nuke-cursor-app

Compare original and translation side by side

🇺🇸

Original

English

🇨🇳

Translation

Chinese

Nuke Cursor App

Why this exists

为什么开发这个工具

Cursor (the AI IDE) is an Electron-based VS Code fork. It has a known, Cursor-acknowledged memory leak: the renderer process accumulates tool-call state (diffs, file contexts) during long agent sessions and never frees it, especially in the Agents window. The UI gets laggy, then freezes. The only reliable recovery is a full restart — that is what this skill does, end to end: snapshot → kill → relaunch → before/after report.

Cursor（一款AI IDE）是基于Electron的VS Code分支版本。它存在一个官方已确认的内存泄漏问题：在长时间的Agent会话期间，渲染器进程会不断累积工具调用状态（差异、文件上下文），且永远不会释放这些内存，尤其是在Agents窗口中。UI会逐渐卡顿，最终冻结。唯一可靠的恢复方法是完全重启——这正是本工具的功能，端到端完成：快照→终止进程→重新启动→报告前后状态。

How Cursor runs

Cursor的运行方式

One main process at

/Applications/Cursor.app/Contents/MacOS/Cursor

plus helper processes (Renderer, GPU, extension host, network service, crashpad) that all live under the

/Applications/Cursor.app

bundle path.

一个主进程位于

/Applications/Cursor.app/Contents/MacOS/Cursor

，加上辅助进程（Renderer、GPU、扩展宿主、网络服务、crashpad），所有进程都属于

/Applications/Cursor.app

包路径。

Safety rules

安全规则

Match processes ONLY by the bundle path
```
/Applications/Cursor.app
```
— never by the bare word "cursor".
Do NOT touch
```
CursorUIViewService
```
— despite the name it is a macOS system text-input service, not part of the Cursor app.
Warn the user first if you have reason to think an important agent run is in flight; killing Cursor kills its local agent sessions.
The metrics collector OWNS
```
cursor-metrics.sqlite3
```
. Read it ONLY with
```
sqlite3 -readonly
```
. Never write to it.
The snapshot is best-effort: if the DB is missing or a query fails, note that in the log and continue — never block the nuke on it.

仅通过包路径
```
/Applications/Cursor.app
```
匹配进程——绝不要仅通过“cursor”这个词匹配。
不要触碰
```
CursorUIViewService
```
——尽管名称相似，但它是macOS系统文本输入服务，不属于Cursor应用。
如果有理由认为正在运行重要的Agent任务，需先警告用户；终止Cursor会同时终止其本地Agent会话。
指标收集器拥有
```
cursor-metrics.sqlite3
```
文件。仅以
```
sqlite3 -readonly
```
模式读取它，绝不要写入。
快照为尽力而为：如果数据库缺失或查询失败，在日志中记录该情况并继续——绝不要因快照问题阻止终止操作。

Procedure

操作流程

1. Snapshot BEFORE killing

1. 终止前的快照

bash

DB="$HOME/Library/Application Support/<metrics-collector>/cursor-metrics.sqlite3"
LOG_DIR="$HOME/Library/Application Support/<metrics-collector>/nuke-logs"
mkdir -p "$LOG_DIR"
LOG="$LOG_DIR/$(date +%Y-%m-%d-%H%M).md"

bash

DB="$HOME/Library/Application Support/<metrics-collector>/cursor-metrics.sqlite3"
LOG_DIR="$HOME/Library/Application Support/<metrics-collector>/nuke-logs"
mkdir -p "$LOG_DIR"
LOG="$LOG_DIR/$(date +%Y-%m-%d-%H%M).md"

Live per-process readout — these are the "before" numbers

实时进程读取——这些是“重启前”的数据

ps -axo pid,rss,comm | grep "/Applications/Cursor.app" | grep -v grep

Totals for the summary line (rss is in KB on macOS)

摘要行的总计数据（macOS上rss单位为KB）

ps -axo rss,comm | awk 'index($0, "/Applications/Cursor.app") {n++; s+=$1} END {printf "%d processes, %.1f GB", n, s/1048576; print ""}'

Last 30 min from the collector (read-only)

收集器中最近30分钟的数据（只读）

sqlite3 -readonly "$DB" "SELECT datetime(timestamp,'unixepoch','localtime'), role, process_count, ROUND(cpu_percent,1), ROUND(resident_bytes/1073741824.0,2) || ' GB' FROM raw_samples WHERE timestamp > strftime('%s','now') - 1800 ORDER BY timestamp;" sqlite3 -readonly "$DB" "SELECT datetime(timestamp,'unixepoch','localtime'), kind, role, reason, memory_pressure FROM events WHERE timestamp > strftime('%s','now') - 1800 ORDER BY timestamp;"


Write all four outputs into `$LOG` with headings: process list, totals,
metrics (last 30 min), events (last 30 min). The after-restart reading is
appended in step 4 — so one file tells the whole story of this nuke.


将以上四个输出内容写入`$LOG`文件，并添加标题：进程列表、总计数据、指标（最近30分钟）、事件（最近30分钟）。重启后的读取数据会在步骤4中追加到该文件——因此一个文件即可记录本次终止操作的完整过程。

2. Kill every Cursor process

2. 终止所有Cursor进程

bash

undefined

bash

undefined

Graceful quit first — lets Cursor save session state

先尝试优雅退出——让Cursor保存会话状态

osascript -e 'tell application "Cursor" to quit' 2>/dev/null sleep 3

Kill anything still alive under the bundle path

终止所有仍在运行的包路径下的进程

pkill -f "/Applications/Cursor.app" 2>/dev/null sleep 1

Verify; force-kill leftovers by PID if needed

验证；如有残留则通过PID强制终止

ps -axo pid,comm | grep "/Applications/Cursor.app" | grep -v grep

if any remain: kill -9 <pid> ...

如果有残留： kill -9 <pid> ...

Final check — must print "all Cursor processes gone"

最终检查——必须输出“all Cursor processes gone”

ps -axo pid,comm | grep "/Applications/Cursor.app" | grep -v grep && echo "STILL RUNNING" || echo "all Cursor processes gone"


The graceful quit often fails exactly when this skill is needed — a leaked
renderer blocks the main thread — which is why the pkill/kill steps exist.
Note in the log whether graceful quit worked or force-kill was needed.

ps -axo pid,comm | grep "/Applications/Cursor.app" | grep -v grep && echo "STILL RUNNING" || echo "all Cursor processes gone"


当本工具需要运行时，优雅退出往往会失败——泄漏的渲染器会阻塞主线程——这就是为什么需要pkill/kill步骤。在日志中记录优雅退出是否成功，或者是否需要强制终止。

3. Relaunch and verify

3. 重新启动并验证

Skip this step only if the user asked to keep Cursor closed ("nuke cursor and keep it closed").

bash

sleep 2
open -a Cursor

仅当用户要求保持Cursor关闭（"nuke cursor and keep it closed"）时，才跳过此步骤。

bash

sleep 2
open -a Cursor

Wait up to 15 s for the main process to come back

等待最多15秒，直到主进程恢复

for i in $(seq 1 15); do ps -axo comm | grep -q "/Applications/Cursor.app/Contents/MacOS/Cursor" && break sleep 1 done ps -axo pid,comm | grep "/Applications/Cursor.app/Contents/MacOS/Cursor" | grep -v grep


If the poll times out, report that — never claim Cursor is back without
seeing the main process.


如果轮询超时，需报告该情况——在未检测到主进程的情况下，绝不要声称Cursor已恢复。

4. After-restart reading

4. 重启后的读取数据

bash

sleep 10   # let Cursor settle and restore windows

ps -axo pid,rss,comm | grep "/Applications/Cursor.app" | grep -v grep
ps -axo rss,comm | awk 'index($0, "/Applications/Cursor.app") {n++; s+=$1} END {printf "%d processes, %.1f GB", n, s/1048576; print ""}'

Append both outputs to

$LOG

under an "after restart" heading.

bash

sleep 10   # 让Cursor稳定并恢复窗口

ps -axo pid,rss,comm | grep "/Applications/Cursor.app" | grep -v grep
ps -axo rss,comm | awk 'index($0, "/Applications/Cursor.app") {n++; s+=$1} END {printf "%d processes, %.1f GB", n, s/1048576; print ""}'

将这两个输出内容追加到

$LOG

文件的“重启后”标题下。

5. Report to the user

5. 向用户报告

End with one plain-English summary line, for example:

Killed 14 processes using 21.3 GB. Cursor is back with 9 processes using 2.1 GB. Snapshot saved to nuke-logs/2026-08-14-1832.md.

Never claim success without the step 2 final check and (unless skipped) the step 3 relaunch check.

最后输出一条简洁的英文总结，例如：

Killed 14 processes using 21.3 GB. Cursor is back with 9 processes using 2.1 GB. Snapshot saved to nuke-logs/2026-08-14-1832.md.

在未完成步骤2的最终检查和（除非跳过）步骤3的重启检查的情况下，绝不要声称操作成功。

Known failure modes

已知失败模式

After a force-kill, Cursor may show a "restore windows?" dialog on relaunch. That is expected — tell the user to click Restore.
If the metrics DB has no recent rows (collector stalled), say so in the log and the report; the
```
ps
```
readouts still give valid before/after.

强制终止后，Cursor在重启时可能会显示“是否恢复窗口？”的对话框。这是预期情况——告知用户点击“恢复”即可。
如果指标数据库没有最近的记录（收集器停滞），需在日志和报告中说明；
```
ps
```
读取的数据仍能提供有效的重启前后对比。