Grok CLI + opencode.ai Gateway 400 Error

Description

When using Grok CLI to call the DeepSeek V4 Flash model through the opencode.ai gateway (https://opencode.ai/zen/go/v1), requests return a 400 Bad Request error when the conversation history contains image content.
Error Message

API error (status 400 Bad Request): invalid_request_error: Error from provider (Console Go):
Upstream request failed: [invalid_request_error] Failed to deserialize the JSON body into
the target type: messages[132]: unknown variant image_url, expected text at line 1 column 568961

---

Environment Configuration

Grok Configuration (~/.grok/config.toml)

[models]
default = "og-deepseek-flash"

[model.og-deepseek-flash]
model = "deepseek-v4-flash"
base_url = "https://opencode.ai/zen/go/v1"
name = "DeepSeek V4 Flash (opencode go)"
env_key = "OPENCODE_GO_API_KEY"
api_backend = "chat_completions"
context_window = 1000000

MCP Plugins

[mcp_servers.qwen-mm-plugins-core]
command = "uvx"
args = [
    "--from",
    "qwen-mm-plugins[core] @ git+https://github.com/QwenLM/Qwen-MM-Plugins.git@main",
    "qwen-mm-plugins-core",
]
enabled = true

---

Root Cause Analysis

Format Transformation Flow

┌─────────────────────────────────────────────────────────────────────────┐
│ Current Flow (Causing Error)                                             │
├─────────────────────────────────────────────────────────────────────────┤
│ 1. Grok Internal Storage:                                               │
│    {"type": "image", "url": "data:image/jpeg;base64,..."}                │
│                                                                          │
│ 2. Grok Builds API Request → Converts to OpenAI Standard Format:       │
│    {"type": "image_url", "image_url": {"url": "data:..."}}              │
│                                                                          │
│ 3. Sends to opencode.ai/zen/go/v1 (chat_completions backend)            │
│                                                                          │
│ 4. DeepSeek API Deserialization Fails ❌                                 │
│    Error: unknown variant `image_url', expected 'text'                  │
└─────────────────────────────────────────────────────────────────────────┘

Root Cause

1. opencode.ai gateway's chat_completions backend is not fully compatible with OpenAI's multimodal format

• OpenAI standard uses {"type": "image_url", "image_url": {...}}
• DeepSeek API's deserializer only accepts {"type": "text"}

2. Images in conversation history are accumulated

• Error occurs at messages[132], indicating a long conversation history
• Images from historical messages are sent together with new requests

3. MCP plugins cannot solve this issue

• qwen-mm-plugins-core provides MCP tools but does not intercept API requests
• Images are directly included in chat_history and sent to the API

---

Log Evidence

unified.jsonl Log

{
  "ts": "2026-08-11T02:50:57.715Z",
  "src": "shell",
  "lvl": "error",
  "msg": "shell.turn.inference_failed",
  "ctx": {
    "kind": "api",
    "status_code": 400,
    "is_retryable": false,
    "message": "API error (status 400 Bad Request): invalid_request_error: Error from provider (Console Go): Upstream request failed: [invalid_request_error] Failed to deserialize the JSON body into the target type: messages[132]: unknown variant `image_url`, expected `text` at line 1 column 568961"
  }
}

Image Budget Log

{
  "msg": "shell.image_budget",
  "ctx": {
    "inline_images": 5,
    "body_bytes": 731237
  }
}

---

Steps to Reproduce

1. Use og-deepseek-flash model in Grok CLI
2. Upload an image or use a tool that produces image results
3. Continue the conversation to build up message history
4. Send a new request when history contains images
5. Receive 400 error

Expected Behavior

Option A: Gateway-level transformation
The opencode.ai gateway should convert OpenAI-formatted image_url to a format accepted by DeepSeek API before forwarding (e.g., convert images to text via OCR, or use a DeepSeek-supported image format).

Option B: Clear documentation
Clearly document in the documentation that the chat_completions backend of opencode.ai gateway does not support OpenAI's multimodal format, and clients should avoid sending image content.

Plugins

No response

OpenCode version

No response

Steps to reproduce

No response

Screenshot and/or share link

No response

Operating System

No response

Terminal

No response

Start by reproducing the request with image content in conversation history through the chat_completions backend, then trace where OpenAI image_url parts are transformed before forwarding to the gateway. Done means the request is accepted with supported image handling, or the documentation clearly states that this multimodal format is unsupported and how clients should avoid it.