[Linux desktop] Codex helpers exhaust fs.inotify.max_user_instances while using zero watches

What version of the Codex App are you using (From “About Codex” dialog)?

26.810.52044

What subscription do you have?

Plus

What platform is your computer?

Fedora 44 KDE latest

What issue are you seeing?

<p data-pm-slice="1 1 []"><span>After using the Codex desktop app, KDE began displaying this warning:</span></p><blockquote><p><span>You have too many applications wanting to monitor file changes! When the capacity is exhausted it will prevent further file monitoring from working correctly. Either close some applications or increase the limit. Currently using 114% of instances and 0% of watches.</span></p></blockquote><p><span>Process inspection shows that the overwhelming majority of the inotify usage comes from ChatGPT/Codex processes. In particular, 15 Codex </span><code><span>cua_node/bin/node_repl</span></code><span> helper processes each hold eight </span><code><span>anon_inode:inotify</span></code><span> file descriptors, despite reporting zero active watches.</span></p><p><span>This appears to exhaust the system’s </span><code><span>fs.inotify.max_user_instances</span></code><span> limit and may prevent unrelated applications from monitoring file changes.</span></p><h2><span>Observed behavior</span></h2><p><span>At the time of inspection:</span></p>
Component | Processes | Inotify file-descriptor references | Watches
-- | -- | -- | --
/usr/lib/chatgpt/resources/cua_node/bin/node_repl | 15 | 120 total, 8 each | 0
Main ChatGPT process | 1 | 10 | 0
Codex app server | 1 | 10 | 0
codex-code-mode-host | 1 | 8 | 0
ChatGPT network service | 1 | 1 | 0

<p><span>ChatGPT/Codex processes therefore accounted for approximately 149 inotify file-descriptor references at inspection time. Some descriptors could refer to shared underlying inotify objects, but Codex is clearly the dominant consumer.</span></p><p><span>The configured system limit is:</span></p><pre><code><span>fs.inotify.max_user_instances = 128
fs.inotify.max_user_watches = 267284
fs.inotify.max_queued_events = 16384</span></code></pre><p><span>The desktop warning reported 114% instance usage and 0% watch usage.</span></p><p><span>Other desktop applications were using comparatively few instances. For example, Chrome used six across its inspected processes, while individual KDE components generally used between one and four.</span></p><h2><span>Expected behavior</span></h2><p><span>Codex helper processes should release inotify instances when the associated task, runtime, or code-mode operation no longer requires them.</span></p><p><span>Normal use of Codex should not exhaust the operating system’s per-user inotify instance limit, particularly when none of the instances have active watches.</span></p><h2><span>Actual behavior</span></h2><p><span>Multiple </span><code><span>node_repl</span></code><span> helper processes remain running and each retains eight inotify descriptors with zero watches. Their combined usage exhausts the system limit and triggers a desktop warning.</span></p><p><span>Once the limit is exhausted, file monitoring in Codex or unrelated applications may stop working correctly.</span></p><h2><span>Steps leading to the issue</span></h2><p><span>I do not yet have a minimal deterministic reproduction. The observed sequence was:</span></p><ol data-spread="false" start="1"><li><p><span>Install and run the Codex/ChatGPT Linux desktop app.</span></p></li><li><p><span>Use multiple Codex tasks normally.</span></p></li><li><p><span>Leave the application running as tasks and code-mode helpers are created.</span></p></li><li><p><span>Eventually receive the KDE inotify-capacity notification.</span></p></li><li><p><span>Inspect </span><code><span>/proc/&lt;pid&gt;/fd</span></code><span> and </span><code><span>/proc/&lt;pid&gt;/fdinfo</span></code><span>.</span></p></li><li><p><span>Observe 15 Codex </span><code><span>node_repl</span></code><span> helpers, each retaining eight inotify descriptors and zero watches.</span></p></li></ol><p><span>The notification first appeared on August 18, 2026. The installed package was installed on August 16, 2026.</span></p><h2><span>Environment</span></h2><ul data-spread="false"><li><p><span>Codex/ChatGPT package: </span><code><span>chatgpt-26.810.52044-1.x86_64</span></code></p></li><li><p><span>App version: </span><code><span>26.810.52044</span></code></p></li><li><p><span>Package source: OpenAI RPM</span></p></li><li><p><span>OS: Fedora Linux 44</span></p></li><li><p><span>Architecture: x86_64</span></p></li><li><p><span>Kernel: </span><code><span>7.1.8-200.fc44.x86_64</span></code></p></li><li><p><span>Desktop: KDE Plasma </span><code><span>6.7.4</span></code></p></li><li><p><span>Display session: Wayland</span></p></li><li><p><span>CPU: AMD Ryzen 7 9700X</span></p></li><li><p><span>Memory: 30.5 GiB</span></p></li><li><p><span>glibc: </span><code><span>2.43-8.fc44</span></code></p></li></ul><h2><span>Relevant process sample</span></h2><pre><code><span>INSTANCES WATCHES APPLICATION
10 0 /usr/lib/chatgpt/ChatGPT
10 0 /usr/lib/chatgpt/resources/codex ... app-server
8 0 /usr/lib/chatgpt/resources/codex-code-mode-host
8 0 /usr/lib/chatgpt/resources/cua_node/bin/node_repl
8 0 /usr/lib/chatgpt/resources/cua_node/bin/node_repl
8 0 /usr/lib/chatgpt/resources/cua_node/bin/node_repl
... 15 node_repl processes in total
1 0 /usr/lib/chatgpt/ChatGPT --type=utility --utility-sub-type=network.mojom.NetworkService</span></code></pre><p><span>The values were obtained by finding file descriptors whose targets were </span><code><span>anon_inode:inotify</span></code><span>, then counting </span><code><span>inotify</span></code><span> entries in the corresponding </span><code><span>/proc/&lt;pid&gt;/fdinfo/&lt;fd&gt;</span></code><span> files.</span></p><h2><span>Possible cause</span></h2><p><span>This may be a lifecycle or cleanup issue involving code-mode </span><code><span>node_repl</span></code><span> helpers. Each helper appears to allocate eight inotify instances even though it has no active watches, and multiple helpers accumulate during app use.</span></p><p><span>Possibilities include:</span></p><ul data-spread="false"><li><p><span>Completed or inactive task runtimes not being terminated.</span></p></li><li><p><span>Inotify instances not being closed when a helper becomes idle.</span></p></li><li><p><span>Inotify descriptors being inherited unnecessarily by child processes.</span></p></li><li><p><span>A fixed group of monitoring descriptors being created for every helper process.</span></p></li></ul><p><span>This is only a hypothesis based on the process inspection.</span></p><h2><span>Workarounds</span></h2><p><span>Potential temporary workarounds are:</span></p><ul data-spread="false"><li><p><span>Completely restart the ChatGPT/Codex desktop app.</span></p></li><li><p><span>Increase </span><code><span>fs.inotify.max_user_instances</span></code><span>.</span></p></li></ul><p><span>Increasing the kernel limit would only mask the accumulation, however, and could allow the same behavior to consume a larger number of instances over time.</span></p><p><span>I have not yet confirmed whether restarting the application permanently resolves the issue or how quickly the helpers accumulate again.</span></p>

What steps can reproduce the bug?

Steps leading to the issue

Install and run the Codex/ChatGPT Linux desktop app.

Use multiple Codex tasks normally.

Leave the application running as tasks and code-mode helpers are created.

Eventually receive the KDE inotify-capacity notification.

Inspect /proc/<pid>/fd and /proc/<pid>/fdinfo.

Observe 15 Codex node_repl helpers, each retaining eight inotify descriptors and zero watches.

The notification first appeared on August 18, 2026. The installed package was installed on August 16, 2026.

What is the expected behavior?

Expected behavior

Codex helper processes should release inotify instances when the associated task, runtime, or code-mode operation no longer requires them.

Normal use of Codex should not exhaust the operating system’s per-user inotify instance limit, particularly when none of the instances have active watches.

Additional information

Report generated by gpt5.6 sol but verified by me ( human)