This server has solid structure with clear tool names, good descriptions, and explicit Pydantic schemas with ToolAnnotations. However, output schemas are undocumented (no structured response definitions visible), many optional parameters lack actionable constraints, and error handling is minimal. The tool names are action-oriented (search_, get_, download_, manage_) and follow verb_noun convention well. Descriptions are present and mostly 50-200 chars, which aligns with LLM optimization. Tool annotations (READ_ONLY_TOOL, MUTATING_TOOL, DESTRUCTIVE_TOOL) are correctly applied. The main gaps are: (1) no documented return schemas, (2) optional parameters (offset, limit, view) lack min/max constraints or detailed format guidance, (3) error responses appear minimal with no recovery guidance visible in the code shown, and (4) the 'update_setting' tool accepts free-form string keys/values with no enum constraint on valid keys despite RUNTIME_SETTING_KEYS being defined.

• Document the output schema for every tool using a structured TypedDict or Pydantic model. Example: 'Returns: {"illusts": [{"id": int, "title": str, "user": {"id": int, "name": str}, "tags": [str]}], "pagination": {"has_more": bool, "next_cursor": str|null}}'.
• Replace free-form 'key' parameter in update_setting with an enum: key ∈ {"download_path", "filename_template", "ugoira_format", "default_limit"}. This is already defined in RUNTIME_SETTING_KEYS; expose it in the schema.
• Add explicit min/max and default constraints to all pagination parameters. Example: 'offset: integer, default=0, minimum=0' and 'limit: integer, default=20, minimum=1, maximum=100'.
• For download tool, document the parameter relationship: 'Either webp_preset (for static images) or gif_preset (for animations) will apply based on the content type. gif_fps only applies when ugoira_format is "gif".'
• Implement and document error responses. Show examples: '{"error": "user_not_found", "message": "User ID 999999 not found. Try search_user() with a partial name.", "retryable": false}' and '{"error": "rate_limited", "message": "Request limit exceeded. Wait 60s before retrying.", "retryable": true}'.
• Add a dry_run parameter to download tool: 'If true, validate the request and return file counts/sizes without writing. Use this to preview before committing.'
• In the search_illust description, explicitly state the default limit and when pagination is necessary: 'Results are capped at 100 per page (default offset=0, limit=20). Use next_page or increment offset to retrieve more.'

Inferred effective spec: <=2025-11-25.

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