AIHubMix provider: streaming completes output but errors near end with ZodError (invalid_union: reason/part/delta)

Description

When using the built-in AIHubMix provider, OpenCode can stream most of the assistant output, but near the end of the stream it throws a ZodError and marks the request as failed.

The error is an invalid_union with fields like:

path ["reason"]: expected string, received object
path ["part"]: expected object, received undefined
path ["delta"]: expected string, received undefined

I use the built-in provider XIAOMI now, and it just works well.

Plugins

at first with oh-my-opencode, and it can be reproduced with none

OpenCode version

1.1.8, and can be reproduced in 1.1.6 and 1.1.1

Steps to reproduce

Use the bulit-in AIHubMix provider:

Select an AIHubMix model.
Send any prompt (e.g. hello / 你好？).
Observe streaming output appears, then near completion OpenCode throws ZodError.
Model tested:
Claude Haiku 4.5 (built-in)
GPT-5.2 (built-in)
glm-4.7 (built-in)
grok-4-1-fast-non-reasoning(added by myself)
gemini-3-flash-preview (added by myself)

Screenshot and/or share link

Operating System

Ubuntu 24.04, in WSL2

Terminal

Windows Terminal

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

All Mistral models fail with ZodError: "finish" expected string, received object #5981: All Mistral models fail with ZodError: "finish" expected string, received object - Same type of ZodError with streaming validation failures (expected string, received object)
Mistral provider fails: auto-install ENOENT + ZodError on streaming responses (Homebrew install) #5997: Mistral provider fails: auto-install ENOENT + ZodError on streaming responses - Contains the same ZodError validation pattern during streaming responses
AI_TypeValidationError on Grok Fast 1 (Opencode Zen mode) #5759: AI_TypeValidationError on Grok Fast 1 (Opencode Zen mode) - Related validation errors with invalid_union patterns

These issues all involve streaming response validation failures where the expected types don't match the actual API response formats. If your issue is specific to AIHubMix provider or involves different API response structures, feel free to ignore these references.

cat ~/.cache/opencode/package.json
{
  "dependencies": {
    "opencode-copilot-auth": "0.0.9",
    "opencode-anthropic-auth": "0.0.5",
    "oh-my-opencode": "2.14.0",
    "@aihubmix/ai-sdk-provider": "1.0.1"
  }
}%

Can confirm this issue on latest opencode desktop (1.0.152).
Possible duplicated as #7674 .

将 provider 从 @aihubmix/ai-sdk-provider 切换到 @ai-sdk/openai-compatible 后，一切就正常工作了。
既然 aihubmix 本身就是 OpenAI 兼容的，可能是专属的 provider 实现本身存在某些兼容性问题。建议排查一下。

After switching the provider from @aihubmix/ai-sdk-provider to @ai-sdk/openai-compatible, everything works fine.
Since aihubmix is already OpenAI-compatible, there might be some compatibility issues in the dedicated provider implementation. This might be worth investigating.

Hi team,

I was wondering if there are any updates or a tentative timeline for a fix? I rely heavily on this service for my daily workflow, so I’m looking forward to getting it back up and running. > Thanks for all your hard work on this!

cat ~/.cache/opencode/package.json
{
  "dependencies": {
    "opencode-copilot-auth": "0.0.9",
    "opencode-anthropic-auth": "0.0.5",
    "oh-my-opencode": "2.14.0",
    "@aihubmix/ai-sdk-provider": "1.0.1"
  }
}%

In my case, I have this:

{
  "dependencies": {
    "opencode-copilot-auth": "0.0.5",
    "opencode-anthropic-auth": "0.0.2",
    "@ai-sdk/openai-compatible": "latest",
    "@ai-sdk/openai": "latest",
    "@ai-sdk/anthropic": "latest"
  }
}

and I also see those errors.

Update Nix dependencies and opencode configuration

@rekram1-node

Experiencing the same issue with a custom OpenAI-compatible provider.

I'm building a provider using @ai-sdk/openai-compatible (v2.0.30). Streaming works fine—text renders correctly—but validation fails at the end with:

[{"expected":"string","code":"invalid_type","path":["reason"],"message":"Invalid input: expected string, received object"}]

Environment

OpenCode: 1.2.10
OS: macOS (Darwin 25.2.0)
Provider: Custom via @ai-sdk/openai-compatible
Backend: Qwen3-32B on vLLM

What I've ruled out

Server response is clean — returns {"type":"completion"} with no reason field
SSE format is correct — "finish_reason":"stop" as a string, matching the OpenAI spec exactly
AI SDK works standalone — using the same provider with streamText directly passes without issues
Only fails inside OpenCode — the ["reason"] path in the validation error doesn't exist in @ai-sdk/openai-compatible's schema (which uses finish_reason), so this appears to be OpenCode's own validation layer expecting a different shape

This is blocking adoption of custom providers. Happy to provide logs or a minimal repro—any workaround or timeline for a fix would be appreciated. @rekram1-node

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

Navigation Menu

Repository navigation

AIHubMix provider: streaming completes output but errors near end with ZodError (invalid_union: reason/part/delta) #7439

Description

Plugins

OpenCode version

Steps to reproduce

Screenshot and/or share link

Operating System

Terminal

Metadata

Assignees

Labels

Type

Projects

Milestone

Relationships

Development

AIHubMix provider: streaming completes output but errors near end with ZodError (invalid_union: reason/part/delta) #7439

Description

Description

Plugins

OpenCode version

Steps to reproduce

Screenshot and/or share link

Operating System

Terminal

Activity

github-actions commented on Jan 9, 2026

qiyuanhuakai commented on Jan 9, 2026

xrh0905 commented on Jan 11, 2026

qiyuanhuakai commented on Jan 11, 2026

zguiyang commented on Jan 17, 2026

elzorrorebelde commented on Jan 18, 2026

Rakan-AlZagha commented on Feb 22, 2026

github-actions commented on Apr 24, 2026

Metadata

Metadata

Assignees

Labels

Type

Projects

Milestone

Relationships

Development

Issue actions