Reviewing the branch turned up defects, several of them worse than the
problems the original commits set out to solve.
Destructive: deleting an automation removed its global triggers and only then
discovered a rundown event also referenced it. The delete failed, the user
cancelled, and their triggers were gone for good. The server reports trigger
references before event ones, so this was reachable whenever both existed.
Deleted triggers are now recreated when the delete is refused.
Silently wrong: a template read the project from disk while writes are
debounced by three seconds, so "make some automations, save them as a
template" could produce a template without them. Verified by hand: an
automation created milliseconds earlier now appears.
The demo shipped half a pair. The danger warning set a stage message visible
and nothing ever cleared it, so from the first event that hit its danger
window the message covered the countdown for the rest of the session. The
recipe library ships the clearing counterpart; the demo now does too, and the
whole cycle is verified against a running server.
Flood control undid itself. resetAutomationLogState ran on every onLoad, and
roll mode loads at every event boundary, so the "logging suppressed" notice
was re-emitted once per cue: exactly the flooding it exists to prevent. It
also wiped the throttle for onLoad and onStop immediately before writing to
it. Reset now happens on stop only, outside the early returns that made the
first attempt at this a no-op.
Trigger reconciliation diffed a mount-time selection against a live prop.
Settings are polled, so a trigger created in another tab while the form was
open would be deleted by a save that never saw it. It now diffs against the
snapshot, and says which triggers a save will remove rather than removing
several same-lifecycle triggers silently.
Also: a rundowns template no longer carries a phantom empty rundown from
makeNewProject; the partial-duplicate endpoint returns the name it actually
used, since collisions get renamed; the template flow invalidates the project
list rather than relying on a refetch on mount; the last-fired label drops to
a one minute cadence instead of holding a 1Hz timer per automation forever;
e2e cleanup moved to afterEach so a mid-test failure stops leaking state.
The e2e spec has now been run against a real server, green on both the demo
project and a blank one, leaving no residue.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01LpbLJVVT26tzWkduck1M9H
The panel started empty with no examples, so the first thing a new user saw
was a form that assumes they already know what OSC, a lifecycle and a template
string are. Eight recipes now cover the common integrations plus three that
need no external software at all, so the library is useful on a bare laptop.
A recipe is not a new kind of object. Installing one posts an ordinary
automation and its triggers through the same endpoints the form uses, then
opens it in the editor: seeing that it is editable is the point.
Every external target defaults to this machine, so a mis-click cannot put
traffic on a venue network, and a test asserts that along with the rest of
the server's automation contract, since recipes are constants that would
otherwise rot silently into something that 400s on install.
OSC recipes are hidden in the cloud build, where OSC output is skipped.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01LpbLJVVT26tzWkduck1M9H
Claude