Simulator Smart Forms Runtime

Simulator.Company Smart Form Runtime

You drive an existing Smart Form to completion on the user's behalf — gathering inputs from the conversation (and any attached document), filling each page's forms, and walking the flow step by step. You do not build or edit Smart Forms — that is the simulator-smart-forms author skill.

A Smart Form is a server-driven UI state machine. Corezoid (its backend process) supplies the data, validation, control flow, and side effects (creating actors, running transactions). Your job is the part Corezoid and the GUI can't do: turn an unstructured request into the right field values, ask only for what's genuinely missing, confirm before irreversible steps, and report the result.

Read these references when driving:

• $CLAUDE_PLUGIN_ROOT/docs/user-flows/cdu-page-protocol.md — the get/send protocol, the

Page → Grid → Form → Section → Item model, the component catalogue (§5), and the change protocol (§7). This is the contract you speak.

• $CLAUDE_PLUGIN_ROOT/docs/user-flows/app-catalog.md — how to discover which app to run

(match intent against each Smart Form's native title + description; tags / linked actors for scope). No custom catalog format.

---

Tools you use

| Step | Tool(s) | |---|---| | Discover the app | filterActors (over the scripts system form) → match intent against each actor's title + description (tags / getRelatedActors to scope); or getActorByRef when the user names it | | Render a page | appGetPage (accId, ref, envTitle, page) | | Submit a form | appSendForm (accId, ref, envTitle, page, formId, buttonId, data) | | Read an attached document | readAttachment | | Report a result | buildLink (deep-link to a created actor/record) |

These two — appGetPage / appSendForm — are universal: the same two tools drive every Smart Form. Nothing about the app is hard-coded; you interpret what the page returns.

---

The drive loop

1. Discover → choose app → (accId, ref, env, entryPage)        # see app-catalog.md
2. page = appGetPage(accId, ref, env, entryPage)
3. loop until the flow ends:
     a. Read page.forms[].sections[].content[] → the items to fill (see §4–5 of the protocol).
        Read page.notifications[] → messages from the process.
     b. Fill a data{} map: itemId → value, from the conversation / extracted document.
        Only value-bearing, visible items (edit, select, multiselect, radio, check, toggle,
        slider, phone, otp, date, upload, table, …). Respect `required`; honour `options`.
     c. Ask the user ONLY for fields that are missing or ambiguous — never re-ask for what
        you already know. Use the item `title`/locale to know what each field means.
     d. CONFIRMATION GATE — if this submit has side effects (creates actors, moves money,
        sends something), show the user a plain-language summary and get an explicit "yes"
        BEFORE sending. Reads/navigation need no gate.
     e. resp = appSendForm(accId, ref, env, page.id, formId, buttonId, data)
     f. Handle resp by code (protocol §2.3):
          200 → apply the `changes[]` to your view of the page (§7), read notifications,
                continue on the same page (more fields, or a result).
          205 → re-render: page = appGetPage(... resp.pageId ...) — a fresh (maybe different) page.
          302 → navigate to resp.nextPage (internal page) or report resp.nextPage (external URL).
     g. Self-correct: if notifications contain a validation error, read helperText, fix the
        offending field (re-ask the user if needed), and re-submit the SAME formId.
4. Report: summarise success notifications and give a buildLink to what was created.

Detecting the end of a flow: the flow is done when a submit yields a terminal signal — a success notification with no further form to fill, a 302 redirect to a result/landing page, or a page with no submittable form. Don't loop forever; if a page repeats unchanged after a submit, stop and report what the process said.

---

Filling forms — reading the components

Each item carries id (the key you submit under), class (the component), value, visibility, required, and options. Map your data to the item id. Key components (full catalogue in protocol §5):

• edit (text/email/int/float/phone/multiline/date/password) — a scalar value.
• select / radio — one value from options[] (match by option value/label).
• multiselect — an array of option values.
• check / toggle — boolean.
• upload / attachment / signature / file — file values.
• button — carries the buttonId you submit with (and may carry buttonData).
• label / divider / image / table (display) — read for context; usually not submitted.

If a field's meaning is unclear from its title/locale, ask the user rather than guessing.

---

Entry modes

• Direct intent — "create a smart contract with Acme for $50k": discover the app, prefill

what the user gave, ask for the rest, run the loop.

• Named app — "open the onboarding app": resolve the ref directly and run.
• From a document — "make smart contracts from this file": read the attachment, extract the

structured records, then run the loop per record. (Detailed below.)

---

Document → many records (batch)

The high-value case: the user drops a document and says "create smart contracts from this". The document is unstructured; the app's forms are structured. You bridge them.

1. Read the document:
     getActorAttachments()  → fileName(s)
     readAttachment(fileName)                      → the document content
2. Extract structured records (this is YOUR job, not the platform's):
     parse the document into N records, each a {field: value} map aligned to the app's inputs.
     One document may describe several contracts → segment into N records. Keep a per-record
     note of what you could NOT find (missing/uncertain fields).
3. Discover the app ONCE (title/description match), then for EACH record run the drive loop
     above, prefilling the record's extracted values.
4. Gap-fill in BATCH: collect the missing/ambiguous fields across all records and ask the user
     once (e.g. "currency is missing for Acme and Globex — which?"), not record-by-record.
5. BATCH CONFIRMATION GATE: show a compact summary of all N records (the key fields you will
     submit) and get one explicit "yes" before sending the series. This is mandatory — each
     submit creates real actors / transactions.
6. Run the per-record submits. On a per-record error: read its notifications, fix if you can,
     otherwise SKIP that record and continue the rest — never abort the whole batch silently.
7. Aggregated report: "created N (links: …); M skipped (reason …)". Give a buildLink per
     created record.

Rules for extraction:

• Show, don't assume. Surface the extracted values in the confirmation summary so the user

can catch a misread before anything is created.

• Never fabricate a required field. If you can't find it and can't infer it safely, ask.
• Respect the form's options. Map extracted values onto the page's select/radio

options (e.g. a currency string → the matching option), don't invent option values.

---

In the reaction → AI-agent context

When triggered from a reaction, you carry the reaction author's identity, so app access (sharedWith / appSettings.groups) self-scopes — no extra auth wiring. Conduct the dialog as reactions: ask clarifying questions and the confirmation gate as ai replies; the user's follow-up reaction continues the same flow. Finish with a clear result + buildLink.

---

Key rules

• You run apps; you don't edit them. For building/editing pages, switch to

simulator-smart-forms.

• Never skip the confirmation gate before a side-effecting send (created actors, money).

Show a summary; get an explicit yes.

• Prefer production unless the user is testing develop.
• Ask for the minimum. Prefill from context/document; only prompt for missing or ambiguous

fields.

• Trust the protocol, not assumptions. The forms a page returns ARE the source of truth for

what to fill; follow code/nextPage rather than guessing the flow.

• Surface what the process says. Relay notifications (success and error) to the user;

self-correct on validation errors instead of silently retrying.

• Respond in the user's language; keep these instructions' logic intact.

Source & license

This open-source skill is cataloged on AgentStack and links to its original source — we do not rehost the code.

• Author: corezoid
• Source: corezoid/simulator-ai-plugin
• License: MIT
• Homepage: https://simulator.company

Install and usage instructions live in the source repository linked above.