Preface

When performing agent tasks in DSH, if a model inherits web_search, it can directly return answers. However, when dealing with variable facts, it is more critical to define: which sources are allowed, what evidence is retained, whether JSON selection is deterministic, and where gaps remain unresolved.

dsh-plugin-verified-search addresses these issues: it replaces the inherited web_search with a bounded workflow, making the source scope, retained evidence, deterministic JSON selection, and unresolved gaps visible.

This plugin verifies the post-conditions of the workflow and structured sources; it is designed to be fail-closed or to explicitly remain unresolved. It does not prove the authenticity of the publisher, nor does it guarantee that the upstream search index is up to date.

What is this

f0909172434/dsh-plugin-verified-search is a plugin for DeepSeek Harness, positioned as a Verified current-source search workflow for DeepSeek Harness. It is maintained by f0909172434 and is licensed under MIT.

The stable version v0.1.1 only contains verified_search and is suitable for narrow, variable-fact queries. It performs post-filtering on structured source hostnames and makes citation-excerpt gaps visible.

The experimental snapshot 0.3.0-experiment.0 additionally includes:

  • verified_research
  • verified_json_selection
  • verified_json_numeric_extrema
  • verified_json_projection

Experimental snapshots are intended only for development and evaluation. When conducting reproducible tests, do not replace the moving main branch with a fixed commit.

Core Capabilities

The goal of verified_search is not to expand search capabilities but to constrain a query to a verifiable scope. The following points are explicitly retained within the result boundary:

  • Source Scope: Post-filtering on the returned structured source hostnames;
  • Retained Evidence: Which excerpts are kept as evidence;
  • Deterministic JSON Selection: Avoiding implicit selection based solely on text semantics;
  • Unresolved Gaps: Remaining unresolved when no answer excerpts exist, rather than using old answers or memory to fill in.

Experimental tools target more JSON selection and projection tasks, such as verified_research, verified_json_selection, verified_json_numeric_extrema, and verified_json_projection. Since these are still in experimental snapshots, version boundaries should be confirmed before production use.

Installation and Enablement

First, confirm the prerequisites:

  • DeepSeek Harness 0.1.0-rc.6
  • Cordis 4.0.1
  • Node.js 22.19.x or 24.x
  • DEEPSEEK_API_KEY available via Harness credential service or startup environment
  • An existing search-capable preset

Regarding platforms, CI coverage includes Ubuntu and Windows; macOS is not currently within the support contract.

First install the stable version, then dump the config, and finally start web:

dsh plugin --profile web add github:f0909172434/dsh-plugin-verified-search#v0.1.1
dsh --profile web --dump-config
dsh web

A single-command format:

dsh plugin --profile web add github:f0909172434/dsh-plugin-verified-search#v0.1.1; dsh --profile web --dump-config; dsh web

If your deployment has already mounted the time-context row from Discussion #344, set the environment variable before starting Harness:

export DSH_VERIFIED_SEARCH_DISABLE_TIME_CONTEXT=1

Typical Usage

Ask bounded, absolute-date questions to a search-capable agent. For example:

Find DeepSeek’s current flagship API model as of 2026-08-14. Use only api-docs.deepseek.com.

The corresponding verified_search parameters are:

{
  "query": "DeepSeek current flagship API model as of 2026-08-14",
  "allowed_domains": ["api-docs.deepseek.com"]
}

Here, allowed_domains is a post-condition on the returned structured source hostnames, not a network egress point or privacy boundary.

Use Cases and Considerations

Suitable for:

  • Limiting current-fact queries to specific hostnames;
  • Viewing retained evidence and citation-excerpt gaps;
  • Evaluating extended workflows like JSON selection, numeric extrema, and JSON row projection in an experimental environment.

Not suitable or requires vigilance:

  • Treating plugin results as proof of publisher authenticity;
  • Assuming the upstream search index is definitely up to date;
  • Treating domain filtering as network egress control or privacy boundary;
  • Using the moving main branch in a production reproducible chain.

Regarding verification, external independent verification is missing. The repository provides internal deterministic testing, CI, package-reproducibility checks, and maintainer-run conformance evidence. These are not third-party reviews.

Security-wise, this plugin runs with the permissions of the current dsh process. Before installation, it is recommended to check the source code, dependencies, and the MIT license, and to ensure that the allowed_domains post-filter does not replace the security boundaries you need.

References

  • Community Directory Page: https://www.skillhub.cn/plugins/f0909172434/dsh-plugin-verified-search
  • GitHub: https://github.com/f0909172434/dsh-plugin-verified-search