Pause all dispatch and catch up with the feed
Two daemon-wide controls added in v0.2.8 (#213): a dispatch pause that stops everything new from starting, and an event feed with a cursor that lets a client which was offline — a closed Mac app window, a phone, a relay — catch up on what happened.
Pause all dispatch
Section titled “Pause all dispatch”Pause when you need the daemon to stop starting work without stopping it: a deploy freeze, a broken main branch, a tracker cleanup, an incident.
- Dashboard: the status strip’s Pause dispatch button (with a confirm). While paused the strip shows a red Dispatch paused chip with the reason and the end time, and the button reads Resume dispatch.
- Mac app: the tray’s Pause all agents pauses every running local daemon; once all of them are paused it reads Resume all agents. Each project’s line shows dispatch paused.
- Terminal:
itervox pause [--reason r] [--for 2h]anditervox resume(see the CLI reference). - API:
POST /api/v1/dispatch/pauseand/resume(see the API reference).
What a pause does:
- Running agents continue. Their turns finish, their results are recorded, their PRs are opened exactly as without a pause.
- Nothing new starts: no dispatch of a waiting issue, no retry, no reviewer, no automation and no resume of an answered question.
- Nothing is dropped. Retries stay due (without spending an attempt),
reviews wait as pending reviews, automations wait in the queue with the
reason
dispatch_paused, answered questions wait to resume, and idle issues show Dispatch paused as the reason they are not running. On resume they start as slots allow. - It survives a restart and a
WORKFLOW.mdreload. With--for(or anuntilin the API) it ends by itself; a pause whose end passed while the daemon was down ends on its first tick.
The pause does not stop operator actions that do not start an agent, the tracker poll, reconciliation or the dependency audit. An on-demand dependency analysis you start yourself still runs.
Catch up with the event feed
Section titled “Catch up with the event feed”GET /api/v1/feed?since=<seq> returns the notable events after a cursor:
runs started, finished and failed, questions asked and answered, review
verdicts, PRs opened and merged, guardrail blocks and approvals, lessons
proposed, and pauses and resumes. Each event has a seq that only grows,
a time, a kind, the issue and a small payload — never agent output or
prompt text.
A client keeps the last seq it has seen and, when it comes back, reads
from it:
curl -s -H "Authorization: Bearer $ITERVOX_API_TOKEN" \ "http://127.0.0.1:8090/api/v1/feed?since=1287" | jq '.events[] | [.seq, .kind, .identifier] | @tsv'- Read again from
nextSeqwhilehasMoreis true. truncated: truemeans the client was away longer than the window (the newest 5000 events, none older than 7 days): reloadGET /api/v1/stateand continue fromheadSeq.- A live dashboard connection receives
event: feedwith the newest seq on the/api/v1/eventsstream, so it reads the feed only when it moved.
The feed is stored in feed.jsonl and the pause in dispatch_paused.json,
both in the daemon log directory (~/.itervox/logs/<project>/). They are
runtime state: never commit them.