@guosheng_047/dsh-erp

Adapt ERP · dsh-erp

简体中文 · npm · Releases · Changelog

Learn your ERP through its browser interface, and keep the knowledge on your machine. Adapt ERP is a plugin for DeepSeek Harness (DSH). Configure your ERP, log in yourself, and let DSH observe menus, pages, fields and business data, build evidence-backed knowledge, and reuse it in later conversations.

The plugin starts without site-specific knowledge. It does not require a particular ERP vendor, API path or login-token format. Compatibility depends on the site's interface and authentication; this preview does not claim successful operation on every ERP.

Preview interaction policy: page observation and local knowledge accumulation are automatic after you confirm your login scope. Every click, fill, selection or scroll asks for individual approval, including queries—an unfamiliar control may save data. This is AI-guided learning with confirmations, not unattended crawling or a transaction engine.

Version 0.1.0-alpha.8 targets DSH 0.1.6-alpha.2 (prerelease). Browser Use / Computer Use providers are not enabled by this upgrade.

Install

Requirement	Current baseline
Node.js	>=24.18.0 in this ERP's UI, inspect its stock and related orders, and record the steps and uncertainty.”

• “Continue the unfinished learning round. Review changed fields and export the knowledge as Markdown and JSON.”

The second example is a model-planned UI task, not a built-in product/order API or a guaranteed accounting reconciliation.

Knowledge belongs to your installation

Each system has systems/ / under the local data root, containing SQLite, evidence, exports and private browser profiles. erp_system_status reports its exact scope and directory. With the same configuration/data root, later conversations can read saved knowledge without opening a browser. Live data requires a current login; old balances and orders remain historical observations.

User knowledge, browser login data, test fixtures and benchmark answers are not included in npm releases. The package ships the learning tools, generic schemas and user guides. SCM/USA examples now live only in the development test baseline. Local knowledge sharing is an explicit user export, not part of publishing the plugin. Markdown/JSON exports are readable projections; SQLite retains revisions and evidence links.

Observations supplied to DSH may be sent to your configured model provider. Snapshots exclude password/hidden/file inputs, obvious credential fields and data-erp-private regions, but redaction is best effort. Tables and ordinary fields can contain business data. No credentials are requested through the interaction tool.

Controls and limits

• Every page interaction needs a separate DSH approval. This also covers write buttons and autosaving inputs; denying the request sends no interaction. Observation authorization alone never permits an action.
• Actions use short-lived snapshot refs, not model-provided selectors, JavaScript or request URLs. Changed targets, navigation, stale revisions or reused refs are rejected. A failed action may already have reached the site: inspect the result before trying again.
• Observation grants expire after 10 minutes and pause on human input/navigation. Action approval includes reading its result and a renewed bounded observation window. Login and grants are rechecked after restart.
• One configured system and one browser page per instance. Same-application frames are supported; cross-origin frames, canvas/closed-shadow interfaces, multi-window login and QR/SSO flows may require manual handling and are not broadly accepted yet. Native browser dialogs remain for the human.
• Snapshots are bounded, non-transactional samples. Multi-step business transactions, automatic retries, unattended write workflows and automatic selection among multiple ERPs are not provided.

Upgrade and remove

Before upgrading, call erp_storage_backup, stop DSH, repeat the install command and restart. Keep the same data root and configuration.

From alpha.5: system IDs and stored knowledge remain. Fixed erp_scm_* tools and SCM menu import are removed; use the browser learning flow. Existing SCM-derived records are historical evidence, not preverified methods for another ERP. Reobserve pages before using an old procedure.

From alpha.4 or earlier: configuring system uses a separate store. The original root-level data is retained, not automatically imported. With DSH stopped, temporarily omit system to query/export/back up the old store under its original scope; browser connection is disabled in that mode. See storage and recovery. For alpha.1/alpha.2 TGZ installations, first remove the old unscoped plugin entry and enable only the scoped package.

After stopping DSH:

npm exec --yes --package=pnpm@11.7.0 -- dsh plugin --profile web remove @guosheng_047/dsh-erp

Uninstalling retains your data. One process owns a store at a time; independent DSH processes cannot concurrently share it.

Development and validation

npm ci --ignore-scripts
npx playwright install chromium --no-shell
npm run verify
npm pack

The build cleans dist so removed adapters cannot survive into a later package. Full tests run locally; the manual Release Action builds, publishes through Trusted Publishing and checks registry/GitHub bytes.

The generic path is exercised with distinct synthetic ERP UIs, empty stores, real Chromium/worker IPC and native DSH approvals. Deterministic test decisions are not a real-model compatibility benchmark. Historical live SCM acceptance is retained as a separate baseline and is not proof that the new UI workflow works on every ERP. See learning and boundaries and development documentation.

MIT. Dependencies retain their own licenses; see notices. browser-use and browser-harness informed the research; Playwright is the current browser runtime.