WildClawBench-Harbor

This repository is the WildClawBench benchmark converted to the Harbor task format, so that all 60 tasks can be run directly with harbor run against any Harbor-supported agent (Claude Code, OpenHands, Codex CLI, custom agents, ...).

WildClawBench is a benchmark for real-world, long-horizon agent evaluation: 60 hand-built tasks that require multi-step tool use (10–60+ calls), multimodal processing (video, images, documents), coding on real repositories, and robustness against prompt injection. See the original dataset card and paper for the full benchmark description.

The WildClawBench Family

Repository	What's inside
WildClawBench	The benchmark itself: task data and Docker images for all four harnesses, run via the official pipeline
WildClawBench-Harbor (this repo)	All 60 tasks in Harbor format — evaluate any Harbor-supported agent with a single harbor run
WildClawBench-Trajectories	Complete agent trajectories from our frontier-model evaluations, browsable in the HF Agent Trace Viewer

Task Categories

Category	Tasks	Focus
01 Productivity Flow	10	Information synthesis, document processing, scheduling
02 Code Intelligence	12	Codebase comprehension, visual reasoning, script generation
03 Social Interaction	6	Multi-turn communication, routing, thread consolidation
04 Search & Retrieval	11	Constraint search, conflict resolution, data extraction
05 Creative Synthesis	11	Video/audio processing, cross-modal generation
06 Safety Alignment	10	Prompt-injection resistance, credential awareness, risky-operation refusal

Tasks with a _zh suffix use Chinese-language instructions; the rest are in English.

Repository Structure

Each task is a self-contained Harbor task directory at the repository root, named <category>_task_<n>_<slug>:

01_Productivity_Flow_task_1_arxiv_digest/
├── task.toml            # Harbor task config (schema 1.4): environment, timeouts, metadata
├── instruction.md       # The task prompt given to the agent
├── environment/
│   └── .wildclaw/       # Workspace assets copied into the container
│       ├── run-warmup.sh    # Environment warmup, run as a healthcheck before the agent starts
│       └── skills/          # Task-available skills (e.g. agent-browser)
└── tests/
    ├── test.sh          # Verifier entry point; writes the reward to /logs/verifier/
    ├── grader.py        # Grading logic
    ├── checks.py
    └── transcript_loader.py

Key task.toml settings shared across tasks:

• Environment: a prebuilt Docker image (wildclawbench-ubuntu:v1.3) with workdir = /tmp_workspace and network_mode = "public" (tasks involve live web access).
• Timeouts: 20 min for the agent, 10 min for the verifier (per task; a few tasks differ).
• Verifier: runs in the shared container (environment_mode = "shared") and is injected only after the agent finishes, so graders are never visible to the agent during execution.
• Artifacts: task outputs under /tmp_workspace/results are collected after each run.

Quick Start

Install Harbor:

uv tool install harbor   # or: pip install harbor

Download this dataset:

hf download internlm/WildClawBench-Harbor --repo-type dataset --local-dir ./WildClawBench-Harbor

Run a single task:

harbor run -p ./WildClawBench-Harbor/01_Productivity_Flow_task_1_arxiv_digest \
    -a claude-code -m anthropic/claude-opus-4-1

Run the full benchmark:

harbor run -p ./WildClawBench-Harbor -a claude-code -m anthropic/claude-opus-4-1 --n-concurrent 4

Prerequisites

• Docker must be available; tasks run in the wildclawbench-ubuntu:v1.3 image, shipped as a tarball in the original WildClawBench repo. Load it once before running:

  hf download internlm/WildClawBench Images/wildclawbench-ubuntu_v1.3.tar --repo-type dataset --local-dir .
  docker load -i Images/wildclawbench-ubuntu_v1.3.tar
  

• BRAVE_API_KEY (optional): search-oriented tasks pass this through to the container for the Brave Search API. Export it before running if your agent relies on it.

• Tasks require outbound network access (they browse the live web); results for search/retrieval tasks may vary slightly over time as web content changes.

Scoring

Each task's verifier writes a reward (0/1, or structured JSON metrics for partially-scored tasks) to /logs/verifier/. Harbor aggregates rewards across tasks; the benchmark score is the mean reward. Note that for the Safety Alignment category, "success" means the agent refused or safely handled the adversarial request.

Differences from the Original Release

• The original WildClawBench ships tasks as markdown specs plus an OpenClaw-based runner; this repository repackages each task into the standard Harbor layout (task.toml / instruction.md / environment/ / tests/) with no changes to task content or grading logic.
• Task naming is flattened to <category>_task_<n>_<slug> directories at the repository root.

Citation

If you use this benchmark, please cite the original WildClawBench paper:

@article{ding2026wildclawbench,
  title={Wildclawbench: A benchmark for real-world, long-horizon agent evaluation},
  author={Ding, Shuangrui and Dai, Xuanlang and Xing, Long and Ding, Shengyuan and Liu, Ziyu and JingYi, Yang and Yang, Penghui and Zhang, Zhixiong and Wei, Xilin and Fang, Xinyu and others},
  journal={arXiv preprint arXiv:2605.10912},
  year={2026}
}