cqai-dsh-plugin-xiaobao

e销宝 (e-Xiaobao)

A realtime sales-coaching copilot plugin for 易宝工坊 (DeepSeek Harness), aimed at small teams serving consumers.

While a salesperson is on a call, they open the e销宝 panel in the right sidebar. Each time the customer finishes a sentence, the panel shows one primary reply, one alternative, one suggested follow-up question, risk warnings, and the sources behind them. The salesperson keeps full control of what gets said — e销宝 never speaks for them.

中文文档是主文档：README.zh.md · 安装说明 · 验收对照

What it does

• Realtime coaching panel (right sidebar tab): live transcript with provisional/final distinction, customer/sales speaker tags, intent + confidence, primary reply, alternative, follow-up, risk warnings, sources, and accept / edit / skip / human-handoff actions.
• Knowledge management page (this plugin's own page in the Plugins manager): phrase cards, review queue, product facts and policies, version health, a read-only preview sandbox, and connection/privacy details.
• Sales strategy conversation (left sidebar and coaching tab): a full-width conversation with one fixed composer. Each response includes customer assessment, communication goal, strategy, reply, follow-up, and cautions. Dictation fills the editable composer; optional continuous listening submits recognized segments. Feedback stays attached to its response. Switching tabs retains the conversation and pauses capture; failed submissions keep their text for retry.
• Five agent tools: search_phrase_cards, get_product_policy, get_customer_context, save_coaching_feedback, request_human_handoff.
• Seven runtime skills: b2c-sales-context, b2c-need-discovery, product-recommendation, realtime-sales-coach, objection-handling, human-handoff, conversation-debrief.

It deliberately does not register any tool that sends messages, places calls, changes prices or discounts, or issues refunds. That boundary lives in code, not in a prompt.

The core design decision

Everything that can be computed deterministically is computed by code, not delegated to the model:

Code decides	Model does
Which phrase cards are retrieved (keyword + character n-gram + metadata, fused)	Organise the retrieved evidence into one sentence a salesperson can say out loud
Content gating — only approved, in-effect content is retrievable	—
The eight human-handoff conditions	—
Forbidden-claim and unauthorised-number interception	—
Source allow-listing (a model citing an unprovided source is discarded whole)	—
Duplicate-suggestion suppression, idempotency, audit	—

Model output passes three gates: constrained input, structural validation plus a source allow-list, and pre-display interception. Any failure falls back to a template suggestion built directly from the top approved phrase card — and the panel labels it as degraded. Degradation is a designed path, not an error path.

Install

See docs/install.zh.md. Short version: point the Plugins manager's "add plugin" at this directory's absolute path, then restart the app.

No npm install is needed — dependencies is empty; only Node built-ins and injected platform services are used.

The npm package is @cqaiclub/cqai-dsh-plugin-xiaobao. Install it through the app's plugin manager or dsh plugin --profile <profile> add @cqaiclub/cqai-dsh-plugin-xiaobao@0.2.1. When migrating from the old unscoped package, disable it before enabling the new package; existing data stays under $DSH_HOME/xiaobao/. See release instructions and changes.

Data location

$DSH_HOME/xiaobao/knowledge/ — phrase-cards.json, facts.json, audit.jsonl. Seed knowledge (30 phrase cards, 12 facts) is injected only into an empty store. The seed content exists to make the plugin usable out of the box; it is not trustworthy business fact — replace it before real use.

Develop

node scripts/build-seed.mjs      # regenerate seed knowledge
node scripts/build-fixtures.mjs  # regenerate evaluation fixtures
node scripts/eval.mjs            # offline evaluation → docs/eval-report.md
node --test-isolation=none --test "test/*.test.mjs"
node scripts/selfcheck.mjs       # pre-delivery check (includes all tests + evaluation)

--test-isolation=none is required: the default isolation mode spawns a child process per test file, which the DSH Windows sandbox forbids (EPERM).

Known limitations

Stated plainly so the boundaries are not misjudged:

• Voice is segmented. The platform accepts one complete 16 kHz mono WAV and returns its transcript. A pause of about 700 ms submits a segment, with a roughly 12 s cap during continuous speech; finishing dictation also submits the remaining audio. Recognition latency depends on the local model.
• Speaker labels are manual. The platform does not diarise; customer/sales is a manual toggle and can be wrong.
• Retrieval is in-process JSON plus character n-grams, not pgvector or BM25. Fine into the low thousands of cards.
• Evaluation is over-fitted. Samples and seed cards come from the same fixture, so the reported metrics are 100%. That shows the pipeline and gates work; it does not show generalisation to real calls. See docs/eval-report.md for the held-out re-measurement procedure.

License

MIT — see LICENSE.