Custom Skill stopped executing in ChatGPT and Codex after August 20, 2026 — possible Skill runtime issue?

ChatGPT Bugs
,
1

Hi everyone,

I am experiencing a serious issue with my custom Skill execution in both ChatGPT and Codex, and I would like to know whether anyone else has encountered the same problem.

Background

I am a ChatGPT Pro user and I created my own custom Skills for structured workflows.

Before August 20, 2026 (Beijing Time), these Skills worked normally in the ChatGPT web interface and Codex app.

The Skills were used repeatedly for complex tasks, including:

  • Processing uploaded files;
  • Following predefined workflow instructions;
  • Translating CSV files;
  • Maintaining strict CSV structures;
  • Applying quality control rules;
  • Generating final output files.

The same Skill packages had been working correctly for many previous runs without any changes.

Timeline

August 20, 2026 (Beijing Time)

Suddenly, Skill execution in ChatGPT web stopped working.

When I invoked my custom Skill:

Expected behavior:

  1. Recognize the Skill;
  2. Load Skill instructions;
  3. Execute the workflow;
  4. Process the files;
  5. Generate the required output.

Actual behavior:

  1. The Skill is recognized;
  2. ChatGPT repeats the Skill description or my task requirements;
  3. It explains what should be done instead of actually executing the workflow;
  4. Sometimes it returns unrelated errors and stops.

The important point is that I did not modify the Skill files before this happened.

The same Skill had worked normally before.

Codex behavior

On August 20:

  • ChatGPT web Skill execution failed;
  • Codex app was still able to execute the Skill.

However, starting August 21:

  • Codex also started showing the same failure;
  • The Skill instructions are no longer properly followed;
  • The agent either repeats the instructions or fails before execution.

Why I think this may not be a Skill file problem

I have tested and observed the following:

  1. The Skill worked correctly for a long time before the issue.
  2. No changes were made to the Skill package before the failure.
  3. The issue appeared first in ChatGPT web and then affected Codex.
  4. The failure behavior is similar across different environments.

This makes me suspect a possible issue with:

  • Skill runtime;
  • Skill loading mechanism;
  • Agent execution layer;
  • Session capability;
  • Recent backend changes or regression.

Current symptoms

The main symptom is:

The Skill can be detected, but it cannot actually execute.

The model behaves as if it only sees the Skill description instead of applying the Skill instructions.

For example:

Instead of processing the uploaded CSV according to the Skill workflow, it responds with:

  • A summary of the task requirements;
  • A restatement of the instructions;
  • A description of what it should do.

It does not enter the execution stage.

Questions

Could this be related to a recent update or regression?

I would like to know:

  1. Has there been any recent change to Custom Skills runtime or execution behavior?
  2. Is there any known issue affecting Skill loading or execution?
  3. Could this be an account/session-level problem?
  4. Are other users experiencing similar problems with their custom Skills?

I can provide:

  • Skill package files;
  • Conversation logs;
  • Screenshots;
  • Error messages;
  • Reproduction steps.

Any feedback would be appreciated.

Thanks.

2

I am experiencing an issue where my custom Skills, which worked normally for a long time, suddenly stopped executing around August 20, 2026 (Beijing Time). In ChatGPT web, the Skill is recognized but instead of following the workflow and performing the task, the model repeatedly explains the instructions or returns errors without execution. At that time, Codex App could still run the same Skills, but starting August 21, the same execution failure appeared in Codex as well. No Skill files were changed before the issue occurred, and the failure happened across different environments, so I suspect this may be related to a recent Skill runtime, Skill loading mechanism, agent execution behavior change, or backend regression rather than the Skill package itself. Has anyone experienced similar problems with Custom Skills recently?

3

I opened a new chat and asked the model which version it was using, but it told me that it was GPT-5.5-mini.

4

I’m seeing a very similar issue, although mine is with Custom GPT Actions rather than Skills.

The pattern is:

  • The configuration/editor side shows the capability correctly.
  • The schema/action definition validates successfully.
  • Authentication is accepted.
  • However, in a new chat session the runtime says the action/tool is unavailable.

Re-saving the API key, updating the GPT, and starting fresh conversations does not resolve it.

This makes me wonder if there is a broader issue with capability/tool manifest propagation or runtime session initialization rather than the individual Skill/Action configuration itself.

In my case this has happened multiple times over the past week and has interrupted an active development workflow, so it seems like there may be a recurring runtime availability issue rather than a one-off setup problem.

5

For the Codex CLI / desktop side of this, “recognized” and “executed” are different records.

A skill can appear in the picker, or be restated in the reply, without Codex loading the playbook. After a run that should have used one custom skill, keep four columns:

skill | listed | invoked | check command + exit

listed = the skill is in the session skill list.

invoked = the retained Codex session jsonl contains `YOUR-SKILL` on a response_item, not merely a read of AGENTS.md.

Synthetic check: session only read AGENTS.md → invoked 0. One commit-message skill block → invoked 1. Built-in skill-creator / skill-installer should not count as a custom skill.

I represent the open-source project Skilled (GitHub av/skilled).

It can fill the invoked column from local Codex history:

skilled calls --source codex --json --no-index

It does not cover ChatGPT web custom Skills, and it does not judge whether the workflow succeeded. If history was cleared or the format is unrecognized, 0 is inconclusive. The four columns work without installing anything.

If you still have the Codex sessions from 20–21 Aug, a sanitized listed=yes / invoked=no row would distinguish a runtime miss from “the model talked about the skill but never loaded it.”