AI transform generator

Describe a transform in plain language and OpexFlow writes the DSL or JS, server-validated before it lands, via any OpenAI-compatible endpoint.

Describe the transform in plain language; OpexFlow writes it. The generator talks to any OpenAI-compatible /chat/completions endpoint, is steered to emit DSL first (falling back to JS when the rule isn’t expressible), and server-validates the output before returning it.

ai:
  enabled: true                            # master switch (default false)
  base_url: "https://api.openai.com/v1"   # any OpenAI-compatible base; plain http:// works (e.g. a local Ollama / stub)
  model: "gpt-4o-mini"
  api_key: "${OPEX_AI_API_KEY}"            # required; read from the environment, never returned by the API
  max_tokens: 4096                         # output budget floor; edits of long scripts ask for more (cap 16384)

Key	Default	Meaning
`ai.enabled`	`false`	Master switch. Off = the generator reports “not configured”.
`ai.base_url`	`https://api.openai.com/v1`	Base URL ending in `/v1`; `/chat/completions` is appended.
`ai.model`	`gpt-4o-mini`	Model id sent to the endpoint.
`ai.api_key`	none	Bearer token (`${OPEX_AI_API_KEY}`). Required.
`ai.max_tokens`	`4096`	Floor for max output tokens. When editing, the request grows with the current script’s length (up to 16384), since the model returns the whole script. A reply that still runs out fails with stage `truncated`.

GET /api/script/ai-status returns { configured, missing?, model?, base_url?, provider_host? } (never the key), so the editor can show a setup hint: missing lists ai.enabled / ai.api_key when they are not set.

In the flow editor, select a transform node and open Generate with AI under its script. Describe the rule (e.g. “emit only when temp exceeds 80 or vibration crosses 5.0”), pick a target language (auto prefers DSL) and hit Generate. The model maps your fields onto metadata.* using a sample record (the last real record seen for the node’s source, unless you pass one) and returns a runnable program; Replace script with this puts it into the node, and Test with last record checks it before you Deploy.

Editing and iterating. When the node already has a script, your request is an instruction to change it (“also drop records above 90 °C”): the model gets the current script and returns the complete updated one, with explanation summarising what changed. The last ~6 turns of the conversation are sent along so you can refine step by step.

Every answer is validated before you see it (sandboxed compile, then a dry run on the sample record). A rejected answer is not hidden: ok:false with stage (validation, dry_run, …), the error, and the rejected source, so you can ask for a fix.

POST /api/script/generate body: { instruction, current?: { backend, script }, history?: [{ role, content }], sample_record?, match?, target? } (description is the older name of instruction). Response: { ok, backend, source, explanation, sample_used, sample_key } or { ok:false, stage, error, backend?, source?, explanation? }.