Laya Browser Agent

Treat them as historical diagnostics only: raw evidence stays in committed reports and release verification docs, with no current latency, throughput, cost, calibration, or size guarantee, and offline accuracy is not end-to-end task success.

The recommended default is the project's own open checkpoint selected with the browser model name and v32b subfolder, while browser-legacy selects the upstream v10s path with its pinned historical revision; see the model compatibility doc for the verified source boundary.

No. Output is validated against the offered option set before anything acts: out-of-set keys, non-summing vectors, or non-argmax choices are rejected to fail open, and output is only ever an index into a table your code built.

The package installs from a repository clone with platform extras rather than a working PyPI URL, the import and CLI name is localdecide, and the browser loop needs a separate Playwright driver install or an attached Chrome.

The split is explicit. The model decides which operation, from CLICK, TYPE_TEXT, SELECT, SCROLL, WAIT, DONE and BLOCKED; which element index to act on; and how confident it is. Your code decides what each operation means, what that index maps to in the DOM, and whether the confidence is good enough to act on. A decision model can still be wrong or confidently wrong, which is why validation, outcome checks and human confirmation stay in place.

criteria is required for every question type; the API rejects questions without it. For choice it is a map of option to rubric description, not a string, and for score it is an array of level names. A working payload was recorded against api.typesafe.ai/v1/systemone on 2026-09-22, and the same payload pointed at the bundled localdecide serve endpoint produces the same answer shape from the local Laya checkpoint, so code written against one works against the other by changing one base URL.

The documented deployment limits are authentication and password fields, long or dynamic forms, collapsed menus, canvas and shadow-DOM content, sites that block headless browsers, and language or vocabulary mismatches. The model-free test suite asserts no current model result at all, and offline MiniWoob or fixture accuracy is called a diagnostic signal rather than end-to-end browser task success. Anything stronger than that would be reading past what the repository claims.