Skip to content

README.md

Latest commit

 

History

History
93 lines (63 loc) · 3.35 KB

README.md

File metadata and controls

93 lines (63 loc) · 3.35 KB

English | 中文

MindSpore Model Agent

MindSpore Model Agent is a training-focused AI agent solution for the MindSpore ecosystem. It is designed for the high-frequency engineering work around model training, where users need more than general code generation and need help with domain-specific training tasks.

It is built on two closely related parts:

  • mindspore-skills: the domain capability layer for model training and debugging tasks. It provides reusable skills for readiness checking, failure diagnosis, accuracy analysis, performance analysis, model migration, algorithm adaptation, and operator implementation. These skills can work not only with MindSpore Model Agent, but also with other agentic CLI environments such as Claude Code, OpenCode, and Codex.
  • mindspore-cli: the official CLI of MindSpore Model Agent. It provides better integration with related skills and is optimized for model training use cases, offering a more unified end-to-end experience for training-oriented workflows.

Latest Version

Latest version: MindSpore Model Agent v0.1.3. See changelog.md for update history.

Highlights:

  • [skills] Added baseline analysis support for Ascend A2 training runtime failures, accuracy drift, and performance bottlenecks.
  • [skills] Added Hugging Face Transformers model migration support and integrated mhc / attn-residual into the Qwen3 skill template.
  • [skills] Integrated openjiuwen claw, with precision-location examples and deployment guidance.
  • [cli] Improved live task progress feedback, including during hidden tool-call assembly.
  • [cli] Added diff view for edit-style tool results and improved transcript layout and readability.
  • [cli] Fixed shell interrupt handling in truncated streaming output scenarios, unified bug / issue data structures, and fixed GitCode-incompatible install examples.

MindSpore CLI

MindSpore CLI is the official end-to-end interface of MindSpore Model Agent. It is designed to provide a unified CLI experience for training-oriented workflows, with tighter integration with the related skills behind the solution.

Installation

Install from script

curl -fsSL https://raw.githubusercontent.com/mindspore-lab/mindspore-cli/main/scripts/install.sh | bash

Build from source

Go 1.24.2+:

git clone https://github.com/mindspore-lab/mindspore-cli.git
cd mindspore-cli
go build -o mscli ./cmd/mscli
./mscli

Quick Start

Use the free built-in model

mscli
# Choose "mscli-provided" → "kimi-k2.5 [free]" on first run

Bring your own API key

export MSCLI_API_KEY=sk-...
export MSCLI_MODEL=deepseek-chat
mscli

Use OpenAI / Anthropic / OpenRouter

# OpenAI
export MSCLI_PROVIDER=openai-completion
export MSCLI_API_KEY=sk-...
export MSCLI_MODEL=gpt-4o

# Anthropic
export MSCLI_PROVIDER=anthropic
export MSCLI_API_KEY=sk-ant-...
export MSCLI_MODEL=claude-sonnet-4-20250514

# OpenRouter
export MSCLI_PROVIDER=openai-completion
export MSCLI_API_KEY=sk-or-...
export MSCLI_BASE_URL=https://openrouter.ai/api/v1

mscli

Documentation

Contributing

See the Contributor Guide for code style, dependency rules, and testing conventions.