BUG: `deepseek-v4-pro` on zen/go `/v1/messages` returns 400 `Empty input messages` for string content (flash works on same request)

Summary

On the OpenCode Zen gateway https://opencode.ai/zen/go, the Anthropic-compatible endpoint POST /v1/messages fails for deepseek-v4-pro when messages[].content is passed as a string. The exact same request with deepseek-v4-flash returns 200 OK. Switching deepseek-v4-pro to the array form of content also works.

This only affects deepseek-v4-pro on the Anthropic endpoint — the OpenAI-compatible endpoint and deepseek-v4-flash are unaffected.

Environment

Reproduction

# Fails: 400 Empty input messages
curl -s https://opencode.ai/zen/go/v1/messages \
  -H "x-api-key: $KEY" \
  -H "anthropic-version: 2023-06-01" \
  -H "content-type: application/json" \
  -d '{"model":"deepseek-v4-pro","max_tokens":30,"messages":[{"role":"user","content":"hi"}]}'

# Control: same request with deepseek-v4-flash returns 200 OK
curl -s https://opencode.ai/zen/go/v1/messages \
  -H "x-api-key: $KEY" \
  -H "anthropic-version: 2023-06-01" \
  -H "content-type: application/json" \
  -d '{"model":"deepseek-v4-flash","max_tokens":30,"messages":[{"role":"user","content":"hi"}]}'

Error response

{
  "error": {
    "param": null,
    "type": "invalid_request_error",
    "code": "invalid_request_error",
    "message": "Error from provider (Console Go): Upstream request failed: [invalid_request_error] Empty input messages"
  }
}

The error is thrown by the Console Go upstream provider.

Test matrix (same endpoint, same payload, only the model changes)

Exact trigger condition

• Endpoint: only the Anthropic-compatible /v1/messages
• Model: only deepseek-v4-pro (same request works for deepseek-v4-flash)
• Trigger: messages[].content is a string ("content":"hi"); the array form ("content":[{"type":"text","text":"hi"}]) works

Impact

The Anthropic SDK accepts a string shorthand for content, e.g. client.messages.create(model="deepseek-v4-pro", messages=[{"role":"user","content":"hi"}]). Serialized as string content, this directly triggers the 400. Any string-content Anthropic-SDK call to deepseek-v4-pro fails, while deepseek-v4-flash or the array form does not.

Related

This overlaps with #34331 (same deepseek-v4-pro error observed there), but that issue is broader and predates this specific flash-vs-pro contrast. This report isolates the deepseek-v4-pro-only regression against a working control model.

Workaround: pass content as the array form [{"type":"text","text":"..."}] for deepseek-v4-pro.