@lovedolove/veyra
Veyra——DeepSeek Harness 的工程大脑。持久化、具备证据意识且项目隔离的工程记忆，可在 DSH 会话之间进行观察、记忆和回忆。
插件会安装到这里；不确定时保持 web。
npx -y @deepseek-ai/dsh plugin --profile web add @lovedolove/veyra@0.1.16Veyra 以 @lovedolove/veyra 发布，当前版本为 0.1.16。Plugin Hub 会校验它的 manifest，并保存精确安装来源，便于复现安装结果。
Persistent engineering memory & Knowledge Observatory for DeepSeek Harness.
Install it, restart DSH, keep working. Veyra observes useful context, remembers durable lessons, indexes RAG documentation, and recalls them via unified Hybrid Search. It does not write into your repo.
dsh plugin --profile web add @lovedolove/veyra
Restart dsh web. A good boot looks like this:
[veyra] plugin loaded (home=...)
[veyra] registered /veyra
[veyra] registered skill: veyra
[veyra] registered tools: veyra_remember, veyra_recall, veyra_inspect, veyra_forget, veyra_promote
Requires Node 22.5+ and DSH >=0.1.2-rc.1. Memory lives in $DSH_HOME/veyra/.
You do not manage memory for ordinary work.
symptom → rootCause → remedy → verifiedOutcome). Durable, grounded lessons become derived memory and are linked to neighbors — extended, updated, or contradicted. Near-duplicates are not cloned; instead, repeated observations and verified test outcomes strengthen existing knowledge (unverified → reviewed → verified) and accumulate evidence anchors. Nothing is merged silently.Ask the agent to remember something important, or just keep going — automatic recall is already on.
Veyra unifies Engineering Memory (kind: 'memory') and Knowledge Base / RAG documentation (kind: 'knowledge'):
relevance, semantic, evidence_strength, validation_tier, scope_proximity, freshness_tier, confidence, intent_affinity, relationship).| Command | What it does |
|---|---|
/veyra | Status and memory counts for this workspace |
/veyra observatory | Open the Human Knowledge Observatory dashboard |
/veyra observatory search <query> | Inspect hybrid search signals and score breakdowns |
/veyra observatory record <id> | Deep inspection answering the 6 questions (What, Why, Where, When, Evidence, Validation) |
/veyra observatory causality | Causal knowledge map (symptom → rootCause → remedy → verifiedOutcome) |
/veyra observatory relationships [id] | Directed graph projection of knowledge links |
/veyra observatory local <id> | 1-hop Local Graph around a record (default graph view) |
/veyra observatory graph [id] | Local Graph when an id is given; otherwise the project edge list |
| Network Graph WebUI | Host page at /veyra?id=<record> (Local Graph, 1 hop default). Search → record → graph |
/veyra observatory contradictions | Active conflicting claims shown side-by-side |
/veyra recall <query> | Hybrid search project (+ reusable) memory |
/veyra recent | Last 8 records, including candidates |
/veyra inspect <id> | Read one record with deep provenance and causal facets |
/veyra forget <id> | Soft-forget (leaves recall, stays inspectable) |
/veyra promote <id> canonical | Mark as project truth — only this is canonical |
Tools with the same names (veyra_remember, veyra_recall, veyra_inspect, veyra_forget, veyra_promote) are available to the agent. Load the bundled veyra skill when you need full guidance.
| What | Authority | Auto-recalled? |
|---|---|---|
| Auto-captured observation | candidate | no |
| Remembered / learned lesson | derived | yes |
| Explicitly promoted truth | canonical | yes |
Canonical is never assigned automatically. Repo files stay authoritative. Similarity is not identity: overlapping memories are linked, not merged. Secrets are redacted. Projects are isolated; reusable memory is opt-in.
DSH Settings exposes Recall limit and Include reusable. recallLimit: 0 disables automatic recalled-memory context only — tools, /veyra, Observatory, and stored memory stay available.
In the profile cordis.patch.yml:
- id: veyra
name: '@lovedolove/veyra'
config:
recallLimit: 5 # 0 = no automatic recall context
includeReusable: true
observe: true
learn: true
VEYRA_HOME overrides the storage root. observe / learn remain yaml-only.
MIT