Keep throwing "I'm sorry, but I cannot assist with that request"

Running into the same issue.

Using own Azure Foundry gpt-5.4 model.

I've been experiencing the same issue also using Codex CLI and Azure Foundry gpt-5.4 model.

The reason for these extremely excessive errors is Stream disconnected before completion: Incomplete response returned, reason: content_filter. But this didn't use to happen last week. I've changed the content policy from DefaultV2 to Default but the issue seems to persist.

If you haven't already done so, please report the issue to Azure.

The reason for these extremely excessive errors is Stream disconnected before completion: Incomplete response returned, reason: content_filter. But this didn't use to happen last week. I've changed the content policy from DefaultV2 to Default but the issue seems to persist.

@matys1 Dude, try to switch a different Foundry resource in another region, instead of eastus2.

> The reason for these extremely excessive errors is Stream disconnected before completion: Incomplete response returned, reason: content_filter. But this didn't use to happen last week. I've changed the content policy from DefaultV2 to Default but the issue seems to persist. @matys1 Dude, try to switch a different Foundry resource in another region, instead of eastus2.

tried different regions and is still getting the same error so I don't think this is region specific.

I've created a custom content-filter and set everything to low and applied to my model, gpt-5.3-codex and I feel like the issue is less frequent now but still happening.

Are all of you using Azure Foundry, or are some of you using an OpenAI-hosted model?

Azure Foundry for me

Same. Happens even with very simple requests

<img width="801" height="492" alt="Image" src="https://github.com/user-attachments/assets/6f2576bb-950d-413f-939e-9116514e3483" />
Sometimes it even fails completely

<img width="753" height="607" alt="Image" src="https://github.com/user-attachments/assets/f55f98c3-36d9-4e16-95e7-c93767efb087" />

Any workaround for it?

Facing the same issue via Azure Foundry models. I have increased the retry and reconnect limits to around 20 while using them, as it was affecting reasoning sessions, and that is how I am getting long-running tasks done. However, this message is popping up far more often than expected, and the system is taking hours and hours to produce a response. Have there been new changes made on the platform?

FYI - I am seeing this with the codex cli, but not with the codex app.

Same issue...

codex-cli 0.120.0
Azure Foundry gpt-5.4

Seems like the issue is related to Azure and some change they made to content filter policy. Does anyone know where to submit an issue with Azure regarding this? The models are unusable right now.

Same,both with codexli and codex app.
I guess only way to fix this is to modify your content filter action to no block,but that requires you to apply for ability to modify content filter.🙂

Like could you elaborate how to do that?

First you should approve for modified content filtering：Doc link
Then you can create a custom content filter and set action to no block.
Haven't try yet.But if this issue is caused by content filter ,it should help.

I have created and tested a custom content filter with minimal restrictions and still seeing the issue.

I resolved the issue by creating a custom guardrail policy in Azure AI Factory. I've been vibe-coding the entire morning and i haven't seen a single issue yet. In Azure portal go to Azure AI Foundry instance -> Build -> Guardrails -> Create. Delete all guardrails except for the mandatory required Content Safety ones (Hate, Self-harm, Sexual, Violence) but set them to the lowest blocking threshold. Once created edit your deployed model and apply the new custom content safety policy to the target model instead of the Default or DefaultV2.

For comparison, this is what the DefaultV2 policy looks like that makes the models unusable for normal coding:

<img width="523" height="698" alt="Image" src="https://github.com/user-attachments/assets/00ae9b5c-8cf1-4045-87a3-112222287802" />

This is the most minimal policy that resolved my issue:

<img width="520" height="334" alt="Image" src="https://github.com/user-attachments/assets/eca12e79-1911-4ecb-82cf-0269e36c2ce1" />

Found this message in "New Foundry" guardrail UI:
"Assigning a guardrail may take up to 15 minutes."
May explain why I didn't see an improvement.

I resolved the issue by creating a custom guardrail policy in Azure AI Factory. I've been vibe-coding the entire morning and i haven't seen a single issue yet. In Azure portal go to Azure AI Foundry instance -> Build -> Guardrails -> Create. Delete all guardrails except for the mandatory required Content Safety ones (Hate, Self-harm, Sexual, Violence) but set them to the lowest blocking threshold. Once created edit your deployed model and apply the new custom content safety policy to the target model instead of the Default or DefaultV2. For comparison, this is what the DefaultV2 policy looks like that makes the models unusable for normal coding: <img alt="Image" width="523" height="698" src="https://private-user-images.githubusercontent.com/99846054/577921254-00ae9b5c-8cf1-4045-87a3-112222287802.png?jwt=eyJ0eXAiOiJKV1QiLCJhbGciOiJIUzI1NiJ9.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.371XxGAh_tyqWTfmaAs_w33e9ydA_aqKXp8MeCIKfsI"> This is the most minimal policy that resolved my issue: <img alt="Image" width="520" height="334" src="https://private-user-images.githubusercontent.com/99846054/577921374-eca12e79-1911-4ecb-82cf-0269e36c2ce1.png?jwt=eyJ0eXAiOiJKV1QiLCJhbGciOiJIUzI1NiJ9.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.jyUdlYqjjtg7HifwreT5FFu9VoiZL5SEYw4Ncjs8lrw">

This seems to do the trick (for now) 🎆

This from Azure Foundry it's unbearable. It's very unreliable to run OpenAI models. Claude works fine but GPT 5.4(Pro totally gets stuck but even the normal) it rarely finishes a task without the stream getting interrupted. Today and yesterday is extreme but overall this is the experience.

<img width="618" height="312" alt="Image" src="https://github.com/user-attachments/assets/5c81a375-fb91-475d-bcc2-e3b58a873709" />

Was facing the same issue on v0.120.0. Created the custom content filter and it no longer throws the content_filter message as often, but it hangs on the first tool call (waited 10 minutes for it to finish reading git status with no response back). I'm running this on Windows and the behavior is consistent in powershell and cmd.

This is an Azure issue. Please report to them. There's not really anything we can do in the Codex harness to work around a server-side content policy.

Are all of you using Azure Foundry, or are some of you using an OpenAI-hosted model?

I'm not using Azure Foundry. I'm using gpt 5.2 hosted in litellm server

Been getting this issue too. The fixes the guys proposed where you create a new Guardrail set with minimum required fixes the issue slightly. It now works SOME of the time, but often, it stutters and stammers. Often you'll get the result you want, but it does weave a good amount of failed attempts into the mix.

Is anyone tracking this on the Azure side of things?

<img width="566" height="646" alt="Image" src="https://github.com/user-attachments/assets/832c6653-4e95-4bc2-ac0e-540b1310d46d" />

EDIT: After updating VS Code, the stuttering issue went away. Coincidence? Not sure. It's gone back to it's old reliable self though. Weird.

Hi @matys1 , @RobertHaslinger and others -

I applied suggested content filter in a new guardrail against my models created in the old foundry portal. Since then I see the thinking is taking lot of time. Every prompt has slowed down in responding. Are you experiencing this ?
I would appreciate your input. Thank you.

@npagare sorry I didn't have that experience.

Still happening to me, across the board, in codex/opencode, super annoying, with 5.5 and even 5.4/mini/nano...

So ways to solve this

• use the latest openai/v1 end point provided in the new Azure foundry platform
• Above you will be seeing ppl have how apply changes to content filter do both should work!
• also try keeping the codex app server as false!

This made these changes made v0.128 work the same way like open ai end point