Proposal: background subagents, monitors, cron, worktree isolation #49842

Activity

1. rndmcnlly commented on Sep 23, 2026

   

   rndmcnlly

   on Sep 23, 2026

   More actions

   The monitor part is the one I would use immediately. I built opencode-perk for V1 because I wanted a running job to send progress back to an idle agent, without blocking the conversation or waiting for the job to finish. A CI watcher, for example, should be able to report a failing step while the run is still going, and I should still be able to talk to the agent in the same session. Claude Code’s Monitor tool captures the workflow well.

   V2 already owns background shell jobs, so I’d rather see progress events attached to those jobs than another job runner. Perk used an append-only drip stream with quiet-gap coalescing; stdout lines might be simpler here. I’m likely to prototype a V2 monitor plugin in the meantime to learn what works. The earlier #33806 looks like useful prior art.
2. KapilSareen commented on Sep 26, 2026

   

   KapilSareen

   on Sep 26, 2026

   Author

   More actions

   I kinda needed the background tasks without waiting on V2, so thought about sharing here so somebody could benefit if they're looking for it
3. opencodejr commented on Oct 28, 2026

   

   opencodejrBot

   on Oct 28, 2026

   Contributor

   More actions

   Thanks for sharing this, and for checking it against dev first!

   In short, you're proposing a set of agent-orchestration features from your fork:

   • non-blocking background subagents with durable completion notices
   • background and notify_parent tools
   • monitor, cron and sleep tools
   • automatic per-agent worktree isolation
   • ESC draining queued input instead of stranding it
   • completed diffs staying visible when tool details are hidden

   The maintainers will look at the proposal and decide which parts, if any, fit upstream. It's a large bundle, so if it comes to PRs, small separate ones per feature will be much easier to review. The tool-only pieces like monitor, cron and sleep might also work as an OpenCode plugin in the meantime. I can't promise anything about direction or timing, but thanks for writing it up so clearly.
4. gkurt commented on Nov 3, 2026

   

   gkurt

   on Nov 3, 2026

   More actions

   We embed OpenCode 2.2.5 through the SDK and built monitor and notify_parent as an Effect plugin. That works, but here are the gaps we hit, in case they help scope this:

   1. Monitor. We send output through steering. I mostly used @rndmcnlly's approach, with a few changes to batching, flushing, and stop conditions.
   2. Background commands that end look like ones that don't. Currently, tasks that finish (a build) and tasks that run indefinitely (a dev server) have the same shape. There is no way to tell them apart, and same rules apply to both of them. For example, I may terminate a sandbox if a dev server is the only remaining task, but not when a build task is running. I would say there should be a background: 'daemon' mode. Maybe monitor tool can also be folded in like background: 'monitor'.
   3. Warming doesn't know about pending wake-ups. It runs for a fixed time after the last request and skips child sessions. A session idling on a monitor or a background subagent goes cold. Our own warmer reuses the built-in keep-alive prompt word for word, so the built-in warming doesn't count our requests as activity. That's brittle. Maybe there should be a built-in helper for warming a session, or the background tasks and subagents should warm the main session regularly by default during their lifetime.
   4. Subagents can't report progress. Our notify_parent sends the parent a <subagent state="running"> message in the shape of the completion notice. Worth having built in?
   5. Need a feature to automatically move a task to background. When a task runs for too long, we move it to background automatically so that the agent can work on other things in parallel. I think it's worth having this as a built-in, with default warm up as mentioned in 3.
   6. A subagent is reported finished too early. When sessions.background moves a subagent's long tool call to the background, the subagent goes idle, and its parent gets state="completed" while the subagent's shell is still running.
   7. Editing a built-in tool's description doesn't reach the model. ctx.tool.transform with editor.update("shell", t => { t.description += "…" }) changes the registry; ctx.tool.list() shows the edit when the request is made. The tool definitions in that request still have the original description.

   Happy to send small PRs if you agree.
5. neriousy commented on Nov 12, 2026

   

   neriousy

   on Nov 12, 2026

   Member

   More actions

   Thanks for the suggestion. This isn't something we plan to build, so closing to keep the tracker focused.
6. neriousy

   closed this as not plannedon Nov 12, 2026
7. gkurt commented on Nov 13, 2026

   

   gkurt

   on Nov 13, 2026

   More actions

   @neriousy Are you not planning to fix the bugs too, or only the proposed feature?