Bring back token-by-token streaming (line-by-line streaming since v2.1.78 feels worse)

Description

Since v2.1.78 introduced line-by-line streaming, the visual experience of receiving text from Claude feels noticeably less smooth than the previous token-by-token streaming.

Before (≤ v2.1.77)

Text streamed token-by-token, similar to claude.ai in the browser. Words appeared fluidly as they were generated, giving a "live writing" feel. Smooth and pleasant to follow.

After (v2.1.78+)

Text streams line-by-line. Whole lines appear at once, then a pause, then the next whole line at once. Within a line there is no progressive build-up — it pops into existence. This feels chunky/staccato and is harder to follow.

Why this matters

• Reading along while Claude generates is harder when output appears in bursts.
• Token-streaming made it easier to start parsing the response as it arrives.
• The previous behavior was a quality-of-experience signature of Claude Code that matched the API's actual streaming granularity.

Request

Please add an opt-in setting / env var to revert to token-streaming, e.g.:

• streamingMode: "token" in settings.json, or
• CLAUDE_CODE_STREAMING_MODE=token

Defaults can stay as-is (line-streaming) — just give power users the choice.

Environment

• Claude Code v2.1.119
• macOS (Apple Silicon), Ghostty terminal
• Default TUI renderer

Related

This is not the same as #37569 (which requests disabling streaming entirely). This issue requests reverting the granularity of streaming, not turning it off.