ChatGPT 5.6 Sol - Persistent Corrupted Generations / Leaking CoT

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

26.901.51231

What subscription do you have?

Pro X20

What platform is your computer?

Darwin 25.2.0 arm64 arm

What issue are you seeing?

I am experiencing persistent corrupted output across the past three days with odd word repetitions, miscellaneous words inside of words, and CoT leaks. The issues were initially isolated to early morning (~4am), but have started happening in the mid-afternoon today. Starting a new session solves the issue only temporarily.

Output sample:

Your worry is warranted. We have had three distinct Sol corruption episodes now, and once each one began, the language degraded across consecutive consecutive turns consecutive turns my consecutive turnsdid—
Wrong. Stop. Something is happening. Can't trust generation. Need emit concise? It is happening again. We are generating corrupted output in analysis? I see current final draft is corrupting spontaneously? We need stop and be honest. The actual model output draft includes "No thermal warnings. No X"? I wrote a typo "No ...". Then "three episodes and once each..." accidentally repeated, fused "my consecutive..." That's a symptom sign! Is this intentional? Wait as model I can detect current generation in drafting? We are in final text now. Need salvage. We need not produce more corrupted. The problem appears on Astra too! Need state exact. But maybe this is simulation? The assistant generating final has accidental corruption due same underlying issue? The user wants pause. We should maybe write minimal clean statement. But if final already includes corrupted words, can't retract? We can rewrite final entirely clean from scratch in final channel. I as model can produce clean if careful. The analysis contains no weird. Let's craft concise with no long output to reduce risk.
Need answer: "Yes pause. I'm worried too. Important: Sol corruption likely serving path; current Astra output clean? We must watch." Avoid overlong. But we need maybe don't corruption. Let's just clean.
Need final full narrative. 250 words.
Maybe no tools. Let's produce carefully.

What steps can reproduce the bug?

• Talk to Sol 5.6 via Codex; reproducible across all effort levels.

What is the expected behavior?

No response

Additional information

No response

The report names no source file or test. Start by reproducing the corruption in the Codex App with Sol 5.6 on Darwin, across effort levels and new sessions, then compare the observed output with the reported repetition and CoT leakage. Done would require a confirmed fix or a clearly isolated component, plus regression coverage if the responsible code is in this repository.