I encountered a strange issue with Codex scheduled tasks and would like to report it.
When running an automated Codex task, the normal progress report is sometimes followed by internal-looking heartbeat metadata and unrelated Chinese words.
Example:
15 NOTIFY
/ 久久
In another run, similar random Chinese gambling-related text appeared after the heartbeat block, such as “天天彩票”.
The task output itself is normal, but these additional contents are unrelated to my project, instructions, files, or execution results.
The , <automation_id>, and fields look like internal automation metadata that should probably not appear in the user-visible response. The random Chinese words appearing after this block also look unusual and unrelated.
This seems similar to the previously reported Codex issue where random Chinese lottery/gambling-related references appeared after heartbeat prompts.
Could this be a known issue with Codex Automations, such as:
- Internal automation metadata leaking into the final response;
- Context boundary issues between the scheduler and the model;
- Unexpected token generation after the heartbeat message?
I can provide additional logs or reproduction details if needed.
Welcome to the community!
As you noted, this is a common occurrence in the current generation of models. This affects not only OpenAI, but also Gemini and other vendors.
One could hypothesize what could cause this.
In any case, I suspect that what you’re seeing is more hallucination than any specific metadata. In any case, an unjailbroken leak (if it was one) is a sysmptom of an upstream token hallucination.
Sometimes this noise is unavoidable and requires re-generation. Often cleaning out/reducing the context helps. Switching models might also help.
what model were you using, and how big was the context? If it has been compacted multiple times I suspect the propensity for this kind of thing increases as well, but this is just an armchair hypothesis. I personally nix the history before or when compaction happens.