Claude Code gains in fluidity and stability after a series of fixes.

Anthropic fixes Claude Code with a new full-screen display, continuous streaming for reasoning and tool calls, and automated session self-repair.

Anthropic is rolling out a set of fixes for Claude Code, aimed at improving its responsiveness and reliability.

The most visible change is a new full-screen display, designed to eliminate flickering and adjusted to adapt to different devices and environments. It is activated via the /tui feedback command and is expected to become the default setting soon. Regarding responsiveness, the model's reasoning and tool calls are now streamed continuously, which eliminates the impression of freezing that could occur when Claude was thinking for an extended period.

Several error messages have been clarified, starting with the classic « tool result doesn't match tool use », whose root causes have been tracked down. Compaction, which condenses a session's history to free up space, now displays its progress and no longer encounters the « prompt too long » error. The MCP layer has also received a set of corrections for connection failures, OAuth flows, and proxy-level rate limiting.

Sessions are also now capable of self-repair: an image or media file that is too large or unreadable, which previously blocked the tool until a restart, is now automatically detected and bypassed. Sending feedback is finally simplified, with the /feedback command allowing the transmission of sessions from the last day or the last week without having to locate the problematic one.