Text typed in the permission-deny dialog is lost when denying one of several parallel Agent tool calls (VS Code extension, 2.1.263)

Summary

When several Agent tool calls are launched in parallel (5 subagents in a single assistant turn) and the user denies one of them via the permission dialog while typing feedback in the "tell Claude what to do differently" input, the typed text never reaches the model. The model only receives the generic denial string (The user doesn't want to proceed with this tool use. ... STOP what you are doing and wait for the user to tell you how to proceed.). The other four subagents keep running (one ran for ~13 minutes), so from the user's point of view Claude ignored the question and continued working.

The user's typed text was lost entirely; it was not delivered as a later user turn either. The user's next actual message arrived prefixed with [Request interrupted by user for tool use], but its content was a complaint about the unanswered question, not the question itself.

Environment

• Claude Code 2.1.263 (VS Code extension anthropic.claude-code-2.1.263-linux-x64)
• VS Code 1.136.1, remote WSL2 (Linux 6.18.33.2-microsoft-standard-WSL2)
• Model: claude-fable-5-1
• Permission mode: auto mode (most Bash/Read/Edit calls were auto-approved; the permission prompt appeared only for the Agent tool)

Steps to reproduce

1. Run a skill/prompt that makes the model launch multiple Agent subagents in one turn (in my case 5 Agent calls with subagent_type: sonnet-model agents, run_in_background: false, sent as parallel tool uses in a single assistant message).
2. When the permission dialog appears for one of them, choose the deny option and type a question/instruction in the free-text input, then submit.
3. Observe:

• The denied subagent's tool result contains only the generic denial text; the typed text is absent.
• The remaining parallel subagents continue running to completion.
• The model never sees the typed text, so it cannot answer or change course until the subagents finish.

Expected

• The text typed in the deny dialog should be delivered to the model together with (or immediately after) the denial, as it is when a single tool call is denied.
• Ideally, denying one call in a parallel batch should also cancel or pause the sibling calls, so the user's instruction is acted on promptly.

Actual

• Only the denial marker reaches the model. The typed text is dropped.
• Sibling subagents keep running for minutes; the user perceives their input as ignored.

Impact

The free-text input on the deny dialog becomes unusable in exactly the situation where it matters most (long-running parallel work). The user in this session concluded the input field is worse than having none, since it silently discards what they typed.