cursor-bridge exposes 4 tools with reasonable descriptions (100-200 chars) and clear naming conventions (verb_noun pattern). However, input schemas are not visible in the provided source code, only parameter names and descriptions are shown. Tool definitions appear to be registered but actual JSON Schema validation is not evident. cursor_agent and cursor_resume are high-risk WRITE operations that lack explicit error handling guidance, confirmation patterns, and recovery instructions. cursor_list_skills and cursor_models are READ_ONLY discovery tools with minimal complexity. Descriptions are adequate but could be more prescriptive about when to use each tool vs. alternatives.

• Publish full JSON Schema definitions for all tool inputs, including type constraints, enums for model IDs, and validation rules for working_dir (must be absolute path, must exist or be creatable).
• Add explicit error handling documentation: 'If agent times out after 5 min, retry with a shorter prompt. If skill injection fails, call cursor_list_skills to verify skill names.'
• Implement a dry-run or preview mode for cursor_agent: return proposed changes without executing them, allowing the agent to review before committing.
• Document output schema for cursor_agent: specify that agent_id is required for cursor_resume follow-ups, include status enum (pending|running|success|failed), and clarify what 'final assistant text' contains (code diffs, explanations, or both).
• Add parameter constraints: model must be from cursor_models result, working_dir must be absolute, load_skills array must not exceed 10 items, prompt must be 10-5000 chars.