feat(benchmark): add MiniMax as LLM provider for HomeSec-Bench

[octo-patch]

Summary

Add MiniMax Cloud API as a first-class LLM provider for the Home Security AI Benchmark (HomeSec-Bench), enabling users to benchmark MiniMax models (M2.7, M2.7-highspeed, M2.5, M2.5-highspeed) against local Qwen/DeepSeek and other cloud providers like OpenAI.

Changes

• Provider presets system with auto-configured base URL and default model for MiniMax
• MiniMax auto-detection via API type, base URL pattern, or MINIMAX_API_KEY env var
• Temperature clamping [0, 1.0] for MiniMax API compatibility
• Cloud API recognition enables stream_options for token tracking
• config.yaml llmProvider selector and minimaxModel dropdown for Aegis UI
• Documentation updated SKILL.md with providers table, README.md mention
• 52 tests (49 unit + 3 integration) covering all provider logic

Files Changed (5 files, 538 additions)

File	Changes
scripts/run-benchmark.cjs	Provider presets, auto-detection, temp clamping
config.yaml	llmProvider and minimaxModel parameters
SKILL.md	Provider docs, env vars, standalone examples
README.md	MiniMax benchmark mention
tests/minimax-provider.test.cjs	49 unit + 3 integration tests

Usage

AEGIS_LLM_API_TYPE=minimax MINIMAX_API_KEY=your-key node scripts/run-benchmark.cjs

MiniMax Models

Model	Context	Notes
MiniMax-M2.7	1M tokens	Latest flagship
MiniMax-M2.7-highspeed	1M tokens	Optimized for throughput
MiniMax-M2.5	204K tokens	Previous generation
MiniMax-M2.5-highspeed	204K tokens	Fast inference

Test Plan

• All 49 unit tests pass
• All 3 integration tests pass with live MiniMax API
• Existing benchmark functionality unchanged
• Run full HomeSec-Bench suite with MiniMax-M2.7

[solderzzc]

@octo-patch , thanks for your PR, I understand the requirements is to enable MiniMax.
We are now working on swift version mlx-server. https://github.com/SharpAI/mlx-server
Will check the whole pipeline for MiniMax 2.7 integration.

[octo-patch]

Thank you @solderzzc! Sounds good — looking forward to hearing back once you've had a chance to review the full pipeline. Happy to adjust anything on this PR in the meantime.

[solderzzc]

Just give you a heads up, we are working on more API providers' integration, and testing them. Will have a release soon.

[octo-patch]

Thanks for the update @solderzzc! Great to see the progress on multi-provider API integration. Looking forward to the release — let me know if there's anything I should adjust on this PR to align with the new architecture.

[solderzzc]

Hi, @octo-patch
We are preparing a new release, it has much more API integrated, so require a bit more time for test, following is an update, the integration doesn't need to change the benchmark code since the API will be set with the existing environment variables:

Provider	Model	Score (Total 96)	Accuracy %	Avg TTFT (ms)	Avg Speed (tok/s)
OpenAI	gpt-5.4-2026-03-05	94	97.9%	601	73.4
Anthropic	claude-opus-4-20250514	93	96.9%	1336	1.8
OpenAI	gpt-5.4-mini-2026-03-17	92	95.8%	553	234.5
Alibaba Cloud	qwen3-max	92	95.8%	1170	5.9
Anthropic	claude-sonnet-4-20250514	91	94.8%	1223	2.6
MiniMax	MiniMax-M2.7-highspeed	90	93.8%	1492	3.0
Moonshot AI	kimi-k2-0905-preview	90	93.8%	2200	62.5
Anthropic	claude-opus-4-6	90	93.8%	1876	2.3
OpenAI	gpt-5.4-nano-2026-03-17	89	92.7%	508	136.4
MiniMax	MiniMax-M2.5-highspeed	89	92.7%	1792	2.8
Alibaba Cloud	qwen-plus	89	92.7%	535	11.7
Anthropic	claude-haiku-4-5	89	92.7%	530	5.3
Anthropic	claude-sonnet-4-6	89	92.7%	1368	2.6
Other Cloud	grok-4-1-fast-non-reasoning	89	92.7%	447	496.1
DeepSeek	deepseek-chat	88	91.7%	1481	21.4
MiniMax	MiniMax-M2.7	88	91.7%	3791	1.5
MiniMax	MiniMax-M2.5	86	89.6%	3230	199.1
Alibaba Cloud	qwen-flash	86	89.6%	398	28.5
Alibaba Cloud	qwen3.5-plus	82	85.4%	1255	16.3
Alibaba Cloud	qwen3.5-flash	81	84.4%	804	26.7
MiniMax	MiniMax-M2.1	77	80.2%	4530	195.7
Moonshot AI	kimi-k2-thinking	70	72.9%	1119	28.3
Moonshot AI	kimi-k2-turbo-preview	66	68.8%	773	134.9
DeepSeek	deepseek-reasoner	63	65.6%	1137	34.5
OpenAI	gpt-5-mini-2025-08-07	60	62.5%	7248	72.6
Moonshot AI	kimi-k2.5	58	60.4%	836	52.6