dsh-plugin-xparse

In LLM agent workflows, document parsing is often the weakest link—information buried in PDFs, scans, and Office files resists structured extraction, async task states are hard to track inside a conversation, and paid-quota approvals lack fine-grained control. dsh-plugin-xparse wraps the full TextIn xParse parsing engine into a single Harness Tool so you can parse documents, extract content, and manage the entire task lifecycle directly inside an agent session, without switching to a terminal or a web console.

The capability set spans the full parsing pipeline: parse submits a document; search_text and read_pages pin down specific content; get_outline and get_doc_info expose structural metadata. Twelve task operations—task_run, task_rerun, task_status, task_resume, task_export, task_debug and more—give you complete control over asynchronous job states. Results are written as workspace artifacts rather than dumped into the conversation, keeping context windows clean. Authentication supports both OAuth device-code login and AppKey, with the Secret Code remaining write-only and tokens auto-refreshed; none of these secrets ever leak into settings, tool arguments, or transcript text. Paid-capacity approval is granular: when a job triggers a paid tier, the service returns a waiting state first, and the agent requests one approval only for the exact task_resume action, covering the entire Task Run rather than asking per file.

This plugin is built for developers and teams constructing intelligent document-processing pipelines in DeepSeek Harness—contract review, report summarization, knowledge-base ingestion, or bulk multi-document extraction. As long as your workflow needs to turn unstructured documents into structured data that an agent can consume, the xParse Tool slots directly into your existing orchestration. The bundled CLI binaries, SHA-256 integrity manifests, and session-scoped state directories ensure deployment consistency and auditability out of the box.