kivi-claude-skills

Autonomous ML experiment loop inspired by Karpathy's autoresearch. Modifies training code, runs experiments with fixed time budgets, evaluates metrics, keeps improvements and discards failures via git. Use when the user wants to autonomously optimize ML training, run experiment loops, auto-tune hyperparameters, or set up autonomous research agents. Triggers include "autoresearch", "experiment loop", "autonomous training", "auto-optimize", "hyperparameter search".

AI research pipeline that orchestrates HF Papers API, web search, and content extraction into actionable practitioner briefings. Teaches research thinking — problem decomposition, query diversification, strategic paper reading, and cross-source synthesis. Use when the user wants to research an AI/ML topic, understand a technique, survey recent papers, find state-of-the-art approaches, or bridge academic findings to practical implementation. Triggers include "context-research", "research this AI topic", "what papers exist on", "survey recent work on", "literature review", "SOTA for", "what's the latest research on".

Causal-chain debugging that traces from symptoms to root cause BEFORE making any code changes. Prevents shallow fixes and wrong initial approaches.

Deployment workflow optimized for China-hosted servers with Docker/GHCR/SSH. Handles common failure modes like Docker Hub timeouts and GHCR auth issues.

Fully automated test-fix-retest loop for full-stack projects. Auto-detects tech stack (Java/Spring Boot, Go, Python, Rust, Node.js + Vue/React/Angular/Next.js/Svelte), runs tests in pyramid order (compile → backend unit → frontend typecheck → frontend unit → API E2E → browser E2E), analyzes failures with causal-chain debugging (symptom → proximate cause → root cause), clusters related failures, auto-fixes code, and iterates until all pass or max iterations reached. Also identifies missing test coverage gaps and auto-generates tests when none exist for a phase. Use this skill whenever the user wants to run tests, fix failing tests, or verify code changes haven't broken anything — including: "run tests", "run all tests", "make tests pass", "fix failing tests", "test loop", "fullstack test", "测试", "跑测试", "跑一下所有测试", "测试全挂了帮我修", "帮我修测试", or after code changes touching both frontend and backend (e.g., API contract changes, field naming migrations, new service dependencies). Proactively suggest when the user finishes implementing a feature that spans multiple layers, or when they mention NullPointerException in tests, type errors after refactoring, or cross-module contract mismatches.