AI Segmentation modes
Three modes — None, Auto, and Force Fresh — that control how Mavster's AI looks at the screen for a step. Trade off cost vs. reliability per step.
The AI Segmentation picker controls how Mavster's AI processes the screen for a given step. It appears in two interaction editors: Tap and Scroll.
Three modes:
None
"Skip AI segmentation, use on device detection only."
Mavster uses on-device detection only. Cheapest mode.
Use when: Always try this first. On-device detection handles a wide range of targets — if it finds yours, you've paid zero credits for segmentation on this step.
Auto
"Automatic: Uses cache if screen unchanged, regenerates if changed."
Mavster reuses a cached result when the screen still matches; otherwise it re-segments. Pays the cost once per visual state rather than once per run.
Use when: None can't find your target. Auto is the next step up — caches per distinct visual state and reuses on subsequent runs.
Force Fresh
"Always regenerate, ignoring cache. May run up credits usage. Use with care."
Every run re-segments the screen from scratch. Use sparingly — useful when you suspect cache staleness is the bug you're hunting, or when the visual state genuinely differs in ways Auto's similarity check doesn't catch.
Use when: You're actively debugging and need to rule out cache staleness. Switch back to Auto once you've narrowed the issue.
Picking a mode per Tap/Scroll step
Try the modes in order — None first, then Auto if None can't find the target, Force Fresh only for debugging.
| Step | Try | Why |
|---|---|---|
| Tap on a stable nav icon | None | On-device detection is enough |
| Tap on a button described in plain language | None first, fall back to Auto | Many targets resolve on-device |
| Scroll with Until on a list | None first, fall back to Auto | Same |
| Anything you're debugging | Force Fresh (temporarily) | Rule out cache effects |
Cost framing
From cheapest to most expensive: None → Auto → Force Fresh. If your credit usage is climbing, the two highest-leverage moves are (1) flipping any Force Fresh steps back to Auto, and (2) checking whether any Auto steps would work just as well on None.
How it relates to memory
Segmentation and Memory are independent settings. A Tap step with Auto segmentation and Position Memory on is the cheapest configuration that still falls back to a fresh resolution when needed.