DeepSeek Harness (DSH) follows an “everything is a plugin” architectural philosophy. When developing agents with DSH, teams often need capabilities such as security auditing, collaborative knowledge accumulation, task orchestration, and Prompt optimization. DeepSeek Companion Enterprise is an enterprise-level companion plugin built on top of DSH, designed to provide development teams with a complete capability matrix covering security compliance and collaboration efficiency. This plugin is maintained by beijingwahw and is open source under the MIT License.

Core Features

Security Auditing and DLP

The plugin includes a built-in rule engine that can automatically identify sensitive information such as API Keys, phone numbers, email addresses, and passwords. It supports custom rules and toggles. Additionally, it provides prompt injection detection, capable of recognizing six attack semantic patterns (instruction override, role jailbreak, etc.), and supports risk scoring and strict-mode blocking. Security auditing provides full operation logs, with export and retrieval support.

Team Collaboration and Knowledge Management

The plugin supports team configuration, including member preferences, default policies, and configuration import/export/diff comparison. Its knowledge management features include session snapshot archiving, team experience accumulation and notes, and intelligent recommendations. The “automatic experience distillation” feature extracts “error → fix” experience cards from session trajectories through signal mining, semantic deduplication, and evidence-chain tracing, forming a reusable knowledge base. The review flow supports multi-round reviews, comments, and decision-making.

Task Orchestration and Automation

The task orchestrator supports multi-step Pipeline definitions, conditional branching, timeout retries, and dependency management. The self-healing execution module provides error classification and exponential backoff retry mechanisms, with a built-in tri-state circuit breaker model and automatic fallback when the primary model fails. It also provides scheduled tasks (Cron scheduling, checkpoint resumption) and task queues (status tracking).

Developer Efficiency and Foundational Capabilities

For developer efficiency, the plugin provides a multi-model arena for parallel comparison and historical evaluation; it uses SPC statistical process control to monitor trajectory metric drift; the Prompt engineering workbench supports version management and template libraries; API cost governance provides real-time pricing and budget control; and adaptive model routing optimizes model selection using multi-armed bandit algorithms. Foundational capabilities include intelligent conversation export (Markdown/PDF/JSON/HTML), context handoff summaries, structured hierarchical handoff, global conversation search, and semantic neighborhood retrieval.

Installation and Enablement

Before installation, ensure that DeepSeek Harness and the corresponding environment requirements are installed. The installation command is as follows:

dsh plugin add beijingwahw/dsh-companion-enterprise --profile web

After startup, the plugin panel will load automatically. Common advanced commands include:

# 升级插件
dsh plugin upgrade dsh-companion-enterprise --profile web

# 卸载插件
dsh plugin remove dsh-companion-enterprise --profile web

# 从本地源码安装
dsh plugin add ./dsh-companion-enterprise --profile web

Permissions and Privacy Notes

  • Network permissions: Accesses only api.deepseek.com for model calls, and five official pricing pages (GET requests only, used for dynamic pricing).
  • Storage: Uses only the companion storage domain, with data stored locally in the Harness plugin sandbox.
  • Privacy: No user behavior tracking, no telemetry, and no data reporting; conversation content and API Keys are stored locally only.

Notes

This plugin runs with the permissions of the current DSH process. It is recommended to review the source code and license before use. All modules can be independently toggled in cordis.patch.yml.

For more technical details, refer to the GitHub repository.