dsh-weknora-ops

The write path and governance for WeKnora in your DeepSeek Harness agent — document ingest, gated deletes, tag trees — under a three-layer safety model.

A DeepSeek Harness plugin (skill + weknora-ops CLI) for WeKnora, the open-source knowledge-base platform.

Relation to the official plugin: WeKnora ships its own read-only DSH plugin, @wxg-prc-cpg/dsh-weknora — four query tools (list/search/read/ask), native cordis tools, contract-tested. For querying, use theirs. This project covers what the official one deliberately does not: the write path — ingest, deletion under a double gate, tag trees — plus read verbs for environments where a CLI form factor is preferred. Both can be installed side by side.

Independent community project. Not affiliated with DeepSeek or Tencent. Bundle format follows the MIT-licensed dsh-superagent-prototype; WeKnora API surface per its public docs. 中文文档

---

Why

Coding agents confidently hallucinate exactly the things your team wrote down somewhere: product rules, deploy procedures, compliance constraints, research conclusions. WeKnora holds that ground truth — this plugin puts it one command away and enforces a simple discipline: cite, don't recall.

Capabilities

	Command	What it does
Query	weknora-ops ask <kb> "<q>"	Knowledge-grounded answer with references
	weknora-ops search <kb> "<q>" --top-k 5	Hybrid retrieval (vector + keyword) with cited chunks
	weknora-ops kb list / kb show	Knowledge bases
	weknora-ops doc list <kb>	Documents + parse status
	weknora-ops tagtree <kb>	Hierarchical tag tree (filter hint: subtree tag ids)
Ingest	weknora-ops doc upload / url / manual	Add documents, web pages, manual knowledge
Destructive	weknora-ops doc delete … --yes	Delete a document (double-gated)

Safety model — three layers

1. Skill etiquette (skills/dsh-weknora-ops/SKILL.md): reads are agent-free; ingests only when a person asked; deletes need explicit in-conversation confirmation. Nothing in the skill grants permission — the surrounding conversation does.
2. CLI gate: doc delete refuses without --yes.
3. Enforcement (the real wall): a workspace-scoped API key with capabilities retrieve / chat / ingest. A key without manage_kbs simply cannot do damage, whatever the agent tries — 403 is the answer, not an obstacle to work around.

This mirrors the admission-control philosophy of dsh-super-agent: trust is structural, not persuaded.

Install

npm i -g dsh-weknora-ops        # exposes the weknora-ops CLI
# or from source: git clone https://github.com/vicfei/dsh-weknora-ops && cd dsh-weknora-ops && npm install && npm link

# credentials — config file is the primary channel inside DSH (its agent shell
# scrubs *KEY* env vars); env vars still work for CI/ops
mkdir -p ~/.config/weknora-ops && chmod 700 ~/.config/weknora-ops
cat > ~/.config/weknora-ops/config <<'EOF'   # workspace API key: WeKnora 设置 → API Keys, capabilities retrieve/chat/ingest
export WEKNORA_URL=http://127.0.0.1:8080
export WEKNORA_API_KEY=wk-…
EOF
chmod 600 ~/.config/weknora-ops/config

# smoke — health + auth probe
weknora-ops status

Into a DSH profile: apply this repo's cordis.patch.yml (see examples/profile.cordis.patch.yml), or drop skills/dsh-weknora-ops/ into your profile's skills directory.

Status

v0.1.1 — vertical slice: 7 read verbs, 3 ingest verbs, delete gate, unit tests. Live-tested end to end: DSH 0.2.1-rc.5 (headless, DeepSeek provider) driving weknora-ops against a real WeKnora v3.1.3 server — skill discovery, status, kb list (542 KBs), search with cited hits; the run also caught and fixed two real integration bugs (env scrubbing → config file; query_text/match_count on v0.8.0-line servers). ask follows the v5-line session flow — on older servers the skill falls back to search + synthesis. Streaming answers, agent-chat operations and an MCP-endpoint mode are on the roadmap.

Acknowledgments

• Tencent/WeKnora (MIT) — the platform this plugin talks to.
• deepseek-ai/dsh-superagent-prototype (MIT) — the observed bundle format this repo follows.

License

MIT