[Desktop] computer-use request_access hangs forever — approval side panel never renders after 1.18286.0 auto-update

Preflight Checklist

• [x] I have searched existing issues and this hasn't been reported yet
• [x] This is a single bug report (please file separate reports for different bugs)
• [x] I am using the latest version of Claude Code

What's Wrong?

Since the Claude Desktop (macOS) auto-update to v1.18286.0 (installed 2026-07-02 20:43 UTC), every call to the computer-use MCP tool request_access hangs indefinitely. The approval side panel (cu-side-panel) never renders, so the user can never approve or deny the requested app list, and the tool call blocks until manually interrupted.

list_granted_applications still responds instantly with {"allowedApps":[],...}, so the MCP server itself is alive — the failure is isolated to the approval panel UI. Because the session allowlist can never be populated, ALL computer-use functionality is unusable.

Reproduced across 3 full app relaunches (2026-07-02/03 UTC) and again on 2026-07-04. Single-display setup, no extra Spaces — the panel genuinely never appears.

(Note: this concerns the Claude Desktop app's built-in computer-use MCP, not the CLI — the Platform/Terminal dropdowns below don't quite apply.)

What Should Happen?

The approval side panel should dock and list the requested applications for approve/deny, as it did before the update — main.log shows [cu-side-panel] docked / restored events working through 2026-07-02 18:16 UTC, with grants completing within ~3 seconds of the request.

Error Messages/Logs

~/Library/Logs/Claude/main.log (timestamps JST = UTC+9):
2026-07-03 02:59:54 [info] [cu-side-panel] docked       <- pre-update, working
2026-07-03 03:16:15 [info] [cu-side-panel] restored     <- last normal event
2026-07-03 05:43:20 [info] [cu-side-panel] initialized  <- first launch after update
2026-07-03 06:13:07 [info] [cu-side-panel] initialized
2026-07-03 13:58:42 [info] [cu-side-panel] initialized  <- never docks again

~/Library/Caches/claude-cli-nodejs/<project>/mcp-logs-computer-use/2026-07-04T09-10-36-376Z.jsonl (UTC):
{"debug":"Calling MCP tool: request_access","timestamp":"2026-07-04T09:20:36.501Z"}
{"debug":"Tool 'request_access' still running (30s elapsed)","timestamp":"2026-07-04T09:21:06.506Z"}
{"debug":"Tool 'request_access' still running (60s elapsed)","timestamp":"2026-07-04T09:21:36.507Z"}
{"debug":"Tool 'request_access' still running (90s elapsed)","timestamp":"2026-07-04T09:22:06.507Z"}
{"debug":"Tool 'request_access' failed after 105s: MCP error -32001: AbortError: remote-cancel","timestamp":"2026-07-04T09:22:21.674Z"}

Other hanging attempts ended the same way after 140 s, 19 s, and 202 s (all "AbortError: remote-cancel" = manual user interrupt).

Steps to Reproduce

1. Claude Desktop (macOS) v1.18286.0, any local session.
2. Ask Claude to do something requiring computer use; it calls mcp__computer-use__request_access with a list of apps.
3. The tool call runs forever ("using request access..." spinner in the chat). No approval UI appears anywhere on screen.
4. The user must interrupt manually; the MCP log then records failed after Ns: MCP error -32001: AbortError: remote-cancel.

Timeline evidence that the auto-update is the trigger:

• Through 2026-07-02 18:16 UTC: [cu-side-panel] docked/restored normal, request_access grants succeed (three grants on 2026-07-01, each answered within ~3 s).
• 2026-07-02 20:43 UTC: ShipIt auto-update installs.
• 2026-07-02 20:49 UTC (6 minutes later): first hanging request_access; every attempt since hangs.

Claude Model

None

Is this a regression?

Yes, this worked in a previous version

Last Working Version

Unknown build number — whatever Desktop build was replaced by the 2026-07-02 20:43 UTC auto-update

Claude Code Version

Claude Desktop (macOS) 1.18286.0 — bug is in the Desktop app's bundled computer-use MCP, not the CLI

Platform

Anthropic API

Operating System

macOS

Terminal/Shell

Terminal.app (macOS)

Additional Information

• macOS 26.5.2, arm64 (MacBook Air M2)
• Hit repeatedly for 2 days ("hangs every time") before we root-caused it from logs.
• Possibly related (same request_access/approval boundary, different symptom): #54872 — there the approval UI rendered but froze the whole app UI while the task completed underneath; here the approval panel never renders at all and the tool call never completes. That issue was auto-closed as stale.
• Happy to provide fuller logs on request.