In the DeepSeek Harness (DSH) plugin ecosystem, when agents first enter a large codebase, the result most easily obtained is often a conclusion that is “locally correct but globally wrong.” For example, they may assume business boundaries solely from directory structure, or ignore inter-system dependencies. lvxinrong/dsh-archscope addresses this pain point by providing an evidence-based system architecture reconnaissance solution, aimed at helping AI truly take over unfamiliar software systems, rather than merely browsing directories or generating empty summaries.

Plugin Positioning

dsh-archscope is a workflow plugin for DeepSeek Harness. It focuses on “evidence-driven” system architecture reconnaissance, supporting source code discovery, evidence collection, system modeling, and project profiling. The plugin uses a layered analysis model to control conclusion boundaries, ensuring that every layer of analysis—from system worldview to code execution paths—is grounded in evidence.

Core Features

The plugin provides the following core capabilities:

  • Evidence-driven system architecture reconnaissance: An evidence-driven system architecture reconnaissance approach that rejects unverified conclusions.
  • Source discovery and evidence collection: Automatically discovers source code and collects evidence.
  • System modeling and project profiling: Builds a system worldview and engineering configuration profile.
  • Layer gates, parallel orchestration, deterministic validation: Ensures analysis quality through layered gates, parallel orchestration, and deterministic validation.
  • Architecture portal: Provides a visual architecture portal.
  • Native to plugin runtime: Natively adapts to the DSH plugin runtime (Cordis service, capability providers).
  • Multi-repository support: Supports multi-repository analysis, with project identity based on workspace-relative paths.

Analysis Model

The plugin uses a layered model to control analysis order and conclusion boundaries:

  1. System (system layer): Defines the worldview, determining the system’s boundaries, entry points, projects, and infrastructure.
  2. Project (project layer): Defines the engineering configuration, clarifying the project’s role in the system, how it is built, run, and connected.
  3. Internal map: Identifies candidates for business capabilities, technical components, and contract components.
  4. Module study: Analyzes single-responsibility boundaries, entry points, data, and dependencies.
  5. Code trace: Traces execution paths for specific endpoints, tasks, messages, or exceptions.

Usage

When using the plugin, follow the process below:

  1. Establish the worldview: Before profiling a project, first establish system-level common facts, ensuring all agents share context.
  2. Evidence validation: All important conclusions must point to auditable evidence; unverified parts must be marked.
  3. Envelope protocol: Evidence agents and writer agents receive initial information through the “short envelope” protocol, including run ID, project identity, and tool name; they then obtain system facts and templates through read-only tools.
  4. Bounded batching: For single-project or multi-project requests, use the bounded batching engine, ensuring each project receives isolated evidence and writer agents.

Installation and Status

The plugin is not yet published to the NPM package manager and must be obtained from the GitHub repository. The current version is v0.3.3, and it is under active development.

package.json information

{
  "name": "dsh-archscope",
  "version": "0.3.3",
  "description": "Evidence-driven system architecture reconnaissance for DeepSeek Harness",
  "license": "MIT",
  "type": "module"
}

Notes

  1. Runtime evidence: Runtime evidence and recovery orchestration features are still under development.
  2. Project identity: The plugin’s project identity is based on workspace-relative paths, not directory names. This means directories with the same name or nested repositories are not silently merged.
  3. Permissions and security: The plugin runs as a DSH plugin; source code and license (MIT) should be checked before installation.

For developers who need rigorous architecture analysis in DeepSeek Harness, ArchScope provides a standardized reconnaissance protocol.

  • Catalog page: [catalog_url]
  • GitHub repository: [github_url]