[Bug] suggestion_handler sends system-only message — fails on MiniMax & Anthropic LLMs

[Ptah-CT]

Summary

handle_get_suggestion_queries in src/memos/api/handlers/suggestion_handler.py builds the LLM request as a single system message:

message_list = [{"role": "system", "content": suggestion_prompt.format(memories=memories)}]
response = llm.generate(message_list)

OpenAI accepts this shape. Several other LLM backends do not, including the ones MemOS already documents/configures (minimax, plus Anthropic-compatible endpoints like the MiniMax Anthropic gateway):

Backend	Endpoint	Response
MiniMax Text API	https://api.minimax.io/v1/chat/completions	HTTP 400 invalid params, chat content is empty (2013)
MiniMax Anthropic-compat	https://api.minimax.io/anthropic/v1/messages	HTTP 400 invalid params, messages must not be empty (2013)
Anthropic native	https://api.anthropic.com/v1/messages	(same — system is a top-level field, messages[] must have ≥1 user turn)

Per MiniMax's own minimal example in the official docs, even system-only is rejected:

{
  "model": "MiniMax-M2.7",
  "messages": [
    { "role": "system", "name": "MiniMax AI" },
    { "role": "user",   "name": "user", "content": "hello" }
  ]
}

Reproduction

1. MOS_CHAT_MODEL=MiniMax-M2.7, OPENAI_API_BASE=https://api.minimax.io/v1, valid OPENAI_API_KEY.
2. POST /product/suggestions with mem_cube_id=<existing cube>, language=en.
3. Server returns HTTP 500. With [Bug] timed_with_status decorator silently swallows exceptions and returns None #1523 fixed: openai.BadRequestError: Error code: 400 ... 'chat content is empty (2013)'. Without: AttributeError: 'NoneType' object has no attribute 'replace'.

Proposed fix

Split the prompt into a short system persona and a user message containing the actual instruction. OpenAI accepts this; MiniMax/Anthropic require it.

message_list = [
    {"role": "system",
     "content": "You generate suggestion queries based on the user's recent memories."},
    {"role": "user",
     "content": suggestion_prompt.format(memories=memories)},
]

This is a strict widening — every backend that accepts the original system-only shape also accepts the split shape. No behaviour change for existing OpenAI/Azure deployments.

PR follows.

Related

• [Bug] timed_with_status decorator silently swallows exceptions and returns None #1523, [Bug] clean_json_response crashes with cryptic AttributeError when given None #1525 — the silent-fail / NoneType crash that this bug currently surfaces as.
• Other handlers in src/memos/api/handlers/ should be audited for the same [{role: system, content: ...}] pattern; I only fixed suggestion_handler in this PR but happy to extend.