AI, Screph Code and automation
These are three downstream paths in the complete program. They use a reviewed project but have different permissions, dependencies and confirmation mechanisms.
Automation Runtime Implemented
The runtime and Automation Manager execute prepared actions through a configured input backend. The target window must be open, visible and selected before execution. Project JSON does not click anything by itself. A launched script is not isolated: it inherits the current user's permissions and environment, and stopping the direct child does not guarantee termination of processes it created.
Project loading, API, OCR and input backends →AI Assistant Experimental
The built-in Assistant is implemented but disabled by default. It uses a tool allowlist, previews and confirmation for mutating actions. Sending images and artifacts is also off by default.
Activation, workspace, context and action cards →Screph Code Implemented
Screph Code runs through a separate Pro Agent process and receives selected project/workspace context. The current Builder sends an agent request whose workspace tools can modify files inside the selected workspace immediately. There is no universal mandatory unified-diff Apply before every write, and Changes/Undo/Checkpoints do not automatically cover all direct agent edits. These are the current OS user's permissions, not a sandbox: use a Git branch/backup before the request and inspect the disk diff afterward. File execution remains a separate explicit action.
Open the Screph Code guide →Providers and local operation
LLM profiles support OpenAI, Anthropic, Gemini, OpenRouter, Ollama, LM Studio, a custom endpoint and LiteLLM Proxy. Local profiles avoid cloud upload, but availability depends on the selected model and running service. Profiles are assigned separately to Assistant, CV, Screph Code, semantic selection and test roles.
Connections, roles, voice and privacy →Pre-flight checklist
- The current project is saved.
- The outgoing data and images have been reviewed.
- An available LLM, runtime or input backend is selected.
- The preview is reviewed before applying changes or executing actions.