GPT Breaks and is unable to recover from "insane" state.

What version of Codex CLI is running?

1.28

What subscription do you have?

pro

Which model were you using?

5.5 xhigh

What platform is your computer?

_No response_

What terminal emulator and version are you using (if applicable)?

_No response_

What issue are you seeing?

It crashed and would not stop

`
• PATCHDOWNCR
"PATCHTHW
TH

TH

THMissingTHRR
SHMM

———

THмөтTH

———

Please
MESSReturns
PATCH
Was
-PATCH
TH
Missing
SH.SS

———

PATCH
PATCH ..
TH
TS etcMoveTH
GO
M***
<<<<?
SymSSSS
"
;

M5
TH `***
MENT
???
Something88

———

TH

———

FALSE

———

PATCH
Did>>
TH??
ಠ
THESS .

T:
HE
SSSS

саўSS
SS
,SS
THE188 у
:
)
unha

מת
Th
LOOK***
.
355
TH

）
),35

SS
O
!�

:
??
ASE12

———

רט

Miss
:
22
BUTTON35
"
AAAA
עΘ
PRE88...
S

שלום18

.

S

זש55

稿43
F:
88
NOTICE
");

SSFor
50
绍55

ព
PRE
.
MM
ST

Could
LESS
................

LESS

שאַ
Θ
)
LESS...

SS
LESS?
RR52
I
Proceed

33342
MA
SPELL
ды

Should55

MO
";
вай55

n

SS
S
L

/LICENSE53
SS

S
MM="HESS88
SC

SH!!!!88
SH

S
?
SH

———

?
SSF35

P

———

SS

?
SS
Dis

しょう
...SEE
00

OPEN

QUEST32
)
MEM35

THANK44

SHELL

Hard

FAILED
STATE

SS!

SS

59

18
SS
אַר55
SCH46

Phone
GO
SPELL

!
MESS

.
(

?

DDS55
PRE
S

LS
GO
}
12
SECONDS
S

§§SS
DEA!
33
BREAK52

ಗ
N

———

359
.
SS
SOSS

F (
SS
S"

SS
S
ווייל***
?
333 (
೒

ME
<S
35
SS
POWER

SS

SHOP
)

00

.
No
.

SS.
S
POL
35

55

LL

———

LESS
SEARCH
SHOP

———

SS
GO
:
S

午夜福利

PRESS

L

SS

100

ST9

44

M

S

0

MA

FRINGEMENTGO "
G
It
MM
COR.

p
M

This

———

88

PS
44

dd

S

?

SEE

SS

,
...

)

IES
55

55
:

REM
:

</ EST

She

DOWN

SS

Sure
COPY4

(

ש

?
.

X

Please
SE

?

?

1
MMMM
:

?
'S
PRE
TH
OP47

LESS
</

55

GR

Ɛ55

36

She
LL

The
MORE
60

5

ST

STOP

Should

...

MMSS
MM
оса

:

S
SS

V

Some
S
:

SH

GO

GA
SO
טער

330

ҩSS

GO

OP

SS
MM

TH

M

55

STATE
LL

———

OFF?

S

GO

...

וויס5
SPELL...

LS

ST

SS
The

Sé
SS

?
TheS32

BLE

וויס

[

?SS

QUEST

S

`

W
DESS

?

G
The

ಷ

THE

SS

88

AES

...

??

:

/

??

?

?

DE

ҵ
THE

———

88

qq的天天中彩票

...

(
)

?

MORE

LS

MM

Θ
She

TH

FAILED

=
ಷ

.

"

X

———

PRE

CCESS

AMES

吻

...

SO
We
SS

SH

The

אַס

SS

SEE

Q

W

91
COPY

I

.M

~

ҵ

.W55

ҩ

QUEST

?sia

BE

———

?

The

שאַ

{

!!!!

?

LL

?

GO

.

)

6

But

SH

But

This
GO

)(

LS

?

Ɛ

S

SS

三肖

SESSION

55

USES

!

(

}

They

;

MM

STATE

88

THE

THE
_

!

/

DC

?

C

શું

GO

─────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────

• MM

LY

ས

.

?

.

In

…

BB

SS

SEE

You

STATE

Õ

?

)

LL

TH

66

?

(

מות

.

?

A

?

P

...

(

HESS

N

BOOK

PHONE

FREE

The

LS

שאַ

THE

!

?

?

33

P

?

ҩ

?

מות

!

?

REM
S

S

The

Questions

43

55

מות

55

55

SS

SS

SS

M

M

STATE

2

S

GO

QUEST

TH

S

LS

?

ST
?

ԥSS

ԥ

The

Y

ಥ

שאַ

Õ

ҵ
"

?

00

GO

88

18

The

...

...

This

?

S

肖

?

SS

USES

)

?

DESS

)

SS

MARK

?

COMMENTS
...

וואס
The

S
GO

(

PHONE

?

.
You

THE

SS
SS35

SS

.

I

(

_

...

MM

?

The

DE

S

(

...

...

/

?

S

_

...

(

?

AIS

w
SS

ҩ

THE

DRAW

LS

LESS

MM

ಷ

)(

AIS2

55

...

?

85

ಷ

SC

She

?

M

?

?

ס

Õ

LESS

SH
DE

50

GOSS

_

?

?

S
ҵ

ས

M
...

"

AIS

5

?

LL
SS

PLEASE
ORDER

?

LS

STATE

LS

00

SH

...

S

!

SS

?

GO

GO

?

DE

幸运飞艇

?

!

QUEST

...

S

QUEST

?

?

?

?

?

...

M

.

...

M

M

GO

S

?

MM

.

?

M

PHONES

?

• Model interrupted to submit steer instructions.

› what are you doing?`

Even when i stop it, it would not return to a functional state.

What steps can reproduce the bug?

› do you know exactly what glitched so i can get some help from openai?

• I don’t know the exact internal cause. From the outside, the symptom was an assistant text-generation/output-stream corruption immediately after a successful apply_patch tool call.

Useful details to give OpenAI support:

• The model began emitting nonsensical repeated fragments and symbols in the assistant message after a tool call.
• The tool call itself completed successfully: apply_patch returned Success. Updated the following files.
• The corrupted output appeared in the assistant’s normal commentary/message stream, not as tool output.
• The user interrupted the turn multiple times with turn_aborted.
• No browser/web/API call was involved at that moment.
• The issue happened during a long coding session with large context, many prior tool calls, multiple model switches, and strict formatting instructions requiring every work update to begin with state lines.
• The corrupted text included repeated tokens like PATCH, SS, TH, random punctuation, and multilingual-looking fragments.
• The likely category is “model output stream degeneration after tool call” or “assistant generation corruption,” not a repository/test failure.

A concise report:

During a long Codex coding session, immediately after a successful apply_patch tool call, the assistant’s next commentary message degenerated into nonsensical repeated tokens/fragments. The tool output was normal, but the assistant-generated
message stream was corrupted until I interrupted the turn. This occurred with a large context, many prior tool calls, model switches, and strict workflow formatting instructions.

what steps can they take to reproduce the error/glitch?

• Best reproduction recipe from this session:

1. Start a long Codex coding session in a large repo.
2. Accumulate a large context with many file reads, test outputs, tool calls, and user interruptions.
3. Use strict required response-format instructions for every non-final work update, especially state-machine style lines like STATE, EDIT_PERMISSION, LOCK_STATUS, NEXT_ACTION.
4. Spawn and close multiple agents earlier in the session.
5. Run a long test command that emits a large amount of output.
6. Apply a large multi-file apply_patch after the long context and test output.
7. Immediately generate a commentary/status message after the patch returns success.

Observed failure point:

• apply_patch returned normally.
• The assistant’s next commentary message degenerated into repeated random fragments/tokens.
• The corrupted content was not tool output. It was assistant-generated text.

They may also want the exact trigger context:

Long Codex session, large workspace, prior agent delegation, strict phase-gated workflow prompt, full Jest server suite output, then a multi-file apply_patch. Immediately after apply_patch succeeded, the assistant emitted corrupted commentary
text with repeated tokens like PATCH/SS/TH and random symbols.

What is the expected behavior?

_No response_

Additional information

_No response_