Expand BasicAction types to make the Agent smarter

[NakaokaRei]

Summary

Currently BasicAction only supports 10 action types. The Agent often needs to combine multiple steps for common operations (e.g., move + leftClick to click a UI element), which introduces errors when the screen changes between steps.

Adding 6 frequently-used composite/convenience actions will make the Agent more reliable and efficient.

Proposed New Actions

Action	Parameters	Rationale	Maps to
clickAt	x: Double, y: Double	Atomic move+click — the most common GUI operation. Eliminates race condition between separate move and click.	Action.sequence([.move(to:), .leftClick])
tripleClick	none	Select entire line/paragraph. selectAll is too broad when targeting a specific line.	Action.tripleClick()
typeAndEnter	text: String	Type text and press Enter — very common for search bars, URL bars, form fields.	Action.typeAndEnter(text)
copy	none	Command+C. More explicit than keyShortcut(["command", "c"]) — LLMs generate it more reliably.	Action.copy()
paste	none	Command+V. Same rationale as copy.	Action.paste()
selectAll	none	Command+A. Same rationale as copy.	Action.selectAll()

Files to Modify

• Sources/SwiftAutoGUI/ActionGenerator.swift — Add 6 enum cases, update toAction(), Codable conformance, and ActionType enum
• Sources/SwiftAutoGUI/OpenAIVisionBackend.swift — Update system prompt, parseAction, JSON schema, describeBasicAction
• Sources/SwiftAutoGUI/OpenAIBackend.swift — Update system prompt (JSON schema auto-inherits)
• Tests/SwiftAutoGUITests/ActionGeneratorTests.swift — Add round-trip, parsing, and conversion tests
• FoundationModelsBackend.swift — No changes needed (@Generable handles new cases automatically)

Design Notes

• All new cases use only String and Double parameter types → fully @Generable compatible
• No new JSON schema properties needed: clickAt reuses x/y, typeAndEnter reuses text
• System prompt should recommend clickAt over move + leftClick for clicking UI elements
• Total action count goes from 10 → 16, still lightweight for on-device model context window