[Bug] Claude→OpenAI translator passes content as array instead of string, crashing NVIDIA NIM / vLLM backends with "sequence item 0: expected str instance, list found"

[bingo4916]

Title:

[Bug] Claude→OpenAI translator passes content as array instead of string, crashing NVIDIA NIM / vLLM backends with "sequence item 0: expected str instance, list found"

Body:

## Bug Description

When using `openai-compatibility` to route Claude Code (`/v1/messages`) requests to an OpenAI-compatible backend (specifically **NVIDIA NIM** running DeepSeek V4-Pro), the Claude→OpenAI message translator passes `content` as a **Claude-format array** `[{"type":"text","text":"..."}]` instead of flattening it to an **OpenAI-format string** `"..."`.

The upstream backend (Python-based vLLM/NVIDIA NIM) receives the array-typed content and crashes with:

TypeError: sequence item 0: expected str instance, list found

This is related but distinct from #1136 (which covers empty system message arrays). This bug affects **all message roles** (user, assistant, system) when their `content` field contains Claude's array format.

## Environment

- **CLIProxyAPI Version:** 6.9.38 (commit `2c626efc`, built 2026-04-25)
- **OS:** Windows 11 (24H2), x64
- **Upstream:** NVIDIA NIM (`https://integrate.api.nvidia.com/v1`)
- **Model:** `deepseek-ai/deepseek-v4-pro`
- **Client:** Claude Code CLI v2.1.119

## Configuration

```yaml
openai-compatibility:
  - name: nvidia
    base-url: https://integrate.api.nvidia.com/v1
    api-key-entries:
      - api-key: nvap***
    models:
      - name: deepseek-ai/deepseek-v4-pro

payload:
  override:
    - models:
        - name: deepseek-ai/deepseek-v4-pro
          protocol: openai
      params:
        chat_template_kwargs.thinking: false
        max_tokens: 16384

Steps to Reproduce

1. Configure CLIProxyAPI with openai-compatibility pointing to NVIDIA NIM
2. Point Claude Code at CLIProxyAPI via ANTHROPIC_BASE_URL=http://localhost:8317
3. Set model to deepseek-ai/deepseek-v4-pro
4. Send any message through Claude Code

Actual Behavior

What Claude Code sends (/v1/messages):

{
  "model": "deepseek-ai/deepseek-v4-pro",
  "messages": [
    {
      "role": "user",
      "content": [
        {"type": "text", "text": "hello"}
      ]
    }
  ],
  "max_tokens": 8192,
  "stream": true
}

What gets forwarded to NVIDIA NIM (/v1/chat/completions) — BUG HERE:

The translator converts the request but keeps content as an array (or produces malformed nested structures). The NVIDIA Python backend receives non-string content and crashes:

HTTP 500 Internal Server Error: sequence item 0: expected str instance, list found

Error log from CLIProxyAPI v6.9.38:

openai_compat_executor.go:258] request error, error status: 500,
error message: "Internal server error: sequence item 0: expected str instance, list found"

After this error, the credential enters cooldown and all subsequent requests return:

503 auth_unavailable: no auth available (providers=nvidia, model=deepseek-ai/deepseek-v4-pro)

Expected Behavior

The Claude→OpenAI translator should flatten all content arrays to strings before forwarding:

{
  "model": "deepseek-ai/deepseek-v4-pro",
  "messages": [
    {
      "role": "user",
      "content": "hello"
    }
  ],
  "max_tokens": 8192,
  "stream": true
}

The flattening logic should handle:

• Simple text blocks: [{"type":"text","text":"..."}] → "..."
• Mixed content (text + tool_use + tool_result): concatenate text parts, stringify tool parts
• System messages with array content
• Multi-turn conversations with assistant content arrays containing both text and tool_use blocks

Workaround

I wrote a Node.js translation proxy that sits between Claude Code and CLIProxyAPI, correctly flattening content arrays. The chain is:

Claude Code → :8318 (translator-proxy, flattens content) → :8317 (CLIProxyAPI) → NVIDIA NIM

With this workaround in place, requests succeed with HTTP 200 (50-80s response time on DeepSeek V4-Pro). This proves the issue is specifically in CLIProxyAPI's internal/translator/claude/openai/chat-completions/ path.

Key workaround code — the flatten function:

function flattenContent(content) {
  if (typeof content === 'string') return content;
  if (Array.isArray(content)) {
    return content
      .filter(block => block.type === 'text')
      .map(block => block.text || '')
      .join('');
  }
  return String(content || '');
}

Root Cause Analysis

The relevant code is likely in:

• internal/translator/claude/openai/chat-completions/claude_openai_request.go

The converter needs a flattenContent() function that:

1. If content is a string → pass through unchanged
2. If content is an array → extract all type:"text" blocks' .text fields, join them into a single string
3. For assistant messages with tool_use blocks → convert to OpenAI tool_calls format with function.name / function.arguments
4. For user messages with tool_result blocks → convert to readable string representation

Additional Context

• Direct OpenAI-format calls work fine: When calling /v1/chat/completions directly (bypassing the Claude→OpenAI translation), NVIDIA NIM responds correctly with HTTP 200
• The issue only occurs through the /v1/messages → OpenAI translation path
• Related issues: #1136 (system empty array), #1043 (assistant array dropped), #1670 (tool_result array handling)
• Alternative project that handles this correctly: anyllm-proxy (Rust-based proxy purpose-built for Anthropic↔OpenAI translation)

Impact

This blocks all Claude Code users from using NVIDIA NIM (and likely other strict OpenAI-compatible backends such as vLLM, Ollama, TGI) through CLIProxyAPI's openai-compatibility feature. NVIDIA NIM is one of the most popular free-tier LLM providers, so this affects a significant number of users.

[mdreamfly]

same issue

[wyllwyc]

The same issue,OpenCodezen also seems to have the same problem

[Nasak2]

qwenpaw同样存在该问题