How work plans work

The assistant never draws. It returns a work plan: which operation the engine should run, which library entries to use and which parameters to change. The engine — the same one that runs your recipes offline — draws, and the workshop checks run before you apply.

What is sent

The plugin sends the list of operations it supports, short summaries of your libraries (names, parameter labels, units, current values) and the parameters currently in use. Your .3dm file, client names and file paths are never sent.

Numbers, not names

The model chooses by number from short lists (“operation 1, cabinet entry 2, parameter 14 = 10”). The server maps numbers back to identifiers and validates every value: unknown entries, parameters of another profile and out-of-range values are rejected, and the model gets one chance to correct its plan. If it still fails, nothing is drawn.

Example

{
  "status": "ok",
  "plan": {
    "protocol": 1,
    "action": "apply",
    "reply": "Shelf front gap 6 → 10 mm; push-button latch.",
    "questions": [],
    "steps": [
      { "op": "dolap.propose",
        "params": { "profile": "cevahir-ornek-dolap",
                    "profile_overrides": { "raf_on_bosluk": 10 },
                    "latch": "mandal-pushbutton" },
        "reason": "The user asked for more front clearance." }
    ]
  },
  "usage": { "points": 3 }
}

action is apply (steps to run), ask (questions only) or unsupported (no recipe for this yet — a recipe can be taught from an example).

Quality tiers

Requests run on the gateway's cheap tier by default; you can ask for more effort (standard, smart) for difficult revisions. Which model answers is decided centrally by the Talivio AI gateway.