Scored across 31 tools
Each tool targets a genuinely distinct operation, and the descriptions cross-reference each other extensively (semantic vs pixel vs log vs stability waits are carefully separated). A few pairs could still be confused at first glance — a11y_marks vs a11y_snapshot and find_elements vs wait_for_element — though their descriptions resolve the differences.
The glass_ prefix is uniform and nearly all tools follow a lowercase verb or verb_noun pattern (click_element, list_windows, wait_for_region). Minor deviations: capabilities, logs, window, and doctor are bare nouns rather than verb_noun, and 'do' is a vague generic verb, but the overall style is predictable.
31 tools is heavy and the verbosely worded descriptions add real token cost, but the domain is unusually broad — lifecycle, window management, accessibility observation/actuation, pixel baselines, input, clipboard, logs, and batched actions all need coverage. Each tool earns its place, but several clusters (wait tools, a11y find tools, screenshot/baseline tools) could tempt consolidation.
The lifecycle is fully covered: glass_start → observe (a11y/pixels/logs) → interact (pointer/keyboard/touch/clipboard) → verify (wait_for_element/region/stable/log) → glass_stop, with doctor/capabilities for diagnostics. Minor gaps exist — no explicit exit-code query or hover-specific action — but agents can work around them, and fallback paths for a11y-less apps are built in.