Property-Based TUI Testing Working Notes

Mock LLM Provider

SessionPrompt gets model metadata through Provider.Service.getModel(...).
Actual generation is routed through LLM.Service / streamText(...), so the first mock should return an AI SDK LanguageModelV3 from the provider service.
The normal Provider.layer builds providers from config/models.dev/plugin state. For first pass, the simulated graph can replace Provider.Service with a smaller simulation provider service instead of trying to flow through provider config.
Control state should own an ordered LLM script queue. The provider/model should consume from that queue when the AI SDK calls the language model.
First version can support text-only output. Tool calls and stream chunk fidelity can come next.
Missing script should fail loudly with a typed simulation error, not silently return an empty assistant message.
Implemented SimulationProvider.layer, replacing Provider.Service in createSimulatedRoutes.
The provider exposes provider simulation and model mock.
doGenerate and doStream both consume one queued script through Simulation.Service.nextLLM().
Current script support: text, thinking (treated as text for now), and error.
Snapshot currently records llmQueued and llmConsumed, not per-step details yet.

OpenTUI Solid exposes testRender(...) from @opentui/solid.
The lower-level core API is createTestRenderer(...) from @opentui/core/testing.
createTestRenderer(...) returns renderer, mockInput, mockMouse, renderOnce, captureCharFrame, captureSpans, and resize.
captureCharFrame() is the simple screen-buffer string API used heavily in OpenTUI snapshots.
captureSpans() returns structured lines/spans plus cursor position, which is a better starting point for visible element discovery than parsing raw characters.
mockInput supports interactions like typeText, pressEnter, and pressArrow.
Implemented TuiSimulation.createSimulationRenderer(...) beside thread.ts. It creates a test renderer and exposes renderOnce, screen, spans, and destroy.
thread.ts checks OPENCODE_SIMULATION, creates the fake renderer there, starts the normal worker/backend, and passes the renderer into tui(...).
Backend route assembly checks OPENCODE_SIMULATION_BACKEND; OPENCODE_SIMULATION implies this flag.
OPENCODE_SIMULATION_BACKEND=1 can run a real frontend against a simulated backend.
tui(...) now accepts an injected CliRenderer, test mode, and an onReady callback. Production still creates the real renderer.

Interactable discovery can walk renderer.root.getChildren() recursively.
Renderable.focusable and Renderable.focused are public and enough to discover focus targets.
renderer.currentFocusedEditor identifies active text input/edit-buffer targets for typing/submission.
renderer.hitTest(x, y) maps terminal coordinates through the hit grid to a renderable id.
Renderables have public geometry: screenX, screenY, width, height, and num.
Mouse listener metadata is stored internally on renderables; first pass checks _mouseListener / _mouseListeners at runtime to identify clickable targets. This is pragmatic but not a stable public API.
Test execution uses mockInput.typeText, mockInput.pressEnter, mockInput.pressArrow, and mockMouse.click from OpenTUI testing.
Implemented SimulationActions with elements(...), actions(...), and execute(...).