MIT DSH knowledge-base plugin: archive sessions locally, retrieve with FTS5 and progressive reading, optionally auto-recall, and browse a knowledge graph — with auto-injection still marked experimental.
DSH integration
Native runtime
Author-claimed
Safety audit
Unaudited
Last verified
2026-11-12
License
MIT
01What can it help you accomplish?
Turn a finished DSH conversation into local knowledge notes, method cards and pitfall cards
Markdown notes plus method / pitfall / project cards with recorded sources, duplicates skipped, and an incremental record for new messages instead of overwriting the old note
DSH users who want experience from past sessions available to later ones
Let the current DSH conversation query earlier knowledge without switching models
An answer from the current model backed by `knowledge_search`, `knowledge_read`, `knowledge_timeline` and `tool_catalog_search`, read progressively from summary to body
DSH users who want retrieval inside the running session rather than a separate chat window
Study a bound folder or import existing notes into the knowledge base
Imported Markdown, UTF-8 text / source, folders, ZIP and native content packs, indexed for retrieval and the knowledge graph, with source paths kept
Users bringing Obsidian-style notes or project documentation into DSH
02How to install into DeepSeek Harness
Prerequisites
- A DeepSeek Harness install with a local plugin-package install entry — the README states Windows + Harness Desktop `0.2.1-alpha.1` as the mainly verified environment
- A backup of the knowledge directory before upgrading, after host sessions and background tasks have finished
Installation steps
- 01
Download `dsh-wanxiang-knowledge-0.2.0-alpha.10.tgz` from GitHub Releases (the repo also keeps the same-version package under `artifacts/`) and pick it in the host's local plugin-package install entry
- 02
For a normal Web profile, install with the host's plugin management command: `dsh plugin --profile web add "<absolute path to the package>"`
$ dsh plugin --profile web add "<absolute path to the package>"
- 03
Source build alternative: `git clone`, then `npm ci`, `npm run build`, `npm run typecheck`, `npm test`, `npm run pack:plugin`
$ git clone
Verify the integration
Not specified by the author
03DSH integration and capability boundaries
Native DeepSeek Harness plugin using the host's session, command, tool and prompt-section extension points
Session archiving into local Markdown
A finished conversation turn→Notes and cards generated on demand, with AI-organised or model-free raw export, and incremental records for new messages
The knowledge directory holds `notes/`, `cards/`, `sources/`, `attachments/`, `tools/` and `.wanxiang/`; a compatibility copy of the top level is kept for older versions and takes extra diskLocal retrieval with FTS5 and progressive reading
A query, optionally the current conversation's question→`exact` / `hybrid` / `deep` retrieval over local `node:sqlite + FTS5`, read as L0 summary → L1 overview → L2 body, with `knowledge_read` paging up to 12,000 characters
Vector retrieval is currently Off; there is no embedding-model dependencyOptional auto-recall into the prompt section
The ready cache from an isolated SQL subprocess→Off / hint-only / auto-inject levels; injection only starts after confirming that knowledge summaries go to the host model
Budgets of 150 ms per slot, 190 ms per search and 200 ms for the cache wrapper; the synchronous prompt section only reads the ready cache and never waits, reads bodies or calls a modelImport and folder study
Markdown, UTF-8 text and declared source/config extensions, folders, ZIP, or a Wanxiang content pack→Indexed entries with relative references and Obsidian `[[wiki links]]` resolved, content-pack manifests verified by SHA-256
One batch is capped at 2,000 files and 64 MiB; a single document or attachment at 2 MiB; encrypted ZIP, split archives and ZIP64 are unsupportedKnowledge graph and optional security-analysis tool catalog
Local knowledge entries and registered tool categories→Sphere or network graph with type legend, title/tag filters, rotation and PNG export, plus a tool library showing environment-detection status per tool
The security-analysis extensions (PE/ELF, disassembly, strings, traffic files, APK/DEX, .NET, YARA) require separate installation and are not bundled with the plugin
04Who is it for? When not to use it?
Good for
- DSH users who want experience from past sessions available to later ones
- DSH users who want retrieval inside the running session rather than a separate chat window
- Users bringing Obsidian-style notes or project documentation into DSH
Not for
- Mainly verified on Windows + Harness Desktop `0.2.1-alpha.1`; other platforms and host versions need further testing. There is no vector retrieval, OCR, PDF/Word body parsing, cloud sync or multi-user collaboration.
- Files, indexes and rule-based retrieval stay local, but AI organisation, folder study, metadata completion and confirmed auto-injection send the relevant text or summaries to the model configured in Harness.
- Auto-injection is still experimental: the README reports a 76.25% miss rate at the development threshold 0.65 on a fixed synthetic strategy set, and real-history evaluation plus a production threshold are unfinished.
- Model-organised conclusions need to be checked against their sources, and the optional analysis extensions are not a malware sandbox; their tools must be installed and configured separately.
05Compatibility, maintenance and safety notes
- Mainly verified on Windows + Harness Desktop `0.2.1-alpha.1`; other platforms and host versions need further testing. There is no vector retrieval, OCR, PDF/Word body parsing, cloud sync or multi-user collaboration.
- Files, indexes and rule-based retrieval stay local, but AI organisation, folder study, metadata completion and confirmed auto-injection send the relevant text or summaries to the model configured in Harness.
- Auto-injection is still experimental: the README reports a 76.25% miss rate at the development threshold 0.65 on a fixed synthetic strategy set, and real-history evaluation plus a production threshold are unfinished.
- Imports are capped at 2,000 files and 64 MiB per batch and 2 MiB per document or attachment; encrypted ZIP, split archives and ZIP64 are unsupported, and PDF/Word/image files are kept as attachments without text extraction or OCR.
- Model-organised conclusions need to be checked against their sources, and the optional analysis extensions are not a malware sandbox; their tools must be installed and configured separately.
MIT source (third-party documents under `knowledge/` keep their own licences); claims here come from the repository README at the recorded commit, with no independent safety audit.
06Frequently asked questions
How do I install it into DeepSeek Harness?
Download `dsh-wanxiang-knowledge-0.2.0-alpha.10.tgz` from GitHub Releases and choose it in the host's local plugin-package install entry. A normal Web profile can also use `dsh plugin --profile web add "<absolute path to the package>"`. The README asks you to let host sessions and background tasks finish and to back up the knowledge directory before upgrading.
Where does the knowledge live, and what gets sent to a model?
Notes, cards, sources, attachments, the tool catalogue and indexes live under the knowledge directory (default: a 万象知识库 folder in Documents, switchable in settings); settings are in `<DSH_HOME>/wanxiang/settings.json`. Files, indexes and rule retrieval stay on the machine, while AI organisation, folder study, metadata completion and confirmed auto-injection send the relevant text or summaries to the model configured in Harness.
Does auto-recall work without confirming anything?
No. Auto-recall has off / hint-only / auto-inject levels, and injection only starts after you confirm that knowledge summaries will be sent to the host model. The README still calls auto-injection experimental and reports a 76.25% miss rate at the development threshold on a fixed synthetic set.
Which retrieval modes and tools does the model get?
`exact`, `hybrid` and `deep` retrieval over local `node:sqlite + FTS5`, read progressively as L0 summary, L1 overview and L2 body. The model tools are `knowledge_search`, `knowledge_read`, `knowledge_timeline` and `tool_catalog_search`; `knowledge_read` pages at most 12,000 characters and returns `nextOffset` until it is `null`. Vector retrieval is currently Off.
What are the import and platform limits?
One batch is capped at 2,000 files and 64 MiB, and a single document or attachment at 2 MiB; encrypted ZIP, split archives and ZIP64 are unsupported. PDF, Word and image files are kept as attachments without body extraction or OCR. The README states Windows + Harness Desktop `0.2.1-alpha.1` as the mainly verified environment, with no vector retrieval, OCR, cloud sync or multi-user collaboration yet.
07Related DSH workflows
weknora
by tencent
Open-source LLM knowledge platform: turn raw documents into a queryable RAG, an autonomous reasoning agent, and a self-maintaining Wiki.
honcho
by plastic-labs
Memory library for building stateful agents
webnovel-writer
by lingfengqaq
A long-form Chinese web-novel writing assistant for Claude Code and DeepSeek Harness, designed to reduce memory loss and hallucinations in serialized projects of up to 2 million characters.
reme
by agentscope-ai
ReMe: Memory Management Kit for Agents - Remember Me, Refine Me.
08Data and sources
万象知识库是为 [DeepSeek Harness](https://github.com/deepseek-ai/deepseek-harness) 开发的社区插件。它在 Harness 内提供会话落库、本地检索、可选自动召回、知识图谱、…
| 原生 Harness 插件 | 接入会话、命令、工具与提示词扩展点，在宿主内完成查询和沉淀 |
This page is generated from the project’s public documentation, repository metadata and a structured parse of DSH Plugins; last verified on 2026-11-12. Found an error? Submit a correction.
Best DeepSeek Harness Plugins
Twelve plugins worth installing first — picked from the whole catalog, across every category.