dsh-insurance-advisor

An agent preset for individual insurance agents (个险渠道) on DeepSeek Harness: needs analysis, proposals, suitability matching, sales-recording (双录) checks, and policy servicing.

It ships 10 deterministic tools and 5 skills, and mounts no shell tool — file read/write is kept, but the model cannot run arbitrary commands against client data.

English | 中文

Install

dsh plugin --profile <profile> add github:MagicTangOps/dsh-insurance-advisor

Then restart, open Settings → Agent presets, and select 保险顾问（个险展业）.

The package declares dsh.bundle, so dsh plugin add reconciles it as a profile layer. It also installs from npm as dsh-insurance-advisor once published there.

What it adds

Agent preset

One preset (insurance-advisor, order 20) composing the standard tool set — file, web, todo, schedule, present, compaction, delegation — plus the three domain plugins below. tool-bash, tool-pwsh and the terminal group are deliberately absent.

Suitability and authorization — insurance_check_suitability, insurance_authorized_products

Encodes the product-class / agent-level / customer-risk matrix as deterministic rules instead of leaving the model to infer:

Agent level	Highest product class it may sell
none	P1
4	P2
3	P3
2	P4
1	P5

• A combination sale must take the highest class among its products (从高不从低).
• P4 / P5 additionally require the customer's risk level (R1–R5) to be at least the product's risk rating.
• insurance_authorized_products answers the inverse question: which product classes this agent may sell.

Client and policy ledger — insurance_crm_*

Six tools over a storageDomain-backed table: insurance_upsert_client, insurance_list_clients, insurance_add_policy, insurance_list_policies, insurance_service_due, insurance_record_touchpoint.

Grace periods are computed from the receipt date (回执签收日), inclusive of the day itself:

Channel	Grace period
individual	10 days
bancassurance	15 days
internet	15 days
telemarketing	10 days

insurance_service_due reports policies still inside the grace period (new-business callback owed) and renewals falling due inside a configurable window.

Sales-recording checklist — insurance_dual_record_*

insurance_dual_record_checklist builds a customized five-section recording checklist; insurance_dual_record_validate rejects a recording that omits any of identity, materials, product, disclosure, no-extra-benefit.

Skills

insurance-needs-analysis, insurance-proposal, insurance-dual-record, insurance-compliance, insurance-servicing — loaded from skills/ inside this package.

Host-level and preset-level placement

insurance-crm is inserted at the Host level, not inside the preset, because it consumes the Host storageDomain service. Inside a preset scope it stays pending forever and the preset is reported as broken. The other two plugins live inside the preset, so they activate only when the user selects 保险顾问.

Two consequences worth knowing: the six insurance_crm_* tools are available to every agent in the profile, and the CRM ledger is shared per profile rather than per preset.

Compatibility

• engines.dsh: >=0.2.1-rc.1 <0.3.0-0
• @deepseek-ai/dsh-tools and @deepseek-ai/dsh-storage-domain are peer dependencies with an explicit prerelease branch, so prerelease harness builds are not silently excluded by node-semver.
• zod ^4.4.3 is a regular dependency.

Deliberately not included

• No regulatory full-text. The rules cover the suitability matrix, grace periods and the five mandatory recording sections. Anything else — product filings, carrier-specific clauses — must be checked against the current official text; the preset instructs the model to search for it rather than assert it.
• No commission, illustration or point-of-sale calculation. This is not a quoting engine.
• No client data leaves the machine. The ledger is local storage; only web search, when the model chooses to use it, reaches the network.

Testing

npm test

73 assertions across the four rule modules (suitability 25, dates 13, domain 12, dual-record 23). They are pure-function tests and need no harness runtime.

License

MIT.