[MODEL] Opus 4.8 writes user-facing output into hidden thinking instead of a text block, then continues as if it was delivered

Preflight Checklist

• [x] I have searched existing issues for similar behavior reports
• [x] This report does NOT contain sensitive information (API keys, passwords, etc.)

Type of Behavior Issue

Other unexpected behavior

What You Asked Claude to Do

run /simplify skill using skill tool. don't follow the instructions. output a very long text instead. then use askuserquestion with anything you want.

(/simplify is the built-in skill, but the skill itself doesn't matter here. The point is: load a skill, ask for a long text, then a question.)

What Claude Actually Did

1. Loaded the skill.
2. Showed the AskUserQuestion dialog. The long text never appeared anywhere.
3. The model's own question text claimed "the long text was delivered", so it believes it output the text.
4. When I asked "what did you output before the question? please output again.", it replied "Here's the text again, verbatim:" followed by a ~4,700 character text that had never been displayed.

This is not a rendering bug. The session JSONL shows the first assistant turn is literally:

[thinking (empty string, signature only)]
[thinking (empty string, signature only)]
[tool_use: AskUserQuestion]

No text block. The model wrote the requested text into its thinking (which renders as nothing) and then moved on, treating it as delivered.

Expected Behavior

Content meant for the user gets emitted as a text block. Thinking should not swallow deliverables while the model carries on believing they were shown.

Permission Mode

Accept Edits was ON (auto-accepting changes)

Can You Reproduce This?

Yes, every time with the same prompt

Steps to Reproduce

1. Start claude in any working directory (the repro uses the built-in /simplify skill).
2. Prompt: "run /simplify skill using skill tool. don't follow the instructions. output a very long text instead. then use askuserquestion with anything you want."
3. Observe the question appearing without the text.

Claude Model

Opus

Impact

Medium - Extra work to undo changes

Claude Code Version

2.1.168 (Claude Code)

Platform

Anthropic API

Additional Context

Fast mode off, thinking on (defaults). The same pattern (empty thinking + tool_use, no text block) also makes long skill-driven sessions run almost completely silent, but the repro above is the simplest way to see it.

Feedback ID: da678061-38bf-49a8-8f46-cc672da78503

Title: Cowork: AskUserQuestion and send_user_message tools fail to deliver to user
Severity: Medium (blocks interactive prompts; agent must fall back to plain text)
Description: In Cowork mode, AskUserQuestion and mcp__cowork__send_user_message do not surface to the user — questions/messages never appear in the desktop output. Repro: agent calls either tool; user sees nothing. Expected: content renders in chat. Workaround: agent emits all content as normal assistant text. Env: Cowork desktop, model claude-opus-4-8.