Since #880, the Dashboard lists a tracked repository that is recorded in
metadata but not cloned on this machine, marks it "No checkout on this
machine", and its Uninstall removes only the record and audit acceptances.
The Dashboard row in the ui reference still described only checkout
uninstalls, so readers could expect the repo to move to trash.
* feat(mcp): support Skills directory reads and metadata listing
Expose accepted skill directories and SKILL.md metadata so MCP hosts can
browse served packages without guessing manifest filenames. Reuse catalog
selection and refresh boundaries, including empty directories and stable
pagination, and normalize strict frontmatter for YAML 1.2 Core readback.
Add regressions and a pinned official conformance runner, and update the
runbook and MCP references to describe the browsing contract.
Refs #895
* fix(mcp): preserve supporting file MIME types in directory listings
Record UTF-8 classification from the same bounded bytes used for package
digests so directory metadata agrees with resource reads for extensionless
and unknown-extension text. Reuse the existing bounded reader without
adding listing-time payload reads or changing package limits.
Share pagination with skills/list, trust the SDK's typed request boundary,
and repair the pagination regression to detect stale catalog snapshots.
Refs #895
* fix(mcp): stream supporting file metadata inspection
Buffering supporting files during catalog refresh allocated with asset size.
Restore one bounded stream for digest, size and UTF-8 classification, retaining
at most three incomplete rune bytes and consuming invalid text through EOF.
Preserve the existing same-file checks, package caps and SKILL.md byte reader.
Keep real package-load benchmarks and strengthen the existing MIME fixtures
for split runes and bytes after invalid sequences. Use one first-match sed
command for the conformance URL without a truncated pipeline under pipefail.
Scoped packages, focused race tests, vet, shell syntax/shellcheck and all 45
pinned official checks pass. Allocation stays near 48 KB for 64 KiB and 8 MiB
assets; multibyte UTF-8 validation still costs CPU versus digest-only loading.
* test(mcp): consolidate directory and frontmatter coverage
Check every directory continuation URI so a duplicated middle entry cannot
hide a missing resource. Keep the existing before/after-cursor refresh cases.
Merge nested-skill acceptance fixtures in the catalog owner, preserving all
skill-count, skip-warning and directory assertions across configured and
targetless cases. Move directory capability coverage to initialization and
merge scalar, alias and rejection cases under the existing parser test.
Give cached-page benchmark skills names that match their URI directories.
Narrow tests, benchmark smoke and vet pass; duplicate-result and stale-catalog
mutations both fail, then restored source passes. Production is unchanged.
* test(install): reproduce loss of pulled tracked records
Exercise the two-machine push and pull sequence and reconciliation states so
missing tracked clones cannot silently erase shared install intent.
* fix(reconcile): preserve tracked records missing local clones
Pull transfers tracked repository metadata but not ignored clones. Keep the
missing records during global reconciliation so another install and push cannot
erase shared intent or accepted audit findings on every machine.
* test(reconcile): isolate tracked record fixtures
Use the shared sandbox fixture for both reconciliation tests so their home
and configuration paths remain isolated from the invoking environment.
* fix(reconcile): retain all sourced tracked declarations
Legacy tracked members and unfollowed source links can hide declarations
from the local walk. Preserve sourced tracked records regardless of their
path so unrelated installs cannot erase them or their audit acceptances.
Document restoring a missing checkout before uninstalling it and cover
migration, link collisions, aliases, and explicit removal.
* test(reconcile): drop fixtures existing tests already pin
The tracked-record unit test also seeded a missing plain record and a
live plain record. TestReconcileGlobalSkills_PrunesStaleEntries already
pins both outcomes: a sourced plain record whose folder is gone is
pruned, and one on disk survives. The tracked record with no source
still catches a fix that keeps everything or skips the prune.
The two-machine test passed HOME and SKILLSHARE_CONFIG to RunCLIEnv,
which already sets both from the sandbox for each child process.
Separate homes, XDG paths, client overrides and git config remain.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
* test(reconcile): retain tracked entry for unavailable link
Expect sourced tracked declarations to survive reconciliation with either
source-link policy, so the test matches the cross-machine retention contract.
* fix(config): prune legacy tracked members
Retain only sourced tracked repository declarations during reconciliation.
Migrated member records are not restorable as repositories and retaining them
would leave permanent misleading install requests and audit acceptances.
* fix(reconcile): preserve tracked intent across followed links
A source link can exist before a tracked repository declaration is pulled. Preserve that declaration when the linked folder is not a checkout so later reconciliation does not discard its record and audit acceptances.
* test(reconcile): cover grouped followed declarations
A followed group can contain the path of a pulled legacy repository declaration, which must retain its tracked intent while its key migrates.
* fix(test): scope collision fixture runner correctly
Keep the isolated CLI runner in the collision subtest that uses it so the integration test package can compile.
* docs(manifest): clarify tracked checkout recovery
A retained declaration cannot restore a checkout while another folder or source link occupies its destination, and restoration requires a reachable source. State those prerequisites in each manifest translation.
* fix(config): prune members under recorded repositories
A migrated underscore-named member can otherwise survive a missing checkout as a repository declaration. Check recorded ancestor paths while preserving ambiguous grouped declarations and unrelated aliases.
* docs(manifest): restore fresh-clone project guidance
Explain that project declarations without local install records remain available so teammates can install them after cloning.
* refactor(reconcile): drop the tracked-intent walker policy
Restore baseline normalization for present paths while retaining absent
repository declarations through the shared metadata predicate. Isolate the
reconcile fixtures without building a CLI and share native-client overrides
without changing the original two-machine record and audit regression.
Validation: gofmt, git diff --check and context-router check only.
Runtime, builds and tests were not executed under the source-only gate.
* fix: forget linked checkout records and declarations without checkouts
Extend the metadata and uninstall owners to remove a declaration without
reading or moving its occupant, contacting its source, or changing trash
and gitignore. A failed metadata save makes every metadata-only result fail.
Clean up global unlink records in CLI and API, preserve existing-path name
precedence, report record-only results truthfully, and list declarations in
the dashboard API with noCheckout. Keep project removal behavior unchanged.
Remove the obsolete grouped-link walker-policy assertion while preserving
grouped absent declarations and the original record/audit A/B regression.
Validation: source-only five-pass review, gofmt, git diff --check and the
context-router check. No runtime, builds, tests or frontend tooling ran.
Stop at commit 15 for the mandatory Luna checkpoint before UI and docs work.
* fix(server): reload metadata before removing declarations
Preserve CLI records and audit acceptances written while the dashboard is open. Resolve new declarations from disk and refuse stale-cache saves after a reload failure, retaining unlink success when its link already moved.
* feat(dashboard): remove a tracked repository record without a checkout
Show declaration-only rows without checkout actions or status promises. Reuse Uninstall with record-specific confirmation and feedback, keep real-checkout force handling, and refresh the missing-entry list after repository removal or unlink.
* docs: explain kept declarations and how to remove them
Document direct global declaration removal and optional restore in English and all four translations, including occupied paths and normalization after reconcile. Keep built-in guidance accurate and use the existing API removal result for dashboard success feedback after concurrent CLI changes.
* fix(dashboard): confirm restored checkouts after forceable refusals
A tracked declaration can be restored through the CLI while its dashboard
row still says there is no checkout. Treat a forceable uninstall refusal
as evidence of the current checkout and promote the existing confirmation
to its dirty or unreadable-status warning before allowing explicit force.
Author stale-row regressions for both forceable codes before the source
repair, and retain nonforceable failure and actual-result success feedback.
Validation: five serial source review passes, git diff --check, and the
context-router check. Tests, builds, runtime and frontend tooling were not
run under the source-only gate.
* test(server): persist overview declarations before routed request
The overview fixture added declarations only to the in-memory skills store,
which the real request's config reload replaced with on-disk metadata. The
hosted unit, offline, and Nix jobs consequently returned no tracked rows.
Save the complete fixture through MetadataStore before serving the request
and fail setup on a save error. Keep the middleware, assertions, and
production behavior unchanged.
* fix(server): prefer exact declarations before checkout shorthand
Removing a missing _team declaration from the dashboard could resolve the
different org/_team checkout by basename and trash it instead.
Reload metadata before declaration lookup, give exact declarations priority,
and preserve checkout resolution before non-exact declaration fallback.
Add routed absent/occupied regressions and checkout precedence guards, and
persist existing checkout fixtures so the metadata reload keeps their state.
Validation: gofmt, git diff --check, and context-router check passed.
Tests were authored but not run under the source-only fleet restriction.
* fix: preserve removal intent and report partial failures
Bind dashboard record-only removal to its canonical declaration and refuse
restored checkouts before checkout resolution or force. Refresh the row and
require another confirmation before moving a restored checkout to trash.
Keep unresolved uninstall inputs and their reasons in JSON failed results,
including all-invalid requests, partial successes and dry runs. Report unlink
metadata cleanup failures as partial results, refresh completed unlinks and
show localized warnings. Keep CLI and dashboard guidance aligned.
Regression tests were authored before source changes and remain unrun under
the source-only task restriction. Static formatting, diff and router checks
passed; no build, product runtime, frontend tooling or test suite was run.
* fix(unlink): log partial metadata cleanup failures
Write the unlink result after global metadata cleanup so load and save
failures retain their diagnostic and report partial success. Preserve the
successful link removal, warning, target and exit policy, and the existing
link, project unlink and native removal failure behavior.
Add the corrupt-metadata regression before the fix; leave it unrun under
the source-only authorization. Formatting, diff and context-router checks
pass; runtime verification remains with the coordinator.
* fix(unlink): forget records below the removed link name
Reconcile can normalize a followed link's record to untracked, and linked
collections can retain nested repository declarations. Clean the removed
namespace through the existing metadata path owner in both unlink callers,
including audit acceptances, while keeping declaration uninstall strict.
Author CLI and route regressions before the source fix for normalized and
nested records, sibling preservation, and subsequent bare install. Tests
remain unrun under the source-only task restriction; static checks only.
* fix(uninstall): refuse a restored checkout at execution
A declaration selected while absent can become a checkout before removal
runs. Recheck the resolved source-relative name in the shared metadata-only
execution owner and refuse before touching metadata, regardless of force.
Reuse the existing dashboard conflict response and refresh/reselection flow
for this refusal, preserving the earlier record-only intent guard. CLI
results already carry shared execution failures through text and JSON.
Author restoration, force, absent/occupied control and client boundary
regressions before production edits. Correct the unlink fixture name and
response-key assertion found in serial static review. All tests remain
unrun under the source-only restriction; formatting/diff/router checks pass.
* refactor(server): consolidate repository uninstall result handling
Resolve a declaration or checkout to one item and canonical name before
using the existing uninstall, conflict, log and success response tail.
Keep metadata-only removal unforced, preserve its error and cache recovery
behavior, and report movedToTrash from the selected item.
Preserve exact declaration precedence, record-only checkout refusals,
ordinary checkout errors and project behavior without changing tests.
Five serial source review passes, gofmt, diff checks and the context-router
check completed. Tests and builds remain unrun under the source-only brief.
* fix(unlink): preserve independent records and trash undo metadata
Limit unlink's legacy group matching to members so a retained declaration
such as team/_repo survives unlinking _team. Share snapshot and removal
matching while keeping generic uninstall's legacy contract unchanged.
Save the exact removed records, audit acceptances and target overrides
through uninstall's existing trash sidecar owner before pruning metadata.
Keep metadata on sidecar failure and preserve completed-unlink warnings,
fresh reloads, cache refresh and project-mode behavior.
Author owner, CLI and API regressions for namespace collisions, undo,
partial failures and project/empty controls. Source-only verification:
gofmt, diff check and context-router check passed; runtime tests, builds,
frontend checks and final make check remain unrun for root-owned gates.
* fix(install): preserve independent tracked declarations
Limit declaration-only cleanup to the precise namespace and legacy members already used by unlink. Keep independently installed prefixed repositories, their audit acceptances, and overrides when forgetting a missing checkout.
* fix(cli): explain missing tracked repository recovery by mode
Offer global record removal when restoring an unreachable source is unwanted. Keep project recovery scoped to install -p, share the hint with doctor, and align update documentation in every existing locale.
* refactor(uninstall): share declaration and unlink operations
Keep declaration resolution and unlink metadata ordering in shared owners so CLI and API preserve the same scope and recovery behavior. Consolidate domain regressions while retaining caller translation coverage.
Distinguish record-only removals in API and CLI operation logs and CLI JSON, including mixed batches and dry runs. Document outcome fields, warning semantics, and current API error codes across existing locales.
---------
Co-authored-by: Claude Opus 5.5 <noreply@anthropic.com>
Co-authored-by: Willie <egg358358@gmail.com>
* fix(doctor): stop the plugin check from holding back later checks
The dashboard's Doctor page streams checks as doctor finds them, but the
plugin check asks each Agent's native CLI what is installed and can take
many seconds. Checks run in order, so backups, trash and the version
checks waited behind it while the page showed only an unlabeled skeleton
row, with nothing saying which check was still running.
When doctor streams, the plugin check now runs in the background. It is
streamed first as a "pending" row in its place, the checks after it
stream as soon as they finish, and its result follows. The full result
keeps the check in its original place. Text output and plain --json run
it inline as before.
The dashboard shows a pending row as the check's name with a spinner and
"Checking", replaces it in place when the result arrives, keeps it
visible under every filter, and drops the unlabeled skeleton while it is
shown. The /api/doctor/stream reference documents the pending row.
* fix(doctor): show the skill scan and Git status as running
Doctor scans the skill source before it streams its first check, and
runs git status on the source inline. On a large source either can take
seconds, and the dashboard showed only an unlabeled skeleton meanwhile.
A streamed doctor now sends a pending "source" row before the scan, and
runs the Git check in the background like the plugin check, so the
checks after it no longer wait. A finished background check is recorded
at the next check instead of at the end, so its row does not stay
pending after its result is known. The pending row now holds the
check's place in the result while it runs.
* fix(install): report reinstalls from the same repo as skipped
checkExistingConflict read only the legacy .skillshare-meta.json
sidecar, which installs no longer write, so reinstalling a skill from
the same repo was reported as "already exists" and counted as failed,
contrary to the documented skip. It now reads the skill's record in
.metadata.json, falling back to the sidecar, and skips only the same
skill of that repo: another skill installed under the same name stays
a conflict instead of being reported as installed.
The dashboard logged an existing agent as installed and a same-repo
skill as failed. Batch installs now return those as skipped and log
them as skipped_skills, passing --into so a reinstall into a group
folder finds its record; the install dialog warns that they are
already installed. A same-repo tracked reinstall returns 409 and is
logged as skipped instead of a 500 logged as failed.
ss search installs, batch and single, log the names they skip, and
the Herdr panel reports skipped names like failed ones. The install
docs and their translations say that only the same skill is skipped.
* fix(install): install requested agents when the only skill is skipped
Reinstalling a single-skill repo with `-s alpha -a helper` returned as
soon as the skill was skipped as already installed, so the requested
agent was never installed while the command reported success. Both
single-skill branches now continue to the agent phase, as the
multi-skill path already does.
* fix(install): require explicit force for JSON overwrites
Machine-readable output must not grant destructive consent. Preserve existing skills and agents unless force is explicit, and report existing skip decisions in the JSON result while retaining selection and audit behavior.
Cover edited copies in global and project modes, and update help, automation guidance, and translated command documentation.
Refs: #877
* fix(install): preserve previews and locked JSON copies
Read-only discovery must work before replacement consent, including existing
single skills and cross-path installs. Keep overwrite checks on write paths.
Require existing update or force intent to advance installed project copies
to a changed lock pin. Preserve ordinary CLI and dashboard lock following,
while JSON alone leaves local edits and incoming pins unchanged.
Cover existing-destination previews, edited locked copies and dashboard lock
following. Clarify translated docs and make the consent runbook step runnable.
Refs: #877
* fix(install): resolve installed names during previews
JSON previews must resolve recorded sources without requiring overwrite consent. Permit metadata lookup for dry runs and restore the original unforced preview regression in global and project modes.
* test(install): remove duplicate JSON consent runbook case
The global local-skill integration case already covers JSON overwrite refusal and explicit forced replacement. Remove the duplicate runbook step and restore its 18-step numbering while retaining the existing purity setup and assertion corrections.
* fix(audit): report audit-rules load errors in global scans
The global scan entry points (ScanSkill, ScanSkillWithFollow,
ScanSkillFiltered, ScanFile) dropped the error from Rules() and scanned
with no pattern rules. One bad regex in the global audit-rules.yaml
turned every pattern check off with no warning, so audits reported
fewer findings than they should have (#879). Project mode already
returned the error.
Return the error from the global path as well, and name the failing
file in it: parse errors from loadUserRules and compile errors from
Rules() and RulesWithProject now carry the path. When a project scan
fails to compile, the global file is blamed only if it fails on its
own, so a project file that disables a broken global rule still loads.
The cached markdown split now takes the rules its caller already
loaded instead of loading them again and dropping the error.
* fix(install): block installs whose audit cannot run
Once global scans return rule-load errors, install treated the error as
a warning and installed the skill with no audit at all. update already
fails closed on a scan error, and --force there accepts findings, not
an install nobody could check.
A scan error now blocks skill, agent and tracked-repo installs and
removes what was installed. --force does not override it;
--skip-audit still bypasses scanning. The error wraps the new
install.ErrAuditScanFailed rather than audit.ErrBlocked, so install
output and the dashboard do not offer a --force retry that cannot help.
upgrade reinstalls the built-in skill with --force, which deleted the
old copy before the audit ran. A typo in audit-rules.yaml would have
left users without the skill, so upgrade now stages the download
through the update path and keeps the installed copy when the audit
fails.
update --all counted the new error as a silent skip. It is now listed
under "Blocked by security audit" with its reason and makes the run
exit non-zero. The install warning summary that counted "audit scan
error" warnings is removed, since nothing emits them any more.
Fixes#879
* fix(plugin): recover OMP stale catalog installations
Refresh owned marketplace catalogs before reuse and validate the native cache identity when removing an installation. Preserve that identity through interrupted cleanup and report blocked batch updates as failures so retries remain safe and visible.
* fix(plugin): preserve reviewed versions during OMP recovery
Keep native removal identity separate from reviewed receipts so interrupted cleanup can retry safely. Expose isolated failures in previews and dry-run exit status while allowing healthy targets to proceed.
* fix(plugin): resume interrupted OMP update cleanup
Retry matching pending removals before reinstalling after registry-first interruption, and clarify isolated failure reporting. Strengthen recovery fixtures and restore accurate installation guidance.
* test(plugin): model OMP cached catalog installs
Use the cached marketplace entry for native install version and source so the stale catalog regression exercises the real version-to-cache relationship.
* test(plugin): align OMP batch expectations with safety checks
Require catalog refresh before reinstalling from an owned marketplace and assert isolated failures when native OMP ownership cannot be verified, while healthy targets update.
* fix(plugin): skip native upgrade checks for OMP imports
Imported plugins have no reviewed source to upgrade, so requiring native upgrade capability makes a safe batch skip fail unnecessarily. Cover that case and OMP dry-run failures, and align translations and native verification runbook with the failure and removal behavior.
---------
Co-authored-by: Willie <egg358358@gmail.com>
* feat(audit): honor findings accepted with --force via --respect-accepted
install and update --force record the findings a user accepted in
audit_accepted, but audit ignored that record. A skill installed with
--force therefore kept failing `skillshare audit` in CI with no way to
let the accepted findings pass (#869).
audit --respect-accepted marks those findings as acknowledged. They stay
in every output (JSON acknowledged: true, a text/TUI note, "(accepted
earlier)" in Markdown, SARIF external suppressions), but they are counted
in summary.accepted instead of their severity and no longer fail the
audit. Without the flag nothing changes, so existing CI gates keep their
behavior. A changed file or snippet produces a new key and blocks again.
Tracked repositories record acceptances at the repo root with paths
relative to it, so ApplyAcceptedFindings walks up from each skill to
match them. The install, update and edit gates use the same function and
now honor repo-wide acceptances for skills inside tracked repos too.
The dashboard always honors accepted findings: the Audit page mutes them
and counts them apart, the skill Audit tab collapses them, and the
Dashboard attention card no longer lists a skill whose only blocking
findings were accepted. The CLI JSON summary and the server summary share
Result.CountOpenBySeverity so both count accepted findings the same way.
The oplog records the accepted count and the log views show it.
* fix(audit): match acceptances when a single file is scanned
A single-file scan names its findings by the file's base name, but the
ancestor lookup used the file's whole path as the prefix, so it hashed
SKILL.md/SKILL.md instead of the recorded SKILL.md. Auditing a skill file
directly with --respect-accepted, and auditing agents inside a tracked
repository in the CLI or the dashboard, still failed on unchanged
accepted findings. For a file target the prefix is now the file's own
directory.
* fix(ui): show health checks as they finish instead of after the slowest
The Doctor page waited for the whole `doctor --json` run before showing
anything. One slow check, such as plugin checks waiting on a native CLI
that does not answer, kept the page on a skeleton for 30 seconds, and the
server's 30-second write timeout could cut the response off entirely.
`doctor --json` now prints each check as a {"check": ...} line when the
internal _SKILLSHARE_DOCTOR_STREAM variable is set, before its usual
result. The new GET /api/doctor/stream relays those lines as Server-Sent
Events and sends the full result as `done`; it has no write deadline and
stops doctor when the client leaves. The page shows checks as they
arrive and replaces them with the final result.
* fix(ui): show the server's reason when a stream fails
The audit, diff and doctor streams send an `error` event with the reason,
but the shared EventSource helper treated it like a dropped connection
and showed only a generic message. It now uses the reason the server
sent, and keeps the generic message for a real disconnect.
* fix(doctor): skip Git reads in the hooks check when no hooks are configured
The hooks check loaded the full hooks inventory, which also scans
unmanaged native and Git hooks twice: about 12 git calls even with no
hooks configured, none of which doctor reports. Doctor now reads the
hooks section first and previews a sync only when there is something to
check.
Point every README release notice at check --exit-code and check --all, the headline CI changes in 0.30.6, and credit tomerzcod7 (#852) and ErykJakubik (#858), whose reports led to the git hook and project config fixes.
The security change in #861 added its provider-token paragraph twice to the install command page in all five locales; keep one copy.
The website builds for all five locales and context validation passes.
check <name> --json printed its warnings and the group expansion note on stdout before the JSON, so the output could not be parsed. When no name or group resolved it then printed an empty result and exited 0, while text mode exits 1, so a CI gate with a mistyped name passed.
With --json, check now exits 1 with an error object when nothing resolves, lists skipped names in a warnings field when some do, and reports every other failure (bad flag combination, missing config) as the same error object. Text mode is unchanged.
check --all is documented as checking both skills and agents, but it ran the skills check only. With --exit-code now available, a CI job using check --all --exit-code passed while an agent had drifted.
--all now runs both checks: the text output has a Skills and an Agents section, --json adds an agents array beside tracked_repos and skills, and --exit-code counts both. check --json and check agents --json keep their shapes.
--all is rejected together with names or --group. It used to be ignored there, and agents have no name filter, so honoring it would check one skill and every agent.
check exits 1 only when an item could not be checked, so a CI job that wants to fail on pending updates or local edits had to parse --json.
--exit-code makes a check that ran cleanly exit 1 when any skill, tracked repository or agent is not up to date. Only items whose status is local are left out, and the default exit status is unchanged. The operation log still records such a run as successful. The flag is rejected with --diff.
Bind provider credentials to canonical or explicitly configured HTTPS origins and preserve that policy across remote consumers. Reject cross-site credentialed reads, linked fallback directories, and discovered-agent writes outside the source; audit staged content before rooted atomic publication.
Add boundary regressions, private Git update coverage, stable cross-platform fixtures, and aligned documentation. Verified the integrated main baseline with all executed CI checks passing.
In project mode the committed config.yaml is what teammates install from, but
it could disagree with what a machine actually has (#858):
- `uninstall -p` and the dashboard's single-skill and tracked-repo removal
left the skill's entry in config.yaml, so a later `install -p` brought it
back. They now drop the entry and the lock pin through one shared
config.ForgetProjectSkills, as the uninstall docs already promised.
- Installing a skill again from another source or branch (after an uninstall,
or over the copy with --force) updated only the gitignored .metadata.json.
config.yaml kept the old source and skills.lock.json ended with the old
source and the new commit. Installs from a named source (CLI install,
search, dashboard) now compare the records from before and after the
command and let the config entry follow the skills it actually reinstalled.
Install from config never rewrites an entry from local records, so a
teammate's edit survives on a machine that has not updated yet.
- Reconcile removed every config entry that had no folder and no record. On a
fresh clone nothing is installed yet, so `install <source> -p` deleted all
the other declared skills. An entry now goes only when the skill was
installed here and its folder is gone, when it was moved, or when a command
removed it; `move` passes the old paths and `update --prune` cleans the
config itself.
Refresh all README release notices and route agents to the new updates command. Clarify that desktop policy controls require a compatible App release; the CLI upgrade alone cannot change older desktop builds.
The website builds for all five locales, context validation and skill validation pass. The documentation diff passed five local review passes.
Add a global automatic-update preference with an environment override, CLI controls and dashboard settings. Suppress automatic release checks and cached notices while retaining explicit checks and upgrades.
Recheck policy before publishing in-flight results, preserve structured JSON errors, and reject incomplete manual checks. Include schema, completion, localization, documentation and regression coverage.
Closes#846.
The docs still showed the old `{"skills": [{"name", ...}]}` layout, which
no current release writes; the file is `{"version", "entries", ...}`
keyed by path. Readers comparing the docs with their own file had no way
to tell what `audit_accepted` or `target_overrides` are for.
tracked-repositories also said project mode records `tracked: true` in
`.skillshare/.metadata.json` so teammates get it, but that path does not
exist and project mode gitignores the metadata files (#760); the flag
lives in `.skillshare/config.yaml`.
The example `type` is `github-subdir` because a subdir install records
that, and the field table omits `into`, which nothing writes.
Expose selected skill names and repository-relative paths in dry-run JSON so the panel follows CLI discovery, ignore rules, and canonical names. Refuse ambiguous plans and treat nonzero install exits as failures even with legacy JSON output.
Closes#838
Release-As: 0.30.4
* fix(plugin): let OMP and Pi plugin updates through without uninstalling (#833)
Every OMP update was blocked, because OMP's native upgrade replaces cache
folders that projects outside its inventory may still use. The only way to
move an OMP binding to a new commit was to uninstall first. An update now
runs the same pair a source change already uses: the scoped removal
adapter (cache kept), then an install under a fresh marketplace and cache
identity. A refresh always takes a new identity, because OMP installs from
the catalog it cached when the marketplace was added, and that catalog
still names the old version. If the removal cannot be verified, the
update is blocked; update --all still skips it with the reason.
Pi was blocked whenever the catalog plugin name differed from its
package.json name. That happens after an upstream npm rename, for example
when npm refused the original name. Pi installs by path, and Skillshare
keys the binding by the catalog name, so the npm name never reaches
either, and the check only stopped working installs.
Updating one package across several Agents stopped every Agent as soon as
one was blocked. Each failing Agent is now reported with its reason while
the others update, and the command exits non-zero. A single --target, or
a package with no other Agent to update, stays blocked as before, so a
dry-run still fails when nothing would update.
* fix(plugin): let one update --all refresh several OMP packages
Every OMP removal proves itself against a fingerprint of the whole native
registry and lock. In a single update --all plan, the first package's
removal and reinstall rewrite those shared files. The second package's
removal therefore saw "OMP native files changed", so only one OMP
package could update per run. The same applied to a sync that replaced
the sources of several OMP packages.
A removal now also fingerprints only its own installation: its registry
entry, its lock entry, its package manifest and its runtime link. After
an earlier change in the same apply has written to that OMP target, the
removal re-reads the files and runs every ownership check again. It goes
ahead only if that own fingerprint still matches the preview. A change
outside Skillshare to the package's own entries still blocks it. Without
an earlier write in the apply, the full-file check is unchanged.
The fake OMP in the tests now merges installs into the registry and lock
and lists them in a stable order, as the real one keeps other plugins.
Scripts and the herdr panel could not tell why an uninstall or install
failed: uninstall --json had no error field, project mode ignored --json
(printing text and still prompting), and a multi-skill or config install
in which every skill was blocked exited 0.
- uninstall --json (skills and agents, global and -p) sets "error" with
each failed name and reason; skills exit non-zero when all fail,
agents when any fails. Project mode now emits pure JSON, including on
the first run that initializes the project.
- uninstall -p agents parses its flags instead of treating --help or an
unknown flag as an agent name, and uninstall agents --json is logged.
- install exits non-zero, naming each skill and reason, when every skill
from a source or from config fails; installed or skipped skills make it
a partial success that still exits 0. Config installs show each
failure's reason, and the dashboard logs an all-failed run as an error.
- herdr asks for JSON in project scope and shows the CLI's error.
install --json printed empty skills and failed lists with exit code 1
when the whole install failed, for example when the security audit
blocked a tracked repository or a single skill. Scripts and the herdr
mod could see that it failed but not why, and the mod already reads an
error field that was never set.
The install JSON now carries the command error in an omitempty error
field, in global and project mode. The install docs also stop saying
--json implies --force: it only permits overwriting, and the audit gate
still applies unless --force is given.
Tracked install and update audit the whole repository root, and the
scanner treated every file below a hidden directory as header-only. Skills
that discovery finds in hidden directories, such as .curated/<skill> and,
since #801, a repository-root .agents/skills/<skill>, were therefore synced
to every target without a text scan.
The hidden-directory boundary is now measured inside the skill that holds
a file, so those skills get the text scan they already get when audited one
by one, while target copies such as .claude/skills and hidden folders
inside a skill stay header-only.
Release-As: 0.30.3
A repository-root .agents/skills now counts as a skill anchor only when the repository has no other SKILL.md outside .git and target directories, so tracked checkouts that author skills there are discovered, while repos that commit .agents/skills copies keep discovering only their authored skills.
Fixes#800
Share portable MCP definitions with Hermes through its native YAML configuration while preserving unrelated settings and managed-entry safeguards.
Support explicit Hermes homes and native OAuth options without copying credential state, and expose the target consistently in the dashboard, schemas, completions and documentation.
Refs #825
Credential-access rules match whole lines, so a Python docstring that
named ~/.aws/credentials in a `cat` example counted as a CRITICAL read and
blocked the mcp-sentinel skill at the default threshold (#817).
In .py files, a built-in CRITICAL credential-access match on a comment-only
line, or inside a string that is a whole statement such as a docstring, is
now HIGH: it still shows, and still blocks under a HIGH threshold. A small
tokenizer decides this and keeps CRITICAL wherever it cannot be sure the
line never runs: argument, assigned and bracketed strings, f-strings and
their replacement fields, code before a trailing comment, files that read
__doc__ or use bare CR line breaks, and explicit severity overrides.
A lowered finding carries severityContext "python-inert" (JSON, SARIF,
`audit rules --format json`), and the dashboard, CLI text output, both
audit TUIs, install/update warnings and the markdown report say it was
lowered from CRITICAL, so it is not mistaken for a changed rule.
Closes#817
The Hubs page keeps a hub's skill list for two minutes and does not
refetch on window focus. The server reads the index fresh on every
request (git hubs are re-cloned), so after someone pushes an updated
skillshare-hub.json the only way to see it was to wait or reload the
whole page.
A refresh icon button next to the filter refetches the list on demand.
It spins and is disabled while loading; the current list stays on screen,
and a failed fetch shows the existing error note above it. Your own
(draft) Hubs are local and do not get the button.
* perf(plugin): check every plugin at once and skip sources that did not move
The Plugins page's update check read every Agent CLI, cloned and hashed
every source, and asked npm about every package one after another, so it
grew with each plugin added. A single Agent CLI that stopped answering
held the whole page at "Checking for updates" with no way to tell which
plugin was stuck.
The check now asks every Agent and checks every plugin at the same time.
Plugins share one read of each Agent and each source, so an Agent with
many plugins is still asked once, and a plugin waits only on its own
Agents, including when proving its own leftover installs. Before downloading a remote source with no pinned ref, the
check asks the remote which commit its default branch points at with
`git ls-remote` and reports no update when that is the installed commit.
A check of three unchanged GitHub plugins on two Agents went from 4.3s
with three clones to 0.6s with none.
The dashboard reads the check from a new stream, GET
/api/plugins/check/stream, that sends each plugin as soon as it is
checked. The page marks rows as they arrive and counts progress in the
header. The stream stops waiting after 60 seconds: a plugin bound to an
Agent that has not answered by then shows "Check failed" with the
reason, and plugins on other Agents are unaffected. There is no limit on
how many plugins run together, because plugins waiting on a hung Agent
would otherwise hold every slot and starve the rest.
A cancelled command did not return when the Agent CLI was a shim that
starts the real client, such as an npm, mise or volta wrapper: killing
the shim left the child holding its output, so the command waited until
the child exited and no timeout took effect. Commands now stop waiting
for output two seconds after they are cancelled; a command that exits on
its own while a helper keeps its output open still counts as answered.
* fix(ui): offer every update the Plugins page shows in the update dialog
The header counted updates from the per-plugin marks, but the update
dialog offered only the targets in the last completed check's plan. Two
cases left the header saying Update N with nothing to tick in the dialog:
a first check that broke off after finding an update never set the plan,
and a recheck that could not reach an Agent kept that Agent's earlier
update while the plan said blocked.
The dialog's plan is now derived from each plugin's last answer together
with the kept updates, so a kept update stands in for the blocked or
missing answer and both always agree. A broken-off check also keeps the
answers it did receive.
DNS rebinding lets a page point its own domain at the user's machine,
which makes its requests same-origin to the dashboard, so the CSRF
guard cannot refuse them. Only the MCP, Hooks and Plugins routes
required a localhost or IP Host, so such a page could still install,
sync and push. That rule also blocked those three pages for people who
run the dashboard behind their own reverse proxy (#809).
The dashboard now answers every route only at localhost, an IP address,
the --host name, and origins given with --trusted-origin or
SKILLSHARE_UI_TRUSTED_ORIGINS. Configuration stays outside config.yaml
so the dashboard cannot widen its own trust through its config API. A
trusted origin must match the Host with its scheme, and an Origin
header, when sent, must be that origin. A refused browser gets a page
with the exact setting to copy, the API gets it in the error, and the
server log names each refused Host once.
BREAKING CHANGE: a dashboard opened by a domain name, such as a reverse
proxy, a Tailscale or mDNS name, or a Docker service name, now returns
403 on every page until that address is trusted, for example with
`skillshare ui start --trusted-origin https://skills.example.com`.
Release-As: 0.30.0
A web page can submit a form or a no-cors fetch to the local dashboard
without a CORS preflight, and the handlers accept a text/plain or empty
body. Only the MCP, Hooks and Plugins routes checked where a request
came from, so any site the user visited could add targets, install,
sync, commit, push or upgrade, whenever the browser let the request
reach localhost.
Every POST, PUT, PATCH and DELETE now needs Fetch Metadata, when sent,
to say same-origin, and Origin, when sent, to match the Host. A client
that sends neither, such as curl, is unchanged, because a page cannot
make a browser omit both on a cross-origin request. The update stream
runs updates over GET, as EventSource only opens GET, so it gets the
same check.
The nginx example now forwards $http_host, which keeps the port, since
a proxy that drops it would now fail every change, not only MCP.
The Hooks page uses the same DNS rebinding guard as the MCP page
(requireLocalHooks), so through a domain name or a reverse proxy its
requests return 403. The reverse proxy note only mentioned MCP, so a
user behind a proxy had no warning that Hooks would fail too.
The Claude manifest fallback from #806 offered Codex more than it can
install:
- A Claude package whose only parts are ones Codex does not read, such
as agents, was offered to Codex with no components, so Codex would
install an empty plugin.
- When Codex's own catalog pointed at a folder it rejects, the whole
candidate could be replaced by the Claude catalog's folder, which now
carries a Codex fallback. Codex reads only its own catalog, so it
would never install from that folder.
#806 also stopped every Agent from using a later catalog's folder when
the first one was rejected. Only Codex needs that, since only Codex
reads one catalog exclusively, so other Agents get the old behaviour
back.
Codex accepts Claude plugin manifests natively, but discovery omitted
Codex for packages such as Matt Pocock's skills. Select that supported
fallback without rewriting package resources or invocation policies.
Match native root and manifest precedence, retain rejected own-catalog
packages, and add discovery and snapshot regression coverage.
Refs #805
The dashboard reference walked through each Targets page control but
not the search box #799 added above the include/exclude preview, which
is also shared by project pages and the Extras Edit target dialog.
Release-As: 0.29.2
A named pipe can block collection, integrity hashing or uncached skill
metadata reads indefinitely. Require resolved regular files at each read
boundary while preserving symlink accounting and integrity evidence.
Cover the paths with bounded FIFO regressions and generated filesystem
cases, and update the Nix vendor hash and integrity-rule documentation.
* fix(projects): save new projects in config.yaml again
v0.29.0 wrote every project add, edit and remove to a machine-local
registry, so a project added on one machine never reached config.yaml
or the other machines that share it (#792). Most people share one
config.yaml and expect a new project to follow it.
config.yaml is the default again. Machine-local storage stays as an
explicit choice: --machine-local on add, or projects_storage:
machine-local for every new project, with --shared to override it.
The new share command moves a local project into config.yaml, and
localize still copies one to this machine. Edits and removals stay in
the layer that owns the project; removing a local copy of a
config.yaml project keeps config.yaml's version unless --shared asks
to remove both, so one machine cannot silently drop a project for
the others. Exclusions written by v0.29.0 are left as they are.
The dashboard picks storage once, in the Add project dialog, and
offers the two removal choices only where both copies exist. Moving a
project between layers stays a CLI command, so the detail page keeps
one job.
Closes#792
* fix(projects): recheck both layers after a confirmation, sync the saved key
remove and share ask before taking the config lock, then compared only
the merged entry after locking. A share or localize from another
terminal moves a project between layers without changing that entry, so
a remove confirmed for this machine's copy could then delete config.yaml's
declaration for every machine, and a confirmed share could overwrite
config.yaml settings changed meanwhile. The recheck now compares the
owning layer and config.yaml's declaration too (ProjectLayerState).
Re-adding a project that v0.29.0 hid restores config.yaml's own key,
which can be spelled differently from the folder given to add (~/app
versus an absolute path). add --sync looked the project up by the given
spelling and failed with "project not found" after saving it; it now
syncs the key that was saved.
* test(projects): make the layer tests hold on Windows
The mutation-sequence test looked the shared project up by
filepath.Join, but the fixture spells config.yaml's key $ROOT/app;
on Windows the two spellings differ, so the lookup found nothing.
It now asks the config for the key it uses.
The remove save-failure test blocked the local registry, but removing a
config.yaml project now writes config.yaml. On Linux the registry check
still failed with ENOTDIR first, which hid the mismatch; on Windows that
reads as a missing file and the save went through. The test now blocks
config.yaml, as the edit save-failure test does, and its hook no longer
shadows the config path with the target path.
---------
Co-authored-by: sim <sim@example.invalid>
A skill installed from a non-default branch of an SSH or other-host
remote (git@host:o/r.git, Gitea, self-hosted HTTPS) could not switch
back to main: the Branch picker was greyed out with "This source can
only use its default branch" (#787). The picker reused /api/hub/refs,
which only lists refs for GitHub, GitLab and Bitbucket web URLs because
a hub entry pins its ref in the URL. Reinstalling from a branch does
not need that; the update endpoint already accepts any git source.
/api/hub/refs now takes anyGit, which lists a non-pinnable git source's
refs while still reporting pinnable=false. The detail page asks for it
under its own query key; the hub editors keep the old request, so they
run no extra ls-remote for sources they cannot pin.
Closes#787
Co-authored-by: sim <sim@example.invalid>
Checking every user-visible change since v0.28.4 against the docs
found these gaps:
- ui.md: GET /api/audit now returns warnings and incomplete; an
incomplete scan warns instead of reporting all clear and appears
under Needs attention; a project branch switch records the branch in
.skillshare/config.yaml. Also fixes a lowercase sentence start.
- audit-engine.md: Markdown and SARIF carry the per-resource coverage
counts added in #756.
- audit-rules.md: the files row pointed at a "File Scoping" section
that does not exist; it now links to File Applicability.
- log.md: project localize logs a localize action.
- built-in skill: how an unreadable followed source link is reported.
Translations are updated to match.
* feat(audit): show unfollowed source links in Markdown and the TUI
When a followed source link could not be read, only text output (and
JSON's warnings/incomplete fields) said so. A Markdown report looked
like a full scan, and the TUI covered the warnings printed before it
opened, so a partial scan read as complete there. Item 25 of #746 asks
for incomplete status to stay visible in human reports, including
under --min-severity.
Markdown now ends its summary with a "Source links not followed"
section, led by an "Incomplete scan" line when a link could not be
read. The TUI summary line starts with the count in the warning
colour. JSON and SARIF are unchanged.
* feat(ui): mark a partial audit scan on the dashboard
The audit API dropped the source-link warnings the scan collected, so
the dashboard showed a scan that skipped an unreadable followed link as
complete. Item 25 of #746 asks for incomplete status to stay visible in
every human report, the dashboard included.
GET /api/audit and the stream's done event now carry warnings and
incomplete with the same meaning as `audit --format json`, and the
Audit page shows the count beside the Skills/Agents tabs.
* feat(ui): stop calling an incomplete audit scan all clear
After a scan that skipped an unreadable followed source link, the
Audit page still toasted "no dangerous content" and the Dashboard's
Quick Scan toasted "All clear", while the home page showed nothing,
so a partial scan still read as a clean one there.
Both toasts now warn that the scan is incomplete, and the home page's
"Needs attention" list gets a row linking to the Audit page. One
helper words the note for the Audit page label, both toasts and the
row.
---------
Co-authored-by: sim <sim@example.invalid>
With --min-severity set, the TUI dropped every result without visible
findings, so clean skills disappeared from the list and the tab count
shrank, while text and Markdown output kept them. The rule now lives in
audit.HiddenByFloor and both the TUI and text output use it: a floor
hides only a result whose findings all fall below it and that is not
blocked.
The dashboard severity buttons label themselves with the English
severity, matching the badge they show, so the six translated label
keys added for them are gone. AuditPage.test.tsx again asserts the
exact coverage line and count instead of any three digits, and drops a
rendered test that only checked translated labels. The audit-engine
paragraph about --min-severity now appears in all four translations.
Co-authored-by: sim <sim@example.invalid>
Reduce audit report noise without changing full findings, risk scores, severity gates or machine output. Reuse existing severity ordering and dashboard selection, and keep visible/hidden counts and incomplete-scan status explicit in human reports.
Co-authored-by: GPT-6.1 Sol <noreply@openai.com>
Co-authored-by: Willie <egg358358@gmail.com>
* feat(audit): inspect bundled hooks and settings
Bundled declarations can schedule commands and broaden permissions outside SKILL.md, while ordinary scanning skips hidden settings. Inspect the four explicit paths with bounded root-confined reads and structured findings so existing audit policy can review their potential execution.
Keep activation conditional and redact arbitrary arguments before reports. Register rule metadata through the existing structure analyzer and cover benign twins, parse gaps, source links, policy overrides, CLI/API output and a synthetic E2E runbook.
Co-Authored-By: GPT-6.1 Sol <noreply@openai.com>
* refactor(audit): parse bundled hook commands with a shell parser
The hand-written lexer for hook commands grew six similar quote and
paren scanners, and each review round found another shell form it
missed (comments after line continuations, nested backticks, `time -p`).
Parsing with mvdan.cc/sh/v3/syntax follows quoting, comments,
substitutions, loops, functions and heredocs the way bash does, so the
analyzer walks every call in the syntax tree instead. The wrapper table
(sudo, env, timeout, xargs, bash -c, git) and the POST and destination
rules are unchanged, now applied to unquoted argv words. A script fed to
a shell on stdin (`bash <<'EOF'`, `bash <<<`) is parsed like `bash -c`.
Statements before a syntax error are still inspected. bash runs some
forms the parser rejects (an array element assigned before a command),
so a rejected command also gets a word-level network check next to its
coverage gap, rather than dropping to MEDIUM.
The dependency adds about 0.26 MB to the release binary; vendorHash
follows the new go.sum.
* fix(audit): bound nested shell parsing and skip stdin data
A script nested in a substitution was inspected twice, once by reparsing
the outer shell's argument and once by walking the outer syntax tree, so
each `bash <<< "$(...)"` or `bash -c "$(...)"` level doubled the work and
a hook under 1 KB could stall an audit. Results are now remembered per
script text and depth, and one hook command may parse at most 1 MiB of
source in total, since deep nesting still costs quadratic time with the
cache alone. Past the budget the command is classified word by word and
gets a coverage gap, as a rejected command does.
Stdin was also treated as a script whenever a shell had a heredoc or
here-string, but `bash -c 'cat' <<<` or `bash ./script.sh <<<` hands it
to something else as data, which turned a lifecycle hook into a false
HIGH. Stdin is now parsed only for `-s`, or when there is neither `-c`
nor a script file operand. Option values (`-euo pipefail`, `-O extglob`,
`--rcfile f`) are not mistaken for a script file.
---------
Co-authored-by: GPT-6.1 Sol <noreply@openai.com>
Co-authored-by: sim <sim@example.invalid>
* feat(ui): show a skill's branch and switch it from the detail page
A skill installed with --branch looked the same in the dashboard as one
from the default branch, and moving it to another branch meant a manual
reinstall from the CLI (#783).
The list, tiles, tree pane and Updates tab now tag the branch a skill was
installed from. The detail page lists the source's branches; picking one
confirms, then reinstalls through POST /api/update with a new `branch`
field, so the audit gate and the local-changes check still apply and a
branch without the skill leaves the install in place.
In project mode, reconciling config.yaml only added missing skills, so a
reinstall from another branch (the switch, or install --branch --force)
left config.yaml naming the old branch while the lock pinned the new
commit. Reconcile now moves the declared branch with the record.
Closes#783
* fix(update): make branch switches install the requested branch safely
Codex review of #784 found four ways a switch could misreport or lose work:
- A root skill installed as a full clone took the pull fast path on later
switches, pulling its current branch instead of the requested one. The fast
path now applies only when the checkout is already on that branch.
- A branch whose skill folder lacks SKILL.md replaced the skill with a plain
folder. A staged update now refuses a candidate without SKILL.md when the
installed copy has one, and keeps the install.
- Reconciling config.yaml from local metadata could rewrite a teammate's
newer branch declaration back to a stale one on an unrelated install. The
generic reconcile change is reverted; the switch updates only the switched
skill's declaration.
- The config editor cache kept the old branch after a switch, so a later save
could undo it. A successful switch now invalidates the config query.
---------
Co-authored-by: sim <sim@example.invalid>
Keep machine-specific project changes out of shared configuration while preserving shared registrations and making their origin visible. Handle missing replicated roots and preserve startup maintenance for mutating commands.
Extension and size filters could hide native executables and Python bytecode
inside document, image and cache paths. Classify bounded headers before those
filters, retain source boundaries, and report opaque and oversized file counts
separately from auditable text.
Keep archive signatures and uncertain headers informational rather than claiming
runtime analysis. Honor analyzer selection for file and dashboard scans, and
cover renamed payloads and benign twins with synthetic fixtures and fuzzing.
Implements item 7 of #746.
Co-authored-by: GPT-6.1 Sol <noreply@openai.com>
Co-authored-by: sim <sim@example.invalid>
Existing URL and code-execution signatures miss cloud metadata access and
language-specific deserializers. Add scoped static evidence and inspect safe
YAML and torch options within each call so safe siblings cannot hide risks.
Validate and preserve file applicability across overrides, scans and rule
listings. Collect PHP scripts and propagate rule-load errors so malformed
filters cannot silently disable global scanning. Document the policy and
limits with paired regressions, fuzzing and an isolated CLI runbook.
Co-authored-by: GPT-6.1 Sol <noreply@openai.com>
Co-authored-by: sim <sim@example.invalid>
Co-authored-by: Willie <egg358358@gmail.com>
Required local references can currently escape a skill bundle or fail silently
when written as images and explicit code paths. Resolve them within the declared
root without reading external targets, retain dangling-link compatibility, and
share the findings and rule controls with doctor. Keep metadata and integrity
fallback reads within the same root so they cannot reopen skipped symlinks.
Co-authored-by: GPT-6.1 Sol <noreply@openai.com>
Co-authored-by: Willie <egg358358@gmail.com>
Long blank spans and repeated filler can move meaningful instructions out
of view during skill review. Add four padding signals under the static
analyzer with Markdown exemptions and original-source line mapping.
Reuse rule overlays so global and project toggles and severity/confidence
overrides also govern this file-level detector. Cover threshold boundaries,
benign examples, optimized scans, reports and CLI rule controls.
Implements item 17 of #746.
Co-authored-by: GPT-6.1 Sol <noreply@openai.com>
Co-authored-by: Willie <egg358358@gmail.com>
* fix(update): keep an agent's source when updating it
After the first `update agents`, an agent's recorded source became
`/agents/<name>.md` with type `unknown-subdir` and no repo URL, so
`check agents` stopped looking upstream and later changes never arrived
(#777).
The batch path rebuilt the repository source from the clone URL, subdir
and branch only, leaving Raw and Type empty, and the metadata written
back was built from that empty Raw. The fallback reinstall and the
dashboard update kept the agent's full source as Raw while clearing the
subdir, so the written source repeated the agent path.
AgentRepoSource now rebuilds the repository source once, by cutting the
agent path off the recorded source, and all three paths use it, so the
source written back is the one install recorded.
Agents whose metadata was already rewritten have no repository left in
it and need to be installed again.
* fix(check): compare each agent at its own branch and report unchecked agents
Agent update checks gave wrong answers in several cases next to #777:
- Agents from one repository were compared with the first agent's
version, so an older agent sharing a repo with a current one never
showed its update.
- The installed branch was never sent to the remote, so an agent pinned
to a non-default branch was compared with the default branch head.
- A tracked agent repo records its full commit while the remote hash is
abbreviated, so a fresh clone always read as update_available; its
checked-out branch was also ignored.
- `update agents` printed "All agents are up to date" and exited 0 when
agents could not be checked at all: an unreachable remote, unreadable
metadata, or a source erased by the #777 bug.
Each agent is now compared with its own version at its own branch, and
an agent whose source the old bug erased reports an error asking for a
reinstall, since its metadata no longer names a repository. Update lists
the agents it could not check and exits non-zero, counting them apart
from agents that failed to update. The docs describe the error status
and the reinstall fix.
* fix(update): repair dashboard-doubled agent sources and share the update flow
Review of the previous commits found gaps in recovering agents that
earlier updates broke, and in how the two update handlers report them:
- A dashboard update before the #777 fix appended the agent path to its
source a second time. Those agents keep their repo URL, so check offers
an update, but discovery then failed on the doubled path. Every copy of
the path is now cut, and the update writes the source back correctly.
- On the dashboard, an agent whose source the CLI bug erased failed with
"invalid source: source cannot be empty". AgentRepoSource now returns
the same reinstall error check shows, and both use one detector.
- When every tracked agent was such an erased agent, no remote was asked,
so the "Checking agents" spinner was never stopped and drew over the
error rows. It is now stopped once the check returns.
- The global and project handlers repeated the check-and-update flow,
and the project copy had already drifted (no dry-run log entry, spinner
in --json). Both now run one shared flow.
Tests cover the erased agent next to an updatable one in global, --json,
--dry-run and project mode, and the doubled and erased sources in
AgentRepoSource.