dsh-prime-agent
Make your DeepSeek Harness agents learn. A closed learning loop connecting
Prime Agent's continual memory system and an
OKF/OpenWiki knowledge registry to DeepSeek Harness
— implemented entirely as native plugins on documented seams. No forks, no patches to core.
┌─────────────── dsh session ───────────────┐ ┌────────── Prime Agent ──────────┐
│ skill catalog ◄── okf-knowledge ───────────┼──────│ skills + OKF bundle pages │
│ durable context ◄── prime-memory ────────┼──────│ memories / prompt notes │
│ turn transcript ── prime-memory ─────────┼──────► inbox (refine.run) │
└───────────────────────────────────────────┘ │ ▲ │
│ │ DSPy / MIPROv2 │
└────────┴── OKF draft pages ─────┘
(accepted candidates flow back
through the catalog above)Packages
| Package | What it does | Works standalone |
|---|---|---|
@morewax/dsh-prime-memory |
Serves Prime skills into ctx.skills, injects Prime memories as durable context (session-reference contract: escaped, budgeted, fail-loud), exports turns to Prime's inbox for refine |
✅ |
@morewax/dsh-okf-knowledge |
Serves OKF/OpenWiki bundle pages as skills with provenance — trust tiers (verified/unverified) surface in metadata, with a verifiedOnly trust gate |
✅ |
prime-agent-testkit |
Reusable fixtures + suites verifying every contract above (16 tests) | dev only |
New here? Start with the 5-minute quick start, then browse the cookbook.
Install
dsh plugin --profile default add @morewax/dsh-prime-memory
dsh plugin --profile default add @morewax/dsh-okf-knowledge
# or straight from GitHub:
dsh plugin --profile default add github:moreWax/dsh-prime-agentQuick start
# ~/.dsh/profiles/default/cordis.patch.yml
- id: prime-memory
name: '@morewax/dsh-prime-memory'
config:
primeHome: ~/.prime/agent # any Prime Agent install
sessionId: <prime-session-uuid> # whose memories to surface
exportTurns: true # write turns back for refine
- id: okf-knowledge
name: '@morewax/dsh-okf-knowledge'
config:
bundles: [/srv/wiki/core] # one or more OKF bundles
verifiedOnly: falseThe offline optimizer ships too — see refine/.
See the cookbook for complete recipes:
- Serve your knowledge base as agent skills
- Inject durable memories safely
- Close the learning loop with DSPy
- Trust-gated serving for team wikis
Why this is different
Memory-injection and transcript-export plugins already exist. This project is not
either of those in isolation — it is the loop between them, plus:
- Provenance-first knowledge: pages carry machine-readable trust tiers; the
model catalog distinguishes verified from unverified content. - Injection safety by contract: untrusted memory text cannot forge framing
tags (lossless escaping), budgets fail with explicit codes. - Crash-safe export: append-only records via
ctx.storageDomain, never
corrupt state, never break the agent loop. - Real interop: drives an actual Prime Agent install — not a reimplementation.
Compatibility
Built against DeepSeek Harness 0.1.1-rc.x (developer preview). The harness makes
no stability promises pre-release; we re-verify against every upstream bump and pin
tested versions here.
Development
pnpm install
pnpm test # 16 tests: store readers, provider wiring, export, injection safety
pnpm buildFramework packages come from npm — no harness checkout needed. See
docs/dev-setup.md for testing inside a live harness.