Skip to content
View Harzva's full-sized avatar

Highlights

  • Pro

Block or report Harzva

Block user

Prevent this user from interacting with your repositories and sending you notifications. Learn more about blocking users.

You must be logged in to block users.

Maximum 250 characters. Please don’t include any personal information such as legal names or email addresses. Markdown is supported. This note will only be visible to you.
Report abuse

Contact GitHub support about this user’s behavior. Learn more about reporting abuse.

Report abuse
Harzva/README.md
Harzva Avatar

Harzva — Make Coding Agents Cheaper

Portfolio GitHub ORCID Email
Token Saver Engineering Research

I work on making long-horizon coding agents cheaper by optimizing prompt-cache reuse at the harness level.

我最近主要做 AI coding agent 的 token 成本优化,同时继续关注视觉语言模型、少样本学习和 Claude Code 生态下的 Agent 工具链。

完整作品集 · Agent 项目 · 论文详情 · 文章专题


适合谁看?

  • 想快速了解我是谁? 这里是 GitHub 入口名片,只保留最重要的信息。
  • 想看完整作品集? 访问 harzva.github.io
  • 想看论文图表和详情? 每篇论文都有独立页面和分析入口。
  • 想看 Agent 工程项目? 这里列出核心仓库,完整说明放在作品集网站。
  • 想看 token saver 方向? 我最近关注 make agents cheaper:通过 harness 层的 prompt cache 复用降低 uncached input 成本。
  • 想看我的产品和工程思考? 文章专题会持续记录 agentic engineering、案例分析和项目复盘。

入口导航

入口 内容
Portfolio 完整作品集、论文、项目、文章和个人介绍
Agent Projects Claude Code 生态、MCP、多 Agent 编排、插件脚手架
Token Saver AI coding agent 的 token 成本优化、prompt cache 复用、trace 解析与 A/B 评估
Publications 论文详情页、关键图表、方法分析
Article Analysis 产品、增长、工程和分发相关案例分析
Agentic Engineering 从 token 到产品、APP 即上下文工程、过程即内容等思考

当前方向

Direction Focus
Token Saver / Make Agents Cheaper Prompt-cache reuse at the harness level, trace parsing, prompt layout optimization, A/B evaluation
AI Agent Infrastructure Claude Code ecosystem, MCP, multi-agent orchestration, plugin scaffolds
Vision-Language Research Few-shot learning, zero-shot learning, compositional generalization, multimodal prompts
Engineering Notes Agentic engineering, product thinking, development process, case analysis

我最近主要做 AI coding agent 的 token 成本优化。核心不是简单压缩上下文,而是研究 agent harness 里哪些内容稳定、哪些内容动态,并通过 prompt layout 调整、trace 解析和 A/B 评估,提高 prompt cache hit rate,减少实际付费的 uncached input,同时保持任务成功率。


核心项目

Project Description
learn-likecc
stars
Claude Code 逆向工程、source map 恢复、运行时分析与工作流重建
codex-managed-agent
stars
面向 Codex / coding agent 的多智能体管理实验,关注任务编排、状态管理和执行控制
keep-claude-cheaper
stars
Token Saver 方向:研究 prompt-cache reuse、trace 解析和 harness 层成本优化
HyperMemory
stars
Memory-enhanced AI knowledge system,连接 RAG、Agent、Wiki、GBrain 和层级记忆
ChinaAI-Roadmaps
stars
中国开源 / 开放权重大模型公司的技术路线整理:GLM、Kimi、DeepSeek、MiniMax
LDC
stars
CVPR 2025 论文 Logits DeConfusion with CLIP for Few-Shot Learning 的代码仓库

论文与研究

研究主题覆盖 Few-Shot Learning、Zero-Shot Learning、Vision-Language Models 和多模态理解。完整图表和方法分析放在作品集网站。

# Paper Venue Year
01 G2D: Discriminative-Generative Collaborative Inference for Zero-Shot Image Classification arXiv 2025
02 Preserving Text Space Integrity for Robust Compositional Zero-Shot Learning via Mixture of Pretrained Experts Neurocomputing 2025
03 Text Augmentation for Vision Knowledge-Based Systems 2026
04 Logits DeConfusion CVPR 2025
05 PromptVAD: Prompt-Based Video Anomaly Detection TNNLS 2023
06 MinEnt: Minimum Entropy for Self-Supervised Representation Learning Pattern Recognition 2023
07 LF2CS: Learning Features into Clustering Space for Few-Shot Image Classification ECCV 2022
08 Text Generation and Multi-Modal Knowledge Transfer for Few-Shot Object Detection Pattern Recognition 2025
09 Augmentative Contrastive Learning for One-Shot Object Detection Neurocomputing 2022
10 ViLT-CLIP: Video and Language Tuning CLIP with Multimodal Prompt Learning AAAI 2024
11 LLM Knowledge-Driven Target Prototype Learning for Few-Shot Segmentation Knowledge-Based Systems 2025

文章与思考


技术栈

Python · TypeScript · React · Node.js · PyTorch · Docker


联系

Pinned Loading

  1. learn-likecc learn-likecc Public

    Claude Code 源码逆向恢复项目 | Source Map 逆向 · 架构分析 · 可运行版本 | 1900+ 文件 · 51万行代码 · 12章节课程

    TypeScript 180 39

  2. codex-managed-agent codex-managed-agent Public

    JavaScript 102 24

  3. Just-Agent/like-code Just-Agent/like-code Public

    A customized Claude Code fork for routed models, multi-agent orchestration, and dense terminal control.

    TypeScript 58 2

  4. LDC LDC Public

    Forked from LiShuo1001/LDC

    The code of "Logits DeConfusion with CLIP for Few-Shot Learning" (CVPR 2025)

    Python

  5. make-agents-cheaper make-agents-cheaper Public

    Forked from Just-Agent/make-agents-cheaper

    Rust CLI for improving prompt cache hit rate in coding-agent workflows

    Rust

  6. Oh-Reflective-loop-skills Oh-Reflective-loop-skills Public

    Forked from Just-Agent/Oh-Reflective-loop-skills

    Python