dsh-deepresearch: A Deep Research Workspace Plugin for DeepSeek Harness

dsh-deepresearch is a Deep Research workspace plugin (DSH plugin) for DeepSeek Harness (DSH), maintained by havingautism, written in JavaScript, and last updated in 2026-08. It brings the evidence-first Codemini research workflow to DSH, addressing the problems of chaotic planning, missing evidence, and untrustworthy reports during research. The plugin provides persistent workflow state, model tools, a generated deepResearch Remote namespace, and a "Deep Research" web workspace, while composing the host's existing web and subagent capabilities. Core capabilities include recording research questions and goals, editing sub-questions and dependencies, associating evidence with confidence, tracking coverage and budgets, and saving conclusions and final reports.

What core features does the dsh-deepresearch DSH plugin support?

The plugin supports creating and managing research projects, recording research questions, goals, constraints, seed materials, and research depth. Before confirmation, users can edit sub-questions, dependencies, and explicit success criteria; the plugin refuses to write evidence until the plan is confirmed. Each sub-question can be associated with arguments, excerpts, URLs, confidence, and covered criteria, and it tracks question coverage, search and crawl budgets, limitations, and partial completion status. Finally, conclusions and final reports, either complete or explicitly marked as incomplete, can be saved. In the web repository, users can search, filter, sort, resume, abort, or delete projects.

How do I install dsh-deepresearch?

Install the plugin and start the web interface, then open the "Deep Research" tab to begin. The installation commands are as follows:

dsh plugin --profile web add github:havingautism/dsh-deepresearch
dsh web

What is the model experience and tool design of dsh-deepresearch?

While the plugin is active, each request receives a research workflow instruction: create or resume a project before investigation, refine and confirm its question plan, then use the composed web and subagent tools, save each source-backed claim, mark coverage honestly, and save the final report only after comparing accepted evidence. The model sees six native tools: deep_research_start, deep_research_list, deep_research_confirm_plan, deep_research_add_evidence, deep_research_update_coverage, and deep_research_complete. Remote clients can also edit draft plans, stop tasks, read full projects, and delete projects. Tool results use concise project summaries, and persistent evidence and reports are subject to configured limits.

What are the known limitations and future work for dsh-deepresearch?

The plugin records and validates orchestration state but does not schedule searches itself; investigation is performed by installed web or subagent capabilities. In this version, search and crawl budgets are plan metadata; automatic deduction requires integration from the corresponding capability provider. Evidence is append-only and not editable; incorrect arguments should be re-investigated in a new project or the original project deleted before relying on the report.

Which DeepSeek Harness versions and platforms are compatible with dsh-deepresearch?

As an official DSH plugin, it depends on official packages such as @deepseek-ai/dsh-api-remotes, @deepseek-ai/dsh-client-runtime, and @deepseek-ai/dsh-tools, ensuring compatibility with the DeepSeek Harness ecosystem. The plugin is installed via the dsh plugin command and works in DSH-supported web environments.

View more DSH plugins at dsh-plugin.org