Foreword¶

Conducting “deep research” in DeepSeek Harness (DSH) typically involves chaining searches, reading, and summarization with fixed prompts—steps hard-coded into the conversation. This approach often leads to missing dimensions, over-searching, or premature conclusions when topics become complex. The DSH community plugin omdsh-dev/dsh-deep-research transforms the deep-research process into an independent plugin (distinct from the skill system), integrated with DSH’s official workflow engine. It leverages cybernetics and information theory to constrain planning, retrieval, and synthesis, forming an adaptive and extensible research loop.

What is This¶

dsh-deep-research is maintained by omdsh-dev, with the npm package name @dsh-external/dsh-deep-research. The current version is v0.1.0, classified as a Workflow in SkillHub. Synopsis: Adaptive deep-research orchestrator plugin for DeepSeek Harness (official workflow engine, cybernetics/information-theory design).

The plugin registers the deep_research tool, orchestrates workflows via ctx.workflows / @deepseek-ai/dsh-workflow-workerthread, and uses DSH’s built-in web_search / web_fetch for retrieval. It does not register skills in ctx.skills nor inject TUI prompts; triggering relies on tool descriptions (e.g., deep research, investigation, multi-source information synthesis), allowing the model to invoke it as needed during conversation.

Core Mechanisms¶

Below, aligning with the “Theory → Mechanism” structure in the README, we explain how the plugin turns the research process into a closed loop rather than a fixed script.

Planning: Define the Answer Space First, Then Decompose Sub-questions¶

The planning agent first defines the decision scope (scope) supported by the research and acceptance criteria (acceptance) for each sub-question. It enumerates the topic’s information dimensions and performs a self-check on output coverage (coverage_gaps). This corresponds to reference signal calibration in cybernetics and Ashby’s Law of Requisite Variety—sub-questions must cover the topic’s dimensions; otherwise, blind spots are inevitable.

Research: Three-State Evidence and Marginal Gain Stopping¶

Each research sub-agent maintains three-state evidence: confirmed, uncertain, and gaps. After updating via web_search / web_fetch, it performs marginal gain verification. A consecutive round with zero gain triggers a stop, with a hard round limit as well. Simple topics may converge in one round, while complex topics collect high-priority gaps after each round and automatically dispatch supplementary research. Blind spots declared in the planning phase are also targeted for reconnaissance and verification.

Synthesis and Optional Review¶

The synthesis sub-agent compresses evidence into a final report with lossy compression, retaining information that discriminates between conclusions. If review: true is enabled, an adversarial review sub-agent conducts citation spot-checks, coverage audits, and flags contradictions or overconfidence.

Tool Parameters¶

After installation, the model can control research granularity by passing parameters to the deep_research tool:

Optional configurations include maxParallel (default 4), maxTotalAgents, and model tiering for plannerModel / researcherModel / synthesizerModel / reviewerModel—the README suggests using stronger models for planning and synthesis, and cheaper models for research execution to control costs.

Installation and Enabling¶

The plugin is installed into a profile via dsh.bundle.patch (cordis.patch.yml). The SkillHub registry source is omdsh-dev/dsh-deep-research. During installation, replace <profile> with tui, headless, web, or a custom profile.

dsh plugin --profile <profile> add git+https://github.com/omdsh-dev/dsh-deep-research.git
dsh --profile <profile>

After restarting DSH with the second command, the deep_research tool is injected with the profile. If pnpm rewrites https to ssh due to local git insteadof configuration, maintain the git+https:// format. If dsh plugin prompts for allowBuilds, add the corresponding configuration to $DSH_HOME/profiles/<name>/pnpm-workspace.yaml as indicated.

Dependencies: The profile must include the official workflow engine and built-in web tools; the dsh official base combination includes them by default. Install into a profile that provides the workflows provider (e.g., tui/headless); if some web combinations do not declare this provider, the Loader will remain pending. Runtime requires Node ^22.19.0 || >=24.0.0.

Updates and uninstallation:

dsh plugin --profile <profile> update
dsh plugin --profile <profile> remove @dsh-external/dsh-deep-research

The license is MIT. Before installation, it is recommended to read the GitHub source code and LICENSE. The plugin runs with the current DSH process permissions and will invoke network search and fetching capabilities.

Typical Usage¶

The tool is automatically triggered by the model based on its description; simply state your needs in the conversation without remembering commands. Examples from the README:

• “Conduct a deep investigation into the MCP ecosystem, focusing on comparing several mainstream implementations, and produce a cited report.”
• “Research based on this list of questions: 1. … 2. … ” (If an existing list is provided, auto-decomposition is skipped, and research proceeds in parallel.)
• “Investigate options A and B, with the purpose of deciding which one we should choose.” (The more specific the purpose, the more accurate the answer space.)

For greater rigor, pass depth: 3; for citation error correction and coverage audits, pass review: true. Complex topics automatically expand rounds, while simple topics typically converge in one round.

Suitable Scenarios and Notes¶

Suitable for: Scenarios in DSH requiring multi-source retrieval, structured evidence summarization, and cited reports; when an existing clear question list is available, parallel research can proceed directly; when seeking to reuse the official workflow’s concurrency limits, cancellation propagation, and wf-runs records, instead of writing custom orchestration scripts.

Notes:

• This plugin is a plugin, independent of the .claude/skills/deep-research skill template; the two can coexist.
• Failure in a single sub-question research only marks that section; planning failure causes the tool to error, and the main agent can retry with adjusted parameters.
• When canceling a DSH session, exec.signal is passed to the workflow run, and sub-agents are terminated accordingly.
• SkillHub (skillhub.cn) is a DSH community plugin directory, with no official affiliation to DeepSeek or High-Flyer. The DSH ecosystem philosophy is “everything is a plugin,” and this plugin is one of the community’s workflow-type extensions.

Links¶

• SkillHub Directory: https://www.skillhub.cn/plugins/omdsh-dev/dsh-deep-research
• GitHub Repository: https://github.com/omdsh-dev/dsh-deep-research