Introduction

DeepSeek Harness (DSH) uses a plugin-based architecture, allowing developers to extend its capabilities. In HR technology and recruitment due diligence scenarios, decisions often rely on subjective judgments or unverified promotional statements. dsh-scout (Scout) aims to introduce an “evidence-driven” workflow, helping developers or users establish a standardized due diligence process in DeepSeek Harness, bind conclusions to real sources, and distinguish between “verified” and “pending verification” statuses.

Plugin Overview

dsh-scout is an evidence-driven company and job due diligence/background check plugin for DeepSeek Harness. Its core value is requiring every conclusion to be linked to an authentic source and limiting the strength of a conclusion so it cannot exceed the evidence level of its sources. It does not replace browsers for deep scraping, and it does not treat company self-descriptions or funding news as verified facts.

The plugin is maintained by MaxHou-infinity and is open-sourced under the MIT license.

Core Concept: Evidence Levels

The plugin introduces a four-level evidence constraint mechanism from E0 to E3. The evidenceLevel of every conclusion must not exceed the strongest source it cites:

  • E0: No source / model inference. Unsupported guesses can only be marked as unknown or needs_verification.
  • E1: User-provided. Information that the user obtains from internal channels.
  • E2: Authoritative third parties. Independent media, recruitment platforms (BOSS/Liepin), business registry aggregation services (Tianyancha/Aiqicha), and Wikipedia.
  • E3: Official primary sources. The National Enterprise Credit Information Publicity System (gsxt.gov.cn), regulatory announcements, and official filings.

The plugin enforces the following discipline:
1. The evidenceLevel of a conclusion must not exceed the strongest cited source.
2. High-impact conclusions without sources must be marked as unknown or needs_verification.
3. Identity verification (scout_verify_identity) must be linked to an E3 official registry source.
4. Decisions are conservative by default; as long as the entity has not been verified or a blocking-level conclusion has not been verified, keep the decision as VERIFY.

Toolchain

The plugin provides a complete due diligence toolchain covering the entire process from case creation to report generation:

  • scout_start: Creates a due diligence case.
  • scout_search: Invokes the Web Provider to search and automatically registers result sources.
  • scout_ingest: Bulk-registers collected results and can automatically infer source type and evidence level.
  • scout_add_source: Manually registers a source (specifying sourceType and evidenceLevel).
  • scout_add_claim: Adds a conclusion (subject to evidence boundary constraints).
  • scout_verify_identity: Verifies the company entity using an E3 official source.
  • scout_verify_claim: Promotes a claim to verified using stronger evidence.
  • scout_report: Renders the current Markdown report.
  • scout_questions: Generates an interview question list.
  • scout_compare: Compares multiple cases side by side.
  • scout_export: Performs persistent five-file export.
  • scout_import: Recovers a case from an export directory.

Typical Usage

Taking “determining whether an HR role at a given company is worth pursuing” as an example, the workflow is as follows:

  1. Create a case
scout_start(caseId=acme-hr, companyName=某科技公司, roleTitle=HRBP, location=深圳)
  1. Register sources
scout_add_source(sourceId=s1, sourceType=company_registry, evidenceLevel=E2)
scout_add_source(sourceId=s2, sourceType=company_official, evidenceLevel=E2)
  1. Add a claim
scout_add_claim(claimId=c1, status=verified, evidenceLevel=E2, content="薪资范围符合市场水平")

For high-impact items without sources (such as reporting lines), mark them as unknown.

  1. Verify identity
scout_verify_identity(...)
  1. Export report
scout_export

Installation and Enablement

The plugin requires Node.js version >= 22.19 and targets DeepSeek Harness 0.1.2-rc.1, @deepseek-ai/dsh-tools 0.1.2-rc.1, and @deepseek-ai/cordis 4.0.2+.

The installation commands are as follows:

dsh plugin --profile scout-demo add github:MaxHou-infinity/dsh-scout#<commit>
dsh --profile scout-demo --dump-config   # 验证工具挂载

Configuration (Optional)

Configure dsh-scout in the DSH profile’s cordis.patch.yml:

- id: dsh-scout
  config:
    scoutDir: /path/to/scout-cases   # 默认导出目录
    autoPersist: true                # 每次变更后自动写五文件
  • When scoutDir is not configured, the default path is ./dsh-scout/<caseId>/.
  • When autoPersist: true, automatically persist to disk after operations such as scout_start, scout_add_source, and so on.

Data Persistence

Cases are stored in memory by default and isolated by session. scout_export writes a case as five files:
* case.json: Case body and decision status
* sources.json: Source list
* claims.json: Conclusion ledger
* events.jsonl: Replayable event stream
* report.md: Current report snapshot

scout_import can restore a case from an export directory and recalculate the decision.

Applicable Scenarios and Cautions

This plugin is suitable for recruitment background checks and due diligence scenarios that require rigorous evidence chains. When using it, note the following boundaries:

  • The built-in scout_search only registers sources from search results; it does not replace dedicated browsers or MCP Providers for deep scraping.
  • It does not treat funding information, company self-descriptions, or job postings as verified facts.
  • It does not constitute legal, investment, or medical advice.
  • Entity verification depends on E3 official sources (such as gsxt); if they cannot be accessed, the report will faithfully mark the status as “pending verification”.