rxbyes

[rxbyes @rxbytes]

@thsottiaux reset

[rxbyes @rxbytes]

以前订阅 20x 将在 10月30日后彻底调整，是时候考虑 c c 的时候了。

[rxbyes @rxbytes]

@ItsRealAlec yes

[rxbyes @rxbytes]

# Title

macOS ChatGPT app: bundled `cua_node` helper processes (`node_repl`, codex-app-tools `server.mjs`, unified-computer-use `launch.mjs`) are never reaped — 335 processes / ~3.5 GB RSS after one day of agent sessions

# Summary

The ChatGPT macOS desktop app leaks helper processes spawned by its bundled `codex app-server`. Every agent-mode / computer-use / code-mode session spawns a set of helper processes, and they are never terminated after the session ends. After roughly one workday of normal use without restarting the app, I accumulated **335 leaked processes consuming ~3.5 GB of RSS**, all parented to a single long-running `codex app-server` process. Quitting and relaunching the app clears them (count drops to ~10), confirming they are app-session leftovers rather than user-launched software.

# Environment

- App: ChatGPT for macOS, version **26.903.71938**
- Bundled agent runtime: `/Applications/ChatGPT.app/Contents/Resources/codex -c features.code_mode_host=true app-server ...` (the "codex app-server")
- Bundled plugins involved: `codex-app-tools` (runs `./server.mjs`), `unified-computer-use` version `26.903.71938` (runs `scripts/launch.mjs`), plus `node_repl` REPL helper processes
- Node runtime used by helpers: `/Applications/ChatGPT.app/Contents/Resources/cua_node/bin/node`
- OS: macOS 26.5.1 (Build 25F80), Apple Silicon (arm64)

# Steps to reproduce

1. Launch the ChatGPT macOS app and keep it running.
2. Use agent mode / computer use / code-mode sessions repeatedly throughout a day (each session spawns helpers such as `server.mjs`, `launch.mjs`, `node_repl`).
3. Inspect the process tree, e.g.:
`ps -axo pid,ppid,lstart,rss,command | grep cua_node`

# Actual behavior

- Helper processes accumulate monotonically over time and are never reaped when their session ends.
- Observed after ~18 hours of app uptime (app launched 2026-09-11 23:43):
- **335** `cua_node` processes in total, of which **249** were direct children of the single `codex app-server` PID
- Breakdown: ~166 × `node_repl`, ~84 × codex-app-tools `server.mjs`, ~83 × unified-computer-use `launch.mjs`
- Combined RSS ≈ **3.5 GB**; per-process CPU time near zero (idle but resident)
- Spawn timestamps spread across the entire day (e.g. 00:28, 03:51, 12:06, 15:51, 17:45 local time), i.e. one leak per session, including sessions finished many hours earlier.
- After quitting and relaunching the app, the count dropped to ~10 processes / ~119 MB, confirming the app-server is the parent and the leak is bounded only by app restarts.

Sample lines (home directory redacted):

```
PID PPID STARTED TIME COMMAND
63203 19921 Fri Sep 12 15:11:00 2026 0:00.32 .../cua_node/bin/node ./server.mjs
63202 19921 Fri Sep 12 15:11:00 2026 0:00.19 .../cua_node/bin/node .../unified-computer-use/26.903.71938/scripts/launch.mjs
63244 63221 Fri Sep 12 15:11:00 2026 0:00.21 .../cua_node/bin/node_repl
```

# Expected behavior

Helper processes should be terminated when the session or tool call that started them ends (e.g. the MCP server exits on client disconnect / stdin EOF, and its children are reaped). Long-running app uptime should not cause unbounded process/memory growth.

# Impact

- ~3.5 GB of wasted memory and hundreds of extra processes within a single workday; each Node.js process also carries ~10 OS threads, so this shows up as heavy thread pressure in Activity Monitor and degrades overall system responsiveness.
- Unbounded growth: the longer the app stays open (common on macOS), the worse it gets.

# Workaround

Periodically quit and relaunch the ChatGPT app, then remove any leftovers with a pattern that only matches the app's own helpers:

```
pkill -f 'https://t.co/OCl42pYgXU…'
```

(Use with care; it intentionally does not match unrelated `node` processes such as local dev servers.)

[rxbyes @rxbytes]

# Title

macOS ChatGPT app: bundled `cua_node` helper processes (`node_repl`, codex-app-tools `server.mjs`, unified-computer-use `launch.mjs`) are never reaped — 335 processes / ~3.5 GB RSS after one day of agent sessions

# Summary

The ChatGPT macOS desktop app leaks helper processes spawned by its bundled `codex app-server`. Every agent-mode / computer-use / code-mode session spawns a set of helper processes, and they are never terminated after the session ends. After roughly one workday of normal use without restarting the app, I accumulated **335 leaked processes consuming ~3.5 GB of RSS**, all parented to a single long-running `codex app-server` process. Quitting and relaunching the app clears them (count drops to ~10), confirming they are app-session leftovers rather than user-launched software.

# Environment

- App: ChatGPT for macOS, version **26.903.71938**
- Bundled agent runtime: `/Applications/ChatGPT.app/Contents/Resources/codex -c features.code_mode_host=true app-server ...` (the "codex app-server")
- Bundled plugins involved: `codex-app-tools` (runs `./server.mjs`), `unified-computer-use` version `26.903.71938` (runs `scripts/launch.mjs`), plus `node_repl` REPL helper processes
- Node runtime used by helpers: `/Applications/ChatGPT.app/Contents/Resources/cua_node/bin/node`
- OS: macOS 26.5.1 (Build 25F80), Apple Silicon (arm64)

# Steps to reproduce

1. Launch the ChatGPT macOS app and keep it running.
2. Use agent mode / computer use / code-mode sessions repeatedly throughout a day (each session spawns helpers such as `server.mjs`, `launch.mjs`, `node_repl`).
3. Inspect the process tree, e.g.:
`ps -axo pid,ppid,lstart,rss,command | grep cua_node`

# Actual behavior

- Helper processes accumulate monotonically over time and are never reaped when their session ends.
- Observed after ~18 hours of app uptime (app launched 2026-09-11 23:43):
- **335** `cua_node` processes in total, of which **249** were direct children of the single `codex app-server` PID
- Breakdown: ~166 × `node_repl`, ~84 × codex-app-tools `server.mjs`, ~83 × unified-computer-use `launch.mjs`
- Combined RSS ≈ **3.5 GB**; per-process CPU time near zero (idle but resident)
- Spawn timestamps spread across the entire day (e.g. 00:28, 03:51, 12:06, 15:51, 17:45 local time), i.e. one leak per session, including sessions finished many hours earlier.
- After quitting and relaunching the app, the count dropped to ~10 processes / ~119 MB, confirming the app-server is the parent and the leak is bounded only by app restarts.

Sample lines (home directory redacted):

```
PID PPID STARTED TIME COMMAND
63203 19921 Fri Sep 12 15:11:00 2026 0:00.32 .../cua_node/bin/node ./server.mjs
63202 19921 Fri Sep 12 15:11:00 2026 0:00.19 .../cua_node/bin/node .../unified-computer-use/26.903.71938/scripts/launch.mjs
63244 63221 Fri Sep 12 15:11:00 2026 0:00.21 .../cua_node/bin/node_repl
```

# Expected behavior

Helper processes should be terminated when the session or tool call that started them ends (e.g. the MCP server exits on client disconnect / stdin EOF, and its children are reaped). Long-running app uptime should not cause unbounded process/memory growth.

# Impact

- ~3.5 GB of wasted memory and hundreds of extra processes within a single workday; each Node.js process also carries ~10 OS threads, so this shows up as heavy thread pressure in Activity Monitor and degrades overall system responsiveness.
- Unbounded growth: the longer the app stays open (common on macOS), the worse it gets.

# Workaround

Periodically quit and relaunch the ChatGPT app, then remove any leftovers with a pattern that only matches the app's own helpers:

```
pkill -f 'https://t.co/9WIhac9Cf5'
```

(Use with care; it intentionally does not match unrelated `node` processes such as local dev servers.)

[rxbyes @rxbytes]

征集一下，大家在使用chatgpt或其他Ai工具时，怎么样保证UI设计与UX高度契合，发表你的想法

[Miles Ma @miles_mazy]

https://t.co/0cr9iBGG2Q

[rxbyes @rxbytes]

@benln and me

[rxbyes @rxbytes]

OpenAI 官方确认：明天为所有 Codex 订阅用户重置额度！

划重点时间：太平洋时间 8月24日 14:00 = 北京时间 8月25日 05:00

这次“发福利”是因为官方排查出了 3 个导致额度消耗过快的 Bug：
1️⃣ 长会话中图片多次压缩效率低下
2️⃣ 计算机历史记录（P95+）异常高消耗
3️⃣ 对话标题生成功能超额扣费

专项团队明天就会上线修复补丁。另外，官方还预告下周会开发一套全新的效率提升方案。定好明早 5 点的闹钟，准备满血复活！

#OpenAI #Codex #AI编程 #开发者工具

[Tibo @thsottiaux]

Update on rate limits in Codex. We do see that for some users the cache hit rate has been worse this week than the stable state the weeks before. This could explain that usage is draining somewhat faster for those users as hitting the cache consistently is an important component of being efficient.

We are investigating and will have an update tomorrow.

[Tibo @thsottiaux]

Update on rate limits in Codex. We do see that for some users the cache hit rate has been worse this week than the stable state the weeks before. This could explain that usage is draining somewhat faster for those users as hitting the cache consistently is an important component of being efficient.

We are investigating and will have an update tomorrow.

[rxbyes @rxbytes]

AI 正在从“谁有最好的模型”，进入“谁能以合理成本、安全、可验证地让模型长期完成真实任务”。

第一，Agent 成本治理会成为正式工程领域。 NVIDIA 服务器价格压力和前沿 AI 公司的巨额资本需求说明，未来不能再假定“模型每年自动变强又自动变便宜”。企业需要 Model Router、缓存、Batch、大小模型分层、异构算力和 Cost per Successful Task 指标。

第二，Agent Harness / Model Factory 的战略价值正在上升。 NVIDIA 花巨资获得 Poolside 能力，说明真正长期有价值的不只是模型权重，而是“训练—工具—环境—验证—反馈”的整套生产体系。这与软件研发 Agent、企业 Agent 以及 FutureLens 的方向高度一致：模型可以换，Evidence、Workflow、Evaluation 和 Human Review 才应该成为持久资产。

[rxbyes @rxbytes]

@Xeift1 所以才需要集思广益，发挥人类的想象和技术推演

[rxbyes @rxbytes]

想做个有意思的小测试 👇

你用什么 Prompt，可以判断自己现在的 ChatGPT 到底是不是最新模型？

我先来一个：

«「你是什么模型？请输出当前实际运行的具体模型版本和变体，并告诉我你是不是 mini 版本。」»

大家可以把你们的测试 Prompt + ChatGPT 返回结果贴在评论区。

看看谁能找到一个更靠谱的“模型验明正身”方法 👀

投票：你觉得直接问 ChatGPT，能准确判断底层模型吗？

🔘 能，基本可信
🔘 不能，模型可能不知道
🔘 只能作为参考
🔘 我有更好的 Prompt（评论区见）

#ChatGPT #OpenAI #GPT5 #AI #PromptEngineering

[rxbyes @rxbytes]

@kristaletz @bot thank you！ [email protected]

[rxbyes @rxbytes]

@gkxspace codex app 不是直接可以使用 chatgpt 5.6 pro吗？

[rxbyes @rxbytes]

@abraibrai my account is [email protected]. I haven’t received the reset card yet. Could you please help me with this? I would really appreciate it. Thank you so much!

[rxbyes @rxbytes]

@keco321 有号吗？pro

[rxbyes @rxbytes]

@JohnSmithg7tn @zstmfhy 怎么可以联系到你

[rxbyes @rxbytes]

@Grace_labubu @zstmfhy 还有号吗？

[rxbyes @rxbytes]

@akokoi1 号被封了，有什么稳定的渠道推荐一下，谢谢。