Repository navigation
fix(onboarding): the dashboard stays empty for 30s after the first-run overlay closes #3109
Description
What happens
On a fresh marketplace install: claim the admin at /setup, complete every step of the first-run overlay, and the dashboard Timeline shows no agents. They appear about 30 seconds later, or immediately on a manual refresh.
Reproduced on a clean Vultr one-click install (image sha-128a0f6). Four agents existed server-side the whole time — the acme trio plus Cornelius — seeded by the backend during /setup completion.
Why
Four separate gaps, each of which alone would be enough:
-
The dashboard store has no
agent_createdhandler.stores/network.js:672-712dispatches onagent_collaboration,agent_status,agent_started,agent_tags_changed,agent_label_changed,agent_deletedandagent_activity. Nothing for a created agent. -
The broadcast could not be caught even if it were.
services/agent_service/crud.py:2277emits{"event": "agent_created", "data": {...}}with notypekey, and that dispatcher keys ondata.type. Theagent_deletedbranch atnetwork.js:706is already dead for the same reason (routers/agents.py:702,:800). Events that do work —agent_started,agent_stopped,agent_label_changed— deliberately send both keys. -
Two stores, two websocket connections, one listener on the wrong one.
utils/websocket.js:124-130does handleagent_created, and writes tostores/agents.js. The dashboard and Timeline readstores/network.js. -
The overlay refetches three stores and no agent list.
FirstRunOverlay.vue:375-382forcesloadFeatureFlags,fetchUserProfileandfirstRun.fetchStateafter every step;close()at:303-316refetches nothing.Dashboard.vue:510mounts it with no close handler — unlike the create modal two hundred lines below at:683-686, which does exactly the missing thing:
function onCreateModalClose() {
showCreateModal.value = false
networkStore.fetchAgents()
}The only thing that recovers is the 30-second poll at stores/network.js:1367-1403, which correctly detects additions. That is why the wait works and why it takes half a minute.
Related
ActivationChecklist.vue:146-155 fixed the sibling surface — the getting-started checklist not updating without a reload. It works because it watches agentsStore.agents.length, and agentsStore is the store the websocket actually feeds. stores/network.js has no such feed to watch.
Suggested scope
- Refetch agents when the overlay closes, matching
onCreateModalClose. - Add an
agent_createdhandler tostores/network.js. - Make the backend broadcast carry
typealongsideevent, which also revives the deadagent_deletedbranch.
The last two are worth doing together; a fix to only the overlay leaves any server-side agent creation invisible to an open dashboard.
Activity
dolho commentedon Oct 1, 2026
/claim
github-actions commentedon Oct 1, 2026
Claimed by @dolho. Label
status-in-progressadded.Next steps:
feature/3109-<short-slug>Fixes #3109docs(learnings): a /ws broadcast must carry the key its consumer disp…
dolho commentedon Oct 1, 2026
Fixed in #3158. Today's Workspace fixes, tested together on a local dev stack and reviewed
This issue: PR #3158.
stores/network.jshandlesagent_createdby refetching through the REST list.agent_created/agent_deletednow carrytype, which also revives the deadagent_deletedbranch.Live, with the dashboard open: a new agent appeared after 0.53 s instead of up to 30 s.
Local verification, 2026-10-01
The frontend was built from
dev+ all five PRs; they merge together cleanly. It ran against a local backend carrying the two backend fixes. All checks were in light and dark, with 0 page errors.[object Object]tooltipThe test clients, runs and agents were removed afterwards.
Review (
/review, posted on each PR)0 critical findings across the five. Informational notes are on each PR. The most useful:
PortalAgentDetailsretry bindings aren't covered by a mounted test.dev'sthreadsLoaded, so it doesn't backport to older branches.shadowedbadge names its sources, not "[object Object]" (#3135) #3157:LibrarySkillSets.vuelacks thesource_idfallback.Every PR carries a regression test that fails with its fix reverted. CI: no failures; #3157 and #3158 were still running when this was posted.
Separate observation (not from these PRs)
On every backend reload the default system seed retries and logs
Default system seed FAILED … Agent already existsfor the acme trio. It has done so since at least 10:07 today, before any of these changes. It's harmless, but it's noise worth its own issue if it reproduces on a clean install.fix(dashboard): new agents appear at once, not on the 30 s poll (#3109)…