dsh-cdi-plugin

Synthetic CDI auditing, packaged for DeepSeek Harness
Native model tools for deterministic SOP-rule evaluation over a bundled SQLite rule store — with the CDI agent skills and (fictional) gold sets shipped in the box.

A DeepSeek Harness (DSH) plugin that turns a governed clinical-documentation audit pipeline into two native model tools. Install it, ask for an audit, and the model walks the pipeline end to end — with every deterministic verdict coming from the SQLite SOP store, never from the LLM's judgment.

Every finding stays requires_human_review and routes to a qualified human reviewer. All data is synthetic and fictional — no real patient, provider, or client data; no clinical, coding, billing, or compliance decision.

---

Who is this for

• Documentation-quality & audit teams in healthcare operations who need a governed, traceable way to review clinical documentation against SOP rules — every finding cites its source fields and routes to a qualified human reviewer.
• DSH users building vertical solutions — a working reference for packaging a domain pipeline (skills + rules + data + tools) as a standard plugin on the stock DSH GUI.
• Researchers & students studying governed agentic workflows — everything runs on synthetic data, so it is safe to experiment with end to end.

Use cases

• Full-pipeline audit of a synthetic ICHD record: normalization → evidence review → rule evaluation (deterministic via SQLite, narrative via use-case skills) → 7 draft findings, all requires_human_review.
• Patient-domain checks with the multi-domain gold set (SYN-ICHD-05/06, requires store: "multi-domain").
• Ad-hoc rule lookups — "which treatments finished ≥ 15 min early?" (cdi_query_rule with SYN-ICHD-01, one record at a time).

Features

• Native tools, no shelling out — cdi_query_rule and cdi_list_rules are first-class model tools (cards in the conversation, not bash+python).
• The LLM reports, never judges — deterministic rules evaluate against bundled SQLite SOP stores (shared + multi-domain); the tool's result is final.
• 8 CDI agent skills in the box — the governed pipeline (normalization → evidence review → rule evaluation → human review) ships with the package, and is also discoverable from any repo's .agents/skills.
• Self-contained — tools + SOP stores + gold sets + rules + skills travel inside the npm package.
• Zero UI changes (Step 1) — runs on the stock DSH GUI; Step 2 adds a domain UI (finding cards, evidence detail, review workbench).

How it works

gold set (synthetic) ──► normalize ──► evidence review ──► rule evaluation ──► draft finding ──► HUMAN REVIEW
                                                             │                        │
                                            deterministic: cdi_query_rule  (SQLite)   └─► confirm / reject / clarify
                                            narrative:     use-case skills  (LLM)              │
                                                                                               └─► feedback loop

Quick start

# install into your web profile
dsh plugin --profile web add dsh-cdi-plugin

# restart dsh web, then in any session:

用 cdi_list_rules 列出所有确定性规则,然后对 data/synthetic-ichd-patient-goldset.json 跑一次完整 CDI 审计

The deterministic-rule-audit skill prefers cdi_query_rule for every deterministic check; tools/query_deterministic_rule.py remains as a fallback with the same contract.

Tools

Tool	Purpose
cdi_query_rule	Evaluate one deterministic SOP rule (SYN-ICHD-01/09/06) against one record. store: "multi-domain" selects the patient-domain store.
cdi_list_rules	List every deterministic rule in both bundled stores with trigger descriptions.

Development

node --test "test/*.test.mjs"     # frozen smoke tests — no DSH host needed

The tests cover the failure modes that actually crash profiles at boot (schema compilation at apply() time) and at tool-call time (output-schema validation, including non-triggered results whose draft_question is null). Run them before every release.

Local dev needs one machine-specific wiring symlink so the plugin resolves @deepseek-ai/* from its real location:

mkdir -p node_modules
ln -sfn ~/.dsh/profiles/node_modules/@deepseek-ai node_modules/@deepseek-ai

(node_modules/ is git-ignored.)

Publish

node --test "test/*.test.mjs"      # gate
npm version patch                  # bump
npm publish                        # requires a bypass-2FA granular token or OTP

peerDependencies deliberately pin the working release train (@deepseek-ai/dsh-tools: ^0.1.0-rc.6) — the npm latest dist-tag of @deepseek-ai/dsh-tools points at a broken 0.0.1-rc.1, so never resolve from latest.

Roadmap

• Step 1 — host plugin (this package): native rule tools, packaged skills/data, verified on DSH 0.1.0-rc.6
• npm publish — [email protected] live on the npm registry
• Step 2 — cdi-gui client half: audit-finding cards in the conversation, evidence/rule detail panel, human-review workbench — enterprise domain UI inside the DSH shell
• More verticals — other documentation-audit domains on the same governed-pipeline pattern

License

MIT