ADR-200: MiniMax Provider Integration

Status: Accepted Date: 2026-02-15 Author: Hal Casteel Track: H.3.9

Context

CODITECT currently supports 4 LLM providers:

1. Claude (Anthropic) — primary, via anthropic SDK
2. Codex (OpenAI) — via openai SDK
3. Gemini (Google) — via google-generativeai SDK
4. Kimi (Moonshot AI) — via process monitoring only

MiniMax (a Chinese AI company) now offers an Anthropic Messages API-compatible endpoint, making integration straightforward — the existing anthropic Python/Node SDK works with only a base URL change.

Decision

Add MiniMax as the 5th LLM provider in coditect-core's multi-provider architecture.

Provider Details

Property	Value
Provider Name	MiniMax
API Base URL	https://api.minimax.io/anthropic
API Compatibility	Anthropic Messages API (drop-in)
Authentication	API key via MINIMAX_API_KEY env var
SDK	Standard anthropic Python/Node SDK
Documentation	https://platform.minimax.io/docs/api-reference/text-anthropic-api

Available Models

Model	Context Window	Speed	Use Case
MiniMax-M2.5	204,800 tokens	~60 tps	Flagship
MiniMax-M2.5-highspeed	204,800 tokens	~100 tps	Fast flagship
MiniMax-M2.1	204,800 tokens	~60 tps	Balanced
MiniMax-M2.1-highspeed	204,800 tokens	~100 tps	Fast balanced
MiniMax-M2	204,800 tokens	Variable	Cost-effective

Supported Features

Feature	Supported
Text messages	Yes
Tool use (function calling)	Yes
Thinking/extended thinking	Yes
Streaming	Yes
System prompts	Yes
Temperature (0.0-1.0]	Yes
top_p	Yes
Image/document inputs	No
top_k, stop_sequences	Ignored

Integration Points

MiniMax touches these coditect-core files:

File	Change
config/model-pricing.json	Add 5 MiniMax models with pricing
config/judge-model-routing.json	Add minimax to supported_providers, add fallback entries
config/llm-tools.json	Add minimax CLI tool entry
config/llm-watchers.json	Add minimax watcher config
scripts/moe_classifier/core/multi_model_client.py	Add MINIMAX enum, MiniMaxClient class, model mappings
scripts/moe_classifier/core/provider_detector.py	Add MINIMAX provider with config and tier mappings
minimax/README.md	Per-LLM directory (ADR-122 pattern)

Implementation Pattern

MiniMax uses the Anthropic SDK with a custom base URL. The client implementation follows the same pattern as DeepSeekClient (OpenAI-compatible with custom base URL), but uses the anthropic SDK instead:

class MiniMaxClient(BaseProviderClient):
    """MiniMax API client (Anthropic Messages API-compatible)."""

    BASE_URL = "https://api.minimax.io/anthropic"

    async def complete(self, request: CompletionRequest) -> CompletionResponse:
        import anthropic
        client = anthropic.Anthropic(
            api_key=self.api_key,
            base_url=self.BASE_URL
        )
        response = client.messages.create(
            model=request.model,
            max_tokens=request.max_tokens,
            temperature=request.temperature,
            system=request.system_prompt or "",
            messages=[{"role": "user", "content": request.prompt}]
        )
        # ... standard response mapping

Claude Code Integration

MiniMax can also be used as the backing provider for Claude Code itself via environment variables:

export ANTHROPIC_BASE_URL=https://api.minimax.io/anthropic
export ANTHROPIC_API_KEY=<minimax-api-key>
claude --model MiniMax-M2.5

Or via .claude/settings.local.json (gitignored):

{
  "env": {
    "ANTHROPIC_BASE_URL": "https://api.minimax.io/anthropic",
    "ANTHROPIC_API_KEY": "<minimax-api-key>"
  }
}

Note: This replaces Anthropic entirely for that Claude Code session — it does not run both simultaneously. For simultaneous multi-provider use, use the MoE system.

Provider Tier Classification

Provider	Tier	Notes
Anthropic	flagship	Primary provider
OpenAI	flagship	Strong reasoning (o3)
Google	flagship	1M context, multimodal
MiniMax	cost_effective	Anthropic-compatible, 204K context, high throughput
DeepSeek	cost_effective	Thinking mode, budget option
Alibaba	specialized	Multilingual, domain expertise
Meta	open_source	Self-hostable via Together/Groq

Important Limitations

1. No image/document inputs — text-only; cannot be used for vision tasks
2. Multi-turn tool use — full assistant response (all content blocks) must be appended to conversation history to maintain reasoning chain continuity
3. Temperature range — exclusive of 0.0: use (0.0, 1.0], not [0.0, 1.0]
4. Ignored parameters — top_k, stop_sequences, service_tier, mcp_servers silently ignored

Consequences

Positive

• 5th provider increases MoE diversity — more perspectives for Constitutional Court panels
• Anthropic SDK compatible — minimal new code; reuses existing anthropic SDK dependency
• High throughput — highspeed variants at ~100 tps useful for batch classification
• 204K context — matches Anthropic's context window size
• Cost-effective tier — provides budget option with Anthropic API compatibility

Negative

• No vision — cannot serve as judge for image-related reviews
• Chinese provider — may have data residency considerations for some customers
• Temperature quirk — 0.0 not allowed, requires guard in client code

Neutral

• Follows established ADR-122 (Unified LLM Component Architecture) and ADR-073 (MoE Provider Flexibility) patterns
• No new SDK dependency — uses existing anthropic package

Related ADRs

ADR	Relationship
ADR-073	MoE Provider Flexibility — framework this extends
ADR-122	Unified LLM Component Architecture — per-LLM directory pattern
ADR-194	Codi-Watcher v3.0 — session monitoring (if MiniMax gets a CLI)