ScreenForge defines 5 tools with complete input schemas and descriptions. Naming follows verb_noun conventions (ui_agent_*), but descriptions vary in quality and completeness. All tools have proper JSON Schema definitions with types and enum constraints. However, parameter descriptions are sparse, output schemas are undocumented, and error handling guidance is missing. The server is at the median for community tools but has gaps preventing production-grade adoption.

• Document output schema for each tool. For ui_agent_inspect_ui, define the structure of ui_tree, element_count, candidates, memory fields. For ui_agent_load_run, document resume_context and replay_assets structure. Include field types and whether fields are always present.
• Add recovery guidance to error scenarios. E.g. 'If platform connection fails, check network and retry. If vision mode times out, disable vision and retry with vision=false. If locator_type is invalid, use one of: xpath, css, text, accessibility_id.'
• Clarify parameter relationships in ui_agent_execute. Specify: 'Use EITHER workflow OR action, not both. If mode=plan_only or mode=dry_run, no state changes occur. vision mode requires image processing capability.'
• Implement proper pagination for ui_agent_load_case_memory. Add offset/cursor field to response, return total_count, and state: 'Use offset to page through results. Call repeatedly with offset+=limit until total_count is reached.'
• Standardize all descriptions to English. Translate or remove Chinese parameter descriptions so LLMs from any language background understand constraints.
• Add tool annotations to schema. Mark ui_agent_execute with 'destructiveHint': true since mode=run has side effects. Mark read-only tools with 'readOnlyHint': true.
• Add default values and bounds. Document that limit defaults to 20 in ui_agent_load_case_memory, and explicitly state the valid range (e.g. 1 - 1000).
• Include chaining IDs in responses. If ui_agent_execute returns a run_id, document that this can be passed to ui_agent_load_run. If ui_agent_inspect_ui returns candidates with locator_value, confirm that ui_agent_execute accepts these directly.

Inferred effective spec: 2025-06-18+.

• Relies on Logging (deprecated) - log to stderr or use OpenTelemetry