Azure OpenAI Models not working in the latest update

[mdgdeveloper]

Description

Since the latest update of OpenCode all Azure OpenAI models fail with the same error message

Tested:

• GPT-5.3-Codex
• GPT-5.2-Codex
• GPT-5.4 Mini

All behave the same way. Opencode starts the conversation, tries to interact with the model and fails with a "I'm sorry, but I cannot assist with that request."

Plugins

No response

OpenCode version

1.3.17

Steps to reproduce

1. Launch opencode
2. Interact in Plan/Build mode adding files
3. The answer starts and then fails with "I'm sorry, but I cannot assist with that request."

Screenshot and/or share link

Operating System

Ubuntu 24.04 (WSL2)

Terminal

No response

[github-actions]

This issue might be a duplicate of existing issues. Please check:

• Plan mode <system-reminder> triggers Azure OpenAI jailbreak detection #16442: Azure OpenAI jailbreak detection triggered by plan mode system prompt (same symptom: requests fail with content policy error in Plan/Build mode)
• plan.txt triggers Azure OpenAI content filter #17968: plan.txt triggers Azure OpenAI content filter (same root cause: hardcoded prompt phrases triggering Azure's content safety policy)

The "I'm sorry, but I cannot assist with that request" error is consistent with Azure OpenAI's content filter rejecting the request due to phrases in OpenCode's built-in plan/system prompts.

[ngbrown]

I was having this issue yesterday after updating to OpenCode v1.4.3. I lowered the model filter in Azure OpenAI from DefaultV2 to just Default. I rolled OpenCode back to 1.4.1, and still was having the issue.

GPT-5.4, GPT-5.2-chat, and GPT-5-mini all had the same behavior. Sometimes they would start responding then say "I'm sorry, but I cannot assist with that request". Other times they would just go right for "I'm sorry, but I cannot assist with that request".

I thought maybe it was the system prompt, but it's hard to get that logged. I finally got that outputted through an OpenCode plugin. The same system prompt and user prompt worked fine in the Azure OpenAI playground.

I tried using the LiteLLM proxy to log the interactions, and everything going through that works fine, with no change on the OpenCode side. So right now, my guess is it's something to do with how OpenCode is calling the Azure OpenAI API.

[CarloWood]

If you start a new session and only use agent "Build" (not "Plan"), do you still have this problem?

[mdgdeveloper]

If you start a new session and only use agent "Build" (not "Plan"), do you still have this problem?

Yes. Happens in both modes

I was having this issue yesterday after updating to OpenCode v1.4.3. I lowered the model filter in Azure OpenAI from DefaultV2 to just Default. I rolled OpenCode back to 1.4.1, and still was having the issue.

GPT-5.4, GPT-5.2-chat, and GPT-5-mini all had the same behavior. Sometimes they would start responding then say "I'm sorry, but I cannot assist with that request". Other times they would just go right for "I'm sorry, but I cannot assist with that request".

I thought maybe it was the system prompt, but it's hard to get that logged. I finally got that outputted through an OpenCode plugin. The same system prompt and user prompt worked fine in the Azure OpenAI playground.

I tried using the LiteLLM proxy to log the interactions, and everything going through that works fine, with no change on the OpenCode side. So right now, my guess is it's something to do with how OpenCode is calling the Azure OpenAI API.

I was able to workaround the problem by defining a custom Content Filter in Azure and lowering everything. I agree that it's something from OpenCode way of interacting with Azure OpenAI that triggers these Content Filters

[Anshul261]

Same issue, I even tried changing the filter to the recommended Default V2 to Default, and also some custom filters, but there was no change to the issue. Older version v1.3.12 works fine still with GPT-5.4 Pro

[abetekgeo]

I have done some testing and for me it seems to be a combination of the system prompt with tool descriptions that is triggering the content filter. Disabling one of bash or question solves the problem but that is definitely not a good long-term solution

"permission": {"question": "deny"}

[paulovsm]

I am experiencing persistent Content Filter errors starting from version 1.4.3. Even with a simple "Hello" in a fresh session, the models fail to respond.

Behavior:

• Azure OpenAI Content Filter: I have attempted to modify settings in both Azure OpenAI and Microsoft Foundry (switching between Default and DefaultV2, and creating custom filters), but the error persists.
• Intermittent Success: Occasionally, after multiple attempts, the session progresses. However, as soon as a tool call is triggered (e.g., read, patch, etc.), the error recurs immediately.
• Configuration Attempts: I tried modifying the system prompt in opencode.json to isolate the issue, but it had no effect.

Models Affected (Azure OpenAI):

• GPT-5.4, GPT-5.4-mini, GPT-5.4-nano
• GPT-5.3-codex

Models Working Normally (Azure OpenAI):

• Kimi-K2.5
• Sonnet 4.6
• Opus 4.6

Environment & Troubleshooting:

• Regression: The issue is resolved only by downgrading to a version prior to 1.3.12.
• Isolation: All tests were conducted in clean Docker containers to eliminate any local environment interference.

Snippet of opencode.json used for testing:

"mode": {
    "build": {
        "prompt": "{file:~/build.md}",
        "tools": {
            "write": true,
            "edit": true,
            "bash": true
        }
    }
}

[paulovsm]

I have done some testing and for me it seems to be a combination of the system prompt with tool descriptions that is triggering the content filter. Disabling one of bash or question solves the problem but that is definitely not a good long-term solution

"permission": {"question": "deny"}

I tried this with partial success. It worked sometimes (mostly with simple requests) and failed in most other cases in version 1.4.3.

[blazejzak]

The fix is creating your own content filter in Azure AI Foundry with minimal checks here
https://ai.azure.com/resource/contentfilters/contentFilter