All 5 tools have identical empty input schemas (no parameters, no properties). Descriptions are present but generic and lack actionable guidance for LLM selection. No parameter descriptions exist (N/A since no parameters). Tool names use verb_noun pattern (agent_vision_*) but lack clarity on distinctions. No output schemas documented. Error handling is minimal, only generic error messages without recovery guidance. No tool annotations (readOnlyHint, destructiveHint, idempotentHint) despite clear read/write semantics.

• Add output schemas to toolFrameResponse and toolTextResponse. Document that frame responses return {content: [{type: 'image', data: base64, mimeType, annotations: {audience, priority}}, {type: 'text', text: string}], isError: false} and text responses return {content: [{type: 'text', text: string}], isError: false}.
• Add tool annotations: agent_vision_snapshot, agent_vision_mood, agent_vision_frame should have readOnlyHint: true; agent_vision_start and agent_vision_stop should have destructiveHint: true.
• Expand descriptions to 100-200 chars each. Example for agent_vision_snapshot: 'Capture a single JPEG frame from the built-in macOS camera. Starts camera if needed, returns one frame, then stops camera only if this tool started it. Use for one-off snapshots; use agent_vision_start/frame/stop for continuous streaming.'
• Implement error classification in toolErrorResponse. Return structured errors: {content: [{type: 'text', text: 'User-friendly message'}], isError: true, errorInfo: {code: 'ERROR_CODE', retryable: boolean, suggestion: 'Next step or recovery action'}}.
• Document response field names and types in tool descriptions. E.g., 'Returns JPEG frame with metadata: {timestamp: ISO8601, width: number, height: number, meanBrightness: 0-1, jpegData: base64-encoded bytes}'.