Skip to main content
Version: 4.2 (beta)

Manager Assistant

The Manager Assistant is available inside the HSYCO Manager. It helps administrators and developers inspect HSYCO documentation, runtime state, project context, open editor buffers, files, and other Manager information.

See the Assistant overview for common chat, session, upload, download, and file cache behavior.

Opening the Manager Assistant​

Open the Manager, then click the Assistant button in the Manager toolbar.

The Assistant opens as a side panel. The composer at the bottom is used to write a message, attach files, select a model, and start a new session.

Available context​

The Manager Assistant can use Manager context, including:

  • the active Manager application;
  • open Project Editor context;
  • open Code Editor context;
  • selected model and reasoning settings;
  • files uploaded in the current Assistant session.

When the Project Editor is open, the Assistant can inspect project pages, objects, and selected project context. When the Code Editor is open, it can use the current editor context to help explain, review, or modify code.

Tool access​

The Manager Assistant can call tools to inspect and operate on HSYCO data. Depending on the configured model and available tools, it can:

  • search HSYCO documentation;
  • inspect runtime status;
  • read events and project files;
  • inspect Project Editor state;
  • materially recompose the current Project Editor page from a reviewed visual reference and apply safe layout changes for review;
  • read uploaded files;
  • create downloadable files.

Tool calls are shown in the transcript as work blocks. The final response is shown separately.

See Assistant Tools for the complete list of Manager Assistant tool bundles and tools.

Checking how a page looks​

After building or rearranging a page, the Assistant captures it as Project Editor shows it, with unsaved edits and without editor outlines, and fixes what does not look right, such as text that does not fit, crammed controls or uneven spacing. You can also ask it to check how a page looks on a phone or a tablet. It can capture the area of a single object, also at twice its size, to read small text and icons. The capture shows objects without live values. Page images are sent to the profile's model provider like any other tool result.

Improving a Project Editor page​

With a page open in Project Editor, ask the Assistant to improve its layout and describe the desired hierarchy, grouping, or region emphasis. An OpenAI or Gemini Manager profile with image generation is required.

The Assistant captures the page that is currently rendered at a bounded roughly 2K pixel density, including the current project skin, and removes Project Editor outlines and visual editing noise only in the detached capture. The selected chat model first acts as a design director and creates a composition brief grounded in the real Project Editor object paths. Its compact HSYCO geometry rubric covers one primary region, visual scan order, task grouping, shared alignment edges, differentiated gaps, content-aware sizing, and visible-content integrity. The image provider edits the source into a full-page mockup. A visual critic then checks content preservation, coherent macro-layout novelty, projected current-skin viability, and macro-geometry extractability; arbitrary rearrangement or grouping that depends on candidate-only chrome cannot pass. It can request one corrected candidate before the operation stops without editing. The mockup is an aspirational layout reference, so candidate-only styling is discarded. A separate layout pass uses the same rubric to reconcile only an accepted composition with the real paths and unchanged skin.

In release 4.2, the workflow can change only pos, width, and height on visible existing objects; it cannot create, remove, rename, reparent, restyle, or change component behavior. It rejects simple grid normalization: pages with enough regions require at least three material top-level changes, and pages with five or more compact launcher groups require at least two meaningful vertical bands.

The page is checked again before changes are applied. If it changed while generation was running, the operation stops without editing it. After the first unsaved layout batch, the Assistant captures the real rendered current-skin result and compares it with the accepted mockup. It can apply one revision-checked geometry-only correction and then performs one final comparison. A failure during this post-apply review is reported as verification status because the initial edits already exist; it does not save the page or silently claim that no change occurred.

Inspect the returned brief, acceptance scores, mockup, reconciliation status, and explanation, then use Undo, adjust the page, or save it normally. Initial and corrective geometry changes use the normal object-level Project Editor Undo history, so reverting a result can require several Undo actions. The original page capture and any rejected candidate are not kept as Assistant artifacts; only the accepted mockup is kept in the session. The visual review can send source, candidate, and post-apply images to the provider more than once, and roughly 2K/high-detail processing increases latency, vision tokens, and image cost. Because these images can contain visible project data, treat them according to the selected providers' data policies.

Live voice in a chat​

With a supported direct OpenAI profile, select the Start voice icon button in the composer to talk in the current normal chat. Allow microphone access when prompted. You can keep typing while Live is connected; the same Assistant history and profile permissions apply to spoken and typed requests.

The normal model selector chooses the backend thinking model (Base/Advanced in the Client). Changes apply to subsequent requests without restarting voice. Live uses the preferred compatible voice model from the catalog, independently of the thinking model.

Captions are hidden by default. Use Show captions or Hide captions in the top-right three-dot menu to toggle their display above the composer. This menu item appears only when the current chat has captured Live captions. The voice controls use icon buttons with tooltips. Mute pauses your microphone while allowing playback; End Live ends voice and keeps the chat and any backend work. Live also ends automatically after one minute of silence while the thinking model is idle; the chat remains open. Closing the panel, changing chats, or changing the profile also stops voice capture. If the browser blocks playback, select Play audio. Microphone access requires HTTPS or localhost.

Live uses the profile's existing tools and restrictions. Full backend results remain in the chat. Voice duration is billed separately from backend model usage and included in the session cost; final usage can be incomplete after an unexpected disconnect. Reconnecting restores bounded recent context, not a live audio connection across page reloads.

Uploads and downloads​

Manager Assistant uploads are scoped to the active Manager Assistant session.

When the assistant creates a file, the response shows an inline download link for the generated filename.

Use Download transcript in the Assistant options menu to save the visible conversation as a text file. To save the raw event log instead, use Download session file in the Assistant options menu or in that conversation's session-list menu. The downloaded .jsonl file contains the persisted session events.

See File cache for upload and generated download retention.

The Manager Assistant complements the Project Editor and Code Editor. Use it when you want to ask questions about the current project, code, runtime status, or documentation without leaving the Manager.