Problem

When setting AUXILIARY_VISION_PROVIDER=minimax and AUXILIARY_VISION_MODEL=MiniMax-M2.7 in the environment, the vision_analyze tool silently fails because _resolve_strict_vision_backend() in agent/auxiliary_client.py has no MiniMax branch.

Current code path

_resolve_strict_vision_backend() only knows how to handle:

• openrouter → _try_openrouter()
• nous → _try_nous(vision=True)
• openai-codex → _try_codex()
• anthropic → _try_anthropic()
• custom → _try_custom_endpoint()

All other providers (including minimax) fall through to return None, None, so the vision client is never created.

Why this matters

The MiniMax plan at \$50/mo advertises LLM + text + image + video + voice + music + search capabilities. While their chat-completions API models are text-only, MiniMax does offer multimodal endpoints (e.g. MiniMax-VL for vision) that could be wired up as a vision backend for vision_analyze in Hermes.

Evidence

# From /v1/models API — text models have vision: false
MiniMax-Text-01  → "vision": false
MiniMax-M2.7    → "vision": false
MiniMax-M2.5    → "vision": false
MiniMax-M1      → "vision": false

This means the standard chat endpoint is text-only, but MiniMax has a separate vision-capable model endpoint that is not being used.

Proposed Solution

Add a _try_minimax() function and wire it into _resolve_strict_vision_backend():

def _resolve_strict_vision_backend(provider: str) -> Tuple[Optional[Any], Optional[str]]:
    provider = _normalize_vision_provider(provider)
    if provider == "openrouter":
        return _try_openrouter()
    if provider == "nous":
        return _try_nous(vision=True)
    if provider == "openai-codex":
        return _try_codex()
    if provider == "anthropic":
        return _try_anthropic()
    if provider == "custom":
        return _try_custom_endpoint()
    if provider == "minimax":
        return _try_minimax()   # <-- ADD THIS
    return None, None

And implement _try_minimax() following the same pattern as _try_openrouter():

def _try_minimax() -> Tuple[Optional[OpenAI], Optional[str]]:
    api_key = os.getenv("MINIMAX_API_KEY")
    if not api_key:
        return None, None
    # MiniMax uses OpenAI-compatible endpoint
    return OpenAI(api_key=api_key, base_url="https://api.minimax.io/v1"), "MiniMax-VL"

Note: The exact vision model name (e.g. MiniMax-VL, minimax-vl-01) and base URL should be verified against MiniMax's official API docs — this is a placeholder that needs confirmation.

Additional context

• MiniMax API key is stored in the OpenClaw config at /root/.openclaw/openclaw.json (env.vars.MINIMAX_API_KEY)
• The MINIMAX_API_KEY env var is already set in the Hermes environment
• Current Hermes version: v0.9.0 (1af2e18d, 2026-04-13)
• Related: _PROVIDER_VISION_MODELS only maps xiaomi → mimo-v2-omni today — MiniMax would follow the same pattern
• There is NO issue with the existing MiniMax text provider — that works fine. This is ONLY about adding MiniMax as a vision backend.

Labels

enhancement, provider-integration, vision