[Feature Request] Actor-based output highlighting: distinguish "user must act" from "assistant is handling"

Summary

Provide a way to visually distinguish assistant output that requires user action from output the assistant is handling itself. Today there is no reliable signal, because Claude Code strips ANSI color codes from model responses — so the model cannot self-mark actionable text, and the only surviving signals are emoji and markdown structure.

Use case

In long or auto-mode sessions, the actionable lines ("approve this", "set X before I publish", "confirm the deploy") get buried in narrative and progress updates. Power users need to glance and instantly see what's in their court vs. what's already handled. A workaround using emoji sigils + blockquotes works but is purely conventional and fragile — there's no real color, because ANSI in responses is discarded.

The underlying blocker

Repeatedly requested and repeatedly closed by inactivity, not resolution:

• #16668 — Support ANSI color codes in model response
• #25346 — ANSI escape codes leak into adjacent chars (Windows Terminal)
• #18728 — Support ANSI color codes in bash command output
• #39369 — Support custom theme colors / respect terminal ANSI palette

And a live, complementary request:

• #60230 — Content-type color coding for assistant output (questions, commands, narrative)

Proposed solution (any one is sufficient)

1. Semantic markers the model can emit (e.g. a lightweight :::action / :::handled fence or inline tag) that the renderer maps to theme colors — keeps the model out of raw ANSI.
2. Honor a small ANSI/color subset in responses, gated behind a setting.
3. Extend #60230's theme.contentColors map with an actionRequired content type.

Why it matters

The feedback-loop cost compounds in auto/headless workflows. A single reserved "you need to act" color removes the need to read every line to find the one that needs a human.