Remote Control: background subagent activity and tool calls are invisible in the app UI

Bug report — Remote Control sessions: background subagent activity and tool calls are invisible in the app UI

For filing on anthropics/claude-code. Drafted 2026-07-04.

Summary

When driving a Claude Code session from the desktop/mobile app via Remote Control, background subagent (Task tool) activity is not surfaced in the app UI. The driver sees only the main session's text output — subagents' individual tool calls, progress, and intermediate results are not visible. During multi-agent work this leaves the person driving effectively blind to everything happening below the top-level loop.

Environment

• Claude Code Desktop, Remote Control session (app driving a session running elsewhere)
• Main session fans out background subagents via the Task tool (run_in_background: true)

What happens

1. The main session spawns N background subagents.
2. Each subagent runs its own tool calls (Bash, Read, Edit, MCP tools, etc.) and may itself spawn children.
3. In the Remote Control app UI, none of that subagent tool activity appears. The driver sees the main session's messages (including the spawn confirmations and any summaries the main loop relays), but not the subagents' live tool calls or transcripts.

Impact

• The driver's only window into subagent work is whatever the main loop chooses to relay. If the main session doesn't summarize a finding, the driver never sees it.
• Diagnosing multi-agent problems from the driver's seat is very hard. In one session, a hook misconfiguration caused a recursive delegation loop (~15 subagents spawning children); from Remote Control this presented only as unexplained slowness and silence, because the agent activity that would have made the cause obvious was not visible.
• It undermines the reviewability of agentic runs when driven remotely — the person is accountable for the work but cannot watch it happen.

Expected

Remote Control should surface background subagent activity comparable to a local terminal session — at minimum a live list of running subagents with their status and current tool call, ideally the ability to open a subagent's transcript (as Ctrl+O does locally). Even a compact "N subagents running; last tool: …" indicator per agent would remove the total blindness.

Notes

• Local (non-remote) sessions do surface subagent transcripts (e.g. Ctrl+O), so this is specific to the Remote Control presentation layer, not the agent framework itself.
• Related reports from the same fleet/session: an OOM crash under multi-subagent load, and a plan-mode enforcement gap — both separately filed.