Server has 7 tools with mostly complete schemas and descriptions. Naming follows verb_noun convention (generate_image, process_image, analyze_*). Descriptions are present and substantive (100-300 chars typical). However, several critical gaps: (1) parameter descriptions lack constraint details (e.g., 'timeout_ms' says '1000-300000 ms' but no guidance on when to override); (2) output schemas are not formally documented, responses are JSON strings, not structured objects; (3) error handling is minimal, no recovery guidance or error classification; (4) no tool annotations (readOnlyHint, destructiveHint, idempotentHint) despite clear read/write distinctions. generate_image and composite_layers are WRITE operations but lack destructiveHint. analyze_* tools are READ-only but lack readOnlyHint. Batch tool lacks concurrency failure semantics.

• Add formal output schemas to all tools. Define return type as structured object with typed fields (e.g., {success: boolean, path: string, provider: string, model: string, revisedPrompt?: string}) instead of JSON string.
• Annotate all tools with readOnlyHint/destructiveHint/idempotentHint. Mark generate_image, process_image, generate_batch, composite_layers as destructive=true. Mark analyze_* as readOnly=true.
• Enhance error responses with recovery guidance. Return {success: false, error: {message, code, retryable: boolean}, suggestion: 'Try X instead'} for all failure paths.
• Document valid values for constrained parameters. Add enum or description for 'lang' codes in analyze_ocr (e.g., 'eng, fra, deu, spa, chi_sim, etc.'). Document operation sequences in process_image.
• Clarify batch failure semantics. Document whether generate_batch returns partial results on per-item failure, and include per-item success/failure in response.