[FEATURE] Label and isolate human-action instructions by destination

Preflight Checklist

• [x] I have searched existing requests and this feature hasn't been requested yet
• [x] This is a single feature request (not multiple features)

Problem Statement

When Claude surfaces instructions requiring human action — paste this in Terminal 1, run this in Terminal 2, click this in Browser — the instructions are sometimes combined into a single block without clearly labeling which destination each action belongs to.

This causes real errors. When I have two terminals open (one local, one SSH'd into a remote NAS) and Claude gives me a block of commands, I have to guess which terminal each command targets. Pasting into the wrong one means path errors, permission failures, or worse — running destructive commands on the wrong machine.

This is not a skill-level issue. Ambiguity costs time regardless of experience. Precision in human-action instructions is not patronizing — it benefits everyone.

Proposed Solution

Two layers of fix:

1. UI fix (Cowork/Chat/Code Tab): Render each human-action in a distinct block with a visible destination label. One action per block. Never combine two destinations in one block.

2. Model behavior fix: Claude should always label and isolate human-action instructions in output regardless of whether the UI renders them specially. The model should never combine two terminal commands targeting different destinations in a single code block.

Alternative Solutions

I currently instruct Claude via custom CLAUDE.md rules to never combine two actions in one block and to always label Terminal 1 vs Terminal 2. This works per-session but requires manual setup every time and doesn't survive compaction without a persistence mechanism.

Priority

High - Significant impact on productivity

Feature Category

Interactive mode (TUI)

Use Case Example

1. I have Terminal 1 on my MacBook (local) and Terminal 2 SSH'd into my NAS at work.
2. Claude tells me to create a directory and copy a file — but the directory command is local and the copy command targets the NAS.
3. Both commands are in one code block with no label. I paste both into Terminal 2 by mistake.
4. With labeled and isolated blocks, each command is clearly tagged with its destination. No ambiguity, no errors.