Getting "AI_APICallError: Invalid API parameter, please check the documentation." in opencode with zAI glm 4.6 #2944

Activity

1. github-actions commented on Oct 3, 2025

   

   github-actionsBot

   on Oct 3, 2025 – with GitHub Actions

   Contributor

   More actions

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

   • GLM Coding Plan - Frequent Errors #2898: GLM Coding Plan - Frequent Errors (reports same "Invalid API Parameter. Check the documentation." error with GLM models, especially pronounced in 4.6)
   • AI_APICallError: The messages parameter is illegal. Please check the documentation. - Zhipu AI GLM-4.5 #2557: AI_APICallError: The messages parameter is illegal. Please check the documentation. - Zhipu AI GLM-4.5 (similar API parameter error with z.ai GLM configuration)
   • GLM-4.6 model fails to use web search tools after processing complex mathematical problems, and subsequently loses all tool use capability even when explicitly prompted. #2903: GLM-4.6 model fails to use web search tools (GLM-4.6 specific issues with tool usage, though different symptoms)

   Feel free to ignore if none of these address your specific case with MCP vision server integration.
2. wKich commented on Oct 3, 2025

   

   wKich

   on Oct 3, 2025

   More actions

   I've described the same issue here #2941 (comment)
3. rekram1-node commented on Oct 3, 2025

   

   rekram1-node

   on Oct 3, 2025

   Contributor

   More actions

   does this specifically happen with attachments that are images?
4. rekram1-node

   assigned

   thdxr

   on Oct 3, 2025
5. moinulmoin commented on Oct 3, 2025

   

   moinulmoin

   on Oct 3, 2025

   Author

   More actions

   does this specifically happen with attachments that are images?

   yes yes specifically happen with attachments that are images, because glm 4.6 doesnt support image input and need to use vision mcp server, which works well with droid cli and claude code cli
6. rekram1-node commented on Oct 3, 2025

   

   rekram1-node

   on Oct 3, 2025

   Contributor

   More actions

   hm we will need to update models.dev for that then

   👍React with 👍2Reacted by Moinul Moin and Max Shapira
7. naltun commented on Oct 14, 2025

   

   naltun

   on Oct 14, 2025

   Last edited by naltun

   More actions

   Also hit this error with GLM 4.6 when attaching images:
   
   

   e: it's related to this: https://docs.z.ai/devpack/mcp/vision-mcp-server

   tl;dr Image and Video processing is only available with GLM Coding Pro plans and up. I am using GLM Coding Lite and can validate this.
8. rekram1-node commented on Oct 14, 2025

   

   rekram1-node

   on Oct 14, 2025

   Last edited by rekram1-node

   Contributor

   More actions

   Yeah ig they don't support them, we can make the tui guard against this pretty easily tho, to be able to continue your convo I recommend doing a /undo and then deleting the image attachments

   👍React with 👍2Reacted by Noah Altunian and Jaehoon Choi
9. xshapira commented on Oct 15, 2025

   

   xshapira

   on Oct 15, 2025

   More actions

   Also hit this error with GLM 4.6 when attaching images:

   e: it's related to this: docs.z.ai/devpack/mcp/vision-mcp-server

   tl;dr Image and Video processing is only available with GLM Coding Pro plans and up. I am using GLM Coding Lite and can validate this.

   I'm using GLM Coding Max and still having the same issue.

   👍React with 👍2Reacted by Noah Altunian and Jaehoon Choi
10. rekram1-node commented on Oct 15, 2025

    

    rekram1-node

    on Oct 15, 2025

    Last edited by rekram1-node

    Contributor

    More actions

    Are you sending images?

    It looks like there is a special mcp server they want you to be using, if it is an mcp server then @ ing the file won't work correctly, nor will a drag and drop (because they are unrelated to mcp servers).

    In this case the llm would have to call the mcp server tool right

    👍React with 👍5Reacted by Max Shapira, shekohex, Jensen, Jaehoon Choi and kaanmertkoc
11. kaanmertkoc commented on Nov 4, 2025

    

    kaanmertkoc

    on Nov 4, 2025

    Contributor

    More actions

    Hey! Encountered this too. Wondering if it'd make sense to have opencode
    detect this pattern and smooth it out:

    1. When GLM-4.6 user attaches an image and hits this error
    2. Show a prompt: "GLM-4.6 vision requires the Z.AI Vision MCP Server.
       Would you like to add it now?"
    3. If yes → auto-run the MCP installation
    4. Resume conversation with image routed to MCP tool

    Curious what you think happy to explore implementation if it resonates.

    👍React with 👍2Reacted by Jensen and Noah Altunian
12. WilfredLemus commented on Feb 23, 2026

    

    WilfredLemus

    on Feb 23, 2026

    More actions

    Hi! Quick update from my side: I built and published a plugin that works around the GLM clipboard image issue in OpenCode.

    It keeps the same paste UX (Cmd+V / clipboard image) for GLM models by:

    • detecting pasted data:image/... parts,
    • writing the image to a temp local file,
    • replacing the image part with a text instruction that includes the local file path so image MCP tools can process it.
    • It also includes automatic temp cleanup.

    Plugin:

    • npm: https://www.npmjs.com/package/opencode-glm-clipboard
    • repo: https://github.com/WilfredLemus/opencode-glm-clipboard

    Install:
    npx -y opencode-glm-clipboard@latest

    This is a temporary workaround while we wait for a native fix in OpenCode.

    👍React with 👍1Reacted by shekohex
13. github-actions commented on Apr 24, 2026

    

    github-actionsBot

    on Apr 24, 2026 – with GitHub Actions

    Contributor

    More actions

    To stay organized issues are automatically closed after 60 days of no activity. If the issue is still relevant please open a new one.
14. github-actions

    closed this as completedon Apr 24, 2026
15. skywalker-35 commented on May 12, 2026

    

    skywalker-35

    on May 12, 2026

    More actions

    I'm encountering the exact same error with GLM-5 Turbo in OpenCode TUI (v1.2.x on Windows). Steps to reproduce:

    1. Select glm-5-turbo as the model in OpenCode TUI
    2. Paste an image path (e.g., C:\path\to\screenshot.png)
    3. Send the message
    4. Error: AI_APICallError: API 调用参数有误，请检查文档。 (Invalid API parameter, please check documentation.)

    After investigating, the root cause appears straightforward:

    GLM-5 Turbo is a text-only model — it does NOT support native vision/image input. This is confirmed by:

    • Z.AI official docs: GLM-5-Turbo — no mention of image input support
    • Z.AI official docs: GLM-5V-Turbo — the vision variant with CogViT encoder, released 2026-04-01
    • Zed IDE promote dev -> prod #51892 — same exact error, closed as "GLM-5 was never equipped with visual capabilities"

    When an image is pasted in TUI, OpenCode encodes it as base64 data:image/... and includes it in the API request. GLM-5's API rejects this because it doesn't understand image content blocks in messages. This is consistent with Issue #14673 (data: URL vs file path for MCP) and #5101 (switching from vision to non-vision model breaks thread).

    Workarounds that work today:

    1. Switch to GLM-5V-Turbo — the vision-capable variant of GLM-5, supports image + text input natively
    2. Use @filepath syntax to reference images so MCP tools (look_at, understand_image) can handle them instead of sending images directly to the model
    3. Community plugin: npx -y opencode-glm-clipboard@latest — detects pasted images, saves to temp file, replaces with file path for MCP tools
    4. /compact before switching from a vision model to GLM-5 to clear image history from context

    Suggestion for OpenCode: it would be great if the TUI could detect when the current model doesn't support image input and either:

    • Warn the user before sending
    • Auto-route images through MCP tools transparently
    • Strip image content blocks and warn instead of failing with an API error

    This is essentially the same class of problem as #5101 (which was closed as fixed) — models without vision capability should not have image content sent to them.