Repository navigation
[Feature]: Improved orchestration #4376
Description
Problem or use case
Right now, it seems that the orchestration skill and set up makes it so the orchestrator can initialize and fire off prompts, but they dont continuously monitor workers nor get notified on their status.
As a result, I'm having to prompt the orch to poll worker status every x epoch, which also blocks them from being used otherwise.
Proposed solution
I know thru using Oh My Openagent that its possible for the workers to send a sort of signal to an orchestrator instead: https://github.com/code-yeongyu/oh-my-openagent
Something like this would be very useful
Alternatives or additional context
No response
Reacted by Andrei Elkin and jinleejun-corp
Activity
tizerluo commentedon Sep 2, 2026
Real-world data point for this exact pain. We run a multi-machine agent fleet with Orca 1.4.194 (macOS desktop + Linux VPS runtimes paired as environments): a coordinator agent (Kimi Code) dispatches work to a long-lived Codex worker on a remote VPS and needs to know when the worker reports back.
What we measured today:
orchestration check --waitin a loop, or a cron drivingterminal read). Polling latency (25 min cadence for us) is pure dead time between "worker is done" and "coordinator notices".worker-start --on <env>) and even read-only federation inspection (worker-showon a federated dispatch,orchestration inbox --environment) are rejected from the CLI withorchestration.federationAttachStart / federationShow / inbox is not available to mobile clients— the CLI client class seems to be treated as mobile, so federation flows appear desktop-only right now.worker_done/question/escalation.Our current workaround (works, but is glue we would love to delete): a "proxy" pattern — a Run homed on the worker's own runtime, coordinated by a shell terminal there, with the real coordinator driving it via
terminal send/read --environment; plus a local watcher script that polls the mailbox and injects into the coordinator terminal.The native feature we would want: a subscription/watch that, on new Run mail (optionally filtered by type), delivers a wake nudge into the bound coordinator terminal's input — push instead of poll. Bonus if it works when the Run home is on a paired remote runtime (federated delivery to the coordinator's home runtime inbox.
Happy to test early builds — this is our daily production path.
EOF
)
nwparker commentedon Oct 22, 2026
Implemented on main. Orca now delivers an orchestration mail notice into the bound coordinator agent when a worker reports back, so the coordinator does not need to keep polling to discover new mail. Terminal agents receive the input notice; native chats receive it through their normal send path, with approval/question holds released after the human responds.
Verified the current mailbox notification/delivery code and merged PR #28540 (including real Codex/Grok validation). The paired-runtime permission model from #26483 also allows workspace-level orchestration/federation calls rather than treating a paired desktop CLI as a phone. Closing as completed.
#28540
#26483