* fix(frontmatter): read and rewrite SKILL.md that starts with a UTF-8 BOM
Windows editors such as Notepad write a BOM before the opening ---. The
lenient readers (ParseSkillName, ParseFrontmatterField/Fields/List, ReadSkillBody)
did not skip it, so under target_naming standard/prefixed the skill had no
readable name and was silently skipped. Strip the BOM in the shared frontmatter
locator and the line scanner, and in the hub, search, audit and sync/context
readers that split lines themselves (utils.TrimBOM).
The writers kept working on the unstripped text, so SetFrontmatterValue
prepended a second block and ToggleFrontmatterFlag/RewriteFrontmatterList
could leave a BOM in the middle of the file. They now split the BOM off and
put it back in front, rewriting only the name value (or flag/list) as before.
* fix(frontmatter): cover the dashboard parser and stop auditing a leading BOM
A consumer sweep after the first BOM commit found two more places that
mishandled a SKILL.md written by Notepad:
- audit reported the leading BOM as a HIGH hidden-unicode finding (U+FEFF is
in the zero-width rule), so every such skill showed a risk score. A BOM at
the very start of a file is its encoding mark and hides nothing; one past
the start is still reported.
- the dashboard's parseSkillMarkdown saw no frontmatter, so the skill editor
would treat the whole file as body and write a second block on save.
The audit-engine page and its four translations note the exemption.
* fix(ui): keep the BOM when the skill editor saves a SKILL.md
parseSkillMarkdown dropped the BOM so the editor could see the frontmatter,
but SkillEditor writes the composed text back through PUT .../content, so every
dashboard edit silently removed the marker from a Notepad-written file. The
parse result now carries the BOM and composeSkillMarkdown puts it back in front,
matching the Go rewrite paths.
Appending an inherited (merge-key) name at the end of the frontmatter
kept running into what may follow the mapping: a "..." end marker,
comments after it at any indent. Its indent also came from the first
key node's column, which sits after "? " for an explicit key. Insert
the key before the first key instead, with that line's own indent; an
explicit key overrides a merged one wherever it stands.
Rewriting the whole name line dropped whatever sat on the key side: an
anchor on the key (&f name:) left its aliases undefined, and quoting,
"name :" spacing or an explicit "? name" key were rewritten. Keep the
text before the value as written, located from the YAML node, and
replace only the value span, keeping the value's own anchor.
A name that defines an anchor (name: &n prototype) lost it on rewrite
while aliases to it (description: *n) stayed, so the copied frontmatter
had an undefined alias. Keep the anchor on the new value, in both the
line rewrite and the flow-mapping rewrite.
Only a bare "..." was taken for the end marker, so "... # end" (or
comment lines after the marker) still got the inherited prefixed name
appended outside the mapping. Accept a trailing comment on the marker
and skip comment lines while looking for it.
When the name is inherited (a merge key) and the frontmatter ends with
"...", the new name: was appended after the marker, outside the
mapping, so the copy kept the old name or held an invalid second
document. Insert the key before the marker.
A multi-line quoted name such as "proto\ followed by an indented line
fails to decode alone, and the literal first line was returned, so the
skill was skipped. Decode the whole frontmatter whenever the name line
does not decode to a string; keep the literal only for frontmatter YAML
rejects.
Each YAML form the line scanner did not know (explicit "? name" keys
this time, after aliases, merge keys and flow mappings) came back as an
empty name and the skill was skipped under standard and prefixed
naming. Keep the early stop for a plain name line and decode the whole
block otherwise, which also replaces the per-form special cases.
ParseSkillName accepts frontmatter whose root mapping is indented as a
whole, but SetFrontmatterValue wrote name: at column zero beside the
indented keys, so a prefixed copy got malformed YAML. Write the key at
the column the mapping's keys use.
A name brought in by "<<: *defaults" has no literal top-level name line,
so ParseSkillName returned empty and naming validation skipped the
skill. When a merge key is present and no explicit name is found,
decode the whole block; an explicit name still wins.
ParseSkillName decoded the name line alone, so "name: *anchor" lost the
anchor defined earlier and came back as the literal alias, which naming
validation rejected. Decode the whole frontmatter block for an alias.
ParseSkillName stopped at "name: >-" and decoded only that header, so
the name was empty and the skill was skipped under standard and
prefixed naming. Keep reading the indented lines of a block scalar
before decoding it.
Target naming resolution read name: by literal line prefix, so a quoted
key, "name :", a flow mapping or a trailing comment gave an empty or
wrong name and the skill was skipped under standard and prefixed naming;
an indented metadata.name could also win. Match only top-level keys,
decode the name line as YAML, and decode a flow mapping whole, while
still stopping at the name line.
Keys of a flow mapping share one line, so the span-based rewrite added a
second name: beside {name: ...}, or dropped the whole line and the
fields before name when it came last. Write a flow mapping back whole
with the new value instead.
The prefixed rename matched "name:" line by line, while resolution parses
the frontmatter as YAML. A quoted key, "name :" or a multi-line scalar
was accepted upstream but left a duplicate key or orphaned continuation
lines in the copy. Locate the key through YAML and replace its full span.
Prefixed target naming has to rewrite name: in each copied SKILL.md.
ToggleFrontmatterFlag only adds or removes a boolean, so SetFrontmatterValue
sets a plain scalar the same way: one line edited, key order, comments,
CRLF and the body untouched. Both now share the delimiter scan.
A project config can name any YAML file in sources.mcp. A cloned
repository could point it at a file outside the project, such as the
global config or the global MCP source, and the next project-mode MCP
change or 0.23.0 migration rewrote that file. Saves now check the
external source against the project root, as config.yaml already is.
Reading a shared file outside the project still works.
The check also needs the right root. Deriving it from the external
file's path would accept a sibling of the project, so the root is now
passed in. A global config acting for a project in mcp.projects is no
longer treated as a project config, so a dotfile link to it keeps
working instead of being refused.
The containment check appended a separator to the project root. For a
project at / or C:\ the root already ends with one, so the prefix became
// or C:\\ and every save of that project's config was refused as
linking outside the project.
ConfigWritePath followed only one hop of a dangling link and resolved
only an existing parent directory. A project link to an in-project
link pointing outside, or to a missing file under an in-project
directory link that leaves the project, passed the containment check,
and the save then wrote or created the file outside the project.
Follow every link of a dangling chain, with a hop limit, and resolve
symlinks in the longest existing prefix of each path so a missing
directory cannot hide an escaping link above it.
Saves follow a symlinked config.yaml to its target so dotfile links
survive. For a project config that also lets a cloned repository point
.skillshare/config.yaml at any file the user can write, including one
that does not exist yet, and every project save, plugin change, hooks
change, MCP change, or legacy migration would then rewrite or create it.
All five writers now resolve the destination through one helper,
utils.ConfigWritePath. A global config keeps writing through its link.
A project config must resolve inside its project; otherwise the save
fails and names the link so the user can replace it with a regular
file. The one-time legacy migration skips persisting instead, and the
next save reports the link.
On main these two readers scanned SKILL.md line by line and returned at
the line they were looking for, keeping nothing. After the refactor they
first collected the whole frontmatter block, so a large block cost its
full size on every call even when name came first, and an unclosed one
cost the rest of the file. Sync, target naming and listing call them
once per skill or agent. Reported by Codex review on #457.
The per-line rule (which line opens and which closes the block) moves
into blockWalk, which locateFrontmatter and the new scanLenientBlock
both use, so the streaming path and the byte locator cannot drift.
ParseSkillName and ParseFrontmatterField stop at their key again; the
list and fields readers collect the block as they always did. Results
are unchanged for every entry point.
Refs #449
Rebuilding the path readers on the shared locator made them read the
whole SKILL.md into memory, where they used to scan line by line and
stop at the closing delimiter. They run once per skill or agent during
sync and listing, so a large body, or a large file with no frontmatter
at all, now cost its full size on every call.
readHead goes back to a bufio.Scanner and keeps the lines from the
opening delimiter to the closing one, nothing before and nothing after.
Reading through the real Scanner also replaces the code that imitated
its quirks (64 KiB line cap, one \r dropped per line); ReadSkillBody,
which needs the body, takes them from scanLines. An unclosed block is
still kept to the end of the file, as the list and fields readers always
did.
Review follow-ups in the same files: RewriteFrontmatterList no longer
copies the content to a string and back, ParseFrontmatterField's two
loops no longer share an index, and the tests share one file helper.
Results are unchanged for every entry point.
Refs #449
Each reader in internal/utils found the --- delimiters with its own
loop, under four different rules that were only visible by reading all
of them, and the list reader existed twice with a path copy and a bytes
copy that had drifted apart. A fix to one reader, like the mcp serve
fixes to the strict one, did not reach the others, and nobody could say
which rule an entry point followed without tracing it.
locateFrontmatter now does the locating for every reader, and the rule
is a frontmatterPolicy value: strictBlock, lenientBlock, bodyBlock and
rewriteBlock name the rules that exist and the entry points that use
them. The rules themselves are not unified; every entry point returns
what it returned before, including the bufio.Scanner quirks of the
readers that take a path (a line of 64 KiB or more ends the content,
one \r is dropped per line), which now live in scannable. The strict
reader still walks only to the closing delimiter and copies nothing.
Two more cases found by comparing old and new code on random input are
pinned: a block scalar on the last frontmatter line, and \r\r\n endings.
Refs #449
The readers in internal/utils locate the --- delimiters under different
rules, so the same SKILL.md can give different answers depending on the
entry point. Before they are rebuilt on one locator, these cases record
each entry point's current result on the inputs where the rules differ
(indented or late delimiters, unclosed and empty blocks, trailing
spaces, CRLF, BOM, delimiters in the body) and on a line longer than the
bufio.Scanner cap that the path readers silently stop at.
Refs #449
The closing --- was matched after trimming spaces, so an indented line,
which is YAML content, closed the frontmatter. It now has to be exactly
---, like the opening one; a CRLF ending is still fine.
Splitting the file into lines cost one string header per line, so a
16 MiB newline-heavy body allocated about 256 MiB per parse. The parser
now scans lines up to the closing --- and keeps only offsets.
The review fixes left the per-file size and digest split across two
packages: a stat-based budget in the walk, a second bounded read in the
catalog, a SKILL.md special case hashed once per file, and the digest
format written out three times. skillpkg.Load now reads every file once
within one package budget and records its digest and size, so the
catalog only maps them, and skillpkg.Digest is the one digest format.
Also: Build settles skills in one deepest-first loop, the skills/list
cursor uses sort.Search, ParseFrontmatterMap splits the file once, mcp
serve names each value flag once, and the migration report is not
called twice in main.
The manifest lists no links, so a skill whose SKILL.md is a link was
published without its own file and could not be read. It is now skipped
with a warning.
The Agent Skills reference parser requires the file to begin with ---
exactly, so an indented delimiter no longer opens frontmatter. A CRLF
file still parses.
The catalog is reused for up to 5 s, so a file edited after skills/list
was read live and failed the client's digest check. A read now hashes
what it returns and reports the file as changed when the digest no
longer matches; the next refresh lists it again.
Also: a BOM-prefixed SKILL.md is parsed from the stripped bytes instead
of being skipped, a non-text compatibility is rejected rather than
treated as absent, and an oversized SKILL.md is skipped before it is
read into memory.
The first command after an upgrade reports legacy data it moved. For
mcp serve that notice landed on stdout ahead of the handshake, so it now
goes to stderr like serve's other messages.
A SKILL.md that opens frontmatter but never closes it could still parse,
so serve published a file hosts read as having no frontmatter; it is now
skipped with a warning.
Size and mtime cannot prove a file is unchanged (cp -p, rsync -t), and a
stale digest makes hosts that verify digests reject the file for good,
so every rebuild hashes every file again; 1,000 skills rebuild in about
140 ms, at most once per refresh interval.
Serve errors now go to stderr: stdout is the protocol stream, and a
startup error printed there is not JSON-RPC.
SKILL.md must begin with its frontmatter, as the Agent Skills format
requires; a later --- block no longer passes.
A root under mcp.projects is served from the global config, where its
project targets (app@claude) live, so the Skillshare tab now offers that
project's targets beside the global ones and saves no -p, which would
fail in a folder with no .skillshare/.
A nested skill skipped for any reason, not only one left unselected, now
skips its parents too: they were settled before the child was validated
and published its files as their own.
Also from review: a skill whose frontmatter JSON cannot carry (a
non-string key) is skipped instead of failing the whole skills/list page;
a listed file with a backslash in its name can be read; and stdio serve
exits 0 on SIGINT/SIGTERM without writing to stdout after the session.
Agents on a disposable VM, in a hosted assistant or behind an MCP gateway can
reach an MCP endpoint but not the folder skills sync to, so they got tools and
none of the user's skills (#428).
`skillshare mcp serve` is a read-only server for the skills skillshare already
manages. It declares io.modelcontextprotocol/skills and implements skills/list,
skills/get and resources/read: one entry per skill, its URI from the source
path, the full frontmatter, and every file with a sha256 digest and size.
--target applies that target's selection, and the source is reread at most
every 5 seconds so install, update, enable and disable show up without a
restart; unchanged files keep their digest. stdio is the default; --http
serves Streamable HTTP and needs SKILLSHARE_MCP_TOKEN off loopback.
Skills the extension cannot carry are skipped with a warning instead of being
published incomplete: a name that breaks the Agent Skills rules or differs from
its directory, more than 512 files or 16 MiB, or a nested skill that is not
served. --check lists them without serving. The naming rules moved to
internal/skillpkg so sync and the server share them.
Few clients support the extension yet, so the server also offers list_skills
and read_skill tools. A client that declares the extension does not see them,
so it never gets each skill twice.
sync.SelectTargetSkills is the one place that turns discovered skills into a
target's selection (enabled, include/exclude, frontmatter targets).
Refs: #428
A junction cannot point at a UNC path; mklink /J exits 1 for one. The
reparse-point API would instead write an unusable mount point, so return
an error up front and leave nothing behind, as mklink did.
Refs #420
With winsymlink=0 Go reports a junction as ModeSymlink instead of
ModeIrregular, so IsJunction returned false and the trash fallback
recreated the junction with os.Symlink again. Read the reparse tag for
both modes.
Refs #420
The junction substitute name is an NT path (\??\C:\x). A target given as
a Win32 extended path (\\?\C:\x or \\?\Volume{GUID}\x) kept its prefix,
so prepending \??\ produced an invalid name and the junction could not
be read through. Swap the prefix instead of stacking it.
Refs #420
Quoting the paths for `cmd /c mklink` stops `&` from splitting the
command, but cmd.exe still expands `%VAR%` inside quotes, so a path with
a token such as `%TEMP%` pointed the junction somewhere else or failed.
Create the directory and set its mount-point reparse data with
FSCTL_SET_REPARSE_POINT instead. No shell parses the paths, which also
removes the need for quoting. Like mklink /J it refuses an existing
link path, and it removes the directory it made if setting the reparse
data fails.
Refs #420
CreateJunction passed both paths to `cmd /c mklink` as plain exec
arguments. Go quotes an argument only when it has whitespace, so a path
with a cmd metacharacter such as `&` was split into a second command:
the junction was not created, and a crafted name could run a command.
Build the command line ourselves and quote both paths. A Windows path
cannot contain a double quote, so the quoting cannot be broken out of.
The regression test uses unspaced paths with `&`, which fail without
this change.
Refs #420
When the skills source and the trash live on different volumes, the
cross-device fallback in moveLink and copyDir recreated a Windows
junction with os.Symlink. That needs Developer Mode or elevation, which
the junction was chosen to avoid, so `unlink` and `trash restore` could
fail after `link` had succeeded.
Recreate a junction as a junction through a shared helper, and keep the
original link in place when creation fails. CreateJunction moves from
internal/sync to internal/utils so trash can reuse it.
Refs #420
Rewrite frontmatter as bytes and read and write target assignments through sourcefs with the discovery follow policy. Refuse nested SKILL.md links with per-skill batch errors or HTTP 400 while regular followed skills remain editable and outside files stay unchanged.
* perf(upgrade): download release assets over parallel ranges
release-assets.githubusercontent.com throttles each connection on some
networks. The 9.6 MB darwin_arm64 archive took 122-171 s over one
connection but 25 s as 8 ranges, each on its own connection, with an
identical checksum. `skillshare upgrade` and the first `skillshare ui`
of a new version both waited minutes on that download.
DownloadToFile probes with `Range: bytes=0-0`; a 206 answer splits the
file into ranges fetched in parallel from the redirected URL, and a 200
answer streams the single response as before, so servers without range
support cost no extra request. Size limits and checksum verification are
unchanged.
The UI asset download is about 2 MB now, so the docs no longer say 1 MB.
* fix(upgrade): keep the caller's transport for parallel ranges
Each range built its client from http.DefaultTransport, so a caller's
custom transport (private CA, authenticated proxy, custom dialer) only
applied to the probe. The ranges then failed or bypassed the caller's
network setup even though the probe succeeded. Each range now copies the
caller's client with a clone of its *http.Transport, which keeps those
settings and still gives the range its own connection. A transport that
cannot be cloned is shared as is.
Since Go 1.23, os.Lstat reports NTFS junctions as irregular files instead
of symlinks (winsymlink, active for go 1.25 in go.mod). Every check that
looked only at os.ModeSymlink therefore missed the junctions skillshare
itself creates on Windows: linked skills could be treated as local folders,
backed up, copied through, or left behind on prune.
Detect links through utils.IsLinkMode, which also asks the OS about the
reparse tag of irregular entries, and count only real symlinks and
junctions as links so OneDrive placeholders and deduplicated files stay
regular files. restore.go also used os.Stat for its link check, which
follows the link and never matched.
Refs #299
On a slow GitHub CDN the ~9MB release archive can take minutes, and a
bare spinner made upgrade look hung, prompting users to abort it.
Show bytes read / total for the CLI and UI asset downloads on a TTY.
yaml.Marshal defaults to 4 spaces, and only the config package forced 2.
Saving a plugin re-encodes the whole config.yaml, so installing one
reindented a user's entire file; skill/agent frontmatter and
audit-rules.yaml were written with 4 spaces too.
Route all YAML writers (config, registry, plugins, MCP, frontmatter,
audit rules) through one utils.MarshalYAML so they cannot drift apart.
* feat(mcp): sync portable server configuration across six clients
Add inline and external MCP sources with guided CLI and dashboard setup, native imports, conflict previews, ownership tracking, backups, and restore. Support Claude Code, Codex, Cursor, VS Code, OpenCode, and Grok CLI in global and project scopes.
Format YAML on save, document MCP fields in the config assistant, and keep its layout stable for scalar list items. Include schemas, localized UI copy, documentation, and regression and end-to-end coverage.
* fix(config): keep loading config when the MCP section is invalid
A mistake in the mcp section (an unknown key, a targets value that is
not a list, an invalid server field) made config.Load, LoadProject,
Save, SaveIn and ValidateConfig fail. Every command that reads config,
including skills sync, install, list and the dashboard, stopped working
because of settings it never uses.
MCPConfig now keeps a section that does not parse exactly as written and
records the error instead of returning it, so commands unrelated to MCP
load and save the config without touching that section. MCP commands
still validate their source on every run, and the dashboard config
editor calls the new ValidateMCP, so saving an invalid MCP section there
is still rejected.
Also correct the ProjectSources comment: sources.mcp resolves from the
directory containing config.yaml, not from the project root.
* fix(mcp): protect Agent files during concurrent edits and recovery
A release review of the MCP sync engine found ways it could lock users
out, report false conflicts, or change entries it does not own.
Recovery
- A failed write left pending.json behind, and every later MCP command
refused to run. Failed writes now remove their journal, and a journal
whose Agent file changed again is discarded, because ownership hashes
still flag entries that no longer match.
- Preview no longer writes. It reads the ownership that recovery will
record, byte for byte, so a preview stays valid across the recovery
that the next write performs under the lock.
Concurrent Agent writes
- Claude Code rewrites ~/.claude.json on its own, so hashing whole files
expired previews and failed applies. Revisions and pre-write checks
now cover only the MCP entries, and apply rereads each file and applies
the previewed entry edits to its current content, keeping settings the
Agent wrote after the preview.
Comparison and ownership
- Only fields Skillshare renders are hashed, after dropping defaults
Agents add or omit (type stdio/http, empty env, args or headers,
enabled: true) and header name case, so equivalent entries are not
reported as drift. enabled and disabled are compared, so turning off a
managed server is a conflict instead of being overwritten.
- An unmanaged entry that already matches the source is reported as
unchanged but no longer claimed, so removing the server never deletes
an entry the user wrote. Import adopts it explicitly, and without
--replace refuses to rewrite a differing entry, such as a literal token
that import turned into an environment reference.
- Sync keeps Agent-only fields such as timeouts in existing entries.
Import
- Agent-only fields become warnings instead of blocking the import.
- streamable-http is accepted, and enabled/disabled flags are checked
for every client.
- Values carrying a URL password, such as database DSNs, become
environment references, and credential-like arguments are flagged.
Writes and state
- Save only renders the definition first, so a server that could never
sync, such as one without targets, is rejected before it is saved.
- A symlinked config.yaml is written through to its target.
- TOML entries are updated in place, keep CRLF line endings, and repeated
add/remove cycles no longer grow blank lines.
- JSON edits no longer HTML-escape characters such as &.
- GROK_HOME is honored.
- Backups keep the newest 20 per Agent file, pruned by file name, and
damaged backup files are skipped instead of failing the list.
- GET /api/mcp no longer fails when an Agent that no server uses has an
unresolvable path.
* feat(mcp): detect the client format of pasted JSON
Import without --from, the add wizard and the dashboard import API
assumed Claude's mcpServers shape unless a client was named, so VS Code
"servers" and OpenCode "mcp" snippets copied from server docs failed to
import. The client is now chosen by the top-level key, and a bare server
object still reads as Claude's single-entry shape.
TOML stays explicit because Codex and Grok share the format; a .toml
file without --from is still read as Codex.
* refactor(mcp): print CLI results through internal/ui
MCP commands wrote plain fmt output and duplicated statusFromErr, unlike
every other command. Plans, backup IDs and save/apply messages now use
the shared ui helpers, so they get the standard status prefixes and
colors; each change line shows the server name, action and target. The
operation log uses statusFromErr. --json output is unchanged.
* fix(server): serve MCP settings only to localhost or IP hosts
MCP settings include stdio commands that Agents execute, so a web page
must not be able to read or change them. The existing Origin check does
not stop DNS rebinding: once an attacker's domain resolves to 127.0.0.1,
the browser sends that domain as both Origin and Host from a loopback
address, and the dashboard also binds 0.0.0.0 in Docker.
Rebinding always needs a domain name, so every MCP route now also
requires a Host that is localhost or an IP literal, whatever the bind
address. Other dashboard routes are unaffected. Opening the dashboard
through a domain name, including a reverse proxy, therefore returns 403
for MCP settings; the message says to use localhost or an IP address,
because rewriting Host at the proxy would still fail the Origin check
and would reopen the hole.
Remove the unused POST /api/mcp/sync endpoint. Configuring with sync
covers it, and the guard wraps every remaining MCP route.
* fix(sync): make sync --all report MCP results reliably
- "sync -g mcp" and "sync --dry-run mcp" ran skills sync, because mcp
was only recognized as the first argument. It now matches in any
position.
- sync --all --json printed plain-text errors when the MCP preflight
failed, and only an error on conflicts. Preflight errors now use the
JSON error shape, and conflicts also return the credential-free plan so
scripts can see which entries conflict.
- When resource sync failed, MCP was skipped without notice. Text output
now warns that MCP settings were not applied; JSON output omits the mcp
field, which --all otherwise always includes.
- MCP was applied with the preflight revision, which failed whenever the
resource sync changed config or state in between. It is now planned
again under the MCP lock, and conflicts found then still stop the
write.
* fix(sync): recognize agents and extras after flags
sync only checked the first argument for a resource kind, so
"sync --dry-run agents" and "sync -g extras" silently synchronized
skills instead. Sync flags take no values, so the kind now matches in
any position.
The shared parseKindArg keeps its first-argument rule: uninstall, update
and similar commands accept resource names, and a name such as "agents"
must not be read as a kind.
* feat(ui): redesign the MCP dashboard around a server-by-agent matrix
The MCP page listed servers and planned changes as flat text, which made
it hard to see where each server is written and what needs attention.
- A matrix shows each server's status per Agent, summarizes pending
changes and conflicts, and can hide synced entries. Agents whose MCP
file or config directory exists are marked as detected; GET /api/mcp
now returns the detected clients.
- Setup is a two-step dialog: add a remote URL, paste JSON or TOML, or
import from an Agent, then preview which Agent files change. Importing
an entry the Agent already has adopts it, Replace overrides only that
entry, and a target selection equal to the inherited mcp.targets is
saved without pinning targets on the server.
- Conflicts explain their cause and offer Import from <agent> or
Replace. Removal lists the affected Agent files and offers removing
from the source only or removing and syncing. Backups are grouped by
day, with a preview before restoring a file.
- Agent icons appear on the MCP and Targets pages, and the universal
target is marked as a path shared across agents. Brand icons come from
@lobehub/icons-static-svg, with bundled SVGs for Warp and Zed.
- Segmented controls wrap instead of overflowing narrow containers.
- New copy is translated for every supported locale.
* fix(ui): let Sync all apply MCP after syncing resources
Sync all previewed MCP, synced skills, agents and extras, then applied
MCP with the revision from the first preview. When the resource sync
changed config or state, that revision was stale and MCP failed with
"preview again" after everything else had already been synced.
MCP is now rechecked before and after the resource sync by comparing the
reviewed changes (target, name and action) instead of the revision, and
applied with the fresh revision. A different set of changes still stops
before any MCP file is written.
* fix(ui): keep flow-style YAML collections when saving config
Formatting on save forced every collection into block style, so a
hand-written "targets: [claude, codex]" was expanded into a list each
time the config was saved. Each collection now keeps its flow or block
style, indentation is still normalized and comments are preserved, and
flow collections are not padded with spaces.
* docs(mcp): document ownership, normalization and dashboard host rules
- Explain that matching entries stay unmanaged until imported, so
removing a server leaves them in place, and that import adopts an
identical entry while --replace rewrites a differing one.
- Note that Agent defaults and header case are not treated as changes,
that disabling a managed server is a conflict, and that previews stay
valid while an Agent rewrites unrelated settings.
- Describe Agent-specific fields, streamable-http, GROK_HOME, backup
retention, recovery by the next MCP write, and the VS Code limitation
on environment references in headers.
- Document that MCP settings require opening the dashboard by localhost
or an IP address, with SSH port forwarding for remote machines.
- Update the built-in skill for all six clients and the adoption rule,
and the MCP runbook for import refusal, Agent-specific fields and
pulled changes.
* docs(skill): streamline guidance and refresh resource workflows
Keep the built-in skill entrypoint focused and route agent and MCP details to dedicated references. Correct scope, audit, automation, and recovery guidance against the current CLI.
* feat(ui): redesign the dashboard around two styles and a flatter layout
The dashboard had grown page by page, so the same idea looked different
depending on where you met it and pages held controls with nothing to act on.
This rebuilds it on one token set that drives both the Clean and Playful styles
in light and dark, and reshapes the pages around what a person actually does.
Notable changes:
- MCP: replace the server-by-agent matrix with one row per server. The agents a
server writes to are chips inside the row, so the list grows downwards as more
CLIs gain MCP support instead of scrolling sideways.
- Settings: collect General, Backup, Log, Health, Extensions and Files into one
page. The Files tab keeps a panel beside the editor showing what the field
under the cursor does, the file's shape, or the unsaved changes.
- Audit rules: the YAML editor gets the same panel, plus a Test tab that runs a
rule's regex against pasted lines.
- Skills and Agents become separate pages; install, search, updates and trash
move into them. Old routes redirect.
- Tour: rewrite every step's copy, add MCP and Install steps, and anchor each
step to a control that is always rendered so no step silently times out.
- Give every button a pointer cursor; the browser default does not say a button
is clickable and each component class had to opt in by hand.
Docs follow the new Settings tabs and the MCP list.
* chore(devcontainer): move dev servers to uncommon ports
5173, 19420 and 8888 are defaults that other projects on the same
machine already hold, so `make devc` failed to bind and the container
never started. Use 45173, 49420 and 48888 instead, identical inside and
outside the container so every banner, log and doc prints the URL that
actually works.
* feat(ui): link dashboard source counts to their pages
The counts in the playful dashboard's source note named a number but led
nowhere, so reaching Skills or MCP from there meant going back to the
sidebar. Each row is now a link to its own page.
The hover fill uses currentColor because .ss-pinnote.src overrides the ink
tokens for the yellow note.
* feat(ui): build and browse hubs from the dashboard
A hub is a JSON index that lists skills, and until now the web UI could
only consume one someone else had already hosted. Writing the index by
hand was the only way to produce one, and the two halves never met: the
dialog that searched hubs and the page that would have built one shared a
config file but nothing else.
The new /hubs page holds both halves behind two tabs:
Browse pick a hub, filter its contents by tag or text, install from it
without leaving the page
My hubs assemble an index from installed skills, validate it, export it
A draft is offered as a source in Browse straight away, read from the
draft API rather than a file, so you can see what a recipient sees before
deciding whether to host anything. Where the index is hosted is stored on
the draft and stripped from the exported document, so a reopened hub
remembers where it was published and the published file stays clean.
Hub management moves out of the install dialog, which keeps only the
search-source dropdown and links here.
Routing switches to createBrowserRouter for useBlocker, which guards the
builder against navigating away from unsaved entries.
* docs(extras): add shared instructions recipe for AGENTS.md across agents
Answers a recurring question about sharing one instruction file between AI
agents: extras already covers it, since each agent keeps its user-level
copy in its own directory and one extra can target all of them.
Naming the source file AGENTS.md rather than something neutral matters —
Codex concatenates AGENTS.md files and has no import syntax, so it only
recognises the file by name. Claude Code and Antigravity read a fixed
filename of their own but resolve @-imports, so a one-line CLAUDE.md or
GEMINI.md picks up the synced file.
Says plainly that this is not agent memory. Agents keep their own
learnings in private stores — a Markdown directory for Claude Code, a
database for Codex, non-file storage for Cursor — so that side is a
migration problem, not something syncing files can solve.
* feat(ui): filter skills and agents by target
The Targets column showed which targets a resource reaches, but there was
no way to narrow the list down to one. The filter reads the sync matrix —
the same source that column renders — so a result always matches the icons
the rows show.
Options and counts are derived from the matrix, so no target is offered
that would filter to nothing, and the control stays hidden until at least
two targets have something. A selected target that disappears on a kind
switch or an uninstall falls back to All rather than emptying the list.
The toolbar had 24px to spare at 1440px, so the new control wrapped the
row. Source moves from a 325px segmented control to a dropdown like its
neighbours, and every filter now sizes to its label instead of a fixed
width that truncated the longer values.
* fix(ui): tighten page headers and reach hubs from skills
PageHeader carried mb-7 while .ss-wrap already separates its children with a
28px flex gap. Margins and flex gaps add up rather than collapse, so every
page inside the wrapper opened with a 56px hole under its title. Move the
default onto .ss-pgh so the wrapper can zero it, keeping the 28px for pages
that lay out their own children.
Hubs is a source you install from and a catalog you publish, not inventory
you own, so it no longer sits in the Library group next to Skills, Agents and
Plugins. The Skills page header already links to it and the page already
carries a back link there. Skills stays highlighted while the hub page is
open. Plugins takes the Package icon its own page identifies with, so the
sidebar no longer shows two identical puzzle pieces.
* fix(skill): stop implement-feature from running --no-tui as a command
The dispatch-order note wrote "not --no-tui" as !`--no-tui`. In a skill file
that is the syntax for running a shell command at load time, so the skill died
with "command not found: --no-tui" before any of it could be read. Spell the
condition out in words instead.
* feat(list): toggle model invocation for a skill with M
The only cheap gesture in the list TUI was E, which hides a skill from every
harness. Wanting a skill installed but not auto-loaded meant leaving the TUI
to edit SKILL.md, so people disabled skills they only meant to make
manual-only. M flips disable-model-invocation in place and the detail panel
shows a "manual only" badge so the state is never a blind toggle.
Unlike E this edits the skill file, which upstream may own: update skips
tracked repos with local changes, and reinstalling a skill drops the edit. M
says so and asks before writing in those cases. The flag is written as a
single-line edit rather than a YAML round-trip, and turning it back off
removes the line, so two presses return the file byte for byte and a tracked
repo is clean again. Agents are left alone; their harnesses do not read the key.
Refs #283
* feat(plugin): manage complete plugins across agent tools
Add CLI, TUI, and web flows for discovery, adoption, sync selection, updates, and removal while preserving native package formats and unrelated installations.
Support Claude, Codex, Cursor, Antigravity desktop, Pi, and OpenCode. Update help, completions, built-in skill, and documentation; keep plugins outside sync --all.
* docs(skill): describe the redesigned dashboard in the style skill
The style skill still documented the pre-redesign hand-drawn system, so
pages written with it came out on legacy tokens and outside the ss-*
classes; the new Plugins page was the first casualty.
The rewrite names tokens and classes and points to components.css for
their values, because copied values are what went stale last time. The
ui command docs described Playful by the old look as well.
* fix(ui): keep Playful ghost buttons free of stray shadows
The hover and disabled rules outranked the ghost variant's
box-shadow: none, so a frameless button grew a lone shadow arc in
Playful. Bulk toolbar and toast buttons shared the problem.
Ghost also gets a soft outline in Playful: beside 2px ink frames a bare
label such as Cancel did not read as a button.
* refactor(ui): lay out the Plugins page like the MCP page
The page was built on legacy tokens and plain markup, so it missed the
Playful frames and drifted from every other page. A per-agent column
layout would not have scaled either: each agent that gains plugin
support would add a column.
Plugins now follow the MCP server list: one row per plugin with its
agents as toggles inside the row, so the list only grows downwards.
Per-agent sync and update move to package-level actions in the row
menu, and the dialogs use the standard header, body and footer.
* feat(ui): give plugins a kind colour and per-agent actions
Plugins borrowed a neutral grey tile because no kind colour existed,
which made them the one resource without an identity on the page.
Magenta sits furthest from the five hues already taken.
Laying plugins out as rows had dropped the per-agent sync and update.
Retrying one agent after a partial failure should not re-run the rest,
so each agent in the expanded row gets its own menu. Update is left out
for a deselected agent, which has nothing installed to update.
* feat(ui): count plugins on the Dashboard
The source summary listed every resource kind except plugins, so the
Dashboard under-reported what a source holds. The Playful source note
grows to fit a fifth row.
MCP now sits above Plugins in the sidebar to match that order, and its
Dashboard label reuses the sidebar's "MCP" instead of a longer string
of its own.
* feat(scripts): add a host verification environment
Plugin and MCP work talks to the real Agent CLIs, and those live on the
host, not in the devcontainer, where every target reports "not installed".
Verifying by hand meant either pointing a build at real config or
assembling a dozen environment variables each time.
`make verify` builds a host binary and runs it against .verify-home/ in
the repo, redirecting HOME, the XDG dirs and each agent's own config
variable. The environment starts from `env -i`, so a stray
SKILLSHARE_CONFIG in the caller's shell cannot pull the run back to real
paths. Living inside the repo is safe because project-mode detection
only inspects the working directory and never walks up to a parent.
It also seeds the dashboard cache from ui/dist. A dev build serves a
placeholder rather than the working tree, because runtime download only
applies to numbered versions, so the UI was previously unverifiable
without a release.
* fix(ui): show progress while a plugin selection is applied
Toggling an agent shells out to that agent's own CLI, which takes
seconds. The row gave no sign of it: the toggle only greyed out, and
the confirmation badge is hidden while a target is deselected, so
turning one on looked like nothing happened at all.
The toggle being applied now lifts out of the greyed state and spins in
its badge. Agent versions also get room for their real strings, since
"2.1.276 (Claude Code)" was truncated to a width chosen for a bare
version number.
verify.sh gains a 'run' subcommand, so the same throwaway environment
can be driven from a script rather than only an interactive shell.
* fix(ui): spin the control that started a plugin action
Reading the inventory shells out to six agent CLIs and takes about two
and a half seconds, and every action refreshes it afterwards. Nothing on
the page said so: a menu action closed the menu and left the page still,
so the wait read as a freeze.
Each control now tracks its own work. Check, sync and import spin in
place, a menu action spins the row it acts on, and the agents refresh
follows the query itself rather than the mutation flag.
* feat(ui): split MCP add and import by where the configuration comes from
"Import from a target" opened a dialog whose second tab pasted a snippet,
which is not a target and usually comes from a provider's setup page rather
than a tool the user already configured. Pasting is a way to add a server,
so it moves under "Add server" beside the form, and importing now means one
thing: read what an installed Agent already has.
The paste side gains a file picker, matching `mcp import --file`; the API
only ever took content, so the browser reads the file and fills the editor.
The target picker no longer waits for a successful parse, and a footer that
said "writes 0 config files" next to a greyed-out button now names what is
missing instead.
* fix(ui): label the plugins result strip as the last action
The strip and the plugin inventory below it used the same list styling with
no heading on either, so two rows carrying the same plugin name read as one
list. The status word made it worse: "imported" is one of the outcomes
lifecycle.go returns for an action, but it sat where an inventory row shows
state, so it looked like a property of the plugin.
Give it the section heading the Agents list already has, plus a count and a
dismiss button, so it reads as the report it is.
* feat(plugins): expand compatibility and clarify setup guidance
Support additional native targets, safe source links, Git refs, and explicit OpenCode entries while exposing per-target capabilities across CLI, TUI, and UI.
Separate unsupported integrations from environment errors, explain next steps in plain language, and link official documentation. Update built-in guidance, website docs, completions, and regression coverage.
* feat(plugin): serve a plugin's files and translatable discovery problems
The dashboard could only show a plugin's name and targets, so reviewing
what a plugin ships meant opening the snapshot by hand. Files and
ReadFile expose the reviewed local copy read-only through os.OpenRoot,
which keeps reads inside the snapshot even across symlinks.
Discovery problems now carry a key and arguments next to the English
sentence, so the dashboard can translate them while the CLI keeps its
wording. Inventory asks every Agent CLI in parallel and Packages answers
from config alone, so the plugin list no longer waits on the slowest CLI.
* feat(mcp): sync to Pi, show native config, and lay out written entries
Pi has no built-in MCP support, so a server targeting it must name the
third-party extension that will read the file. `--pi-extension` and the
`piExtension` source field record that choice, and sync writes the file
format of the chosen package instead of guessing.
RenderNative and POST /api/mcp/render return what each Agent would
receive for a definition, with secrets kept as references, so the
dashboard can show the result before anything is written.
Sync used to insert entries on a single line and never revisit them. A
managed entry that still sits on one line is now reported as an update
and rewritten one field per line; entries formatted by hand are left
alone.
* feat(ui): choose plugin Agents like MCP, browse plugin files, view MCP config
The plugin fold listed bound Agents with a per-Agent menu, while MCP let
you simply tick the Agents you want. Plugins now work the same way: the
fold discovers where the source can also go, shows those Agents as
unticked toggles, and a tick previews an install. The per-Agent menu is
gone, and updating a plugin targets only the Agents that support it
instead of hiding the action when one of them does not.
A plugin with a local copy can be browsed in a read-only files dialog,
using the same tree as a skill's Files tab. The MCP "view config" moved
out of a fold in the server form into its own view, reachable from the
server menu, so the native result has room to be read. The MCP dialogs
also ask for the Pi extension when Pi is selected.
Playful tile names drop from 25px to 22px: "brainstorming" was 1px wider
than the tile and lost its last letter to the next line.
* docs: reshoot dashboard screenshots for the redesigned UI
The README and website still showed the pre-redesign dashboard, which no longer matches what users see after upgrading.
* docs: describe MCP, plugins and the new dashboard for v0.21.0
The README still announced v0.20.0 and never mentioned plugins, and its
MCP section listed four clients. The built-in skill and the docs pick up
the Pi extension setup, the formatting-only MCP update, and the plugin
fold and files dialog, so agents and users are not surprised by them.
Add PreserveTildeOnSave config flag (default false). When enabled, the
Save path folds $HOME prefixes back to ~ when serializing the in-memory
Config to YAML so the on-disk config stays portable across machines
(e.g. when shared via dotfiles).
The in-memory Config is left untouched; cloneForSave() produces a
shallow copy with path fields rewritten. Mirrors the existing
expandPath() calls performed in Load():
- cfg.Source, cfg.ExtrasSource
- cfg.Targets[*].Path / Skills.Path / Agents.Path
- cfg.Extras[*].Source and Extras[*].Targets[*].Path
Adds:
- utils.FoldHomePath() helper with unit tests (exact-home, child,
sibling-non-home, unrelated-abs, relative, Windows case-insensitive)
- 3 Save-path tests covering default-off, fold-on, and Load→Save
idempotency.
Default off keeps the change backward compatible: existing users see no
behavior difference. Users who share their config via dotfiles can opt
in with:
preserve_tilde_on_save: true
- doctor: replace inline frontmatter parser with version.ReadLocalSkillVersion
which correctly reads nested metadata.version (fixes#119)
- frontmatter_write: use TrimRight instead of TrimSpace for --- delimiter
so indented --- inside YAML block scalars is not mistaken for end marker
- Update changelog and website changelog for v0.18.7
New utility: SetFrontmatterList() writes/removes YAML list fields
in SKILL.md frontmatter with dot-notation support (metadata.targets).
Normalizes legacy top-level targets field. Preserves body content.
New endpoints:
- POST /api/skills/batch/targets — set targets for all skills in folder
- PATCH /api/skills/{name}/targets — set target for single skill
Both use s.mu.Lock() for safe concurrent access.
Includes 15 unit tests covering set/remove/traverse/root-only/not-found.
Part of issue #115.