Smart Actions · Eidos Plugins

About this app

Use Jev to classify requests, assess urgency and score feedback—then write the results back to your .eidos file.

Select inputs → define rules → choose output fields.

Create reusable actions by hand, or describe what you need and review an AI-generated draft. One action can read multiple fields and produce multiple results.

What you can build

Input	Judgment	Output
Customer message	Which department should handle it?	Category · Select
Message and subject	Does this need immediate attention?	Urgency · Number, 0–1
Customer feedback	How dissatisfied is the customer?	Score · Number
Customer message	Is a refund explicitly requested?	Refund requested · Checkbox

Jev performs classification, scoring and probability judgments. Actions do not generate free-form replies or summaries, create records, or add fields.

Before you start

Smart Actions requires plugin API 1.1.0 and uses package format 2. Upgrade older clients first. CLI Serve does not yet support its action and configuration APIs. The plugin API contract is independent of the npm SDK version.

Eidos Lite 0.17.0 or later, with plugin API 1.1.0 support.
A TypeSafe AI API key to access Jev.
A .eidos table with the input and output fields you want to use.

An additional generation model is optional; manual configuration does not need it.

Install and connect

Drag the .eidos-plugin package into Lite's Plugins page and enable it for your Space. To update, drag the new package into the same page.
Open Plugins → Smart Actions → Settings → TypeSafe AI and save your API key.
Right-click a .eidos file and choose Open with → Smart Actions.

Create an action

Select a table in the sidebar, then ＋新建动作 → 手动创建.

Inputs: select the fields the action should read.
Outputs: add a destination and choose an existing writable field.
Rules: select an output in the flow map to edit its judgment type and rule below. For example: “Use Message to choose the department that best matches the customer's main request.”
Name the action, optionally choose an icon, and Save.

Actions are grouped under their tables. Drag the sidebar's right edge to resize it; double-click to reset. Switching tables or outputs preserves unsaved edits while the editor remains open.

Generate a draft with AI

Configure Plugins → Smart Actions → Settings → Action Generator with:

The full public HTTPS Chat Completions endpoint, including /v1/chat/completions.
The provider's model ID.
An API key for that endpoint.

Choose ＋新建动作 → 用 AI 生成动作 and describe the result you want:

Classify each customer message using Category's existing options, and put the probability that it needs immediate attention in Urgency.

Review the fields and rules, edit as needed, then save. Generating a draft does not execute it or modify records. The generation model designs the configuration; Jev still performs the judgments when you run it.

Run, stop and undo

Open the file with the built-in editor. In Grid, right-click a record or selected records and choose your action. Select all includes matching records not yet loaded into the grid.

Validated results apply directly. The task window shows progress and estimated Jev execution cost. Stopping a run preserves completed writes.

Undo this run restores completed writes without overwriting later edits.
Redo this run reapplies stored results without another model request.
Undo is available in the current table/run session, not after restarting Lite.
Closing the task window hides it; use the stop control to stop processing.

Targets are fixed when a run begins. Later filter changes and newly added records do not expand them. Deleted records are skipped. Errors or conflicts stop the run and retain previously completed writes.

Output types

Judgment	Destination	Behavior
Classification	Select	Reuses current field options; manage them in field settings.
Classification	Text	Chooses from categories defined in the action.
Score	Number	Uses ordered levels starting at 0; results can be fractional.
Probability	Number	Writes a value between 0 and 1.
Yes / No	Checkbox	Applies a configurable probability threshold.

Classification needs 2–255 distinct, nonempty choices; scoring needs 2–10 ordered levels. Outputs are independent: rules cannot rely on other outputs' results from the same run.

Data and connections

Action definitions are saved with their table in the .eidos file. API keys are stored separately in Lite's encrypted connection store, scoped to the current Space. Changing a generation endpoint requires entering its key again.

Operation	Data sent to the provider
Generate a draft	Your description and field structure/options; no record contents.
Run an action	Selected input values, judgment rules and choices.
Undo or redo	No model request.

Execution batches up to 10 records per request, with two batches in flight. Large inputs may use smaller batches. The displayed cost is an estimate from returned usage, not an invoice; unavailable usage is not treated as free. Generation-provider charges are separate, and undo does not reverse charges.

Troubleshooting

Cannot generate: configure Action Generator and check the full endpoint, model ID and credentials.
No suitable output field: create it in the built-in editor first.
Missing classification options: add them in the Select field's settings. Each run reads the latest options.
Only some records updated: inspect the task error and resolve the cause, or undo completed writes before running again.

Development

Source code and co-located tests live in src/. plugin.json declares the extension and settings entry points; build and smoke-test scripts live in scripts/.

npm ci
npm run check
npm test
npm run pack:plugin

The package and checksum are written to dist/. There are no runtime dependencies. Action configuration lives in settings_json.plugins["eidos.smart-actions"]. Runs are limited to 100,000 records and the host's 64 MiB undo storage limit.

scripts/live-smoke.mjs accepts TYPESAFE_API_KEY or a key via stdin and tests with synthetic data. Do not commit credentials.

Jev API documentation