Alert rules already auto-migrate into automations on boot and the old
engine stands down after. Deletes the rule/firing tables, routes, UI,
and AI tools; keeps notification channels since automations still use
them. Also fixes automations' notify step, which still called the
deleted repository.
Repoints core consumers (dispatcher, collab, session-sharing, tab
utils) at the plugin, adds isTabTypeAvailable gating for rdp/vnc/
telnet, widens hostCapability to accept an array, and renames the
guacamole i18n namespace to remoteDesktop.
Moves src/backend/ai and its routes into plugins/ai, mounted through a
small /ai dispatcher instead of its own port. Moves the ai-enabled and
ai-private-endpoints settings routes out of user-settings-routes.ts
into the plugin too, so it has no leftover core dependents.
Require auth on /plugin-api, gate plugin enable/grant/revoke routes
behind admin.plugins.manage, and thread the calling user through
worker ctx.hosts/ctx.ssh calls instead of always using the install
owner. Also adds the plugin permissions grant/revoke UI.
Moves docker backend and frontend into plugins/docker/, alongside
ssh-terminal. Adds a host editor tab seam, a dashboard card registry,
and a build step to compile plugin backends into dist.
The test still asserted transferToHost was called with the destination
directory "/srv". Commit 7727fa09 (#1304) deliberately changed a single
file transfer to send the full destination path including the filename,
so the correct expectation is "/srv/remote-1.txt". The code was right
and the test was left behind.
ctx.http.route now registers a real router on the /plugin-api/:pluginId
dispatcher, replacing its stub 404. A plugin never touches the Express
req/res: it gets a plain summary and returns a plain object, so sockets
and session cookies stay on this side of the boundary. Authorization and
cookie headers are not forwarded.
Plugins load last in the start-up sequence, after the server is fully
wired, and are terminated before the database on shutdown so none can
outlive it. A plugin that fails to load cannot stop the backend.
fix: correct the nginx plugin-api location regex
The pattern was ^/plugin-api/(/.*)?$, which needs a double slash and so
never matched /plugin-api/<id>/<route>. Every other block in the file
uses ^/prefix(/.*)?$. Left alone, plugin routes would work in dev and
404 only behind Docker.
Every ctx call a plugin makes is checked against its granted
capabilities, performed by the main thread, then audited. A capability
must be both declared in the manifest and granted in the database, so
widening a plugin's reach always needs a new manifest the user can see,
never just a database row.
Audit attribution is set by the broker, not the plugin. The plugin
supplies only the details, so it cannot forge the actor.
fix: stop plugin workers seeing plaintext credentials in error messages
A failing ctx.ssh.connect handed the worker the raw error from deep in
the SSH stack, which embeds the connect config. The worker received, in
full:
Authentication failed for root using password <SECRET>
where <SECRET> was the host's real plaintext password. Any plugin with
ssh.exec could read it by catching the error, defeating the whole point
of the handle-based ssh API.
Errors crossing the boundary are now sanitised. Only PluginFacingError
and PluginPermissionError, both authored here, reach a plugin verbatim;
anything else becomes a generic message and the real one goes to the
log. This uses a marker class rather than matching on message text,
because matching would start leaking again the moment an upstream error
happened to contain a familiar phrase.
Covered by credential-isolation.test.ts, which plants a sentinel in
every secret-bearing field and asserts it appears in no message the
worker ever receives.
Formalises three ad-hoc publish paths that already existed:
automation-events.ts, metrics/automation-bridge.ts and the
hostSessionStatus singleton, now the host.session.status topic.
Both invariants the old code relied on are preserved. Publishing stays
fire-and-forget, so a failing subscriber cannot disturb a metrics poll or
a host delete. The automations engine subscribes to the bus at scheduler
start instead of being imported ad hoc, which keeps the edge
one-directional: automations may import repositories, hosts modules must
not import automations.
Validates an unpacked plugin, spawns a worker_threads worker and manages
activate/deactivate/crash-restart with backoff, disabling after three
tries.
The crash counter clears only after a plugin stays up past a stability
window, not on activation. Clearing it on activation let a plugin that
activates cleanly and then dies restart forever, because every attempt
looked like the first.
loader.ts documents what the worker boundary does and does not buy you.
Node's permission model is process-wide, not per-worker, so it cannot be
enabled for plugin workers alone and is not claimed here.
LukeGus