DSH plugin that turns research into a persisted project with a confirmed plan, per-sub-question Scout / Evaluator agents, an append-only evidence ledger and a URL-cited final report in a web workspace.
DSH integration
Native runtime
Author-claimed
Safety audit
Unaudited
Last verified
2026-11-14
License
Not specified by the author
01What can it help you accomplish?
Run research where the plan has to be confirmed before evidence can be written, instead of letting the chat model wander
A project holding the research question, goals, constraints, seed material and depth; an editable sub-question tree with dependencies and success criteria; evidence refused until the plan is confirmed
People who need an auditable research trail — what was asked, which sources, which criteria were met — rather than a one-shot answer
Get per-sub-question investigation with Scout and Evaluator agents, then a written report
Each sub-question carries claims, excerpts, URLs, confidence and covered criteria; Scouts run web search and fetch, Evaluators review criterion candidates, and a Writer drafts the report with URL-cited sources
Researchers who want parallelism and a review step between searching and writing
Keep track of where a long research project stands and come back to it later
Coverage of the question, search and fetch budget, limitations and partial-completion status, saved conclusions and the final report — browsable in a web library where you can search, filter, sort, resume, abort or delete projects
People running several research threads over days who need state that survives a session
02How to install into DeepSeek Harness
Prerequisites
- A DSH profile you can add plugins to (the README installs into the `web` profile) and the `dsh web` command to start the web UI
- No separate storage setup — the plugin attaches sqlite / HTTP fetch itself if the host does not already have them
Installation steps
- 01
Install the plugin: `dsh plugin --profile web add github:havingautism/dsh-deepresearch`
$ dsh plugin --profile web add github:havingautism/dsh-deepresearch
- 02
Start the web UI: `dsh web`
$ dsh web
- 03
Open 「深度研究」 at the bottom of the left sidebar and create a project
- 04
The plugin builds a private research-only DSH Agent in the background to generate the plan; the page keeps refreshing, and once the plan appears you can edit it and click 「确认并开始」
Verify the integration
- The plan appears in the 「深度研究」 page after creation — the page refreshes every 750 ms and shows the question list, Scout cards and accepted evidence
- If planning fails, the page stays on the 「计划」 step with the persisted error instead of jumping to an empty investigation board
03DSH integration and capability boundaries
A DSH plugin (`@deepseek-ai/dsh-deepresearch`) added to a profile: it patches in a runner, creates private planning / investigation agents, exposes a generated `deepResearch` Remote namespace and mounts a 「深度研究」 web workspace in the sidebar.
Confirmed-plan workflow
research question, goals, constraints, seed material and depth→an editable plan of sub-questions with dependencies and success criteria; evidence writes are refused until you confirm it
creating a project spawns a private planning Agent that only submits the planconfirming starts a new private investigation Agent that writes back to the workspaceEvidence ledger with criteria coverage
claims, excerpts, URLs and confidence submitted by Scouts→per-sub-question evidence plus coverage of the question, search / fetch budget, limitations and partial-completion status
evidence is append-only and cannot be editedPrivate per-role agents with scoped tools
a confirmed plan→a Planner with only `deep_research_submit_plan`, Scouts with `research_web_search` / `research_web_fetch` / `read_artifact` / `submit_criterion_candidates`, Evaluators with `read_artifact` / `submit_criterion_review`, and a Writer with only `deep_research_complete`
ordinary chat does not receive the research prompt, the research tools or a general `web_fetch`the fixed input cost is borne only by the private research agents' requestsSQLite-backed project storage
project, question, criterion, evidence and report records→state persisted in the same SQLite as scratch notes, in a dedicated `deepresearch` domain (default `~/.dsh/storages/dsh.sqlite`)
on startup it migrates the old `deepresearch.sqlite`, and only if that is empty imports older JSON filesit does not re-insert `storage-sqlite` in YAML when the host already has it
04Who is it for? When not to use it?
Good for
- People who need an auditable research trail — what was asked, which sources, which criteria were met — rather than a one-shot answer
- Researchers who want parallelism and a review step between searching and writing
- People running several research threads over days who need state that survives a session
Not for
- Running private agents die with the host process: projects, evidence and reports persist, but an interrupted investigation is not resumed automatically after a restart.
- Evidence is append-only: a wrong claim cannot be edited in place — you re-investigate in a new project, or delete the original project before relying on its report.
- Installing it means accepting SSRF-class risk from web fetching, because the plugin mounts a fetch provider at startup if the host lacks one — the README says to use it in a trusted environment.
05Compatibility, maintenance and safety notes
- Running private agents die with the host process: projects, evidence and reports persist, but an interrupted investigation is not resumed automatically after a restart.
- Evidence is append-only: a wrong claim cannot be edited in place — you re-investigate in a new project, or delete the original project before relying on its report.
- Installing it means accepting SSRF-class risk from web fetching, because the plugin mounts a fetch provider at startup if the host lacks one — the README says to use it in a trusted environment.
- The research workspace is a web UI: the README installs into the `web` profile and starts it with `dsh web`, and the private agents need the profile's installed Web and subagent capabilities plus sqlite and an HTTP fetch provider.
No license declared in the repository metadata · README evidence only, no GitHub release, no security audit
06Frequently asked questions
How do I install Deep Research into DeepSeek Harness?
Run `dsh plugin --profile web add github:havingautism/dsh-deepresearch` and then `dsh web`. Open 「深度研究」 at the bottom of the left sidebar and create a project. No extra YAML is needed — the plugin explicitly enables the runner in its patch and attaches sqlite / HTTP fetch itself if the host lacks them.
Why can't I add evidence right after creating a project?
By design. A private planning agent submits only the plan, and evidence writes are refused until the plan is confirmed. The page refreshes every 750 ms while the plan is generated; once it appears you can edit sub-questions, dependencies and success criteria, then click 「确认并开始」 to start the private investigation agent.
Does ordinary chat get the research tools or web fetching?
No. Only the private agents created by the plugin receive the stage-specific guidance and tools: the planner only gets `deep_research_submit_plan`, Scouts get `research_web_search` / `research_web_fetch` / `read_artifact` / `submit_criterion_candidates`, Evaluators get `read_artifact` / `submit_criterion_review`, and the Writer only gets `deep_research_complete`. Chat's `web_fetch` stays off, and ordinary chat requests do not carry the fixed input cost.
Where is my research data stored?
In the same SQLite as scratch notes — by default `~/.dsh/storages/dsh.sqlite` — under a dedicated `deepresearch` domain. On startup the plugin migrates the old `deepresearch.sqlite` first, and only imports the older JSON files if that is also empty. Do not add another `storage-sqlite` insert to your YAML.
What happens to a running investigation if DSH restarts?
The private agents stop with the host process. Projects, evidence and reports stay persisted, but interrupted work is not resumed automatically after a restart. Note also that evidence is append-only — to correct a wrong claim you re-investigate in a new project or delete the original project before depending on its report.
07Related DSH workflows
archify
by tt-a1i
Turn any idea, plan, or codebase into a beautiful interactive diagram. An agent skill for Claude Code, Codex, and more.
openviking
by volcengine
Self-evolving Context Database for AI Agents. Unify Agent Memory, Knowledge RAG and Skills.
aitoearn
by yikart
Let's use AI to Earn!
learn-harness-engineering
by walkinglabs
Harness engineering beginner tutorial, from 0 to 1
08Data and sources
`@deepseek-ai/dsh-deepresearch` 把证据优先的 Codemini 研究工作区带到 DSH。它提供持久工作流状态、模型工具、生成的 `deepResearch` Remote namespace 和“深度研究”W…
插件 patch 显式启用 runner，并设置项目、问题、标准、证据和报告上限。单独安装即可使用，不必先装随手记或手写 yml。
This page is generated from the project’s public documentation, repository metadata and a structured parse of DSH Plugins; last verified on 2026-11-14. Found an error? Submit a correction.
Best DeepSeek Harness Plugins
Twelve plugins worth installing first — picked from the whole catalog, across every category.