feat(minimax): add MiniMax provider with tier-aware rate limiting

[Societus]

Summary

Add MiniMax as a built-in LLM provider using the generic tier framework from #82.

This PR is a straightforward application of the same pattern as #83. Both MiniMax and Z.AI are OpenAI-compatible APIs with subscription tiers and built-in reasoning models. The generic tier framework made this provider almost mechanical to implement -- the only provider-specific code is the model names, the reasoning_split parameter vs Z.AI's thinking toggle, and the tier definitions.

Depends on: #82 (generic tier framework -- merge that first)

Why This Was Inconsequential

MiniMax shares the same architectural profile as Z.AI:

• OpenAI-compatible API at https://api.minimax.io/v1
• Bearer token auth using the same openai SDK
• Reasoning models with a thinking separator parameter
• Published subscription tiers with conservative rate limits

The generic framework from #82 eliminated all boilerplate for tier resolution. Adding MiniMax was just: define RATE_LIMIT_TIERS, set the base URL, and pick the reasoning parameter name. Everything else is inherited.

Changes

New: MiniMax Provider (minimax.py)

• RATE_LIMIT_TIERS with Starter/Plus/Max/Ultra configs from published limits
• resolve_rate_limiter() from BaseProvider (zero custom tier code)
• reasoning_split=True by default (separates thinking from content)
• Retry budget: 5 retries / 30s max wait
• Models: MiniMax-M2.7 (default), M2.7-highspeed, M2.5, M2.5-highspeed, M2.1, M2.1-highspeed, M2

Registry (registry.py)

• Register minimax -> MiniMaxProvider with openai package hint

Rate Limiter (rate_limiter.py)

• PROVIDER_DEFAULTS["minimax"] = Starter-tier conservative (5 RPM / 25K TPM)

CLI Helpers (helpers.py)

• MINIMAX_API_KEY, MINIMAX_BASE_URL, MINIMAX_REASONING_SPLIT, MINIMAX_TIER env vars
• Auto-detect from MINIMAX_API_KEY
• Added to provider validation list

Tests (test_minimax_provider.py)

• 30 tests: constructor, tier resolution (4 tiers + edge cases), generate with mock, stream_chat, reasoning_split, registry integration

Rate Limit Tiers

From published MiniMax docs (5-hour rolling window):

Tier	Requests/5hrs	RPM	TPM
Starter	1,500	5	25,000
Plus	4,500	15	75,000
Max	15,000	50	250,000
Ultra	30,000	100	500,000

Highspeed variants (e.g., MiniMax-M2.7-highspeed) share the same rate limits as their base plan. The difference is model selection (faster inference), not quota.

Ref: https://platform.minimax.io/docs/token-plan/intro

Configuration

export MINIMAX_API_KEY="***"
export MINIMAX_TIER="plus"              # starter | plus | max | ultra
export MINIMAX_BASE_URL="..."           # override default
export MINIMAX_REASONING_SPLIT="true"   # default true

Test Plan

uv run pytest tests/unit/test_providers/test_minimax_provider.py -v
# 30 passed

All 121 provider tests pass with zero regressions.

PR Stack

#	PR	Status
1	#82 -- Generic tier framework	Ready for review
2	#83 -- Z.AI adopts the framework	Depends on #82
3	This PR -- MiniMax provider	Depends on #82

Related

• Ref: Feature: Add Z.AI (Zhipu AI) provider support #68