rename project from robot-harness to roboharness

Rename the package before its initial PyPI release for a catchier,
more memorable name aligned with projects like roboflow and robocorp.

Changes span:
- pyproject.toml: package name, optional deps, wheel config
- src/robot_harness/ → src/roboharness/: directory and all imports
- All test, example, and documentation files updated
- CI workflows, issue templates, and SVG assets updated

Closes #21

https://claude.ai/code/session_01LA2Uo7hCVdpo5gjGXief4H

Author: claude

.github/ISSUE_TEMPLATE/bug_report.md

@@ -1,6 +1,6 @@
 ---
 name: Bug Report
-about: Report a bug in Robot-Harness
+about: Report a bug in Roboharness
 title: "[Bug] "
 labels: bug
 assignees: ''
@@ -14,7 +14,7 @@ A clear description of the bug.
 
 - OS:
 - Python version:
-- Robot-Harness version:
+- Roboharness version:
 - Simulator (MuJoCo/Isaac Lab/ManiSkill/etc.):
 - Simulator version:
 

.github/workflows/publish.yml

@@ -20,7 +20,7 @@ jobs:
       - name: Verify package
         run: |
           pip install dist/*.whl
-          python -c "import robot_harness; print('Install OK')"
+          python -c "import roboharness; print('Install OK')"
       - uses: actions/upload-artifact@v4
         with:
           name: dist

CONTRIBUTING.md

@@ -1,6 +1,6 @@
-# Contributing to Robot-Harness
+# Contributing to Roboharness
 
-Thank you for your interest in contributing! Robot-Harness is in early development and we welcome contributions of all kinds.
+Thank you for your interest in contributing! Roboharness is in early development and we welcome contributions of all kinds.
 
 ## How to Contribute
 
@@ -34,7 +34,7 @@ Thank you for your interest in contributing! Robot-Harness is in early developme
 
 We especially welcome new simulator backends! To add one:
 
-1. Create a new file in `src/robot_harness/backends/`
+1. Create a new file in `src/roboharness/backends/`
 2. Implement the `SimulatorBackend` protocol (see `core/harness.py`)
 3. Add an example in `examples/`
 4. Add the simulator's dependencies to `pyproject.toml` as an optional dependency group

README.md

@@ -1,6 +1,6 @@
 <div align="center">
 
-# Robot-Harness
+# Roboharness
 
 **A Visual Testing Harness for AI Coding Agents in Robot Simulation**
 
@@ -20,26 +20,26 @@
 
 </div>
 
-## What is Robot-Harness?
+## What is Roboharness?
 
-Robot-Harness is a framework that lets AI Coding Agents (Claude Code, OpenAI Codex, [OpenClaw](https://github.com/openclaw/openclaw), etc.) control robot simulations through a **visual feedback loop**:
+Roboharness is a framework that lets AI Coding Agents (Claude Code, OpenAI Codex, [OpenClaw](https://github.com/openclaw/openclaw), etc.) control robot simulations through a **visual feedback loop**:
 
 <p align="center">
-  <img src="assets/architecture.svg" width="800" alt="Robot-Harness Architecture"/>
+  <img src="assets/architecture.svg" width="800" alt="Roboharness Architecture"/>
 </p>
 
-**Key insight**: Modern coding agents are already multimodal — they can write code AND see images AND make decisions. We don't need a separate VLM. Robot-Harness just needs to present simulation visuals in a format agents can directly consume.
+**Key insight**: Modern coding agents are already multimodal — they can write code AND see images AND make decisions. We don't need a separate VLM. Roboharness just needs to present simulation visuals in a format agents can directly consume.
 
 ## Installation
 
 ```bash
-pip install robot-harness
+pip install roboharness
 
 # With MuJoCo + Meshcat backend
-pip install robot-harness[mujoco]
+pip install roboharness[mujoco]
 
 # Development
-pip install robot-harness[dev]
+pip install roboharness[dev]
 ```
 
 ## Quick Start
@@ -49,7 +49,7 @@ pip install robot-harness[dev]
 Run a complete grasp simulation with zero external dependencies:
 
 ```bash
-pip install robot-harness[mujoco] Pillow
+pip install roboharness[mujoco] Pillow
 python examples/mujoco_grasp.py --report
 ```
 
@@ -73,7 +73,7 @@ Wrap any Gymnasium-compatible environment with one line:
 
 ```python
 import gymnasium as gym
-from robot_harness.wrappers import RobotHarnessWrapper
+from roboharness.wrappers import RobotHarnessWrapper
 
 env = gym.make("CartPole-v1", render_mode="rgb_array")
 env = RobotHarnessWrapper(env,
@@ -102,8 +102,8 @@ For custom simulator integrations:
 <summary><b>Show code example</b></summary>
 
 ```python
-from robot_harness import Harness
-from robot_harness.backends.mujoco_meshcat import MuJoCoMeshcatBackend
+from roboharness import Harness
+from roboharness.backends.mujoco_meshcat import MuJoCoMeshcatBackend
 
 backend = MuJoCoMeshcatBackend(
     model_path="robot.xml",
@@ -130,7 +130,7 @@ For tasks with multiple grasp positions, each with multiple agent retry trials:
 <summary><b>Show code example and directory structure</b></summary>
 
 ```python
-from robot_harness.storage import GraspTaskStore
+from roboharness.storage import GraspTaskStore
 
 store = GraspTaskStore(base_dir="./output", task_name="pick_and_place")
 store.add_grasp_position(position_id=1, xyz=(0.5, 0.0, 0.05), object_name="red_cube")
@@ -177,7 +177,7 @@ harness_output/
 <summary><b>Project structure</b></summary>
 
 ```
-robot_harness/
+roboharness/
 ├── core/
 │   ├── harness.py         # Main Harness class + SimulatorBackend protocol
 │   ├── checkpoint.py      # Checkpoint management & state snapshots
@@ -218,7 +218,7 @@ We especially welcome:
 - Real-world usage examples
 - Integration with popular RL libraries (SB3, CleanRL, etc.)
 
-**AI agents are welcome contributors!** We actively encourage contributions from AI coding agents such as [Claude Code](https://docs.anthropic.com/en/docs/claude-code), [OpenAI Codex](https://github.com/openai/codex), [OpenClaw](https://github.com/openclaw/openclaw), and other autonomous coding tools. If your agent can improve Robot-Harness, send a PR!
+**AI agents are welcome contributors!** We actively encourage contributions from AI coding agents such as [Claude Code](https://docs.anthropic.com/en/docs/claude-code), [OpenAI Codex](https://github.com/openai/codex), [OpenClaw](https://github.com/openclaw/openclaw), and other autonomous coding tools. If your agent can improve Roboharness, send a PR!
 
 ## License
 

assets/architecture.svg

@@ -1,6 +1,6 @@
 # CI Strategy: CPU + GPU Testing
 
-> 讨论背景：随着 Robot-Harness 集成 cuRobo 规划、WBC 运控、Policy 推理等 GPU 依赖组件，
+> 讨论背景：随着 Roboharness 集成 cuRobo 规划、WBC 运控、Policy 推理等 GPU 依赖组件，
 > CI 需要从纯 CPU 扩展到 GPU 测试。本文档记录分层方案和平台选型。
 
 ## 当前状态

docs/ci-strategy.md

@@ -1,27 +1,27 @@
-# Robot-Harness: Context Document
+# Roboharness: Context Document
 
 > **Document Version**: v0.1-draft | **Date**: 2026-04-02
-> **Purpose**: This is the complete context document for the robot-harness project, intended as a reference for Claude Code, Codex, and other AI Agents during code review, architecture design, and feature development.
+> **Purpose**: This is the complete context document for the roboharness project, intended as a reference for Claude Code, Codex, and other AI Agents during code review, architecture design, and feature development.
 
 ## Part 1: Project Overview and Motivation
 
-### 1.1 What is Robot-Harness
+### 1.1 What is Roboharness
 
-Robot-Harness is a **visual testing framework for AI Coding Agents in robot simulation**. Its core goal is to enable Claude Code, OpenAI Codex, and similar coding agents to:
+Roboharness is a **visual testing framework for AI Coding Agents in robot simulation**. Its core goal is to enable Claude Code, OpenAI Codex, and similar coding agents to:
 
 1. **Control simulation stepping** (step-by-step execution) — pause simulation at critical moments
 2. **Capture multi-view screenshots** — acquire RGB/depth images from different camera positions at the same simulation moment
 3. **Autonomously judge task results** — agent directly observes screenshots to determine whether motion is reasonable, grasps are successful, etc.
 4. **Iteratively optimize algorithms** — based on visual judgment results, the agent autonomously modifies control code and reruns
 
-**Fundamental difference from traditional approaches**: We don't need a separate VLM model for visual evaluation. Claude Code and Codex themselves are multimodal agents — they can write code, see images, and make decisions. Robot-Harness's responsibility is to **efficiently present simulation visual information in a format that agents can directly consume**.
+**Fundamental difference from traditional approaches**: We don't need a separate VLM model for visual evaluation. Claude Code and Codex themselves are multimodal agents — they can write code, see images, and make decisions. Roboharness's responsibility is to **efficiently present simulation visual information in a format that agents can directly consume**.
 
 ### 1.2 Core Use Case
 
 Taking a grasping task as an example, the complete Agent-in-the-loop workflow is:
 
 1. Agent writes/modifies grasp control code
-2. Robot-Harness runs the simulation, automatically pausing at predefined checkpoints (plan start, plan end, contact point, lift complete)
+2. Roboharness runs the simulation, automatically pausing at predefined checkpoints (plan start, plan end, contact point, lift complete)
 3. At each checkpoint, the Harness captures screenshots from multiple viewpoints and saves them as files
 4. Agent examines screenshots + structured state data, judging whether the current phase is normal
 5. If problems are found, the agent modifies code and reruns from the appropriate checkpoint
@@ -165,7 +165,7 @@ This workflow is running and producing real results, but is currently a custom i
 - **VLM-RMs** (ICLR 2024): CLIP cosine similarity as zero-shot reward signal
 - **RL-VLM-F** (ICML 2024): VLM compares image pairs to learn reward functions
 - **RoboCLIP** (NeurIPS 2023): Video-language model computes trajectory similarity
-- **StepEval** (2025): Sub-goal decomposition + VLM stage-by-stage evaluation — closest to the evaluation granularity needed by robot-harness
+- **StepEval** (2025): Sub-goal decomposition + VLM stage-by-stage evaluation — closest to the evaluation granularity needed by roboharness
 - **Robo2VLM** (2025): Generates 684,710 VQA questions from 176K real trajectories
 
 ### 3.4 AI Agent-Driven Simulation Iteration Frameworks

docs/context.en.md

@@ -1,27 +1,27 @@
-# Robot-Harness: Context Document
+# Roboharness: Context Document
 
 > **文档版本**: v0.1-draft | **日期**: 2026-04-02
-> **用途**: 本文件是 robot-harness 项目的完整上下文文档（Context Document），供 Claude Code、Codex 等 AI Agent 在进行代码 Review、架构设计和功能开发时作为背景参考。
+> **用途**: 本文件是 roboharness 项目的完整上下文文档（Context Document），供 Claude Code、Codex 等 AI Agent 在进行代码 Review、架构设计和功能开发时作为背景参考。
 
 ## 第一部分：项目概述与动机
 
-### 1.1 什么是 Robot-Harness
+### 1.1 什么是 Roboharness
 
-Robot-Harness 是一个 **为 AI Coding Agent 设计的机器人仿真视觉测试框架**。它的核心目标是让 Claude Code、OpenAI Codex 等编程 Agent 能够：
+Roboharness 是一个 **为 AI Coding Agent 设计的机器人仿真视觉测试框架**。它的核心目标是让 Claude Code、OpenAI Codex 等编程 Agent 能够：
 
 1. **控制仿真步进**（step-by-step execution）— 在关键时刻暂停仿真
 2. **采集多视角截图** — 在同一仿真时刻从不同相机位置获取 RGB/深度图像
 3. **自主判断任务结果** — Agent 直接观察截图，判断运动是否合理、抓取是否成功
 4. **迭代优化算法** — 基于视觉判断结果，Agent 自主修改控制代码并重跑
 
-**与传统方案的根本区别**：我们不需要单独的 VLM 模型来做视觉评估。Claude Code 和 Codex 本身就是多模态 Agent，它们既能写代码、又能看图片、又能做决策。Robot-Harness 的职责是 **把仿真的视觉信息以 Agent 能直接消费的格式高效地呈现出来**。
+**与传统方案的根本区别**：我们不需要单独的 VLM 模型来做视觉评估。Claude Code 和 Codex 本身就是多模态 Agent，它们既能写代码、又能看图片、又能做决策。Roboharness 的职责是 **把仿真的视觉信息以 Agent 能直接消费的格式高效地呈现出来**。
 
 ### 1.2 核心使用场景
 
 以抓取任务为例，完整的 Agent-in-the-loop 流程是：
 
 1. Agent 编写/修改抓取控制代码
-2. Robot-Harness 运行仿真，在预定义的检查点（plan 开始、plan 结束、接触点、抬升完成）自动暂停
+2. Roboharness 运行仿真，在预定义的检查点（plan 开始、plan 结束、接触点、抬升完成）自动暂停
 3. 在每个检查点，Harness 从多个视角截图并保存为文件
 4. Agent 查看截图 + 结构化状态数据，判断当前阶段是否正常
 5. 如果发现问题，Agent 修改代码并从适当的检查点重跑
@@ -165,7 +165,7 @@ Codex 分析截图 → 判断抓取效果
 - **VLM-RMs**（ICLR 2024）：CLIP 余弦相似度作为零样本奖励信号
 - **RL-VLM-F**（ICML 2024）：VLM 对比图像对学习奖励函数
 - **RoboCLIP**（NeurIPS 2023）：视频-语言模型计算轨迹相似度
-- **StepEval**（2025）：子目标分解 + VLM 逐阶段评估 — 最接近 robot-harness 需要的评估粒度
+- **StepEval**（2025）：子目标分解 + VLM 逐阶段评估 — 最接近 roboharness 需要的评估粒度
 - **Robo2VLM**（2025）：从 176K 真实轨迹生成 684,710 个 VQA 问题
 
 ### 3.4 AI Agent 驱动的仿真迭代框架

docs/context.zh-CN.md

@@ -1,4 +1,4 @@
-# Technical Survey of Open-Source Robot Control Projects and Robot-Harness Integration Analysis
+# Technical Survey of Open-Source Robot Control Projects and Roboharness Integration Analysis
 
 > The Gymnasium Wrapper approach can cover 60% of mainstream projects, but two major ecosystems (legged_gym family and JAX/Brax family) require dedicated adapters.
 
@@ -45,7 +45,7 @@ Video recording is supported via `gymnasium.wrappers.RecordVideo` (requires `--e
 
 Known issues: headless mode rendering hangs (#324), camera-enabled environments not rendering (#3250), recordings missing debug markers (#2233), WebRTC errors in Docker (#3192).
 
-### Robot-Harness Integration Feasibility: Very High
+### Roboharness Integration Feasibility: Very High
 
 The standard Gymnasium Wrapper approach is directly applicable:
 
@@ -77,7 +77,7 @@ The configuration system uses nested Python classes (neither dataclasses nor YAM
 
 **No automated tests, no CI/CD** — this is a notable weakness of the project.
 
-### Robot-Harness Integration Feasibility: Moderately Difficult
+### Roboharness Integration Feasibility: Moderately Difficult
 
 Five fundamental incompatibilities exist:
 1. **Vectorized vs. single-instance**: Isaac Gym runs 4096+ parallel environments simultaneously; Gymnasium expects a single instance
@@ -102,7 +102,7 @@ Entirely based on JAX functional programming. Environments are pure functions: `
 
 Batch parallelism is achieved via `jax.vmap()`, and training loops complete full rollouts on-device using `jax.lax.scan`. Single-threaded MJX on GPU is 10x slower than CPU MuJoCo — its advantage comes entirely from massive parallelism (batch sizes of 1024–8192+).
 
-### Robot-Harness Integration Feasibility: Difficult
+### Roboharness Integration Feasibility: Difficult
 
 The standard Gymnasium Wrapper cannot be used directly because:
 1. **Explicit state passing vs. internal state**: The Wrapper must maintain `self._state` internally
@@ -137,13 +137,13 @@ env = ManiSkillVectorEnv(env, auto_reset=True)
 env = gym.make("PickCube-v1", num_envs=N)
 ```
 
-### Robot-Harness Integration Feasibility: Very High
+### Roboharness Integration Feasibility: Very High
 
 The CPUGymWrapper mode is directly compatible with the standard Gymnasium Wrapper. The ManiSkillVectorEnv mode requires handling batched GPU data, similar to Isaac Lab. The RecordEpisode wrapper provides built-in recording functionality.
 
 ---
 
-## Robot-Harness Integration Summary
+## Roboharness Integration Summary
 
 | Project | Integration Feasibility | Integration Method | Estimated Effort |
 |---------|------------------------|-------------------|-----------------|

docs/simulator-survey.en.md

@@ -1,4 +1,4 @@
-# 机器人控制开源项目技术调研与 Robot-Harness 适配分析
+# 机器人控制开源项目技术调研与 Roboharness 适配分析
 
 > Gymnasium Wrapper 方案能覆盖主流项目的 60%，但两大生态系统（legged_gym 系与 JAX/Brax 系）需要专门适配器。
 
@@ -45,7 +45,7 @@ Isaac Lab 是目前最成熟的 Gymnasium 原生机器人学习框架，GitHub 
 
 已知问题：无头模式渲染卡死（#324）、启用相机后不渲染（#3250）、录制不包含调试标记（#2233）、Docker 中 WebRTC 错误（#3192）。
 
-### Robot-Harness 接入可行性：极高
+### Roboharness 接入可行性：极高
 
 标准 Gymnasium Wrapper 模式直接适用：
 
@@ -77,7 +77,7 @@ env = Sb3VecEnvWrapper(env)     # RL 库 wrapper 必须在最后
 
 **无自动化测试，无 CI/CD** — 这是该项目的显著弱点。
 
-### Robot-Harness 接入可行性：中等偏难
+### Roboharness 接入可行性：中等偏难
 
 存在五个根本性不兼容：
 1. **向量化 vs 单实例**：Isaac Gym 同时运行 4096+ 并行环境，Gymnasium 期望单实例
@@ -102,7 +102,7 @@ Google DeepMind 的 GPU 加速机器人学习框架，覆盖 50+ 环境（DM Con
 
 批量并行通过 `jax.vmap()` 实现，训练循环通过 `jax.lax.scan` 在设备上完成完整 rollout。单线程 MJX 在 GPU 上比 CPU MuJoCo 慢 10 倍 — 其优势完全来自大规模并行（batch 1024–8192+）。
 
-### Robot-Harness 接入可行性：困难
+### Roboharness 接入可行性：困难
 
 标准 Gymnasium Wrapper 无法直接使用，原因是：
 1. **状态传递 vs 内部状态**：需在 Wrapper 内维护 `self._state`
@@ -137,13 +137,13 @@ env = ManiSkillVectorEnv(env, auto_reset=True)
 env = gym.make("PickCube-v1", num_envs=N)
 ```
 
-### Robot-Harness 接入可行性：极高
+### Roboharness 接入可行性：极高
 
 CPUGymWrapper 模式直接适用标准 Gymnasium Wrapper。ManiSkillVectorEnv 模式需要处理批量 GPU 数据，与 Isaac Lab 类似。RecordEpisode wrapper 已内建录制功能。
 
 ---
 
-## Robot-Harness 接入总结
+## Roboharness 接入总结
 
 | 项目 | 接入可行性 | 接入方式 | 预估工作量 |
 |------|-----------|---------|-----------|