Device Screen Verification (adb → local HTTP → browser_vision)

Proven pipeline (2026-08-08, CarSah Stage 4) for seeing a connected Android device's screen and producing verifiable visual evidence. The active LLM is often text-only; browser_vision falls back to the auxiliary vision model configured for the platform — this pipeline feeds it the screenshot.

Vision model (pinned 2026-08-09): auxiliary.vision → stepfun/step-3.7-flash:free via provider nous (multimodal: text + image + video → text). Pin it explicitly in the Hermes config (auxiliary.vision.model), do not rely on auto-detect. The accuracy measurements below predate this pin (auto vision) — re-measure with the pinned model if a small-text/quality doubt resurfaces.

When to use