Where does the bug appear (feature/product)?
Cursor IDE
Describe the Bug
Auto-review is enabled and selected, but it no longer behaves like semantic review. Routine Shell commands that the agent was explicitly instructed to run (e.g. stage + commit after “create a git commit”) are blocked with an approval card whose footer is:
Not in allowlist: git add, git commit
Buttons offered: Skip / Always Run git / Run, plus an Auto-review dropdown on the card.
That is allowlist-miss UX, not useful Auto-review. Last week (pre-update), Auto-review was smart enough to allow these obvious, user-requested actions without requiring every git subcommand on a static allowlist.
Related broken behavior (same system):
PowerShell command chains (;, if ($LASTEXITCODE…), pipelines) get tokenized into junk allowlist candidates ($LASTEXITCODE, .Count, -not, $dir, exists, etc.).
Clicking Always Run pollutes Command Allowlist with those fragments instead of treating the action as a normal git / dotnet operation.
Result: Auto-review feels like a brittle allowlist with zero of the previous “smart” path.
Steps to Reproduce
Settings → Agents → Approvals & Execution → Run Mode = Auto-Review (confirmed selected).
Windows + PowerShell agent shell.
Ask the agent to stage specific files and create a commit (explicit user request).
Agent proposes something like:
git add ; git status --short; git commit --trailer “Co-authored-by: Cursor [email protected]” -m “…”
Observe approval card: Not in allowlist: git add, git commit — must Skip / Always Run git / Run.
Optionally: approve similar PowerShell probes via Always Run and watch Command Allowlist fill with fragments (.Count, $LASTEXITCODE, -not, variable names, etc.).
Expected Behavior
With Run Mode = Auto-review:
Allowlist hit → run.
Else sandbox when possible.
Else classifier reviews context (user asked to commit; command is ordinary git add / git status / git commit) → auto-run without a human prompt.
Allowlist miss for git commit should not by itself become a blocking approval card for an explicitly requested commit.
Actual Behavior
Classifier path effectively fails closed / never absorbs the action. UI surfaces Not in allowlist: git add, git commit and waits for human approval / Always Run git, even though Auto-review is the active Run Mode.
Screenshots / Screen Recordings
Operating System
Windows 10/11
Version Information
Version: 3.20.21 (user setup)
VS Code Extension API: 1.128.0
Commit: f09fca384ceca23f7bf21f9c23655b162641d740
Date: 2026-09-13T19:02:51.781Z
Layout: IDE
Build Type: Stable
Release Track: Default
Electron: 42.10.0
Chromium: 148.0.7778.280
Node.js: 24.18.1
V8: 14.8.178.38-electron.0
xterm.js: 6.1.0-beta.291
OS: Windows_NT x64 10.0.26200
Additional Information
Need a solution before my whole team just turns on YOLO mode. I went from a workflow doing work independently to having to constantly watch it if it got blocked waiting for approval approval to run silly things like “Exists” and wanting to add them to the allow list ($dir = “some dir”; if (Test-Path -LiteralPath $dir) { Write-Host “exists” } else { Write-Host “not exists” }).
With many background subagents running that have absolutely no indicator of waiting other than click into every one of them cursor IDE becomes unusable without YOLO mode.
With no working way to have a workspace level allow list that’s sharable to the team, it’s a very annoying excercie of “hey everyone, please copy this allowlist every time something changes”.
Does this stop you from using Cursor
No - Cursor works, but with this issue
