[FEATURE]: Agent Activity Panel -- sidebar view for monitoring running background agents

[zh-zimoer]

Summary

When orchestrating multiple background agents in parallel (e.g., explore, librarian, oracle), there's currently no way to monitor their status without calling background_output() manually. A dedicated sidebar/panel showing live agent activity would greatly improve visibility into multi-agent workflows.

Problem

Right now, when agents are spawned with run_in_background=true, the user has zero visibility into:

• How many agents are currently running
• What each agent is doing (task description / current step)
• Their status (running / completed / failed)
• Their partial output (live progress)

The only way to check is to wait for a system notification and then call background_output(task_id="...") -- which is a pull-based, reactive model with no live visibility.

Proposed Solution

Add an Agent Activity Panel in the sidebar (similar to a CI/CD pipeline view) that shows all background agents with their real-time status.

Mockup

+- Agents ──────────────────────────────+
|                                       |
|  🟡 explore     "Find auth patterns…" |  <- click to expand
|  🟡 librarian   "JWT security docs…"  |  <- click to expand
|  ✅ oracle      "Architecture review" |  <- click to see output
|  ❌ explore     "Find error handlers" |  <- click to see error
|                                       |
+───────────────────────────────────────+

Clicking an agent entry expands to show:

• Full task description / prompt
• Live streaming output (if still running)
• Final result (if completed)
• Error details (if failed)
• Start time + elapsed duration

User Story

As a power user running complex multi-agent workflows, I want to see all active background agents in the sidebar at a glance, so I can monitor their progress and click into any agent to view real-time output -- without having to wait for completion notifications.

Acceptance Criteria

• Sidebar panel lists all background agents spawned in the current session
• Each entry shows: agent type, task description snippet, status icon (running / done / failed)
• Clicking an entry opens a detail view with full output
• Running agents show live/streaming output as it arrives
• Panel updates in real-time (no manual refresh needed)
• Completed/failed agents remain visible (dismissible)

Context

This is especially valuable for workflows where 3-5 agents run in parallel (e.g., parallel codebase exploration + external docs research + architecture consultation). Currently the UX feels like "fire and forget" with no control panel.

Related to the overall goal of making multi-agent orchestration more transparent and manageable from the UI.

[Snailflyer]

I would make the panel prove liveness, not just list spawned background jobs.

For each row, the minimum useful receipt could be:

• background task id
• parent session / conversation id
• agent type
• current turn id or output cursor
• last event observed by the UI
• whether the detail view is live, stale, completed, or failed

The failure mode to avoid is a sidebar row that still looks active while background_output(task_id=...) would return a completed, failed, or disconnected stream. For multi-agent orchestration, "visible in the panel" and "still advancing in the same session" should be separate states.

Adjacent reference: Faryo has the same control-surface boundary in a smaller tmux-first product. The phone/browser workbench only counts an action as successful when the same live terminal session advances:
https://github.com/Snailflyer/faryo/releases/download/v1.0.7/faryo-public-redacted-same-session-handoff-walkthrough-20260528-0120.gif

[wusai2333]

👍 Strongly support this feature.

When orchestrating multi-agent workflows (parallel explore + librarian + oracle), the lack of real-time visibility into background subagents is a significant UX gap. Users are essentially flying blind after run_in_background=true — relying solely on system-reminder notifications and reactive background_output() polling.

The proposed Agent Activity Panel design (CI/CD pipeline style sidebar) is exactly the right approach. The server-side SSE infrastructure (session.status events) is already there; this is primarily a rendering concern.

Would love to see this prioritized — it's a key unlock for any serious multi-agent use case.

[yukihamada]

I'd like to implement this. Before I start, a quick design check with the core team (per CONTRIBUTING.md):

Proposed approach:

1. Server: expose a GET /session/:id/background-agents endpoint (or extend the existing event stream) that lists background tasks with status, agent type, task snippet, start time — sourced from the existing background task registry rather than a new store.
2. TUI: add a collapsible "Agents" section in the sidebar, updated via the existing event bus (no polling).
3. Click/enter on an entry opens the existing background output view.

Scope for a first PR: read-only panel (list + status + open output). Live streaming of partial output can be a follow-up.

Does this direction match what the core team has in mind, or is someone already working on it? Happy to adjust the design before writing code.

[makslanies]

One implementation detail that may prevent this panel from showing ghost agents: make it a projection of a durable execution ledger, not just the live event stream. Each row needs a stable run_id / agent_id, parent identity, explicit lifecycle transitions, a heartbeat timestamp, and a terminal result; otherwise a disconnect can leave an agent looking running forever even though its process completed or died.

VOLY's A2A canvas uses this pattern: nodes are upserted from pending to running, then completed / failed / blocked, with route, duration, cost, files, error, and heartbeat-backed run records: https://github.com/voly-codes/voly/blob/main/docs/backend/a2a.md

For the proposed read-only first PR, I'd include stale as a distinct display state even if restart/cancel controls come later. It preserves honest observability without expanding the initial scope.