* chore(main): release 0.26.1
* chore(release): synchronize release metadata
* docs(changelog): describe 0.26.1 for users
Replace the generated entry with user-facing notes so the release body
explains what changed in the dashboard and in the tracking guidance.
---------
Co-authored-by: github-actions[bot] <41898282+github-actions[bot]@users.noreply.github.com>
Co-authored-by: Willie <egg358358@gmail.com>
The Agents Updates tab updates through /api/update/stream, which resolved
every name against the skills source, so agent updates failed there. The
agent detail page names a tracked repo by its folder, which updateAgent
did not resolve, so updating an agent inside a repo failed with "agent not
found".
The stream now takes a kind and routes agent names to updateAgent, which
also resolves an agent repo folder under the agents source. The tab sends
its kind to the stream and to Prune, and matches update results by kind and
by the agent's path, so a nested agent is marked up to date after it
updates.
Closes#600
* fix(update): honor accepted findings in dashboard tracked updates
The dashboard ignored findings accepted through the CLI, so a tracked repo
that `update --force` let through was blocked again on the next dashboard
update, and every dashboard Force Retry had to be repeated. Regular skill
updates in the dashboard already honored and recorded acceptances, so only
tracked repos differed.
Tracked-repo updates now always honor accepted findings and record the ones
an override lets through, for both callers. The opt-in option and the
AuditGate flag it fed are removed, since no caller turns them off.
Dashboard agent tracked repos passed an empty source dir, so a nested repo
kept acceptances in its group folder under a different key than the CLI.
They now pass the agents source, as the CLI does.
* fix(update): reload the agents store after agent repo refresh
Dashboard agent tracked repos now refresh metadata in the agents source,
but a metadata change reloaded the skills store, leaving the agents store
as it was until the next request. Reload the store that matches the source.
Every caller of update.TrackedRepo now passes a source dir, so the
empty-dir guard and its doc no longer describe a supported mode. The new
agent test reuses the tracked fixture instead of rebuilding it.
Correct Unix-only resolution expectations on Windows and pin terminal and renderer state so the package is deterministic. Add theme to the Windows test job to catch future platform regressions. Related to #344.
Trimming the plugin section dropped the instructions to run only the
requested operation and to read plugins.md before applying changes, so
agents lost the cue to stop at what the user asked for.
The tracking docs fix (#596) and this follow-up ship as 0.26.1.
Release-As: 0.26.1
* docs: clarify upstream skill and plugin tracking
The built-in skill recommends --track --all, which install rejects. Show valid scoped workflows for following upstream collections and native plugins, distinguish update from sync, and explain selection flags before users hit their conflicts.
* docs: clarify default-branch and retained-ref boundaries
Qualify default-branch guidance for unpinned sources so update examples do not imply that an existing ref is cleared. Clarify mixed resource collections and filter modes alongside the workflow.
* docs: address tracking workflow review feedback
Remove rejected tracking/selection combinations from enterprise recipes in all five languages, document incompatible flags in the install tables, and state the selection errors explicitly. Keep the built-in skill concise by routing detailed upstream workflows to its references.
Update the release highlights in every README and fill the missing hooks, pull conflict, and memory guidance translations before reviewing the Release PR.
* fix(plugins): honor native OMP marketplace builds
Resolve OMP entries from their native marketplace when present and inspect package.json extension-only plugins. Keep target-specific source paths, marketplace identities and components so Claude adoption and hybrid manifests remain valid.\n\nShow separate builds in text discovery without changing logical JSON selection, and cover native precedence, invalid catalogs, manifestless and hybrid packages, global/project materialization and CLI output with regression tests. Update the plugin docs and existing translations.\n\nFixes #587
* fix(plugins): preserve native build metadata across consumers
Keep OMP catalog versions in target discovery metadata and materialized marketplace entries so previews, cache guards, and native installs use the same version, including legacy catalog fallback. Cover catalog versions that differ from package manifests.
Show each dashboard discovery target's own version, components, and source path instead of presenting a merged candidate as one build.
* fix(plugins): avoid ambiguous shared build versions
Derive candidate versions from agreement across supported targets and recompute after catalog overrides. No-Agent records and lists no longer inherit an arbitrary native build version; equal versions remain available and unknown versions stay unknown.
Keep Cursor and Antigravity inventory target-specific when manifests disagree. Cover no-Agent preview and persistence, equal and unknown versions, catalog overrides, and mixed-manifest local inventories. Align English and translated plugin references.
* fix(plugins): accept Pi extension manifests for OMP catalogs
Reuse the existing OMP extension detector instead of requiring an OMP-specific manifest key. Native OMP accepts Pi extension manifests too, so the extra gate rejected supported native catalog entries.
Cover Pi-only discovery with scoped package names and global/project materialization. Align the English plugin references and all existing translations.
After a release changed the guidance text, outdated agents showed an
unchecked box in "Connect to agents", so updating looked like it
required disconnecting and reconnecting. An "Update to latest" button
now opens the same review with every outdated agent preselected and
its mode kept. Hand-edited blocks are "modified", not "outdated", so
they are never preselected and the backend still refuses to overwrite
them.
A pull with dozens of conflicting files meant clicking "Keep this
version" once per file. Two buttons now set every file to the same
side; individual files can still be switched afterwards.
* feat(hooks): read hook code and scripts from files next to the config
Code Agent extensions and command scripts lived as strings in
config.yaml, so a longer Pi extension or shell script was edited inside
a YAML block scalar with no editor support, and nothing type-checked or
linted it. A breaking change in an Agent's extension API went unnoticed
because the generated file is not part of any tsconfig.
A binding now takes `codeFile: path` instead of `code`, and a `files`
value may be `{from: path}` instead of the contents. Paths resolve
against the folder of the config that holds the entry and are read
every time the config is read, so the regular typecheck or shellcheck
can cover the real source. The revision and drift checks use the
resolved contents, so a file edited after a preview fails the apply as
stale.
Saving keeps the references: enable/disable, `hooks edit` and the
dashboard write `codeFile` and `{from}` back instead of inlining the
contents. JSON carries the contents plus `codeFile`/`filesFrom`, and an
entry echoed back with both must match the file, so an edit is refused
rather than silently dropped. Paths must stay inside the config folder
without symlinks, so a cloned project cannot pull in files from
elsewhere. References are taken from the YAML tree before one strict
decode, which keeps unknown-field checks and anchors shared across
bindings working.
Closes#585
* fix(hooks): refuse an emptied file sent with its reference
An empty files value stood for "contents not sent", so a file cleared in
the dashboard's Git YAML editor while its filesFrom reference stayed was
treated as unsent: the file was read again from disk and the save
reported success with the user's edit discarded.
A value read from {from: path} now keeps a dedicated mark until it is
resolved, so an empty value is compared with the file like any other
contents and refused when they differ. Code keeps using empty as unsent,
since a code binding can never be empty.
* fix(hooks): render referenced files as they are on disk
RenderNative rendered the contents the caller sent. The dashboard sends
the entry it loaded, so after a referenced file changed, View
configuration still showed the old contents while describing them as
what sync would write.
Render now drops sent contents for codeFile and {from} references and
reads every reference again, like preview and save. Draft, render and
config decoding are the only paths that take an entry, and all three now
resolve references.
* fix(hooks): accept file references in JSON entries and anchored values
Two inputs that config.yaml accepts failed to load elsewhere.
A `{from}` value carrying its own YAML anchor, aliased from another
binding, lost its anchor because the value node was replaced instead of
rewritten, so the alias no longer resolved. The node is now rewritten in
place with its anchor kept.
An entry JSON file (`hooks add|edit --file x.json`) could not write a
files value as `{"from": path}`, because JSON decoded straight into
string contents. Binding now decodes JSON files values as contents or a
reference, strictly, like YAML, so both entry formats and the Web API
take the same shapes.
ParseEntry trimmed the whole document before decoding it, which was
meant only to detect a leading `{` for JSON. The YAML branch decoded
the trimmed bytes, so a `code: |` block scalar ending the file lost
its final newline: under clip chomping the line break survives only
if it is present. A Pi or OMP extension added from such a file was
written without a trailing newline and failed formatters such as
`prettier --check`. JSON input was unaffected because its newline is
escaped inside the string.
The YAML branch now decodes the original bytes.
* feat(mcp): serve an Executor skill index from mcp serve --http
Executor apps load remote skills with wellKnownSkills, but could only
reach Skillshare's skills through a copy in Git, so an edit saved with
apply_skill_edit never reached them without a commit and push. With
--http, mcp serve now also answers /.well-known/agent-skills/index.json
and the files it lists, built from the same catalog and token, so
Executor sees an edit on its next check.
The format is the one Executor's loader reads, not the agent-skills
discovery draft v0.2. One skill that loader rejects fails its whole
load, so the index leaves out skills it would reject (non-text metadata,
license or allowed-tools, files over 2 MB, non-UTF-8 files, file names
decodeURIComponent cannot handle, names used twice) and lists them in
--check and on stderr at startup. Versions are digests of each skill's
paths and file digests, and the output is stable while nothing changes,
because the loader reads the index twice and compares the bytes. An
index over Executor's file or byte limits answers with an error naming
--target.
Manifest files now record their path within the listing skill: a
parent's entry otherwise took its nested skill's paths, which would
have listed SKILL.md twice. The index route repeats the MCP handler's
DNS rebinding check for loopback servers without a token.
Closes#574
* fix(mcp): accept proxied hosts with a token and use Executor's app accounts
Behind the documented TLS proxy, a loopback mcp serve sees the proxy's
public Host, and the DNS rebinding check answered 403 even with a valid
bearer token. The index route had it, and so did the MCP endpoint through
the go-sdk's localhost protection. A rebinding page cannot send the token,
so both checks now apply only when no token is set; the docs say a
tokenless server behind a proxy needs a loopback Host.
The Executor example used the account form apps 0.0.1-beta.90 removed
(`{ accounts }` and `ctx.accounts`, now typed as never). It declares the
token as an app account and reads ctx.app.accounts, in all five locales.
The wellKnownSkills loader itself is unchanged at the v2 head (b98424d9).
* feat(mcp): write per-Agent native fields from the source
Some MCP settings have no portable form, such as Codex's
default_tools_approval_mode or Claude Code's per-server timeout. Sync
kept them only while an entry survived, so a recreated entry (for
example after a rename) lost them, and every machine needed them added
by hand again.
A server's `options` now holds these fields per Agent, and sync writes
them as given. The field ledger that piOptions used is extended to
every Agent, so sync removes an option it wrote once the source drops
it, and reports a conflict when the Agent changed it since.
Options are refused when they could break the entry or leak a secret:
fields Skillshare writes (including portable names some Agents rename),
values a native format cannot write back (null, arrays of tables),
options.pi (piOptions covers Pi), disabled entries, and credentials as
plain text, since the Agent's file may be committed.
The dashboard has no editor for options yet but keeps them on save.
Refs: #578
* fix(mcp): close gaps in per-Agent options found in review
- A credential option passed if it merely contained "${", so text around
a reference, or another Agent's syntax, was written out as plain text.
It must now be a whole-value reference in the Agent's own syntax, and
Codex and Antigravity options, which read no references, cannot hold one.
- The Kilo Code project check looked only at portable env references, so
an {env:VAR} in options reached a project file Kilo refuses to load.
- A replace resolution left the dropped-option conflict in place, so the
plan stayed blocked; replace now lets the source win and prunes those
fields.
- Adopting an entry that already matched skipped writing its options until
a later sync; the takeover now writes them.
* fix(mcp): refuse numeric credentials and constrain option values in the schema
- YAML reads an unquoted secret such as clientSecret: 123456 as a number,
which the credential check skipped, so sync wrote it out as plain text.
A mapping under a sensitive name such as credentials also hid its
plainly named fields. Sensitivity now carries down through mappings and
lists, and a number under a sensitive name is refused.
- The schema allowed any value inside an Agent's options, so editors
accepted null or nested lists that LoadSource then rejects. A recursive
optionValue definition now matches the runtime rules.
* feat(ui): edit other Agents' MCP options in the server dialog
The dialog only carried options through unchanged, so setting a field
such as Codex default_tools_approval_mode or Claude Code timeout still
meant editing mcp.yaml by hand. One Agent picker with a JSON editor keeps
the dialog short however many Agents are ticked; ticked Agents come first,
and an unticked Agent that still has settings stays listed so they can be
cleared. Anything but a JSON object blocks saving; the server checks the
fields themselves.
* fix(ui): offer every Agent in the options picker and document the editor
- The picker listed targets from the order, which includes accounts, so a
ticked account produced an options key the server refuses, and the
Agent the account writes to was not offered. Options are keyed by Agent,
so every Agent but Pi is offered, ticked ones first.
- An unticked Agent left the list as soon as its text was emptied, so its
settings could not be cleared and retyped.
- The editor was hidden on the Skillshare tab although its server takes
options too, and hidden invalid text could block Save there.
- The docs still said the dashboard had no options editor.
* fix(ui): open the options editor on a ticked account's Agent
The picker ranked Agents by the ticked targets, which hold account names,
so a server reaching only codex-work opened the Claude editor and what was
typed there was saved under options.claude. The MCP page now provides
each account's Agent, and the ranking uses it.
The retype test now asserts the saved options rather than the editor
staying visible.
* fix(ui): show the server's refusals of Agent options in the dashboard language
The dialog leaves field rules to the server, but its refusals (a field
Skillshare writes, a plain-text credential, a credential in an Agent that
reads no references, an unsupported value) reached the dialog in English.
describeError now maps each to a translated message naming the Agent and
field. The other option errors cannot come from the dialog: it offers no
Pi, no non-Agent key, and no options on an off switch.
* feat(mcp): write per-Agent native fields from the source
Some MCP settings have no portable form, such as Codex's
default_tools_approval_mode or Claude Code's per-server timeout. Sync
kept them only while an entry survived, so a recreated entry (for
example after a rename) lost them, and every machine needed them added
by hand again.
A server's `options` now holds these fields per Agent, and sync writes
them as given. The field ledger that piOptions used is extended to
every Agent, so sync removes an option it wrote once the source drops
it, and reports a conflict when the Agent changed it since.
Options are refused when they could break the entry or leak a secret:
fields Skillshare writes (including portable names some Agents rename),
values a native format cannot write back (null, arrays of tables),
options.pi (piOptions covers Pi), disabled entries, and credentials as
plain text, since the Agent's file may be committed.
The dashboard has no editor for options yet but keeps them on save.
Refs: #578
* fix(mcp): close gaps in per-Agent options found in review
- A credential option passed if it merely contained "${", so text around
a reference, or another Agent's syntax, was written out as plain text.
It must now be a whole-value reference in the Agent's own syntax, and
Codex and Antigravity options, which read no references, cannot hold one.
- The Kilo Code project check looked only at portable env references, so
an {env:VAR} in options reached a project file Kilo refuses to load.
- A replace resolution left the dropped-option conflict in place, so the
plan stayed blocked; replace now lets the source win and prunes those
fields.
- Adopting an entry that already matched skipped writing its options until
a later sync; the takeover now writes them.
* fix(mcp): refuse numeric credentials and constrain option values in the schema
- YAML reads an unquoted secret such as clientSecret: 123456 as a number,
which the credential check skipped, so sync wrote it out as plain text.
A mapping under a sensitive name such as credentials also hid its
plainly named fields. Sensitivity now carries down through mappings and
lists, and a number under a sensitive name is refused.
- The schema allowed any value inside an Agent's options, so editors
accepted null or nested lists that LoadSource then rejects. A recursive
optionValue definition now matches the runtime rules.
* docs(mcp): name every Agent whose options cannot hold a credential
The docs named only Codex and Antigravity, but Claude Desktop, Goose,
Junie, LM Studio and Warp read no environment references there either,
so a source such as options.goose.apiKey: "${GOOSE_KEY}" looked valid
while LoadSource refused it.
AgentIcon had no entry for git, so places that show a Git source fell
back to the generic icon. The logo comes from Simple Icons (CC0); the
Git Logo itself is by Jason Long under CC BY 3.0, credited next to the
import.
* chore(main): release 0.25.13
* chore(release): synchronize release metadata
* docs(changelog): write the 0.25.13 entry for users
Replace the generated notes with user-facing descriptions and an example
config for mcp serve skill editing, and drop the Release-As request line.
---------
Co-authored-by: github-actions[bot] <41898282+github-actions[bot]@users.noreply.github.com>
Co-authored-by: Willie <egg358358@gmail.com>
Codex reviews often took four to six rounds because each round found
one more place where the same bug class was unhandled, and it re-raised
cases outside the supported workflows that only our internal wiki ruled
out. AGENTS.md now carries code review rules Codex reads: supported
scope, one finding per bug class with every location, and re-reviews
limited to new commits. The pull request workflow now fixes every
instance of a confirmed class at once and, after three rounds, explains
why findings keep coming instead of only stopping.
The mcp serve read-only hints and opt-in skill editing (#577) ship as
0.25.13 instead of the computed 0.26.0, per maintainer decision.
Release-As: 0.25.13
* feat(mcp): let mcp serve clients edit listed skills
A remote assistant that reaches skills only through `mcp serve` could
read them but had to hand every change back to a person (#540). Skills
listed in `mcp.serve.edit.skills` can now be changed one existing file
at a time: `preview_skill_edit` writes nothing and returns the full
diff, the audit result and a preview id, and `apply_skill_edit` saves
exactly that preview, once.
Every save is tied to the revision the client read, so a file changed
in the meantime fails with a conflict instead of being overwritten.
Skills that `update` would overwrite stay read-only: tracked repos,
upstream installs, and skills nested inside an install. The skill is
validated and audited in a temporary copy before anything is written;
as with install, a finding at or above `audit.block_threshold` refuses
the edit unless it was accepted with `--force`, and findings are marked
new or existing so the client can see what its change adds.
The write is atomic, keeps the file mode, goes through sourcefs link
checks, and is recorded in the operation log. Native Skills extension
clients still see the edit tools, since the extension cannot write.
* fix(mcp): close gaps in mcp serve skill editing
Codex review of #577 found five ways the edit tools could do more, or
record less, than they promise:
- A blank pattern in mcp.serve.edit.skills was dropped by the filter,
and no patterns match every skill, so `skills: [' ']` made every skill
editable. Blank patterns are now a config error.
- preview_ttl was only checked when skills were listed, so the dashboard
could save an invalid value that later stopped mcp serve. It is now
checked whenever edit is present.
- Apply compared only the edited file's digest, so two previews of
different files, or an edit made elsewhere in between, could save a
skill state that was never validated or audited. Apply now checks and
audits the skill again as it will be saved.
- Edit audits were not written to audit.log, so blocked attempts left no
trace in `skillshare log --audit`. Each preview and apply audit is now
logged like install's audit gate.
- The writability check migrated legacy metadata, so a "writes nothing"
preview (and --check) could rewrite .metadata.json and delete
registry.yaml. It now reads the store without migrating; a legacy
registry.yaml keeps every skill read-only, and a legacy sidecar marks
its skill as installed.
Clients and gateways that gate tools by annotation treat a tool without
readOnlyHint as one that changes data. Executor v2 does, so every
list_skills or read_skill call could stop for approval even though both
only read the skills source. Marking them read-only lets those clients
call them without asking. Refs: #540.
The windows-test job lists only the packages verified to pass on Windows, so
that the job is green on arrival. internal/memory never reached that list: it
is absent from #344's set of failing packages, and no commit ever added it,
which leaves its 36 tests unrun on the platform whose binaries are published.
Cross-compiled the package's test binary for windows/amd64 and ran it on a
real Windows host: 36 top-level tests passed, 0 failed, exit code 0. The
symlink call sites in store_test.go skip or drop their subcases when the
privilege is missing, so a runner without Developer Mode cannot fail on them.
* chore(main): release 0.25.12
* chore(release): synchronize release metadata
* docs(changelog): write user-facing 0.25.12 notes
The generated entry listed commit subjects only. Describe each fix by what a
user sees: the token in ps, unchanged skills reinstalled by update --all, the
silent fetch failure, missed updates behind symlinks, and the dashboard
breaking after upgrade, with examples and issue references.
---------
Co-authored-by: github-actions[bot] <41898282+github-actions[bot]@users.noreply.github.com>
Co-authored-by: Willie <egg358358@gmail.com>
0.25.12 changes two things users can notice that the docs never covered.
Token auth now goes through a host-scoped credential helper (#566), so the
token stays out of ps and git skips the user's own helper or keychain for
that host. Grouped updates now warn when a repository cannot be fetched
(#570), with credentials removed from the URL and error. Note both on the
install and update pages in every locale.
* fix(update): stop trusting tree hashes for skills with symlinks
Install copies a file symlink's target, but git stores the link as a blob
holding the target path, so a skill's subtree hash stays the same when a
target outside the folder changes. Batch update skipped such skills as
current and check reported them up to date, leaving stale content installed.
A subdir that holds a symlink now has no usable tree hash: install records
none, update falls back to a reinstall, and check reports the skill as
update_available. Check lists entries with -z so links with non-ASCII names
are not missed behind git's path quoting.
Closes#572
* fix(check): split ls-tree -z output on NUL only
parseLsTreeOutput also split on newlines so the old newline fixtures kept
passing, but with -z a newline belongs to the path. A directory or symlink
whose name holds a newline was cut in two, and a symlink under it was missed.
The fixtures now use NUL-terminated entries like real -z output.
Also writes both shortcut comments in the repo's one-line form and uses
testutil.RunGit in the new update test instead of a raw exec call.
* docs(update): note that skills with symlinks always reinstall
The update page says grouped updates leave skills whose folder did not
change alone. A skill with a symlink no longer has a usable tree hash, so it
is reinstalled whenever the upstream commit moves. State that exception on
the English page and its four translations.
* fix(update): read store in batch fast path
The grouped update fast path read per-skill .skillshare-meta.json
sidecars. The migration to .metadata.json deletes those files, so the
check never matched and update --all reinstalled every skill on every
run.
Read the central store instead, and refresh version and commit there
instead of writing a sidecar. Require SKILL.md on disk so the update
restores a deleted install. --force now skips the check and reinstalls
everything.
* test(update): cover legacy grouped metadata keys
Grouped skills can still be stored under a basename key with a Group
field. Check that the current-skill lookup finds that entry and that the
version refresh updates it in place instead of adding a new key.
* feat(update): report up-to-date skills
The summary lumped skills the fast path skipped in with failures. Count
them as "up to date" in the summary and as up_to_date in --json
output.
* feat(update): log up-to-date count in oplog
The operations log recorded updated, skipped, pruned, and blocked counts
for update but not the new up-to-date count. Record it as up_to_date
so the log matches the summary and --json output.
* test(update): cover up-to-date batch skills
update --all reinstalled every grouped skill on every run. Check that
unchanged skills report as up to date in global and project mode, that
an upstream change updates only the changed skill, and that --force
reinstalls everything.
* docs(update): describe up-to-date batch skills
Describe how update --all and --group leave unchanged skills alone and
report them as up to date, the up_to_date JSON field, and the new
meaning of --force.
* fix(update): keep lock pins on tree-only matches
A tree-hash match after upstream HEAD moved refreshed the full commit in
the store. A project lock pin follows that field, so an unrelated
upstream commit rewrote skills.lock.json even though the skill's files
did not change. Refresh only the version, as the old sidecar refresh
did. check compares the version, so it still reports the skill as up to
date.
The help text, docs and skill reference said --all and --group skip
current skills. The skip lives in the batch path, which runs for any
update of several skills, and a single skill always takes the reinstall
path. Say that instead.
* fix(update): load the store once per batch
Each skill skipped as current reloaded .metadata.json before comparing
its version, so an unchanged batch reparsed the whole store once per
skill. Compare the version already read from the snapshot first, and
refresh the skills that do lag HEAD with one load and save after the
loop. An unchanged batch of 400 skills drops from about 520ms to 64ms.
The help text and docs said any update of several skills takes this
path. In global mode, named and glob targets carry no metadata and
reinstall on the standalone path. Name what does take it: --all,
--group, and several skills named in project mode.
---------
Co-authored-by: Willie <egg358358@gmail.com>
* test(update): cover batch fetch failure warning
A batch update that cannot fetch a repo counts its skills as skipped and
prints nothing about it. Remove the remote after installing two grouped
skills and check that update --all names the repo in a warning.
* fix(update): warn when a repo fetch fails
When a clone failed during update --all, the summary counted its skills
as skipped and printed nothing. That hid a repo that timed out on every
run. Print the repo and the error.
* fix(update): redact credentials in fetch warning
The fetch warning printed the stored repo URL as is. An HTTPS URL can
carry a token as userinfo, and git error text can echo that URL, so the
token reached stdout and CI logs. Name the repo with DisplayURL, which
already strips userinfo for submodule URLs, and redact the userinfo
from the error text.
* fix(update): redact query tokens in fetch warning
The fetch warning redacted only the userinfo of the repo URL. A token in
the query survived in the error text, since git echoes the full URL.
And DisplayURL treats the last "@" anywhere as the end of userinfo, so
an "@" in the path named a different repo.
Parse http(s) repo URLs the way git does: userinfo ends at the last "@"
before the first "/", and the query and fragment go too. Remove the
same parts from the error text. Other URLs print unchanged. DisplayURL
goes back to being private to submodule.go.
---------
Co-authored-by: Willie <egg358358@gmail.com>
An already-running UI server still reads lazy-loaded chunks from its versioned cache after the CLI binary is upgraded. Preserve the pre-upgrade version alongside the downloaded version so upgrade cleanup does not break page navigation.
Add a multi-version cleanup regression test and document the retention policy and accepted consecutive-upgrade limitation across all website locales.
Token auth rewrote the clone URL with url.<user>:<token>@host.insteadOf.
git hands the rewritten URL to git-remote-https as an argument, so any
user on the machine could read the token with ps for as long as a clone
or fetch ran. Every platform token (GitHub, GitLab, Bitbucket, Azure
DevOps, Gitea, CNB) went through the same rewrite.
git now gets the token from a credential helper scoped to the host,
which reads it from SKILLSHARE_GIT_AUTH_PASSWORD; printf is a shell
builtin, so the token never reaches an argv. An empty helper entry
first clears the user's own helpers for that host, so git neither asks
their keychain nor erases from it on a 401, and no longer stores the
skillshare token there as the URL rewrite did.
An http.extraHeader was tried first and dropped: the header goes out
before any 401, so a stale token broke installs from public repos, and
after a rejection git fell back to the user's helper and erased their
stored credentials.
Refs #565
The callout in every README now names the 0.25.11 changes: the MCP pin
and tool-change checks and the list detail panel's Targets row.
aleps001 reported the backup staging leak fixed in #558 (#556).
* chore(main): release 0.25.11
* chore(release): synchronize release metadata
* chore(release): write user-facing notes for 0.25.11
Replace the generated commit list with notes that explain what users
see: the MCP pin and tool-change checks with examples, the backup
cleanup fix, and the list detail panel's Targets row. The internal
skills code-review change is left out because it does not reach users.
---------
Co-authored-by: github-actions[bot] <41898282+github-actions[bot]@users.noreply.github.com>
Co-authored-by: Willie <egg358358@gmail.com>
* fix(list): show every target sync placed a skill or agent in
The Targets row of the list detail panel only looked for a symlink named
after the skill's flat name in the target's skills folder. It therefore
left out targets that sync fills in copy mode, targets linked whole in
symlink mode, targets using standard or prefixed naming, and every agent
target (agents were looked up in the skills folder). Targets with skills
turned off were still listed.
The row now asks sync's own name resolution and manifest which entry a
target holds for the skill, and checks agents in their resolved agents
folder by mode or extension output. Each target is resolved once per load
so moving through a large source under standard naming stays fast.
* fix(list): drop a skill's targets as soon as it is disabled in the TUI
Pressing t updated the row's disabled flag but not the discovered skills
that sync's name resolution reads, so the open list kept showing the
skill's targets while a reopened list, which rediscovers it as disabled,
showed none. The toggle now updates that flag too, so both agree: a
disabled skill lists no target, since the next sync removes its entries.
The Playwright example now pins @0.0.83, so readers asked whether a
version is required. It is not: without an exact version (or with
latest, a tag or a range) mcp check only warns, and the server still
syncs and starts.
* feat(mcp): pin tool definitions in mcp check --live and warn when they change
An Agent reads a server's tool names and descriptions as instructions,
and a server can change them after the user reviewed it (a "rug pull").
The MCP spec 2026-07-28 local-server security guide notes the protocol
does nothing about this and advises clients that re-prompt on a changed
definition; most clients do not.
`mcp check --live` now records one SHA-256 per tool over name,
description and inputSchema the first time a server answers, in
mcp/tool-pins.json of the state directory, keyed by scope and server.
Later probes report a changed or new tool as a `tool-change` warning and
a removed one as info, until `--accept-tools` (only with --live) records
the current definitions. Without --live the pins are never read or
written. An unreadable pin file is never rewritten, since it may hold
other servers' pins. The dashboard words both warnings in every locale.
* fix(mcp): keep schema numbers exact, name the scope to accept in, and log acceptance
Codex review on #560:
- Tool definitions were decoded through float64 before hashing, so a
schema bound changed from 9007199254740992 to 9007199254740993 hashed
the same and raised no warning. Numbers now keep their literal text.
- The suggested accept command left out -g/-p; run from inside a project
after a global check, it would accept the project's server instead.
It now names the scope of the check that warned.
- Accepting changed tool definitions approves a possible rug pull, but
left no record. --accept-tools now writes an operations log entry with
the servers it accepted.
* fix(mcp): log a tool acceptance only when its baseline was saved
Codex review on #560: when comparing succeeded but writing tool-pins.json
failed (lock held, state directory unwritable), the report kept its
"accepted" finding, so the operations log listed the server as accepted
with status ok although its old pins remained. Such a server is now
logged under `failed` and the entry's status is error.
* fix(mcp): log first pins as accepted and report errors in the accept entry
Codex review on #560:
- A check that accepted one server but found an error on another still
logged status ok, while the command itself exited 1. A server with an
error is now listed under `failed` and the entry's status is error.
- --accept-tools on a server's first probe records its baseline, but the
finding says "pinned", so the entry listed nothing as accepted. A first
pin counts as accepted now.
* chore(release): request version 0.25.11
The MCP pin checks (#559, #560) and the backup staging fix (#558) ship
as 0.25.11 instead of the computed 0.26.0, per maintainer decision.
Release-As: 0.25.11
* feat(mcp): warn when a local server runs an unpinned package or image
An entry such as `npx -y some-mcp` fetches whatever the registry serves
at every Agent start, so a compromised release reaches the machine on the
next launch without anyone noticing. The MCP local-server security guide
(spec 2026-07-28) names version pinning as a baseline and points out
that a container tag can move while a digest cannot.
`mcp check` now adds a `pin` finding from command and args alone, with no
network: npx, uvx and `pipx run` with no version or `latest`, and
`docker`/`podman run` with no tag or `latest`, are warnings; an image
pinned by tag but not by `@sha256:` digest is info. Value-taking run
flags are skipped so hardened lines such as `--cap-drop ALL` are not read
as the image. Warnings keep the exit code at 0, as before.
* fix(mcp): check every npx package, word pin warnings in the dashboard, document them in doctor
Codex review on #559:
- npx accepts --package more than once and fetches each one, but the pin
check stopped at the first, so `npx --package=lodash@4.17.21
--package=cowsay cowsay` passed although cowsay is unpinned. Every
--package/-p value is now checked.
- The dashboard words env, command and dns findings in the viewer's
language but showed the new pin warning in English. It now has its own
sentence for an unversioned package and for an untagged image in every
locale. The dashboard hides info findings, so the digest note needs none.
- doctor promotes check warnings, so an unpinned launcher now shows there;
its docs list the pin check.
The mcp check runbook gains a step for the pin check.
* fix(mcp): check uvx --with packages and ignore environment markers in pins
Codex review on #559:
- uvx --with (-w) installs each named package beside the tool, so an
unpinned one moves just like an unpinned tool, but the pin check
skipped it as a flag value. Every --with requirement is now checked
next to the --from or command package.
- In `--from 'mcp-server; python_version == "3.12"'` the == belongs to
the environment marker, not to the package, yet it read as a pinned
version. Only the requirement before ';' is inspected now.
* fix(mcp): treat only exact versions as pins, and pin the docs' own examples
Codex review on #559:
- npm resolves a dist-tag (`@next`) or a range (`@^1.2.3`) to whatever
release matches at launch, yet any selector but `latest` passed as
pinned. An npx package is pinned now only by an exact semver version,
the rule Pi bindings already use.
- `==1.*` is a prefix match in Python version specifiers, so
`uvx --from 'mcp-server==1.*'` can pick up a new 1.x release; it is no
longer read as a pin.
- The Context7, Playwright and filesystem examples in the docs and the
built-in skill launched unpinned packages, so copying them produced the
new warning at once. They now name the current registry version.
* fix(backup): reclaim abandoned staging directories safely
Interrupted backup copies leave hidden .snapshot-* directories that retention cleanup never discovers. Hold a shared process lock during copies and use an exclusive cleanup lock to reclaim orphaned staging without deleting active work.
Share retention planning with non-mutating dry runs, report busy manual cleanup in the CLI and API, and align dashboard messages and translated docs. Cover process interruption and automatic cleanup with regression tests and an executable E2E runbook.
Closes#556
* fix(backup): address staging cleanup review findings
Keep cleanup reachable for staging-only dashboard backups and disclose abandoned incomplete directory deletion in every confirmation locale. Cover the empty-state confirmation with a regression test.
Read staging entries before a read-only lock probe so first-use dry runs cannot count a newly active writer. Refresh entries only under the acquired lock and classify both backup-root reads in the filesystem guard.
* fix(backup): align cleanup guides and doctor status
Disclose abandoned incomplete directory removal in backup help, completion descriptions, and the daily backup guide with all four translations. Add help and completion regressions so these consumers stay aligned with cleanup.
Keep the doctor Backups row visible when only the persistent lock remains after cleanup, preserving the existing JSON check and covering the lock-only state.
* fix(backup): complete cleanup documentation disclosure
Cleanup reclaims abandoned incomplete directories regardless of retention, so retention-only summaries understated the action. Align the backup synopsis, option table and previews, related workflow guides, built-in reference and all four website translations. Scope the retention defaults to completed snapshots.
None of these paths exists in the checkout or is written by a script, build or workflow in the repo: old planning and tool folders (docs/plans, docs/superpowers, openspec, .spectra, .superpowers, .feature-radar, showcase), MEMORY.md and internal/server/dist. !/proposals/ re-included a folder no rule ignores. git status --ignored lists the same paths before and after.
Entries had piled up under "Test" and after "Air hot-reload" regardless of what they ignore, and /specs/ was listed twice. Group them by purpose and drop the duplicate and openspec/.vector-search.db*, which openspec/ already covers. git status --ignored lists the same paths before and after.
skillshare sync -p rewrites both manifests on every run, if only to bump updated_at, so each sync left a diff to discard. The links they describe stay tracked; the manifests stay on disk for sync.
The skill syncs to .agents/skills, so Pi and Codex read it too. Naming Sonnet made Pi fail to start every review subagent, since its model registry has no such model. Use the tool's default subagent model, and run the passes in sequence where a tool has no subagents.
Contributors had to run `skillshare sync -p` before the project skills
showed up in Claude Code or in agents that read .agents/skills. Tracking
the links makes them available right after cloning. The manifests drop
skillshare-codebase-audit, which was removed in b9304e01.
The code review skill also treats translated docs as consumers, so an
English page change without its website/i18n copies is reported.
The skill restated the audit topic in an older form: its flag grep matches
no file because the CLI parses its own arguments, and its fixed list of
mutating commands contradicts the topic's rule to judge from current
behavior. The audit topic stays the source for consistency audits.
The contributor page listed it, so it now lists skillshare-code-review in
its place, and uses the skills' real names, which gained the skillshare-
prefix in an earlier rename.
Reviews so far ran /simplify and a generic /code-review separately. Neither
knows AGENTS.md, the wiki topics, or the consumer sweep the pull-requests
topic asks for, so Codex kept finding missed consumers one round at a time.
skillshare-code-review runs correctness, rules, spec, consumer and cleanup
passes in parallel, verifies each finding against the code, and reports by
pass. AGENTS.md now asks agents to run it with `fix` on their own change
after finishing a request and before opening a pull request; a review the
user requests stays read-only.
* chore(main): release 0.25.10
* chore(release): synchronize release metadata
* docs(changelog): describe 0.25.10 for users
Replace the generated commit list with what changed on the dashboard: Open folder on the MCP page and a project's Agent files on its MCP tab.
---------
Co-authored-by: github-actions[bot] <41898282+github-actions[bot]@users.noreply.github.com>
Co-authored-by: Willie <egg358358@gmail.com>
The README callouts still described 0.25.6. They now describe the MCP
page's Open folder and the project MCP tab's Agent files, in every
translated README.
The built-in skill said the MCP page has a Projects tab; that tab
redirects to the Projects page, where each project's MCP tab now also
lists the project's Agent files.
* feat(mcp): open an Agent's MCP config folder, and project files, from the dashboard
A user who wants their own editor (Notepad++ on Windows) found setting
both PATH and $EDITOR too much just to open a config file, and could not
find any entry point for the MCP files of an mcp.projects root (#538).
Each Agent row on the MCP page now also has Open folder, which opens the
folder holding the file in the file manager, so any program can open it
from there. Before the Agent has written its file, it opens the closest
folder that exists. The project MCP tab gets the same Agents list for
that project's files. The open endpoint resolves project paths from the
config, as it does global ones, so a request still cannot name another
file, and folder and file opens keep the same-machine and headless
guards. The file manager launch is shared with the extensions folder.
* fix(mcp): keep a project's existing Agent files listed for remote viewers
Codex review: the project MCP tab built its Agents list from the
project's targets plus the files this viewer could open. A viewer on a
LAN address or behind a proxy can open none, so an existing file of an
Agent outside the project's targets dropped off the list, although the
docs say it shows with its path and only the open buttons go.
The list now returns which files exist apart from whether this viewer
may open them, and the page shows the open buttons only when local.
* fix(mcp): list the Agents a project's own servers write to in its file rail
Codex review: a server that exists only in one project can pick Agents
outside the project's targets, and sync writes their project files, but
the project tab listed only the project's targets and files that already
existed. Before the first sync such an Agent had no row, so its path
could not be copied and its folder could not be opened.
The list now also takes the targets of the project's own servers. Global
servers reach a project only through its targets, so they add nothing.
The rule lives in mcpView as projectFileTargets, with a unit test.