Pi reads as one are refused — Pi 0.99.2 reads names that differ only in - and _, such as my-docs and my_docs, as one server and skips the second with a config error. Skillshare now refuses the second before writing it.
• Pi's error says how to turn off a server — a disabled entry for Pi that Skillshare cannot write now fails with a message pointing at a complete server with piOptions: {"enabled": false}, instead of only listing the clients that support a switch. The MCP docs explain the same.

Dashboard​

• Beautify keeps config.yaml in order — Beautify and save in Settings → Files now keep config sections in a consistent order with a blank line between them, and keep the # yaml-language-server: $schema=… line at the top of the file, moving one that was saved in the wrong place.

Website​

• Desktop app guide — a new Desktop App page covers installing Skillshare App on macOS, Windows and Linux, in all five documentation languages. The homepage, navigation and README now show the desktop app before the CLI installation. On Apple Silicon Macs:

  brew tap runkids/tap
  brew install --cask skillshare-app
  

[0.23.0] - 2026-10-01​

New Features​

Hooks​

• Manage native hooks — skillshare hooks keeps named hooks in hooks.entries and writes them into each Agent's own hook configuration: Claude Code, Codex, Gemini CLI, Copilot CLI, Cursor, Factory Droid, Qwen Code, Antigravity (and its CLI, agy), Pi, Amp and OpenCode, in global and project scope. Each binding keeps that Agent's own event names and format; Skillshare does not translate between Agents. Sync previews every file it changes, keeps your other settings and hooks, reports a hook you edited by hand as a conflict, and backs files up before writing. It never runs a hook or changes an Agent's trust.

  skillshare hooks add check --file ./check.yaml
  skillshare hooks sync --dry-run
  skillshare hooks sync
  skillshare sync --all              # now includes hooks
  

• Import the hooks you already have — hooks import --from AGENT lists that Agent's existing hooks. Saving one takes over those registrations in place, so the next sync does not add a duplicate.

  skillshare hooks import --from claude
  

• Disable, stop managing and restore — hooks disable keeps the definition and removes it from the Agents on the next sync. hooks remove --keep-files stops managing a hook and leaves its entries in the Agent files as yours; hooks import offers them again. hooks restore brings back a backup of an Agent file without touching unrelated later edits.

  skillshare hooks disable check --sync
  skillshare hooks remove check --keep-files
  skillshare hooks restore BACKUP_ID --dry-run
  

• Project hooks from the global config — hooks.projects holds hooks for other project folders, so one sync from the global config reaches each of them.
• Removal cleans up after itself — removing a hook deletes a hook file Skillshare created once nothing else is left in it, together with the folders it created that are now empty. Files and folders that existed before, and anything you added, stay.

MCP​

• Tool policy per server — tools.allow and tools.deny say which of a server's tools reach the model. Write them once; sync translates them into Pi's toolExposure, Codex's enabled_tools and disabled_tools, and Copilot's tools. Parts an Agent cannot hold, such as the whole policy for OpenCode and Kilo Code, are named in the sync plan and by mcp check instead of being dropped.

  skillshare mcp add github --target pi --target codex --tools-allow 'get_*,search_code' --tools-deny get_secret -- github-mcp
  skillshare mcp edit github --tools-allow ''    # clear the allow list
  

Dashboard​

• Hooks page — add and edit hooks by picking each target's documented events, copy commands from another target's tab, or edit the native JSON with completion and checks. Every sync shows a diff of each file it changes. Import lists each target's hooks that Skillshare does not manage yet. Hooks also appear on target and project pages, on the Sync page and in Settings → Backups.
• Clear remove choices — removing a hook offers Remove and sync, Remove from source only and Stop managing; hovering or focusing a button explains what it does to the target files. In the hook and MCP remove dialogs, Remove and sync names the other pending hooks or servers that go out with it.
• Tools in the MCP server dialog — load a server's tool list with the dialog's current settings, even before saving, and tick the tools the model gets. The dialog says which selected Agents follow the policy fully, partly or not at all.
• Plan notices before an MCP sync — the MCP sync dialog and the Sync page list the plan's notices, such as Pi's built-in MCP needing Pi 0.99.0, before you sync.
• Open config.yaml at the right place — the config.yaml button in the Hooks and MCP page headers opens Settings → Files at the hooks: or mcp: section. The field panel there explains every hooks key.

Bug Fixes​

• Script installs no longer need administrator access — install.sh now installs to ~/.local/bin by default; INSTALL_DIR still overrides the location. When that folder is not on your PATH, or an older skillshare such as /usr/local/bin/skillshare comes first, the installer says so and prints the PATH line to add. Remove the old copy (sudo rm /usr/local/bin/skillshare) so the new one runs.
• Stop managing an MCP server works in the dashboard — Stop managing in the MCP remove dialog always failed with a "changed since preview" error. It now succeeds, and its message says the entries stay in the Agent files.
• Beautify unfolds one-line YAML — in Settings → Files, Beautify left a section squeezed onto one line, such as servers: {docs: {url: …}}, unchanged and said there was nothing to tidy. It now unfolds nested one-line sections; short lists such as targets: [claude, codex] stay on one line.

Breaking Changes​

Pi MCP​

• Pi uses only its built-in MCP — Skillshare now writes Pi's servers only to Pi's own mcp.json (~/.pi/agent/mcp.json, or .pi/mcp.json in a project). It no longer writes for pi-mcp-adapter or pi-mcp-extension, and the piExtension choice is gone. The first sync after upgrading moves your servers: entries Skillshare wrote to mcp-adapter.json are removed and written to mcp.json, and pi-mcp-extension entries are rewritten in place. Entries you added to mcp-adapter.json yourself are left alone, and mcp import --from pi still reads them.
• Needs Pi 0.99.0 or later — Pi added its built-in MCP in 0.99.0. On older Pi the moved servers stop loading until you update Pi. Skillshare does not check Pi's version; the sync that moves servers prints a warning.
• Remove the old extension from Pi — if pi-mcp-adapter or pi-mcp-extension is still installed in Pi, uninstall it. Pi's docs say an installed extension that registers /mcp replaces the built-in MCP, and pi-mcp-adapter 3.0.0 and later no longer read mcp.json.
• Adapter-only server options are dropped — Pi's built-in MCP does not read these piOptions fields, so the next sync removes them from your config: approveTools, auth, bearerToken, bearerTokenEnv, bearerTokenStore, caFile, debug, exposeResources, idleTimeout, inheritEnv, lifecycle, protocolVersion, requestHeadersCommand, requestTimeoutMs, searchKeywords, socket, tasks, toolPrefix, trace.
• No per-project off switch for Pi — Pi cannot turn off one global server in one project. The next sync removes pi from the targets of disabled entries, so that server is on again in that project. To keep a server off for Pi, set piOptions: {enabled: false} on a complete entry.
• directTools becomes Pi's exposure — directTools: true becomes piOptions.exposure: direct, "search" becomes deferred, and a list of tool names becomes piOptions.toolExposure with those tools direct. mcp.directTools and a project's default are copied into each Pi server that sets none. piOptions.includeTools / excludeTools become tools.allow / tools.deny.
• Removed flags — --pi-extension, --direct-tools and --pi-options-prune now fail with a message saying what to use. Use --pi-options '{"exposure":"direct"}' or --pi-options '{"toolExposure":{"TOOL":"direct"}}' instead of --direct-tools. Sync always removes unchanged Pi fields Skillshare wrote earlier, so --pi-options-prune is no longer needed.
• Your config is updated on the first sync — the old config still loads, and sync mcp --dry-run names each retired setting. The first sync mcp, sync --all or dashboard sync saves config.yaml (or the sources.mcp file) without them, keeping the previous version in the file history.

[0.22.2] - 2026-09-30​

New Features​

MCP​

• Check servers before an Agent starts them — skillshare mcp check tells you whether each server will work as synced: a fromEnv variable that is unset, a command not found on PATH, a remote host that does not resolve, a server an Agent's rule refuses, and an Agent entry that conflicts with the source or is not synced yet. It starts nothing, writes nothing and never prints variable values. It exits 1 when it finds an error, so it fits in scripts. In the global config it also checks the servers under mcp.projects and names their project.

  skillshare mcp check
  skillshare mcp check docs github --json
  skillshare mcp check --no-dns    # skip the DNS lookup of remote hosts
  

• Probe servers live — mcp check --live also starts each local server, or sends one request to each remote server, reports its name, version, protocol version and number of tools, and stops it again. A server that needs sign-in is a warning; Skillshare never signs in. Each probe gets 10 seconds, or --timeout, and configured values are removed from every message.

  skillshare mcp check --live --timeout 30s
  

• Pi built-in MCP — Pi 0.99.0 includes MCP, and piExtension: builtin syncs servers into ~/.pi/agent/mcp.json (.pi/mcp.json in a project) with no extension to install. piOptions takes Pi's per-server fields such as exposure and toolExposure, and import keeps them. A new server that goes to Pi without a mode now uses builtin; before, mcp add --target pi without --pi-extension failed. Existing servers keep their mode.

  skillshare mcp add docs --url https://example.com/mcp --target pi --pi-options '{"exposure":"deferred"}' --no-tui
  

• Remove cleared Pi settings — clearing a Pi setting stops managing it and leaves its value in Pi. --pi-options-prune, or Remove cleared settings from Pi in the dashboard, removes the fields Skillshare wrote that nobody has changed since.

  skillshare mcp edit docs --pi-options '{}' --pi-options-prune --no-tui
  

Dashboard​

• Check MCP servers — the MCP page and each project's MCP tab have a Check button in the Sync box once there are servers to check. It runs the same check as mcp check without --live, with a summary above the list and each problem under its server.
• See every file Sync writes for a server — View what each Agent gets lists the Skillshare source and every target Agent's file, with its path and format (JSON, JSONC, TOML, YAML), and shows the one you pick with syntax highlighting. It counts the files Sync writes and names the scope: global, or the project.
• Pi settings in the server dialog — a Pi server chooses its Pi MCP mode, built-in by default, its tool exposure, other Pi settings as JSON, and Remove cleared settings from Pi. When a pasted snippet holds one server, the paste tab shows the same Pi settings.
• Tidier MCP page — the header keeps only Import from a target and Add server; Check and Backups and restore moved into the Sync box. Each server row shows its command next to its name and the Agents it goes to as chips below, and the Pi chip shows the Pi mode. Project MCP tabs use the same rows, and the MCP dialogs are wider.

Bug Fixes​

• Shell completion matches the CLI — completion offered commands that did not work: backup restore backed up a target named restore, extras mode only printed help, install --source failed as an unknown option, and hub index --audit-skills was rejected. Those are gone. bash, zsh, fish, PowerShell and Nushell now complete mcp check with --live, --timeout and --no-dns, subcommands such as backup files and audit rules, and flags no shell offered before.
• SSH failures say why — installing or loading a hub over SSH that failed showed only "Could not read from remote repository.". The message now keeps SSH's reason, such as Permission denied (publickey), Host key verification failed or an unresolved host, and a denied key suggests ssh -T user@host.
• Dashboard target links open the target — target rows on the dashboard, the Playful pin notes and the "needs attention" entries opened the target list. They now open that target's page.
• Chinese dashboard says 目標 and 目标 — buttons and messages such as "從 target 匯入" mixed the English word into Traditional and Simplified Chinese. They now say 目標 and 目标.
• Hubs page opens on the default hub — with no default hub saved, the Hubs page opened with nothing selected and Skillshare Hub had no default star, although search --hub already falls back to it. The page now stars Skillshare Hub in that case and opens on the default hub when you have no hub of your own.
• Find skills no longer installs when it finds nothing — in the Install dialog, Find skills on a source with no skills or agents went on to install the whole source as one skill without a SKILL.md, and so did Install on such a search result. The dialog now says nothing was found and installs nothing.
• MCP restore preview keeps its size — switching backups resized the restore dialog while the preview loaded. It now stays the same size.

[0.22.1] - 2026-09-30​

Bug Fixes​

• Dashboard file editor no longer crashes — opening a file tab on a target page, such as APPEND_SYSTEM.md on pi, showed "Something went wrong — Unrecognized extension value in extension set". The dashboard bundled two copies of its code editor library; it now bundles one.

[0.22.0] - 2026-09-30​

New Features​

Dashboard​

• Richer markdown preview — skill and file previews now render HTML such as <details>, GitHub alerts, ::: containers, emoji shortcodes, ==mark==, ++ins++, ^sup^ and ~sub~, and footnote links resolve. HTML goes through a sanitizer, since previews show skills from anyone. Tables no longer squeeze short columns until words break, file previews keep single line breaks as the tool reads them, a leading --- block counts as frontmatter only when it is YAML, and outside links open in a new tab.
• Failed sync targets listed first — when a dashboard sync loses a target, the Sync page now lists each failed target with its part (Skills, Agents, Extras or Config) and error above the other warnings, marks it in the change list, and shows a warning toast instead of Sync complete. Each failure gets a plain-language explanation for common causes (a symlink pointing elsewhere, permission denied, a read-only disk, a file where a folder belongs, a missing path, invalid target settings). A skills symlink that points elsewhere offers Turn on Force. The Last sync card names the failed targets, and the project sync dialog shows the same list instead of a success note.

Bug Fixes​

Sync​

• Dashboard sync keeps going after a target fails — the first skills target that failed stopped a dashboard sync with an error, leaving every later target unsynced and agents, extras and MCP not run. Like skillshare sync, the failed target is now reported as <target>: sync failed: <err> and the rest still sync; the operation log records it as partial.
• Invalid target settings fail only that target — one target with invalid settings, such as a path that is a file instead of a folder, or a project target without a path, stopped the whole sync, sync -p and dashboard sync before anything ran. That target is now skipped with its error and the others still sync.
• Project dashboard works with a target that has no path — skillshare ui -p would not start, and a running project dashboard failed every request, when a custom target had no skills.path. It now starts with a warning, and its sync lists that target as failed.
• sync -p agents checks settings and is logged — it synced agents into targets with invalid settings and left no entry in skillshare log. It now skips those targets, exits non-zero, and logs the run like sync agents.
• Dashboard sync reports every failed target when all fail — when every skills target failed, the dashboard showed only the first error, and agents, extras and MCP were not synced. It now lists every failed target the same way as a partial failure.
• Agent failures counted in the operation log — a target whose agents failed to sync was left out of the sync entry's targets_failed, so a partly failed sync could be logged as ok. The entry now counts it and lists failed_targets.
• Prune failures reported — when removing orphaned skills or agents from a target failed, sync and the dashboard said nothing and stale links stayed behind. They now show <target>: agents prune failed: <err> (or prune failed for skills) along with the prune's own warnings.
• Dashboard respects symlink conflicts — in symlink mode, a target folder linked somewhere else was replaced by a dashboard sync without the conflict message skillshare sync shows. Without Force it now fails with conflict - symlink points to X (use --force to override).
• sync --all exits non-zero when an extras target fails — it printed a warning and exited 0, and with --json ignored extras errors entirely. It now fails in text and JSON, global and project, and the dashboard's operation log records extras failures as partial instead of ok.
• An extras failure no longer stops a dashboard sync — the first failed extras target ended the run: the skills result and warnings already on screen were dropped, MCP was never written, and the error named neither the extra nor the target. The sync now finishes and lists every failed extras target.
• Missing extras source is skipped, not created — global sync and the dashboard created an empty source folder for an extra whose source did not exist, which hid a misconfigured path. Every sync now skips that extra and says Source directory does not exist: <path>.
• Dashboard skips extras targets the agents sync writes — like the CLI, a dashboard sync no longer writes the agents extra into a target folder that the agents sync already manages, so the two stop fighting over it.
• Context-cost warnings name the right target — the dashboard's sync warnings showed {target} instead of a name, and when several targets tied, listed offenders from a different target than the one named. They now name the target, and ties go to the alphabetically first one with its own offenders.

Tracked repositories​

• Unreadable git status no longer treated as clean — when skillshare could not read a tracked repo's git status (for example, a corrupt index), update pulled over it, uninstall moved it to trash after a warning, and status and list showed it as up to date. update and uninstall now fail that repo with failed to check git status: <err> unless --force, while other items in the batch still run; status and list show it as unknown with a warning, and status --json reports "status": "unknown". The error now includes what git printed, not just exit status 128.

  skillshare update --all --force   # pull even when git status cannot be read
  

• Dashboard checks private repos with your token — the dashboard's update check fetched without the HTTPS token, so private tracked repos and skills without a pinned branch reported errors that skillshare check did not. It now uses the same check as the CLI.

Project mode​

• Project uninstall reports the real problem — every name that could not be resolved was reported as "not found", even when it matched several nested skills or pointed at a file. Project mode now gives the same specific errors as global.
• Project uninstall logs like global — the operation log recorded --help, dry runs and declined prompts as uninstalls, always with zero succeeded. It now records one entry when the uninstall ran, with the real count, and uninstalls from the project list view are logged too.
• logs/, trash/ and backups/ stay out of your project's .gitignore — in projects created before 0.17.3, project uninstall added them to the project's own .gitignore instead of .skillshare/.gitignore.

CLI​

• search shows mixed-case skill names — skills named like MySkill, which install accepts, never appeared in search results. Search now uses the same name rule as install.

Dashboard​

• Revert on ignore-file tabs — on the .skillignore and .agentignore tabs of the Config page, Revert reset config.yaml and left the ignore file's edits in place. It now resets the file the tab has open.
• Install dialog marks the right skills as installed — the dialog matched installed skills by name only, so a skill with the same name from another repo showed as Installed. It now matches by source.
• Path overlap warning translated — the Sync page showed the overlap warning in English in every language. It is now translated and has a button that opens Health Check.
• Target-only resources grouped by target — the Sync page's list of resources that exist only in targets was one comma-joined line. Each target now has its own group with its logo and count.
• Failed target settings no longer reported as saved — batch target updates reported success even when saving the settings failed, so an override could disappear on the next reload. They now show the error.
• Sync page in Chinese, Japanese and Korean — the Playful theme's handwritten note on the Sync page repeated the subtitle and, with no CJK glyphs in its font, showed as stray large text. It has been removed.
• Selected tool in Add target — the picked tool now uses the same highlighted row as the collect and install dialogs, instead of an inset ring.

Breaking Changes​

• update and uninstall stop on an unreadable git status — a tracked repo whose git status cannot be read now fails instead of being treated as clean. Scripts that run update --all unattended should expect a non-zero exit for such a repo, or pass --force to pull anyway.
• sync --all exits non-zero when an extras target fails — it used to warn and exit 0. sync --json also reports the failure.
• sync -p agents exits non-zero for a target with invalid settings — it used to sync agents into that target and exit 0.
• Missing extras sources are no longer created — sync used to create an empty source folder for an extra whose source did not exist. Create the folder yourself or fix the path in the config.
• POST /api/sync returns 200 when every target fails — the dashboard API used to answer 500 with the first error. It now returns the usual result with every failed target in failed; check that list instead of the status code.

[0.21.17] - 2026-09-29​

New Features​

Dashboard​

• Edit Pi's APPEND_SYSTEM.md on its target page — the pi and omp target pages have an APPEND_SYSTEM.md tab next to the instruction file, so you can edit the text Pi appends to its system prompt without leaving the dashboard. + adds a tab for any other file a tool reads, including files in subfolders; the dashboard saves them under the target's files. Three tabs show at a time and the rest move to a menu; removing a tab doesn't delete the file. Refs: #301.

  targets:
    pi:
      files:
        - SYSTEM.md
        - prompts/review.md
  

  Paths stay inside the tool's folder (~/.pi/agent for pi, .pi in a project). Share with Extras on a tab opens Add extra filled in with the file's folder and name, so one file can be shared across tools.

Extras​

• Single-file extras named after their file — in Add extra, a single file's Name follows its file name without the extension (APPEND_SYSTEM.md gives APPEND_SYSTEM) until you type one, and the hint suggests using the file name. Refs: #300.

Bug Fixes​

• Ticking an Agent after importing from it takes its entry over — importing an MCP server from an Agent without ticking that Agent, then ticking it later, left the Agent's entry unmanaged: no sync was offered, nothing was recorded in state.json, and unticking the Agent removed nothing. The preview now lists the entry as Take over; sync records it as managed without changing the file, and unticking the Agent afterwards removes it. A server you turned off in the Agent itself is still never claimed. Refs: #303.
• MCP tab for antigravity-cli — the antigravity-cli target page now has an MCP tab, showing the ~/.gemini/config/mcp_config.json file it shares with Antigravity.
• Dialogs keep focus on their first field — dialogs that focus a field when they open, such as Add extra, moved focus to the close button instead.
• Pi's logo in its brand colors — Pi now shows its coral, blue and yellow logo instead of a black mark.
• Sync page lists you can scan — the expanded targets in sync list was one comma-separated line mixing global targets with <project>@<tool> ones. It now shows logo chips, global targets apart from each project's. The ignored list is grouped by .skillignore and .agentignore and by folder, so a prefix such as security/ is written once.

[0.21.16] - 2026-09-29​

New Features​

Sync​

• Targets that undo each other's sync — when two targets sync skills into the same folder with different include or exclude filters, each sync adds what one wants and removes what the other filters out, so sync keeps showing the same changes. sync, doctor and the dashboard's Sync page now name both targets and the folder and suggest keeping one (universal when it is one of them); the Sync page has a button that stops syncing skills for the other.

  skillshare target codex --skills=false   # let universal alone write ~/.agents/skills
  

  Targets with the same filters in one folder are not reported.

Dashboard​

• Plugin version and logo without an Agent — a plugin added without picking an Agent now shows its version, and the logo from its Codex manifest (interface.logo, up to 1MB) in the Plugins list and the add preview, instead of the default icon.

Bug Fixes​

• No false "also reads" note for a shared folder — a target sharing its skills folder with another, such as codex and universal both on ~/.agents/skills, was told it also reads the other's skills and sees each one twice. It holds one copy, so the note no longer appears.
• Duplicate skills described accurately — the dashboard and docs said each skill "shows up twice" when a tool reads two skills folders. Pi keeps the first and warns, and Gemini and others pick one, so they now say the tool finds each skill twice.
• Antigravity no longer counted as reading ~/.agents/skills — Antigravity documents only ~/.gemini/config/skills as the desktop app's global skills folder, so switching its skills off no longer claims it still sees universal's skills.

[0.21.15] - 2026-09-29​

New Features​

Targets​

• Switch skills off for a target — a target can now stop syncing skills while skillshare keeps managing its agents, MCP servers and instructions. Use it for a tool that also reads another target's folder, such as Pi reading ~/.agents/skills of universal, so each skill no longer shows up twice. Turning skills off saves skills.enabled: false and removes only the links into your source: your own skills stay, copies from copy mode are kept and listed apart, and a folder another enabled target writes to is left alone. sync, diff, status and doctor then skip the target's skills, and analyze counts what the folder still holds.

  skillshare target pi --skills=false --dry-run   # preview what is removed
  skillshare target pi --skills=false
  skillshare target add gemini ~/.gemini/skills --no-skills
  

  Turn skills back on with --skills=true; the next skillshare sync syncs them again. Works in project mode with -p.
• 10 more targets — autohand-code, fx, jazz, kimchi, kimi-code, ona, posit-assistant, qoder-cn, reasonix and zcode, with the folders each tool also reads recorded so the dashboard can tell you when two targets overlap.
• More vendor logos — more targets show their vendor's colored logo instead of a letter.

Dashboard​

• Stop syncing skills from the target page — Stop syncing skills on a target's Skills tab lists what will be removed and what stays before anything changes, and warns when other tools read the same folder and would lose those skills. The dashboard's target list shows such targets as Skills off.
• See who else reads a skills folder — a target's Skills tab names the targets that stopped syncing skills and read its folder instead, and the other folders a tool also reads, so you can spot skills that load twice.
• Adding a target shows what it writes — the add dialog pins universal as the shared folder when it is not configured yet, and for the picked tool lists every place skillshare will write: the skills folder (with a switch to add it with skills off), the agents folder, the MCP config and the instruction file.
• Where a target's AGENTS.md comes from, in one card — the AGENTS.md tab now shows one card with the shared file the target follows or imports, the other targets using it, and the actions for it. Change switches the target to another shared AGENTS.md in place; Edit opens the shared file. The tab opens in Preview when the file has content, Show all expands it to its full length, and Save appears once there is something to save.

Extras​

• Several single-file extras in one source folder — in the Add extra dialog, a single file's Source file is one path: the source folder, then the file name. The folder defaults to the extra's name; click a folder another single-file extra uses, or type a new one, and each extra syncs its own file from it. Project extras now accept --source too, relative to the project root, so a project can share a folder the same way. Removing an extra keeps files the others still use. Refs: #300.

  skillshare extras init review -p --source .skillshare/extras/prompts --file review.md --target .pi
  skillshare extras init append -p --source .skillshare/extras/prompts --file append.md \
    --target .pi --as APPEND_SYSTEM.md
  

Bug Fixes​

• Windows paths of single-file extras use backslashes throughout — the Extras page showed paths such as ~\.pi\agent/APPEND_SYSTEM.md, with a slash before the file name. Refs: #300.
• Kilo Code's AGENTS.md location — the dashboard suggested ~/.kilocode/AGENTS.md, a folder Kilo never reads. The kilocode target now uses ~/.config/kilo/AGENTS.md globally and AGENTS.md in a project, as Kilo documents; Change location on its AGENTS.md tab sets another path such as ~/.kilo/AGENTS.md. Refs: #302.

Breaking Changes​

• The zencoder target moved to .agents/skills — ~/.agents/skills globally and .agents/skills in a project, because Zencoder now documents that location and reads .zencoder/skills only for backward compatibility. A global config stores full paths, so an existing one keeps its path. A project config stores target names only, so it follows the new default on its next sync -p, which also removes the links skillshare left in the old folder; folders you made by hand there are kept.
• replit is project-only — Replit documents only the project .agents/skills folder, so the target no longer has a global default path. An existing global replit target keeps the path in your config.

[0.21.14] - 2026-09-29​

New Features​

Install​

• Pin a skill to a tag or commit from its web URL — the branch, tag or commit SHA after tree/ or blob/ in a GitHub URL (-/tree/ on GitLab, src/ on Bitbucket) is now the install ref, so a pasted URL installs the version it names. skillshare update keeps the pin. A ref the remote no longer has, such as tree/master/ after a rename to main, fails the install instead of falling back to the default branch. Branch names containing / are matched against the remote's branches and tags, tree/HEAD/ links use the default branch, and --branch still overrides the URL. Refs: #293.

  skillshare install github.com/team/skills/tree/v1.2.0/skills/foo
  

  Skills installed earlier from a tree/<ref>/ URL got the default branch, and update and skillshare install from config keep them there.
• Pin hub entries — a hub index entry whose source names a ref installs that revision for everyone, from skillshare search --hub and the dashboard's Hubs page. Move the pin by editing the ref in the index. Pinned skills from one repo and ref still share one clone. Refs: #293.

  { "name": "reviewer", "source": "github.com/owner/repo/tree/v1.2.0/skills/reviewer" }
  

Extras​

• Sync any single file — skillshare extras init takes --file to sync one file from the source folder instead of the whole folder, and --as to give it a different name at the targets. Use --add-target <path> --as <name> for a different name per target. init only writes the config: it does not create the source file or sync. Refs: #300.

  skillshare extras init pi-prompt --source ~/dotfiles/prompts --file system.md \
    --target ~/.pi/agent --as APPEND_SYSTEM.md
  

  In global mode, several single-file extras can share one --source folder, each syncing only its own file. extras list shows the full source and target file paths, and the extras init wizard asks whether to sync a folder or a single file.
• Single files on the dashboard — the first Extras tab is now Folders & files and lists every extra except shared AGENTS.md files. Add extra can create a single file, with a file name per target and merge, copy or import mode.

Dashboard​

• One Hubs page for browsing and building — browsing a hub and building your own are now the same page: your hub is shown the way others will see it and edited in place. Entries can be added from any repository URL, with a branch or tag picked from the remote, and skills added from installed ones keep the version they were installed from.
• Backups grouped by day — Settings → Backup groups folder backups by day with one compact row per backup. Opening a row lists each folder with its file count and size, and its own Restore button. The page says what it is doing while backups load and while Back up now runs.

Docs site​

• llms.txt for AI tools — the docs site now serves llms.txt, an index of every page, and llms-full.txt, the full English docs in one file, so you can point an AI assistant at the docs. Both are linked from the site footer.

Bug Fixes​

Dashboard​

• Previewing your own hub shows it — previewing a draft from the hub builder opened an empty "pick a hub" page instead of the draft.
• Counts of one read correctly — labels such as "1 skills", "1 of 1 targets" and "1 backups" now use the singular in every language.
• The dashboard no longer looks stuck after an upgrade — restarting after skillshare upgrade deleted the UI files the upgrade had just downloaded, so the server fetched them from GitHub again. On a slow connection that outlasted the reconnect wait.
• The project-mode note no longer says Backup is hidden — Settings → Backup appears in project mode, scoped to the project.

Performance​

• The Backup page opens faster — each snapshot folder is read once for its size and file count, instead of twice.

[0.21.13] - 2026-09-28​

New Features​

Dashboard​

• Choose how each tool gets a shared AGENTS.md — on Extras → AGENTS.md, every connected tool now has a mode dropdown: import adds one @import line and keeps your own lines, symlink links the file, and copy writes a copy. Saving the shared file in the dashboard updates the copies right away. Switching back to import brings back your own content, including edits you made while in import mode. Refs: #299.

  skillshare ui
  

• See what a restore puts back before it happens — turning a tool off, or Restore all, first shows the file the tool will have afterwards, with a diff against what it has now.
• Move a tool's instruction file — Change location on a target's AGENTS.md tab sets a custom path and file name for any supported tool, not only custom targets. A folder is refused, and the option is off while a shared file is connected.
• Put a shared AGENTS.md in any folder — under Other locations, Add location writes the shared file into a folder that is not in the targets list, under its own name or another one such as instructions.md, as symlink, copy or import. Each location has its own mode, and Remove shows the restore preview first. In a project, shared files move from the Folders tab to the AGENTS.md tab, with the same locations relative to the project root. Refs: #299.
• Preview everywhere, and Cmd/Ctrl+S on the target page — every box that shows an AGENTS.md has Preview and Source (or Edit) tabs, long lines wrap, and the shared file's card opens in Preview. The target page editor saves with Cmd/Ctrl+S and asks before you leave with unsaved edits.
• Update progress — the Updates tab on Skills and Agents shows a progress bar, marks the row being updated, and moves blocked or failed updates into their own section with a one-line reason.

Backups​

• Manage every backup from Settings → Backup — the tab now has three sections. Target folders lists snapshots with a filter by target or agents, and can delete one. Files lists each AGENTS.md, CLAUDE.md or shared-file location skillshare backed up before rewriting it, with each version's reason (converted, @AGENTS.md added, edited, collected, overwritten…), a diff preview, and restore. MCP lists MCP config backups by agent and what each changed. The tab now appears in project mode too, scoped to the project. Refs: #299.

  skillshare backup files                      # files with backups
  skillshare backup files show ~/.claude/CLAUDE.md
  skillshare backup files restore ~/.claude/CLAUDE.md <id>
  skillshare backup --delete 2026-09-28_10-52-00
  

  skillshare backup files is now a subcommand; to back up a target named files, use skillshare backup -t files. In a project, backup --list -p and backup --cleanup -p work on the project's agent snapshots.

Extras​

• Attach a single-file extra under another name — --add-target takes --as to pick the file name in the target folder, such as instructions.md in a notes folder. skillshare extras <name> --help now prints that extra's options.

  skillshare extras personal --add-target ~/work/notes --as instructions.md --mode symlink
  

Bug Fixes​

Windows​

• AGENTS.md no longer shows up as a folder — single files were linked with directory junctions, which tools read as a folder. Files are now linked with file symlinks, and when Windows can't create them (Developer Mode off and not an administrator), skillshare writes copies and keeps them up to date instead. The dashboard marks symlink as unavailable and explains why. Refs: #299.
• Junctions are recognized as links again — since Go 1.23, a junction is no longer reported as a symlink, so skills synced as junctions could be reported as local folders by status, doctor, sync and the dashboard.
• The AGENTS.md tab shows the file name on Windows — a target's instructions tab was labeled with the whole Windows path instead of its file name, such as CLAUDE.md.

Extras​

• One shared file per linked or copied tool — a tool whose file is a link to, or a copy of, one shared file could also be connected to another shared file. The import line was then written through the link into the other shared file, and every tool reading it picked it up. The dashboard, Connect all, --add-target, --mode and sync extras now refuse this and name the file that holds the tool.
• Your files are kept in more cases — an instruction file that is a link into your dotfiles is left as it was after restore; a missing end marker no longer duplicates the managed block; replacing a link you pointed elsewhere is reported and recorded; and the managed block follows the file's CRLF line endings.
• Project imports use relative paths — in project mode the @import line written into a committed CLAUDE.md pointed at an absolute path in your home folder. It is now relative to the file, so it works for everyone and after the project moves.
• Clearer status for single-file extras — extras list -p no longer reports correctly synced relative links as drift; an edited copy shows modified and can be collected from the dashboard; an invalid mode in config.yaml is no longer shown as healthy; and sync extras --dry-run says when an edit would be backed up and replaced.
• sync extras --json exits with an error when a target fails — it used to exit 0 while the plain output exited 1.

Dashboard​

• Agents backups can be restored from the dashboard — restoring a <target>-agents snapshot failed with "target not found".
• Messages about shared AGENTS.md files are translated — warnings and errors from mode changes, connecting, moving and restoring used to appear in English in every language.
• Empty rules folders are left out of the read order — a target's AGENTS.md tab listed ~/.claude/rules/ even when it held no rule files.

CLI​

• NO_COLOR is honored everywhere — status, doctor, extras list and about twenty other commands still printed colors with NO_COLOR set.

[0.21.12] - 2026-09-28​

New Features​

Dashboard​

• Edit the instruction files of tools that get skills through universal — Codex, Gemini CLI, Pi and other tools read skills from ~/.agents/skills, but none of them reads ~/.agents/AGENTS.md, so with only universal as a target their own files, such as ~/.codex/AGENTS.md, had no page. Universal's AGENTS.md tab now has a dropdown with each of these tools that is installed and isn't a target of its own; pick one to see and edit its file. They also appear in the shared AGENTS.md list under Extras, so a shared file can be connected to them. Cline and the Warp Agent CLI read ~/.agents/AGENTS.md itself, and are now shown as sharing universal's file.

  skillshare ui
  

Bug Fixes​

Targets​

• Removing a target no longer takes skills away from another target in the same folder — codex and universal both write to ~/.agents/skills. Removing one of them turned the other's synced skills into local copies that skillshare no longer managed. The folder is now left alone while another target still uses it, and target remove says so, including with --dry-run.

  skillshare target remove codex --dry-run
  

Upgrade​

• A skill left owned by root now says how to fix it — an upgrade run with sudo before v0.21.10 wrote the built-in skill as root, and every later upgrade failed with only permission denied. The error now includes the chown command that gives the skills folder back to you.

[0.21.11] - 2026-09-27​

New Features​

Dashboard​

• More tools' instruction files are filled in — the instruction file tab on a target page now knows the file 22 more targets read, instead of asking you for its path. For example, universal reads ~/.agents/AGENTS.md, pi reads ~/.pi/agent/AGENTS.md, copilot reads ~/.copilot/copilot-instructions.md and qwen reads ~/.qwen/QWEN.md. A file you already have at that path shows up with no setup. The full list is in Share one AGENTS.md across your tools.

  skillshare ui
  

Bug Fixes​

Dashboard​

• Reloading a page other than the first one no longer hangs — opening or refreshing an address such as /targets/claude asked for the page's scripts in the wrong folder, so it stayed on its loading placeholder and the browser console showed Failed to load module script errors.

Plugins​

• Opening a plugin no longer downloads its source again — expanding a plugin to see which other Agents can take it cloned its whole repository every time, which could take tens of seconds. Skillshare now reads the copy it reviewed when the plugin was added, and downloads the source only if that copy was changed.
• Clearer messages while a plugin source is read — when a source can't be reached over the network, or the repository or branch doesn't exist, the dashboard now says which, instead of a general git failure. It also says when you are offline, and when reading a source or asking an Agent's CLI is taking a while. For example, codex plugin list can wait on Codex's remote marketplace; Skillshare waits up to 90 seconds for each Agent.

[0.21.10] - 2026-09-27​

New Features​

Dashboard​

• Share one AGENTS.md across your tools — the Extras page has a new AGENTS.md tab. Create a shared AGENTS.md and choose which targets read it. Claude has no user-level AGENTS.md, so it gets an @ import in ~/.claude/CLAUDE.md and keeps its own content; Codex, Gemini and other targets get a link in place of their file, which is backed up first. Several shared files can sit side by side, such as one for personal and one for work, and targets that follow @ imports can use more than one. Restore asks first, then puts back exactly what the target had before it was attached.

  skillshare ui
  

• Edit each tool's instruction file — every target page has a tab named after the file that tool reads, such as CLAUDE.md, GEMINI.md or AGENTS.md. It shows the read order, an editor, and warnings such as Windsurf reading only the first 6,000 characters. Convert… moves the content of CLAUDE.md into AGENTS.md by import, rename or copy, and backs up the file first.
• AGENTS.md in projects — in project mode the tab shows whether each target reads ./AGENTS.md, and adds a small fix for tools that only read their own file, such as @AGENTS.md in CLAUDE.md.
• Tools skillshare doesn't know — a custom target can say which instruction file it reads, in the Custom target dialog when you add it or later from the same tab. The setting is saved as instructions on the target:

  targets:
    myagent:
      path: ~/.myagent/skills
      instructions:
        path: ~/.myagent/AGENTS.md
        import: true        # the tool follows @path lines
  

• See which projects a skill reaches — on the Skills page, the Targets column shows your global tools as icons, followed by a folder badge with the number of projects the skill is synced into. Before, each project repeated its tools' icons, so claude plus two projects showed three Claude icons. Hover to see the global tools and each project's tools, including projects the skill does not reach. The Target filter now groups global tools and projects, with one entry per project, and in the tree view the Targets row shows where the selected skill actually goes. Refs: #297.

Extras​

• Single-file extras — an extra with file syncs one file instead of a whole folder, as renames it per target, and the new import mode writes an @ line into the target's own file instead of replacing it. extras list shows modified when a linked target was replaced by a different file, and extras remove puts back what each target had before the first sync.

  extras:
    - name: personal
      file: AGENTS.md
      targets:
        - path: ~/.codex
        - path: ~/.claude
          as: CLAUDE.md
          mode: import
  

Bug Fixes​

Git sync​

• A pull that fails midway no longer leaves the remote's files behind — when git could not finish a pull, for example because a file in the source folder was owned by root, the files it had already written showed up as local changes. The dashboard then blocked the next pull and suggested committing them, which would have pushed stale content back. A failed pull now restores those files and leaves your own edits alone, and a permission failure shows the chown command that gives the source folder back to you.

Upgrade​

• Upgrading with sudo no longer leaves root-owned files in your home — when the binary lived in a root-owned folder, upgrade ran entirely under sudo, so the built-in skill, dashboard assets and logs were written as root and a later git pull of the skills source failed with Permission denied. Only the binary replacement now runs with sudo.

  skillshare upgrade
  

MCP​

• Pi servers load again with pi-mcp-adapter 3.0 — the adapter stopped reading Pi's mcp.json and now reads mcp-adapter.json in the same folder, so servers synced for it were ignored. Skillshare now writes them to ~/.pi/agent/mcp-adapter.json, .pi/mcp-adapter.json in a project, or mcp-adapter.json in a Pi account's folder. The next sync removes the entries Skillshare had written to mcp.json and leaves your own there. If you already renamed the file as Pi's warning suggests, Skillshare keeps managing the entries you moved. pi-mcp-extension still uses mcp.json. Refs: #298.

  skillshare sync mcp
  

[0.21.9] - 2026-09-26​

New Features​

Dashboard​

• Turn a whole folder or tracked repo on or off — the tree view on the Skills and Agents pages now shows the folders on the left and what you selected on the right. Selecting a folder or tracked repo gives one switch that enables or disables everything inside it, and each skill is listed with its own switch. Cmd/Ctrl-click adds to the selection and Shift-click selects a range, so the same switch works on any set of skills. A disabled skill is marked with a power-off icon in place of its usual one. The divider between the two sides can be dragged. Refs: #295.
• Targets for skills in tracked repos — setting targets on a tracked repo, one of its subfolders or a single skill in it used to be refused. It now works, and the setting is kept outside the cloned repo, so the repo stays clean and skillshare update keeps it. Refs: #295.
• Filter and group by folder — the list and cards views have a Folder filter, and grouping has a Folder option, so skills installed into a folder can be seen one folder at a time or side by side. A tracked repo counts as one folder, and skills at the top of the source are grouped under Root.

  skillshare install ~/my-skill --into frontend
  

• A shorter toolbar on the Skills and Agents pages — the Source, Status, Target and Folder filters are now chips that show only their name until set; a set filter shows its value and a button to clear it. Grouping and sorting share one menu, and expand all / collapse all in the tree view is one button.

Bug Fixes​

Sync​

• Skills under a dot folder are no longer reported missing — a source skill in a folder such as .system/ syncs to a target entry starting with a dot, which target scans skipped. diff, doctor, status and the dashboard reported these synced skills as missing, and sync did not prune them after the source skill was deleted. Refs: #294.

Dashboard​

• One Agent's plugin error no longer breaks the Plugins page — when an Agent's plugin list could not be read, for example OpenCode with both opencode.json and opencode.jsonc, the whole Plugins page failed to load. That Agent now shows its error and the other Agents work as usual. Refs: #296.
• The dashboard server no longer risks crashing under overlapping requests — changing a skill while another request was being handled could crash the server.

[0.21.8] - 2026-09-25​

New Features​

Git sync​

• See what the remote has before you push — opening the Git Sync page fetches from the remote, and Pull shows how many commits are waiting, such as Pull 2 commits, so a machine that is behind finds out before a push is refused.
• .metadata.json conflicts resolve themselves — every install or update rewrites .metadata.json, so pulling after two machines had both installed or updated skills almost always conflicted there. pull now merges that file skill by skill; when both machines changed the same skill, the one installed later wins. A conflict in any other file stops the pull, undoes the merge and names the files, so the repository is never left half-merged.

  skillshare pull
  

Bug Fixes​

Git sync​

• Pulling after both machines committed — once this machine and the remote each had commits the other lacked, pull failed with Need to specify how to reconcile divergent branches, in the terminal and on the Git Sync page, and the only way out was git in a terminal. pull now merges the two histories. skillshare update does the same for a tracked repository with local commits.
• A refused push offers Pull — when the remote had newer commits, Push on the Git Sync page showed git's raw rejected message. It now says to pull first and puts a Pull button next to the error.

[0.21.7] - 2026-09-24​

New Features​

Extras​

• Pull target edits back with extras collect --force — collect skipped every file that already existed in source, so an edit made directly in a copy-mode target could not be brought back. With --force, the target's version overwrites the source file. Files whose content already matches are still skipped, and with flatten, a second target file with the same name is reported instead of overwriting the first.

  skillshare extras collect rules --force
  

  Refs: #291.

Bug Fixes​

Extras​

• Collecting from a copy-mode target keeps its files — extras collect, in the terminal and in the extras list, replaced each collected file in the target with a symlink, even when the target uses mode: copy. Copy-mode targets now keep their files. Refs: #291.
• --from finds the target however its path is written — extras collect --from only matched a target written exactly as in config.yaml, so /Users/me/.claude/rules or a trailing slash did not match ~/.claude/rules, and the target's mode, flatten and extension settings were ignored. Paths are now compared after expanding ~. Refs: #291.

[0.21.6] - 2026-09-23​

New Features​

Targets​

• MCP servers on each target's page — a target whose Agent has an MCP config file now has an MCP tab with one row per server: its endpoint, whether this Agent gets it, and whether a change is waiting to be synced. Clicking a row adds the server to that Agent or takes it out, saved right away. As on the MCP page, this only changes the source; Sync all targets then writes the MCP files of every target at once. Adding and importing servers, resolving conflicts and restoring backups stay on the MCP page. The Targets list also counts the MCP servers each Agent gets, next to its skills.
• A project's MCP on its targets in project mode — with skillshare ui -p, the tab lists the project's servers, which go to its own files such as .mcp.json. For Claude Code, OpenCode, Kilo Code and Pi, it also lists the switches that turn a global server off in this project.

Sync​

• MCP and project sync dialogs look like the Skills one — the dialogs behind Sync MCP and Sync project show one row per target instead of one line per item, as the Skills sync dialog does. An MCP row names the project of the Agent file and the servers that change, with counts such as 1 to add or 1 to turn off. Sync project still says which MCP entry is in conflict and why. The Sync page keeps the item-by-item list.

Bug Fixes​

MCP connections​

• Quick clicks on Agent toggles no longer undo each other — ticking two Agents in quick succession on the MCP page or a project's MCP tab could fail with an error and untick the first one again. Each toggle now waits until the previous one is saved.
• Turning a server off in a project for another Agent is saved — on the MCP page of skillshare ui -p, ticking an Agent on a server that is off in this project also listed the project's other Agents, including ones that cannot turn a server off, such as Cursor, so the save was refused. Only Agents that can turn a server off in a project are listed now.

[0.21.5] - 2026-09-22​

New Features​

Targets​

• A second account of Claude Code, Codex or Pi as its own target — when you run an Agent with a second config folder, such as a work account started with CLAUDE_CONFIG_DIR=~/.claude-work, add that folder as a target and Skillshare works out where its skills and agents go. Codex (CODEX_HOME) and Pi (PI_CODING_AGENT_DIR) work the same way. In the dashboard, choose Add target → Another account; in the terminal:

  skillshare target add claude-work --agent claude --config-dir ~/.claude-work
  

  In config.yaml the target only records the folder:

  targets:
    claude-work:
      agent: claude
      config_dir: ~/.claude-work
  

  Refs: #289.
• Accounts receive MCP servers and plugins too — an account's name works wherever an Agent's does in mcp.targets, a server's targets, --target and --from. Servers are written into the account's own file (.claude.json, config.toml or mcp.json in its folder), and plugins are installed with the Agent's CLI pointed at that folder, so one account can have a server or plugin the other does not. skillshare mcp import --from claude-work reads that account's servers. A Pi account needs piExtension: pi-mcp-adapter, since pi-mcp-extension always reads ~/.pi/agent/mcp.json. Refs: #289.
• Removing an account says where MCP still names it — skillshare target remove and the dashboard remove the target, then warn when mcp.targets or a server's targets still lists it, since the next MCP sync would fail. Refs: #289.

Sync​

• Sync one part from its own page — Skills, Agents and MCP each have a sync button that previews and writes only that part, and a project page has Sync project, which writes that project's skills, agents and MCP without touching global targets or other projects. The sync prompts after an update, uninstall or collect open the same dialog. Refs: #289.

MCP connections​

• Stop managing a server without touching Agent files — removing a server always cleared its entries from the Agents on the next sync. Stop managing drops it from Skillshare and leaves the entries where they are, so the Agent keeps using it. It is the third choice in the dashboard's remove dialog and in the remove wizard:

  skillshare mcp remove context7 --keep-files
  

  Refs: #290.
• Servers added to an Agent directly are pointed out — the MCP page and each project's MCP tab say which Agent files hold servers Skillshare does not manage, and Import opens on that file. Refs: #290.
• Pi adapter settings show on the server row — Direct tools are listed on a line under the endpoint instead of only on hover. Other adapter settings only say that they are set. Refs: #289.

Bug Fixes​

MCP connections​

• Importing a project's conflict uses that project's files — Import from an Agent on a project's conflict read the Agent's global file and wrote to the global source, so the project's own entry could never be taken over. Refs: #290.
• Pi adapter settings survive unticking Pi — unticking every Agent dropped Direct tools and piOptions, so ticking Pi again reset the adapter. Refs: #289.
• The MCP page no longer crashes on a cold load.

Dashboard​

• Skill and agent pages list their projects — the Skills and Agents lists counted project targets, but the detail page only showed global ones. It now lists each project the skill or agent syncs to, with a link.
• Skills and Agents sync previews match the Sync page — they showed changes on targets that were already in sync.

CLI​

• Global setup errors say they are about the global config — skillshare ui -g with a missing source folder said "source directory not found" and suggested skillshare init, which then said Skillshare was already initialized. The error now says it is the global source, names the config, and tells you to create the folder or point sources.skills at an existing one.

[0.21.4] - 2026-09-22​

New Features​

Plugins​

• Add a plugin without choosing Agents yet — leave every Agent unticked in Add plugin to keep the plugin in Skillshare without installing it anywhere. Its row shows No Agents yet; tick an Agent there when you want it, and the install preview uses the source as it is then. In the terminal, leave out --target, or confirm the picker with nothing selected:

  skillshare plugin add ./my-plugin --no-tui    # keep it in Skillshare; choose targets later
  

  A plugin added this way stays managed when its last Agent is removed. skillshare plugin remove NAME without --target removes it from Skillshare.
• Versions are easier to see — each plugin's version is now a tag next to its name. After Check updates, a plugin whose source has a new version shows 1.0.0 → 1.1.0 in the list and in the preview.

MCP connections​

• Keep a server without sending it to any Agent — untick every Agent in the server dialog to keep a server in Skillshare while no Agent receives it, for example to take it out of use for a while without losing its settings. The next sync removes the entries it had, and ticking an Agent brings it back. Project servers work the same way. In the terminal, use --target none; in config.yaml, an empty list:

  skillshare mcp edit context7 --target none    # stays in Skillshare, written to no Agent
  

  Refs: #289.
• Set other pi-mcp-adapter fields from Skillshare — the adapter has per-server fields that Skillshare has no setting for, such as excludeTools and approveTools. Put them under piOptions and they are written into the server's entry in Pi's file as given. The server dialog has a JSON box for them under Direct tools, which checks that the text is a JSON object before saving. In the terminal:

  skillshare mcp edit github --pi-options '{"excludeTools":["*emulator*"]}'
  

  A field removed from piOptions stays in Pi's file until you delete it there. Refs: #289.

Upgrade​

• skillshare upgrade shows download progress — on a slow connection the release download can take minutes, and the spinner alone made the upgrade look stuck. In a terminal it now shows how much has been downloaded, for the CLI and for the dashboard assets:

  Downloading v0.21.4...  3.2 MB / 9.1 MB
  

Bug Fixes​

Plugins​

• Pi and OpenCode are no longer blocked by a scoped npm name — a package.json named like @owner/pkg made both Agents show the plugin as blocked, although they install it by path.
• Pi packages whose package.json has no name can be installed — they were blocked for Pi even when another manifest named the plugin.

Config​

• Installing a plugin no longer reindents config.yaml — saving a plugin rewrote the whole file with 4-space indentation. Every YAML file Skillshare writes, including skill and agent frontmatter and audit-rules.yaml, now uses 2 spaces.

MCP connections​

• An entry left behind by a deleted configuration can be taken over — when the Skillshare configuration that owned an Agent's MCP entry was moved or deleted, the conflict still said the entry was managed by another configuration and named a file that is no longer there, leaving no way forward in the dashboard. Such a conflict now reports the entry as left over, and Import from the Agent or Replace with source takes it over, in the terminal and from the conflict in the dashboard. A configuration that only cannot be read, such as one on a drive that is not mounted, still counts as present. Refs: #288.
• OpenCode project settings kept in .opencode/ are used — OpenCode also reads opencode.json and opencode.jsonc from a project's .opencode/ folder. A file kept there was ignored, and sync created a second one at the project root. It is now the file Skillshare writes to; a new file is still created at the root. Refs: #289.
• Turning a global server off in a project follows the project's Agents — the dashboard saved the switch with every Agent the global server reaches, including Agents the project does not use, and the list went stale when the project's targets changed. The switch now stores no Agent list and applies to the Agents the project uses that have a per-project switch. The row shows those Agents as logos and says which of the project's Agents still load the server. An entry saved by an earlier version shows Match the project to update it. Refs: #289.
• The Sync page says when a change only turns a server off — a project's switch for Claude Code is kept in ~/.claude.json, and it read like the global server being added or removed there. Such a change now says it turns the server off or back on, and names the project. A conflict on a switch offers only Replace with source, since importing reads the Agent's global file. Refs: #289.

Dashboard​

• Agent logos show on Windows machines that report the wrong type for SVG files — a few larger logos, Antigravity among them, appeared as broken images on some Windows machines while every other logo was fine. Refs: #289.
• Going back to the global targets asks first — switching a project's MCP targets back to Same as global saved at once and dropped the project's own list.
• Status labels and Kilo Code messages follow the dashboard language — enabled on the Skills page, In sync on the Extras page, the Kilo Code project warning and the error for an unreadable pasted snippet stayed in English. Refs: #289.

[0.21.3] - 2026-09-21​

New Features​

Dashboard​

• Analyze only the skills Skillshare manages — when a target also holds skills you added by hand, an Only skillshare-managed switch on the Analyze tab hides them, and the always-loaded and on-demand totals follow the switch. The dashboard remembers your choice. Local skills now show their folder name instead of an internal synced/… path; the full path appears on hover and in the skill's details.

Bug Fixes​

Plugins​

• One bad entry no longer blocks a whole marketplace — a marketplace failed to load when one entry used an external source, pointed at a missing folder, or had an invalid or duplicate name. Such an entry is now marked in its own row with the reason, in your language, and the rest of the marketplace can be installed. When one catalog points a plugin at an external source and another at a folder inside the repository, the folder is used.
• Claude Code plugins without plugin.json can be installed — Claude Code does not require the manifest. Such a plugin is now read from its marketplace entry and its default folders, and a SKILL.md at the plugin root counts as its only skill. Definitions in the entry, such as strict: false, skills and lspServers, and its version, are carried into the install.
• Entries that share one folder install their own skills — when several entries in a marketplace pointed at the same folder and each listed its own skills, every one of them installed the same content.
• Plugins whose marketplace name differs from their manifest install in Claude Code — Claude Code installs such a plugin under the marketplace name, so it is no longer blocked there. Codex refuses the mismatch, so other Agents still show it as blocked.
• Each Agent installs from its own folder — when a marketplace ships a separate copy of a plugin for each Agent, every Agent now gets the copy made for it instead of the first one found.
• Broken links no longer block a plugin source — a single broken symlink anywhere in a repository, or one that pointed outside it, made every plugin in it fail. Such links are now left out of the install and listed in a warning. They are still never followed.
• Add plugin no longer preselects a plugin that cannot be installed — when the only plugin in a source was blocked, the dialog chose it for you.
• Pi's version is shown — Pi prints its version to stderr, so the plugin list showed it without one.

MCP connections​

• View what each Agent gets shows a spinner while it loads — an empty dialog gave no sign whether the preview was still coming.

[0.21.2] - 2026-09-21​

New Features​

Projects​

• Manage many project folders from the global config — a projects: key declares a folder once, with the tools used there. sync then treats each one as a <project>@<tool> target with its own mode and filters, so you no longer list every project skills folder as a custom target or run project mode in each repo.

  # ~/.config/skillshare/config.yaml
  projects:
    ~/work/project01:
      targets: [claude, codex]
      skills:
        mode: copy
        include:
          - myskill-*
      agents: {}
  

  skillshare sync --dry-run   # preview
  skillshare sync
  

  Tools that share a project skills folder are written once. A project whose folder is missing is skipped with a warning instead of failing the sync, and collect and target leave project targets alone. Refs: #286.

• MCP servers for many projects in one place — mcp.projects lists project roots, each with its own targets, servers and directTools. One sync mcp plans the global files and every root together, and removing a root cleans up what was written there.

  # ~/.config/skillshare/config.yaml
  mcp:
    projects:
      ~/work/project01:
        targets: [opencode, pi]
        servers:
          context7:                  # off in this project only
            disabled: true
  

  mcp.projects is refused in a project config. A disabled entry works for Claude Code, OpenCode, Kilo Code and Pi with pi-mcp-adapter. For Claude Code the switch is written to ~/.claude.json, where it keeps its per-project off list, and the server definitions in that file are left as they are. Codex is refused, because a switch for a server its global config lacks stops Codex loading its config at all. Refs: #286.

• Projects page in the dashboard — in global mode, a Projects page lists each project with Skills, Agents and MCP tabs. Adding a project takes a folder and its targets, and an ordinary target that already points inside a project can be converted. On a project's MCP tab, every global server has a switch that turns it off for that project.

Install​

• Project lockfile — project mode now writes .skillshare/skills.lock.json next to config.yaml. The config records which skills the project wants; the lockfile records the exact commit each one resolved to. Commit both, and skillshare install -p gives every teammate the same commit even after upstream moves on.

  skillshare install github.com/team/skills --all -p   # writes the lockfile
  git add .skillshare/ && git commit -m "Add team skills"
  
  skillshare install -p          # a teammate gets the locked commits
  skillshare update --all -p     # moves the pins forward; commit the lockfile again
  

  Tracked repos are pinned too and stay on their branch, so they can still be updated. update -p and the dashboard move a pin only for skills whose commit changed, a tracked repo with uncommitted changes is left alone, and uninstall -p drops the pin.

• Pin an install to a tag or a commit — --branch now accepts a tag or a commit SHA as well as a branch name. check resolves tags and skips the network for full SHAs, which can never move.

  skillshare install github.com/team/skills --branch v1.2.0 --all
  skillshare install github.com/team/skills --branch 8f14e45fceea167a5a36dedd4bea2543ce848564 --all
  

  --track is refused with a tag or a SHA, in the CLI and the dashboard, because a tracked repo needs a branch to pull. Refs: #281.

• Gitea and CNB sources — skills install from Gitea (gitea.com and self-hosted) and CNB (cnb.cool), including subdirectories and web URLs. Private repos use GITEA_TOKEN and CNB_TOKEN, and self-hosted instances are listed under gitea_hosts and cnb_hosts in the config, or in SKILLSHARE_GITEA_HOSTS and SKILLSHARE_CNB_HOSTS. Subdirectory installs download through each platform's contents API and fall back to a git clone. Refs: #168.

MCP connections​

• directTools for Pi — pi-mcp-adapter can register a server's tools as individual Pi tools. Set it per server, as a default under mcp, or per project. It accepts true, false, search, or tool names separated by commas.

  skillshare mcp add docs --url https://example.com/mcp --target pi --pi-extension pi-mcp-adapter --direct-tools search --no-tui
  

• Edit MCP defaults in the dashboard — a Defaults section on the Servers tab edits mcp.targets and mcp.directTools, which could only be changed by hand before.

Analyze​

• analyze measures what each target actually loads — a symlink target exposes the whole source folder, so skills disabled in .skillignore are now counted for it and flagged disabled. For merge and copy targets, skills you dropped into the target folder by hand are counted and flagged local. Token estimates charge one token per wide character, so Chinese, Japanese and Korean text is no longer undercounted by roughly four times. The CLI, the TUI, the dashboard and the sync summary now report the same numbers.

Targets​

• antigravity-cli is its own target — the Antigravity CLI reads ~/.gemini/antigravity-cli/skills and never the app's folder, so as an alias of antigravity it received skills in a folder it does not read. A path-less antigravity-cli: entry in the config is now valid.

• Convert agents for a tool that reads a different format — an extension: on a target's agents block converts each agent while it syncs, reusing the extension mechanism of extras. The new built-in opencode-agents turns Claude-style agents into OpenCode agents. It keeps only the fields OpenCode documents and adds mode: subagent when missing. It refuses agents that limit their tools (tools, disallowedTools, permissionMode), because dropping those would give the OpenCode agent every tool. codex-agents works here too and writes .toml files. Converted targets are never collected back into your source. The dashboard sets the extension from the target's Agents tab, and the targets list tags a target that uses one. Refs: #267, #241.

  targets:
    opencode:
      agents:
        extension: opencode-agents
  

Dashboard​

• Beautify button in the config editor — reformats config.yaml in the editor, keeping comments, so the result is visible and revertable before you save.
• Expand the config editor — an Expand button opens the same editor and assistant panel in a near-fullscreen dialog, for files too long to read in the inline row. The dialog ignores backdrop clicks and Escape, so an edit in progress is never dismissed by accident.
• The install dialog opens on the URL tab — installing from a known repo URL is the common path. A ?install=search link still opens the search tab.
• Copy the config path from the sidebar — a copy button appears on hover next to the truncated path.
• Analyze follows the target in the URL — target and project pages link straight into the analysis of that target, and project targets are listed after your own with their tool icon.

Documentation​

• The docs site is available in four more languages — Traditional Chinese, Simplified Chinese, Japanese and Korean, at full parity with the English docs. The README is available in the same languages.

Bug Fixes​

• A skill installed into a group reaches the project config — install <git source> --into <group> -p, and installing a config entry that has group:, wrote the skill's metadata into the group folder. The skill was then missing from config.yaml, so teammates running install -p did not get it. Config installs were affected in global mode too.
• doctor suggests removing the right target for a discovery overlap — it pointed at the target that owns the shared folder, such as universal, and removing that hides skills from every tool reading the folder. It now suggests removing the scanning target, such as codex. Refs: #135.
• Overlap warnings match each tool's documentation — the list of extra folders each tool scans covered nine targets and had drifted. Kimi no longer gets a false warning for ~/.agents/skills, and OpenCode, Goose, Copilot, Crush, Droid, Pi, Cline, Command Code, Deep Agents, Kilo Code, OpenClaw, OpenHands and Kode now get a warning when paired with a target whose folder they also read. Refs: #135.
• The dashboard's Update now no longer hangs waiting for a password — when the binary lives in a root-owned folder, the upgrade waited up to ten minutes for a sudo prompt nobody could see. It now fails at once and names the terminal command to run. Cached credentials and NOPASSWD setups still upgrade. The release download also has a timeout.
• A rejected target save in the dashboard changes nothing — when one field was invalid, such as a malformed include pattern, fields sent in the same save were still applied in memory and written by the next successful save.
• The Gitea token stays on the API host — a download URL on another origin is refused and the install falls back to a git clone, so a hostile server cannot collect GITEA_TOKEN.

Breaking Changes​

• The codex, goose and openhands targets moved to .agents/skills — ~/.agents/skills globally and .agents/skills in a project, because each tool now documents that location and reads its old folder only for backward compatibility. With universal also configured, the old default showed every skill twice. A global config stores full paths, so an existing one keeps its path and its doctor overlap warning. A project config stores target names only, so it follows the new default on its next sync -p, which also removes the links Skillshare left in the old folder; folders you made by hand there are kept. A fresh init sets up only universal for Codex. Refs: #135.

[0.21.1] - 2026-09-20​

New Features​

MCP connections​

• Turn off a global MCP server in one project — a server in an Agent's global config loads in every project. In project mode, add an entry with the same name and mark it disabled to stop it loading in that project only. The Agent keeps the command or URL from its global entry.

  cd my-project
  skillshare mcp add company-docs --disabled --target claude --target opencode
  skillshare sync mcp
  

  # .skillshare/config.yaml
  mcp:
    servers:
      company-docs:
        disabled: true
        targets: [claude, opencode]
  

  Works with Claude Code, OpenCode, Kilo Code, and Pi with pi-mcp-adapter. OpenCode, Kilo Code and Pi get a lone switch in their project file. Claude Code takes a whole entry from one scope, so Skillshare adds the name to that project's off list in ~/.claude.json, the one the /mcp panel edits; a name you turned off yourself is never claimed or removed. Other clients are refused, because a lone switch would replace the server instead of turning it off. In the dashboard, open it from the project folder, choose Add server and pick Off in this project. Refs: #286.

• Kilo Code is an MCP client — kilocode joins the supported clients, in global and project scope. It uses OpenCode's format; Skillshare writes to whichever kilo.jsonc or kilo.json already exists and creates kilo.jsonc when there is none. Kilo ignores a project config that holds an environment reference, so fromEnv and bearerToken are refused for Kilo in project mode before anything is written. Refs: #287.

• Claude Code local scope servers are reported — a server added with claude mcp add and no --scope wins, whole, over one of the same name in .mcp.json or the user scope. In project mode the plan and the dashboard now point out such a server next to the entry it hides, with the command that removes it. The sync is not blocked.

• Limits of each client are caught before the write — Claude Code skips the reserved names workspace, claude-in-chrome and computer-use, and reads its own credentials such as ANTHROPIC_API_KEY as empty in a remote server's url and headers; both are now refused for Claude with a message that says what to do. The dashboard's config preview refuses the same things saving would.

Dashboard​

• Manual only skills are visible — a skill with disable-model-invocation: true carries a manual only tag in the skills list, on its tile and on its detail page, matching the badge the list TUI shows after M. Refs: #283.
• Add field explains each frontmatter field — the skill editor's Add field menu shows what each field does under its name, instead of bare keys such as context and shell.
• Agent icons in the MCP import picker — the Agent dropdown in Import from a target shows each Agent's logo. Factory, LM Studio, Kilo Code and Claude Desktop now have their own icons.

Bug Fixes​

• The MCP page detects Agents that have no MCP file yet — detection looked only for the MCP file, so a fresh Claude Code install showed as not detected. An Agent's own settings folder now counts. In project mode the page listed every Agent as detected; it now lists those with a project MCP file or a global install.
• Cline's MCP settings are found again — Cline moved its settings to ~/.cline/data/settings/, shared by the VS Code extension, the CLI and the SDK. Skillshare now writes there and honors CLINE_MCP_SETTINGS_PATH, CLINE_DATA_DIR and CLINE_DIR. The old VS Code extension path is still used while it is the only one that exists.
• An MCP entry owned by a deleted config can be taken over — an entry managed by a Skillshare config that was later moved or deleted stayed blocked forever as "managed by another Skillshare config". An explicit import or replace now takes it over, and the conflict names the owning file.
• --disabled is refused where it does nothing — only mcp add uses the flag, but the other mcp commands and sync mcp accepted it and ignored it.
• A relative XDG_CONFIG_HOME is ignored for MCP paths — as the XDG specification requires, instead of producing a path relative to the current directory.
• Skill detail and new skill pages — the file viewer no longer leaves a gap above it when it sticks while scrolling, and the info column beside a long SKILL.md stays in view.

Breaking Changes​

• The kilocode skills target moved to .kilo/skills — from .kilocode/skills, in both global and project scope, because that is the only location Kilo Code's documentation lists now. Run skillshare sync once to write skills to the new location.

[0.21.0] - 2026-09-19​

New Features​

MCP connections​

• Define an MCP server once, sync it into each Agent's own config — a server lives in config.yaml (or a separate mcp.yaml) as a portable definition, and sync mcp writes it in the format each client expects: JSON, JSONC, TOML or YAML, under that client's own key. Nothing is started or proxied; Skillshare only manages the settings.

  skillshare mcp add                                   # guided: URL, command, or pasted JSON
  skillshare mcp add docs --url https://example.com/mcp --target claude --target cursor
  skillshare mcp import docs --from claude --sync      # adopt what an Agent already has
  skillshare sync mcp --dry-run                        # preview every file that would change
  skillshare sync mcp
  

  Supported clients: Claude Code, Codex, Cursor, VS Code, OpenCode, Grok CLI, Antigravity, Amp, Claude Desktop, Cline, Copilot CLI, Factory, Gemini CLI, Goose, Junie, Kiro, LM Studio, Warp, Windsurf and Pi. Global and project scope are both supported where the client has them. skillshare mcp without arguments opens a TUI for browsing, editing and syncing.

• Secrets stay out of the files — headers and environment values are written as {fromEnv: VARIABLE} references and rendered in each client's own reference syntax. On import, literal tokens and values carrying a URL password, such as database DSNs, are turned into references, and credential-like arguments are flagged.

• Only entries Skillshare wrote are ever changed — ownership is tracked per entry. An entry you wrote by hand is left alone even when it matches the source, and stays that way until you import it. A managed entry edited in the Agent is reported as a conflict instead of being overwritten, Agent-only fields such as timeouts survive a sync, and previews stay valid while an Agent rewrites unrelated settings in the same file.

• Backups and restore — every write backs up the Agent file first, keeping the newest 20 per file.

  skillshare mcp restore BACKUP_ID --dry-run
  skillshare mcp restore BACKUP_ID
  

• Pasted snippets are recognised by shape — mcp import --file and the add wizard pick the client format from the top-level key, so VS Code servers and OpenCode mcp snippets copied from a provider's docs import without naming a client. TOML still needs --from, because Codex and Grok share the format.

• Pi through a chosen extension — Pi has no built-in MCP support, so a server targeting Pi names the third-party package that reads the file, pi-mcp-adapter or pi-mcp-extension. Install one in Pi yourself; Skillshare only writes its configuration.

  skillshare mcp add docs --url https://example.com/mcp --target pi --pi-extension pi-mcp-adapter --no-tui
  

• sync --all includes MCP — it syncs skills, agents, extras and MCP settings together, and --json reports the MCP plan alongside the rest.

Plugins​

• Install a complete plugin and choose which tools receive it — a plugin bundles skills, hooks, MCP settings and scripts that only work together. plugin keeps that package intact and installs it through each Agent's own CLI, so native formats and unrelated installations are untouched.

  skillshare plugin add                                          # guided: source, plugin, targets, review
  skillshare plugin add owner/repo --target claude --target codex --no-tui
  skillshare plugin import review@team --from claude             # adopt an existing native installation
  skillshare plugin disable review --target codex                # save the selection only
  skillshare sync plugins --dry-run                              # then install or remove on sync
  skillshare plugin check review
  skillshare plugin update review --target claude
  

  Install targets: Claude Code, Codex, Cursor, Antigravity Desktop, Antigravity CLI, GitHub Copilot CLI, Pi and OpenCode. Grok Build supports import and removal, with install and trust handled in Grok. Sources can be a local folder, owner/repo or an HTTPS Git URL, pinned with --source-ref. Plugins are synced with sync plugins and are not part of sync --all.

Dashboard​

• Redesigned in two styles — the dashboard now comes in Clean and Playful, each in light and dark, switched from the theme menu. Skills and Agents are separate pages that hold their own install, search, updates and trash, and General, Backup, Log, Health, Extensions and Files are collected into one Settings page. Old routes redirect.
• MCP page — one row per server with its Agents as toggles. Add a server from a URL, a command, a pasted snippet or a file, or import what an installed Agent already has. Conflicts explain their cause and offer Import or Replace, backups are grouped by day with a preview before restoring, and View what each Agent gets shows exactly what each Agent would receive, including edits you have not saved yet. MCP settings are only served when the dashboard is opened through localhost or an IP address.
• Plugins page — one row per plugin with its Agents as toggles, the same as MCP. Expanding a row also lists the other Agents the source supports, so ticking one previews an install there. View files opens a read-only browser of the local copy Skillshare reviewed.
• Hubs page — browse a hub and install from it, or assemble your own index from installed skills, validate it and export it, all at /hubs. A draft can be browsed straight away, before it is hosted anywhere.
• Filter skills and agents by target, and a Remote source filter for skills from GitLab, Gitea, self-hosted and SSH sources, which previously matched no filter. Refs: #278.
• Config editors explain themselves — the Files tab and the audit rules editor keep a panel beside the YAML showing what the field under the cursor does and the unsaved changes. Audit rules also get a Test tab that runs a rule's regex against pasted lines.
• Dashboard source counts link to their pages, and plugins are counted alongside the other kinds.

Audit​

• Findings accepted with --force are remembered — a skill that legitimately quotes attack strings, such as a security scanner or red-team notes, no longer has to be forced on every update --all. The accepted findings are recorded for that skill, and later updates treat the same rule matching the same text as acknowledged. Any new finding, or the same rule matching different text, still blocks. Refs: #279.

List TUI​

• Press M to make a skill manual only — toggles disable-model-invocation in the selected skill's SKILL.md, so the skill stays installed and can be invoked by name but is no longer loaded by the model on its own. The detail panel shows a manual only badge. For a tracked or installed skill the TUI asks first, because the edit counts as a local change; pressing M again restores the file exactly. Refs: #283.

Bug Fixes​

• sync --dry-run agents and sync -g extras no longer sync skills — the resource kind was only recognised as the first argument, so putting a flag before it silently synced skills instead. The kind now matches in any position.
• A mistake in the mcp section no longer breaks every command — an invalid MCP setting is reported by MCP commands and the config editor, while sync, install, list and the dashboard keep working.
• Web UI: saving the config keeps flow-style YAML — a hand-written targets: [claude, codex] is no longer expanded into a block list on every save. Indentation is still normalised and comments are preserved.

[0.20.29] - 2026-09-11​

Bug Fixes​

• install -p no longer removes the skills it just installed from config.yaml — a project install driven by config.yaml reconciled against the metadata it had read before installing, so a skill that arrived as a plain copy — any source pointing at a subdirectory, which has no .git to fall back on — looked absent and its entry was pruned from the declarative list. The install itself reported success and the skill and its metadata landed correctly, but the entry was gone from config.yaml, which broke the install -p && sync flow for anyone setting the project up from a clean checkout. Metadata is now re-read from disk before reconciling, which also covers project-mode installs made through search. Refs: #280.

[0.20.28] - 2026-09-09​

New Features​

• Visible skillshare/ project directory — repositories that treat skills as reviewable content, rather than tool state, can now use a visible skillshare/ directory instead of the hidden .skillshare/.

  skillshare init -p --visible    # create skillshare/ instead of .skillshare/
  

  Everything lives inside whichever directory is in use: config.yaml, skills/, agents/, extras/, and the operational trash/, backups/ and logs/. Detection checks .skillshare/config.yaml first and skillshare/config.yaml second, so existing projects are unaffected and .skillshare/ wins when both exist. init -p without the flag still creates .skillshare/. To move an existing project, run mv .skillshare skillshare followed by skillshare sync -p to repair target symlinks, and update any sources paths in config.yaml that name .skillshare/ explicitly. Refs: #256.

Bug Fixes​

• .skillignore matches again when the file uses CRLF line endings — rules were compiled with a trailing carriage return that could never equal a path segment, so a .skillignore saved on Windows silently ignored nothing while status still reported its patterns as loaded. Editing patterns was affected the same way: existing entries were not found, adding one duplicated it, and removing one failed. A regression since v0.17.4. Refs: #275.
• A directory symlink inside a skill no longer discards every file hash — hashing stopped at the first entry that could not be read as a file and threw away the hashes computed so far, which left installed skills without the metadata used to detect local edits. Directory symlinks are skipped; file symlinks are still hashed, and real failures such as broken links still surface. Refs: #272.

[0.20.27] - 2026-09-02​

New Features​

• Agents can declare which targets they belong to — a targets list in an agent's frontmatter now restricts that agent to the listed tools, the same way metadata.targets works for skills. Agent files are copied verbatim and Claude Code, OpenCode, Cursor and Copilot read different frontmatter fields, so this lets you keep a per-tool variant of the same agent side by side. Agents without the field still sync everywhere; target aliases such as claude-code match claude. Applies to sync agents, the dashboard, doctor and the target summary. Refs: #267.

  # ~/.config/skillshare/agents/reviewer.md
  ---
  targets: [claude]
  tools: Read, Write, Bash(git log)
  ---
  
  # ~/.config/skillshare/agents/reviewer-opencode.md
  ---
  targets: [opencode]
  mode: subagent
  permission: { read: allow, write: allow, bash: ask }
  ---
  

Bug Fixes​

• update --all no longer deletes a skill when the audit blocks its update — when several skills from one repository were updated together, the existing skill directory was removed before the new content was copied in, and an audit block then cleaned up the new content as well, leaving nothing on disk and a stale metadata entry that later updates could not find. Grouped updates now stage the new content in a temporary directory, run the audit there, and only swap it into place once the audit passes, matching what update <name> already did. If a skill was already lost this way, skillshare update <name> --force reinstalls it. Refs: #271.
• Batch updates and the web UI follow the branch a skill was installed from — update --all cloned the remote default branch for skills installed with -b <branch>, so they were reported stale or downgraded, and --prune moved them to trash. The dashboard's check and update did the same, comparing against the remote HEAD instead of the installed branch. All of these now group by repository and branch and fetch from the installed branch. Refs: #268.
• Windows drive-letter paths install as local skills — skillshare install D:\skills\my-skill failed with unrecognized source format because only paths starting with /, ~, ./ or ../ were treated as local. Drive-letter paths (D:older, C:/Users/...) and backslash-relative paths (.\skill) are now recognised in the CLI and the dashboard install page. Refs: #269.
• Dashboard labels non-GitHub git sources as Remote — skills installed from Gitea, GitLab, self-hosted or SSH sources showed a Local badge because only GitHub metadata types were recognised. Any non-local git source now shows Remote, matching skillshare list. Tracked repos and GitHub sources keep their existing badges. Refs: #270.

[0.20.26] - 2026-08-27​

Bug Fixes​

• update --force can override the security audit again — when an update was blocked by an audit finding, the error told you to pass --force, but the flag never reached the audit gate. Update stages new content in a temporary directory and had to leave its internal overwrite flag off to keep the gate active, which discarded your --force along with it, so a flagged update could not be applied short of --skip-audit — which turns scanning off entirely. --force now reaches the gate on every update path: regular skills, tracked repos, agents, and the web UI's Force Retry button.

  skillshare update my-skill --force       # apply despite audit findings
  skillshare update my-skill --skip-audit  # skip scanning entirely
  

  Audit scan failures stay fail-closed regardless of --force: accepting findings you have seen is not the same as proceeding when the scanner could not run.

• install --json no longer bypasses the security audit — --json set the internal overwrite flag so installs could run non-interactively, and the audit gate read that same flag, so JSON-mode installs silently accepted content that would have been blocked interactively.

• Grouped batch updates are covered by the audit gate — updating several skills from a single repository inherited the same overwrite flag and skipped the block threshold.

• Web UI: Force Retry only appears where it can help — a failure such as failed to remove existing skill: ... permission denied offered a Force Retry button that retried with force and failed identically, with no hint of what would actually fix it. The button now shows for audit blocks and for pulls the server would retry with force, and is hidden elsewhere.

• Web UI: update errors no longer quote CLI flags — messages ending in Use --force to override or --skip-audit to bypass scanning were rendered verbatim in the dashboard, where there is no command line to type them into.

Breaking Changes​

• A --force at install time no longer exempts later updates from the audit gate — --force is a per-command decision. A skill installed with --force is scanned again on its next update, and needs --force (or --skip-audit) again to apply findings at or above the block threshold.

[0.20.25] - 2026-08-10​

Bug Fixes​

• Batch updates no longer mark skills under target dot-directories as stale — skills installed from subdirectories such as .claude/skills/... or .codex/skills/... were skipped by discovery during batch updates and reported as deleted upstream, which could incorrectly suggest --prune. Batch update now resolves the requested subdirectory directly before declaring it missing. Refs: #261.
• Explicit target dot-directory installs discover their requested content — commands such as skillshare install user/repo/.claude no longer return zero results just because .claude is normally excluded from repository-wide discovery. Explicitly requested roots are scanned for both skills and agents, while repository-root scans continue to skip synced target copies.
• Project commands preserve existing projects when config.yaml is missing — when .skillshare/ already contains skills or agents but .skillshare/config.yaml has disappeared, project-mode commands now stop with recovery guidance instead of silently re-initializing an empty config, dropping target configuration, and leaving stale links. Fresh projects and shared repositories that intentionally gitignore config.yaml still initialize automatically.
• Windows can uninstall nested skills again — uninstalling a skill stored under a folder no longer passes Windows backslashes into the trash-name validator and fails with trash name must not contain backslash. Nested skill names are normalized to slash-separated paths in global and project mode without weakening traversal checks. Refs: #264.

[0.20.24] - 2026-08-03​

Bug Fixes​

• Gemini CLI and Antigravity are separate targets again — gemini had been folded into antigravity as an alias, but the two runtimes read different global skill directories, so one of them was always pointed at the wrong path. Antigravity now resolves to ~/.gemini/config/skills (global) and .agents/skills (project); gemini is a target of its own resolving to ~/.gemini/skills and .gemini/skills, with ~/.agents/skills and .agents/skills still scanned as fallbacks. Aliases: gemini-cli for Gemini CLI, antigravity-cli for Antigravity. Refs: #255.

  skillshare target add gemini        # Gemini CLI
  skillshare target add antigravity   # Antigravity
  skillshare sync
  

  Antigravity's skill scanner also skips symlinked skill directories entirely — silently on macOS and Linux, and as Incorrect function on Windows. The troubleshooting docs now cover this along with its two workarounds: switch the target to copy mode, or point Antigravity at your skillshare source directory via its Skill Custom Paths setting.

• doctor's symlink compatibility hint is deterministic — the hint chose its example target by iterating a map, so the same config produced a different suggestion on each run, and it could name a target whose runtime handles symlinks fine — telling users to switch something that was not broken. The example is now drawn only from the targets known to skip symlinked skill directories, in a fixed order. The sync docs are also corrected: the hint is printed by doctor, not sync.

Breaking Changes​

• gemini no longer resolves to the Antigravity target — configs that used gemini (or gemini-cli) to reach Antigravity now sync to Gemini CLI's own directory instead. Use antigravity for Antigravity. Existing antigravity targets change global path from ~/.gemini/skills to ~/.gemini/config/skills; run skillshare sync after upgrading.

[0.20.23] - 2026-07-30​

New Features​

• list --status filters enabled or disabled entries outside the TUI — the enabled/disabled filter previously existed only inside the list TUI (s key, status: search tag), so scripts had to know that disabled is omitted for enabled entries and filter the JSON themselves. --status combines with the pattern and --type using AND semantics, and works in project mode and for agents.

  skillshare list --status disabled
  skillshare list --status enabled --json
  

  --status all is the default and produces output identical to omitting the flag. Refs: #244.

Bug Fixes​

• Backups no longer copy your source, and old snapshots are pruned automatically — pre-sync backups followed merge-mode skill symlinks and copied the resolved content, so every snapshot duplicated the source directory, including files kept out of Git but still present on disk such as model weights, .venv, and browser profiles. Retention only ran when backup --cleanup was invoked by hand, so nothing removed the accumulating snapshots; backup directories could reach hundreds of gigabytes until sync failed with no space left on device. Snapshots now capture only local target content, and the existing retention policy (10 snapshots, 30 days, 500 MB) runs after every automatic backup. Refs: #252.
• backup --cleanup keeps the newest snapshot when it exceeds the size cap — a single snapshot larger than the 500 MB cap previously removed every backup including the most recent one, leaving no restore point at all.
• Backup previews and cleanup now agree — backup --cleanup --dry-run no longer reports that it will delete an oversized newest snapshot when the actual cleanup keeps it. Retention also counts only snapshots that remain after cleanup, avoiding unnecessary deletion of older snapshots that still fit.
• Failed backups are discarded instead of appearing as restore points — if a file cannot be copied, the partial snapshot is removed and cannot consume the newest retention slot. Automatic cleanup failures are now shown as warnings instead of being silently ignored.
• Targets holding only symlinks no longer produce empty snapshots — an empty restore point consumed a retention slot and evicted older snapshots that did have content.
• doctor handles linked target directories correctly — in project symlink mode the valid link ../.skillshare/skills was reported as pointing to the wrong location, because relative links were resolved against the current working directory instead of the link's parent. Symlink-mode targets no longer report every source skill as a duplicate target-local copy, while copy-mode targets reached through a symlink still report real duplicate copies. Refs: #251.

Breaking Changes​

• Backups capture local target content only — merge-mode skill symlinks are skipped, so restoring a target recovers its local skills and then needs skillshare sync to recreate symlinks for synced skills. Copy-mode targets are unaffected because they hold real files. The documented backup location is also corrected to ~/.local/share/skillshare/backups/ (the XDG data directory), which had been listed under ~/.config in some places.

[0.20.22] - 2026-07-21​

New Features​

• Grok CLI target — added xAI's Grok CLI as a built-in target, syncing skills to ~/.grok/skills (global) and .grok/skills (project), with legacy fallback to ~/.agents/skills. Aliases: xai, grok-cli.

  skillshare target add grok
  skillshare sync
  

Bug Fixes​

• collect --json no longer forces overwrites — JSON mode previously implied --force, silently overwriting existing skills and agents in the source. It now still skips confirmation prompts but keeps the overwrite guard, so existing resources are preserved unless --force is passed.
• uninstall --json no longer bypasses the uncommitted-changes guard — JSON mode previously implied --force, removing tracked skills that had uncommitted changes without warning. Dirty repositories now return a structured error unless --force is passed.
• Batch uninstall explains all-dirty failures — uninstalling multiple tracked skills where every repository has uncommitted changes now reports why nothing was removed instead of failing without explanation.
• Editing a skill's source no longer rewrites the source directory's Git remote — in the dashboard, changing a nested skill's source URL could overwrite the origin of the skills source directory itself when that directory is a Git repository (for example, backed by Git Sync). Source edits now only affect a tracked skill's own repository.
• Freshly installed skills show their correct source instead of "Local" — after installing a skill from the dashboard, the resource now immediately shows its GitHub source and type, rather than appearing as a local skill with an empty source until the server was restarted.

[0.20.21] - 2026-06-29​

Bug Fixes​

• Copy-mode sync respects file ignore patterns — sync in copy mode now skips configured ignore: artifacts, plus default .DS_Store, .git/, and __pycache__/ files, across CLI sync, diff, and dashboard sync/diff paths. This prevents ignored cache and build artifacts from being copied into targets or changing copy-mode checksums.
• Global sync uses the default skills source when source is omitted — global configs that only define targets now work with skillshare sync --global, using the default ~/.config/skillshare/skills source instead of failing with source path is empty. Refs: #238.

[0.20.20] - 2026-06-19​

Bug Fixes​

• Self-managed GitLab URLs with deep project paths install correctly — generic HTTPS sources such as https://domain.com/dir1/dir2/dir3/dir4 now retry deeper repository boundaries when the initial dir1/dir2 clone is not a Git repository, so nested GitLab projects can install without adding .git or configuring gitlab_hosts. Authentication, SSL, branch, and network errors still fail directly instead of retrying unrelated paths.

[0.20.19] - 2026-06-17​

Bug Fixes​

• init shows shared skills directories as a single universal target — when a detected CLI uses the shared ~/.agents/skills directory, skillshare init now presents the shared directory guidance instead of listing each matching CLI target separately.
• Trash restore handles nested entries and current-directory restores again — nested trashed skills returned by skillshare trash list can be restored with their slash-separated names, and restoring to . no longer fails the path safety check while sibling-prefix escapes are still rejected.

[0.20.18] - 2026-06-16​

Bug Fixes​

• Dashboard project-root installs are rejected before copying into themselves — the web dashboard now rejects project-mode installs where the local source resolves to the project root, matching skillshare install ./ -p and returning the same "install a specific skill subdirectory" guidance instead of recursively copying .skillshare/skills into itself.
• Trash operations reject traversal-style names — moving skills or agents to trash, restoring from trash, and automatic trash cleanup now validate trash-relative names and destination paths before filesystem writes or removals. Traversal segments, absolute paths, backslashes, NUL-containing names, and empty names are rejected while safe nested names like org/team-skill and demo/my-agent continue to work.

[0.20.17] - 2026-06-15​

Bug Fixes​

• update --all no longer crashes on unreadable metadata — when .metadata.json is corrupt or unreadable, update now reports a metadata warning and continues scanning local skills instead of panicking with a stack trace.
• Project root installs are rejected before copying into themselves — skillshare install ./ -p now fails with a clear message telling users to install a skill subdirectory, preventing .skillshare/skills from recursively copying into itself on Windows.

[0.20.16] - 2026-06-15​

Bug Fixes​

• Repository subdir installs reject traversal paths — source parsing now rejects repository subdirectories with absolute paths, ./.. segments, backslashes, NUL/control characters, or encoded traversal before install and download flows use them. Inputs such as github.com/owner/repo/../../etc/passwd and unsafe blob SKILL.md paths now fail instead of resolving outside the repository boundary. Refs: #224.
• Metadata files stay readable after atomic saves — install and update operations now write .metadata.json with 0644 permissions, so Git and other tooling can read metadata after Skillshare replaces the file.

[0.20.15] - 2026-06-14​

Bug Fixes​

• Git branch refreshes now fail visibly when fetch fails — the dashboard no longer serves stale remote branches or continues a checkout after git fetch fails. Branch listing and checkout requests now report the fetch failure so users can fix connectivity, authentication, or remote problems before switching branches.
• Source URL edits keep metadata and remotes in sync — updating a tracked skill or agent source now updates the Git remote before saving metadata. If the remote update fails, the API returns an error and leaves the existing source metadata unchanged instead of reporting success with an old on-disk remote.
• Target removal preserves config when cleanup fails — removing a target from the dashboard now stops if Skillshare cannot inspect the target, remove the target symlink, remove the target manifest, or clean managed symlinks. The target remains configured so users can fix the filesystem issue and retry instead of losing the target entry.
• Version checks handle release tag formats correctly — update checks now accept versions with a leading v prefix while still rejecting malformed version segments. Local metadata builds only advertise a release version when built from a clean exact tag; non-release builds stay in dev mode so update checks do not compare against commit-describe strings.
• JSON-mode automation stays clean during cleanup warnings — temporary Git clone cleanup failures are still logged for human-readable flows, but cleanup warnings no longer leak into --json stderr output.
• Skill linting reports rule load failures instead of panicking — malformed embedded lint rules now return explicit errors through analysis/discovery paths, and repeated lint runs keep the load error instead of losing it after the first attempt.
• Audit finding severity dots are vertically centered — severity indicators in the dashboard Audit findings list now align with their badges and messages.

[0.20.14] - 2026-06-13​

Bug Fixes​

• Push failures redact token-auth URLs without losing diagnostics — failed Git push flows now sanitize credential-bearing error output before it reaches CLI/API/UI callers, while still preserving useful Git and pre-push hook diagnostics. Refs: #214.

[0.20.13] - 2026-06-11​

New Features​

Web Dashboard​

• Rehydrate missing tracked repos from the dashboard — the Updates page now shows a warning banner listing tracked repos declared in .metadata.json whose clone directories are missing on disk, with a one-click Rehydrate button that re-clones them from metadata. The Dashboard's Update All also warns about missing repos and points to rehydrate, instead of reporting that there is nothing to update. Refs: #212.

Bug Fixes​

• update --all reports missing tracked repos in batch and project mode — reporting a missing tracked repo previously only worked when it was the single update target; when update --all covered multiple items (the common case) the batch path skipped missing repos silently, and project mode (-p) never detected them at all. Both now surface each missing repo with a warning and a one-shot rehydrate hint:

  skillshare update --all          # ! _team-skills  clone directory absent
  skillshare install               # rehydrate from metadata
  

  update --all --json now carries an aggregated missing_tracked_repos summary (names + hint), and the per-item error is the concise clone directory absent. Refs: #212.

[0.20.12] - 2026-06-11​

New Features​

• Droid syncs custom droids as agents — the droid target now distributes custom droids (.md files with YAML frontmatter) alongside skills, mapping them to ~/.factory/droids (global) and .factory/droids (project) through the existing agents sync. The target also accepts factory as an alias:

  skillshare target add factory   # same as: skillshare target add droid
  skillshare sync agents
  

  Refs: #213.

Bug Fixes​

• Project-mode agent symlinks are now relative — skillshare sync agents created absolute symlinks in project mode, which broke when the repository was moved or checked out on another machine. Agent symlinks now use relative paths, matching how project skill symlinks already work.
• Factory alias syncs Droid agents correctly — adding the Droid target by its factory alias now also resolves the built-in agents path, so skillshare sync agents writes custom droids to ~/.factory/droids or .factory/droids instead of skipping the target as agentless.
• Web UI sync respects agent filters — syncing from the dashboard now honors target-level agents.include and agents.exclude filters, matching the CLI. Agents that become excluded are pruned from target directories on the next sync. Refs: #211.

[0.20.11] - 2026-06-10​

Bug Fixes​

• Grouped tracked repositories rehydrate at the correct path — when .metadata.json contains tracked repos installed with --track --into <group>, skillshare install now restores the missing clone at the original grouped path instead of applying the group twice and failing on paths like anthropics/anthropics/_skills. Refs: #212.

[0.20.10] - 2026-06-10​

Bug Fixes​

• Missing tracked repositories are no longer silently ignored — when .metadata.json declares a tracked repo but the local _repo/ clone is missing (common after a fresh clone on another machine), status, check, update --all, and doctor now report it as missing instead of showing no tracked repos. The message points to the existing recovery path:

  skillshare install
  skillshare sync
  

  update --all --json now counts the missing repo as skipped and includes an item explaining the recovery step. Refs: #212.

[0.20.9] - 2026-06-05​

New Features​

• Batch enable/disable in the web dashboard — the Resources page now has a selection mode. Click Select, tick multiple skills or agents across the grid, folder, or table view (folders offer a select-all checkbox), then enable or disable them all at once from the bottom action bar. Enabling applies immediately; disabling asks for confirmation first. Works for both skills (.skillignore) and agents (.agentignore). Refs: #203.

Bug Fixes​

• Fixed enabled: false being ignored for tier and cross-skill audit rules — disabling one of these rules in audit-rules.yaml marked it disabled in the rule listing, but the scan still fired it at full severity. Both the per-skill and single-file scan paths now honor the disabled rule. Refs: #204.

[0.20.8] - 2026-06-05​

Bug Fixes​

• SSH GitHub Enterprise hub entries inherit the hub SSH login — when an SSH hub returns same-host GitHub or GitHub Enterprise domain-prefixed sources, search results now install them over SSH using the hub username and host. For example, a hub loaded from acme@acme.ghe.com:Org/skills.git//hubs/team.json can return acme.ghe.com/Org/skills/skills/reviewer, and Skillshare installs it as acme@acme.ghe.com:Org/skills.git//skills/reviewer. Explicit HTTPS/SSH sources, cross-host entries, local paths, and in-memory indexes keep their existing behavior. Refs: #196.
• SSH skill previews can read private hub results — previews for SSH GitHub/GHE sources now fall back to a shallow clone when no token is available or the Contents API rejects the request, so the dashboard can show full SKILL.md content for SSH-only private hub results instead of only index metadata.

Performance​

• Faster .skillignore globstar matching — repeated ** patterns no longer trigger exponential backtracking, so commands that scan ignored skills stay responsive with complex ignore rules.

[0.20.7] - 2026-06-03​

Bug Fixes​

• Fixed single-skill uninstall for disabled skills in the dashboard — a skill hidden by .skillignore still appeared on the Resources page, but uninstalling it from the item menu or detail page could return "skill not found". Single-resource uninstall now resolves disabled skills the same way the list and batch uninstall flows do.

Performance​

• Trash page virtualization — the dashboard Trash page now renders long trash lists incrementally, so large skill or agent trash folders stay responsive instead of rendering every trashed item at once.

[0.20.6] - 2026-06-03​

New Features​

• Clearer hub errors in the web dashboard — when a hub fails to load, the Search page now names the failing hub and explains the likely cause (malformed URL, missing index file, authentication required, or invalid JSON) instead of showing a bare HTTP 400.

Bug Fixes​

• Fixed the skill preview showing only index metadata (name, description, tags) for skills from non-github.com hubs — the web dashboard always fetched SKILL.md via api.github.com, so GitHub Enterprise, GitLab, and other sources never rendered their full content. The preview now reads from the source's own host (the GHE Contents API, or a shallow clone for other platforms), and degrades with a clear notice when a source genuinely can't be fetched.
• Fixed the hub selector dropdown being clipped behind the search box on the dashboard's Search page.
• Fixed skillshare audit ignoring .skillignore inside tracked hub repos — skills excluded via a tracked repo's .skillignore were still scanned and reported. Audit now skips those skills, matching how sync and the rest of the CLI treat them.

[0.20.5] - 2026-06-03​

New Features​

• Zed editor target — zed is now a supported sync target. Add it, then sync; skills go to ~/.agents/skills (global) and .agents/skills (project):

  skillshare target add zed
  skillshare sync
  

Bug Fixes​

• Fixed disabled skills failing to uninstall from the web dashboard — a skill disabled via .skillignore still appeared in the dashboard list but couldn't be removed, reporting "skill not found". Uninstall now resolves disabled skills the same way the list does. The CLI was unaffected.

[0.20.4] - 2026-06-02​

New Features​

SSH hub sources​

• Fetch a hub index over SSH — skillshare search --hub and skillshare hub add now accept SSH URLs, so a shared hub index can live in a private or GitHub Enterprise repo that teammates reach over SSH without cloning it first. Skillshare shallow-clones the repo with your SSH keys and reads the index from it:

  skillshare search react --hub git@ghe.corp.com:team/skills.git
  skillshare hub add git@ghe.corp.com:team/skills.git//hubs/team.json --label ghe
  

  The index path inside the repo comes from the //path suffix and defaults to skillshare-hub.json at the repo root. Both scp-style (git@host:org/repo.git) and scheme-style (ssh://git@host/org/repo.git) URLs work. In the web dashboard, SSH hub sources must be saved first — the server only clones saved hubs.

Extras per-target management​

• Add or remove a single target on an existing extra — extras <name> --add-target and --remove-target manage one target without recreating the whole extra. Both update config only; run skillshare sync extras afterward to apply:

  skillshare extras rules --add-target ~/.cursor/rules --mode copy
  skillshare extras rules --remove-target ~/.cursor/rules
  

  Removing a target leaves already-synced files in place by default. Add --prune to also delete the skillshare-managed files under that target — in merge mode only symlinks are removed, so your own files are preserved. Removing the last remaining target is rejected (use extras remove <name> for the whole extra). The web dashboard's Extras page gains matching per-target add/remove controls.

Bug Fixes​

• Fixed extension downloads in the web dashboard's Config page clearing each other's progress — starting a second download no longer wipes the first one's loading spinner; each download now tracks its own state.

Breaking Changes​

• Removed the extras mode subcommand — change a target's sync mode or flatten setting with the extras <name> shorthand instead (same behavior, one less command):

  skillshare extras rules --mode copy --target ~/.claude/rules
  skillshare extras agents --flatten
  

[0.20.3] - 2026-06-01​

New Features​

• Copilot CLI agents — agents now sync to Copilot CLI alongside skills. Copilot uses the same .agent.md format Skillshare already manages, so skillshare sync agents symlinks your agents into ~/.copilot/agents (global) and .github/agents (project) with no conversion:

  skillshare sync agents    # now includes copilot
  

Bug Fixes​

• Installing a specific skill no longer drags in every agent — when installing from a repo that contains both skills and agents, -s/--skill now installs only the named skills (no agents), and -s with -a/--agent installs only the named agents. An unknown -a name fails the whole command up front, so automation never sees a half-completed install. --all/--yes still install everything.

[0.20.2] - 2026-05-31​

New Features​

Doctor remediation suggestions​

• skillshare doctor now suggests how to fix what it flags — when doctor detects targets writing to a shared path, or one target's runtime discovering another target's skills, it prints a remediation suggestion next to the warning instead of only reporting the overlap. Suggestions are also included in doctor --json under a new suggestions field.
• Suggestions point at a ready-to-run target removal command — overlap suggestions now include the exact command to preview removing a duplicate target:

  skillshare target remove <name> --global --dry-run
  

• Health Check page surfaces suggestions — the dashboard Health Check page renders each check's remediation suggestions inside its expandable detail block, localized across all supported languages.

Bug Fixes​

• GitHub Enterprise Cloud data residency hosts are now recognized — repositories on *.ghe.com tenants are detected as GitHub over both HTTPS and SSH, and the GitHub API base is resolved as https://api.<tenant>.ghe.com, so installing and updating from data residency accounts works.
• SSH remote URLs with a custom username now work — clone URLs such as acme@acme.ghe.com:org/repo.git (any username, not only git) are parsed and normalized correctly when installing and updating.

[0.20.1] - 2026-05-31​

Bug Fixes​

• Codex agent transforms reject incomplete agents — the bundled codex-agents extension now fails clearly when the resolved name, description, or Markdown body is blank. Missing name still falls back to the source filename, and the extension docs link to Codex's custom agent schema for the required fields.
• Dashboard Extras sync shows extension errors — when an extras transform fails, the Extras page toast now shows the first file-level error instead of only an error count, so users can fix the specific source file without opening logs.

[0.20.0] - 2026-05-30​

New Features​

Git scope control (git_root)​

• git_root scope — choose which directory skillshare commit, push, and pull version. The default stays your skills source, but you can point git at agents, extras, or root (skills + agents + extras together in a single repo). Set it during init, or switch later on an existing setup:

  skillshare init --git-root root      # version skills, agents, and extras in one repo
  skillshare init --git-root agents    # switch scope headlessly later
  

  A root-scope repo automatically keeps config.yaml out of version control (it holds machine-specific paths), and nested git repositories are detected and blocked before they would upload as empty submodules. If git_root points to a scope whose directory has no repo, commit/push/pull print a "Git root mismatch" error with the exact commands to fix it.
• Switch scope from the dashboard — the Git Sync page can change the git_root scope, set the git remote during the switch, and offers a one-click action when the scoped directory isn't a repository yet.

Extras extension transforms​

• extension field on extras targets — convert Markdown into a tool's native format during sync, for tools that don't read Markdown. Reference extensions ship for Gemini CLI (TOML commands) and Codex CLI (TOML agents):

  extras:
    - name: commands
      targets:
        - path: .gemini/commands
          extension: gemini-commands    # transforms .md → .toml during sync
  

  Transforms run source → target only (extras collect skips them), use copy semantics, and never overwrite a local file or directory without --force. The Codex agents extension maps name, description, and model from frontmatter.
• Manage extensions from the dashboard — the Config page lists installed extensions with descriptions and guards against removing one that is still in use; the Extras page and Add Extra modal include a per-target extension picker.

List filtering​

• Filter skills by enabled/disabled status — press s in the list TUI to cycle All → Enabled → Disabled, or use the s:enabled / s:disabled tag to combine status with other filters. The dashboard Resources page gains the same status filter.

Bug Fixes​

• Hardened git remote handling — remote URLs beginning with - (which git could misinterpret as a flag) are now rejected when setting or adding a remote, including via skillshare init --remote.
• Dashboard pull keeps skill paths correct with scoped git roots — pulling from a root, agents, or extras git scope no longer causes the follow-up sync to flatten paths such as skills/foo into the wrong skill name.
• Transformed extras no longer show false drift — dashboard diff/status checks now compare transformed filenames (for example .md → .toml) consistently, and missing extension definitions surface as warnings instead of silently falling back.
• Safer reference transforms — bundled TOML transforms now escape control characters correctly and stop hung transform commands instead of blocking sync indefinitely.

[0.19.24] - 2026-05-27​

New Features​

Local skill checkpoints​

• skillshare commit — create a local git commit for source skills without pushing to the remote. This is useful when iterating locally and wanting a restore point before you are ready to share changes across machines. It stages all source changes and commits them with the provided message, and it works even when no git remote is configured.

  skillshare commit -m "Update writing skill"
  skillshare commit --dry-run
  

• Dashboard local commit action — the Git Sync page now has a Commit locally button alongside Push and Pull. It uses the same commit message and dry-run preview area, but only creates the local commit and never pushes.

Bug Fixes​

• Dashboard update results stay current after updating — after updating skills from the dashboard, successfully updated or already-current items are now marked Up to date and keep their latest check status instead of falling back to Unchecked.

[0.19.23] - 2026-05-26​

Bug Fixes​

• Nested GitHub-installed skills stop reappearing as updateable after update — skillshare update and the dashboard Update page now refresh the stored metadata for skills installed under a subdirectory, so items such as tools/agent-browser no longer keep showing Update available immediately after a successful update
• Dashboard Update checks are remembered between visits — the Update page now keeps the last completed check status in browser storage and shows the previous check time, so returning to the page no longer resets every item to Unchecked

[0.19.22] - 2026-05-26​

Bug Fixes​

• Web UI install now handles mixed-track repos — installing a repository that contains both skills and agents with Track enabled (for example github/awesome-copilot) used to fail with a tracked install is ambiguous; pass --kind skill or --kind agent error toast and no way forward. The dashboard now opens a kind picker showing skill and agent counts; choose Skills or Agents and the install proceeds with the chosen --kind. Refs: #167

Performance​

• One fewer clone when recovering from a mixed-track install — the install API now reports skill/agent counts in the ambiguity error itself, so the dashboard no longer re-clones the repository via /api/discover before showing the kind picker

[0.19.21] - 2026-05-24​

Bug Fixes​

• Update page checks now finish for nested GitHub-installed skills — the dashboard now matches check results by relative path as well as display name, so items such as tools/agent-browser no longer stay stuck on Checking after Check All completes
• Update Selected now targets nested GitHub-installed skills correctly — selecting an item installed under a subdirectory now sends its relative path (for example tools/agent-browser) to the update endpoint instead of the flattened display name
• Remote update checks no longer wait indefinitely on Git prompts — lightweight git ls-remote probes now disable interactive credential prompts and time out after 15 seconds, surfacing an error in the UI/CLI instead of leaving the update check spinner running forever

[0.19.20] - 2026-05-24​

New Features​

Web UI process management​

• skillshare ui start / skillshare ui stop — run the web dashboard as a managed background process instead of holding a foreground shell. Re-running start reuses the existing healthy process; stop shuts it down using the remembered host/port. The legacy foreground skillshare ui and the --no-open & shell-backgrounding workaround keep working unchanged.

  skillshare ui start                 # start in background, return to shell
  skillshare ui start --clear-cache   # clear cached UI assets, then start
  skillshare ui stop                  # stop the background server
  

• --app flag — skillshare ui start --app opens the dashboard in a Chromium app-mode window (Chrome, Edge, Brave) so it gets its own Dock/taskbar entry and chrome-less frame on macOS, Windows, and Linux.

In-place upgrade from the dashboard​

• Update without leaving the browser — the Update dialog and the Doctor page's Version card now show an Update now button. The dashboard runs skillshare upgrade on the host, restarts the local UI server, then auto-reloads the page once the new server reports healthy. Two new endpoints back this flow:

  POST /api/upgrade   # run skillshare upgrade in place
  POST /api/restart   # restart the UI server (optional { "clearCache": true })
  

  If the dashboard cannot reconnect on its own, it surfaces a message asking you to run skillshare ui start to bring the background server back up.
• Doctor Version card redesign — the version section now leads with a state-coloured icon (blue when an update is available, green when up to date), shows the version delta as current → latest, and hides the Update now button when there is nothing to upgrade.

Installable as a Progressive Web App​

• PWA support — the dashboard now ships a site.webmanifest, app icons (192px / 512px), and a minimal service worker. Browsers that support it (Chrome, Edge, Safari, Brave) let you install Skillshare as a standalone desktop app from the address bar; offline cache covers the static shell so the dashboard opens even before the local server is reachable.

[0.19.19] - 2026-05-23​

Bug Fixes​

• Tracked root-level SKILL.md repos now get integrity hashes — skillshare install <repo> --track for single-skill repos with SKILL.md at the repo root now writes file_hashes into .metadata.json, so skillshare doctor can verify them instead of warning that integrity checks are unavailable. Running skillshare update _repo also backfills hashes for already-installed tracked root-skill repos that were created by v0.19.18. Refs: #165

[0.19.18] - 2026-05-22​

Bug Fixes​

• Tracked repos with a root-level SKILL.md are fully supported — installing a single-skill repo with --track (where SKILL.md sits at the repo root, not nested) used to report Found 0 skill(s), skip metadata persistence, and omit the repo from skillshare status and skillshare list. Now the install reports the correct count, writes the tracked entry to .metadata.json, and the repo appears in status output with skill_count: 1. Cross-machine recovery via metadata works again for this layout. The misleading Run skillshare sync to distribute skills next-step hint is also suppressed when the install produced zero skills or agents. Refs: #163
• GitLab SSH URLs with nested subgroups now produce a clean tracked-repo name — skillshare install git@gitlab.example.com:org/subgroup/my-skills.git --track previously stored the repo under a directory name containing the subgroup path; the resolved name is now just my-skills, matching the existing HTTPS subgroup behavior

[0.19.17] - 2026-05-22​

New Features​

• doctor warns on overlapping skill paths — skillshare doctor now flags two classes of duplicate-skill problems before they reach the runtime picker. Refs: #135

  ! Shared path ~/.agents/skills ← universal, warp
  ! codex will see content from: universal
      ~/.agents/skills ← universal
  

  The first warning fires when two enabled targets resolve to the same primary path. The second fires when an enabled target's runtime also scans a directory another enabled target writes to (e.g. Codex Desktop reads ~/.agents/skills in addition to ~/.codex/skills). Both checks are pure metadata — no filesystem probing, no runtime calls. skillshare sync also prints a one-line hint when overlap is detected, pointing back to doctor for the full breakdown:

  ! Skill path overlap across 2 target(s) — run `skillshare doctor` for details
  

Bug Fixes​

• cline target path reverted to ~/.cline/skills — the cline target was briefly pointing at the shared ~/.agents/skills root, which doesn't match the official cline documentation. Restored to the brand-specific ~/.cline/skills (global) and .cline/skills (project). If you synced cline between v0.19.14 and v0.19.16, re-run skillshare sync to move skills to the correct location

[0.19.16] - 2026-05-21​

New Features​

• Global config sources: map — global mode now accepts the same sources: map shipped for project mode in v0.19.15. Override skills, agents, or extras source directories from a single place:

  # ~/.config/skillshare/config.yaml
  sources:
    skills: ~/work/skills
    agents: ~/work/agents
    extras: ~/work/extras
  

  Each key is optional. Existing top-level source: / agents_source: / extras_source: fields continue to work unchanged; when both formats are set, sources.<key> wins. Fresh skillshare init -g writes the new sources: shape; existing configs are never auto-rewritten

Behavior Changes​

• extras init no longer silently backfills extras_source: — running skillshare extras init <name> on a global config without an explicit extras source used to write the derived extras_source: path back into your config.yaml. The runtime now resolves the extras parent on the fly, so the legacy field stays empty unless you set it yourself. Same change applies to the equivalent server endpoint (POST /api/extras)
• extras source <path> writes to whichever field the user already configured — if sources.extras is already set, the command updates that key; otherwise it falls back to the legacy extras_source field. This prevents the new value from being silently shadowed by sources.extras

[0.19.15] - 2026-05-20​

New Features​

• Custom project source directories — project mode can now read skills, agents, and extras from any directory in your repo. Useful for co-locating skill content with existing project documentation. Refs: #153, #162

  # .skillshare/config.yaml
  sources:
    skills: ./docs/skills
    agents: ./docs/agents
    extras: ./docs/extras
  targets:
    - claude
  

  Each key is optional — omit to fall back to .skillshare/<type>/. Paths are resolved from the project root; absolute paths and ~ work too. skillshare init -p still seeds the default .skillshare/ directories. Trash, backups, and operation logs always stay under .skillshare/ regardless of sources settings

Behavior Changes​

• Project commands fail closed on malformed config.yaml — uninstall, new, enable/disable, and check now return failed to load project config instead of silently falling back to .skillshare/skills. With custom sources, the old fallback could have operated on the wrong directory. Fix any YAML errors (e.g. targets: {} → targets: []) and the command will succeed
• Sync rejects source/target path overlap — skillshare sync -p errors when sources.skills or sources.agents aliases or nests with a target path. Without this check, sync --force could delete the configured source directory. Common safe layout: sources.skills: ./docs/skills with a claude target (no overlap)

[0.19.14] - 2026-05-20​

Refactoring​

• Unified Gemini and Antigravity targets — the standalone gemini target has been merged into antigravity. The shared skill path ~/.gemini/skills is now served by the antigravity target. Old names gemini, gemini-cli, and antigravity-cli continue to work as aliases

Bug Fixes​

• Audit: Swift files are now scannable — .swift files are included in security audit scans alongside .go, .py, .ts, and other source files
• Audit: reduced false positives — tightened publisher-claim extraction to avoid flagging ordinary phrases like "from CSV data" as organization claims. Bracket placeholders like [Count] ([Percentage]%) no longer trigger dangling-link warnings. Common documentation domains (MDN, Apple Developer, etc.) are excluded from external-link findings
• Audit: uppercase URI schemes no longer flag as dangling links — custom URI schemes like VSCode:// are now correctly recognized as external links

[0.19.13] - 2026-05-17​

Bug Fixes​

• Tracked skills now retain custom name and branch from config — previously, skillshare update on a tracked repo re-derived the directory name from the clone URL (e.g. _owner-repo), ignoring any custom name: set in config.yaml. Now the resolution priority is: --name flag > config name: > URL-derived name. Similarly, a non-default branch: in config is now respected during clone and update. Refs: #158

[0.19.12] - 2026-05-14​

Bug Fixes​

• Project skills: in config.yaml no longer stripped — previously, project-mode commands silently removed the skills: section from .skillshare/config.yaml during an internal migration, leaving no committable record of remote skill dependencies. Now skills: stays in config.yaml as the declarative source of truth. Teammates can clone the repo and run skillshare install -p to install all listed skills. Refs: #157

  # .skillshare/config.yaml — committed to git
  targets:
    - claude
    - cursor
  skills:
    - name: pdf
      source: anthropic/skills/pdf
    - name: review
      source: github.com/team/skills/code-review
      group: frontend
  

  skillshare install <source> -p automatically adds the skill to config.yaml. skillshare uninstall removes it. Runtime metadata (hashes, timestamps) stays in .metadata.json (gitignored)

[0.19.11] - 2026-05-14​

New Features​

• preserve_tilde_on_save — opt-in config flag that folds $HOME prefixes back to ~ when saving config.yaml. Keeps the on-disk config portable across machines when shared via dotfiles (stow, chezmoi, yadm, bare git repo). Refs: #155

  preserve_tilde_on_save: true
  

  Non-home absolute paths (e.g. /opt/shared/skills) are passed through unchanged. The in-memory config is unaffected — Load() still expands ~ as usual

Bug Fixes​

• Fixed skillshare init -p not gitignoring .skillshare/backups/ — backup artifacts from project-mode agent sync could be accidentally committed

[0.19.10] - 2026-05-12​

Bug Fixes​

• Fixed skillshare update --all creating duplicate .metadata.json entries with ../../... relative-path keys when the skills source directory is a symlink or custom location. Existing metadata keys such as browser/agent-browser now stay stable during update. Refs: #152

[0.19.9] - 2026-05-11​

New Features​

• Context cost summary after sync — skillshare sync now displays a one-line token cost summary showing always-loaded and on-demand context usage per target. Targets with identical token counts are grouped on a single line. Refs: #150

  ✔ Synced 47 skill(s) to 4 target(s) in 312ms
    Context: ~12.4K always-loaded · ~58.2K on-demand (claude, cursor, codex, opencode)
  

• Configurable budget warnings — set token budget thresholds in config.yaml. When sync or analyze detects a target exceeding the budget, a warning shows the top 3 offenders by token count

  context_budget:
    warn_always_loaded_tokens: 10000   # default; 0 = disabled
    warn_on_demand_tokens: 100000      # default; 0 = disabled
  

  Defaults to 10K always-loaded / 100K on-demand. Set to 0 to disable
• --quiet / -q flag for sync — suppresses the token summary and budget warnings. JSON output (--json) always includes context_cost regardless of --quiet
• Context cost in Web UI — the Sync page now displays token cost groups and budget violation warnings after each sync

Bug Fixes​

• Fixed skillshare install failing with http:// protocol URLs (contributed by @eekryuos)

[0.19.8] - 2026-05-08​

New Features​

• Shell completion — new skillshare completion command generates tab-completion scripts for bash, zsh, fish, PowerShell, and Nushell. Supports --install to auto-write the script to the correct platform path. Refs: #148

  skillshare completion bash --install    # one-line setup
  skillshare completion zsh --install
  skillshare completion fish --install
  skillshare completion powershell --install
  skillshare completion nushell --install
  

  Covers all commands, subcommands, and per-command flags. Bash, zsh, and PowerShell scripts auto-detect aliases (e.g. alias ss=skillshare) and register completions for them

[0.19.7] - 2026-05-05​

New Features​

• Azure DevOps Server (on-premises) support — added azure_hosts config parameter for self-hosted Azure DevOps Server instances on custom domains. Previously, only dev.azure.com and *.visualstudio.com were recognized; custom-domain URLs would fail with an incorrect .git suffix in the clone URL. Refs: #147

  azure_hosts:
    - azuredevops.mycompany.com
  

  skillshare install https://azuredevops.mycompany.com/Org/Project/_git/Repo
  

  Works in both global and project configs. For CI/CD, use the SKILLSHARE_AZURE_HOSTS environment variable (comma-separated, merged with config file values)

[0.19.6] - 2026-05-05​

New Features​

• 8 new agent targets — added AiderDesk, CodeArts Agent (Huawei), Code Studio (Syncfusion), Devin for Terminal (Cognition), Dexto, ForgeCode, Rovo Dev (Atlassian), and Tabnine CLI. Total built-in targets: 64+
• Cursor project path updated to .agents/skills — Cursor officially supports .agents/skills as a project-level skill path alongside .cursor/skills. The project path now uses the ecosystem-standard .agents/skills convention, matching Cursor's official documentation. The global path (~/.cursor/skills) and agents paths are unchanged
• Automatic target inference from directory structure — skills organized under a target's project path are now automatically scoped to that target during sync, without needing targets: in SKILL.md frontmatter. For example, placing a skill at .cursor/skills/my-skill/ in the source directory automatically restricts it to the cursor target

  source/
    .cursor/skills/browse/     → syncs only to cursor
    .factory/skills/bench/     → syncs only to droid
    openclaw/skills/investigate/ → syncs only to openclaw
    shared/my-tool/            → syncs to all targets (no inference)
  

  Inference matches against the project paths defined in targets.yaml. When multiple targets share the same path (e.g. universal, codex, amp all use .agents/skills), the skill is correctly scoped to all of them

Bug Fixes​

• Fixed host path inference only returning the canonical target name when multiple targets share the same project path — .agents/skills/ skills now correctly infer all sharing targets instead of just universal
• Fixed skillshare update failing for skills originally installed at the repo root of an orchestrator (multi-skill) repository — the update now correctly detects the repo layout and re-applies subdirectory extraction

[0.19.5] - 2026-04-23​

New Features​

Search Experience​

• Skill preview modal — clicking a search result card or table row now opens a preview modal that renders the remote SKILL.md so you can read the full content before installing. The Install button has moved inside the modal

  skillshare ui              # open Search, click any result
  

  Backed by a new GET /api/preview?source=owner/repo/path endpoint that fetches SKILL.md from the GitHub Contents API with a Git Tree fallback for hub sources, plus a 5-minute in-memory cache. If the fetch fails but hub metadata exists, the modal falls back to showing the description and tags instead of an error. Refs: #141

• Higher-quality search results — the scoring formula that ranked search hits was dominated by name matches (45%) while stars — the strongest quality signal — only contributed 5%, so zero-star personal experiments often outranked well-known skill repos. Weights are now rebalanced (name 30%, description 20%, stars 25%, source 25%), with a better star-count curve in the 5–50 range where most useful skills live

• Low-quality result filtering — search now filters out 0-star non-preferred repos, results with very short descriptions, and known spam orgs, and boosts preferred hubs like anthropics/skills and vercel-labs/skills. Results are also deduplicated when the same skill appears from multiple sources

• Repo-scoped search queries — you can now search inside a specific repo (or subdirectory) using owner/repo or owner/repo/subdir patterns

• Frontmatter validation — search results without valid SKILL.md frontmatter are filtered out, so non-skill .md files no longer appear in results

Skill Detail Editor​

• Editable source URL on skill detail page — the skill editor now exposes a Source URL field so you can correct or reassign the upstream URL without touching disk

  skillshare ui              # open any skill's detail page → edit Source URL
  

  For tracked repos, updating one skill's source URL updates every sibling skill from the same repo and rewrites the git remote origin to match. Saving source-URL-only changes skips the diff review dialog
• Backed by a new PATCH /api/resources/{name}/source endpoint

Bug Fixes​

• Fixed a skill lookup collision when a standalone skill and a tracked-repo skill shared the same base name — detail-page requests now prefer an exact FlatName match before falling back to the base name, so the correct skill is always returned
• Fixed spurious frontmatter diffs when only the body or source URL changed — the YAML serializer now preserves the original formatting (quote style, line wrapping) unless a field's value actually changed
• Fixed IntersectionObserver on the Search page not reconnecting after "load more", so infinite scrolling now continues to fire as you reach the bottom of the list
• Fixed background page scrolling when a modal was open — DialogShell now locks body scroll for all modal dialogs

Improvements​

• Keyboard-accessible search results — search result cards and table rows now have role="button" with tabIndex=0 and respond to Enter/Space, so the preview modal can be opened without a mouse
• Bounded preview cache — the preview cache is now capped at 200 entries with TTL purge plus a full reset on overflow to prevent unbounded memory growth on long-running UI sessions
• Shared input components in the frontmatter editor — the editor now uses the same Input / Textarea primitives as the rest of the UI (with a new size="sm" variant) for visual consistency, and the editor header now uses the shared KindBadge / SourceBadge components

[0.19.4] - 2026-04-22​

Bug Fixes​

• UI install no longer deletes local source files — installing a skill from a local path via the Web UI would permanently delete the original source directory after discovery. The cleanup logic now only removes temporary directories created by git clones, leaving user directories untouched. CLI installs were unaffected. Refs: #139

Improvements​

• $ARGUMENTS tokens highlighted on detail page — skill detail pages in the Web UI now render $ARGUMENTS placeholders with the same warning-badge style used in the editor, making them easier to spot when reviewing a skill's content

[0.19.3] - 2026-04-18​

New Features​

Web-Based Skill Editor​

• Edit skills directly in the Web UI — the resource detail page now has an Edit action that opens a full in-browser editor, replacing the previous read-only view. Changes are saved back to the source SKILL.md through a new POST /api/resources/:name/content endpoint

  skillshare ui              # then open any skill and click Edit
  

• Two-pane Markdown editor — side-by-side textarea and live preview with synced scrolling. A mode toggle in the status bar switches between Edit / Split / Preview (⌘P cycles), and ⌘S saves while Esc cancels. An outline drawer lets you jump to any heading, and the status bar shows token / word / line / file counts with a 5K-character budget warning

• Structured frontmatter editor — all 13 official SKILL.md fields are grouped into Identity / Invocation / Execution / Metadata sections with field-appropriate controls: switch toggles for booleans, segmented control for enums, and chip inputs for list-valued fields. A shared 1,536-character budget is enforced across description + when_to_use. A raw YAML mode is also available, and round-trips cleanly with the Fields view. Legacy root-level targets: is automatically migrated to metadata.targets on load

• Diff review before save — saving opens a side-by-side diff modal so you can confirm every change before writing to disk. The YAML serializer emits plain scalars when safe, so the diff no longer shows spurious quote wrapping for values containing :, *, #, or "

• Open in local editor — a new POST /api/resources/:name/open endpoint opens the skill file in your preferred local editor (via $EDITOR), useful when you prefer vim / VS Code over the browser editor

• Targets visible in detail sidebar — the resource detail sidebar now shows a Targets row when metadata.targets is set, so you can see at a glance which targets a skill is scoped to

Web UI Localization​

• 11 languages in the Web UI — every page is now fully translated. Pick your language from the language switcher in the top navigation bar; the preference is saved to your browser and auto-detected on first visit from navigator.languages
  • Supported: English, 中文, 日本語, 한국어, Español, Français, Deutsch, فارسی, Português (BR), Bahasa Indonesia
  • Persian (فارسی) automatically switches the layout to right-to-left

[0.19.2] - 2026-04-14​

New Features​

• Local agent count in Targets — the Targets page and skillshare target --json now show how many local (non-source) agents exist per target alongside the linked/expected counts. The Collect button appears for targets with local agents and routes directly to the Collect page with the correct scope pre-selected

  skillshare target --json   # agentLocalCount field in each target
  

• Richer target list --json output — each target entry now includes targetNaming (effective naming strategy, e.g. flat), sync (skill sync summary), and agentSync (agent sync summary), making JSON output more useful for scripted workflows and external tooling

  skillshare target list --json   # targetNaming, sync, agentSync per target
  

Bug Fixes​

• Fixed batch uninstall for agents — selecting multiple agents from the same tracked repo on the Uninstall page sent only the repo name instead of individual agent names, causing all uninstalls to fail with "agent not found". The confirmation dialog now correctly lists each agent and the API receives individual names
• Fixed empty state messages on the Resources and Uninstall pages — switching to the Agents tab with no agents installed now shows "No agents installed" instead of the generic "No skills installed" text
• Fixed Health Check theme display — the Theme check on the Web UI always showed a warning with technical internals (fallback-dark-no-tty, no_color: false) that users couldn't understand. Now the CLI and UI both show human-readable messages like "auto-detected from terminal" or "set via SKILLSHARE_THEME", and the UI no longer shows a false warning caused by the server subprocess lacking a terminal

[0.19.1] - 2026-04-13​

Bug Fixes​

• Orchestrator repos no longer copy the entire repository — installing a multi-skill repo with a root SKILL.md (e.g. skillshare install user/project) previously copied the entire repo root into the root skill directory, including source code, assets, CI configs, and build scripts. Now the root skill installs only its SKILL.md, and each child skill installs as an independent flat directory. Refs: #124

  # Before: skills/MyProject/ contained the entire repo
  # After:
  skills/MyProject/SKILL.md       ← root skill (SKILL.md only)
  skills/child-a/...              ← independent child
  skills/child-b/...              ← independent child
  

• Structured output no longer corrupted by update notices — --json, -j, and --format json/sarif/markdown modes could emit a trailing human-readable update notification into stdout, producing invalid JSON for downstream consumers. The update check is now skipped entirely in structured-output modes, and the notification itself writes to stderr as a safety net. Refs: #129

[0.19.0] - 2026-04-11​

New Features​

Agent Management​

Agents are now a first-class resource type alongside skills. You can install, sync, audit, and manage agent files (.md) across agent-capable targets (Claude, Cursor, OpenCode, Augment) with the same workflow as skills.

• Agents source directory — agents live in ~/.config/skillshare/agents/ (or .skillshare/agents/ in project mode). skillshare init creates the directory automatically, and agents_source is a new config field that can be customized

  skillshare init              # creates skills/ and agents/
  skillshare init -p           # same for project mode
  

• Positional kind filter — most commands accept an agents keyword to scope the operation to agents only. Without it, commands operate on skills (existing behavior is unchanged)

  skillshare sync agents             # sync agents only
  skillshare sync --all              # sync skills + agents + extras
  skillshare list agents             # list installed agents
  skillshare check agents            # detect drift on agent repos
  skillshare update agents --all     # update every agent
  skillshare audit agents            # scan agents for security issues
  skillshare uninstall agents foo    # uninstall a single agent
  skillshare enable foo --kind agent
  skillshare disable foo --kind agent
  

• Install agents from repos — install auto-detects agents in three layouts:

  • agents/ convention subdirectory
  • mixed-kind repos with both SKILL.md and agents/
  • pure-agent repos (root .md files, no SKILL.md)

  skillshare install github.com/team/agents            # auto-detect
  skillshare install github.com/team/repo --kind agent # force agent mode
  skillshare install github.com/team/repo --agent cr   # specific agents
  

  Conventional files (README.md, LICENSE.md, CHANGELOG.md) are automatically excluded

• Tracked agent repos — agents can be installed with --track for git-pull updates, including nested discovery. check, update, doctor, and uninstall all recognise tracked agent repos, and the --group / -G flag filters by repo group

• Agent sync modes — merge (default, per-file symlink), symlink (whole directory), and copy are all supported. skillshare sync agents skips targets that don't declare an agents: path and prints a warning

• .agentignore support — agents can be excluded via .agentignore and .agentignore.local using the same gitignore-style patterns as .skillignore. The Web UI Config page now has a dedicated .agentignore tab

• Agent audit — skillshare audit scans agent files individually against the full audit rule set, with Skills/Agents tab switching in both the TUI and Web UI. Audit results carry a kind field so tooling can filter by resource type

• Agent backup and restore — sync automatically backs up agents before applying changes, in both global and project mode. The backup TUI and trash TUI tag agents with an [A] badge and route restores to the correct source directory

• Project-mode agent support — every agent command works in project mode with -p. Agents are reconciled alongside skills into .skillshare/

• JSON output for agents — install --json and update --json now emit agent-aware payloads and apply the same audit block-threshold gate as skills. Useful for scripted agent workflows

  skillshare update agents --json --audit-threshold high
  

• Kind badges — TUI and Web UI surface [S] / [A] badges throughout (list, diff, audit, trash, backup, detail, update, targets pages) so you can tell at a glance what kind of resource you're looking at

Unified Web UI Resources​

• /resources route — the old /skills page is now /resources, with Skills and Agents tabs. Tab state persists to localStorage, and the underline tab style follows the active theme (playful mode gets wobble borders)

• Targets page redesign — equal Skills and Agents sections, with a modal picker for adding targets. Filter Studio links include a ?kind= param so you jump directly to the right context

• Update page redesign — a new three-phase flow (selecting → updating → done) with skills/agents tabs, group-based sorting, and status cards. EventSource streaming is properly cleaned up on page change

• Filter Studio agent support — agent filters can be edited via PATCH /api/targets/:name (agent_include, agent_exclude, agent_mode) and via the CLI through new flags on skillshare target <name>:

  skillshare target claude --add-agent-include "team-*"
  skillshare target claude --remove-agent-include "team-*"
  skillshare target claude --agent-mode copy       # merge | symlink | copy
  

  The UI Filter Studio is a single-context view driven by ?kind=skill|agent

• Audit cache — audit results are now cached with React Query and invalidated on mutation. The audit card icon colour follows the max severity, and the count no longer mixes agent totals with finding counts

• Collect page scope switcher — a new segmented control lets you collect skills or agents from targets

Theme System​

• internal/theme package — unified light/dark terminal palette with WCAG-AA-compliant light colours and softened dark primary. Resolution order: NO_COLOR > SKILLSHARE_THEME > OSC 11 terminal probe > dark fallback. All TUIs, list output, audit output, and plain CLI output now route through the theme

  SKILLSHARE_THEME=light skillshare list
  SKILLSHARE_THEME=dark skillshare audit
  

  skillshare doctor includes a theme check to help debug unreadable colours

Install & TUI Polish​

• Explicit SKILL.md URLs resolve to one skill — pasting a direct blob/.../SKILL.md URL now installs only that skill, bypassing the orchestrator pack prompt. Previously, the URL would trigger the full multi-select picker even though the intent was clear

  skillshare install https://github.com/team/repo/blob/main/frontend/tdd/SKILL.md
  

  Refs: #124

• Radio checklist follows the cursor — the single-select TUI (used for orchestrator selection, branch selection, and similar flows) now auto-selects the focused row. No more confusing empty-selection state — pressing Enter always confirms the item your cursor is on

• Diff TUI — single-line items with group headers instead of the old verbose-per-item layout. Agent diffs are shown with an [A] badge

• List TUI — entries are now grouped by tracked repo root and local top directory, with a new k:kind filter tag for quick agent/skill filtering inside the fuzzy filter

Centralized Metadata Store​

• .metadata.json replaces sidecar files and registry.yaml — installation metadata is now stored in a single atomic file per source (~/.config/skillshare/skills/.metadata.json). This fixes long-standing issues with grouped skill collisions (e.g. two skills both named dev in different folders) where the old basename-keyed registry would mix them up
  • Automatic migration — the first load after upgrade reads any existing registry.yaml and per-skill .skillshare-meta.json sidecars, merges them into .metadata.json, and cleans up the old files. Idempotent — safe to run repeatedly
  • Full-path keys — lookups use the full source-relative path, so nested skills never collide
  • No user action required; existing installs continue to work

Bug Fixes​

• Sync extras no longer flood when agents target overlaps — targets that declare an extras entry called agents are now skipped automatically when agent sync is active, preventing duplicate file writes
• Nested agent discovery — check agents now uses the recursive discovery engine, so agents in sub-folders (e.g. demo/code-reviewer.md) are detected correctly
• Doctor drift count excludes disabled agents — agents disabled via .agentignore no longer count toward the drift total reported by skillshare doctor
• Audit card mixes counts — the Web UI audit card no longer mixes agent counts with finding counts, and excludes _cross-skill from the card total (shown separately)
• Audit scans disabled agents too — the audit scan walks every agent file regardless of .agentignore state, so hidden agents still get checked
• List TUI tab bar clipping — the tab bar no longer gets cut off in the split-detail layout on narrow terminals
• Sync extras indent — removed the stray space between the checkmark and the path in sync extras output; summary headers are now consistent across skills, agents, and extras
• UI skill detail agent mode — the detail page hides the Files section for agents (single-file resources), remembers the selected tab via localStorage, and shows the correct folder-view labels
• Sync page layout — the stats row and ignored-skills grouping are now easier to scan
• Button warning variant — the shared Button component now supports a warning variant that was already referenced by several pages
• Check progress bar pop-in — removed the loading progress bar that caused layout shift on the Skills page
• Tracked repo check status — the propagated check status is now applied to every item within the repo, not just the root
• Target name colouring in doctor — only the status word is coloured in doctor target output, not the full line

Breaking Changes​

• audit --all flag removed — use the positional kind filter instead:

  skillshare audit                # skills (default, unchanged)
  skillshare audit agents         # agents only
  

  The old --all flag is gone because audit now runs per kind and the Web UI has dedicated tabs

[0.18.9] - 2026-04-07​

New Features​

• Relative symlinks in project mode — skillshare sync -p now creates relative symlinks (e.g., ../../.skillshare/skills/my-skill) instead of absolute paths. This makes the project directory portable — rename it, move it, or clone it on another machine and all skill symlinks continue to work. Global mode continues to use absolute paths. Existing absolute symlinks are automatically upgraded to relative on the next sync

Bug Fixes​

• Status version detection — skillshare status no longer reports ! Skill: not found or missing version when the version is stored under metadata.version in the SKILL.md frontmatter. Previously, the status command used its own local parser that only checked for a top-level version: key, while doctor (fixed in v0.18.7) and upgrade already used the correct shared parser

[0.18.8] - 2026-04-06​

Bug Fixes​

• Sync no longer deletes registry entries for installed skills — running skillshare sync (or project-mode sync -p) would silently remove registry.yaml entries for skills whose source files were not present on disk. This meant that installing a skill and then syncing could erase the installation record entirely. Sync now leaves the registry untouched — only install and uninstall manage registry entries

[0.18.7] - 2026-04-04​

New Features​

Folder-Level Target Display & Bulk Editing​

• Folder target aggregation — the Skills page grouped view now shows aggregated target info on each folder row. If all skills share the same target, it shows that target; mixed targets show the union with a warning badge; All is shown when no targets are set

  • Compact display: folders with 4+ targets show N targets with full list in tooltip

• Bulk set target — right-click any folder to set or remove the target for all skills in that subtree at once

  Right-click folder → Available in... → claude
  

  Writes metadata.targets to every SKILL.md in the folder. Selecting All removes the field. Disabled skills are skipped

• Single skill target editing — right-click any skill in grouped or grid view, or use the inline Available in dropdown in table view, to change which targets receive that skill

Right-Click Context Menu​

• Context menu on all views — right-click skills in grouped, grid, or table view for quick actions:

  • Available in... — submenu to set target (hover-expand with 180ms intent delay)
  • View Detail — navigate to skill detail page
  • Enable / Disable — toggle skill visibility
  • Uninstall — with confirmation dialog

• Folder context menu — right-click folders in grouped view for Folder available in... (batch target) — only target actions, no uninstall

• Submenu pattern — top-level items with sub-options expand on hover. Future actions (e.g. Move, Rename) can be added as flat items alongside

Table View Improvements​

• Inline target selector — the table now has an Available in column with an inline dropdown for one-click target switching — no context menu needed
• Actions column — ⋯ button opens a flat menu with View Detail, Enable/Disable, and Uninstall
• Simplified layout — reduced from 7 columns to 5 by merging Path and Source into the Name cell. Path shows as a subtitle when different from the name; source shows as a clickable globe icon linking to the repo
• Persistent page size — the selected page size (10/25/50) is remembered across sessions

UX Polish​

• Right-click tip banner — a one-time dismissible tip appears on first visit, explaining that right-click is available for actions. Styled with playful theme support (wobble borders, paper-warm background, slight tilt)
• Optimistic updates — all mutations (set target, enable/disable, uninstall) update the UI instantly with automatic rollback on error
• Active item highlight — when a context menu is open, the targeted skill or folder gets a hover-matching highlight

Bug Fixes​

• Config page dashed border clipping — in playful theme, the Structure panel's dashed borders were cut off at the right edge because the panel wrapper kept overflow-hidden after expanding. Now uses overflow-visible when expanded

• Tracked repos in project-mode dashboard — the Web UI dashboard now shows tracked repositories when running in project mode (skillshare ui -p), with Update and Uninstall actions

• Nested tracked repo update and uninstall — repos installed with a nested path (e.g. org/_team-skills) can now be updated and uninstalled from both the CLI and Web UI. Previously, the server failed to resolve nested repo paths for these operations

• Project-mode uninstall cleans correct .gitignore — uninstalling a tracked repo in project mode now removes entries from .skillshare/.gitignore instead of the global source .gitignore. Previously, stale ignore rules were left behind

• Registry prune no longer affects sibling repos — uninstalling a nested tracked repo (e.g. org/_team-skills) no longer accidentally removes registry entries belonging to a sibling with the same basename (e.g. dept/_team-skills)

• Nested trash lifecycle — trash, restore, cleanup, and listing now work correctly for nested tracked repo names. Parent directories are created on restore and cleaned up after expiry

• Dashboard tracked repo row polish — status indicators (clean / modified) moved next to the repo name as compact badges. Action buttons use a smaller xs size to reduce visual weight

• Bulk target folder matching — setting targets on a folder with a trailing slash (e.g. frontend/) no longer silently skips all skills. The server now normalizes folder paths before matching

• Doctor version detection — skillshare doctor no longer reports ! Skill: missing version when the version is stored under metadata.version in the SKILL.md frontmatter. Previously, the inline parser only checked for a top-level version: key

• Target display on Skills page — the Skills page now correctly shows saved targets for each skill. Previously, the API did not parse SKILL.md frontmatter, so targets always appeared as All even after being set

• Target editing respects tracked repos — setting targets via the context menu or batch folder action now skips tracked-repo skills. Previously, writing to SKILL.md inside a tracked repo would make the repo dirty and block future updates. Audit hash integrity is also preserved after target edits

• Uninstall Repo from context menu — right-clicking a tracked-repo skill now shows Uninstall Repo instead of the individual Uninstall action (which would always fail with an error)

• Enable/disable with glob patterns — enabling a skill that was disabled by a glob or directory pattern in .skillignore (or .skillignore.local) now correctly returns an error with guidance, instead of silently showing a success toast while the skill stays disabled

• Disabled tracked skills stay in Tracked view — disabling a tracked-repo skill via .skillignore no longer removes it from the Tracked tab. The discovery engine now correctly preserves the isInRepo flag for all disabled skills

[0.18.6] - 2026-04-01​

Bug Fixes​

• UI batch uninstall now removes nested skill registry entries — previously, uninstalling grouped skills (e.g. frontend/vue/vue-best-practices) from the Web UI left stale entries in registry.yaml because the flat name (__) didn't match the stored path name (/). Uninstall now tracks the exact resolved path for accurate cleanup

• Sync prunes stale registry entries — skillshare sync and the Web UI Sync page now automatically remove registry.yaml entries for skills that no longer exist in the source directory. Covers manual deletions, not just uninstall. Skills hidden by .skillignore are preserved

• Uninstall page search works as substring match — typing matt in the filter box now matches mattpocock/tdd (substring search). Previously, only glob patterns like *matt* worked. Glob syntax (*, ?) still works when present

• Uninstall page shows path format — the confirmation dialog and result summary now display frontend/vue/vue-best-practices instead of frontend__vue__vue-best-practices

• Updates page removes redundant status line — the "0 repo(s) and 20 skill(s) already up to date" line no longer appears when everything is already current (the empty state already says this)

[0.18.5] - 2026-04-01​

New Features​

• --help for all commands — every command now supports --help / -h to show usage info, flags, and examples. Previously, commands like push, pull, sync, status, collect, doctor, and ui would execute instead of showing help when --help was passed

  skillshare push --help     # shows usage instead of pushing
  skillshare sync -h         # shows flags and examples
  skillshare ui --help       # shows port/host options
  

[0.18.4] - 2026-03-31​

New Features​

Branch Support (--branch / -b)​

• Install from a specific branch — new --branch / -b flag lets you clone from any branch instead of the remote default:

  skillshare install github.com/team/skills --branch develop --all
  skillshare install github.com/team/skills --track --branch frontend
  

  • Works with both tracked repos (--track) and regular skill installs
  • Branch is persisted in metadata — update and check automatically use the correct branch
  • Same repo on different branches: use --name to avoid collisions:

    skillshare install github.com/team/skills --track --branch frontend --name team-frontend
    skillshare install github.com/team/skills --track --branch backend --name team-backend
    

  • Supported in project mode (-p) and config-driven rebuild (skillshare install with no args)
  • registry.yaml stores the branch for cross-device reproducibility

• Branch in Web UI — the Install form shows a Branch input field (inline with Source) when a git source is detected. Skills page shows a branch badge on cards, and the Skill Detail page includes branch in the metadata section

• Branch in CLI list — skillshare list detail panel shows the tracked branch when non-default

• Branch-aware check — skillshare check compares against the correct remote branch ref, not just HEAD. JSON output includes a branch field for tracked repos

Sync Accuracy​

• Accurate skill counts on Targets page — the expected skill count now reflects what sync actually resolves (after name collision and validation filtering), instead of the raw source count. Previously, targets using standard naming could show 39032/64075 shared when all resolved skills were actually in sync

• Skipped skill visibility — when skills are excluded by naming validation or collisions, the Targets page and Sync page now show a summary (e.g. "12345 skill(s) skipped, 456 name collision(s)") instead of silently dropping them. Suggests switching to flat naming to include all skills

Target Naming Mode​

• target_naming config option — choose how synced skill directories are named in targets. Set globally or per-target:

  target_naming: standard    # use SKILL.md name as directory name
  targets:
    claude:
      skills:
        target_naming: flat  # override: keep flattened prefix (default)
  

  • flat (default) — nested skills like frontend/dev become frontend__dev in targets
  • standard — uses the SKILL.md name field directly (e.g. dev), following the Agent Skills specification
  • Standard mode validates that SKILL.md names match their directory name, warns and skips invalid skills
  • Name collisions (e.g. two skills both named dev) are detected and both are skipped with a warning
  • Switching from flat to standard automatically migrates existing managed entries (symlinks and copies are renamed in place)
  • If a local skill already occupies the bare name, the legacy flat entry is preserved with a warning
  • target_naming is ignored in symlink mode (the entire source directory is linked)

• Collision output redesigned — name conflict warnings are now deduplicated across targets and displayed as a compact summary instead of repeating each collision per target

Folder Tree View (Web UI)​

• Skills folder tree — the second layout on /skills is now a true folder tree with multi-level expand/collapse, matching your actual directory structure from --into:

  • Click any folder to expand/collapse its children
  • Expand All / Collapse All buttons in the toolbar
  • Sticky folder header — scrolling through a long folder keeps the folder name pinned at the top; click it to jump back
  • Search-aware — filtering or searching auto-expands all matching folders; clearing restores your previous collapse state
  • Hover tooltip on skill rows shows path, source, and install date (follows cursor, 1.5s delay)
  • Collapse state persists across page reloads via localStorage
  • Virtualized rendering handles 10,000+ skills with no performance impact

• Skill detail button layout — the Enable/Update/Uninstall buttons no longer wrap awkwardly on narrow screens

Agent Target Paths​

• Agent-specific paths — targets that support agents (Claude, Cursor, OpenCode, Augment) now declare separate agents: paths in their configuration. This is groundwork for upcoming agent sync support

GitHub Actions​

• setup-skillshare action — install skillshare in CI with a single step:

  - uses: runkids/setup-skillshare@v1
  

Bug Fixes​

• CLI sync output no longer floods terminal — targets with thousands of naming validation warnings (common with standard naming and large skill sets) now print a compact summary instead of one line per skipped skill
• Target dropdown lag removed — changing sync mode or target naming in the Web UI Targets page now updates instantly via optimistic cache update, instead of waiting 2 seconds for the API round-trip

Performance​

• Cached branch lookups — skillshare list caches git branch queries per tracked repo, so listing 500 skills from the same repo runs 1 git command instead of 500

[0.18.3] - 2026-03-29​

New Features​

Enable / Disable Skills​

• skillshare enable / skillshare disable — temporarily hide skills from sync without uninstalling them. Adds or removes patterns in .skillignore:

  skillshare disable draft-*        # hide from sync
  skillshare enable draft-*         # restore
  skillshare disable my-skill -p    # project mode
  skillshare disable my-skill -n    # dry-run preview
  

• TUI toggle — press E in skillshare list to toggle the selected skill's enabled/disabled state. The change is written immediately without leaving the TUI. Disabled skills show a red disabled badge in the detail panel and a [disabled] suffix in compact view

• Web UI toggle — the Skill Detail page now shows an Enable/Disable button. The REST API exposes POST /api/skills/{name}/enable and POST /api/skills/{name}/disable endpoints

Target Config — Skills Sub-Key​

• Per-resource-type configuration — target configs now support a skills sub-key with its own path, mode, include, and exclude settings. Existing flat-field configs are auto-migrated on first load — no manual editing needed:

  targets:
    claude:
      skills:
        path: ~/.claude/skills
        mode: merge
  

Upgrade Improvements​

• Auto-sudo for protected paths — skillshare upgrade now auto-detects when the binary is in a write-protected directory (e.g., /usr/local/bin) and transparently re-executes with sudo (#105)

Bug Fixes​

• Fixed diff showing skills with unsupported targets values (e.g., targets: ["*"]) as "source only" pending items — these skills are now correctly filtered out, matching the behavior of sync
• Fixed collect ignoring inherited sync mode when a target's mode was set at the global level — the command now resolves the effective mode before deciding whether to scan for local skills
• Fixed collect using a stale manifest after switching a target from copy to merge mode
• Fixed pull info message referencing git stash instead of git stash -u
• Fixed config migration writing directly to the config file — now uses atomic write-to-temp + rename to prevent corruption on disk errors

[0.18.2] - 2026-03-28​

New Features​

Update Page Improvements​

• Sticky progress bar — the Update page now shows a real-time progress bar during batch updates with percentage, completed/total count, and ETA. The bar stays pinned to the top while scrolling through the item list

• Auto-scroll to active item — during batch updates, the page automatically scrolls to the item currently being updated so you can follow the progress without manual scrolling

• Purge stale skills — when a skill fails to update because its subdirectory no longer exists in the repository, a Purge button appears instead of Force Retry. Clicking it removes the stale skill from your source directory

Analyze & Install UX​

• Install picker improvements — the skill picker modal in Install and Analyze pages now auto-focuses the search field, shows clearer skill descriptions, and handles keyboard navigation better

• Tooltip enhancements — tooltips across the dashboard now follow the cursor and stay within viewport bounds

Bug Fixes​

• Fixed analyze page crash when a target has no skills (null targets array from API)
• Fixed redundant path line showing in the skill detail dialog on the Analyze page
• Fixed install source field not clearing after a successful installation

[0.18.1] - 2026-03-27​

New Features​

Analyze — Filter & Token Budget​

• --filter flag — filter skills by name or group path with case-insensitive substring matching. Works across all output modes:

  skillshare analyze claude --json --filter frontend   # JSON with filtered_summary
  skillshare analyze --filter marketing                 # TUI with pre-populated filter
  

• Dynamic token subtotals — the TUI stats line and Web UI now show always-loaded, on-demand, and total token sums for the current filtered set. In the TUI, the format changes to 5/50 skills when a filter is active

• Web UI filtered summary bar — when searching or filtering on the Analyze page, a summary bar appears above the table with token counts per category. The bar slides in/out with a smooth animation

• JSON filtered_summary — when --filter is used with --json, the output includes filter, matched_count, total_count, and a filtered_summary object with always_loaded, on_demand, and total token counts

Registry Location​

• Registry moved to source directory — registry.yaml now lives at ~/.config/skillshare/skills/registry.yaml (inside the source directory) instead of ~/.config/skillshare/registry.yaml. This means git sync automatically includes the registry, so tracked skill metadata is preserved across machines. Migration is automatic on first run (#103)

Bug Fixes​

• init --remote skips skill prompt — when init --remote clones a repo that already contains skills, the interactive "choose skills to install" prompt is now skipped since the remote already defines the skill set (#102)

Improvements​

• Analyze Web UI polish — improved empty state with icon and helper text, fixed table height to prevent layout jumps when filtering, lint filter badge now shows readable rule names (e.g., "No Trigger Phrase" instead of no-trigger-phrase)

[0.18.0] - 2026-03-26​

New Features​

Analyze Command — Context Window & Skill Quality​

• skillshare analyze — new command that calculates context window token usage for each target's skills. Shows two layers of cost: "always loaded" (name + description, loaded every request) and "on-demand" (skill body, loaded when triggered). Token estimates use chars / 4:

  skillshare analyze               # interactive TUI
  skillshare analyze claude        # single target (auto-verbose)
  skillshare analyze --verbose     # top 10 largest descriptions
  skillshare analyze --json        # machine-readable output
  skillshare analyze -p            # project mode
  

• Skill quality lint — analyze runs 7 built-in lint rules against every skill, checking SKILL.md structure and description quality:

  • Errors: missing name, missing description, empty body
  • Warnings: description too short (<50 chars), too long (>1024 chars), near limit (900–1024), missing trigger phrases (e.g., "Use when…")

  Lint issues appear in the TUI (✗ for errors, ⚠ for warnings) and in --json output as lint_issues per skill

• Interactive TUI — full-screen bubbletea TUI with left skill list and right detail panel. Features include:

  • Color-coded dots (red/yellow/green by percentile) indicating relative token cost
  • Tab/Shift+Tab to switch between targets; identical targets are merged into groups
  • / to filter skills, s to cycle sort (tokens↓ → tokens↑ → name A→Z → Z→A)
  • Quality section in detail panel showing all lint findings with icons

• --no-tui flag — disable the interactive TUI and print plain text summary

Web UI — Analyze Page​

• Analyze dashboard — new page in the web dashboard showing per-target token usage with a chart, skill table with token breakdown, and lint issue indicators
• Skill detail token breakdown — the Skill Detail page now shows always-loaded and on-demand token counts
• GET /api/analyze endpoint — REST API returning per-target context analysis with lint issues, skill paths, tracked status, and descriptions

Web UI — Update Page Improvements​

• Sticky search filter — the search input on the Update page now sticks to the top when scrolling, making it easy to filter skills in long lists
• SplitButton actions — Update page action buttons replaced with a SplitButton component for cleaner interaction

Improvements​

• Unified dialog styling — all modal dialogs (Confirm, File Viewer, Hub Manager, Keyboard Shortcuts, Skill Picker, Sync Preview, Update) now share consistent styling via a shared DialogShell component

[0.17.11] - 2026-03-25​

New Features​

Extras — Flatten Option​

• flatten for extras targets — when flatten: true is set on an extras target, all files from subdirectories are synced directly into the target root instead of preserving the directory structure. This is useful for AI tools (e.g., Claude Code's /agents) that only discover files at the top level:

  extras:
    - name: agents
      targets:
        - path: ~/.claude/agents
          flatten: true    # source/curriculum/tactician.md → target/tactician.md
  

• --flatten flag for extras init — enable flatten when creating a new extra:

  skillshare extras init agents --target ~/.claude/agents --flatten
  

• --flatten / --no-flatten for extras mode — toggle flatten on existing targets:

  skillshare extras agents --flatten
  skillshare extras agents --no-flatten
  

• Flatten in TUI wizard — the extras init interactive wizard now includes a "Flatten files into target root? (y/N)" step after mode selection (skipped for symlink mode)

• Filename collision handling — when flatten causes files from different subdirectories to share the same name (e.g., team-a/agent.md and team-b/agent.md), the first file wins (sorted alphabetically) and subsequent collisions are skipped with a warning

• F flatten toggle in TUI — press F in the extras list TUI to toggle flatten on/off for a target. Single-target extras toggle directly; multi-target extras show a target picker first

Web UI — Flatten Support​

• Flatten checkbox — the Extras page shows a flatten checkbox per target, both when creating extras and on existing targets. Disabled when mode is symlink
• Config editor validation — the YAML config editor warns when flatten: true is combined with mode: symlink
• Target name field docs — clicking a target name in the config editor (both name: claude and short-form - agents) now shows the correct "target name" documentation instead of unrelated field docs

[0.17.10] - 2026-03-24​

New Features​

• Update notification in Web UI — a dialog appears on first visit when a newer CLI or skill version is available. Shows current and latest versions with a copyable skillshare upgrade command. Dismissed once per browser session

Bug Fixes​

• Skill cards equal height — skill cards on the Skills page now stretch to equal height within each row
• Tour step target fix — the skill-filters tour step now correctly highlights its target element

[0.17.9] - 2026-03-20​

New Features​

• Force toggle on Extras page — a new Force button in the Extras page header lets you overwrite existing files when the sync mode has changed. Hover for a tooltip explaining what it does. Previously, skipped files could only be force-synced via the CLI (skillshare sync extras --force)

Bug Fixes​

• Config page assistant panel now scrollable — the right-side Structure/Diff panel now has a fixed 500px content area matching the editor height, enabling vertical scrolling when the YAML structure is long
• Removed false "Unknown target" warnings — the config editor no longer flags custom target names as unknown. Target names are user-defined and freely configurable — any name is valid
• Audit rules assistant panel scrollable — same fixed-height scrolling fix applied to the Audit Rules page's assistant panel
• Extras sync API response key — fixed the JSON response key from results to extras, which prevented the Extras page from displaying sync results

Improvements​

• Richer targets field docs — the targets field documentation example now shows all sub-fields (path, mode, include, exclude) with multiple targets
• Filter Studio virtual scrolling — the skill preview list now uses virtual scrolling for smooth performance with large skill collections
• Force auto-resets after sync — the Force option on both the Sync and Extras pages automatically disables after a successful sync, preventing accidental overwrites on subsequent runs
• Smarter skip toast — when extras files are skipped, the toast suggests "enable Force to override" instead of a CLI command. If Force is already enabled, the hint is omitted

[0.17.8] - 2026-03-19​

New Features​

Extras — Configurable Source Paths​

• Custom extras source directory — extras source paths are now configurable instead of hardcoded. Add extras_source to config.yaml to set a global default, or use per-extra source for individual overrides:

  extras_source: ~/my-extras               # all extras default to here
  extras:
    - name: rules
      source: ~/company-shared/rules       # this one overrides extras_source
      targets:
        - path: ~/.claude/rules
    - name: commands                        # uses extras_source (~/my-extras/commands/)
      targets:
        - path: ~/.cursor/commands
  

  Resolution priority: per-extra source > extras_source > default (~/.config/skillshare/extras/<name>/)

• --source flag for extras init — specify a custom source directory when creating an extra:

  skillshare extras init rules --target ~/.claude/rules --source ~/company-shared/rules
  

• Source input in TUI wizard — the extras init interactive wizard now includes a source directory step between name and target input. Leave empty to use the default

• extras source command — show or set the global extras_source directory from the CLI instead of editing config.yaml manually:

  skillshare extras source                          # show current value
  skillshare extras source ~/company-shared/extras  # set new value
  

• --force flag for extras init — overwrite an existing extra without needing to extras remove first:

  skillshare extras init rules --target ~/.cursor/rules --force
  

• --source rejected in project mode — extras init --source now returns a clear error in project mode instead of silently ignoring the flag. Project mode always uses .skillshare/extras/<name>/

• source_type in JSON output — extras list --json and GET /api/extras now include a source_type field (per-extra, extras_source, or default) indicating which level resolved the source path

Web UI — Extras Source​

• Source type badge — the Extras page shows a (per-extra) or (extras_source) badge next to non-default source paths
• Source input in Add Extra modal — optional "Source path" field when creating extras from the dashboard
• API accepts source field — POST /api/extras now accepts an optional source field in the request body

Web UI — Config Editor Assistant Panel​

• Context-aware assistant panel — the Config page now has a right-side panel that shows relevant information as you edit config.yaml:
  • Field docs — move your cursor to any field and see its description, type, allowed values, and an example snippet. Covers all 28+ config fields (source, mode, targets, extras, audit, hub, log, tui, gitlab_hosts, and sub-fields)
  • Structure tree — visual outline of your YAML structure with line numbers. Click any node to jump to that line in the editor
  • Real-time validation — inline error markers for YAML syntax errors. Schema validation warns about unknown target names (with typo suggestions), invalid sync modes, and invalid audit settings
  • Diff preview — see what changed since last save, with colored add/remove lines. Includes a "Revert All" button to reset to the last saved version
  • The panel auto-switches between views by priority (errors → field docs → structure), or lock to Structure/Diff via the bottom bar
  • Collapse the panel with the toggle button or Cmd+B; save with Cmd+S
• Empty config guide — when the editor is empty, the panel shows all available top-level fields as a quick reference
• .skillignore panel — the .skillignore tab shows a simplified panel with change count and the list of currently ignored skills (from all sources, including tracked repos)

Web UI — Audit Rules Assistant Panel​

• YAML editor assistant panel — the Audit Rules page's YAML editor now has the same assistant panel as the Config page:
  • Field docs — cursor-aware documentation for all audit rule fields (id, severity, pattern, message, regex, exclude, enabled) with allowed values and examples
  • Regex tester — move your cursor to a regex: field and the panel auto-switches to an inline regex tester. Paste test lines and see match results instantly with highlighted matches. Supports (?i) flag conversion from Go to JavaScript
  • Real-time validation — warns about invalid severity values, uncompilable regex patterns (with Go-specific syntax detection), and YAML syntax errors
  • Structure tree + Diff preview — same as Config editor
  • Three bottom-bar locks: Structure / Diff / Test
• Save button in header — the Save button is now in the page header (top-right) for better visibility

Bug Fixes​

• .skillignore save no longer shows sync preview banner — the "Preview Sync" prompt after save is now only shown for config.yaml changes, not .skillignore
• Tracked repo ignores visible without root .skillignore — previously, if the root .skillignore file didn't exist, the API returned no ignore stats at all, hiding tracked repos' own .skillignore entries. Now always reports all ignored skills regardless of whether a root file exists
• Dirty state guard on tab switch — switching between config.yaml and .skillignore tabs with unsaved changes now shows a confirmation dialog instead of silently discarding edits

[0.17.7] - 2026-03-19​

New Features​

Web UI — Uninstall Skills Page​

• Batch uninstall from the dashboard — a new "Uninstall Skills" page lets you remove multiple skills at once with filtering and multi-select. Filter by group directory, glob pattern (*react*, frontend/*), or type (Tracked / GitHub / Local), then check the skills you want to remove and confirm:
  • Group dropdown narrows to a specific directory
  • Glob pattern input with real-time matching (supports * and ?)
  • Type filter tabs with counts (matching the Skills page style)
  • Select All / Deselect All for the current filtered view
  • Tracked repos auto-escalate — selecting any skill inside a tracked repo selects the entire repo, with a force option for uncommitted changes
  • Results show per-item success/failure with a "Run sync" reminder
• Batch uninstall API — POST /api/uninstall/batch accepts multiple skill names in a single request with skip-and-continue semantics. Includes registry cleanup, config reconciliation, and .gitignore batch removal — matching the CLI's behavior

TUI — Target Remove Action​

• Remove targets with R key — the target list TUI (skillshare list --targets) now supports pressing R to remove a target from your config

Bug Fixes​

• Sidebar no longer shifts on long skill lists — fixed a layout issue where scrolling to the bottom of the Skills page caused the sidebar to shift horizontally

[0.17.6] - 2026-03-19​

Bug Fixes​

• Sync auto-creates missing target directories with notification — v0.17.5 introduced a strict check that blocked sync when a target directory didn't exist (e.g., ~/.claude/skills on a fresh Claude Code install). This prevented first-time users from syncing without manually creating directories (#87). Sync now auto-creates missing directories and shows what it did:

  ✓ claude: merged (99 linked, 0 local, 0 updated, 0 pruned)
  ℹ   Created target directory: ~/.claude/skills
  

  Dry-run mode previews which directories would be created without actually creating them
• skillshare init creates target directories — when init detects an installed CLI tool (e.g., ~/.claude/ exists) but the skills subdirectory is missing, it now creates it automatically instead of leaving it as "not initialized"

Web UI​