WB-Bench-Web Report Workflow

Generate an analysis report from one WB-Bench-Web single Harbor run artifact directory. Analyze only the evaluation results under the input path. Use WB-Bench-Web as the display name, with no version suffix.

Dataset Context

The public dataset id is wb-bench-web. If the user provides a WorkBuddy Bench repository or dataset directory, read task metadata from a path shaped like /path/to/workbuddy-bench/datasets/wb-bench-web. If no dataset source path is provided, analyze only the RUN_DIR artifacts and do not assume a dataset source path.

The current task set has 70 tasks:

• Categories: page-interaction 21 tasks, data-visualization 15 tasks, visual-design 9 tasks, analytical-report 7 tasks, code-testing 7 tasks, page-implementation 6 tasks, and document-conversion 5 tasks.
• Task modes: From Scratch 35 tasks, Bug Fix 8 tasks, Extend Existing 8 tasks, Review & Analysis 7 tasks, Test Generation 7 tasks, and Format Conversion 5 tasks.
• Interaction complexity: no interaction, light interaction, single-flow state, multi-step workflows, persistence/offline behavior, and cross-state behavior.

Web tasks use the CompositeVerifier web profile. Evaluation artifacts may include rule, VLM chat, and agent-judge evidence. Prefer structured score.json and eval-report evidence. Use screenshots only as supporting evidence.

Input

Required:

• RUN_DIR: a single Harbor run directory. Standard layout: <RUN_DIR>/result.json, config.json, job.log, and