tui: session freezes after model thinking dump, unresponsive to input, reconnect gets stuck in compaction #36537

Summary

During a medium-length session (~10-30 messages), the model output a long
thinking block that was dumped into the thread, then the TUI became completely
unresponsive — ENTER, Ctrl-C, Ctrl-X, and Esc all had no effect. The session
had to be killed by closing the PowerShell window. On reconnecting with
--continue, the history was truncated and the session got stuck in
"compaction" indefinitely.

Environment

• opencode version: 1.17.17
• OS: Windows_NT 10.0.26200 (win32 x64)
• Terminal: VS Code integrated terminal (TERM_PROGRAM=vscode, COLORTERM=truecolor,
  TERM=xterm-256color)
• Shell: C:\WINDOWS\system32\cmd.exe (via PowerShell)
• Install/channel: release (1.17.17)
• Active plugins: none found

Reproduction

This was a one-time occurrence and has not been reproduced consistently.
Attempted reproduction steps:

1. Run opencode in a PowerShell terminal (VS Code integrated terminal)
2. Engage in a session of ~10-30 messages until the model produces a large
   thinking block
3. Observe the thinking content being dumped into the thread output
4. Attempt to submit, cancel, or interrupt (ENTER, Ctrl-C, Ctrl-X, Esc)
5. Observe no response to any keyboard input
6. Force-close the terminal / PowerShell process
7. Reopen opencode with --continue
8. Observe truncated history and session stuck in "compaction"

Expected Behavior

• The TUI should remain responsive to keyboard input even after large model
  output
• Ctrl-C / Esc should interrupt model output and return to prompt
• --continue should restore the full session without getting stuck in
  compaction

Actual Behavior

• TUI becomes completely unresponsive — no keybinds work
• No way to recover the session without force-killing the process
• Reconnect via --continue produces truncated history and hangs on compaction

Additional Context

• No logs available due to the session being killed
• Occurred once; session length was medium (~10-30 messages)
• The pre-crash event was a model thinking dump (the reasoning block was
  printed into the thread)
• No recent opencode updates, config changes, or plugin installations