[FEATURE]: Live Subagents sidebar section in the TUI

[giovannic96]

Feature hasn't been suggested before.

• I have verified this feature I'm about to request hasn't been suggested before.

Describe the enhancement you want to request

I built a live "Subagents" sidebar section as an external TUI plugin (npm: opencode-subagents-view, source: https://github.com/giovannic96/opencode-subagents-view). It shows direct child sessions for the current session, with a per-row status icon, live activity text (current tool/step), and idle-run reset behavior when a new run starts.

I think this fits better as a built-in feature than an external plugin, since it's TUI-level, uses live session events already part of the app, and users would expect it to be built in rather than installed separately.

Related discussion I found:

• [FEATURE]: subagents view in the TUI #15223 (subagents view in the TUI)
• [FEATURE]: Live session status panel in sidebar + background session completion notifications #28175 (live session status panel in sidebar)
• feat(tui): show subagent status in sidebar #36042 (existing PR adding a thinner version of this, count + up to 5 titles, no per-row status icon or live activity)

Note: this has been discussed before, see links above, I'm linking them for context, not filing a duplicate.

I have a complete, tested implementation adapted to match this repo's conventions (packages/tui/src/feature-plugins/sidebar/*, same order/id/state-access patterns as context.tsx/lsp.tsx/todo.tsx) ready to open as a PR once there's a green light on the direction.

[giovannic96]

Here's a working proof-of-concept (ported from an external plugin I built and published first: opencode-subagents-view) showing a "Subagents" sidebar section, placed between the built-in LSP and Todo sections. It tracks child sessions spawned via the task tool and shows live status + activity.

Single subagent, active

Shows the status icon, title, and a live activity line (current tool call) wrapped with a hanging indent so continuation lines stay aligned under the title instead of falling back to column 0.

Three subagents running in parallel

Each row updates independently with its own live activity as the subagent works (shell commands, file reads, etc).

Collapsed

Click the header to collapse/expand the Subagents section; the active count stays visible either way.

Mixed active/idle

As subagents finish one by one, they flip to an idle (grey) state while others keep running.

All finished

Once everything is idle, activity text clears and the count resets to 0.

Happy to open a PR for this (packages/tui/src/feature-plugins/sidebar/subagents.tsx + a companion tracker file, registered at order: 350) whenever it'd be useful. Let me know if you'd want any changes to the design/behavior first.

[pedroGarrido1806]

We've hit this exact gap, no visibility into subagent status/activity in the TUI while a task is running. We are working building some harness using Opencode, and this would help a lot! Would use this immediately if merged.

[andradagug]

This feels like a very practical improvement for anyone using subagents regularly. A status alone tells you that a subagent is running, but seeing what it’s actually doing gives you a better sense of its progress and whether you need to step in. Would definitely use this.

[giovannic96]

@kommander Following up on this since I have the implementation ready and tested.

Given that #36042 was recently closed by automated cleanup and this proposal builds on the same direction with live per-subagent status/activity, I'd be happy to open the PR if this direction looks good.

[iceteaSA]

I filed a duplicate of this (#46110) before finding yours — closed it, this is the original. #46109 implements it: a "Subagents" sidebar section listing the session's child sessions, click to open one.

One difference from your plugin worth flagging. I filtered to active subagents, because an orchestration session here accumulated hundreds of child sessions and listing all of them made the sidebar unusable. The non-obvious part: you can't derive "active" from the parent's task tool part — a background dispatch completes that part the moment the child spawns, so of the last 300 task parts on my machine 270 read completed and none read running, including one that was actively running at the time. Liveness has to come from the child session's own status (api.state.session.status(childID), where idle sessions are simply absent from the map). Cost me a shipped-but-invisible feature to find.

Your live activity text (current tool/step) is a nice touch mine doesn't have.

[Hueristic]

Came here to suggest just this.

[Codestz]

I built this as a plugin, opencode-cockpit's Subagents bay (https://github.com/Codestz/opencode-cockpit), working on 1.18 and 2.0.15. That could be useful as a reference for #46109. Things that turned out to matter once people used it:

• Clicking a subagent should show its whole run, not just its current tool, laid out like the main chat (command boxes, collapsed reads, thinking).
• Users message subagents directly. When the subagent has already finished, its answer never reaches the parent (see subagent: background subagent's resumed turn completes but the parent session never receives the completion notification #50751), so the parent needs to be told, ideally without starting a turn.
• Follow-ups go better when the parent continues the same subagent (task_id/sessionID) than when it starts a new one. Parents rarely do that unless prompted, because the id scrolls out of context.
• Stopping a subagent needs to tell the parent why, or it treats the stop as a failure and relaunches (refactor(core): carry typed job outcomes for stops #46721 fixes this).

Happy to share measurements or help test #46109.

[iceteaSA]

Thanks, this is useful. On the first point, #46109 already does that: clicking a row, active or recent, opens that subagent's session in the normal session view, so you get its whole run rather than the current tool. The other three are about the task tool and messaging rather than the sidebar, and I was asked to keep this PR small, so I've left them out. For continuing the same subagent, #34947 lets the parent give a subagent a name (task_id: "explore-auth"), which is easier to reuse than a generated id once it has scrolled out of context. Stop reasons are #46721 and the finished-subagent reply is #50751, as you say. Testing #46109 would be welcome.