dsh-deepresearch

English | 中文

@deepseek-ai/dsh-deepresearch brings the evidence-first Codemini research workspace to DSH. It provides durable workflow state, model tools, the generated deepResearch Remote namespace, and the 深度研究 Web workspace while composing the host's existing Web and subagent capabilities.

✨ Features

• 🧭 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.
• 🔎 Attach claims, snippets, URLs, confidence, and covered criteria to each sub-question.
• 📊 Track question coverage, search/fetch budgets, limitations, and partial completion.
• 📝 Save conclusions and a full or explicitly incomplete final report.
• 🗂️ Search, filter, sort, resume, abort, or delete projects from the Web library.
• 🤖 Planning submits a reviewable plan only. Confirming starts Scout / Evaluator per criterion (max 3 questions in parallel), then a Writer that cites writing-pack URLs. Normal chat does not inherit research tools or generic fetch. Projects persist in the sqlite file shared with Notebooks (default ~/.dsh/storages/dsh.sqlite, domain deepresearch). A leftover deepresearch.sqlite is imported first; JSON is imported only if that table is still empty.
• 🔄 The research view polls progress every 750ms and draws the question list plus Scout cards (fuse, recent tools, verification, handoff).
• ✨ Match Codemini's research modal, loading motion, and context-specific button shapes.

🚀 Quick Start

Install this bundle by itself:

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

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. It uses the profile's Web and subagent capabilities and writes searches, evidence, coverage, and the final report back to the research workspace. The normal chat receives no research prompt, tool calls, or model output. The patch enables the runner and sets explicit limits for projects, questions, criteria, evidence, and reports. A deepresearch-only install is enough — you do not need Notebooks or a hand-written yml.

Private research sessions record the host launch directory as cwd so DSH can assemble the persona and runtime context. A planning failure stays on the Plan step and displays the persisted error instead of opening an empty investigation board.

At boot, if the host has no sqlite / HTTP fetch yet, this plugin mounts them; if they already exist, it reuses them. Research data lives in the deepresearch domain on that sqlite backend. Do not YAML-insert storage-sqlite again.

Web fetch and security

DSH does not mount a fetch provider or expose chat web_fetch by default (see dsh-base cordis.patch.yml). At boot this plugin mounts @deepseek-ai/dsh-web-fetch-http when it is missing, for private Scout research_web_fetch and host ctx.web.fetch; an already-mounted provider is left alone. Installing alongside @deepseek-ai/dsh-notebooks therefore does not duplicate loader ids. Chat web_fetch stays off.

Installing it opts into web retrieval and associated SSRF-style risk. Use in a trusted environment. See the default Web search decision note.

Model Experience

System prompt

What the model sees

Only the private planning and investigation Agents receive their phase-specific research guidance.

The planning Agent only has deep_research_submit_plan. The orchestrator then spawns Scout (research_web_search / research_web_fetch / read_artifact / submit_criterion_candidates) and Evaluator (read_artifact / submit_criterion_review) roles. The Writer only has deep_research_complete and must cite URLs from the writing pack.

Token effect

The fixed input cost applies only to private research Agent requests, not ordinary chat requests.

KV Cache effect

The section remains prefix-stable while its text and registration scope are unchanged.

Native tools

What the model sees

Each private Agent sees only its role tools. The orchestrator writes durable state. Remote clients create, edit, confirm, stop, inspect, and delete projects.

Token effect

Fixed schema cost applies while the tools are visible. Tool results are concise project summaries; durable evidence and reports remain bounded by configuration.

KV Cache effect

Tool definitions are stable while unchanged. New evidence arrives through later calls and results rather than rewriting earlier request content.

Known Limitations and Deferred Work

• Private 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.