[BUG] Android app: web-search permission modal fails to appear in remote-control Code session, leaving the model stuck in an execution loop (recoverable from PC, not from Android)

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?

When a remote-controlled Code session invokes the web-search skill and requests permission, the permission prompt is not reliably presented in the Android app. The model then becomes stuck in an execution loop that cannot be interrupted from the Android app. The loop can be broken from the PC controlling the same session, but not from Android. No internal error is shown during the loop — it only surfaces afterward, in the next prompt's output, where the model retrospectively diagnoses the previous turn's hang.

What Should Happen?

The permission modal should appear in the chat view; the "Allow once" action from the session list should open or resolve the same modal; and the user should be able to interrupt/stop execution from the Android app.

Error Messages/Logs

Sample transcript corroborating the "transient internal error, diagnosed only on the next turn" behavior (this instance involved research sub-agents rather than the web-search permission modal specifically):

● Running 3 agents…
  ├ Research ... · Initializing…
  ├ Research ... · Initializing…
  └ Research ... · Initializing…
● The three research agents hit a transient internal error and returned nothing. Re-running them now.
● 3 agents finished — all Done.

Steps to Reproduce

1. From the Android app, open/control the Claude Code session running on the Debian PC.
2. Have the session invoke the web-search skill, which requests permission (the "Allow once" prompt).
3. The in-chat permission prompt does not appear in the chat view on Android.
4. In the Code section's session list/menu view, an "Allow once" button appears for the session.
5. Open the chat from that menu — the selection/permission modal does not come up.
6. The model enters an execution loop and cannot be stopped from Android (it can be stopped from the PC).
7. After stopping the loop, send a follow-up prompt. The next response retrospectively identifies why the prior turn hung and surfaces an internal error message at that point.

Claude Model

Opus

Is this a regression?

I don't know

Last Working Version

_No response_

Claude Code Version

2.1.158 (Claude Code)

Platform

Anthropic API

Operating System

Ubuntu/Debian Linux

Terminal/Shell

Xterm

Additional Information

• Claude Android app: 1.260521.20 (26052120)
• Claude Code: 2.1.158
• OS (PC side): Debian Linux, terminal
• Session type: remote-control of the Debian Claude Code session from the Android app
• Frequency: frequent / reproducible