dsh-deep-research: An Adaptive Deep-Research Orchestration Plugin for DeepSeek Harness

dsh-deep-research is a deep-research orchestration plugin (DSH plugin) for DeepSeek Harness, built on the official workflow engine (ctx.workflows / @deepseek-ai/dsh-workflow-workerthread) with a cybernetics and information-theory design. It is not a fixed-prompt pipeline but a living, adaptive research loop. It addresses the problems of blind searching, non-convergence, and hiding uncertainty in traditional research tools. Core capabilities include: a planning agent that defines the answer space and acceptance criteria, research sub-agents that maintain three-state evidence (confirmed/uncertain/gaps), marginal gain validation (stop after one round of zero gain), closed-loop re-planning (auto-dispatch supplementary research), and adversarial review (citation spot-check and coverage audit).

What core features does the dsh-deep-research DSH plugin support?

• Adaptive research loop: Round 1 researches all sub-questions in parallel; after each round, high-priority gaps are collected and automatically dispatched for the next round. Simple topics converge in one round; complex topics expand automatically until marginal gain ≈ 0.
• Information-theory driven: Research sub-agents predict → act (web_search/web_fetch) → update evidence → validate marginal gain; stop after one consecutive round of zero gain, with a hard round limit.
• Coverage self-check: The planning agent enumerates information dimensions, maps each sub-question to a dimension, and outputs coverage_gaps; blind spots are verified by targeted reconnaissance.
• Adversarial review (optional): Citation spot-check (URL reachability/support), coverage audit, contradiction and overconfidence labeling, to counter hallucinations.
• Model tiering: Configurable plannerModel/researcherModel/synthesizerModel/reviewerModel; use strong models for planning/synthesis and cheap models for research, significantly reducing cost.

How to install dsh-deep-research?

Use the following command (replace <profile> with tui / headless / web or your own profile):

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

If pnpm rewrites the https URL to git+ssh (due to a global git insteadof config), use the git+https form above; if dsh plugin prompts for allowBuilds, add a line to $DSH_HOME/profiles/<name>/pnpm-workspace.yaml as instructed.

How to use dsh-deep-research?

The tool is automatically triggered by the model based on its description (deep research/investigation/multi-source information synthesis/research report/literature collection). Just speak naturally in conversation, for example:

• "Deep-dive into the current MCP ecosystem, compare several mainstream implementations, and produce a report with citations."
• "Research according to this question list: 1. ... 2. ..." (if a list exists, skip automatic decomposition and research in parallel)
• "Research A/B options, purpose is to decide which one we choose" (the clearer the purpose, the more accurate the answer space)

Complex topics automatically expand rounds (adaptive loop); simple topics converge in one round. For more rigor, pass depth: 3; for citation correction and coverage audit, pass review: true.

What DeepSeek Harness components does dsh-deep-research depend on?

The profile combination must include the official workflow engine and built-in web tools—the official dsh base combination includes them, no extra installation needed; peer dependencies (like @deepseek-ai/dsh-tools) are provided by the combination, and setting autoInstallPeers: false in the profile avoids looking up unpublished @deepseek-ai/* packages from the registry.

How to update and uninstall dsh-deep-research?

dsh plugin --profile <profile> update
dsh plugin --profile <profile> remove @dsh-external/dsh-deep-research
# Or: remove the dependency from the profile's package.json and then run dsh plugin --profile <profile> update

This plugin is maintained by omdsh-dev, licensed under MIT, last updated in 2026-08. The source is the runtime (the package entry points directly to src/index.ts, no build step), and it requires erasable-only TS syntax.

See more DSH plugins at dsh-plugin.org