Claude Code v2.1.235 major updates - Addition of spellcheck settings and security fixes around permissions

This is Ishikawa from the Cloud Business Division. Claude Code v2.1.235 (released 2026-08-18) has been released. The update includes one new feature addition, with the remainder allocated to fixes and improvements. I verified the Shift+Tab behavior in the permission prompt.

The previous update article is here.

https://dev.classmethod.jp/articles/20260818-cc-updates-v2-1-234/

Update Summary

v2.1.235 includes 19 changes. The breakdown shows fixes accounting for the majority at 12, with one new feature addition. The remainder consists of improvements to existing features such as permissions, terminal UI, and resource usage during background execution.

Notable Updates

Spellcheck Setting for Prompt Input Field

An optional spellcheck setting has been added to the prompt input field that displays underlines on misspelled words while typing. It uses installed aspell / hunspell / ispell for spell checking.

For those who frequently write prompts and commit messages in English, this offers the advantage of catching errors at the input stage. However, since the feature uses whatever spell checker is already installed, it seems like you'd first need to check whether aspell or similar is available on your machine.

Fix for Unintended Permission Grants in Permission Prompts

A bug was fixed where pressing Shift+Tab in the comment field of a permission prompt would approve the edit instead of closing the field, while also granting edit permissions for the entire session.

For those who use the comment field in permission prompts, this closes off a path that could unintentionally grant session-wide edit permissions, making it worth prioritizing as an upgrade.

Alignment Between Permission Dialog Display and Actual Permission Scope

The permission dialog has been improved. The display text and "don't ask again" options now always match the actual scope of the permission being granted, and "don't ask again" will no longer be presented when the content cannot be fully displayed.

For those who use "don't ask again" to narrow down permissions, having the display match the actual permission scope carries significant meaning.

Fix for Prompt Cache Invalidation on Language Server Reconnection

A bug was fixed where the entire prompt cache would be invalidated when a language server disconnected and reconnected mid-session.

For those running long sessions in environments with a language server, the reduced frequency of cache drops should have a positive impact on both wait times and costs.

Improved Memory and CPU Usage During Cloud Session Execution

Memory and CPU usage during background execution of cloud sessions such as /ultrareview and /autofix-pr has been improved. The event stream no longer re-scans and re-renders on every update.

For those who continue working locally while running cloud sessions in the background, this should translate directly to a noticeably lighter experience.

Fix for Embedded grep Memory Exhaustion and Context Output

The grep embedded in the macOS / Linux native build has been improved. Pathological patterns now fail immediately before exhausting memory, and combining -m N with -A / -C now outputs the correct context lines.

This affects those using the macOS / Linux native build, and should help prevent incidents where searches cause the local environment to become unresponsive.

Update Details

New Features

• An optional spellcheck setting has been added to the prompt input field that displays underlines on misspelled words while typing. It uses installed aspell / hunspell / ispell. This is a welcome update for English-speaking users, but unfortunately it does not support Japanese.

Security

• Fixed a bug where pressing Shift+Tab in the comment field of a permission prompt would approve the edit instead of closing the field, granting edit permissions for the entire session.
• The display text and "don't ask again" options in the permission dialog now always match the actual scope of the permission being granted. "Don't ask again" will no longer be presented when the content cannot be fully displayed.

Improvements & Performance

• Memory and CPU usage during background execution of cloud sessions such as /ultrareview and /autofix-pr has been improved. The event stream no longer re-scans and re-renders on every update.
• The context limit error message has been improved to notify users when auto-compact is disabled and guide them to /config to re-enable it.
• In Vim mode, NORMAL mode and cursor position are now preserved when toggling the detailed transcript display (ctrl+o) or closing panels.
• Remote Control's claude rc now applies the same enterprise gateway availability check as interactive startup.

Fixes

• Fix for full prompt cache invalidation: Fixed a bug where the entire prompt cache would be invalidated when a language server disconnected and reconnected mid-session.
• Fix for embedded grep memory exhaustion and context output: Fixed issues in the macOS / Linux native build where pathological patterns would exhaust memory, and where combining -m N with -A / -C would not correctly output context lines.
• Fix for misaligned nested list display: Fixed a bug in the terminal UI where markdown list items nested three or more levels deep would appear misaligned. Hanging indentation has also been added for wrapped list items.
• Fix for misaligned highlights in prompt input field: Fixed a bug where highlights (slash commands, keywords, mentions) in the input field would appear offset by one or more characters for certain multi-line prompts.
• Fix for default agent display in Agent tool: Fixed a bug where the Agent tool would present the general-purpose agent as the default even in sessions where it is unavailable. In such sessions, omitting subagent_type will now return a clear error listing the available agents.
• Fix for missing content in notebook approval dialog: Fixed a bug where the approval dialog for cell deletion or replacement would silently omit existing cell content when a notebook or cell could not be read. The dialog now displays the reason.
• Fix for accidental dialog selection: When pressing arrow keys and Enter in quick succession, the option at the destination is now selected instead of the previously highlighted one.
• Fix for SendMessage message discarding: SendMessage now proactively rejects messages that are too large for cross-session delivery instead of silently discarding them.
• [VSCode] Fix for Claude tab focus shifting: Fixed a bug where focus would unexpectedly shift between open Claude tabs when restoring or reloading a window with multiple Claude panels open.
• A quietly welcome fix: Fixed a bug where the "Update installed" restart notification would not appear in the prompt footer after an automatic background update. This should reduce cases where users unknowingly continue working on an outdated process.
• Additionally, various minor bugs have been fixed, including character display for slash commands executed during responses and the display state of task lists when resuming sessions.

Verifying the Shift+Tab Behavior in the Permission Prompt

Let's verify the Shift+Tab behavior from this security fix. I prepared a file containing hello in advance.

% mkdir -p ~/tmp/cc-shift-tab-check && cd ~/tmp/cc-shift-tab-check
% printf 'hello\n' > sample.txt
% ls -la
total 16
drwxr-xr-x@ 4 ishikawa.satoru  staff  128  8月 19 13:56 .
drwxr-xr-x@ 5 ishikawa.satoru  staff  160  8月 19 13:56 ..
-rw-r--r--@ 1 ishikawa.satoru  staff    6  8月 19 14:11 sample.txt

To ensure the permission prompt is displayed every time, launch claude in manual mode, which prompts for permission on every action.

% claude --permission-mode manual

Once launched, confirm that a grey ⏸ badge appears in the footer. As noted in the CHANGELOG with Added a grey ⏸ badge to the footer when in manual permission mode, making the active mode always visible, this badge indicates manual mode is active. In the steps that follow, the badge not changing serves as evidence that no permissions have been granted.

Make the following request to the session.

Please rewrite the contents of sample.txt from hello to hello world

A prompt requesting permission to edit the file will appear. Here, select No.

In this state, pressing Tab enables input in the permission prompt. Enter Test here.

In this state, pressing Shift+Tab returns to the state without comment field input (No), and the edit is not approved. Prior to the fix, it is said that instead of closing, the edit would be approved and the file would be modified.

Pressing Shift+Tab in a comment field feels like the kind of key you might hit unconsciously when trying to exit the input area. The fact that this was leading to approval and session-wide permission grants means there is a possibility that permission scopes were being expanded without users realizing it.

Closing

This release has a composition of 12 fixes out of 19 changes, feeling more like a release that solidifies the fundamentals rather than introducing prominent new features. The alignment between the permission dialog display and the actual permission scope is something I see as directly improving the ease of understanding permissions.

This is relevant not only for those who regularly interact with permission prompts, but also for those who make heavy use of search in native builds. Why not update and give it a try?

References

https://github.com/anthropics/claude-code/blob/main/CHANGELOG.md

https://code.claude.com/docs/en/changelog

https://dev.classmethod.jp/articles/20260818-cc-updates-v2-1-234/