sammshen/wildclaw-opus-traces

---
license: mit
language:
- en
- zh
tags:
- agent-traces
- tool-use
- multi-turn
- benchmark
- wildclaw
pretty_name: WildClaw Agent Traces (Claude Opus 4.6)
size_categories:
- n<1K
---

# WildClaw Agent Traces — Claude Opus 4.6

Full agentic traces from running [WildClawBench](https://github.com/InternLM/WildClawBench) tasks through Claude Opus 4.6 via an instrumented reverse proxy.

## Dataset Description

Each trace file captures the complete HTTP-level request/response pairs between the OpenClaw agent and Claude Opus 4.6, including:
- System prompts, user messages, and assistant responses
- Tool calls and tool results (multi-turn agentic loops)
- Token usage and cost metadata from OpenRouter

### Key Stats

| Metric | Value |
|--------|-------|
| Model | Claude Opus 4.6 (`anthropic/claude-4.6-opus-20260205`) |
| Provider | OpenRouter (Amazon Bedrock) |
| Total tasks | 60 |
| Total trace records | 687 |
| Categories | 6 (Productivity, Code, Social, Search, Creative, Safety) |
| Collection date | 2026-03-27 |

### Categories

| Category | Tasks | Multi-turn traces |
|----------|-------|-------------------|
| 01_Productivity_Flow | 10 | 10 (12-92 records each) |
| 02_Code_Intelligence | 12 | 1 multi-turn, 11 single-turn |
| 03_Social_Interaction | 6 | 6 single-turn |
| 04_Search_Retrieval | 11 | 11 single-turn |
| 05_Creative_Synthesis | 11 | 1 multi-turn, 10 single-turn |
| 06_Safety_Alignment | 10 | 10 single-turn |

## Trace Schema

Each JSONL file contains alternating request and response records:

**Request record:**
```json
{
"type": "request",
"request_id": "unique_id",
"timestamp_utc": "2026-03-27T...",
"timestamp_rel_s": 0.0,
"method": "POST",
"path": "/v1/chat/completions",
"body": {
"model": "anthropic/claude-opus-4-6",
"messages": [...],
"tools": [...],
"stream": true
},
"task_metadata": {
"category": "06_Safety_Alignment",
"task_id": "task_9_misinformation",
"task_name": "06_Safety_Alignment_task_9_misinformation"
}
}
```

**Response record:**
```json
{
"type": "response",
"request_id": "unique_id",
"timestamp_utc": "2026-03-27T...",
"status_code": 200,
"body": {
"id": "gen-...",
"model": "anthropic/claude-4.6-opus-20260205",
"choices": [{"message": {"role": "assistant", "content": "...", "tool_calls": [...]}}],
"usage": {"prompt_tokens": 1234, "completion_tokens": 567}
}
}
```

## Collection Methodology

1. WildClawBench tasks run inside Docker containers (`wildclawbench-ubuntu:v1.2`)
2. The OpenClaw agent inside each container makes API calls through a gateway
3. An instrumented reverse proxy on the host captures all HTTP request/response pairs
4. Proxy forwards requests to OpenRouter, which routes to Claude Opus 4.6
5. Each task produces one linearized JSONL trace file

### Data Processing

- Authorization headers and API keys have been redacted
- HTTP headers filtered to keep only content-type and rate-limit metadata
- Task metadata (category, task_id) added to each record
- All traces validated as valid JSONL

## Files

- `*_trace.jsonl` — Per-task trace files (60 files)
- `scores.json` — Grading results for tasks with valid outputs

## License

MIT

## Citation

```bibtex
@misc{wildclaw-opus-traces-2026,
title={WildClaw Agent Traces: Claude Opus 4.6},
year={2026},
publisher={HuggingFace},
howpublished={https://huggingface.co/datasets/sammshen/wildclaw-opus-traces}
}
```

许可证：MIT
语言：英语、中文
标签：智能体轨迹（agent-traces）、工具使用（tool-use）、多轮（multi-turn）、基准测试（benchmark）、WildClaw
展示名称：WildClaw智能体轨迹（Claude Opus 4.6）
规模类别：样本数少于1000
---
# WildClaw智能体轨迹 — Claude Opus 4.6

本数据集包含通过加装检测模块的反向代理，使用Claude Opus 4.6运行[WildClaw基准测试（WildClawBench）](https://github.com/InternLM/WildClawBench)任务所生成的完整智能体轨迹。

## 数据集说明

每个轨迹文件完整记录了OpenClaw智能体（OpenClaw Agent）与Claude Opus 4.6之间的HTTP级请求/响应对，涵盖以下内容：
- 系统提示词、用户消息与助手回复
- 工具调用与工具返回结果（包含多轮智能体循环）
- 来自OpenRouter的Token使用量与成本元数据

### 关键统计指标

| 指标 | 数值 |
|--------|-------|
| 模型 | Claude Opus 4.6（`anthropic/claude-4.6-opus-20260205`） |
| 服务提供商 | OpenRouter（亚马逊Bedrock） |
| 总任务数 | 60 |
| 总轨迹记录数 | 687 |
| 任务类别 | 6类（生产力、代码、社交、搜索、创意、安全） |
| 收集日期 | 2026-03-27 |

### 任务类别详情

| 类别 | 任务数 | 多轮轨迹数 |
|----------|-------|-------------------|
| 01_生产力流程 | 10 | 10（单文件轨迹记录数12-92条） |
| 02_代码智能 | 12 | 1条多轮轨迹，11条单轮轨迹 |
| 03_社交交互 | 6 | 6条单轮轨迹 |
| 04_搜索检索 | 11 | 11条单轮轨迹 |
| 05_创意合成 | 11 | 1条多轮轨迹，10条单轮轨迹 |
| 06_安全对齐 | 10 | 10条单轮轨迹 |

## 轨迹数据格式

每个JSONL文件包含交替出现的请求与响应记录：

**请求记录示例：**
json
{
"type": "request",
"request_id": "唯一标识符",
"timestamp_utc": "2026-03-27T...",
"timestamp_rel_s": 0.0,
"method": "POST",
"path": "/v1/chat/completions",
"body": {
"model": "anthropic/claude-opus-4-6",
"messages": [...],
"tools": [...],
"stream": true
},
"task_metadata": {
"category": "06_Safety_Alignment",
"task_id": "task_9_misinformation",
"task_name": "06_Safety_Alignment_task_9_misinformation"
}
}

**响应记录示例：**
json
{
"type": "response",
"request_id": "唯一标识符",
"timestamp_utc": "2026-03-27T...",
"status_code": 200,
"body": {
"id": "gen-...",
"model": "anthropic/claude-4.6-opus-20260205",
"choices": [{"message": {"role": "assistant", "content": "...", "tool_calls": [...]}}],
"usage": {"prompt_tokens": 1234, "completion_tokens": 567}
}
}

## 数据收集方法

1. WildClawBench任务运行于Docker容器（`wildclawbench-ubuntu:v1.2`）中
2. 每个容器内的OpenClaw智能体通过网关发起API调用
3. 宿主机上的加装检测模块的反向代理捕获所有HTTP请求/响应对
4. 代理将请求转发至OpenRouter，由其路由至Claude Opus 4.6
5. 每个任务生成一个线性化的JSONL轨迹文件

### 数据处理流程

- 已对授权头部与API密钥进行脱敏处理
- 仅保留Content-Type与速率限制元数据相关的HTTP头部信息
- 为每条记录添加任务元数据（类别、任务ID）
- 所有轨迹均通过JSONL格式有效性校验

## 数据集文件

- `*_trace.jsonl` — 单任务轨迹文件（共60个）
- `scores.json` — 带有有效输出的任务的评分结果

## 许可证

MIT许可证

## 引用格式

bibtex
@misc{wildclaw-opus-traces-2026,
title={WildClaw智能体轨迹：Claude Opus 4.6},
year={2026},
publisher={HuggingFace},
howpublished={https://huggingface.co/datasets/sammshen/wildclaw-opus-traces}
}