- Resolve the Windows admin config via the ProgramData known folder
instead of the user-controlled ProgramData variable, so a user cannot
redirect the admin lookup and drop pinned labels
- vendor rename rewrites matching agent-lib.labels.json entries, so a
renamed item keeps its labels instead of going to every client
- validate checks the labels file in repositories without a lockfile
- A kept modified copy of a deselected item is no longer drift in status,
and the toast explains it as "no longer assigned to you"
- Labels parser rejects trailing data such as an extra closing brace
- Tests: admin pin wiring (config chain, --repo, admin as chosen config),
reselect redeploy with byte-level no-op, status labels JSON, validate
without lockfile, rename keeps labels
- README: document unreadable admin config, drift semantics, rename
Refs: agent_lib-fh4
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
- Add internal/labels: parse and validate the optional work-repo file
agent-lib.labels.json (version 1; unlabeled items go to every client;
names match across types)
- Add "labels" to the client config: absent = no filtering, [] = unlabeled
items only; labels from the admin default config are pinned and cannot be
overridden by user, env or flag configs; a broken admin config aborts
- Filter in prepare() before planning; deselected items follow the
upstream-deletion rules (unmodified removed, modified kept + warn,
--force removes) with reason "not selected for this client's labels"
- Abort sync before any change when the labels file does not validate;
names that match no item only warn on the client
- Report the active selection in sync output and status (--json: labels
object with client, source, labels_file, not_selected, unknown_names)
- validate checks the labels file, including that every name resolves to
an own, vendored or MCP item
- Document in README ("Client labels"), fleet handoff and the proposal;
revise the "client-side selection" non-goal
- Tests: labels parser, config precedence, planner deselection, e2e for
filter/remove/keep-modified/abort/unknown names, validate
Refs: agent_lib-fh4
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Managed items with local modifications were kept (skip + warn) with no
escape hatch other than deleting the deployed file by hand — after an
upstream change the deployed content silently stayed old, which reads
exactly like a stale cache.
--force turns the two skip cases into update (repository version wins)
and remove (upstream deleted + modified). Unmanaged items and never-
managed path occupancies stay untouched even with --force; the default
behavior is unchanged.
A cached bare clone's HEAD points at refs/heads/<branch>, which is written
once at clone time and never updated again — fetch only touches
refs/remotes/*. Syncs without --ref therefore resolved the frozen
clone-time revision forever: the literal "cached version of the repo"
symptom.
Resolve the default branch via remote.List instead: the advertised HEAD
(plain hash or symref target) is the only source of truth. Verified
against the production Forgejo instance; e2e test fails on the pre-fix
binary and passes now.
sync/status now resolve the work repository from a fallback chain:
--repo flag > --config flag > AGENT_LIB_CONFIG env > user-level config
(XDG / AppData) > admin platform default. --repo works without any
config file; tokens fall back to AGENT_LIB_GIT_TOKEN/AGENT_LIB_GIT_USER
without ever overriding a configured token. A completely empty chain
yields an error listing every searched location, a minimal config
example and the --repo shortcut.
Closes beads: agent-lib-9o8
- status: read-only report of deployed/repository revision, pending changes,
skipped items (with reasons), unrecognized local items, binary version;
--json machine-readable; exit 1 when local modifications block repository
updates (headless drift checks)
- sync/status share a read-only prepare() front half
- notify package: windows toast via PowerShell script (go-toast mechanism,
zero deps, build-tag gated), noop on linux/macos; cross-compile verified
- toast fires exactly once per sync with warnings, never for routine syncs;
notification failures are logged and never fail the sync (fail-open)
- Plan is a pure function (inventory, manifest, disk facts, injected path
mapping) emitting exactly six actions: deploy, update, skip+warn,
leave+warn, remove, keep — deterministic order, no filesystem access
- full 3x3 matrix (disk unmodified/user-modified/never-managed x upstream
added/changed/removed) table-tested plus missing-disk restore cases,
a mixed fixture and a purity/determinism guarantee test
- executor applies plans item by item (atomic deploy/remove per item),
rewrites the manifest once, appends timestamped SKIP/LEAVE lines with
item name and reason to the sync log
- sync now plans before mutating: unmanaged occupants block new deploys,
upstream deletions remove only unmodified copies, local modifications
always win and are warned about
- admin-protected client config (flag > AGENT_LIB_CONFIG > platform default),
token used for basic auth only, redacted from all errors and files
- work-repo pull via go-git (bare cache clone + fetch, remote re-pointed on
config change), clean failure keeps previous state
- inventory: own items + external areas from lockfile; MCP inventoried but
never deployed
- manifest v1 records name/type/origin/revision/content-hash per item,
deterministic bytes; folder hashes over sorted relpath+filehash lines so
tree and disk hashes agree
- shared deploy package (atomic temp+rename, exec-bit preserving) now backs
both curator materialization and client deployment
- sync validates and hashes everything before the first mutation; second run
is a byte-level no-op