Windows Terminal: Claude Code shows block cursor despite `cursorShape = "bar"`

Found 3 possible duplicate issues:

1. https://github.com/anthropics/claude-code/issues/20878
2. https://github.com/anthropics/claude-code/issues/53619
3. https://github.com/anthropics/claude-code/issues/30082

This issue will be automatically closed as a duplicate in 3 days.

• If your issue is a duplicate, please close it and 👍 the existing issue instead
• To prevent auto-closure, add a comment or 👎 this comment

🤖 Generated with Claude Code

Thanks for the duplicate detection. This is related to #20878, #53619, and #30082, but I think this issue adds a specific implementation detail/workaround that is not captured there.

The new finding is that on Claude Code 2.1.183 with Windows Terminal, setting:

{
  "cachedGrowthBookFeatures": {
    "tengu_native_cursor": true
  }
}

in ~/.claude.json, together with removing explicit UI/input-mode overrides from ~/.claude/settings.json, made Claude Code use the native terminal cursor successfully.

So this is not just another request for configurable cursor style. It suggests the native cursor path already exists behind an internal GrowthBook flag and works in practice. The request is to expose/document that path as a stable setting or environment variable, or clarify whether CLAUDE_CODE_ACCESSIBILITY=1 is the supported public workaround and what UI tradeoffs it has.

#20878 can still be the umbrella tracking issue, but I think this issue is useful to keep because it documents a working native-cursor path on current Claude Code.

[!WARNING] Superseded on 2026-07-17. Do not use the old script.

The original comment documented an experimental composite workaround. It edited cachedGrowthBookFeatures.tengu_native_cursor, which is a server-managed feature-gate cache rather than a supported configuration surface, and it removed several settings whose relationship to cursor rendering was not isolated.

The pasted version also lost path separators such as the one in .claude\settings.json, so it should not be copied or executed.

Current Claude Code documentation provides a stable, cursor-specific option. Add this to the existing env object in ~/.claude/settings.json:

{
  "env": {
    "CLAUDE_CODE_NATIVE_CURSOR": "1"
  }
}

Then fully close all Claude Code processes and restart. This uses the terminal's native cursor and respects its blink, shape, and focus settings. It does not require editing ~/.claude.json, removing editorMode / tui / theme / showTurnDuration, changing showSpinnerTree, or touching MCP/plugin settings.

Documentation: https://code.claude.com/docs/en/env-vars

For historical context, the earlier composite operation produced a thin cursor on three Windows devices, but that observation did not establish which individual change caused it. The old script is therefore retained only in GitHub's edit history, not as a recommended workaround.

Same symptom on macOS + Ghostty, so this isn't platform:windows — the label probably undersells it. More usefully: step 1 of your workaround doesn't hold here, and I think I found why, plus a lever that does work.

tengu_native_cursor can't be pinned from ~/.claude.json

cachedGrowthBookFeatures is a cache of a server-side gate, not a setting. I set tengu_native_cursor = false, fully quit every claude process, restarted — and it was back to true on the next launch, written by Claude Code itself. Editing it in the other direction will be just as unstable; if it worked on your machine, I'd guess the gate simply already evaluated that way for your account rather than your edit being what stuck.

Worth checking on your side: does your edit still hold after a few days, or did the gate just happen to agree with it?

What actually works

// ~/.claude/settings.json
{ "env": { "CLAUDE_CODE_ACCESSIBILITY": "1" } }

Thin native cursor is back, survives restarts, and doesn't depend on the gate or on removing any of your settings. From the minified 2.1.212 bundle (best-effort read, mangled symbols):

function Iwe(){                                   // -> { nativeCursor: Iwe() }
  if (ye.CLAUDE_CODE_ACCESSIBILITY) return true;  // first line: overrides everything below
  if (TL()) return true;
  return !x5e() && _Vi();
}
function _Vi(){
  ...
  return et("tengu_native_cursor", false);        // the server gate you were editing
}

Because CLAUDE_CODE_ACCESSIBILITY short-circuits Iwe() before the gate is consulted, it's stable in a way the cache edit can't be.

On removing editorMode / tui / theme / showTurnDuration

I couldn't reproduce this part as a cause. I had tui: fullscreen set in both of my installs (and theme + editorMode in one), and CLAUDE_CODE_ACCESSIBILITY=1 gives a thin cursor with all of them still in place. I can't see a mechanism by which theme would reach cursor rendering, and the code path above doesn't consult any of these. Possible those removals were incidental to whatever fixed it for you.

Environment

• Claude Code 2.1.212 (homebrew cask claude-code@latest)
• Ghostty 1.3.1, cursor-style = bar, TERM=xterm-ghostty
• macOS 26.5.2 (arm64), zsh

Cross-referencing #20878 (canonical) where I've posted the full findings.

@mrsombre, thanks — this is a very useful cross-platform data point and code-path analysis. I agree with the main conclusion: cachedGrowthBookFeatures is not a stable configuration surface, and my original composite workaround did not isolate the causal variable.

One clarification on the Windows evidence: this was not based only on the first machine where the thin cursor appeared unexpectedly. I then applied the same composite workaround on a new Windows desktop and got the thin native cursor, and a classmate followed the same Chinese guide on a third Windows laptop with the same result. All three devices were Windows. That is why I initially attributed the result to the configuration differences. It remains a real three-device empirical reproduction, but I agree that it does not prove which individual edit was responsible. Your false -> restart -> true test demonstrates that the GrowthBook value can be rewritten; it does not strictly rule out a transient effect from writing true in the older 2.1.183 Windows build, but it does show why the cache must not be recommended as a setting.

There is also a more targeted supported lever in the current documentation:

{
  "env": {
    "CLAUDE_CODE_NATIVE_CURSOR": "1"
  }
}

The docs say this displays the terminal's own cursor and respects its blink, shape, and focus settings: https://code.claude.com/docs/en/env-vars

That is preferable here to both editing the GrowthBook cache and enabling the broader accessibility mode. It can be set under env in ~/.claude/settings.json, so the UI settings can remain unchanged.

I've updated the issue and replaced my old script comment with a superseded notice, including a correction for the path separators that were lost in the pasted GitHub version. I've also revised the Chinese guide to use CLAUDE_CODE_NATIVE_CURSOR=1 and kept the three-Windows-device result only as historical context.

Thanks again for tracing the 2.1.212 path and cross-referencing the canonical issue.