Claude Code: Line-by-Line Streaming and Auto-Retry on Mid-Thinking Drops

Smoother streaming and fewer interruptions

Two improvements in Claude Code 2.1.181 change how responses are received and displayed, one visual and one reliability focused. Together they make Claude Code feel more responsive during active use.

Line-by-line streaming for long paragraphs

Previously the terminal renderer waited for the first line break before displaying a long paragraph. Extended explanations and analyses would appear to hang while the spinner ran, then land all at once. The effect was worst for dense prose where natural line breaks come late.

Now text appears incrementally, line by line, as it arrives. Responses feel more alive, and you can start reading while generation is still in progress. The visual cadence matches what Claude is actually producing in real time.

Auto-retry on mid-thinking connection drops

Claude Code's extended thinking involves a phase where the model reasons internally before producing visible output. If an API connection dropped during that phase, users saw the message "Connection closed while thinking" and were left with no response, forcing a manual retry of the full prompt.

In 2.1.181, connection drops that occur mid-thinking now trigger an automatic retry instead of an error. This is a quiet but meaningful reliability win for anyone using extended reasoning on unstable connections or during periods of API instability, which were common across the Claude ecosystem in June 2026.

A complementary pair

These changes address different layers of the same experience. The first makes good responses better to receive. The second recovers responses that would otherwise have been lost. For long-form reasoning tasks like architecture discussions and code reviews, both reduce friction where it matters most. Neither requires any configuration: they apply automatically to all 2.1.181 sessions.