AI_APICallError: The messages parameter is illegal. Please check the documentation. - Zhipu AI GLM-4.5

[jgwhite0111]

Zhipu AI GLM-4.5 regularly hangs mid edit and waits for a response after the message "AI_APICallError: The messages parameter is illegal. Please check the documentation."

Edit: I don't think it is mid-edit. It seems to happen after session compaction. And it seems to then start its memory again from the last session.

Only settings that are none default are z.ai's recommended setup for 'Coding Plan' on opencode. In ~/.config/opencode/opencode.json I have:

{
    "$schema": "https://opencode.ai/config.json",
    "provider": {
        "zhipuai": {
            "api": "https://api.z.ai/api/coding/paas/v4"
        }
    }
}

[github-actions]

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

• Opencode with GLM 4.5 String issue #1880: GLM 4.5 String issue - Similar GLM 4.5 model problem with different error message but same hanging behavior
• Add support for Z.ai GLM Coding Plan (Anthropic-compatible provider) #2431: Add support for Z.ai GLM Coding Plan - Discusses the same Z.ai Anthropic-compatible endpoint configuration you're using
• Please adapt to the official GLM-4.5 direct connection version instead of Openrouter #1559: Please adapt to the official GLM-4.5 direct connection version - Related to GLM-4.5 connectivity issues and official API support

Feel free to ignore if none of these address your specific case.

[rekram1-node]

if this was happening at random intervals (like after compacts) can you try latest? There were some fixes shipped today

[jgwhite0111]

I'm still getting the message in the latest version. It's a shame, because apart from that, glm-4.5 runs very fast and intelligently in opencode compared to other cli tools I'm testing. Sadly, this error makes it pretty unusable, as I can't leave it unattended, and need to re-prompt the AI constantly to get it to continue after this error. My workflow is to set all permissions to allow and run it sandboxed in a VM, then connect through a mosh session on my phone. I'm not the only person that uses this type of workflow.

[rekram1-node]

I haven't tested but someone in the discord said:

use the anthropic endpoint and it will fix it

ex:

"zai": {
      "name": "zai-anth-base",
      "npm": "@ai-sdk/anthropic",
      "models": {
        "glm-4.5": {
          "id": "glm-4.5",
          "name": "GLM-4.5",
          "attachment": true,
          "tool_call": true,
          },
          "limit": {
            "context": 131072,
            "output": 98304
          },
        },
      "options": {
        "baseURL": "https://api.z.ai/api/anthropic/v1"
      },
    },

[jgwhite0111]

Yea, I just saw that and tested it. It solves the issue. Thanks! I won't close for now in case anyone has any tweaks or recommendations to offer. Otherwise, feel free to close.

[shekohex]

But you now would lose the CoT and reasoning from the model.

[jgwhite0111]

OK so the Anthropic API doesn't suport reasoning? I think I'd rather put up with re-prompting it than lose reasoning for now.

[vinaychilukuri]

Have been facing the same issue! And, I also don't want to lose reasoning at this point.

[0serna]

Same problem here, today I've been getting this error repeatedly in opencode v0.9.5

[rekram1-node]

Where does it say you lose CoT reasoning? just curious maybe I missed it

[jgwhite0111]

Where does it say you lose CoT reasoning? just curious maybe I missed it

In the replies above from shekohex:

But you now would lose the CoT and reasoning from the model.

[rekram1-node]

Yeah I meant like where does the provider say this

[shekohex]

Yeah I meant like where does the provider say this

Nowhere in the docs, but looks like the OpenAI format shows the CoT/Thinking but using the Anthropic compatible API DOES NOT show it, or maybe it is missing it altogether.

[aemr3]

would love to see this fixed as well, I am using GLM-4.5 heavily as well, with subscription. I didn't switch to anthropic, this model is better with reasoning.