Frequently asked questions

A Windows desktop workbench for visual acquisition, manual and CV-assisted markup, result review and handing a canonical project to automation, AI tools, IDEs and Screph Code.

No. Markup and Project JSON do not execute input automatically. Actions require a selected automation-runtime scenario, a visible target window and a configured input backend.

No. This is a narrow runtime preflight returning (ok, errors), which must be checked explicitly. It does not replace canonical schema/artifact diagnostics or validate hashes/path containment, target-window visibility/foreground state or the selected input monitor. The monitor guard runs separately during actual pointer actions, and the workflow must verify each action's postcondition.

Parameter-driven live preview is not saved. Manual Apply in the main panel runs analysis and may attach or replace a cv_result in the selected element's image series by default, but it does not accept a candidate as canonical geometry or semantics. That requires Add as elements or a confirmed mode review/apply flow.

It creates a technical run bundle and legacy/debug projection in a named runtime session under ScrephData/screen_selector/cv/results. This is not a second project format or automatic markup apply. Only CV entries attached to element image series enter the canonical package; the recent list keeps up to ten IDs but does not automatically delete old session directories.

Apply/edit writes the revisioned Selector sidecar <project>.temporal-markup.json, so it needs a stable project path. This is a separate file, not embedded canonical JSON. Copy it with the project and, after current Project Manager duplicate/rename, manually match its basename to the new project JSON. Video CV runs remain separately under ScrephData/screen_selector/cv/results/temporal, and Hide active run does not delete them.

The 28 entries are single methods in the main panel. The other 14 are preprocessing, transform, mask and contour utilities for pipelines, not 14 extra detectors. GUI Automation uses shared CV and adds domain types and review/apply rather than a second catalog of 42 algorithms. See the CV guide for the full list.

No. The Assistant is off by default, sees only allowed context and proposes registered actions. A policy gate can block an action entirely, while medium/high/destructive actions require explicit confirmation. See the Assistant guide for details.

Yes. Screph Code provides Builder through a separate Pro Agent process. The current Builder is a write-capable agent inside the selected workspace: sending a request can create or modify files immediately, without a universal mandatory diff/Apply. Make a backup/commit first and inspect the disk diff afterward. Running the generated file is a separate action.

Capture, Selector, most CV, local Ollama/LM Studio and Vosk recognition can work without a cloud service after their components and models are provisioned. On startup, an installed build may automatically check for a product update and refresh a six-page help cache; with a stored account, the profile refreshes every 30 seconds. First opening the Vosk list fetches the Alpha Cephei catalogue. Activating OpenAI/LiteLLM Speech settings checks the provider, with the LiteLLM path sending a synthetic silent WAV. Model archives and user cloud operations require network access separately.

For device login, profile/balance, support chat, server speech and manual diagnostics submission. With a stored token, the desktop refreshes the profile every 30 seconds; a visible support chat fetches its thread about every 20 seconds. The core local authoring workflow operates independently of the account. See the exact boundary →

Do not rely on this as a published workflow. The balance and payment backend are implemented, but the current purchase page redirects to the homepage; no current price or checkout is published on the site.

The runtime, local weights and active backend are three separate checks. Provision the required files on the model-family page, refresh status or SHA verification and select a downloaded model for the method. Follow the CV model guide for the complete map.

Local projects and sources are not uploaded on their own. Automatic update/help requests send ordinary HTTP metadata; account refresh sends the stored token; the Vosk list only requests its catalogue. OpenAI Speech settings send the key in a model-list GET, while LiteLLM settings may send a generated 0.1-second silent WAV—not a microphone recording. User content is transferred by a selected cloud LLM/speech, support or diagnostics workflow. Assistant attachments are off by default and Action CV applies a privacy gate. A telemetry flag alone does not start an exporter; no such client was found in the current runtime.

No. Project Manager does not delete a project directory: this command only changes the local recent-path list, while Clear missing removes paths that no longer exist. Manager tags and notes live in a separate path-keyed metadata store and are outside the canonical package.

No. These fields control the preview and an explicit feature projection for downstream tools, while canonical JSON stores the complete feature payload, assets, summary and original screenshot coordinates. The primary save flow does not currently create a separate projection file.

General and GUI Automation have ready end-to-end workflows. Industrial and UAV are available as limited foundation/domain workspaces. Ground Robotics remains an internal deferred scaffold.

The user-facing summary is on the status page; source changes, licenses and detailed engineering documents are on GitHub.

The current installer and supported product target Windows 10/11 x64. Other operating systems are roadmap items and should not be considered supported.