Claude Code 2.1.208 Is One of the Bigger Patch Releases in Months

Anthropic shipped Claude Code 2.1.208 early this morning, followed by a quick 2.1.209 a few hours later. The pair represents the most substance in a single release since the Bedrock/Vertex auto mode rollout last week.

The new features

Screen reader mode is the headline addition. Users who need a screen reader can now enable plain-text rendering via three methods: the --ax-screen-reader flag at launch, the CLAUDE_AX_SCREEN_READER=1 environment variable, or adding "axScreenReader": true to settings. The mode strips out the formatted TUI output and delivers clean text that works with standard assistive technology.

Vim insert mode remaps lets you define two-key insert-mode sequences through the new vimInsertModeRemaps setting. The canonical example is mapping jj to Escape, which is a common enough vim habit that its absence in earlier versions was a real friction point for vim users.

Process wrapper support via CLAUDE_CODE_PROCESS_WRAPPER is aimed at corporate environments where every executable needs to run through a sanctioned launcher. When this variable is set, both the agent view and the background service route every Claude Code self-spawn through the wrapper. This has been a gap that prevented some enterprise deployments from complying with IT policy.

Mouse input in fullscreen mode is a smaller ergonomic improvement. Multi-select menus and “Other” input rows in fullscreen now respond to mouse clicks, which removes the need to navigate those fields exclusively by keyboard.

The bug fixes that matter

The fix list is long. A few stand out.

Context window reset on auto-update: After CLI updates, the context window was silently reverting to 200k rather than keeping your configured value. This is fixed.

Background agent reply loss: Replies to background agents were being lost when delivery failed, with no recovery. Now the text is saved and delivered on the next restart.

Markdown table rendering: Tables with 200 or more rows were stalling the TUI. The fix introduces a truncated view with a row count notice, similar to how long file reads are handled.

Session transcripts are 79x smaller for edit-heavy sessions. This is significant for memory and cost. Edit-heavy sessions were accumulating redundant file content in the transcript. The fix eliminates that redundancy.

Memory and performance improvements:

MCP stdio server stderr was accumulating to 64 MB per server with no bound. Now capped.
LSP documents now use an LRU cache capped at 50 documents instead of staying open indefinitely.
File edit read cache is bounded to 16 MB (down from effectively unlimited).
Tool-call CPU overhead is up to 7x lower at high tool counts.
Multi-second per-turn slowdowns with many permission rules are addressed.

File tool fixes:

Edit tool was failing on files modified after reading.
Read was incorrectly reporting empty files as “shorter than offset.”
Grep was silently returning “No files found” for invalid regex patterns instead of surfacing the error.
Grep count mode was under-reporting totals when paginated.

AWS/Bedrock fixes: Bedrock streaming was producing a misleading “Truncated event message received” error. AWS SSO profiles where sso_region differs from the Bedrock region were breaking auth, a regression introduced in 2.1.207.

Auto mode safety: Destructive shell operations using $(...), backticks, or process substitution now prompt for confirmation in auto mode. Previously these could execute without review.

Version 2.1.209

The follow-up release, tagged a few hours after 2.1.208, fixes one regression: /model and other dialogs were being blocked in background sessions due to an overly broad guard added in 2.1.208. That’s reverted and working again.

Update via npm install -g @anthropic-ai/claude-code or the auto-updater if you have it running.

Sources: Claude Code GitHub Releases — v2.1.208 · Claude Code GitHub Releases — v2.1.209