Skip to main content
Workspace pages include a top-right Preview button. Use it to chat with the page agent while the app runs the current Procedure blocks and shows live activity beside the page.

​
What it does

  • Opens a right-side chat panel from the page header.
  • Keeps the browser on the workspace page instead of navigating to /chats/new.
  • Lets the user drag the Preview panel’s left edge to resize the page and chat columns.
  • Includes a new-thread action in the Preview header so a page test can be restarted cleanly.
  • Lists prompts from the page’s ## Test prompts section.
  • Sends each chat message to the direct Agents SDK Preview runner with the current Markdown page source. The server re-parses that Markdown into executable page nodes before running Preview.
  • Saves the Preview thread to the same active AI chat list used by the sidebar when WorkOS and Supabase chat persistence are configured.
  • Marks saved Preview chats as waiting for approval in the Chats sidebar when an approval-required step pauses, and lists the same request in the Approvals side panel.
  • Opens saved Preview chats in the full chat route with the same visible activity feed and pending approval card shown in the page side panel.
  • Lets the user reject a pending approval with inline feedback; Preview cancels the paused run and sends that feedback as the next user message in the same thread.
  • Compiles inline @ tags through the Procedure tag registry before execution.
  • Treats @read_person, compatibility @find_person, @create_person, and @create_position_sf as supported Preview tools, asks the Agents SDK model planner to call them with their approved JSON schemas, and records their results in the activity trace.
  • Shows a visible registry error when the Markdown contains an unknown @tag.
  • Compiles authored :::if / @condition blocks into Classify Agent runtime nodes. The classifier returns a structured branch id, then the app routes deterministically.
  • Returns a conversational assistant reply from the page Preview agent after the visible activity finishes.
  • Shows an inline assistant reasoning feed for the Start input, agent steps, tool calls, Classify Agent decisions, database reads, and errors. The reasoning feed stays open while Preview is working, collapses after completion, and leaves the final assistant reply as a normal chat message.
  • Keeps the raw event log boxed below the assistant reasoning feed for exact trace details and metadata.

​
What it does not do yet

  • It does not publish immutable Procedure versions.
  • It does not run the full simulation library yet.
  • It does not capture microphone or system audio.
  • It does not use Inngest for the direct page Preview MVP path.
  • It only wires the first person-check Preview tool path, @read_person / @find_person then conditional @create_person, and one Classify Agent branch pattern for the first MVP slice.

​
Requirements

Preview uses the workflow page run route, the Markdown procedure runtime compiler, the direct Agents SDK Preview runner, and the Agents SDK Preview responder. Live model behavior needs WorkOS AuthKit, a signed-in user session, and OPENAI_API_KEY configured on the server. Saved Preview threads also need the Supabase chat-thread environment used by the main chat surface. If the key is missing, Preview fails visibly when it reaches model-backed tool planning or Classify Agent evaluation instead of silently extracting tool arguments with deterministic parsing. When the selected branch asks for @create_person or @create_position_sf, the model planner must call the schema-backed tool with only schema-required fields unless the workflow Markdown asks for more. People writes store identity facts; SuccessFactors org facts such as title, worker classification, start date, and reporting line belong in positions_sf. The Preview agent response falls back to a deterministic summary of the visible trace only after trace-backed work exists; it must not claim hidden model reasoning or missing tool calls.

​
Hosted documents and models

Hosted previews save document content and table data in the database, so edits survive refreshes and redeployments. A new hosted environment starts with an empty workspace; the local demo’s example documents are kept separately. Saved revisions are retained while the document exists; a revision-history interface is not yet available. The default text agents use GPT-6 Astra. Chat offers Extra High, High and Balanced reasoning profiles. Previously saved GPT-5.5 profile selections use the matching Astra profile. Voice transcription uses a separate audio model.