havingautism/dsh-deepresearch
dsh-deepresearch brings an evidence-first deep research workspace to DeepSeek Harness. It adds a 'Deep Research' panel to the DSH sidebar footer where you create structured research projects: capture the research question, goal, constraints, seed material, and depth; edit sub-questions, dependencies, and explicit success criteria before confirmation; require plan confirmation before evidence can be recorded. The plugin uses two private DSH Agents that never touch your normal chat sessions: a planning Agent that generates a reviewable plan (you can edit it before confirming); and an investigation Agent launched after 'Confirm & start' that uses the profile's Web tools, saves source evidence, updates coverage, and submits a cited report. Features: track question coverage, search/fetch budgets, limitations, and partial completion; attach claims, snippets, URLs, confidence levels, and covered criteria to each sub-question; save conclusions and a full or explicitly incomplete final report; search, filter, sort, resume, abort, or delete projects from the Web library; research view refreshes continuously with search budgets, questions, evidence, coverage, and report progress visible in real time.
Install
dsh plugin --profile web add github:havingautism/dsh-deepresearchGitHub source install (no npm package as of 2026-08-23): dsh plugin --profile web add github:havingautism/dsh-deepresearch, then dsh web. After install, open Deep Research from the sidebar footer and create a project. The plugin creates a research-only DSH Agent to generate the plan while the page refreshes. Review or edit that plan, then choose 'Confirm & start' to launch a new private investigation Agent. The normal chat receives no research prompt, tool calls, or model output.
Compatibility
DeepSeek Harness web profile. Requires the host profile's Web and subagent capabilities. Private research Agents stop with the Host process — projects, evidence, and reports remain durable, but an interrupted run does not resume automatically after restart. Evidence is append-only; correct a mistaken claim by starting a new project or deleting the project before relying on its report. The research view refreshes continuously while running. Private planning and investigation Agents stay scoped to the research project and do not write into normal chat.
Details
- Repo: havingautism/dsh-deepresearch
- Category: Learning & Research
- Stars: 9
- Version: GitHub source install from havingautism/dsh-deepresearch (no npm package; 2026-08-23)
- Last push: 2026-08-26
- First seen: 2026-08-12
Recent updates
The current English README (README.en.md, primary English) documents: two-Agent architecture (planning Agent + investigation Agent), evidence-first workflow (question → plan → confirmation → investigation → report), durable workflow state, tool set (deep_research_get, deep_research_submit_plan, deep_research_add_evidence, deep_research_update_coverage, deep_research_complete), GitHub install, known limitations (no auto-resume after restart, append-only evidence).
FAQ
- Does deep research use my normal chat session?
- No — the planning and investigation Agents run as separate private DSH Agents. The normal chat receives no research prompt, tool calls, or model output. Research activity is fully isolated.
- Can I edit the research plan before the investigation starts?
- Yes — after the planning Agent generates the plan, you review it (and can edit sub-questions, dependencies, and success criteria) before clicking 'Confirm & start'. The investigation Agent only launches after your confirmation.
- What happens if DSH restarts during a research run?
- Projects, evidence, and reports remain durable and are preserved. However, a running investigation Agent stops with the Host process and does not auto-resume — you would need to continue the project manually from where it left off.
Alternatives
omdsh-dev/dsh-deep-research · PerryLink/dsh-checkpoint-rewind · omdsh-dev/dsh-annotation